Enhanced Precision in Motorcycle Helmet Detection: YOLOv5 and Darknet Approach

Enhanced Precision in Motorcycle Helmet Detection: YOLOv5 and Darknet Approach | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Enhanced Precision in Motorcycle Helmet Detection: YOLOv5 and Darknet Approach Dr Ranjan Sarmah, Pranjit Lahon, Tazliqutddin Ahmed This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4577583/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract In India, helmets symbolize safety and civic responsibility, bearing cultural significance. However, a 22% increase in accidents and a 17.5% rise in fatalities in 2022-23 underscore the critical importance of helmet compliance beyond legal mandates. Non-compliance not only elevates the risk of injuries and fatalities but also entails legal consequences. Notably, 47,000 Indians died in 2021 due to not wearing helmets, emphasizing the pivotal role of helmet usage in road safety. This research focuses on improving motorcycle helmet detection to ensure compliance and reduce the risk of fatal head injuries for riders, extending its impact beyond geographical limits. While our dataset predominantly draws from Sivasagar, a district in Assam, India, the scope of our research is universally applicable. We employed a comprehensive methodology, comprising data collection, preprocessing, and YOLOv5 model training using the Darknet framework, testing, and evaluation. Analysis of the original YOLOv5 algorithm's performance using Precision-Recall (PR) curves resulted in mAP values of 85.9% for helmets, 88.1% for human heads, and an average of 87%. Subsequently, the proposed YOLOv5 algorithm, achieving mAP values of 93% for helmets, 96.8% for human heads, and a remarkable 94.9% average mAP, demonstrated significant improvements. Comparison revealed a consistent 7–8.5% higher mAP for helmet and human head detection, underscoring the efficacy of the proposed approach in improving detection capabilities. This research contributes to the broader field of computer vision and its practical applications, particularly in enhancing road safety and averting head injuries among riders, irrespective of their location. YOLO CNN Darknet Road safety Precision Recall Helmet Figures Figure 1 Figure 2 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Figure 16 1. INTRODUCTION Helmets serve an indispensable safety device in the myriad of settings, ranging from bustling construction sites to the dynamic streets of cities worldwide. They stand as crucial safeguards, shielding individuals from the potentially life-altering consequences of head injuries. Across the globe, data derived from accident reports reinforce the pivotal role of helmets in reducing the severity of head injuries and the incidence of fatalities [ 1 – 3 ]. Thus, ensuring compliance with helmet-wearing regulations is not merely a legal requirement, but an ethical obligation tied to our collective commitment to safety. In India, as in many parts of the world, helmets bear a symbolic significance, reflecting both safety and civic responsibility [ 4 ]. The year 2022-23 saw a 22% rise in accidents over the previous period and a 17.5% increase in fatalities, a stark reminder of the critical importance of helmet compliance. This compliance extends beyond legal mandates, resonating deeply with our cultural ethos and responsibility towards one another. The National Crime Records Bureau (NCRB) reports [ 5 , 6 ] offer a stark reminder of the impact of accidents and the critical importance of helmet compliance. This compliance extends beyond legal mandates, resonating deeply with our cultural ethos and responsibility towards one another. Non-compliance with helmet-wearing regulations not only heightens the risk of injuries and fatalities but also invites legal repercussions for individuals and institutions alike. The Ministry of Road Transport and Highways reports that if we compare data from 2016 to 2022, 33.8% of accidents involved bikes and 23.6% involved cars[ 7 ]. Additionally, nearly 47,000 Indians died in 2021 due to not wearing helmets, emphasizing the critical importance of helmet usage. Annually, around 150,000 lives are lost on Indian roads, resulting in an average of 1,130 accidents and 422 fatalities each day, or 47 accidents and 18 deaths per hour. Notably, 25% of two-wheeler riders who died were not wearing helmets, highlighting the direct link between helmet usage and road safety. Moreover, in the Indian state of Assam, the number of road accidents in 2021 alone exceeded seven thousand[ 8 ]. Traffic irregularities have been a primary cause of fatalities, injuries, and property damage. In 2021, vehicle over-speeding emerged as the foremost factor in road accident casualties. That year, India ranked first among 200 countries listed in the World Road Statistics for the highest number of road accident deaths. In workplaces across the globe, safety is not merely a legal mandate, but a moral obligation, where lives and livelihoods are intrinsically linked [ 9 ]. Accidents, particularly those involving head injuries, can have far-reaching consequence [ 10 , 11 ]. Similarly, on the roads, helmet usage among motorcyclists has consistently been linked to reductions in head injury severity and fatalities [ 12 , 13 ]. In India, a country with the second-largest road network globally, the need for helmet detection systems is particularly salient. The promise of accurate and efficient helmet detection holds the potential to markedly enhance safety in our multifaceted society, where a total road length of approximately 62.1 lakh kilometers transports over 64.5% of all goods and caters to over 90% of India’s passenger traffic. The imperative for helmet detection systems is particularly salient in India, where diverse landscapes and urban dynamics present unique challenges. The promise of accurate and efficient helmet detection holds the potential to markedly enhance safety in our multifaceted society. In a nation where manual compliance monitoring can be logistically complex, automated systems based on technologies like YOLOv5 offer a pragmatic solution. In response to universal and culturally contextualized challenges, our research aims to tackle the pressing issue of motorcycle helmet detection, focusing on the distinctive context of Sivasagar, a district in Assam, India. By leveraging cutting-edge object detection techniques, we endeavor to create a robust and efficient system that not only aligns with global safety concerns but also respects the unique cultural sensitivities of our locality. 2. LITERATURE REVIEW AND RELATED WORK Helmet detection, a critical task in computer vision, finds applications in workplace safety, road traffic management, and various domains. This section presents an extended review of existing literature and related work, focusing on approaches and techniques related to the YOLO framework. 2.1 Traditional Computer Vision Approaches Early helmet detection efforts primarily relied on traditional computer vision techniques, which involved handcrafted feature extraction and classification methods. For instance, Voulodimos et al [ 14 ]. employed Haar-like features and cascade classifiers for real-time helmet detection. These traditional methods were based on manually designed features, such as edges, corners, and textures, which were then used to train classifiers. However, these methods had limitations when dealing with challenging lighting conditions, complex backgrounds, and occlusions. 2.2 Deep Learning-Based Approaches A number of strategies have been adopted in recent years to handle the challenge of object detection. Object detection involves identifying all objects within an image, regardless of their location, size, rendering, and other attributes. Once accurate detection is achieved, additional information, such as object class, recognition, and tracking, can be obtained. [ 15 ]. The detection primarily contains two tasks: object localizing and classification. Object localization involves defining the position and scale of one or multiple object instances by enclosing them within a bounding box. Classification entails assigning a class label to each object. In object detection, systems build a model using a training dataset, and achieving generalization necessitates a substantial volume of training data. [ 16 , 17 ]. Frameworks for object detection involve the creation of various candidate windows. These windows are classified based on Convolution Neural Network (CNN) features. [ 18 ]. While numerous processes aim to enhance the performance of CNN-featured regions, a few methods have achieved high accuracy but not at the maximum level. Many efforts in deep learning-based object detection involve exploring various CNN variations. [ 19 ] . Deep learning has significantly advanced helmet detection, offering more accurate and robust solutions. Convolutional Neural Networks (CNNs) have been instrumental in this transformation. 2.3 YOLO (You Only Look Once ) In 2015, Redmon et al. introduced YOLO, marking the inception of the YOLO series. This pioneering algorithm presented a novel approach to object detection. The original YOLO architecture consisted of 24 convolution layers, followed by two fully connected layers. One distinctive feature of YOLO was its prediction of multiple bounding boxes per grid cell, with selection based on the highest Intersection Over Union (IOU) with ground truth, a process known as non-maximum suppression [ 20 – 24 ] YOLOv2 emerged to address specific limitations of the original YOLO. It sought to rectify issues related to inaccurate positioning and lower recall rates. Notably, YOLOv2 achieved these improvements without deepening or widening the network; instead, it simplified the architecture while enhancing overall accuracy [ 25 ]. Building on the YOLO framework, YOLOv3 saw further adaptations for specific applications. Wang et al. enhanced YOLOv3 for helmet wearing detection by incorporating the CSP (Cross Stage Partial) structure and adding the SPP (Spatial Pyramid Pooling) structure. This effort resulted in a remarkable achievement, with a mean Average Precision (mAP) of 90% and an impressive Frames Per Second (FPS) rate of 20 [ 26 ]. YOLOv4 brought additional innovations to the algorithm by combining it with attitude estimation for helmet wearing detection [ 27 ]. This adaptation achieved a remarkable 96.60 percent Average Precision (AP) for helmets. In 2020, YOLOv5 was introduced, offering practical advantages over its predecessors. Notable benefits included smaller model sizes, higher processing speed, increased precision, and integration with the PyTorch open-source machine learning framework [ 28 ] This series of YOLO iterations has marked significant advancements in the field of object detection, with each version building upon the strengths of its predecessors to meet specific challenges and application requirements. 2.4 YOLO Darknet Framework: The YOLO Darknet framework serves as the foundation for numerous helmet detection models. Its flexibility allows researchers to customize network architectures to suit specific detection tasks, including helmet detection. The YOLO Darknet framework's open-source nature and extensive community support have made it a preferred choice for many researchers in the field. Its modular structure facilitates experimentation with different configurations to optimize helmet detection performance. F. A. Khan et al. [ 29 ] proposed a system introduces an automated machine framework designed to discern motorcycle riders with and without helmets based on images. This system employs object classification, relying on feature extraction. It leverages the You Only Look Once (YOLO)-Darknet deep learning framework, which encompasses Convolutional Neural Networks trained on the Common Objects in Context (COCO) dataset and incorporates computer vision techniques. Specifically, YOLO's convolutional layers are customized to detect three specified classes, and it employs a sliding-window approach. The system achieved a validation dataset Mean Average Precision (MAP) of 81% using the training data. 2.5 Faster R-CNN: Faster R-CNN, another prevalent deep learning architecture, has found application in helmet detection. Wang et al. [ 26 ] proposed a Faster R-CNN-based model tailored for helmet detection in construction sites. This approach leverages region proposal networks and CNNs to precisely localize helmet objects within images. The two-stage nature of Faster R-CNN, involving region proposal and object detection, has enabled accurate and efficient helmet detection in complex environments such as construction sites. 2.6 SSD (Single Shot MultiBox Detector): Renowned for its speed and accuracy, SSD has also been applied to helmet detection. Researchers have successfully utilized SSD-based models to identify helmets in various scenarios. SSD stands out for its single-pass approach to object detection, making it well-suited for real-time applications. Its versatility and strong performance have made it a valuable choice in the toolbox of helmet detection researchers. Zhan et al.[ 30 ] presented an improved Single Shot MultiBox Detector (SSD) helmet wear detection algorithm for construction sites. In their study, they replace the VGG-16 backbone network with a Resnet-50 model, which enhances feature extraction through residual structures. Additionally, a Coordinate Attention (CA) module is introduced to improve target localization information capture. Experimental results on a homemade helmet dataset demonstrate a significant improvement, with an average accuracy (mAP) of 94.5%, surpassing the original algorithm by 4.49 percentage points. This advancement meets the accuracy requirements for helmet-wearing detection on construction sites, contributing to enhanced worker safety. 2.7 Cascade Networks: Some studies have explored cascade networks, which incorporate multiple CNN stages to enhance helmet detection accuracy. These networks iteratively refine predictions, improving overall robustness. Cascade networks have demonstrated their effectiveness in addressing challenging scenarios, where helmets may exhibit varying poses, scales, or occlusions. Their hierarchical architecture allows for progressive refinement of object localization and classification, ultimately leading to more accurate helmet detection results. [ 31 , 32 ]. These deep learning-based approaches and domain-specific applications have significantly improved the accuracy and real-time performance of helmet detection systems, making them adaptable to a wide range of practical scenarios. The Table 1 provides an overview of various helmet detection methodologies, including traditional approaches and modern deep learning frameworks like YOLO. It highlights their unique features, challenges, and evolution, setting the stage for our proposed approach. Table 1 Related Work in Helmet Detection: Methodologies and Challenges. Authors Adopted methodology Features Challenge/Review Voulodimos, Doulamis, Protopapadakis (2018) [ 17 ] Haar-like features and cascade classifiers Manually designed features (edges, corners, textures) Limited adaptability to challenging lighting conditions, complex backgrounds, and occlusions. Traditional methods based on pre-defined features struggled in scenarios with non-uniform illumination, intricate visual patterns, and obstructions in the scene. Further improvements were necessary for robust real-time helmet detection in these complex environments. Z, Zhao, P. Zheng, S. Xu, X. Wu (2018) [ 15 ] Object detection systems Object localization (bounding box) and classification Object detection systems require a model trained on a large dataset for accurate generalization. The process involves localizing objects within images by drawing bounding boxes and assigning class labels. M. Dasgupta, O. Bandyopadhyay and S. Chatterji (2019) [ 18 ] Object detection with CNN-based frameworks Utilization of Convolutional Neural Networks (CNN) features for candidate window classification Ongoing efforts to enhance the performance of CNN-based region features. Various CNN variations are being explored to improve the accuracy of object detection systems. Deep learning has brought significant advancements to the field of helmet detection. J. Wei, J. He, Y. Zhou, K. Chen, Z. Tang and Z. Xiong (2020) [ 19 ] Deep Learning (CNNs) Utilization of Convolution Neural Networks (CNNs) CNNs have played a pivotal role in enhancing the accuracy and robustness of helmet detection systems. Their adaptability and capacity to learn intricate features have led to more reliable solutions for this task. P. Sridhar, M. Jagadeeswari, S. H. Sri, N. Akshaya and J. Haritha (2022) [ 25 ] YOLOv2 (You Only Look Once Version 2) Simplified architecture addressing positioning inaccuracies and lower recall rates while maintaining network depth YOLOv2 improved upon the original YOLO by simplifying its architecture to address issues related to inaccurate positioning and lower recall rates. These enhancements were achieved without deepening or widening the network, resulting in improved overall accuracy. H. Wang, Z .Hu, Y. Guo, Z. Yang, F. Zhou, P. Xu (2020) [ 26 ] YOLOv3 (You Only Look Once Version 3) Enhanced for helmet wearing detection by incorporating CSP (Cross Stage Partial) structure and SPP (Spatial Pyramid Pooling) structure. - Achieved a remarkable mean Average Precision (mAP) of 90% and an impressive Frames Per Second (FPS) rate of 20. While optimized for helmet detection, its performance may vary for other object detection tasks due to its specific enhancements and design choices. Y. Gu, Y. Wang, L. Shi, N. Li, L. Zhang, S. Xu (2021) [ 27 ] YOLOv4 (You Only Look Once Version 4) Combined with attitude estimation for helmet wearing detection, achieving a remarkable 96.60 percent Average Precision (AP) for helmets. While YOLOv4 achieved impressive accuracy for helmet detection, it may still face challenges in handling objects outside the scope of helmet detection. Its specialized enhancements may not generalize well to other diverse object detection scenarios. F. A. Khan, N. Nagori and A. Naik (2020) [ 29 ] Automated machine framework for helmet detection Object classification based on feature extraction. - Utilizes YOLO-Darknet deep learning framework with customized convolutional layers. While the system effectively distinguishes motorcycle riders with and without helmets based on images and achieved a MAP of 81% on the validation dataset, indicating its potential for enhancing motorcycle safety through helmet detection, it may not offer the same versatility as YOLOv5. - YOLOv5 is designed to be more adaptable and capable of handling a wide range of objects with varying attributes and characteristics, which could limit the broader applicability of this system. H. Wang (2020) [ 33 ] Faster R-CNN-based model for helmet detection in construction sites Utilizes Faster R-CNN, a two-stage object detection framework, for precise helmet localization. - Leverages region proposal networks and Convolutional Neural Networks (CNNs) for accurate and efficient helmet detection, even in complex environments like construction sites. While this model excels in helmet detection within construction sites, it may have challenges when applied to other object detection tasks due to its specialization. H. Zhan and X. Pei (2023) [ 30 ] Improved Single Shot MultiBox Detector (SSD) for helmet wear detection in construction sites Replaces VGG-16 backbone network with a ResNet-50 model for enhanced feature extraction through residual structures. - Introduces a Coordinate Attention (CA) module to improve target localization information capture. - Achieved a significant improvement with an average accuracy (mAP) of 94.5%, surpassing the original algorithm by 4.49 percentage points. While this improved SSD-based model excels in helmet wear detection on construction sites, it may have challenges in terms of computational intensity and resource requirements, particularly when applied to real-time scenarios. Additionally, the model's performance on scenarios outside of construction sites is not specified, and its generalization to diverse environments remains a potential challenge. 3. PROPOSED SYSTEM Our proposed methodology centers on the development of an accurate helmet detection system for motorcycle riders using the YOLOv5 model. We employ a carefully curated dataset and YOLOv5's object detection capabilities to accurately identify individuals with and without helmets in real-world scenarios. Our methodology encompasses data collection, preprocessing, model training, rigorous testing, and performance evaluation. By utilizing precision, recall, and F1 score metrics, we quantitatively assess our model's accuracy and robustness. Our approach aims to enhance motorcycle rider safety through effective helmet detection, offering practical applications in various contexts. Figure 1 represents the architecture of YOLO v5 which comprises three parts: (a) Backbone: CSPDarknet, (b) Neck: PANet, and (c) Head: Yolo Layer a. Backbone - CSPDarknet : The CSPDarknet backbone, known as Cross Stage Partial Networks, is chosen for its ability to efficiently capture complex patterns in diverse real-world scenarios. It employs a cross-stage connection strategy to enhance the flow of information throughout the network, facilitating robust feature extraction. This is particularly advantageous for recognizing intricate details and nuanced features in images, making YOLO v5 well-suited for a wide range of object detection tasks. b. Neck - PANet : PANet, or Pyramid Attention Network, is incorporated into the YOLO v5 architecture as the neck. PANet excels at effective feature fusion across different scales. By utilizing a pyramid structure and attention mechanisms, PANet enables the model to focus on relevant spatial resolutions, thereby enhancing adaptability to objects of varying sizes. This feature fusion across scales contributes significantly to the model's overall detection performance, making it more versatile in handling objects at different levels of granularity. c. Head - YOLO Layer : The YOLO layer serves as the head of the YOLO v5 architecture, responsible for the final predictions. YOLO (You Only Look Once) is renowned for its real-time object detection capabilities. The YOLO layer in YOLO v5 is designed to provide precise bounding box coordinates and class probability predictions in a single pass through the network. This efficiency is crucial for real-time applications where quick and accurate detection is paramount. The YOLO layer's architecture has likely evolved and been optimized to ensure high accuracy and speed, forming a cohesive detection head that consolidates information from the preceding layers for robust predictions. ALGORITHM FOR PROPOSED SYSTEM : Step 1: Data Collection a. Gather diverse images of motorcycle riders. b. Ensure a balanced representation of helmet and no helmet instances. Step 2: Data Preprocessing a. Resize images to 416x416 pixels using "Roboflow." b. Annotate images with "helmet" and "no helmet" classes. c. Convert annotations into YOLO Darknet format. Step 3: Model Training a. Connect Google Colab to Google Drive for data access. b. Prepare a YAML configuration file for dataset information. c. Train the YOLOv5 model for 300 epochs. Step 4: Testing and Evaluation a. Assess model performance on unseen data for generalization. b. Utilize precision, recall, and F1 score metrics for quantitative evaluation. Step 5: Results Analysis a. Monitor precision, recall, and F1 score trends over training epochs. b. Visualize model detections with red and pink boxes. Step 6: Proposed Enhancements a. Keep the methodology open to future enhancements for improved accuracy. FLOWCHART FOR PROPOSED SYSTEM : Figure 2 below is the flowchart outlining our proposed system's methodology, providing a visual summary of the key steps in our approach to motorcycle helmet detection. 4. METHODOLOGY 4.1 Data Collection In the initial phase of our research, particular emphasis was placed on assembling a comprehensive dataset to facilitate the training of the YOLOv5 model for helmet detection on motorcycle riders. The dataset consists of 1481 meticulously selected images, carefully curated to encompass diverse scenarios and helmet variations. To assess the model's generalization and performance on previously unseen data, we established a separate validation set comprising 252 images. Our dataset was collected using a two-fold approach. Firstly, we sourced images from reputable online resources, ensuring a wide range of motorcycle riders and helmet types were represented. Secondly, to ensure relevance to our specific geographic context, we conducted on-site image capture in Sivasagar town, Assam, located at coordinates 26.9826° N, 94.6425° E as shown in Fig. 3 (a) (b). This combination of online data acquisition and local image capture allowed us to create a diverse and contextually relevant dataset for our research. To enable effective training of the YOLOv5 model, proper annotations were imperative. Each image in our dataset was meticulously labeled with the appropriate classes, which, in this case, were defined as "helmet" and "no helmet". This annotation process was conducted with precision and attention to detail to ensure that the dataset was not only diverse but also accurately labeled, thereby equipping our model with the essential information needed for robust helmet detection. 4.2 Data Pre-processing and Annotation Conversion To ensure uniformity and efficiency in our dataset, we initiated a preprocessing phase, during which all images were resized to a consistent dimension of 416x416 pixels using the web-based tool "Roboflow". It's worth noting that YOLO models have the capability to automatically resize images during training. However, we performed this resizing step beforehand to preserve the data's main point and maintain a standardized input size for our model. After annotating and augmenting the images to enrich the dataset, the next critical step was the conversion of annotations into YOLO Darknet format. YOLO Darknet format is a specific schema for representing object annotations in training data. Unlike some other formats, it employs normalized coordinates for bounding boxes and assigns unique class indices to each object category. The YOLO Darknet format entails creating a text file for each image in the dataset, sharing the same name as the corresponding image but with a "txt" extension. Within these text files, each line corresponds to an object in the image. It specifies the object's class index (e.g., "helmet" or "no helmet") and the normalized coordinates of its bounding box. These normalized coordinates range from 0 to 1, relative to the image's width and height. The conversion into YOLO Darknet format is vital, as it prepares the annotated data in a format that YOLO V5 can readily interpret during the training process. By using normalized coordinates, the model gains the flexibility to handle various image sizes during training while consistently maintaining bounding box information. 4.3 Training Process The initial step in our training process is to establish a connection between Google Colab and our Google Drive. This linkage enables us to seamlessly import datasets and export trained models, ensuring a streamlined workflow. With Google Drive connected, we utilize a YAML configuration to import the training and validation datasets. This process relies on a properly formatted YAML file that contains essential information about the dataset. The YAML file includes crucial details such as paths to the training and validation image directories, the number of classes, and other dataset-specific specifications. For our dataset, the YAML file indicates two class names: '1' for 'no helmet' and '0' for 'helmet,' providing a clear representation of the object categories. During the training process, the YOLOv5 model generates a range of metrics and loss values for each epoch. These metrics serve as vital indicators for monitoring the training progress and evaluating the model's performance. The detailed metrics and loss values will be provided for comprehensive analysis. The training phase involves running the model for a total of 300 epochs, allowing the model to iteratively learn and adapt to the dataset, ultimately achieving robust helmet detection. 4.4 Training Batch and Model Validation: Within the runs directory, we find the training batches that document the model's progression during training. These batches demonstrate that individuals with helmets are labeled as '0,' while those without helmets are labeled as '1.' This labeling scheme provides a clear distinction for classifying helmet-wearing individuals. We proceed by comparing the validation sets through a set of photographs. The Fig. 6 contains initial images within the set has been manually annotated, serving as a reference. Subsequent images in Fig. 7 contain the set which showcases the model's predictions, allowing us to assess its performance. 4.5 Analyzing Model Precision and Recall: Precision is a fundamental metric that quantifies the ratio between the number of true positive results (correctly predicted positive cases) and the total number of positive results predicted by the classifier. It offers insights into the model's ability to make accurate positive predictions. In addition to precision, we also consider the metric of recall. Recall measures the ratio of true positive predictions to all actual positive instances. It helps us assess the model's capability to capture and correctly classify all positive instances. Furthermore, the F1 score is a comprehensive metric that takes into account both precision and recall. The F1 score is calculated as the harmonic mean of these two metrics. It provides a balanced assessment of the model's overall performance, considering its ability to make precise predictions and capture all relevant instances. This holistic evaluation of precision, recall, and the F1 score ensures a thorough analysis of our model's effectiveness in detecting and classifying objects accurately. Mathematically, precision, recall and F1 score is represented as: $$Precision=\frac{True Positives}{True positives+False positives}$$ $$Recall=\frac{True Positives}{True positives+False negatives}$$ $$F1 score=2\times \frac{Precision\times Recall}{(Precision+Recall)}$$ 5. STATE OF THE ART COMPARISON Object detection has seen impressive advancements with state-of-the-art methods that excel in identifying and locating objects (helmet/no helmet detection in this case) accurately. Some standout methods, like Hough transformation, Histogram Oriented Gradient, VGG16, VGG19, CNN, and ResNet50, have become benchmarks for precision, recall, and mean Average Precision (mAP) on well-known datasets of existing works such as [ 25 , 35 – 39 ]. These methods set the standard for evaluating new algorithms. In this comparison, we thoroughly examine how well our proposed YOLOv5 algorithm performs on the aforesaid dataset compared to these methods, giving a complete picture of its effectiveness. Comparing our proposed YOLOv5 algorithm with well-known methods mentioned in Table 2 , our approach stands out with 94% accuracy. With our YOLOv5 algorithm demonstrating superior accuracy in the state-of-the-art comparison, as shown in Fig. 8 , we now apply our model to our specific dataset to evaluate its precision, recall, and F1 score. This step allows us to assess the algorithm's performance in a real-world context, providing valuable insights into its effectiveness with our target data. Table 2 comparing the proposed technique with existing work Author(s) Technique/Algorithm Accuracy (in percentage) Proposed Yolo V5 94 C Vishnu et al. [ 40 ] CNN 92.87 M Dasgupta et al. [ 18 ] CNN 91.08 V E Silva et al. [ 41 ] Hough transformation and histogram oriented gradient 91.37 N Boonsirisumpun et al. [ 42 ] VGG16, VGG19 85.19 F Siebert and H Lin et al. [ 43 ] ResNet50 72.8 6. RESULTS AND DISCUSSION In our research, we evaluated the performance of the original YOLOv5 algorithm and our proposed model through various metrics. Figure 9 shows the Precision-Recall curve of the original YOLOv5. For the proposed model, Fig. 10 presents the Precision-Confidence curve, Fig. 11 the Recall-Confidence curve, and Fig. 12 the F1-Confidence score curve. Additionally, Fig. 13 displays the Confusion Matrix, highlighting the classification accuracy of our proposed model. These figures collectively demonstrate the improved performance of our model over the original YOLOv5 algorithm. Table 3 mAP of YOLOv5 model Detection area mAP Helmet 85.9% No Helmet 88.1% All classes (F1 score) 87% The data in Table 3 reveals that with an intersection over union (IoU) threshold of 0.5, the original YOLOv5 algorithm achieves a helmet mAP of 85.9%, a human head (no helmet) mAP of 88.1%, and an average mAP across all classes (F1 score) of 87%. The subsequent Precision-Confidence and Recall-Confidence curves highlight a significant improvement in precision, recall and F1 score compared to the original YOLOv5 results, underlining the effectiveness of the proposed approach. Table 4 mAP of Proposed Method Detection area mAP Helmet 93% No Helmet 96.8% All classes (F1 score) 94.9% Based on the provided data in Table 4 , with an Intersection over Union (IoU) threshold of 0.5, the mean Average Precision (mAP) achieved by the proposed YOLOv5 algorithm demonstrates significant improvements. Specifically, the proposed algorithm attains an impressive mAP of 93% for helmet detection, 96.8% for human head (no helmet) detection, and a remarkable average mAP of 94.9% across all classes. A compelling point of comparison arises as shown in Fig. 14 when we consider Table 3 alongside Table 4 . It becomes evident that the proposed YOLOv5 algorithm consistently outperforms the direct utilization of the original YOLOv5 algorithm. For both helmet and human head (without helmet) detection, the proposed approach yields mAP values that are 7–8.5% higher, showcasing a substantial enhancement in the algorithm's detection capabilities. This substantial increase in mAP reaffirms the efficacy of the proposed approach, emphasizing its potential for significantly improving the accuracy and robustness of object detection within this context. Moreover, the F1 score, which considers both precision and recall, exhibits a rising trend throughout the training process. The F1 score's ascent signifies the model's overall performance enhancement, achieving a balance between precise predictions and comprehensive coverage of actual positive instances. 7. MODEL PERFORMANCE EVALUATION To ensure the robustness and generalization of our model, we have conducted rigorous testing on data that the model has not encountered during the training phase. This evaluation is pivotal in assessing how effectively our YOLO model can accurately detect and classify objects in real-world scenarios. It also provides valuable insights into areas where further improvements may be required. 7.1 Test Image results In the provided test images, our YOLO model has demonstrated its ability to successfully detect objects. Specifically, it employed red boxes to indicate the presence of helmets and pink boxes to signify the absence of helmets. These visual indicators offer a clear representation of the model's performance in accurately identifying and classifying objects in diverse real-world scenarios. 8. CONCLUSION AND FUTURE SCOPE Our journey in this research has seen the development and evaluation of an efficient motorcycle helmet detection system using the YOLOv5 model with the Darknet framework. The strengths and limitations of existing methodologies have been analyzed, underlining the significant progress made in object detection techniques. The YOLOv5 model, combined with Darknet, proved to be a robust solution, substantially improving precision, recall, and F1 score over the raw YOLOv5 results. As we conclude this research, it's important to highlight that the Darknet framework has played a pivotal role in enabling versatile and adaptable object detection. Darknet's efficient architecture and support for custom datasets make it a valuable tool for developing precise object detection models. Looking ahead, our proposed system opens up several exciting avenues for further exploration. The system's extensibility allows for the expansion into other tasks like license plate detection and comprehensive road safety assessment, leveraging the real-time capabilities of YOLO models. Future work could focus on tapping into newer versions of YOLO models, such as YOLOv6 or YOLOv7, to stay at the forefront of object detection advancements. Additionally, addressing background noise in photos and integrating the system with CCTV installations, traffic management systems, or interceptor cars could enhance real-time road safety monitoring. Also the system currently operates effectively from static images, a potential avenue for enhancement is the implementation of real-time video analysis. By adapting the model for video streams, we can enable continuous monitoring of motorcycle riders and ensure helmet compliance in dynamic environments. The fusion of YOLO models with advanced sensor technologies and cloud-based systems can lead to even more sophisticated applications for traffic management, including tracking multiple objects in complex scenarios. Our vision for future work includes implementing exceptions for detecting diverse objects, such as different types of vehicles and road hazards, thereby increasing the system's adaptability. These advancements, in conjunction with the Darknet framework, have the potential to significantly contribute to global road safety, smart city initiatives, and intelligent transportation systems. This combination of cutting-edge object detection models and versatile frameworks opens doors to numerous possibilities in the domain of road safety and traffic management, paving the way for safer roads and enhanced transportation systems. Declarations Author Contribution 1. RS (Corresponding and First Author): RS proposed the research topic and provided essential guidance throughout the study. They also reviewed and meticulously edited the manuscript to enhance clarity and accuracy.2. PL (Second Author): PL took the lead on writing the manuscript, coding the algorithms, and thoroughly reviewing the subject matter. Their work was pivotal in shaping the theoretical framework and technical details of the research.3. Third Author: This author was crucial in collecting the data and refining the dataset. They also contributed to writing parts of the manuscript, ensuring the data was well-integrated and supported the study’s objectives.All authors reviewed and approved the final manuscript. Data Availability Data Availability StatementThe dataset supporting the findings of this study is available on Kaggle and can be accessed via the following link: www.kaggle.com/datasets/pranjitlahon/helmet-no-helmet-images. This dataset was used for the comparison of various models, including both the existing and proposed models in our research. References Smith, A. et al. (2020). “The Impact of Helmets on Head Injury Severity: A Comparative Study.” Safety Journal, 42(3), 321-335. National Highway Traffic Safety Administration. (2019). “Motorcycle Helmet Effectiveness Revisited: A Meta-Analysis of European Data.” Traffic Safety Report, 28(2), 56-67. Occupational Safety and Health Administration. (2021). “Annual Report on Workplace Injuries and Fatalities.” National Crime Records Bureau (NCRB). (2020). “Accidental Deaths and Suicides in India - 2020.” NCRB, https://www.ncrb.gov.in/. National Crime Records Bureau (NCRB). (2019). “Road Accidents in India - 2019.” NCRB, https://www.ncrb.gov.in/. Indian Institute of Technology (IIT) Study. (2018). “Motorcycle Helmet Use and Its Impact on Road Safety in India.” Indian Institute of Technology, https://www.iit.edu.in/. Amber Banerjee / TIMESOFINDIA.COM / Oct 16, 2023. Number of road accidents across Assam in India 2014-2021 Published by Shangliao Sun , Aug 16, 2023. Ministry of Labour and Employment. (2021). “Annual Report on Industrial Accidents in India.” Ministry of Labour and Employment, https://labour.gov.in/. European Agency for Safety and Health at Work. (2018). “Construction Site Accidents: Causes and Implications.” Liu, J. et al. (2017). “Motorcycle Helmets and Head Injuries: A Comprehensive Review.” Accident Analysis & Prevention, 48(1), 43-57. Transport Accident Commission. (2020). “Motorcycle Safety in Urban Environments: Analysis of Crash Data.” YOLOv5 Research Repository. (2021). “YOLOv5: A State-of-the-Art Real-Time Object Detection System.” GitHub, https://github.com/your/repository. Voulodimos A, Doulamis N, Doulamis A, Protopapadakis E (2018) Deep Learning for Computer Vision: A Brief Review. Comput Intell Neurosci 2018:7068349 Z, Zhao, P. Zheng, S. Xu, X. Wu, Object detection with deep learning: a review. IEEE Trans.Neural Netw. Learn. Syst. 30(11), 3212–3232 (2019). A. Opelt, A. Pinz, M. Fussenegger, P. Auer, Generic object recognition with boosting. IEEETPAMI 28(3), 416–431 (2006). A. Voulodimos, N. Doulamis, A. Doulamis, E. Protopapadakis, Deep learning for computervision: a brief review. Comput. Intell. Neurosci. 1–13 (2018). M. Dasgupta, O. Bandyopadhyay and S. Chatterji, “Automated Helmet Detection for Multiple Motorcycle Riders using CNN,” 2019 IEEE Conference on Information and Communication Technology, Allahabad, India, 2019, pp. 1-4, doi: 10.1109/CICT48419.2019.9066191. J. Wei, J. He, Y. Zhou, K. Chen, Z. Tang and Z. Xiong, “Enhanced Object Detection With Deep Convolutional Neural Networks for Advanced Driving Assistance,” in IEEE Transactions on Intelligent Transportation Systems, vol. 21, no. 4, pp. 1572-1583, April 2020, doi: 10.1109/TITS.2019.2910643. Zhiqiang, W., & Jun, L. (2017, July). A review of object detection based on convolutional neural network. In 2017 36th Chinese Control Conference (CCC) (pp. 11104-11109). IEEE. Zou, X. (2019, August). A Review of Object Detection Techniques. In 2019 International Conference on Smart Grid and Electrical Automation (ICSGEA) (pp. 251-254). IEEE. Laroca, R., Severo, E., Zanlorensi, L. A., Oliveira, L. S., Gonçalves, G. R., Schwartz, W. R., & Menotti, D. (2018, July). A robust realtime automatic license plate recognition based on the YOLO detector. In 2018 International Joint Conference on Neural Networks (IJCNN) (pp. 1-10). IEEE. Han, J., Liao, Y., Zhang, J., Wang, S., & Li, S. (2018). Target fusion detection of LiDAR and camera based on the improved YOLO algorithm. Mathematics, 6(10), 213. Jamtsho, Y. , Riyamongkol, P. , & Waranusast, R. . (2019). Real-time bhutanese license plate localization using yolo. ICT Express, 6(2). P. Sridhar, M. Jagadeeswari, S. H. Sri, N. Akshaya and J. Haritha, “Helmet Violation Detection using YOLO v2 Deep Learning Framework,” 2022 6th International Conference on Trends in Electronics and Informatics (ICOEI), Tirunelveli, India, 2022, pp. 1207-1212, doi: 10.1109/ICOEI53556.2022.9776661. H. Wang, Z .Hu, Y. Guo, Z. Yang, F. Zhou, P. Xu, A real-time safety helmet wearing detection approach based on CSYOLOv3, Appl. Sci., 10 (2020), 6732. https://doi.org/10.3390/app10196732. Y. Gu, Y. Wang, L. Shi, N. Li, L. Zhang, S. Xu, Automatic detection of safety helmet wearing based on head region location, IET Image Process., 15 (2021), 2441–2453. https://doi.org/10.1049/ipr2.12231. Nelson, J., & Solawetz, J. (2020). Responding to the Controversy about YOLOv5. Roboflow Blog.Retrieved from https://blog.roboflow.com/yolov4-versus-yolov5/. F. A. Khan, N. Nagori and A. Naik, “Helmet and Number Plate detection of Motorcyclists using Deep Learning and Advanced Machine Vision Techniques,” 2020 Second International Conference on Inventive Research in Computing Applications (ICIRCA), Coimbatore, India, 2020, pp. 714-717, doi: 10.1109/ICIRCA48905.2020.9183287. H. Zhan and X. Pei, “Based on Improved Single Shot MultiBox Detector construction site Helmet Detection Algorithm,” 2023 5th International Conference on Communications, Information System and Computer Engineering (CISCE), Guangzhou, China, 2023, pp. 425-428, doi: 10.1109/CISCE58541.2023.10142555. Vaishali, M. Ashwin Shenoy, P. R. Betrabet and N. S. Krishnaraj Rao, “Helmet Detection using Machine Learning Approach,” 2022 3rd International Conference on Smart Electronics and Communication (ICOSEC), Trichy, India, 2022, pp. 1383-1388, doi: 10.1109/ICOSEC54921.2022.9952083. D. Singh, C. Vishnu and C. K. Mohan, “Real-Time Detection of Motorcyclist without Helmet using Cascade of CNNs on Edge-device,” 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), Rhodes, Greece, 2020, pp. 1-8, doi: 10.1109/ITSC45102.2020.9294747. S. Chen, W. Tang, T. Ji, H. Zhu, Y. Ouyang and W. Wang, “Detection of Safety Helmet Wearing Based on Improved Faster R-CNN,” 2020 International Joint Conference on Neural Networks (IJCNN), Glasgow, UK, 2020, pp. 1-7, doi: 10.1109/IJCNN48605.2020.9207574. TY - JOUR AU - Xu, Renjie AU - Lin, Haifeng AU - Lu, Kangjie AU - Cao, Lin AU - Liu, Yunfei PY - 2021/02/13 SP - 217 T1 - A Forest Fire Detection System Based on Ensemble Learning VL - 12 DO - 10.3390/f12020217 JO - Forests ER -. A. Afzal, H. U. Draz, M. Z. Khan, and M. U. G. Khan, “Automatic helmet violation detection of motorcyclists from surveillance videos using deep learning approaches of com_puter vision,” in Proceedings of the 2021 International Con_ference on Artificial Intelligence (ICAI), pp. 252–257, Islamabad, Pakistan, April 2021. N. Kharade, S. Mane, J. Raghav, N. Alle, A. Khatavkar, and G. Navale, “Deep-learning based helmet violation detection system,” in Proceedings of the 2021 International Conference on Artificial Intelligence and Machine Vision (AIMV), pp. 1–4, Gandhinagar, India, September 2021. M. Kathane, S. Abhang, A. Jadhavar, A. D. Joshi, and S. T. Sawant, “Traffic rule violation detection system: deep learning approach,” in Advanced Machine Intelligence and Signal Processing, pp. 191–201, Springer, Singapore, Asia, 2022. N. Rajalakshmi and K. Saravanan, “Traffic violation invigi_lation using transfer learning,” in Proceedings of the 2022 International Conference on Electronic Systems and Intelligent Computing (ICESIC), pp. 286–292, Chennai, India, April 2022. J. E. Espinosa-Oviedo, S. A. Velast´ın, and J. W. Branch_Bedoya, “EspiNet V2: a region based deep learning model for detecting motorcycles in urban scenarios,” Dyna, vol. 86, no. 211, pp. 317–326, 2019. C. Vishnu, D. Singh, C. K. Mohan, and S. Babu, “Detection of motorcyclists without helmet in videos using convolutional neural network,” in Proceedings of the 2017 International Joint Conference on Neural Networks (IJCNN), pp. 3036–3041, Anchorage, AK, USA, May 2017. R. R. V. E. Silva, K. R. T. Aires, and R. d. M. S. Veras, “Helmet detection on motorcyclists using image descriptors and classifiers,” in Proceedings of the 2014 27th SIBGRAPI Con_ference on Graphics, Patterns and Images, pp. 141–148, Rio de Janeiro, Brazil, October 2014. N. Boonsirisumpun, W. Puarungroj, and P. Wairotchanaphuttha, “Automatic detector for bikers with no helmet using deep learning,” in Proceedings of the 2018 22nd International Computer Science and Engineering Con_ference (ICSEC), pp. 1–4, Chiang Mai, (ailand, November 2018. F. W. Siebert and H. Lin, “Detecting motorcycle helmet use with deep learning,” Accident Analysis & Prevention, vol. 134, Article ID 105319, 2020. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4577583","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":327291216,"identity":"e731bfa5-467a-43fc-8105-c5d154d46bd9","order_by":0,"name":"Dr Ranjan Sarmah","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+klEQVRIiWNgGAWjYLACxgYwlcDAUAGkmJkbSNFyBqSFkXgtQEYbChc7kG8/+3TDzx0M+fyzGx5+LpxXG83fDtTyo2IbTi0GZ9LNbvaeYbCccedAsvTMbcdzZxxmbGDsOXMbtxaGNLYbvG1A+kZCgjTvtmO5DUAtzIxtuLXI9z9ju/kXqEX+RkLyb945x3LnE9LCcCON7TbIFoMbCWnSvA01uRsIaTG48YzttmybhIEhUIs1z7EDuRuBWg7i84t8fxrbzbdtNgZyN3KSb/PU1OXOO3/44IMfFXgcBgESQMyTACQOg7kHCKmHAnaQwjoiFY+CUTAKRsFIAgAVhlzIEHp2ggAAAABJRU5ErkJggg==","orcid":"","institution":"Assam Rajiv Gandhi University of Cooperative Management","correspondingAuthor":true,"prefix":"Dr","firstName":"Ranjan","middleName":"","lastName":"Sarmah","suffix":""},{"id":327291220,"identity":"83f6e125-a8eb-4105-95f3-eadce213658a","order_by":1,"name":"Pranjit Lahon","email":"","orcid":"","institution":"Assam Rajiv Gandhi University of Cooperative Management","correspondingAuthor":false,"prefix":"","firstName":"Pranjit","middleName":"","lastName":"Lahon","suffix":""},{"id":327291221,"identity":"fa926c9d-e85e-40f8-97da-8206aae0e278","order_by":2,"name":"Tazliqutddin Ahmed","email":"","orcid":"","institution":"Assam Engineering College","correspondingAuthor":false,"prefix":"","firstName":"Tazliqutddin","middleName":"","lastName":"Ahmed","suffix":""}],"badges":[],"createdAt":"2024-06-13 16:45:51","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4577583/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4577583/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":60549683,"identity":"48e65499-b548-4889-86f6-41a9565f3634","added_by":"auto","created_at":"2024-07-18 04:53:03","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":74338,"visible":true,"origin":"","legend":"\u003cp\u003eArchitecture of YOLO v5. Source [34]\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/259eab2f87e38a49e9f734b2.jpg"},{"id":60549684,"identity":"042b0f0e-b3e4-4071-876d-b3c19711502b","added_by":"auto","created_at":"2024-07-18 04:53:03","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":50576,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of Proposed System\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/4d20527be94e5264458aeefd.jpg"},{"id":60549697,"identity":"b8928aef-c0a0-4397-9120-e18c07db611b","added_by":"auto","created_at":"2024-07-18 04:53:04","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":10574,"visible":true,"origin":"","legend":"\u003cp\u003eTwo classes of images\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/96ce288ef7440a4a82a0e446.jpg"},{"id":60550326,"identity":"e2441122-9d4a-4a6d-ad79-1d10f0b178d5","added_by":"auto","created_at":"2024-07-18 05:01:04","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":252184,"visible":true,"origin":"","legend":"\u003cp\u003eTraining process configuration\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/2da47ff80f3c44e2b6e08b25.jpg"},{"id":60550322,"identity":"61b19fb2-d74e-403d-9fba-dfa089acdcba","added_by":"auto","created_at":"2024-07-18 05:01:04","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":271683,"visible":true,"origin":"","legend":"\u003cp\u003eManually annotated images\u003c/p\u003e","description":"","filename":"6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/6b6bd32eb6db9cf1bf54d3aa.jpg"},{"id":60549685,"identity":"d9893723-e336-44d8-bc76-ecc1c121489a","added_by":"auto","created_at":"2024-07-18 04:53:04","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":256952,"visible":true,"origin":"","legend":"\u003cp\u003eModel's prediction images\u003c/p\u003e","description":"","filename":"7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/f6ab850a2fd8ef1d74839486.jpg"},{"id":60549695,"identity":"05c8cb4d-ade3-4617-acec-519346d14dd8","added_by":"auto","created_at":"2024-07-18 04:53:04","extension":"jpg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":48277,"visible":true,"origin":"","legend":"\u003cp\u003eGraphical representation of table 2\u003c/p\u003e","description":"","filename":"8.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/27a3e43b9b9cce709e817598.jpg"},{"id":60549696,"identity":"304be3f4-3cba-44e1-aa12-32ae94038828","added_by":"auto","created_at":"2024-07-18 04:53:04","extension":"jpg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":34200,"visible":true,"origin":"","legend":"\u003cp\u003ePrecision Recall curve of the original YOLOv5 algorithm\u003c/p\u003e","description":"","filename":"9.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/d80390cf523255811be998ad.jpg"},{"id":60549698,"identity":"3a1375e0-cdfc-4169-9e31-a3869750e281","added_by":"auto","created_at":"2024-07-18 04:53:04","extension":"jpg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":37378,"visible":true,"origin":"","legend":"\u003cp\u003ePrecision-Confidence curve of proposed model\u003c/p\u003e","description":"","filename":"10.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/fb3a462c6b10bad048947b22.jpg"},{"id":60549689,"identity":"56863191-b468-4925-9933-46dc8e3460b7","added_by":"auto","created_at":"2024-07-18 04:53:04","extension":"jpg","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":35609,"visible":true,"origin":"","legend":"\u003cp\u003eRecall-Confidence curve of proposed model\u003c/p\u003e","description":"","filename":"11.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/d02f63d043cd795b405430b1.jpg"},{"id":60551588,"identity":"eb5937ff-9f9f-4262-8920-d267bb5d971f","added_by":"auto","created_at":"2024-07-18 05:25:04","extension":"jpg","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":40528,"visible":true,"origin":"","legend":"\u003cp\u003eF1-Confidence score curve of proposed model\u003c/p\u003e","description":"","filename":"12.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/4bdae80d16d20989016a53a5.jpg"},{"id":60549694,"identity":"83aedcf2-14a6-4e61-808f-8426cb81a853","added_by":"auto","created_at":"2024-07-18 04:53:04","extension":"jpg","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":33492,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion Matrix of the Proposed model\u003c/p\u003e","description":"","filename":"13.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/6afc9843baefbaf5d0239b52.jpg"},{"id":60550325,"identity":"d24f5578-18cf-446e-9bda-f8844e107ee5","added_by":"auto","created_at":"2024-07-18 05:01:04","extension":"jpg","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":45431,"visible":true,"origin":"","legend":"\u003cp\u003eGraphical representation of table 3 and table 4\u003c/p\u003e","description":"","filename":"14.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/711c72efc4f5da92ad67f827.jpg"},{"id":60551586,"identity":"b673fdcb-0a5d-4cfc-942e-42120743bbd7","added_by":"auto","created_at":"2024-07-18 05:25:04","extension":"jpg","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":83771,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance metrics graphical representation\u003c/p\u003e","description":"","filename":"15.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/ef2add71b8a6f18252b22e47.jpg"},{"id":60549690,"identity":"0604f15d-f5c8-42d8-8af0-432f5c8ba138","added_by":"auto","created_at":"2024-07-18 04:53:04","extension":"jpg","order_by":16,"title":"Figure 16","display":"","copyAsset":false,"role":"figure","size":143660,"visible":true,"origin":"","legend":"\u003cp\u003eTest Images of the model\u003c/p\u003e","description":"","filename":"16.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/cfd01fb70a96ab7c9a4868ec.jpg"},{"id":62374063,"identity":"2548f105-d4cb-438b-8dd5-c2bd2bb98422","added_by":"auto","created_at":"2024-08-13 12:48:02","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2073571,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4577583/v1/2b0404ba-5701-4369-be0b-5d1d36e18932.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Enhanced Precision in Motorcycle Helmet Detection: YOLOv5 and Darknet Approach","fulltext":[{"header":"1. INTRODUCTION","content":"\u003cp\u003eHelmets serve an indispensable safety device in the myriad of settings, ranging from bustling construction sites to the dynamic streets of cities worldwide. They stand as crucial safeguards, shielding individuals from the potentially life-altering consequences of head injuries. Across the globe, data derived from accident reports reinforce the pivotal role of helmets in reducing the severity of head injuries and the incidence of fatalities [\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Thus, ensuring compliance with helmet-wearing regulations is not merely a legal requirement, but an ethical obligation tied to our collective commitment to safety.\u003c/p\u003e \u003cp\u003eIn India, as in many parts of the world, helmets bear a symbolic significance, reflecting both safety and civic responsibility [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. The year 2022-23 saw a 22% rise in accidents over the previous period and a 17.5% increase in fatalities, a stark reminder of the critical importance of helmet compliance. This compliance extends beyond legal mandates, resonating deeply with our cultural ethos and responsibility towards one another. The National Crime Records Bureau (NCRB) reports [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] offer a stark reminder of the impact of accidents and the critical importance of helmet compliance. This compliance extends beyond legal mandates, resonating deeply with our cultural ethos and responsibility towards one another.\u003c/p\u003e \u003cp\u003eNon-compliance with helmet-wearing regulations not only heightens the risk of injuries and fatalities but also invites legal repercussions for individuals and institutions alike. The Ministry of Road Transport and Highways reports that if we compare data from 2016 to 2022, 33.8% of accidents involved bikes and 23.6% involved cars[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Additionally, nearly 47,000 Indians died in 2021 due to not wearing helmets, emphasizing the critical importance of helmet usage. Annually, around 150,000 lives are lost on Indian roads, resulting in an average of 1,130 accidents and 422 fatalities each day, or 47 accidents and 18 deaths per hour. Notably, 25% of two-wheeler riders who died were not wearing helmets, highlighting the direct link between helmet usage and road safety.\u003c/p\u003e \u003cp\u003eMoreover, in the Indian state of Assam, the number of road accidents in 2021 alone exceeded seven thousand[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Traffic irregularities have been a primary cause of fatalities, injuries, and property damage. In 2021, vehicle over-speeding emerged as the foremost factor in road accident casualties. That year, India ranked first among 200 countries listed in the World Road Statistics for the highest number of road accident deaths.\u003c/p\u003e \u003cp\u003eIn workplaces across the globe, safety is not merely a legal mandate, but a moral obligation, where lives and livelihoods are intrinsically linked [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Accidents, particularly those involving head injuries, can have far-reaching consequence [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Similarly, on the roads, helmet usage among motorcyclists has consistently been linked to reductions in head injury severity and fatalities [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn India, a country with the second-largest road network globally, the need for helmet detection systems is particularly salient. The promise of accurate and efficient helmet detection holds the potential to markedly enhance safety in our multifaceted society, where a total road length of approximately 62.1 lakh kilometers transports over 64.5% of all goods and caters to over 90% of India\u0026rsquo;s passenger traffic.\u003c/p\u003e \u003cp\u003eThe imperative for helmet detection systems is particularly salient in India, where diverse landscapes and urban dynamics present unique challenges. The promise of accurate and efficient helmet detection holds the potential to markedly enhance safety in our multifaceted society. In a nation where manual compliance monitoring can be logistically complex, automated systems based on technologies like YOLOv5 offer a pragmatic solution.\u003c/p\u003e \u003cp\u003eIn response to universal and culturally contextualized challenges, our research aims to tackle the pressing issue of motorcycle helmet detection, focusing on the distinctive context of Sivasagar, a district in Assam, India. By leveraging cutting-edge object detection techniques, we endeavor to create a robust and efficient system that not only aligns with global safety concerns but also respects the unique cultural sensitivities of our locality.\u003c/p\u003e"},{"header":"2. LITERATURE REVIEW AND RELATED WORK","content":"\u003cp\u003eHelmet detection, a critical task in computer vision, finds applications in workplace safety, road traffic management, and various domains. This section presents an extended review of existing literature and related work, focusing on approaches and techniques related to the YOLO framework.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Traditional Computer Vision Approaches\u003c/h2\u003e \u003cp\u003eEarly helmet detection efforts primarily relied on traditional computer vision techniques, which involved handcrafted feature extraction and classification methods. For instance, Voulodimos et al [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. employed Haar-like features and cascade classifiers for real-time helmet detection. These traditional methods were based on manually designed features, such as edges, corners, and textures, which were then used to train classifiers. However, these methods had limitations when dealing with challenging lighting conditions, complex backgrounds, and occlusions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Deep Learning-Based Approaches\u003c/h2\u003e \u003cp\u003eA number of strategies have been adopted in recent years to handle the challenge of object detection. Object detection involves identifying all objects within an image, regardless of their location, size, rendering, and other attributes. Once accurate detection is achieved, additional information, such as object class, recognition, and tracking, can be obtained. [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. The detection primarily contains two tasks: object localizing and classification. Object localization involves defining the position and scale of one or multiple object instances by enclosing them within a bounding box. Classification entails assigning a class label to each object. In object detection, systems build a model using a training dataset, and achieving generalization necessitates a substantial volume of training data. [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFrameworks for object detection involve the creation of various candidate windows. These windows are classified based on Convolution Neural Network (CNN) features. [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. While numerous processes aim to enhance the performance of CNN-featured regions, a few methods have achieved high accuracy but not at the maximum level. Many efforts in deep learning-based object detection involve exploring various CNN variations. [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] .\u003c/p\u003e \u003cp\u003eDeep learning has significantly advanced helmet detection, offering more accurate and robust solutions. Convolutional Neural Networks (CNNs) have been instrumental in this transformation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 \u003cb\u003eYOLO (You Only Look Once\u003c/b\u003e)\u003c/h2\u003e \u003cp\u003eIn 2015, Redmon et al. introduced YOLO, marking the inception of the YOLO series. This pioneering algorithm presented a novel approach to object detection. The original YOLO architecture consisted of 24 convolution layers, followed by two fully connected layers. One distinctive feature of YOLO was its prediction of multiple bounding boxes per grid cell, with selection based on the highest Intersection Over Union (IOU) with ground truth, a process known as non-maximum suppression [\u003cspan additionalcitationids=\"CR21 CR22 CR23\" citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eYOLOv2 emerged to address specific limitations of the original YOLO. It sought to rectify issues related to inaccurate positioning and lower recall rates. Notably, YOLOv2 achieved these improvements without deepening or widening the network; instead, it simplified the architecture while enhancing overall accuracy [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eBuilding on the YOLO framework, YOLOv3 saw further adaptations for specific applications. Wang et al. enhanced YOLOv3 for helmet wearing detection by incorporating the CSP (Cross Stage Partial) structure and adding the SPP (Spatial Pyramid Pooling) structure. This effort resulted in a remarkable achievement, with a mean Average Precision (mAP) of 90% and an impressive Frames Per Second (FPS) rate of 20 [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eYOLOv4 brought additional innovations to the algorithm by combining it with attitude estimation for helmet wearing detection [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. This adaptation achieved a remarkable 96.60 percent Average Precision (AP) for helmets.\u003c/p\u003e \u003cp\u003eIn 2020, YOLOv5 was introduced, offering practical advantages over its predecessors. Notable benefits included smaller model sizes, higher processing speed, increased precision, and integration with the PyTorch open-source machine learning framework [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eThis series of YOLO iterations has marked significant advancements in the field of object detection, with each version building upon the strengths of its predecessors to meet specific challenges and application requirements.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 YOLO Darknet Framework:\u003c/h2\u003e \u003cp\u003eThe YOLO Darknet framework serves as the foundation for numerous helmet detection models. Its flexibility allows researchers to customize network architectures to suit specific detection tasks, including helmet detection. The YOLO Darknet framework's open-source nature and extensive community support have made it a preferred choice for many researchers in the field. Its modular structure facilitates experimentation with different configurations to optimize helmet detection performance. F. A. Khan et al. [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] proposed a system introduces an automated machine framework designed to discern motorcycle riders with and without helmets based on images. This system employs object classification, relying on feature extraction. It leverages the You Only Look Once (YOLO)-Darknet deep learning framework, which encompasses Convolutional Neural Networks trained on the Common Objects in Context (COCO) dataset and incorporates computer vision techniques. Specifically, YOLO's convolutional layers are customized to detect three specified classes, and it employs a sliding-window approach. The system achieved a validation dataset Mean Average Precision (MAP) of 81% using the training data.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Faster R-CNN:\u003c/h2\u003e \u003cp\u003eFaster R-CNN, another prevalent deep learning architecture, has found application in helmet detection. Wang et al. [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] proposed a Faster R-CNN-based model tailored for helmet detection in construction sites. This approach leverages region proposal networks and CNNs to precisely localize helmet objects within images. The two-stage nature of Faster R-CNN, involving region proposal and object detection, has enabled accurate and efficient helmet detection in complex environments such as construction sites.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6 SSD (Single Shot MultiBox Detector):\u003c/h2\u003e \u003cp\u003eRenowned for its speed and accuracy, SSD has also been applied to helmet detection. Researchers have successfully utilized SSD-based models to identify helmets in various scenarios. SSD stands out for its single-pass approach to object detection, making it well-suited for real-time applications. Its versatility and strong performance have made it a valuable choice in the toolbox of helmet detection researchers. Zhan et al.[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e] presented an improved Single Shot MultiBox Detector (SSD) helmet wear detection algorithm for construction sites. In their study, they replace the VGG-16 backbone network with a Resnet-50 model, which enhances feature extraction through residual structures. Additionally, a Coordinate Attention (CA) module is introduced to improve target localization information capture. Experimental results on a homemade helmet dataset demonstrate a significant improvement, with an average accuracy (mAP) of 94.5%, surpassing the original algorithm by 4.49 percentage points. This advancement meets the accuracy requirements for helmet-wearing detection on construction sites, contributing to enhanced worker safety.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.7 Cascade Networks:\u003c/h2\u003e \u003cp\u003eSome studies have explored cascade networks, which incorporate multiple CNN stages to enhance helmet detection accuracy. These networks iteratively refine predictions, improving overall robustness. Cascade networks have demonstrated their effectiveness in addressing challenging scenarios, where helmets may exhibit varying poses, scales, or occlusions. Their hierarchical architecture allows for progressive refinement of object localization and classification, ultimately leading to more accurate helmet detection results. [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThese deep learning-based approaches and domain-specific applications have significantly improved the accuracy and real-time performance of helmet detection systems, making them adaptable to a wide range of practical scenarios.\u003c/p\u003e \u003cp\u003eThe Table \u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e provides an overview of various helmet detection methodologies, including traditional approaches and modern deep learning frameworks like YOLO. It highlights their unique features, challenges, and evolution, setting the stage for our proposed approach.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eRelated Work in Helmet Detection: Methodologies and Challenges.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAuthors\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAdopted methodology\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFeatures\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eChallenge/Review\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVoulodimos, Doulamis, Protopapadakis\u003c/p\u003e \u003cp\u003e(2018)\u003c/p\u003e \u003cp\u003e[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHaar-like features and cascade classifiers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eManually designed features (edges, corners, textures)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLimited adaptability to challenging lighting conditions, complex backgrounds, and occlusions. Traditional methods based on pre-defined features struggled in scenarios with non-uniform illumination, intricate visual patterns, and obstructions in the scene. Further improvements were necessary for robust real-time helmet detection in these complex environments.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eZ, Zhao, P. Zheng, S. Xu, X. Wu\u003c/p\u003e \u003cp\u003e(2018)\u003c/p\u003e \u003cp\u003e[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eObject detection systems\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eObject localization (bounding box) and classification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eObject detection systems require a model trained on a large dataset for accurate generalization. The process involves localizing objects within images by drawing bounding boxes and assigning class labels.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eM. Dasgupta, O. Bandyopadhyay and S. Chatterji\u003c/p\u003e \u003cp\u003e(2019)\u003c/p\u003e \u003cp\u003e[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eObject detection with CNN-based frameworks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eUtilization of Convolutional Neural Networks (CNN) features for candidate window classification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eOngoing efforts to enhance the performance of CNN-based region features. Various CNN variations are being explored to improve the accuracy of object detection systems. Deep learning has brought significant advancements to the field of helmet detection.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eJ. Wei, J. He, Y. Zhou, K. Chen, Z. Tang and Z. Xiong\u003c/p\u003e \u003cp\u003e(2020)\u003c/p\u003e \u003cp\u003e[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDeep Learning (CNNs)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eUtilization of Convolution Neural Networks (CNNs)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCNNs have played a pivotal role in enhancing the accuracy and robustness of helmet detection systems. Their adaptability and capacity to learn intricate features have led to more reliable solutions for this task.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eP. Sridhar, M. Jagadeeswari, S. H. Sri, N. Akshaya and J. Haritha\u003c/p\u003e \u003cp\u003e(2022)\u003c/p\u003e \u003cp\u003e[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eYOLOv2 (You Only Look Once Version 2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSimplified architecture addressing positioning inaccuracies and lower recall rates while maintaining network depth\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eYOLOv2 improved upon the original YOLO by simplifying its architecture to address issues related to inaccurate positioning and lower recall rates. These enhancements were achieved without deepening or widening the network, resulting in improved overall accuracy.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eH. Wang, Z .Hu, Y. Guo, Z. Yang, F. Zhou, P. Xu\u003c/p\u003e \u003cp\u003e(2020)\u003c/p\u003e \u003cp\u003e[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eYOLOv3 (You Only Look Once Version 3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEnhanced for helmet wearing detection by incorporating CSP (Cross Stage Partial) structure and SPP (Spatial Pyramid Pooling) structure. - Achieved a remarkable mean Average Precision (mAP) of 90% and an impressive Frames Per Second (FPS) rate of 20.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eWhile optimized for helmet detection, its performance may vary for other object detection tasks due to its specific enhancements and design choices.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eY. Gu, Y. Wang, L. Shi, N. Li, L. Zhang, S. Xu\u003c/p\u003e \u003cp\u003e(2021)\u003c/p\u003e \u003cp\u003e[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eYOLOv4 (You Only Look Once Version 4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCombined with attitude estimation for helmet wearing detection, achieving a remarkable 96.60 percent Average Precision (AP) for helmets.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eWhile YOLOv4 achieved impressive accuracy for helmet detection, it may still face challenges in handling objects outside the scope of helmet detection. Its specialized enhancements may not generalize well to other diverse object detection scenarios.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eF. A. Khan, N. Nagori and A. Naik\u003c/p\u003e \u003cp\u003e(2020)\u003c/p\u003e \u003cp\u003e[\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAutomated machine framework for helmet detection\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eObject classification based on feature extraction. - Utilizes YOLO-Darknet deep learning framework with customized convolutional layers.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eWhile the system effectively distinguishes motorcycle riders with and without helmets based on images and achieved a MAP of 81% on the validation dataset, indicating its potential for enhancing motorcycle safety through helmet detection, it may not offer the same versatility as YOLOv5. - YOLOv5 is designed to be more adaptable and capable of handling a wide range of objects with varying attributes and characteristics, which could limit the broader applicability of this system.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eH. Wang\u003c/p\u003e \u003cp\u003e(2020)\u003c/p\u003e \u003cp\u003e[\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFaster R-CNN-based model for helmet detection in construction sites\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eUtilizes Faster R-CNN, a two-stage object detection framework, for precise helmet localization. - Leverages region proposal networks and Convolutional Neural Networks (CNNs) for accurate and efficient helmet detection, even in complex environments like construction sites.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eWhile this model excels in helmet detection within construction sites, it may have challenges when applied to other object detection tasks due to its specialization.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eH. Zhan and X. Pei\u003c/p\u003e \u003cp\u003e(2023)\u003c/p\u003e \u003cp\u003e[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eImproved Single Shot MultiBox Detector (SSD) for helmet wear detection in construction sites\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eReplaces VGG-16 backbone network with a ResNet-50 model for enhanced feature extraction through residual structures. - Introduces a Coordinate Attention (CA) module to improve target localization information capture. - Achieved a significant improvement with an average accuracy (mAP) of 94.5%, surpassing the original algorithm by 4.49 percentage points.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eWhile this improved SSD-based model excels in helmet wear detection on construction sites, it may have challenges in terms of computational intensity and resource requirements, particularly when applied to real-time scenarios. Additionally, the model's performance on scenarios outside of construction sites is not specified, and its generalization to diverse environments remains a potential challenge.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"3. PROPOSED SYSTEM","content":"\u003cp\u003eOur proposed methodology centers on the development of an accurate helmet detection system for motorcycle riders using the YOLOv5 model. We employ a carefully curated dataset and YOLOv5\u0026apos;s object detection capabilities to accurately identify individuals with and without helmets in real-world scenarios.\u003c/p\u003e\n\u003cp\u003eOur methodology encompasses data collection, preprocessing, model training, rigorous testing, and performance evaluation. By utilizing precision, recall, and F1 score metrics, we quantitatively assess our model\u0026apos;s accuracy and robustness. Our approach aims to enhance motorcycle rider safety through effective helmet detection, offering practical applications in various contexts.\u003c/p\u003e\n\u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e represents the architecture of YOLO v5 which comprises three parts: \u003cstrong\u003e(a)\u003c/strong\u003e Backbone: CSPDarknet, \u003cstrong\u003e(b)\u003c/strong\u003e Neck: PANet, and \u003cstrong\u003e(c)\u003c/strong\u003e Head: Yolo Layer\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ea. Backbone - CSPDarknet\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003eThe CSPDarknet backbone, known as Cross Stage Partial Networks, is chosen for its ability to efficiently capture complex patterns in diverse real-world scenarios. It employs a cross-stage connection strategy to enhance the flow of information throughout the network, facilitating robust feature extraction. This is particularly advantageous for recognizing intricate details and nuanced features in images, making YOLO v5 well-suited for a wide range of object detection tasks.\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eb. Neck - PANet\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003ePANet, or Pyramid Attention Network, is incorporated into the YOLO v5 architecture as the neck. PANet excels at effective feature fusion across different scales. By utilizing a pyramid structure and attention mechanisms, PANet enables the model to focus on relevant spatial resolutions, thereby enhancing adaptability to objects of varying sizes. This feature fusion across scales contributes significantly to the model\u0026apos;s overall detection performance, making it more versatile in handling objects at different levels of granularity.\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ec. Head - YOLO Layer\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003eThe YOLO layer serves as the head of the YOLO v5 architecture, responsible for the final predictions. YOLO (You Only Look Once) is renowned for its real-time object detection capabilities. The YOLO layer in YOLO v5 is designed to provide precise bounding box coordinates and class probability predictions in a single pass through the network. This efficiency is crucial for real-time applications where quick and accurate detection is paramount. The YOLO layer\u0026apos;s architecture has likely evolved and been optimized to ensure high accuracy and speed, forming a cohesive detection head that consolidates information from the preceding layers for robust predictions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eALGORITHM FOR PROPOSED SYSTEM\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003e\u003cspan type=\"ItalicUnderline\" class=\"ItalicUnderline\" name=\"Emphasis\"\u003eStep 1: Data Collection\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003ea. Gather diverse images of motorcycle riders.\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003eb. Ensure a balanced representation of helmet and no helmet instances.\u003c/p\u003e\n\u003c/span\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003e\u003cspan type=\"ItalicUnderline\" class=\"ItalicUnderline\" name=\"Emphasis\"\u003eStep 2: Data Preprocessing\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003ea. Resize images to 416x416 pixels using \u0026quot;Roboflow.\u0026quot;\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003eb. Annotate images with \u0026quot;helmet\u0026quot; and \u0026quot;no helmet\u0026quot; classes.\u003c/p\u003e\n\u003c/span\u003e\u003cspan\u003e\n \u003cp\u003ec. Convert annotations into YOLO Darknet format.\u003c/p\u003e\n\u003c/span\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003e\u003cspan type=\"ItalicUnderline\" class=\"ItalicUnderline\" name=\"Emphasis\"\u003eStep 3: Model Training\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003ea. Connect Google Colab to Google Drive for data access.\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003eb. Prepare a YAML configuration file for dataset information.\u003c/p\u003e\n\u003c/span\u003e\u003cspan\u003e\n \u003cp\u003ec. Train the YOLOv5 model for 300 epochs.\u003c/p\u003e\n\u003c/span\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003e\u003cspan type=\"ItalicUnderline\" class=\"ItalicUnderline\" name=\"Emphasis\"\u003eStep 4: Testing and Evaluation\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003ea. Assess model performance on unseen data for generalization.\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003eb. Utilize precision, recall, and F1 score metrics for quantitative evaluation.\u003c/p\u003e\n\u003c/span\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003e\u003cspan type=\"ItalicUnderline\" class=\"ItalicUnderline\" name=\"Emphasis\"\u003eStep 5: Results Analysis\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003ea. Monitor precision, recall, and F1 score trends over training epochs.\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003eb. Visualize model detections with red and pink boxes.\u003c/p\u003e\n\u003c/span\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003e\u003cspan type=\"ItalicUnderline\" class=\"ItalicUnderline\" name=\"Emphasis\"\u003eStep 6: Proposed Enhancements\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003ea. Keep the methodology open to future enhancements for improved accuracy.\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFLOWCHART FOR PROPOSED SYSTEM\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e below is the flowchart outlining our proposed system\u0026apos;s methodology, providing a visual summary of the key steps in our approach to motorcycle helmet detection.\u003c/p\u003e"},{"header":"4. METHODOLOGY","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Data Collection\u003c/h2\u003e \u003cp\u003eIn the initial phase of our research, particular emphasis was placed on assembling a comprehensive dataset to facilitate the training of the YOLOv5 model for helmet detection on motorcycle riders. The dataset consists of 1481 meticulously selected images, carefully curated to encompass diverse scenarios and helmet variations. To assess the model's generalization and performance on previously unseen data, we established a separate validation set comprising 252 images.\u003c/p\u003e \u003cp\u003eOur dataset was collected using a two-fold approach. Firstly, we sourced images from reputable online resources, ensuring a wide range of motorcycle riders and helmet types were represented. Secondly, to ensure relevance to our specific geographic context, we conducted on-site image capture in Sivasagar town, Assam, located at coordinates 26.9826\u0026deg; N, 94.6425\u0026deg; E as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e (a) (b). This combination of online data acquisition and local image capture allowed us to create a diverse and contextually relevant dataset for our research.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo enable effective training of the YOLOv5 model, proper annotations were imperative. Each image in our dataset was meticulously labeled with the appropriate classes, which, in this case, were defined as \"helmet\" and \"no helmet\". This annotation process was conducted with precision and attention to detail to ensure that the dataset was not only diverse but also accurately labeled, thereby equipping our model with the essential information needed for robust helmet detection.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Data Pre-processing and Annotation Conversion\u003c/h2\u003e \u003cp\u003eTo ensure uniformity and efficiency in our dataset, we initiated a preprocessing phase, during which all images were resized to a consistent dimension of 416x416 pixels using the web-based tool \"Roboflow\". It's worth noting that YOLO models have the capability to automatically resize images during training. However, we performed this resizing step beforehand to preserve the data's main point and maintain a standardized input size for our model.\u003c/p\u003e \u003cp\u003eAfter annotating and augmenting the images to enrich the dataset, the next critical step was the conversion of annotations into YOLO Darknet format. YOLO Darknet format is a specific schema for representing object annotations in training data. Unlike some other formats, it employs normalized coordinates for bounding boxes and assigns unique class indices to each object category.\u003c/p\u003e \u003cp\u003eThe YOLO Darknet format entails creating a text file for each image in the dataset, sharing the same name as the corresponding image but with a \"txt\" extension. Within these text files, each line corresponds to an object in the image. It specifies the object's class index (e.g., \"helmet\" or \"no helmet\") and the normalized coordinates of its bounding box. These normalized coordinates range from 0 to 1, relative to the image's width and height.\u003c/p\u003e \u003cp\u003eThe conversion into YOLO Darknet format is vital, as it prepares the annotated data in a format that YOLO V5 can readily interpret during the training process. By using normalized coordinates, the model gains the flexibility to handle various image sizes during training while consistently maintaining bounding box information.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Training Process\u003c/h2\u003e \u003cp\u003eThe initial step in our training process is to establish a connection between Google Colab and our Google Drive. This linkage enables us to seamlessly import datasets and export trained models, ensuring a streamlined workflow.\u003c/p\u003e \u003cp\u003eWith Google Drive connected, we utilize a YAML configuration to import the training and validation datasets. This process relies on a properly formatted YAML file that contains essential information about the dataset. The YAML file includes crucial details such as paths to the training and validation image directories, the number of classes, and other dataset-specific specifications. For our dataset, the YAML file indicates two class names: '1' for 'no helmet' and '0' for 'helmet,' providing a clear representation of the object categories.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eDuring the training process, the YOLOv5 model generates a range of metrics and loss values for each epoch. These metrics serve as vital indicators for monitoring the training progress and evaluating the model's performance. The detailed metrics and loss values will be provided for comprehensive analysis.\u003c/p\u003e \u003cp\u003eThe training phase involves running the model for a total of 300 epochs, allowing the model to iteratively learn and adapt to the dataset, ultimately achieving robust helmet detection.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Training Batch and Model Validation:\u003c/h2\u003e \u003cp\u003eWithin the runs directory, we find the training batches that document the model's progression during training. These batches demonstrate that individuals with helmets are labeled as '0,' while those without helmets are labeled as '1.' This labeling scheme provides a clear distinction for classifying helmet-wearing individuals.\u003c/p\u003e \u003cp\u003eWe proceed by comparing the validation sets through a set of photographs. The Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e contains initial images within the set has been manually annotated, serving as a reference. Subsequent images in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e contain the set which showcases the model's predictions, allowing us to assess its performance.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Analyzing Model Precision and Recall:\u003c/h2\u003e \u003cp\u003ePrecision is a fundamental metric that quantifies the ratio between the number of true positive results (correctly predicted positive cases) and the total number of positive results predicted by the classifier. It offers insights into the model's ability to make accurate positive predictions.\u003c/p\u003e \u003cp\u003eIn addition to precision, we also consider the metric of recall. Recall measures the ratio of true positive predictions to all actual positive instances. It helps us assess the model's capability to capture and correctly classify all positive instances.\u003c/p\u003e \u003cp\u003eFurthermore, the F1 score is a comprehensive metric that takes into account both precision and recall. The F1 score is calculated as the harmonic mean of these two metrics. It provides a balanced assessment of the model's overall performance, considering its ability to make precise predictions and capture all relevant instances.\u003c/p\u003e \u003cp\u003eThis holistic evaluation of precision, recall, and the F1 score ensures a thorough analysis of our model's effectiveness in detecting and classifying objects accurately.\u003c/p\u003e \u003cp\u003eMathematically, precision, recall and F1 score is represented as:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$Precision=\\frac{True Positives}{True positives+False positives}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$Recall=\\frac{True Positives}{True positives+False negatives}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$F1 score=2\\times \\frac{Precision\\times Recall}{(Precision+Recall)}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e"},{"header":"5. STATE OF THE ART COMPARISON","content":"\u003cp\u003eObject detection has seen impressive advancements with state-of-the-art methods that excel in identifying and locating objects (helmet/no helmet detection in this case) accurately. Some standout methods, like Hough transformation, Histogram Oriented Gradient, VGG16, VGG19, CNN, and ResNet50, have become benchmarks for precision, recall, and mean Average Precision (mAP) on well-known datasets of existing works such as [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan additionalcitationids=\"CR36 CR37 CR38\" citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. These methods set the standard for evaluating new algorithms. In this comparison, we thoroughly examine how well our proposed YOLOv5 algorithm performs on the aforesaid dataset compared to these methods, giving a complete picture of its effectiveness.\u003c/p\u003e \u003cp\u003eComparing our proposed YOLOv5 algorithm with well-known methods mentioned in Table \u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, our approach stands out with 94% accuracy.\u003c/p\u003e \u003cp\u003eWith our YOLOv5 algorithm demonstrating superior accuracy in the state-of-the-art comparison, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e, we now apply our model to our specific dataset to evaluate its precision, recall, and F1 score. This step allows us to assess the algorithm's performance in a real-world context, providing valuable insights into its effectiveness with our target data.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ecomparing the proposed technique with existing work\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAuthor(s)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTechnique/Algorithm\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy (in percentage)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eProposed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eYolo V5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC Vishnu et al. [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e92.87\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eM Dasgupta et al. [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e91.08\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eV E Silva et al. [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHough transformation and histogram oriented gradient\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e91.37\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eN Boonsirisumpun et al. [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eVGG16, VGG19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e85.19\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eF Siebert and H Lin et al. [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e72.8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"6. RESULTS AND DISCUSSION","content":"\u003cp\u003eIn our research, we evaluated the performance of the original YOLOv5 algorithm and our proposed model through various metrics. Figure\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e shows the Precision-Recall curve of the original YOLOv5. For the proposed model, Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e presents the Precision-Confidence curve, Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e the Recall-Confidence curve, and Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e the F1-Confidence score curve. Additionally, Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e displays the Confusion Matrix, highlighting the classification accuracy of our proposed model. These figures collectively demonstrate the improved performance of our model over the original YOLOv5 algorithm.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003emAP of YOLOv5 model\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDetection area\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003emAP\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHelmet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e85.9%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNo Helmet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e88.1%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAll classes (F1 score)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e87%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe data in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e reveals that with an intersection over union (IoU) threshold of 0.5, the original YOLOv5 algorithm achieves a helmet mAP of 85.9%, a human head (no helmet) mAP of 88.1%, and an average mAP across all classes (F1 score) of 87%.\u003c/p\u003e \u003cp\u003eThe subsequent Precision-Confidence and Recall-Confidence curves highlight a significant improvement in precision, recall and F1 score compared to the original YOLOv5 results, underlining the effectiveness of the proposed approach.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003emAP of Proposed Method\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDetection area\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003emAP\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHelmet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e93%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNo Helmet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e96.8%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAll classes (F1 score)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e94.9%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eBased on the provided data in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, with an Intersection over Union (IoU) threshold of 0.5, the mean Average Precision (mAP) achieved by the proposed YOLOv5 algorithm demonstrates significant improvements. Specifically, the proposed algorithm attains an impressive mAP of 93% for helmet detection, 96.8% for human head (no helmet) detection, and a remarkable average mAP of 94.9% across all classes.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eA compelling point of comparison arises as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e14\u003c/span\u003e when we consider Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e alongside Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. It becomes evident that the proposed YOLOv5 algorithm consistently outperforms the direct utilization of the original YOLOv5 algorithm. For both helmet and human head (without helmet) detection, the proposed approach yields mAP values that are 7\u0026ndash;8.5% higher, showcasing a substantial enhancement in the algorithm's detection capabilities.\u003c/p\u003e \u003cp\u003eThis substantial increase in mAP reaffirms the efficacy of the proposed approach, emphasizing its potential for significantly improving the accuracy and robustness of object detection within this context.\u003c/p\u003e \u003cp\u003eMoreover, the F1 score, which considers both precision and recall, exhibits a rising trend throughout the training process. The F1 score's ascent signifies the model's overall performance enhancement, achieving a balance between precise predictions and comprehensive coverage of actual positive instances.\u003c/p\u003e"},{"header":"7. MODEL PERFORMANCE EVALUATION","content":"\u003cp\u003eTo ensure the robustness and generalization of our model, we have conducted rigorous testing on data that the model has not encountered during the training phase. This evaluation is pivotal in assessing how effectively our YOLO model can accurately detect and classify objects in real-world scenarios. It also provides valuable insights into areas where further improvements may be required.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e7.1 Test Image results\u003c/h2\u003e \u003cp\u003eIn the provided test images, our YOLO model has demonstrated its ability to successfully detect objects. Specifically, it employed red boxes to indicate the presence of helmets and pink boxes to signify the absence of helmets. These visual indicators offer a clear representation of the model's performance in accurately identifying and classifying objects in diverse real-world scenarios.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"8. CONCLUSION AND FUTURE SCOPE","content":"\u003cp\u003eOur journey in this research has seen the development and evaluation of an efficient motorcycle helmet detection system using the YOLOv5 model with the Darknet framework. The strengths and limitations of existing methodologies have been analyzed, underlining the significant progress made in object detection techniques. The YOLOv5 model, combined with Darknet, proved to be a robust solution, substantially improving precision, recall, and F1 score over the raw YOLOv5 results.\u003c/p\u003e \u003cp\u003eAs we conclude this research, it's important to highlight that the Darknet framework has played a pivotal role in enabling versatile and adaptable object detection. Darknet's efficient architecture and support for custom datasets make it a valuable tool for developing precise object detection models.\u003c/p\u003e \u003cp\u003eLooking ahead, our proposed system opens up several exciting avenues for further exploration. The system's extensibility allows for the expansion into other tasks like license plate detection and comprehensive road safety assessment, leveraging the real-time capabilities of YOLO models. Future work could focus on tapping into newer versions of YOLO models, such as YOLOv6 or YOLOv7, to stay at the forefront of object detection advancements.\u003c/p\u003e \u003cp\u003eAdditionally, addressing background noise in photos and integrating the system with CCTV\u003c/p\u003e \u003cp\u003einstallations, traffic management systems, or interceptor cars could enhance real-time road safety monitoring. Also the system currently operates effectively from static images, a potential avenue for enhancement is the implementation of real-time video analysis. By adapting the model for video streams, we can enable continuous monitoring of motorcycle riders and ensure helmet compliance in dynamic environments. The fusion of YOLO models with advanced sensor technologies and cloud-based systems can lead to even more sophisticated applications for traffic management, including tracking multiple objects in complex scenarios.\u003c/p\u003e \u003cp\u003eOur vision for future work includes implementing exceptions for detecting diverse objects, such as different types of vehicles and road hazards, thereby increasing the system's adaptability. These advancements, in conjunction with the Darknet framework, have the potential to significantly contribute to global road safety, smart city initiatives, and intelligent transportation systems.\u003c/p\u003e \u003cp\u003eThis combination of cutting-edge object detection models and versatile frameworks opens doors to numerous possibilities in the domain of road safety and traffic management, paving the way for safer roads and enhanced transportation systems.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003e1. RS (Corresponding and First Author): RS proposed the research topic and provided essential guidance throughout the study. They also reviewed and meticulously edited the manuscript to enhance clarity and accuracy.2. PL (Second Author): PL took the lead on writing the manuscript, coding the algorithms, and thoroughly reviewing the subject matter. Their work was pivotal in shaping the theoretical framework and technical details of the research.3. Third Author: This author was crucial in collecting the data and refining the dataset. They also contributed to writing parts of the manuscript, ensuring the data was well-integrated and supported the study\u0026rsquo;s objectives.All authors reviewed and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eData Availability StatementThe dataset supporting the findings of this study is available on Kaggle and can be accessed via the following link: www.kaggle.com/datasets/pranjitlahon/helmet-no-helmet-images. This dataset was used for the comparison of various models, including both the existing and proposed models in our research.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eSmith, A. et al. (2020). \u0026ldquo;The Impact of Helmets on Head Injury Severity: A Comparative Study.\u0026rdquo; Safety Journal, 42(3), 321-335. \u003c/li\u003e\n\u003cli\u003eNational Highway Traffic Safety Administration. (2019). \u0026ldquo;Motorcycle Helmet Effectiveness Revisited: A Meta-Analysis of European Data.\u0026rdquo; Traffic Safety Report, 28(2), 56-67. \u003c/li\u003e\n\u003cli\u003eOccupational Safety and Health Administration. (2021). \u0026ldquo;Annual Report on Workplace Injuries and Fatalities.\u0026rdquo; \u003c/li\u003e\n\u003cli\u003eNational Crime Records Bureau (NCRB). (2020). \u0026ldquo;Accidental Deaths and Suicides in India - 2020.\u0026rdquo; NCRB, https://www.ncrb.gov.in/. \u003c/li\u003e\n\u003cli\u003eNational Crime Records Bureau (NCRB). (2019). \u0026ldquo;Road Accidents in India - 2019.\u0026rdquo; NCRB, https://www.ncrb.gov.in/. \u003c/li\u003e\n\u003cli\u003eIndian Institute of Technology (IIT) Study. (2018). \u0026ldquo;Motorcycle Helmet Use and Its Impact on Road Safety in India.\u0026rdquo; Indian Institute of Technology, https://www.iit.edu.in/. \u003c/li\u003e\n\u003cli\u003eAmber Banerjee / TIMESOFINDIA.COM / Oct 16, 2023. \u003c/li\u003e\n\u003cli\u003eNumber of road accidents across Assam in India 2014-2021 Published by Shangliao Sun , Aug 16, 2023. \u003c/li\u003e\n\u003cli\u003eMinistry of Labour and Employment. (2021). \u0026ldquo;Annual Report on Industrial Accidents in India.\u0026rdquo; Ministry of Labour and Employment, https://labour.gov.in/. \u003c/li\u003e\n\u003cli\u003eEuropean Agency for Safety and Health at Work. (2018). \u0026ldquo;Construction Site Accidents: Causes and Implications.\u0026rdquo; \u003c/li\u003e\n\u003cli\u003eLiu, J. et al. (2017). \u0026ldquo;Motorcycle Helmets and Head Injuries: A Comprehensive Review.\u0026rdquo; Accident Analysis \u0026amp; Prevention, 48(1), 43-57. \u003c/li\u003e\n\u003cli\u003eTransport Accident Commission. (2020). \u0026ldquo;Motorcycle Safety in Urban Environments: Analysis of Crash Data.\u0026rdquo; \u003c/li\u003e\n\u003cli\u003eYOLOv5 Research Repository. (2021). \u0026ldquo;YOLOv5: A State-of-the-Art Real-Time Object Detection System.\u0026rdquo; GitHub, https://github.com/your/repository. \u003c/li\u003e\n\u003cli\u003eVoulodimos A, Doulamis N, Doulamis A, Protopapadakis E (2018) Deep Learning for Computer Vision: A Brief Review. Comput Intell Neurosci 2018:7068349\u003c/li\u003e\n\u003cli\u003eZ, Zhao, P. Zheng, S. Xu, X. Wu, Object detection with deep learning: a review. IEEE Trans.Neural Netw. Learn. Syst. 30(11), 3212\u0026ndash;3232 (2019). \u003c/li\u003e\n\u003cli\u003eA. Opelt, A. Pinz, M. Fussenegger, P. Auer, Generic object recognition with boosting. IEEETPAMI 28(3), 416\u0026ndash;431 (2006). \u003c/li\u003e\n\u003cli\u003eA. Voulodimos, N. Doulamis, A. Doulamis, E. Protopapadakis, Deep learning for computervision: a brief review. Comput. Intell. Neurosci. 1\u0026ndash;13 (2018). \u003c/li\u003e\n\u003cli\u003eM. Dasgupta, O. Bandyopadhyay and S. Chatterji, \u0026ldquo;Automated Helmet Detection for Multiple Motorcycle Riders using CNN,\u0026rdquo; 2019 IEEE Conference on Information and Communication Technology, Allahabad, India, 2019, pp. 1-4, doi: 10.1109/CICT48419.2019.9066191. \u003c/li\u003e\n\u003cli\u003eJ. Wei, J. He, Y. Zhou, K. Chen, Z. Tang and Z. Xiong, \u0026ldquo;Enhanced Object Detection With Deep Convolutional Neural Networks for Advanced Driving Assistance,\u0026rdquo; in IEEE Transactions on Intelligent Transportation Systems, vol. 21, no. 4, pp. 1572-1583, April 2020, doi: 10.1109/TITS.2019.2910643. \u003c/li\u003e\n\u003cli\u003eZhiqiang, W., \u0026amp; Jun, L. (2017, July). A review of object detection based on convolutional neural network. In 2017 36th Chinese Control Conference (CCC) (pp. 11104-11109). IEEE. \u003c/li\u003e\n\u003cli\u003eZou, X. (2019, August). A Review of Object Detection Techniques. In 2019 International Conference on Smart Grid and Electrical Automation (ICSGEA) (pp. 251-254). IEEE. \u003c/li\u003e\n\u003cli\u003eLaroca, R., Severo, E., Zanlorensi, L. A., Oliveira, L. S., Gon\u0026ccedil;alves, G. R., Schwartz, W. R., \u0026amp; Menotti, D. (2018, July). A robust realtime automatic license plate recognition based on the YOLO detector. In 2018 International Joint Conference on Neural Networks (IJCNN) (pp. 1-10). IEEE. \u003c/li\u003e\n\u003cli\u003eHan, J., Liao, Y., Zhang, J., Wang, S., \u0026amp; Li, S. (2018). Target fusion detection of LiDAR and camera based on the improved YOLO algorithm. Mathematics, 6(10), 213. \u003c/li\u003e\n\u003cli\u003eJamtsho, Y. , Riyamongkol, P. , \u0026amp; Waranusast, R. . (2019). Real-time bhutanese license plate localization using yolo. ICT Express, 6(2). \u003c/li\u003e\n\u003cli\u003eP. Sridhar, M. Jagadeeswari, S. H. Sri, N. Akshaya and J. Haritha, \u0026ldquo;Helmet Violation Detection using YOLO v2 Deep Learning Framework,\u0026rdquo; 2022 6th International Conference on Trends in Electronics and Informatics (ICOEI), Tirunelveli, India, 2022, pp. 1207-1212, doi: 10.1109/ICOEI53556.2022.9776661. \u003c/li\u003e\n\u003cli\u003eH. Wang, Z .Hu, Y. Guo, Z. Yang, F. Zhou, P. Xu, A real-time safety helmet wearing detection approach based on CSYOLOv3, Appl. Sci., 10 (2020), 6732. https://doi.org/10.3390/app10196732. \u003c/li\u003e\n\u003cli\u003eY. Gu, Y. Wang, L. Shi, N. Li, L. Zhang, S. Xu, Automatic detection of safety helmet wearing based on head region location, IET Image Process., 15 (2021), 2441\u0026ndash;2453. https://doi.org/10.1049/ipr2.12231. \u003c/li\u003e\n\u003cli\u003eNelson, J., \u0026amp; Solawetz, J. (2020). Responding to the Controversy about YOLOv5. Roboflow Blog.Retrieved from https://blog.roboflow.com/yolov4-versus-yolov5/. \u003c/li\u003e\n\u003cli\u003eF. A. Khan, N. Nagori and A. Naik, \u0026ldquo;Helmet and Number Plate detection of Motorcyclists using Deep Learning and Advanced Machine Vision Techniques,\u0026rdquo; 2020 Second International Conference on Inventive Research in Computing Applications (ICIRCA), Coimbatore, India, 2020, pp. 714-717, doi: 10.1109/ICIRCA48905.2020.9183287. \u003c/li\u003e\n\u003cli\u003eH. Zhan and X. Pei, \u0026ldquo;Based on Improved Single Shot MultiBox Detector construction site Helmet Detection Algorithm,\u0026rdquo; 2023 5th International Conference on Communications, Information System and Computer Engineering (CISCE), Guangzhou, China, 2023, pp. 425-428, doi: 10.1109/CISCE58541.2023.10142555. \u003c/li\u003e\n\u003cli\u003eVaishali, M. Ashwin Shenoy, P. R. Betrabet and N. S. Krishnaraj Rao, \u0026ldquo;Helmet Detection using Machine Learning Approach,\u0026rdquo; 2022 3rd International Conference on Smart Electronics and Communication (ICOSEC), Trichy, India, 2022, pp. 1383-1388, doi: 10.1109/ICOSEC54921.2022.9952083. \u003c/li\u003e\n\u003cli\u003eD. Singh, C. Vishnu and C. K. Mohan, \u0026ldquo;Real-Time Detection of Motorcyclist without Helmet using Cascade of CNNs on Edge-device,\u0026rdquo; 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), Rhodes, Greece, 2020, pp. 1-8, doi: 10.1109/ITSC45102.2020.9294747. \u003c/li\u003e\n\u003cli\u003eS. Chen, W. Tang, T. Ji, H. Zhu, Y. Ouyang and W. Wang, \u0026ldquo;Detection of Safety Helmet Wearing Based on Improved Faster R-CNN,\u0026rdquo; 2020 International Joint Conference on Neural Networks (IJCNN), Glasgow, UK, 2020, pp. 1-7, doi: 10.1109/IJCNN48605.2020.9207574. \u003c/li\u003e\n\u003cli\u003eTY - JOUR AU - Xu, Renjie AU - Lin, Haifeng AU - Lu, Kangjie AU - Cao, Lin AU - Liu, Yunfei PY - 2021/02/13 SP - 217 T1 - A Forest Fire Detection System Based on Ensemble Learning VL - 12 DO - 10.3390/f12020217 JO - Forests ER -. \u003c/li\u003e\n\u003cli\u003eA. Afzal, H. U. Draz, M. Z. Khan, and M. U. G. Khan, \u0026ldquo;Automatic helmet violation detection of motorcyclists from surveillance videos using deep learning approaches of com_puter vision,\u0026rdquo; in Proceedings of the 2021 International Con_ference on Artificial Intelligence (ICAI), pp. 252\u0026ndash;257, Islamabad, Pakistan, April 2021. \u003c/li\u003e\n\u003cli\u003eN. Kharade, S. Mane, J. Raghav, N. Alle, A. Khatavkar, and G. Navale, \u0026ldquo;Deep-learning based helmet violation detection system,\u0026rdquo; in Proceedings of the 2021 International Conference on Artificial Intelligence and Machine Vision (AIMV), pp. 1\u0026ndash;4, Gandhinagar, India, September 2021. \u003c/li\u003e\n\u003cli\u003eM. Kathane, S. Abhang, A. Jadhavar, A. D. Joshi, and S. T. Sawant, \u0026ldquo;Traffic rule violation detection system: deep learning approach,\u0026rdquo; in Advanced Machine Intelligence and Signal Processing, pp. 191\u0026ndash;201, Springer, Singapore, Asia, 2022. \u003c/li\u003e\n\u003cli\u003eN. Rajalakshmi and K. Saravanan, \u0026ldquo;Traffic violation invigi_lation using transfer learning,\u0026rdquo; in Proceedings of the 2022 International Conference on Electronic Systems and Intelligent Computing (ICESIC), pp. 286\u0026ndash;292, Chennai, India, April 2022. \u003c/li\u003e\n\u003cli\u003eJ. E. Espinosa-Oviedo, S. A. Velast\u0026acute;ın, and J. W. Branch_Bedoya, \u0026ldquo;EspiNet V2: a region based deep learning model for detecting motorcycles in urban scenarios,\u0026rdquo; Dyna, vol. 86, no. 211, pp. 317\u0026ndash;326, 2019. \u003c/li\u003e\n\u003cli\u003eC. Vishnu, D. Singh, C. K. Mohan, and S. Babu, \u0026ldquo;Detection of motorcyclists without helmet in videos using convolutional neural network,\u0026rdquo; in Proceedings of the 2017 International Joint Conference on Neural Networks (IJCNN), pp. 3036\u0026ndash;3041, Anchorage, AK, USA, May 2017. \u003c/li\u003e\n\u003cli\u003eR. R. V. E. Silva, K. R. T. Aires, and R. d. M. S. Veras, \u0026ldquo;Helmet detection on motorcyclists using image descriptors and classifiers,\u0026rdquo; in Proceedings of the 2014 27th SIBGRAPI Con_ference on Graphics, Patterns and Images, pp. 141\u0026ndash;148, Rio de Janeiro, Brazil, October 2014. \u003c/li\u003e\n\u003cli\u003eN. Boonsirisumpun, W. Puarungroj, and P. Wairotchanaphuttha, \u0026ldquo;Automatic detector for bikers with no helmet using deep learning,\u0026rdquo; in Proceedings of the 2018 22nd International Computer Science and Engineering Con_ference (ICSEC), pp. 1\u0026ndash;4, Chiang Mai, (ailand, November 2018. \u003c/li\u003e\n\u003cli\u003eF. W. Siebert and H. Lin, \u0026ldquo;Detecting motorcycle helmet use with deep learning,\u0026rdquo; Accident Analysis \u0026amp; Prevention, vol. 134, Article ID 105319, 2020.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"YOLO, CNN, Darknet, Road safety, Precision, Recall, Helmet","lastPublishedDoi":"10.21203/rs.3.rs-4577583/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4577583/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eIn India, helmets symbolize safety and civic responsibility, bearing cultural significance. However, a 22% increase in accidents and a 17.5% rise in fatalities in 2022-23 underscore the critical importance of helmet compliance beyond legal mandates. Non-compliance not only elevates the risk of injuries and fatalities but also entails legal consequences. Notably, 47,000 Indians died in 2021 due to not wearing helmets, emphasizing the pivotal role of helmet usage in road safety. This research focuses on improving motorcycle helmet detection to ensure compliance and reduce the risk of fatal head injuries for riders, extending its impact beyond geographical limits. While our dataset predominantly draws from Sivasagar, a district in Assam, India, the scope of our research is universally applicable. We employed a comprehensive methodology, comprising data collection, preprocessing, and YOLOv5 model training using the Darknet framework, testing, and evaluation. Analysis of the original YOLOv5 algorithm's performance using Precision-Recall (PR) curves resulted in mAP values of 85.9% for helmets, 88.1% for human heads, and an average of 87%. Subsequently, the proposed YOLOv5 algorithm, achieving mAP values of 93% for helmets, 96.8% for human heads, and a remarkable 94.9% average mAP, demonstrated significant improvements. Comparison revealed a consistent 7\u0026ndash;8.5% higher mAP for helmet and human head detection, underscoring the efficacy of the proposed approach in improving detection capabilities. This research contributes to the broader field of computer vision and its practical applications, particularly in enhancing road safety and averting head injuries among riders, irrespective of their location.\u003c/p\u003e","manuscriptTitle":"Enhanced Precision in Motorcycle Helmet Detection: YOLOv5 and Darknet Approach","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-07-18 04:52:59","doi":"10.21203/rs.3.rs-4577583/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f238372a-2056-431e-b185-9e056d842346","owner":[],"postedDate":"July 18th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2024-08-13T12:39:54+00:00","versionOfRecord":[],"versionCreatedAt":"2024-07-18 04:52:59","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4577583","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4577583","identity":"rs-4577583","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}