Evaluating the accuracy of a Vision-Based Algorithm for Groundline Estimation in Trotting Horses Using Multiple Camera Angles

Evaluating the accuracy of a Vision-Based Algorithm for Groundline Estimation in Trotting Horses Using Multiple Camera Angles | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 13 February 2025 V1 Latest version Share on Evaluating the accuracy of a Vision-Based Algorithm for Groundline Estimation in Trotting Horses Using Multiple Camera Angles Authors : Karsten Thuren Key 0009-0004-4678-2316 [email protected] , Katja Berg , Jakob Kirkegaard 0009-0001-5571-3863 , Kristian Ringkjær Andresen , and Sabrina Skov Hansen 0000-0003-4227-9890 Authors Info & Affiliations https://doi.org/10.22541/au.173944665.56886579/v1 Published Veterinary Medicine and Science Version of record Peer review timeline 244 views 175 downloads Contents Abstract Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract Background: Equine lameness diagnosis largely relies on subjective visual assessments, which can be biased. Although marker-based methods, force plates, and inertial measurement units (IMUs) provide objective measurements, they require specialized setups. Vision-based algorithms offer a portable, markerless alternative, but their accuracy needs thorough testing. Objectives: To evaluate a custom vision-based algorithm for estimating the groundline across multiple camera angles, including handheld use in horses trotting on a treadmill. Study design: Experimental comparative study. Methods: Eight Standardbred trotter mares were recorded trotting on a high-speed treadmill using seven iPhones positioned at various heights and angles, including a handheld device. A trained deep neural network algorithm placed 2D keypoints on each video frame. Vertical Displacement Signals (VDS) for the eye, withers, and croup were computed relative to either an algorithm-estimated or a fixed treadmill groundline. Maximum (Maxdiff) and minimum (Mindiff) stride values were compared using Bland-Altman analysis, scatter plots, and histograms. The effect of handheld use on variability and accuracy was assessed by comparing results from a handheld camera to those from a static camera. Results: Groundline estimation closely matched the fixed reference, exhibiting near-zero mean angle error and low mean average error (MAE = 0.45°; n = 242.192). Maxdiff and Mindiff stride-level (n = 36.981) MAE were 0.5 mm, with clinically acceptable additional variability introduced by handheld use at the trial level (Maxdiff and Mindiff MAE < 1.8 mm; n = 357). Main limitations: Treadmill-based data and a single breed/coat colour may limit generalizability to other settings. Conclusions: The vision-based algorithm accurately estimates the groundline and stride VDS parameters from various camera setups, including handheld. Further validation in diverse environments and against other objective gait analysis systems is recommended. Evaluating the accuracy of a Vision-Based Algorithm for Groundline Estimation in Trotting Horses Using Multiple Camera Angles Summary Background: Equine lameness diagnosis largely relies on subjective visual assessments, which can be biased. Although marker-based methods, force plates, and inertial measurement units (IMUs) provide objective measurements, they require specialized setups. Vision-based algorithms offer a portable, markerless alternative, but their accuracy needs thorough testing. Objectives: To evaluate a custom vision-based algorithm for estimating the groundline across multiple camera angles, including handheld use in horses trotting on a treadmill. Study design: Experimental comparative study. Methods: Eight Standardbred trotter mares were recorded trotting on a high-speed treadmill using seven iPhones positioned at various heights and angles, including a handheld device. A trained deep neural network algorithm placed 2D keypoints on each video frame. Vertical Displacement Signals (VDS) for the eye, withers, and croup were computed relative to either an algorithm-estimated or a fixed treadmill groundline. Maximum (Maxdiff) and minimum (Mindiff) stride values were compared using Bland-Altman analysis, scatter plots, and histograms. The effect of handheld use on variability and accuracy was assessed by comparing results from a handheld camera to those from a static camera. Results: Groundline estimation closely matched the fixed reference, exhibiting near-zero mean angle error and low mean average error (MAE = 0.45°; n = 242.192). Maxdiff and Mindiff stride-level (n = 36.981) MAE were 0.5 mm, with clinically acceptable additional variability introduced by handheld use at the trial level (Maxdiff and Mindiff MAE < 1.8 mm; n = 357). Main limitations: Treadmill-based data and a single breed/coat colour may limit generalizability to other settings. Conclusions: The vision-based algorithm accurately estimates the groundline and stride VDS parameters from various camera setups, including handheld. Further validation in diverse environments and against other objective gait analysis systems is recommended. Keywords : equine lameness; deep learning; pose estimation; lameness detection; objective gait analysis; vision-based algorithm; handheld; groundline estimation. 1. Introduction Equine lameness is a significant concern among horse owners and veterinary practitioners, often necessitating objective gait analysis for accurate diagnosis. Traditionally, lameness examination has relied on subjective visual assessment by trained professionals, which can be time-consuming and expensive [1,2]. This method, while valuable, is limited by observer bias and variability in detection sensitivity [3–8]. To enhance objectivity, various technologies have been developed for gait analysis, including force plates, inertial measurement units (IMUs), and optical motion capture systems [2,9]. Force plates provide accurate ground reaction force data but are limited to laboratory settings and can disrupt natural gait [10]. IMUs, attached to the horse’s body, offer portability and uses acceleration and angular velocity measurements for gait analysis [11,12]. However, the data quality may be affected by sensor placement and attachment stability [13,14]. Optical motion capture systems, such as marker-based systems using multiple high-speed cameras, provide detailed kinematic data and have been considered the gold standard for motion analysis [15]. These systems require reflective markers placed on anatomical landmarks, which can be time-consuming to set up and may be sensitive to attachment instability. Additionally, they are confined to controlled environments due to their reliance on fixed camera setups [14]. Recent advancements in computer vision and machine learning have enabled the development of algorithms capable of analysing video data from standard smartphones [16–18], paving the way for marker-less motion capture using video data. This reduces the need for specialised equipment and allows for more natural movement analysis. The algorithms measure movement symmetry from key anatomical points or body-parts, offering a more accessible solution for asymmetry and lameness detection for both professionals and horse owners [16]. In human biomechanics, deep learning-based pose estimation models like OpenPose [19], DeepPoseKit [20] and DeepLabCut [21] have demonstrated high accuracy in tracking body landmarks by comparison with manually annotated keypoints. These models have been adapted for animal studies, including rodent and primate movement analysis [17]. In the equine field, few studies have explored marker-less motion capture for gait analysis, demonstrating the potential of deep learning in equine biomechanics [16,18,22]. It is generally recognized that monitoring the cyclical vertical movements (Figure 1) of key anatomical landmarks during trot, particularly the head, withers, and pelvis, is valuable for evaluating overall gait symmetry and detecting lameness. Of particular significance are the indicators of impact and push-off asymmetry within each stride cycle, quantified by differences in minimum vertical displacement (Mindiff) and maximum vertical displacement (Maxdiff). Among the many systems for objective gait analysis, some (like the one in this study) derive a vertical displacement signal (VDS) referenced to the ground (”ground truth”), while others more commonly analyse the relationship between the first and second harmonics of the VDS [2,12,14,15,23–26]. To date, limited research has focused on the impact of camera positioning and movement (e.g., handheld versus stationary) on the reliability of such gait analysis algorithms. Understanding how these factors affect keypoint detection and symmetry measurements is essential when using a system calculating the VDS from ground truth data, from varying angles and with handheld use. This study aimed to evaluate the accuracy of a vision-based algorithm for estimating the groundline in horses trotting on a treadmill by using multiple camera angles and perspectives to thoroughly stress-test its robustness. Specifically, we compared how fixed versus estimated groundlines affected keypoint placement and consequential VDS, as measured by key gait symmetry metrics. We also investigated the handheld application-related imprecision factor by comparing measurements with those from a static camera placed at an approximate equal height and angle. (Insert Figure 1) Figure 1. Example of the vertical displacement cyclical signal (VDS) from eye, withers or croup. Left: A stride vertical displacement cycle is defined as; the minimum displacement after left hoof-on (\(valley\ left=vl\)) → maximum displacement after left hoof-off (\(peak\ left=pl\)) → minimum displacement after right hoof-on (\(valley\ right=vr\)) → maximum displacement after right hoof-off (\(peak\ right=p\)r). Metrics for lameness quantification are calculated as the difference between the two minima (Mindiff) and the two maxima (Maxdiff) per stride. This example has a Mindiff and a Maxdiff deficit on the right leg. Right: The vertical displacements (red, green, blue arrows) are measured in the 2D sagittal plane of the horse (blue square), when seen from the side, as the distance in pixels from the dynamically estimated groundline (orange) or a fixed groundline to the keypoints. The pixel difference is calibrated to a metric value, knowing the metric withers height. 2. Materials and Methods 2.1. Horses and Equipment Eight horses were selected for the study from the teaching herd at the [masked for review], Department of Veterinary Clinical Sciences. All the horses were familiar with exercise on the high-speed treadmill. Local ethical approval was obtained in accordance with the university ethical guidelines (The Local Ethical and Administrative Committee of the Department of Veterinary Clinical Sciences no. 2024-007). All horses were brown, Standardbred trotter mares. Before inclusion in the study, each horse was trotted by hand on hard tarmac while visually examined by an experienced veterinarian (XXX). All horses were considered fit to work in trot on the treadmill. The horses were trotted at an individual pace on the treadmill, where the trot was uniform and stable, without pushing the horses forward. During trot on the high-speed treadmill, each horse was recorded by seven iPhones (Model 15 Pro, iOS-Version 15.5.1, Apple Inc, Palo Alto, California, United States) positioned around the treadmill. The cameras were placed at various heights and angles to capture different perspectives of the horses’ movement. The handheld camera was operated by the same person (XX), who aimed to walk on the spot approximately at 1-second stride intervals, creating clear camera movement. The treadmill was placed in a room with limited distance from the horse of 300 cm on the right side and 250 cm on the left side. The horses had a weight range of 414-580 kg and a height range of 154-168 cm. The average trotting speed was 4.8 m/s [4.5-5.2 m/s]. The specific camera positions are outlined in Table 1 and Figure 2. Table 1. Camera Setup and Positioning (Insert table 1) (Insert Figure 2) Figure 2. Camera positions and angles relative to the treadmill and horse. The dynamic estimated groundline is visible near the horse’s hooves. All statistical analyses were performed using Microsoft® Excel® (Version 2410, Build 16.0.18129.20158, 64-bit) (Microsoft Corporation, Redmond, WA, USA). 2.2. Data Collection (Insert Figure 3) Figure 3. The study population (n = 8) were recorded from 6 angles as described in figure 1 with iPhones recording at HD, 30 FPS. The frames were keypoint annotated by a trained deep neural network. The groundline were either fixed or dynamically estimated by using the hoof keypoints and the VDS were filtered. The resulting VDS signals were stride split and the matching strides were compared statistically by comparing Mindiff and Maxdiff. The VDS curves on the right illustrate matching strides and a Maxdiff calculation as an example. The horses were recorded from six angles and videos were synchronized with one-frame precision using a light source visible to all cameras. The recordings were captured using an ultrawide lens (13 mm, f/2.2) in 4K at 60 FPS and later converted to HD at 30 FPS before analysis, as this is the standard setting used by the algorithm. The ultrawide lens was selected to ensure that the pixel-to-metric scale matched real-world applications. The vision-based algorithm processed the video segments to estimate a frame-by-frame “dynamic groundline” (as seen in Figure 2) based on hoof keypoints. The control dataset consisted of the exacts same videos where the groundline was fixed. After this point the two datasets had the same downstream data processing. The Vertical Displacement Signal (VDS) (Figure 1) relative to each groundline was calculated for three keypoints: eye, withers, and croup. The data were then filtered and analysed by the algorithm without excluding any outlier strides. The algorithm computed the maximum (Maxdiff) and minimum (Mindiff) differences in VDS across all strides for each keypoint, as described by Keegan et al. [11]. Maxdiff and Mindiff values, derived using both the algorithm-estimated groundline and a fixed groundline, were used for subsequent comparative analysis. 2.2.1 Signal Filtering and Symmetry Metrics Computation The core analysis relied on a trained deep neural network to identify keypoints on the horse from a side-view video. These keypoints included the eye, withers, back, and croup, as well as three points on each leg: hoof, fetlock, and carpus/hock. This model was trained on a large dataset (231 videos, 65,000 frames, and over one million manually annotated keypoints) that included a diverse range of horse types trotting from a side perspective in straight lines, from the centre on the lunge and one horse on a treadmill. The model processed each frame of the recorded video, identifying high-confidence segments where all relevant keypoints were visible. This helped exclude segments unsuitable for analysis, such as moments when head movement obscured the eye keypoint, leading to minor variations in detected strides across cameras for the eye, withers, and croup. To standardize the reference frame, the vision-based algorithm estimated a dynamic groundline based on hoof keypoints (Figure 2). A control dataset was created using the same videos but with a marked and fixed groundline. After marking the groundline each frame was rectified using the same transformation (corresponding to the central frame) throughout the video. After this step, both datasets underwent identical downstream processing. The Vertical Displacement Signal (VDS) (Figure 1) was computed relative to each groundline for the eye, withers, and croup. The recordings were calibrated using the known metric withers height of the horse, allowing vertical displacement to be expressed in millimetres. The raw VDS contained both motion-related signals and noise from factors such as camera movement, keypoint imprecision, and non-periodic horse motions (e.g., sudden head movements affecting the eye’s VDS). To extract relevant frequency components, the expected stride frequency was estimated to account for variations in velocity. A Butterworth high-pass filter was applied, removing low-frequency noise while preserving motion characteristics associated with trotting [28,29]. After filtering, individual strides were extracted based on leg keypoint information including algorithm recognition of left versus right, with stride boundaries identified using expected peaks and valleys [26]. The Maxdiff (maximum VDS difference) and Mindiff (minimum VDS difference) values for each stride were then computed, as described by Keegan et al. [11]. These values, derived using both the dynamic and fixed groundlines, were used for comparative analysis. 2.3. Data Analysis The primary objective of this study’s data analysis was to evaluate the agreement between fixed and algorithm-based (estimated) groundlines in quantifying stride-related variables, namely Maxdiff (peak vertical displacement difference) and Mindiff (lowest vertical displacement difference), across multiple anatomical keypoints. The secondary objective was to evaluate the effect of a handheld application on analysis accuracy. The analysis proceeded as follows: 2.3.1. Groundline Angle Error In addition to vertical displacement metrics, we also evaluated the groundline angle error by comparing the vision-based estimate to the known fixed groundline in all frames. Histograms of signed angle error and absolute angle error were created to reveal potential bias (mean error) and overall accuracy (MAE) in groundline orientation. 2.3.2. Descriptive Statistics For each stride and each of the keypoints eye, withers and croup, Maxdiff and Mindiff were extracted under both the fixed and the estimated groundline conditions. The resulting data were pooled across all cameras and all horses, yielding large sample sizes for both Maxdiff and Mindiff comparisons. We computed mean differences, 95% limits of agreement (LoA), mean absolute differences (MAE) and 99% confidence intervals (CI), to provide an overall summary of measurement error. Histograms of differences provided a quick visual check for systematic over- or under-estimation. Histograms of MAE further characterized how large the deviations were, irrespective of sign. 2.3.2. Scatter and Bland-Altman Plots Scatter plots of fixed vs. estimated groundline measurements (Maxdiff or Mindiff) were generated to visualize how closely the values aligned. To quantify agreement in more detail, Bland-Altman plots were generated for both Maxdiff and Mindiff, showing the difference between estimated and fixed measurements plotted against the mean of these two measurements. The mean difference and the 95% LoA were plotted to identify any systematic bias and examine whether the variance of the differences changed with the magnitude of the measured values. Outliers outside the range of the plot were noted in the figure text. 3. Results 3.1. Groundline Angle Evaluation The first analysis examined how closely the estimated groundline angle matched the fixed reference across all recorded frames (n=242.192). Figure 4 shows histograms of the signed and absolute groundline angle errors. The mean signed angle error was 0.01°, with 95% LoA ranging from −1.17° to +1.20°. The MAE of the groundline angle was 0.45° and the 99% CI fell within ±0.45°. These results indicate zero bias and a tight distribution of angle deviations, suggesting the vision-based algorithm reliably estimates the real ground orientation with a wide range of camera angles. (Insert Figure 4) Figure 4 . Histograms of signed groundline angle error (left) and absolute groundline angle error (right). The mean signed angle error was 0.01°, with 95% limits of agreement (LoA) ranging from −1.17° to +1.20°, indicating high accuracy (low bias). The mean absolute error (MAE) of the groundline angle was 0.45°, and the LoA range shows a low spread, indicating good precision (low variability). 3.2. Stride based comparison between estimated and fixed groundlines 3.2.1. Stride Detection and Missing Data A total of eight horses were recorded, yielding a minimally varying number of detected strides after excluding occluded or incomplete data. The theoretical maximum number of detectable strides, estimated using the mean stride length, was 12.672. The actual detected stride sample sizes for stationary cameras 1–6 were as follows: Eye: n=12.256, Withers: n=12.462, and Croup: n=12.261. This corresponds to a stride detection failure rate of 2–3%. Overall, the algorithm demonstrated a high success rate in detecting and tracking strides, even with very challenging camera angles, with only a small proportion of frames and, consequently, strides discarded due to visual obstructions or tracking interruptions. 3.3.2. Mindiff Agreement For Mindiff (the stride cycle lowest vertical displacement difference), the dataset included 36,981 comparisons between fixed and estimated groundlines across all keypoints (Eye, Withers, and Croup) and strides. Figure 5 presents the scatter plot, Bland-Altman plot, and histograms of the differences. The MSE was -0.01 mm, with a 95% LoA of [-1.39, +1.36] mm, and the MAE was 0.50 mm, with a 99% CI of [0.50, 0.51] mm. These findings confirmed a near zero bias and a narrow spread of differences for Mindiff measures, indicating a robust match between the vision-based and fixed groundline methods for capturing the lowest vertical displacement difference of each stride. (Insert Figure 5) Figure 5. Scatter plot (top), Bland‐Altman plot (middle), and histograms of signed and absolute differences (bottom) for MinDiff values comparing estimated vs. fixed groundlines across all keypoints. The mean difference is -0.01 with 95% LoA (-1.39 to +1.36 mm) and the absolute mean is 0.50, indicating high precision and accuracy. 3.3.2. MaxDiff Agreement For Maxdiff, a total of 36,979 stride comparisons were made, as illustrated in Figure 6. The mean signed difference was 0.07 mm, with a 95% LoA of [-1.31, +1.46] mm, and the MAE was 0.50 mm, with a 99% CI of [0.49, 0.51] mm. The scatter and Bland‐Altman plots show a strong correlation, with most data points tightly clustered around the line of identity and minimal systematic bias. Two outlier strides outside the chosen range of the Bland-Altman plot were noted. (Insert Figure 6) Figure 6 . Scatter plot (top), Bland‐Altman plot (middle), and histograms of signed and absolute differences (bottom) for MaxDiff values comparing estimated vs. fixed groundlines across all keypoints. The mean difference is +0.07 with 95% LoA (-1.31 to +1.46mm) and the absolute mean is 0.50, indicating high precision and accuracy. Bland-Altman outliers outside plot: (19.6, -44.1; -18.8,21.6) 3.3. Comparison Between Handheld and Stationary Cameras After excluding occluded or incomplete data (recording frames with missing keypoint detection), the detected matching stride sample sizes between the handheld camera (Camera 7) and the stationary camera (Camera 4) were as follows: Eye: n=2.053, Withers: n=2.081, and Croup: n=2.095. The theoretical number of detectable strides was estimated to be 2.112, resulting in a stride detection failure rate of 1-3%. 3.3.1. Stride-based Results Table 2 summarizes the comparison between handheld and stationary cameras for the keypoints Eye, Withers, and Croup. For both Maxdiff and Mindiff metrics, minimal bias (mean signed differences close to zero) and narrow 95% LoA were observed for all keypoints. Figure 7 illustrate the overall agreement for all three keypoints (n=6229) with scatterplot, Bland-Altman plot and histograms of signed and absolute differences. Table 2. Stride-based agreement between handheld and stationary cameras (Insert table 2) (Insert Figure 7) 3.3.2. Trial-based Results The videos from the 8 horses were divided into trial-like shorter segments of an average of 17 strides creating a sample size of 357 trials, which were compared between handheld and stationary recording (Table 3). The MAE per trial for all keypoints (Figure 8) were considerably lower than the within stride MAE (Figure 7), indicating a strong agreement between the stationary and handheld camera with minimal bias and error in both Maxdiff and Mindiff measurements. Table 3. Trial-based agreement between handheld and stationary cameras (Insert Table 3) (Insert Figure 8) 4. Discussion Quantitative gait evaluation is increasingly recognized as a key element of lameness detection in horses, improving upon subjective observations that may be prone to bias [8,15]. The theoretical basis for using the Vertical Displacement Signal (VDS) in trotting horses to compute asymmetry measures such as Maxdiff (peak vertical displacement difference) and Mindiff (lowest vertical displacement difference), as well as the clinical relevance of these parameters, is well-established in the literature [2,11,12,15,23–27]. These measures provide key insights into equine locomotor function, aiding in the detection and evaluation of gait asymmetries and potential lameness. In the present study, we investigated how a vision-based algorithm performs when automatically estimating the groundline compared to a fixed groundline reference, under controlled circumstances with horses trotting on a treadmill. We evaluated the consistency of groundline angle error and stride based Maxdiff and Mindiff across a range of camera angles, including handheld use, which is especially relevant for clinical and field applications. The algorithm exhibited a low failure rate, with only a small number of strides failing to be analysed due to missing keypoint detection in a frame of the strides; this highlights a robustness in handling challenging camera angles. One of the objectives was to quantify how reliably the vision-based system could estimate a groundline compared to a fixed horizontal reference. Across more than 240.000 frames, the algorithm showed a mean signed angle error of only 0.01°, with 95% LoA ranging from −1.17° to +1.20°. The MAE was 0.45 °, with a 99% CI interval within ±0.45°. These data suggest that, despite stress-testing with varying camera heights and angles, the system consistently detects ground orientation with minimal bias. Previous methods for gait analysis often require markers or additional calibration objects to establish a precise ground reference [11,23]. In contrast, the current approach leverages automated keypoint tracking of the limbs to derive a frame-by-frame continuous, “dynamic groundline”. This approach may simplify clinical setups by eliminating the need for manual calibration and allowing for greater flexibility in real-world conditions. An accurate groundline is crucial for measuring stride-based metrics such as peak (Maxdiff) and lowest (Mindiff) vertical displacement differences, which are widely used for lameness and symmetry assessments in horses [2,12]. When comparing the estimated versus fixed groundline, both Maxdiff and Mindiff metrics showed near-zero mean differences (±0.1 mm) and a narrow mean absolute error (~0.5 mm). The 95% LoA spanned approximately ±1.4 mm, indicating that automated groundline estimation introduces minimal additional error and no bias into the vertical displacement signals. From a clinical perspective, differences of a few millimetres in Maxdiff or Mindiff are within the range of measurement noise in most gait analysis tools [11]. Accordingly, using the algorithm-estimated groundline appears to be a reliable alternative to traditional manual methods, facilitating accurate stride-by-stride analysis under controlled circumstances (horses trotting on a treadmill). The many different camera angles and perspectives in this study were aimed to stress-test the algorithm to better simulate real-world applications. Handheld use appeals to both clinicians and horse owners, and it is essential for this system, as the algorithm relies on side-view recordings of horses trotting in a straight line or from the centre of the circle when being lunged. However, handheld recording inherently introduces camera movement. When evaluating the algorithm’s performance under handheld use, we found that it maintained accuracy, with only small biases observed. Stride-based Maxdiff and Mindiff comparison between a handheld camera and a stationary reference camera, showed small biases (close to zero) and narrow 95% LoA (within ±10–12 mm for the eye, withers, and croup). The MAE remained well under 5 mm, indicating that handheld camera motion did not substantially degrade VDS calculation. When averaged at the trial level, typically consisting of ~17 strides, the measurement error was even lower, with eye and croup below 1.8 mm. While minor extra noise from camera movement can be expected [14], the results confirm that the algorithm’s stability is generally robust enough for handheld use. The current study used a high-speed treadmill with a single breed (Standardbred trotter mares). Although this ensures consistent speed and straight-line motion, treadmill conditions do not capture all aspects of over-ground locomotion. Coat colour may also influence keypoint detection reliability under different lighting conditions or camera settings. Future work should evaluate algorithm performance in outdoor environments and with a broader set of coat colours and breeds. Another valuable step will be direct comparison against established, validated gait analysis technologies, such as multi-camera motion capture or inertial sensor arrays, to further quantify the algorithm’s absolute accuracy. While this study relied on a fixed groundline as a reference standard, correlating these measurements with an external system could provide further assurance for clinical use. The vision-based method evaluated here for groundline estimation accuracy and effect of handheld use, potentially offers a low-barrier solution, requiring only a smartphone camera and no extensive setup or calibration. This adaptability makes it suitable for everyday clinical contexts and remote telemedicine assessments. Further testing under field conditions is warranted. 5. Conclusion This study demonstrated that a vision-based algorithm can accurately estimate the groundline and compute comparable stride-based symmetry measures in trotting horses under controlled circumstances, across a range of stationary camera positions and a handheld setup. The method produced high-accuracy vertical displacement signals with negligible differences relative to a fixed ground reference. The minimal additional noise from handheld recordings highlights the system’s flexibility, which is relevant for real-world clinical applications. While promising, further testing is needed to validate this approach in varied environments, different lighting conditions, different breeds and coat colours. Comparative assessments against established gait analysis systems would provide an external standard for accuracy. Nevertheless, these results support marker-less smartphone-based gait analysis as a practical, accessible technology for objective lameness detection and broader equine locomotor evaluations. Ethics approval and consent to participate: Local ethical approval was obtained in accordance with the university ethical guidelines (The Local Ethical and Administrative Committee of the [masked for review], Department of Veterinary Clinical Sciences no. 2024-007). Consent to participate was not applicable. Consent for publication: Not applicable. Availability of Data: Competing Interests: Funding: Authors Contribution: Acknowledgments: We would like to thank the staff at the [masked for review], Department of Veterinary Clinical Sciences, for their support with the recordings and data collection. 6. References 1. Hardeman, A.M., Van Weeren, P.R., Serra Bragança, F.M., Warmerdam, H. and Bok, H.G.J. (2022) A first exploration of perceived pros and cons of quantitative gait analysis in equine clinical practice. Equine Vet. Educ. 34 , e438–e444.2. Keegan, K.G. (2007) Evidence-Based Lameness Detection and Quantification. Vet. Clin. North Am. Equine Pract. 23 , 403–423.3. Arkell, M., Archer, R.M., Guitian, F.J. and May, S.A. (2006) Evidence of bias affecting the interpretation of the results of local anaesthetic nerve blocks when assessing lameness in horses. Vet. Rec. 159 , 346–349.4. Fuller, C.J., Bladon, B.M., Driver, A.J. and Barr, A.R.S. (2006) The intra- and inter-assessor reliability of measurement of functional outcome by lameness scoring in horses. Vet. J. Lond. Engl. 1997 171 , 281–286.5. Hammarberg, M., Egenvall, A., Pfau, T. and Rhodin, M. (2016) Rater agreement of visual lameness assessment in horses during lungeing. Equine Vet. J. 48 , 78–82.6. Parkes, R.S.V., Weller, R., Groth, A.M., May, S. and Pfau, T. (2009) Evidence of the development of “domain-restricted” expertise in the recognition of asymmetric motion characteristics of hindlimb lameness in the horse. Equine Vet. J. 41 , 112–117.7. Keegan, K.G., Dent, E.V., Wilson, D.A., Janicek, J., Kramer, J., Lacarrubba, A., Walsh, D.M., Cassells, M.W., Esther, T.M., Schiltz, P., Frees, K.E., Wilhite, C.L., Clark, J.M., Pollitt, C.C., Shaw, R. and Norris, T. (2010) Repeatability of subjective evaluation of lameness in horses. Equine Vet. J. 42 , 92–97.8. Hewetson, M., Christley, R.M., Hunt, I.D. and Voute, L.C. (2006) Investigations of the reliability of observational gait analysis for the assessment of lameness in horses. Vet. Rec. 158 , 852–857.9. Serra Bragança, F.M., Rhodin, M. and van Weeren, P.R. (2018) On the brink of daily clinical application of objective gait analysis: What evidence do we have so far from studies using an induced lameness model? Vet. J. Lond. Engl. 1997 234 , 11–23.10. Weishaupt, M.A., Wiestner, T., Hogg, H.P., Jordan, P. and Auer, J.A. (2004) Compensatory load redistribution of horses with induced weightbearing hindlimb lameness trotting on a treadmill. Equine Vet. J. 36 , 727–733.11. Keegan, K.G., Kramer, J., Yonezawa, Y., Maki, H., Pai, P.F., Dent, E.V., Kellerman, T.E., Wilson, D.A. and Reed, S.K. (2011) Assessment of repeatability of a wireless, inertial sensor-based lameness evaluation system for horses. Am. J. Vet. Res. 72 , 1156–1163.12. Pfau, T., Witte, T.H. and Wilson, A.M. (2005) A method for deriving displacement data during cyclical movement using an inertial sensor. J. Exp. Biol. 208 , 2503–2514.13. Serra Bragança, F.M., Rhodin, M., Wiestner, T., Hernlund, E., Pfau, T., van Weeren, P.R. and Weishaupt, M.A. (2018) Quantification of the effect of instrumentation error in objective gait assessment in the horse on hindlimb symmetry parameters. Equine Vet. J. 50 , 370–376.14. Keegan, K.G., Yonezawa, Y., Pai, P.F., Wilson, D.A. and Kramer, J. (2004) Evaluation of a sensor-based system of motion analysis for detection and quantification of forelimb and hind limb lameness in horses. Am. J. Vet. Res. 65 , 665–670.15. Hardeman, A.M., Egenvall, A., Serra Bragança, F.M., Swagemakers, J.-H., Koene, M.H.W., Roepstorff, L., van Weeren, R. and Byström, A. (2022) Visual lameness assessment in comparison to quantitative gait analysis data in horses. Equine Vet. J. 54 , 1076–1085.16. Lawin, F.J., Byström, A., Roepstorff, C., Rhodin, M., Almlöf, M., Silva, M., Andersen, P.H., Kjellström, H. and Hernlund, E. (2023) Is Markerless More or Less? Comparing a Smartphone Computer Vision Method for Equine Lameness Assessment to Multi-Camera Motion Capture. Anim. Open Access J. MDPI 13 , 390.17. Pereira, T.D., Aldarondo, D.E., Willmore, L., Kislin, M., Wang, S.S.-H., Murthy, M. and Shaevitz, J.W. (2019) Fast animal pose estimation using deep neural networks. Nat. Methods 16 , 117–125.18. Wang, Y., Li, J., Zhang, Y. and Sinnott, R.O. (2021) Identifying lameness in horses through deep learning. In: Proceedings of the 36th Annual ACM Symposium on Applied Computing , Association for Computing Machinery, New York, NY, USA. pp 976–985. https://doi.org/10.1145/3412841.3441973. Accessed October 3, 2024.19. Cao, Z., Hidalgo, G., Simon, T., Wei, S.-E. and Sheikh, Y. (2018) OpenPose: Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields. arXiv.org . https://arxiv.org/abs/1812.08008v2. Accessed October 3, 2024.20. Graving, J.M., Chae, D., Naik, H., Li, L., Koger, B., Costelloe, B.R. and Couzin, I.D. (2019) DeepPoseKit, a software toolkit for fast and robust animal pose estimation using deep learning. eLife 8 , e47994.21. Mathis, A., Mamidanna, P., Cury, K.M., Abe, T., Murthy, V.N., Mathis, M.W. and Bethge, M. (2018) DeepLabCut: markerless pose estimation of user-defined body parts with deep learning. Nat. Neurosci. 21 , 1281–1289.22. Kallerud, A.S., Marques-Smith, P., Bendiksen, H.K. and Fjordbakk, C.T. (2024) Objective movement asymmetry in horses is comparable between markerless technology and sensor-based systems. Equine Vet. J. .23. Roepstorff, C., Dittmann, M.T., Arpagaus, S., Serra Bragança, F.M., Hardeman, A., Persson-Sjödin, E., Roepstorff, L., Gmel, A.I. and Weishaupt, M.A. (2021) Reliable and clinically applicable gait event classification using upper body motion in walking and trotting horses. J. Biomech. 114 , 110146.24. Buchner, H.H., Savelberg, H.H., Schamhardt, H.C. and Barneveld, A. (1996) Head and trunk movement adaptations in horses with experimentally induced fore- or hindlimb lameness. Equine Vet. J. 28 , 71–76.25. Rhodin, M., Smit, I.H., Persson-Sjodin, E., Pfau, T., Gunnarsson, V., Björnsdóttir, S., Zetterberg, E., Clayton, H.M., Hobbs, S.J., Serra Bragança, F. and Hernlund, E. (2022) Timing of Vertical Head, Withers and Pelvis Movements Relative to the Footfalls in Different Equine Gaits and Breeds. Anim. Open Access J. MDPI 12 , 3053.26. Starke, S.D. and Clayton, H.M. (2015) A universal approach to determine footfall timings from kinematics of a single foot marker in hoofed animals. PeerJ 3 , e783.27. Serra Bragança, F.M., Roepstorff, C., Rhodin, M., Pfau, T., van Weeren, P.R. and Roepstorff, L. (2020) Quantitative lameness assessment in the horse based on upper body movement symmetry: The effect of different filtering techniques on the quantification of motion symmetry. Biomed. Signal Process. Control 57 , 101674. Table 1. Camera Setup and Positioning 1 Oblique front right 300 50 100 2 Right side 300 90 50 3 Right side 300 90 100 4 Right side 300 90 160 5 Oblique rear right 300 50 100 6 Left side 250 90 100 7 Handheld right side 300 90 ~160 Table 2. Stride-based agreement between handheld and stationary cameras Eye Maxdiff 0.18 −9.60 to +9.96 3.84 3.70 to 3.97 Eye Mindiff −0.27 −10.32 to +9.78 4.02 3.88 to 4.15 Wither Maxdiff 0.33 −11.85 to +12.51 4.85 4.69 to 5.02 Wither Mindiff −0.02 −11.12 to +11.08 4.42 4.27 to 4.58 Croup Maxdiff −0.13 −9.73 to +9.47 3.86 3.73 to 3.99 Croup Midkiff −0.86 −10.53 to +8.80 3.98 3.85 to 4.11 Table 3. Trial-based agreement between handheld and stationary cameras Eye Maxdiff 0.22 −3.27 to +3.72 1.30 1.07 to 1.54 Eye Mindiff -0.33 −3.76 to +3.09 1.38 1.17 to 1.59 Withers Maxdiff 0.10 −6.17 to +6.37 2.33 1.95 to 2.71 Withers Mindiff -0.21 −5.19 to +4.77 1.74 1.42 to 2.06 Croup Maxdiff -0.00 −4.93 to +4.92 1.70 1.38 to 2.02 Croup Mindiff -0.89 −5.14 to +3.35 1.81 1.55 to 2.07 All Keypoints Maxdiff 0.10 −5.00 to +5.20 1.81 1.61 to 2.00 All Keypoints Mindiff -0.49 −4.83 to +3.86 1.66 1.50 to 1.82 Information & Authors Information Version history V1 Version 1 13 February 2025 Peer review timeline Published Veterinary Medicine and Science Version of Record 30 Dec 2025 Published Copyright This work is licensed under a Non Exclusive No Reuse License. Authors Affiliations Karsten Thuren Key 0009-0004-4678-2316 [email protected] Fredensborg Kommune View all articles by this author Katja Berg Fredensborg Kommune View all articles by this author Jakob Kirkegaard 0009-0001-5571-3863 Fredensborg Kommune View all articles by this author Kristian Ringkjær Andresen Fredensborg Kommune View all articles by this author Sabrina Skov Hansen 0000-0003-4227-9890 Kobenhavns Universitet View all articles by this author Metrics & Citations Metrics Article Usage 244 views 175 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Karsten Thuren Key, Katja Berg, Jakob Kirkegaard, et al. Evaluating the accuracy of a Vision-Based Algorithm for Groundline Estimation in Trotting Horses Using Multiple Camera Angles. Authorea . 13 February 2025. DOI: https://doi.org/10.22541/au.173944665.56886579/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.173944665.56886579/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9ff6ed7a4c041b23',t:'MTc3OTQwMTU5OQ=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();