Drone-View Detection of YOLO via Tandem Cross Self-Attention and Contextual-Aware Module

Abstract Unmanned Aerial Vehicle (UAV) target detection involves automatically labeling objects captured during flight based on meaningful features. Compared to tra ditional images, aerial images have a higher viewpoint and often contain small objects and clustered features. Recent studies have demonstrated the advantages of deep learning in recognizing UAV images. However, early methods often had overlapping convolutional layers, which render the original contextual informa tion irrelevant. Besides, the imbalanced distribution of objects in the UAV field of view is prone to introduce the biases in detector attention, which will weaken the model training. This paper proposes an improved YOLOv5 framework for extracting multiple contextual information and enhancing dense region in UAV target detection. First, a context-aware module is introduced into DarkNet so that the multiple contextual information can be integrated skillfully. Addition ally, we propose a novel context semantic fusion approach, which contributes to integrate the varying expansion rates and deformable convolutions. This method effectively expands the perception field and enhances the precision of object boundaries. Finally, a concatenated cross self-attention operation is designed on high-resolution images to focus on dense regions and extract contextual infor mation from each pixel. Experimental results show improvements with average precision (AP50) of 5.9% and 3.89% on the VisDrone and UAVDT datasets, respectively. The findings demonstrate superior performance of the proposed method in UAV detection tasks and its potential application prospects.