Unmanned Aerial Vehicle (UAV) Object Detection in High-Resolution Image Based on Improved YOLO v5

Abstract:
Unmanned Aerial Vehicle (UAV) detection is important to limit illegal UAV activity. The object detection method based on computer vision has the advantages of low cost and strong discrimination. Due to the small size of UAV, existing object detection methods are still insufficient in the detection performance of UAV. In this paper, we make targeted improvements based on the powerful YOLO v5 to improve the detection performance of small objects. First, we take the divide and conquer approach, the high-resolution image is cut into multiple image blocks to be detected separately, and the pure sky region without object is removed through the sky region recognition algorithm. In terms of network structure, to utilize the spatial information in the shallow layers, the neck and detection head of YOLO v5 are shifted to the shallower layer of the backbone. Meanwhile, Image Pyramid Guidance Network is also introduced into the network structure of YOLO v5. Experiments show that the proposed approach achieved Recall of 0.881 and mAP0.5 of 0.83 for small UAVs in our data set, significant improvement compared to before improvement.