Object Detection Algorithm Based on Improved YOLOv7 for UAV Images

The technology of unmanned aerial vehicles (UAVs) has matured and widened the range of applications in engineering, agriculture, transportation, surveillance, and so on. Deep learning techniques for object detection assist operators in identifying targets and enhancing efficiency. Thus, it is important to improve the accuracy of object detection. However, due to the limited resources of the UAV platform and the necessity for real-time recognition, striking a balance between accuracy enhancement and decreasing computation is pivotal for development. Thus, we proposed an algorithm based on a YOLOv7 single-stage object detector with a detection head to improve the detection effect of small objects and using a modified feature layer attention module (M-FLAM) at the high-level feature layer to enhance the attention on small objects. The algorithm employed modified efficient layer aggregation network (M-ELAN) to reduce the number of parameters without significant loss in accuracy. Experiments were conducted on the VisDrone dataset. YOLOv7 achieved mean average precision (mAP) @ 0.5 of 46.6%, mAP @ 0.5:0.95 of 26.2%, and the number of parameters was 37.2 million. In comparison to YOLOv7, the developed algorithm demonstrated a 2.3% increase in mAP @ 0.5 and a 2.3% increase in mAP @ 0.5:0.95, with a notable 4.8% reduction in the number of parameters.