The 'Low Slow and Small' UAV target detection and tracking algorithm based on improved YOLOv7 and DeepSort

To improve the accuracy of Low altitude unmanned aerial vehicle(UAV) target detection and tracking, an improved UAV detection algorithm based on YOLOv7 and DeepSort framework is proposed. The CBAM attention mechanism is introduced into the backbone network of YOLOv7 algorithm to improve feature extraction ability. To improve feature fusion ability at different scales, BiFPN weighted feature pyramid is used to replace PANet, and a small target detection layer is added to improve the detection accuracy of small target UAVs. A "low slow small" human-machine data set is constructed with four types of backgrounds: sky, trees, buildings, and dark conditions. The experimental test is carried out. The results show that the detection part mAP@0.5 of the improved algorithm is improved by 8.6%, and the detection accuracy of small-size and weak-feature targets is improved by about 21%. In the final tracking result, the MOTA index was increased by 24%, and the correct output target box accounted for about 70% of the true target box.