UAV-YOLOv5: A Lightweight Object Detection Algorithm on Drone-captured Scenarios

Aiming at the problems of common object detection algorithms on drone-captured scenarios, such as too large model, difficult deployment, low accuracy of small-scale object detection, this paper proposed a series of improved methods based on YOLOv5, which effectively improved the performance of the algorithm on drone-captured scenarios. A new dual-branch CSPNet (DB-CSPNet) structure was proposed, which effectively reduced the complexity and computation of the model. A new feature fusion path (FS-FPN) was proposed, which effectively improved the detection accuracy of the model. By integrating a attention mechanism (ACmix), the performance of the model is effectively improved. The experimental results shown that the proposed methods have a significant improvement effect on the accuracy of the object detection algorithm on drone-captured scenarios. The mAP@0.5 and mAP@0.5:0.95 of the algorithm which used the method 2 and 3 proposed in this paper can be improved by 2.5% and 1.6%. At the same time, the method 1 proposed in this paper can also achieve good lightweight effect, the model parameters and FLOPs can be reduced by 26.6% and 30.4%. The UAV-YOLOv5 implemented by all methods in this paper can also achieve a good balance between precision and lightweight. Compared with the default YOLOv5s, the mAP@0.5 and mAP@0.5:0.95 increased by 1.5% and 1.0%, and the parameters and FLOPs decreased by 3.7% and 7.0% respectively.