Research of Vehicle Detection under UAV Based on YOLOv4 Improved Algorithm

The improvement of small target detection ability determines the effect of target recognition in the field of target detection. To solve the problems of inaccurate positioning of small target detection, difficulty in feature extraction, and easy-to-miss detection, a small vehicle target detection algorithm based on YOLOv4 was proposed in this paper, namely RC_IYOLOv4, based on VEDAI, a public data set of small vehicle target under UAV. Firstly, the k-Means ++ clustering algorithm retrieves the anchor point box conforming to the VEDAI data set. Then, feature extraction was enhanced by embedding RFB receptive field module and amplifying small target receptive fields by replacing the SPP module with RFB. Finally, this paper adds the CBAM attention mechanism module. It uses one-dimensional convolution to replace the improved combination of complete connection and parallel connection in its interior with focusing on learning the feature representation of small targets. Based on the improved methods above, the experimental results show that the proposed improved algorithm RC_IYOLOv4 improves 4.98% in the VEDAI data set MAP@0.5 compared with YOLOv4. Finally, in the PASCAL VOC data set commonly used for target detection, MAP@0.5 improves by 2.27%.