Improved pedestrian detection algorithm of Yolov5 in dense scenes

Abstract

Occluded, small-scale pedestrians are easy to occur in dense scenes to solve this problem, an improved Yolov5 is proposed detection algorithm. First of all, feature extraction for small-size pedestrians is insufficient the problem uses SPD-Conv convolution to enhance the complex background small target feature extraction ability. Second, fuse an efficient multi-foot the degree of attention module further enhances the visual area of the target pedestrian feature extraction. Finally, focal eiou loss was used as the loss letter number, so that the regression more attention to high-quality samples, improve the accuracy of the model and robustness. Training and testing were performed on the dataset CrowdHuman in the test, the improved Yolov5 algorithm AP50 can reach 83.6 %, phase for the original algorithm, AP50 and AP50−95 are improved by 2.5 %, respectively. Average accuracy improved by 2.2%. The experimental results validate the proposed method effectiveness in crowded scenarios.