Pedestrian Detection Algorithm Based on Improved YOLOv5

Abstract

In order to solve the problem that the object detector is difficult to identify pedestrians at different scales, this paper introduces the MSBlock module into the Backbone layer by improving the YOLOv5 algorithm, enhancing the ability of the real-time object detector to extract multi-scale features and the problem of missed detection and false detection caused by the large difference in the size of the character target, while maintaining the fast inference speed, adding the PConv module to the Head layer to reduce redundant computing and memory access, so as to extract spatial features more effectively. At the same time, the path aggregation structure of multi-feature fusion is realized by fusing the underlying features in the backbone network, and the CIoU loss function is used to reduce the regression error and further improve the accuracy of target detection, so as to achieve accurate identification of pedestrians. Experimental results show that compared with the YOLOv5s algorithm and the original YOLOv5n algorithm, the mAP of the improved YOLOv5 algorithm in the recognition experiment is increased by 1% and 5.3%, respectively. And the improved YOLOv5 algorithm is lighter than YOLOv5s.