Channel Pruning-Based Lightweight YOLOv5 for Pedestrian Object Detection

Channel Pruning (CP) and lightweighting of neural network models has an important place in the field of object detection. In this paper, we introduce a global channel pruning with sparse nature and use this pruning method to achieve model lightweighting in YOLOv5, so as to achieve more lightweight pedestrian object detection while ensuring detection accuracy. We conducted extensive experiments on a large public dataset, CrowdHuman. The results show that the lightened YOLOv5n and YOLOv5m are 72%/40.1% and 80.2%/50% in mAP50 and mAP50-95, respectively, while parameters and GFLOPs are reduced by about 49% and 58%, respectively. The lighter model helps to achieve pedestrian target detection on resource-constrained smart devices.