Abstract:

For the occlusion problem of pedestrian detection in dense scenes, this study proposes the SC-YOLOv4 crowd detection network based on YOLO. Based on the CSPNet structure of YOLOv4 and combined with the idea of ShuffleNetv2 network, the common convolution structure is improved, and the original common residual module is replaced with the Shuffle Module. A backbone network structure based on S-CSPDarkNet53 is proposed, which preserves the accuracy and reduces the number of network parameters. The centroid prediction module is designed on the basis of retaining the original PANet structure, and the original three output feature layers are replaced with a centroid-based prediction method, that is, the regression and training of the target center point are carried out to calculate the loss, and the original NMS operation is discarded to further improve the detection accuracy in the case of occlusion. The experimental results show that YOLOv4 with S-CSPDarkNet53 structure on the CrowdHuamn data set reduces the amount of parameters and improves the detection speed by 5.2 frame·s^-1 when compared with the original network. Compared with YOLOv4, the final SC-YOLOv4 network improves the detection speed by 4.9 frame·s^-1.

Key words: crowd detection, YOLO, Shuffle Module, center point detection, dense crowd, CrowdHuman, CSPNET, YOLOv4