Pedestrian fall detection based on AC-YOLOv5s

Abstract

In view of the serious occlusion phenomenon of pedestrian fall detection, the difficulty of extracting small target details, and the slow detection speed, this paper proposes a high-precision lightweight detection network AC-YOLOv5s. First, the convolution module in the backbone is replaced by ACBConv, and the C3 module is replaced by ACBC3 to improve the detailed feature extraction capability. Secondly, a small target detection layer is added to the feature fusion network (FPN) to improve the detection accuracy of small targets. Finally, use Alpha IoU loss replaces CloU loss to improve the loss and regression accuracy of the High IoU target. Finally, compared with the original YOLOv5s, the network in this paper improves the mAP by 2.33%, and the FPS reaches 21 during detection. The experimental results show that our network achieves better results than other networks.