A Human Motion Detection Algorithm Based on Improved YOLOv5s and ST-GCN

Aiming at the problem that existing human motion detection algorithms have low recognition accuracy on complex backgrounds, this paper proposes a human action recognition algorithm based on improved YOLOv5s and ST-GCN. Firstly, based on YOLOV5s, the human body is extracted by adding a saliency (SANet) module, adding a convolutional attention (CBAM) module, and replacing the loss function to the original network. Secondly, the Alphapose algorithm was used to extract the key points of human bones and input them into the Spatiotemporal graph Convolution network (ST-GCN) to complete the action recognition. Meanwhile, in order to increase the model’s ability of capturing the tiny human actions, the Channel Attention (SENet) module is added on the basis of ST-GCN network. Experiments show that the fusion algorithm can effectively improve the recognition of human behaviors in different scenes, especially the detection of head movements, and the effect is obvious, about 15.3%. The detection of arm movements is the best, with a top5 accuracy of 99%.