Human behavior recognition based on improved slowfast network

Aiming at the problem of low accuracy of human action recognition for small targets and small amplitude actions at present, a human action recognition method with improved SlowFast network structure is proposed. This method uses the SlowFast network as the backbone network to extract features of different scales and inputs the features of multiple scales into the proposed spatiotemporal feature pyramid for processing to increase the network's sensitivity to semantic and spatial information at multiple scales. The experimental results on the public dataset AVA show that this method improves the accuracy of human behavior recognition by 0.77mAP compared with the original network, which is more in line with practical application requirements than similar algorithms.