Kinect-Based Dual-Stream Spatiotemporal Convolution Human Behavior Recognition Technology

Aiming at the problems of low recognition rate and insufficient network depth in dual-stream spatiotemporal convolution, this paper proposes an improved dual-stream spatiotemporal convolution method based on Kinect to recognize human actions. Under the original dual-stream spatiotemporal convolution framework, the VGG-16 model is changed to the ResNet-50 model to convolve the bone data collected by Kinect, and then, the spatial and temporal data is fused, connected, and normalized through subsequent layers. The experimental results show that the recognition rates on the two datasets HMDB-51 and UCF-101 reach 69.4% and 90.3%, respectively. Compared with the original two-stream model, the accuracy rate is significantly improved and it is also significantly improved compared with other research methods.