Multi-Streams Network for Action Recognition

In this paper, combined with depth information and rich texture information in RGB video sequences, a multi-stream behavior recognition algorithm is proposed. The algorithm uses DenseNet as the main network to obtain color texture information, optical flow information and depth information, and uses them as input to the information flow network. Then use LSTMs for feature fusion and behavior classification. Through experiments, the recognition rate on the UTD-MHAD data set of the public action recognition library is 93.88%, which is significantly better than similar algorithms.