Research on Human Behavior Recognition Method Based on Static and Dynamic History Sequence
Feng Xiufang, Dong Xiaoyu · 2020
In this paper, a behavior recognition model based on CNN and RNN is proposed, which has high efficiency and high performance. Firstly, a key frame selected method based on the motion variation curve can intuitively represent the global change of motion, and reduce redundant information to construct the optimal key frame sequence. Secondly, the static and dynamic history sequence as the input characteristics of the model is designed based on sparse sampling method. Finally, a new Progressive network model with pyramid pooling to extract features effectively. Next, a bi-directional gate recurrent unit (Bi-GRU) is designed, which encodes all frame features of video successively. At the same time, an attention mechanism is used to dynamically assign weight to the frame in the video, and make the model can actively focus on the key information in the video clip. Moreover, the method is validated in the data sets Ut-interaction and HMDB51 and showed a better recognition performances.