Sequence Segmentation Attention Network for Skeleton-Based Action Recognition

With skeleton-based action recognition, it is crucial to recognize the dependencies among joints. However, the current methods are not able to capture the relativity of the various joints among the frames, which is extremely helpful because various parts of the body are moving at the same time. In order to solve this problem, a new sequence segmentation attention network (SSAN) is presented. The successive frames are encoded in each of the segments that make up the skeleton sequence. Then, we provide a self-attention block that may record the associated information among various joints in successive frames. In order to better recognize comparable behavior, a model of external segment action attention is employed to acquire the deep interrelation information among parts. Compared with the most advanced approaches, we have shown that the proposed method performs better on NTU RGB+D and NTU RGB+D 120.