Human skeleton behavior recognition method based on self-attention graph convolution

The invention discloses a human skeleton behavior recognition method based on self-attention graph convolution, and the method comprises a space mixing module MGS combining graph convolution and self-attention, a space self-attention module TSSA specific to a time frame, and a multi-scale time convolution module MS-TCN. In the MGS, local and global relations between nodes are modeled by executing graph convolution and self-attention operation in parallel, cross-branch bidirectional interaction is carried out between two branches, information complementation in channel and spatial dimensions is realized, a TSSA module uses an intra-frame spatial relation in a self-attention learning behavior sequence, and the self-attention learning behavior sequence is obtained. The unique spatial features of a single-frame human skeleton are modeled, the MS-TCN module adopts a multi-branch design, and the time receptive field is expanded by using time convolution with different voidage. The MGS, the TSSA and the MS-TCN are c