Skeleton Action Recognition Based on Spatial-Temporal Dynamic Topological Representation

Solving the problem of spatial-temporal invariance has always been a hot research topic in the field of skeleton-based action recognition. However, most of the current methods only solve the invariant problem of spatial dimension without considering both temporal and spatial dimensions together. To address above issue, we propose spatial-temporal dynamic topological representations (ST-DTR) to dynamically learn features of spatial-temporal nodes and topological relation, and employ aggregated module to effectively combine spatial-temporal features. At the same time, we adopt the operation of adaptive selection kernel in the temporal dimension for effective spatial-temporal modeling. Specifically, the spatial-temporal joint attention mechanism is introduced to enhance the feature representation and obtain the joints with the plenty information from key frames in the whole skeleton sequence for improving the network identification performance. The effectiveness of the proposed method is evaluated on three standard datasets NW-UCLA, NTU RGB+D 60 and NTU RGB+D 120. Extensive experiments show that our proposed method outperforms some state-of-the-art methods.