Research On Human Action Recognition Based On Spatial-Temporal Graph Convolution

As one of the important research topics in video understanding,human action recognition has broad prospects in the fields of intelligent surveillance,autonomous driving,and intelligent sports analysis which has received extensive attention from researchers.In recent years,action recognition methods based on spatial-temporal graph convolution have significantly improved the recognition effect,while on the other hand,bringing new concerns worthy of further exploration.This thesis focuses on research on single-person action recognition and interaction recognition between two persons based on spatial-temporal graph convolution.In the research on single-person action recognition,in order to solve two problems existing in current methods that the spatial receptive field of the dynamic graph generation process is restricted and the predefined static graph limits the efficiency of feature aggregation,this thesis proposes a single-person action recognition method based on a novel network called graph involution network.In terms of the improvement of dynamic graph generation,an involution module is proposed to replace the conventional self-attention mechanism based dynamic graph generation and feature aggregation process,providing a global spatial receptive field for each vertex in the graph,and the resolution of dynamic feature along channel dimension is improved at the same time.For the static graph,instead of being restricted by human natural skeleton structure,a novel spiral human connection topology is designed to enhance feature aggregation efficiency between different human body key points.Experiments on public datasets verify the effectiveness and advancement of the proposed method.In the research on human interaction recognition,this thesis proposes a recognition method based on spatial-temporal interaction graph convolutional network in view of the lack of consideration of the spatialtemporal co-occurrence interaction characteristics in current methods.In order to make full use of the spatial-temporal information of both individuals and interactions between individuals,a joint framework of single-person and interaction feature extraction is designed,which combines two parts in a modular form to maximize the extraction effect.For interaction feature extraction,a novel spatial-temporal feature interaction module is proposed.Based on the temporal association and alignment based on the dynamic time warping algorithm,in combination with the interaction spatial graph,a spatial-temporal interaction graph structure is finally constructed to realize the co-occurrence feature extraction.Experiments on public datasets verify the key role played by the proposed spatial-temporal feature interaction module in interaction recognition,and compared with state-of-the-art methods,STI-GCN proposed in this thesis achieves higher recognition performance.