Single Soldier Action Recognition Based on Graph Convolution and Rule Matching
Aiming at the problem that the action recognition method based on skeleton data has insufficient semantic understanding and incomplete acquisition of skeleton data, resulting in low recognition accuracy, a single soldier action recognition method based on graph convolution and rule matching based on fusion semantic analysis is proposed. Firstly, the OpenPose pose estimation model is used to extract the skeletal key points of the soldier combat video. Then, according to the effective skeleton key point extraction, the You Only Look Once ( YOLO) based individual action recognition method or the graph convolution based action recognition method are dynamically selected. Finally, a rule matching algorithm is introduced to further complete individual action recognition and judgment for the low confidence discrimination results of graph convolutional network. Experimental results show that compared with Spatial Temporal Graph Convolutional Networks( ST-GCN) algorithm and Two-Stream Adaptive Graph Convolutional Networks ( 2s-AGCN ) lgorithm, this method improves the accuracy of individual action recognition tasks by about 38％ and 11％.
万方数据
维普
