基于注意力机制的GCN-Bi-LSTM动作识别方法OA

In response to the challenges posed by existing action recognition methods that rely on skeleton information,particularly their low utilization of skeletal sequence data and difficulties in accurate recogni-tion,this paper presents a novel action recognition approach utilizing a graph convolutional bidirectional long short-term memory(GCN-Bi-LSTM)network enhanced by an attention mechanism.Firstly,a non-physical dependency relationship is constructed by leveraging the relative distances between skeleton nodes to enrich skeletal features.Secondly,a spatio-temporal graph convolutional network is employed to extract features from each video frame,thereby obtaining more sophisticated semantic representations.Finally,these frame-specific features are fed into the bidirectional LSTM network augmented with an attention mechanism to capture global temporal characteristics and then action discrimination capabilities can be en-hanced effectively.Experimental results show that the proposed method can significantly improve recogni-tion accuracy and has great potential for application in tactical action analysis and training scenarios.