부산대학교 도서관

Pedestrian fall detection, as one of the important techniques for human behavior recognition tasks, plays a key role in intelligent surveillance systems. However, the current mainstream behavior recognition method, Spatial Temporal Graph Convolutional Network (ST-GCN), is still facing challenges such as high false alarm rate and large leakage rate in practical applications. It is because the existence of neighborhood constraints and the difficulty in modeling the long-distance dependency relationship of inter-frame nodes. In this regard, this paper proposes ST -GCNFormer network on the basis of ST -GCN. First, in order to enhance the flexibility of information transfer between nodes, an adjacency matrix with adaptive learning parameters is introduced into the spatial temporal graph convolutional layer to optimize the graph structure. Then, to solve the long-range dependency problem, the Transformer self- attention layer is designed to enhance the global context extraction capability of the network. Finally, the local information perception advantage of ST -GCN is combined with the long-distance dependency modeling of Transformer self-attention layer. The experimental results show that ST -GCNFormer achieves 95.9% and 81.0% accuracy on the mainstream behavior recognition datasets NTU60 (X-View) and NTU120 (X-Set), respectively. It exceeds the baseline network ST-GCN's 7.6% and 7.8%, and has an accuracy of 97.0 % on the Le2i fall dataset. It verifies that the proposed ST -GCNFormer network can improve the accuracy of fall detection technique in practical applications.