Abstract:

In order to effectively identify the unsafe behaviors of construction workers in high-altitude environments, a recognition model based on MCT-GCN was proposed. Firstly, a data preprocessing module was designed to convert video surveillance data into three-dimensional skeleton data. Secondly, a tri-component dynamic adjacency graph convolution was constructed. It integrated learnable structural prior topology, channel correlation adaptive topology, and relative position encoding topology to dynamically build adjacency matrices adapted to different actions. Furthermore, a multi-scale separable temporal convolution was proposed to decompose standard convolutions into deep temporal convolution and point-by-point convolution, thereby independently modeling the temporal characteristics and spatial distribution characteristics of construction workers' actions. Finally, experimental verification and comparative analysis were conducted on both public datasets and a self-built dataset of workers' unsafe behaviors. The results demonstrate that the proposed model outperforms existing methods in terms of recognition accuracy and cross-scene generalization. On the self-built dataset, the model achieved a peak recognition accuracy of 95.8%, significantly enhancing the detection of workers' unsafe behaviors in complex and dynamic construction environments, and making a significant contribution to the development of intelligent monitoring and control in the construction industry.

Key words: multi-component topology graph convolutional network(MCT-GCN), construction workers, unsafe behavior recognition, multi-scale, skeleton