Two-stream adaptive graph convolutional network with multi-head attention mechanism for industrial safety detection

With the increasing complexity of industrial environments, safety incidents caused by abnormal behaviors have become more frequent, posing significant threats to productivity and personnel safety. Traditional video surveillance systems exhibit limitations such as poor real-time performance, high false alarm rates, and inadequate adaptability to complex scenes for abnormal behavior detection. To tackle these challenges, this paper introduces a two-stream adaptive graph convolutional network (2s-MA-ARGCN) that leverages a multi-head attention mechanism. This model enhances recognition accuracy and robustness in complex environments by partitioning skeleton data into joint point information and bone information, extracting multi-scale temporal features through dilated convolutions, and effectively modeling long-range dependencies between skeletal points using adaptive attention mechanisms. The superiority of the proposed model in behavior recognition tasks is evidenced by its enhanced accuracy and stability, as demonstrated through evaluations on the NTU-RGBD dataset and an abnormal behavior dataset. This approach holds significant potential for application in safety monitoring systems within complex industrial environments, offering a novel solution for abnormal behavior detection.