Real-Time Worker Behavior Recognition in Power Construction Using Selective State Space Model

In the high-risk domain of power construction, real-time recognition of worker behavior is essential for improving safety. This paper introduces a novel approach that utilizes the Mamba framework—an advanced Selective State Space Model—for real-time behavior recognition. The method begins with YOLOv8 for accurate human detection and pose estimation, allowing precise identification of worker poses in video frames. These poses are subsequently analyzed using GCN Encoder to derive feature vectors representing the workers’ poses. Temporal sequences of these feature vectors are then input into the Mamba model, which, due to its linear-time complexity, efficiently processes these sequences to recognize specific behaviors. Experimental results demonstrate the model’s effectiveness, with enhanced recognition rates and real-time processing capabilities. This system’s ability to provide immediate feedback on unsafe practices has the potential to significantly advance safety protocols in construction and related industries.