Enhanced human pose estimation using attention mechanism and dynamic convolution

In modern industrial settings, especially subway construction, automatically monitoring workers' safety postures is vital for personnel safety and robotic assistance. A posture recognition system's effectiveness hinges on its capacity to capture environmental features. However, subway construction sites are intricate and dynamic, often featuring occlusions and machinery that hinder accurate recognition. The varying camera angles and worker movements further complicate the situation, posing significant challenges to the accuracy and reliability of safety monitoring. These factors make the development of advanced and robust monitoring systems particularly urgent.This study proposes an enhanced Multi-spectral Channel Attention (MCA) mechanism and Dynamic Region-aware Convolution (DRC) to improve the feature extraction of the original Lite-HRNet model. These enhancements lead to the development of the Lite-MDHRNet model, which achieves precise real-time safety monitoring of metro construction personnel through accurate pose recognition. The inclusion of the MCA module enhances the modified model's AP by 0.6% and [email protected] by 0.2%. The DRC module alone boosts the modified model's AP by approximately 0.4% and [email protected] by 0.3%. Together, these two modules result in an overall AP increase of 1.8% and a [email protected] increase of 1.4% for the model. These improvements demonstrate the modified model's superior performance over the original unmodified model in complex multi-scale backgrounds.