Abstract:Addressing the critical challenges in identifying on-site workers under occlusion conditions, this paper delves into and proposes an innovative method for global feature representation based on mutual learning. This approach ingeniously constructs a multi-branch network architecture, where each branch independently focuses on processing images of workers in both occluded and non-occluded scenarios. Through a mutual learning mechanism, it achieves complementary and collective enhancement of feature extraction capabilities across these two extreme scenarios. Specifically, we employ category mutual learning and feature contrastive learning strategies, which not only facilitate category consistency between occluded and non-occluded data within the model, but also strengthen the mutual reinforcement at the feature level, thereby significantly boosting the robustness and discriminability of the global feature representation. Furthermore, to better align with real-world operational settings, we introduce a data augmentation technique based on realistic scenario simulation, which generates highly realistic occluded data, effectively elevating the model"s generalization ability in complex occluded environments. Experimental results demonstrate that this method exhibits outstanding performance in identifying on-site workers under occlusion conditions, thereby providing a solid technical foundation and innovative solution for enhancing on-site work safety management efficiency and ensuring personnel safety.