In high-risk industrial environments such as tunnel construction, reliable safety helmet detection is critical for preventing head injuries. However, severe illumination inhomogeneity and multi-scale object appearances pose significant challenges to existing detectors due to static anchor designs and the absence of illumination-aware feature learning. This paper proposes AE-LFOG-YOLO, an end-to-end framework that enhances YOLOv8 through dual physics-informed optimizations. The approach integrates an Illumination-Invariant Module (IIM) that employs a dual-path feature decoupling strategy to suppress lighting artifacts within the network backbone. Concurrently, the Adaptive Evolutionary - Light Field Optimized Generation (AE-LFOG) algorithm replaces static anchors with a dynamic evolutionary process guided by local illumination gradients and thin-lens imaging principles, enabling continuous optimization of anchor parameters during training. Evaluated on a real-world tunnel dataset, the method achieves 94.83% mAP@0.5 and significantly improves robustness under challenging illumination variations, as evidenced by a 35.7% extension in effective operating range. These results demonstrate the effectiveness of integrating physical imaging priors into deep learning for robust visual perception in complex industrial scenarios.
Liu et al. (Tue,) studied this question.