Lightweight mine personnel behaviors recognition based on improved YOLOv8n

Abstract In recent years, the recognition of personnel behaviors in high-risk mining environments has faced challenges such as poor robustness, high computational cost, and insufficient accuracy. To address these issues, this paper proposes an improved lightweight YOLOv8n recognition algorithm: the Dynamic_HGBlock module is introduced into the backbone network to enhance feature extraction capabilities under complex conditions such as low light, dust, and occlusion; the C2f_MLCA module integrates a multi-level contextual attention mechanism, optimizing the structure to reduce parameters and computational cost, while improving the fine-grained recognition accuracy of hazardous actions; the detection head incorporates Detect_LSDECD, which enhances localization ability for ambiguous boundaries and overlapping multi-target regions through local structure perception and edge dynamic regulation. Experimental results on the dsmlf and CUMT-HelmeT datasets show that the improved model achieves a 2.4% increase in average accuracy to 91.0%, a reduction of GFLOPS by 4.2G to 4.0G, and a decrease in parameters by 1.1M to 1.9M, significantly improving detection accuracy and computational efficiency, providing reliable technical support for intelligent mine monitoring.