Accurate analysis and identification of underground monitoring videos in coal mines can prevent safety accidents caused by unsafe behaviors of underground personnel and protect their safety. In light of this context, this study enhances the traditional YOLOv11 algorithm for target detection and proposes a fast and effective method for identifying unsafe behaviors among underground personnel in the complex environment of coal mines. Firstly, a statistical analysis of the most common types of unsafe behaviors in current underground coal mines is conducted, exploring the classification of miners’ unsafe behaviors into item-type, action-type, and area-type categories. Secondly, based on the characteristics of these unsafe behaviors, we propose dataset augmentation and denoising preprocessing techniques to enhance fine-grained feature extraction. Simultaneously, we introduce the parameter-free SimAM to improve the saliency mapping of miners’ behaviors. Finally, we optimize the YOLOv11 algorithm by incorporating a function enhancement module and the K-means + + anchor frame, and we propose a dual-model recognition method for target detection that integrates the YOLOv11 algorithm with the YOLOv11-Pose algorithm. To validate the performance of our non-standard miners’ behavior recognition method, we test it on a self-constructed dataset. The research results demonstrate that our method can quickly and effectively recognize unsafe behaviors among underground personnel. Compared to traditional methods, our approach significantly improves recognition accuracy on both the self-constructed dataset and the public dataset, achieving a mean Average Precision (mAP) of 95.7%, an accuracy rate of 95.3%, and a recall rate of 95.1%. These findings are significant for preventing underground safety accidents.
Juan et al. (Fri,) studied this question.