An efficient YOLOv8-based model with hierarchical feature fusion for enabling real-time detection of miner unsafe behaviors

Abstract Accurate detection of unsafe behaviors in miners is crucial for coal mine safety. Existing underground monitoring systems face three main limitations: difficulty in feature extraction due to insufficient lighting, degradation of small target features in cluttered backgrounds, and localization errors, all of which lead to low recognition accuracy and poor real-time performance. This paper proposes an improved YOLOv8 (Hierarchical Feature Fusion YOLOv8) for high-accuracy real-time detection of unsafe miner behaviors. First, partial convolution (Pconv) is integrated into the backbone network to reduce redundant computations, thereby decreasing the network’s parameter size and complexity. Next, a convolution-attention fusion module (CAFM) is introduced to enhance key features and improve behavior detection performance. Additionally, the hierarchical feature fusion neck (HFFN) enhances small target features through cross-layer aggregation while suppressing background interference. Finally, the proposed Inner-MPDIoU loss function improves localization accuracy and accelerates convergence. Evaluations on the MBD-1500 dataset show that the model achieves 94.9% accuracy with only 9.1 M parameters, maintaining real-time detection at 115 FPS. The balance between computational efficiency and performance significantly enhances coal mine safety monitoring.