ORAF-YOLO: A Lightweight Factory Unsafe Behavior Recognition Method With Occlusion-Robust Adaptive Fusion
Ying Qin, Yuanyuan Chen, Yameng Tong, Tongshan Liu, Hao Cao, Fuqiang Wu, Rui Shi, Guowei Zhang, Yuling HE · IEEE Access · 2026
Unsafe-behavior recognition in factory production lines remains challenging because complex backgrounds, frequent occlusion, and strict real-time requirements often degrade detection accuracy and robustness. To address these issues, this paper proposes ORAF-YOLO, a lightweight unsafe-behavior recognition framework based on YOLOv11 for cigarette manufacturing scenarios. ORAF-YOLO integrates four coordinated improvements: SEAM enhances critical local features under occlusion, CGA-Fusion improves adaptive multi-level feature aggregation while suppressing background interference, grouped dual-kernel convolution reduces computational cost while preserving feature representation, and SlideLoss improves convergence stability by dynamically reweighting samples of different difficulty levels. By jointly optimizing feature enhancement, feature fusion, lightweight computation, and training strategy, the proposed method provides a more complete solution for complex industrial unsafe-behavior recognition. Experiments on a self-constructed dataset show that ORAF-YOLO achieves 89.5% Precision, 94.2% mAP50, 68.4% mAP50–95, and 64.20 FPS, outperforming the original YOLOv11 while maintaining real-time inference capability. These results demonstrate that ORAF-YOLO achieves a favorable balance among detection accuracy, robustness, and deployment efficiency in complex industrial environments.