Detection of Smoking and Phone Usage Behavior in Gas Station Scenes Based on YOLOv8

To address the safety threats posed by smoking and phone usage in petrochemical environments, we propose a behavior detection method called BDAF-YOLOv8, based on the YOLOv8 network, which integrates scene context for detecting smoking and phone usage behaviors. This method employs a dual-backbone network(DBN) to extract features, where the auxiliary branch generates reliable gradient information for the main branch, thereby resolving the information bottleneck problem. Deformable Convolutional Networks v3(DCNv3) are combined to form C2fDCNv3 for feature extraction, enhancing the feature extraction capability. Finally, adaptive spatial feature fusion(ASFF) is employed as the detection head, which enhances scale invariance and improves detection accuracy. Experimental results on a self-constructed dataset validate the effectiveness of the proposed improvements. The BDAFYOLOv8 model improves the mean Average Precision (mAP@50) by 4.1% and mAP@50:95 by 4.2% compared to the original YOLOv8n network. On the publicly available VOC dataset, the proposed BDAF-YOLOv8 model achieves a 6.1% improvement in mAP@50 and a 6.7% improvement in mAP@50:95 compared to the original YOLOv8n network. These results verify the effectiveness of our improvements, demonstrating that the method can achieve real-time and accurate detection of smoking and phone usage behaviors in petrochemical scenarios.