Research on Driving Behavior Detection Algorithm Based on YOLOv10

To address low accuracy and slow speed in detecting abnormal driving behaviors during daily driving, we propose YOLOv10n-DBD, based on the YOLOv10n baseline model. It integrates the lightweight G-ghostNet for the backbone, enhancing efficiency. The addition of the NAMA-Attention module reduces computational load. To tackle issues with small object detection like cigarettes, DSConv replaces traditional convolutions, improving efficiency. Experimental results show YOLOv10n-DBD improves mAP50 by 5.1%, reduces parameters by 23.4%, and decreases GFLOPs by 13.4% compared to YOLOv10n, enhancing both accuracy and real-time performance in detecting abnormal driving behaviors.