Multi-Scale Feature Fusion based Abnormal Driver Behavior Detection

Abnormal driver behavior detection should not only focus on key areas such as the face, but also needs to be analyzed by combining global posture. Therefore, a multi-scale feature fusion based abnormal driver behavior detection method MSFYOLO is proposed in this paper. Firstly, a lightweight Multi-scale Feature Reinforcement and Fusion module (MFRF) is proposed to reconcile scale discrepancies between local and global features. Secondly, a Multi-scale Kernel Decomposition Attention (MKDA) module is introduced for deep fusion between entire local regions and global contexts. The shapes and sizes of objects that need to be focused on in driver abnormal driving behavior detection vary significantly. Therefore, Shape-IoU is adopted to introduce both the shape and size information of objects into the loss function. Comparative experiments show MSF-YOLO achieves a 3.3 % mAP@50 increase over the baseline YOLOv8s on the Abnormal Driver Behavior Detection Dataset, outperforming state-of-the-art models (Faster R-CNN, SSD, YOLOv11s) by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$6.0 \%, 17.1 \%$</tex>, and 2.4 %, respectively. Ablation studies show that the MFRF, MKDA, and Shape-IoU contribute <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1.6 \%, 2.1 \%$</tex>, and 0.7 % mAP@ 50 improvements.