MF-YOLO: A Lightweight Method for Real-Time Dangerous Driving Behavior Detection

Dangerous driving behavior is a serious issue leading to harm drivers and further increase traffic burden. A You Only Look Once (YOLO) model is a commonly used fast detection model suitable for real-time dangerous driving behavior detection with poor detection performance. To address this problem, a lightweight object detection model called multiple fusion YOLO (MF-YOLO) model is proposed to show the superior capability in small target detection and compatibility with mobile chipsets. First, we design a novel backbone using convolution and vision transformer (ViT) multifusion blocks to fuse local and global context information. Second, a lightweight feature pyramid network (FPN) neck is developed to reduce model complexity and enhance feature extraction ability. Third, an attention mechanism is added to the neck for concentrating the YOLO model on relevant information during feature fusion. Finally, the activation function of fractional rectified linear unit (FReLU) equipped with spatial intersection over union (SIoU) loss function to improve model speed and accuracy. Experimental results from our self-built driving scenario dataset indicate that MF-YOLO achieved mean average precision (mAP) of 91.4%, surpassing YOLOv5n by 6.4%, and even outperforming the latest YOLOv8n by 2.3%.