Abstract

In autonomous driving, vehicles are recognized by computer vision and image processing to reduce the risk of accidents. However, traditional vehicle detection algorithms often struggle to deal with the vehicle occlusion problem effectively, necessitating the modification of feature map size as vehicle sizes vary. To address these issues, we proposed RBS-YOLO, a vehicle detection model based on YOLOv5. First, the ResFusion module was designed to expand the model’s receptive field and capture features at various scales. Second, using a bidirectional feature pyramid network, we enhanced the inclusiveness of feature information by fusing them. Finally, the SloU loss function was used instead of the CloU loss function to improve the network’s positioning accuracy and convergence speed. The experimental results indicate that the RBS-YOLO model achieves a precision rate of 97.5 \(\%\) and 72.5 \(\%\) on the UA-DETRAC dataset and BDD-100K dataset, exceeding YOLOv5 by 1.1 \(\%\) and 1.7 \(\%\) .