VDA-YOLOv8 An Improved YOLOv8 Network Model for Vehicle Detection

Vehicle detection is of great significance to intelligent transportation systems and autonomous driving. However, due to many complex uncertainties associated with real traffic scenarios, such as bad weather, occlusion, motion blur and illuminance differences achieving precise and real-time vehicle detection continues to be a challenging task. To minimize the impact of uncertainties in real-world environments and enhance both the real-time and accuracy of vehicle detection, this paper proposes the VDA-YOLOv8 network model. First, the structural re-parameterization technique is employed to build a C2f-DBB module, which combine the Diverse Branch Block with the C2f to reinforce the feature extraction ability of the network. In addition, Bi-Level Routing Attention is introduced to help the model locate the interest region and suppress useless information. Finally, P2 detection layer is added to make the model more effective in detecting and locating small targets. Our algorithm improves the precision by 3.5%, mAP50 by 4.3%. The detection speed reaches 108.69 FPS. Our algorithm reaches better balance between accuracy and speed compared with other algorithms.