Research on road defect detection based on improved YOLOv8

Currently, the use of autonomous driving technology in China is generally limited by strict lane division and road conditions. In contrast to Tesla's vision-based autonomous driving technology, autonomous driving in China generally relies on radar signal reflection for positioning. As target detection algorithms continue to advance, it is a future trend to introduce computer vision methods into autonomous driving. This article proposes an improved YOLOv8 network model for research using relevant Chinese road datasets. By introducing a better backbone network, convnextv2, based on YOLOv8, it can effectively extract features from the data. Additionally, by introducing the novel WIOU computation function, it can better calculate the rectangular boxes. In this experiment, the improved YOLOv8-ATE achieved an average precision of mAPO.5 and mAPO.5:0.95, which were 18.1 % and 19% higher, respectively, compared to YOLOv8-base. The proposed YOLOv8-ATE model can effectively detect road defects and provide theoretical and technical support for future visual research in autonomous driving.