复杂天气条件下基于YOLO-CGT的 自动驾驶车辆检测

打开文本图片集

关键词：自动驾驶；车辆检测；YOLOv8;复杂天气；多尺度特征中图分类号：TP394.1;TP181 文献标识码：Adoi：10.37188/OPE.20253319.3135 CSTR：32169.14.OPE.20253319.3135

Abstract：To mitigate the pronounced decline in vehicle detection performance caused by object blur and occlusion under adverse weather conditions，an enhanced YOLOv8-based vehicle detection algorithm， designated YOLO-CGT，is proposed. Tailored for vehicle-mounted camera imagery，the algorithm incor porates multiple enhancements to the YOLOv8 architecture to substantially improve detection robustness in challenging environments. Specifically，a multi-scale residual aggregation module replaces the original C2f module in the backbone network to increase exploitation of raw feature information and to alleviate gradient vanishing associated with greater network depth. A spatial aggregation module is incorporated to integrate global information extraction with local feature perception. Moreover，a lightweight dynamic detection head is developed to balance detection accuracy and computational eficiency.The conventional IoU metric is supplanted by the Inner-Minimum Points Distance Intersection over Union （Inner-MPDIoU） to reduce bounding-box overlap issues. Trained and validated on a vehicle dataset captured under complex weather conditions，the proposed method attains an average detection accuracy of 81. 4% -an improvement of 6.3% -with 3.259×106 model parameters and a computational cost of 9.7 GFLOPs，demonstrating suitability for lightweight deployment while delivering substantial accuracy gains. These results provide a robust foundation for the safe and reliable operation of autonomous driving systems.

Key words： intelligent driving;vehide detection; YOLOv8;complex weather;multi-scale features

1引言

复杂天气是引发交通事故的重要原因之一。（剩余19408字）