SCU-YOLOv13算法及其路面坑洼检测研究

打开文本图片集

中图分类号：TP391.4； U418.6+2 文献标识码：A

Abstract： For the road pothole detection task，based on the YOLOvl3 model，the region attention mechanism is replaced with a Single-Head Self-Attention （SHSA） module to reduce computational complexity while enhancing global context modeling capability. A Convolution-Attention Fusion module and a multi-scale feedforward network are introduced to construct the CAMixing block （Convolution Attention Mixing block）， strengthening the fusion of local details and global semantics. The Unified-Intersection over Union （UIoU） loss function is employed for bounding box regression，improving localization accuracy through a dynamic scaling mechanism and dual attention strategy. By integrating the three modules—SHSA，CAMixing，and UIoU—into the YOLOv13 model， a novel SCU-YOLOvl3 algorithm is proposed. Experimental results show that SCU-YOLOv13 maintains high detection eficiency while increasing precision and recall by 1.8% （204号 and 56.7% ，respectively. The mean average precision （mAP） at different IoU thresholds improves by 7.2% ， 6.9% ，and 10.8% ，demonstrating its particular suitability for road pothole detection in complex scenarios.

Keywords： YOLOvl3； Single-Head Self-Attention; Convolution-Attention Fusion; Multi-Scale Feed-Forward Network;UIoU

路面环境检测与评估是智能汽车道路环境感知系统的核心组成部分。（剩余10269字）