长沙理工大学学报（自然科学版）2025，Vol.22Issue(5)：1-16,16.DOI:10.19951/j.cnki.1672-9331.clzkb20250304002
基于双目BEV与改进YOLOv8的路面裂缝识别方法
A road crack identification method based on stereo BEV and improved YOLOv8
摘要
Abstract
[Purposes]This study aims to address the issues of inconsistent target scale and insufficient global feature modeling in road crack detection,proposing a road crack identification method based on binocular bird's eye view(BEV)and an improved YOLOv8 model to achieve efficient and accurate crack detection and segmentation.[Methods]High-precision BEV images were generated using binocular stereo vision and inverse perspective mapping(IPM)technology to resolve the scale inconsistency problem in traditional perspectives.The proposed C2f-DRR module captured multi-scale contextual information of cracks effectively by employing a region-residual and semantic-residual decoupling strategy.It combined large kernel convolutions with small kernel dilated convolutions to enrich image detail and reduce background interference.Additionally,a context anchor point attention mechanism was introduced to dynamically focus the model on the central region of cracks and achieve model long-range dependencies between distant pixels.[Findings]To verify the effectiveness of the improved model,comparative experiments are conducted on the test set.The improved model achieves a mean average precision(MAP50)of 83.7％,accuracy(P)of 83.9％,and F1 score of 83.5％,with improvements of 4.4,4.0,and 1.8 percentage points,respectively,over the original YOLOv8n model.The model is also tested on the publicly available UAV-PDD2023 dataset,achieving an MAP50 of 70.5％,recall(R)of 64.8％,and accuracy(P)of 74.1％,with improvements of 3.5,4.5,and 0.6 percentage points,respectively.The improved model outperforms the original model in identification accuracy,robustness,and generalization ability.[Conclusions]The proposed BEV-based crack segmentation method effectively enhances detection accuracy and generalization in complex road environments,providing reliable technical support for automated road damage detection.关键词
路面裂缝识别/双目立体视觉/鸟瞰图/YOLOv8/上下文锚点注意力/特征提取Key words
road crack identification/stereo vision/bird's eye view/YOLOv8/context anchor point attention/feature extraction分类
交通运输引用本文复制引用
谢海波,邱杨航,黄莹颖,张汇祥,蔡生勇,王培玉,萧白..基于双目BEV与改进YOLOv8的路面裂缝识别方法[J].长沙理工大学学报（自然科学版）,2025,22(5):1-16,16.基金项目
国家自然科学基金项目(52478495) （52478495）
湖南省自然科学基金项目(2022JJ50324) Project(52478495)supported by the National Natural Science Foundation of China （2022JJ50324）
Project(2022JJ50324)supported by Hunan Provincial Natural Science Foundation （2022JJ50324）
