基于3DLaneTR的三维车道线检测算法  

3D Lane Detection Algorithm Based on 3DLaneTR

作　　者：邹斌[1,2] 刘晶 ZOU Bin;LIU Jing(Hubei Key Laboratory of Advanced Technology for Automotive Components,Wuhan University of Technology,Wuhan 430070,China;Hubei Collaborative Innovation Center for Auto Parts Technology,Wuhan 430070,China)

机构地区：[1]武汉理工大学现代汽车零部件技术湖北省重点实验室,武汉430070 [2]汽车零部件技术湖北省协同创新中心,武汉430070

出　　处：《武汉理工大学学报》2024年第7期146-151,共6页Journal of Wuhan University of Technology

基　　金：湖北省重点研发项目(2020BAB135)。

摘　　要：车道线的细长结构导致其在三维空间中很难被精准定位,将高低层特征融合以获取全局空间结构关系,有利于精确检测出车道线位置。基于上述问题,提出一种基于改进Anchor3DLane的车道线检测算法——3DLaneTR。3DLaneTR在Anchor3DLane的基础上引入特征优化模块,进一步增强网络对全局特征和细节特征的提取能力。通过锚线筛选模块保留最具代表性和信息量的锚线,提高了模型的预测精度。进一步地,为充分利用三维空间信息,3DLaneTR通过注意力机制引入三维位置编码,增强了网络的空间感知能力。在OpenLane三维车道线数据集上,对所提出的3DLaneTR算法进行实验验证和模块消融实验。结果表明,3DLaneTR的F1-score达到58.3%,相对于Anchor3DLane提高了9.8%。The slender structure of lane makes it difficult to accurately locate in 3D space.Integrating high and low level features to capture global spatial structural relationships is beneficial for enhancing the accuracy of lane detection.To tackle the above issues,3DLaneTR is proposed,which builds upon Anchor3DLane.3DLaneTR introduces a feature optimization module,enhancing the network's ability to extract global and detailed features.The anchor filter preserves the most representative and informative anchors,improving the prediction accuracy of the model.Furthermore,to fully utilize 3D spatial information,3DLaneTR introduces 3D position encoding through attention mechanism,enhancing the network's spatial perception ability.Results on OpenLane dataset show that F1 score of 3DLaneTR reaches 58.3%,which has increased by 9.8% compared to Anchor3DLane.

关 键 词：三维车道线检测 高度估计 锚线筛选 3DLaneTR 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]