基于CPAM-UFLD的车道线检测方法的研究  

Research on lane line detection method based on CPAM-UFLD

作　　者：陈诚 张新闻 李强 CHEN Cheng;ZHANG Xinwen;LI Qiang(School of Mechanical and Energy Engineering,Zhejiang University ofScience and Technology,Hangzhou 310023,Zhejiang,China)

机构地区：[1]浙江科技大学机械与能源工程学院,杭州310023

出　　处：《浙江科技大学学报》2024年第6期515-528,共14页Journal of Zhejiang University of Science and Technology

基　　金：浙江省“尖兵”“领雁”研发攻关计划项目(2023C01254)。

摘　　要：【目的】为改进车道线检测算法的精度,优化自动驾驶系统中的车道保持,提出了一种新的车道线检测算法——基于空洞金字塔池化的通道和空间注意力机制的车道线检测算法(channel and position attention mechanism with atrous spatial pyramid pooling for ultra fast structure-aware deep lane detection,CPAM-UFLD)。【方法】首先,在超快速结构感知深度车道检测(ultra fast structure-aware deep lane detection,UFLD)算法的基础上融入了空洞金字塔池化模块(atrous spatial pyramid pooling,ASPP),以有效地捕捉车道线图像中不同尺度的特征;使用了通道和空间注意力机制(channel and position attention mechanism,CPAM)以关注图像中的关键区域;同时,为平衡类别权重和提高定位精度,使用了包括加权交叉熵损失函数等的四种损失函数;其次,提出了一种亮度改善模块,该模块旨在提升输入图像的质量,从而增强车道线的识别度。【结果】本算法在TuSimple数据集上的检测精度由原来的95.86%提升至96.56%;同时,在CULane数据集上,检测精度由原来的72.2%提升至73.7%。【结论】通过算法的改进,可以有效提高车道线检测的精度,这为智能网联汽车自动驾驶系统的环境感知提供了理论参考。[Objective]To improve the accuracy of lane line detection algorithms and optimize lane keeping in autonomous driving systems,a novel lane line detection algorithm was proposed on the basis of channel and position attention mechanism with atrous spatial pyramid pooling for ultra fast structure-aware deep lane detection(CPAM-UFLD).[Method]First,the atrous spatial pyramid pooling(ASPP)was integrated into the ultra fast structure-aware deep lane detection(UFLD)algorithm to effectively capture features of different scales in lane line images;a channel and position attention mechanism(CPAM)was employed to focus on key regions in the image;additionally,to balance class weights and enhance positioning accuracy,four types of loss function were harnessed including the weighted cross entropy loss function;then,a brightness enhancement module was proposed with a view to enhancing the quality of input images to improve the recognition of lane lines.[Result]The accuracy of the algorithm on the TuSimple dataset is improved from the original 95.86%to 96.56%;simultaneously,on the CULane dataset,the accuracy is increased from the original 72.2%to 73.7%.[Conclusion]The improved algorithm can effectively enhance the accuracy of lane line detection,providing a theoretical reference for the environmental perception of autonomous driving systems in intelligent connected vehicles.

关 键 词：车道线检测 损失函数 CPAM注意力机制 空洞金字塔池化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]