基于改进行方向的车道线检测  

Lane Line Detection Based on Improved Direction

作　　者：鲜阳 谭飞 尹宋麟 XIAN Yang;TAN Fei;YIN Songlin(School of Automation and Information Engineering,Sichuan University of Science&Engineering,Yibin 644005,China)

机构地区：[1]四川轻化工大学自动化与信息工程学院,四川宜宾644005

出　　处：《无线电工程》2023年第7期1587-1595,共9页Radio Engineering

基　　金：国家自然科学基金(61902268);四川省科技计划(2019YFSY0045);省级大学生创新创业训练计划项目(S202010622090,cx2021192)。

摘　　要：近年来车道线检测逐步向深度学习领域发展,然而基于分割的形状描述对车道线效率低下,同时如YOLO系列检测算法不适用于车道线这种细长、弯曲的物体。针对上述问题,提出一种基于改进行方向的车道线检测方法。选取ResNet作为主干网络,加入了特征金字塔网络(Feature Pyramid Network,FPN)用于提取多尺度特征,并使用特征传递架(Feature Transfer Architecture,FTA)模块进行上下文信息的融合。对于车道线实例区分,将实例检测应用于车道线,通过预测热图来检测车道线实例起始点,并回归相对应的一组卷积参数。对于车道线形状预测,采用行分类方法并进行改进。使用逐行位置选择公式来确定车道线的点集,使用车道线在特征图上与真实位置的偏移量来细化每条车道线的形状,降低了FP值,实现了对车道线的预测。在实验平台上使用Tusimple、CuLane两大基准数据集进行验证,取得了良好的指标与检测效果。In recent years,lane line detection has gradually developed into the field of deep learning.However,the shape description based on segmentation is inefficient for lane lines.At the same time,detection algorithms like the YOLO series are not suitable for slender,curved objects such as lane lines.To solve the above problems,a lane line detection method based on improving the direction of travel is proposed.ResNet is selected as the backbone network,the Feature Pyramid Network(FPN)is added to extract multi-scale features,and the Feature Transfer Architecture(FTA)module is used to fuse context information.For lane line instance distinction,instance detection is applied to the lane line,the starting point of the lane line instance is detected by predicting the heat map,and a corresponding set of convolution parameters is regressed.For lane line shape prediction,a row classification method is adopted and improved.The line-by-line position selection formula is used to determine the point set of the lane line,and the offset of the lane line on the feature map and the real position is used to refine the shape of each lane line.This method reduces the FP value,and realizes the prediction of the lane line.Two benchmark datasets,Tusimple and CuLane,are used for verification on the experimental platform,and good indicators and detection results are achieved.

关 键 词：车道线检测 逐行分类 特征传递架 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]