基于实例分割的车道线检测算法  

Lane Detection Algorithm Based on Instance Segmentation

作　　者：武志斐 李守彪 Wu Zhifei;Li Shoubiao(College of Mechanical and Vehicle Engineering,Taiyuan University of Technology,Taiyuan 030024)

机构地区：[1]太原理工大学机械与运载工程学院,太原030024

出　　处：《汽车工程》2023年第2期263-272,共10页Automotive Engineering

基　　金：山西省基础研究计划(202103021224040);山西省回国留学人员科研项目(2021-050)资助。

摘　　要：为实现在自动驾驶复杂场景下检测数量变化的车道线,提出一种基于实例分割的车道线检测算法。首先以ResNet18网络作为主干网络提取图像特征,采用特征金字塔网络进行特征融合。同时设计一种扩张卷积残差模块来提高检测的精度;然后基于车道线的位置进行实例分割,利用语义分割出的车道线点位置预测对应的聚类点位置,通过对聚类点采用DBSCAN聚类算法实现车道线实例区分。结果表明,该算法能够在复杂的自动驾驶场景下有效地进行多车道线检测,在CULane数据集和TuSimple数据集上的调和平均值分别达到75.2%和97.0%。To realize the detection of lane lines with changing numbers in complex scenarios of autonomous driving,a lane detection algorithm based on instance segmentation is proposed.Firstly,the ResNet18 network is used as the backbone network to extract image features,and the feature pyramid network is used for feature fusion,at the same time a dilated convolutional residual module is designed to improve the detection accuracy.Then instance segmentation is carried out based on the location of lane,and the corresponding clustering point locations is predicted using the lane point locations from the semantic segmentation.Finally,the DBSCAN clustering algorithm is used for the cluster points to realize the lane instance distinction.The results show that the algorithm can effectively detect multi-lane lines in complex and changeable autonomous driving scenarios,and the F1 indicators on the CULane dataset and TuSimple dataset reach 75.2%and 97.0%,respectively.

关 键 词：汽车主动安全 车道线检测 实例分割 聚类 

分 类 号：U463.6[机械工程—车辆工程] TP391.41[交通运输工程—载运工具运用工程]