改进UNet与DBSCAN融合的车道线检测方法  

Lane Line Detection Method Integrating Improved UNet and DBSCAN

作　　者：孙浩铭 夏金泽 胡盛辉 梁冬泰[1,2] Sun Haoming;Xia Jinze;Hu Shenghu

机构地区：[1]宁波大学机械工程与力学学院,浙江宁波315211 [2]宁波大学浙江省零件轧制成形技术研究重点实验室,浙江宁波315211

出　　处：《机械制造》2023年第4期68-75,61,共9页Machinery

基　　金：浙江省公益技术应用研究计划项目(编号:LGG21E050008);宁波市交通运输科技项目(编号:202105);宁波市公益性科技计划项目(编号:202002N3133)

摘　　要：室内外标记线和车道线检测是移动作业机器人研究的难点这一。针对现有方法效率低、精度差等问题,提出了改进UNet与DBSCAN融合的车道线检测方法。基于分布式融合结构,将图像输入嵌有并行融合注意力模块的UNet,通过注意力机制加大车道线权重,并抑制干扰信息,以获得更好的识别结果。三维激光雷达的点云数据通过随机抽样一致算法和自适应邻域DBSCAN获得车道线原始点云,经过最小二乘拟合得到点云车道线。在像素坐标系上引入匈牙利算法匹配决策判断,实现车道线拟合。基于KITTI数据集和真实数据集进行试验验证,识别结果优于UNet和LaneNet,检测时间平均约为0.5 s,相比纯视觉方法具有较好的鲁棒性和准确性。Indoor and outdoor marking lines and lane line detection is a difficult point in the research of mobile work robot.In view of the problems of low efficiency and poor accuracy of existing methods,a lane line detection method integrating improved UNet and DBSCAN was proposed.Based on the distributed fusion structure,the image is input UNet embedded with parallel fusion attention module,and the lane line weight is increased through the attention mechanism,and the interference information is suppressed to obtain better recognition result.The original point cloud of lane line is obtained by random sampling consensus algorithm and adaptive neighborhood DBSCAN based on point cloud data of 3D lidar,and the point cloud lane line is obtained by least squares fitting.The Hungarian algorithm is introduced to match the decision and judgment on the pixel coordinate system to realize lane line fitting.Test was carried out based on the KITTI dataset and the real dataset,the recognition results are better than UNet and LaneNet,and the average detection time is about 0.5 s,which has better robustness and accuracy than the pure visual method.

关 键 词：车道线 检测 深度学习 聚类 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]