摘要
为了解决传统的车道线检测算法对光照变化、阴影遮挡等环境干扰较为敏感而导致鲁棒性不足的问题,提出了一种基于实例分割和自适应透视变换算法的多车道线检测方法.该算法首先通过设计的多分支实例分割网络实现多车道线分割,该多分支实例分割网络包括车道线语义分割分支和车道线Id分支;再应用自适应透视变换模型获得鸟瞰图视角下的实例分割后的车道线像素点集合;最后利用最小二乘法二阶多项式完成车道线像素点的拟合.基于Culane车道线数据集进行训练及验证,验证表明,每帧图片检测用时约28 ms,车道线检测准确率达91.4%.将车道线检测模型集成到实车ROS平台进行测试,测试表明,所提算法能够实现各类复杂交通场景下的多车道线实时检测.
To solve the problem that the traditional lane detection algorithm is sensitive to the environment interference such as illumination and occlusion,which leads to insufficient robustness,a multi-lane detection method based on instance segmentation and the adaptive transformation algorithm is proposed.Firstly,the multi-branch instance segmentation network is designed to achieve multi-lane line segmentation,and the multi-branch instance segmentation network includes lane line semantic segmentation branch and lane line Id branch.Then,an adaptive perspective transformation model is employed to obtain the set of lane pixel points from the bird's eye view.Finally,the lane line fitting is completed by using the least square method.Using the Culane dataset for training and verification,each frame takes about 28 ms,and the detection rate is 91.4%.Integrating the lane detection model into the ROS(Robot Operating Systems)platform,the experimental results show that the proposed algorithm can achieve the real-time detection of multiple lane lines in various complex traffic scenarios.
作者
蔡英凤
张田田
王海
李祎承
孙晓强
陈龙
Cai Yingfeng;Zhang Tiantian;Wang Hai;Li Yicheng;Sun Xiaoqiang;Chen Long(Institute of Automotive Engineering,Jiangsu University,Zhenjiang 212013,China;School of Automotive and Traffic Engineering,Jiangsu University,Zhenjiang 212013,China)
出处
《东南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2020年第4期775-781,共7页
Journal of Southeast University：Natural Science Edition
基金
国家重点研发计划资助项目(2017YFB0102603,2018YFB0105003)
国家自然科学基金资助项目(51875255,61601203,61773184,U1564201,U1664258,U1764257,U1762264)
江苏省自然科学基金资助项目(BK20180100)
江苏省战略性新兴产业发展重大专项资助项目(苏发改高技发(2016)1094号)
镇江市重点研发计划资助项目(GY2017006)。
关键词
车道线
深度学习
实例分割
自适应透视变换
lane line
deep learning
instance segmentation
adaptive perspective transformation
