摘要
为了提高车道线损坏检测的效率、准确率和鲁棒性,将人工检测对于车道线损坏程度评定标准定义为车道线的健康度,并对健康度分级,将车道线损坏检测转化为健康度评估分类任务。提出了一种基于深度学习的车道线健康度评估系统,首先利用IPM变换消除车道线透视变形,基于Mask R-CNN网络进行车道线区域分割,然后基于车道线区域分割结果和Res Net网络进行健康度评估。实验证明,该系统在车道线区域检测上的AP指标可以达到0.79,对于车道线健康度评估准确率可达到95%。
To improve the efficiency,the accuracy,and the robustness of the damage detection of lane lines,this article defines the healthy degree of the lane lines according to the standard of manual detection. It converts the damage detection of the lane lines into a classification task. Since deep learning has higher accuracy and robustness than the traditional image processing methods in classification tasks,this paper proposes a health evaluation system for lane lines based on deep learning. It firstly segment the road areas based on Mask R-CNN network. Health assessment is then conducted based on the segmentation results and Res Net network. Experimental results show that the AP index of the system on the lane line area detection can reach 0. 79,and the accuracy rate of the lane line health assessment can reach 95%.
作者
周晓
李新成
王进举
ZHOU Xiao;LI Xincheng;WANG Jinju(School of Mechanical and Electronic Engineering,Wuhan University of Technology,Wuhan 430070,China)
关键词
深度学习
车道线
健康度评估
区域检测
分类
deep learning
lane line
healthy degree assessment
area detection
classification
作者简介
周晓(1979-),男,湖北武汉人,武汉理工大学机电工程学院副教授.
