面向园区场景的车道线局部定位检测方法研究  

Research on Local Positioning Detection Method of Lane Lines in Park Scenes

作　　者：刘丹萍 汪珺[1] 葛文祥 LIU Dan-ping;WANG Jun;GE Wen-xiang(School of Advanced Manufacturing Engineering, Hefei University, Hefei 230000, China)

机构地区：[1]合肥学院先进制造工程学院,合肥230000

出　　处：《重庆工商大学学报（自然科学版）》2022年第4期19-25,共7页Journal of Chongqing Technology and Business University:Natural Science Edition

基　　金：安徽省高校自然科学研究重点项目(KJ2021A0978);合肥学院科学研究发展基金项目(18ZR02ZDB);国家级大学生创新创业训练项目(202111059041);安徽省大学生创新创业训练项目(S202011059206).

摘　　要：面向校园驾驶场景,提出一种综合性能表现较好的通用车道线检测算法,即局部定位检测法。首先,采用经典图形学与基于噪声容忍的递归神经网络的学习模型相结合的方法,完成车道线所在局部区域的检测,对目标车道线的灰度图像进行霍夫变换和灰度拉伸,设计递归神经网络学习模型,以梯度数据作为引导,对算法模型训练学习,以排除梯度信息相似的干扰物,并识别几何属性相关的不完整车道线形态,进而完成补全工作,应用稀疏惩罚,设计具有噪声容忍的递归学习模型,最大效率地利用具有被污染数据标注的自建车道线图像数据集,以此为基础,采用深度强化学习方法,通过6个标识点对目标车道线进行精确定位,并基于6个准确的定位点,检测和绘制车道线。In the campus driving scene,this paper proposes a general lane line detection algorithm with better comprehensive performance,that is,a local positioning detection method.First,a combination of classical graphics and a learning model based on a noise-tolerant recurrent neural network is used to complete the detection of the local area where the lane line is located,perform Hough transform and gray stretching on the grayscale image of the target lane line,design a recurrent neural network learning model,use the gradient data as a guide to train the algorithm model to learn in order to exclude distractors with similar gradient information and identify geometric attributes related of incomplete lane line morphology,and then complete the complementary work,apply sparse penalties,design recursive learning models with noise tolerance,and maximize the efficiency of using the self-built lane line image dataset with contaminated data annotation.Based on this,a deep reinforcement learning method is used to accurately locate the target lane line through six identification points,and based on the six accurate positioning points,the lane line is detected and drawn.

关 键 词：车道线检测 递归神经网络 深度强化学习 自动驾驶 

分 类 号：U471.15[机械工程—车辆工程]