基于MultiRes+UNet网络的车道线检测算法  

Lane line detection algorithm based on MultiRes+UNet network

作　　者：李梅梅 胡春海[2] 龙平 刘少楠 Li Meimei;Hu Chunhai;Long Ping;Liu Shaonan(School of Computer and Communication Engineering,Northeastern University at Qinhuangdao,Qinhuangdao 066004,China;School of Electrical Engineering,Yanshan University,Qinhuangdao 066004,China)

机构地区：[1]东北大学秦皇岛分校计算机与通信工程学院,秦皇岛066004 [2]燕山大学电气工程学院,秦皇岛066004

出　　处：《电子测量与仪器学报》2020年第9期117-122,共6页Journal of Electronic Measurement and Instrumentation

基　　金：国家自然科学基金(61601105,61602099,61701097);河北省自然科学基金(F2016501073)资助项目。

摘　　要：无人驾驶技术改变人类生活方式,带车道线属性的高精地图,是无人驾驶领域的重要一环。针对现有算法在车道线检测时存在准确率低、效率低等问题提出基于MultiRes+UNet检测方法。该方法通过空洞卷积扩大卷积感受野,从而对全局信息统筹,运用MultiRes block和Res path结构减轻编码器-解码器特征之间的差异,大大降低了内存的需求。实验结果表明,此算法在保证检测准确率的同时,提高了算法的稳定性和运行速率,在纯车道、复合车道、阴影污损车道等多情况下,调和平均值分数分别为0.959、0.942、0.891,该算法存在高效性、高鲁棒性。Unmanned technology changes human lifestyle.The high-precision map with the lane line attribute plays a crucial role in the unmanned field.Proposing a detection method based on MultiRes+UNet network,aiming at the problem that the low accuracy and low efficiency of existing algorithms in the detection of composite lane lines.This method expands the convolution receptive field by dilated convolution to co-ordinate global information.The MultiRes block and Res path structure are used to reduce the difference between the encoder and decoder features,which greatly reduces the memory requirement.The experimental results show that the proposed algorithm improves the detection speed of the algorithm while ensuring the detection accuracy.When the pure lane,compound lane and shadow fouling lanes are obtained,the harmonic mean F1 scores are 0.959,0.942,and 0.891,the algorithm is high efficiency and high robustness.

关 键 词：车道线检测 MultiRes+UNet网络 空洞卷积 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]