基于轻量化卷积神经网络的车道线检测  

Lane Detection Based on a Lightweight Convolutional Neural Network

作　　者：胡杰 熊宗权 徐文才 曹恺 鲁若宇 Hu Jie;Xiong Zongquan;Xu Wencai;Cao Kai;Lu Ruoyu(Hubei Key Laboratory of Modern Auto Parts Technology,School of Automotive Engineering,Wuhan University of Technology,Wuhan 430070,Hubei,China;Auto Parts Technology Hubei Collaborative Innovation Center,School of Automotive Engineering,Wuhan University of Technology,Wuhan 430070,Hubei,China;Hubei Technology Research Center of New Energy and Intelligent Connected Vehicle Engineering,Wuhan 430070,Hubei,China;Dongfeng Yuexiang Technology Co.,Ltd.,Wuhan 430058,Hubei,China)

机构地区：[1]武汉理工大学汽车工程学院,现代汽车零部件技术湖北省重点实验室,湖北武汉430070 [2]武汉理工大学汽车工程学院,汽车零部件技术湖北省协同创新中心,湖北武汉430070 [3]新能源与智能网联车湖北工程技术研究中心,湖北武汉430070 [4]东风悦享科技有限公司,湖北武汉430058

出　　处：《激光与光电子学进展》2022年第10期342-350,共9页Laser & Optoelectronics Progress

基　　金：湖北省技术创新专项(2019AEA169);湖北省科技重大专项(2020AAA001)。

摘　　要：针对当前基于语义分割的车道线检测算法速度与精度不平衡等问题,提出一种优化ERFNet的车道线检测算法。首先设计一个高效的核心模块,通过引入通道分离和通道重组等操作,大幅降低了模型参数量与计算量。其次对下采样进行调整,增加单分支下采样,在减少信息损失的同时提高模型并行度。最后在编码器末端引入特征融合模块,利用空洞卷积扩大感受野,提取不同尺度的车道线特征。在CULane数据集上对本文算法和四种基于语义分割的车道线检测算法进行对比实验,结果表明,在交并比阈值为0.5的情况下,本文提出的算法综合F1评分为73.9%,单帧图像的推理时间可达到12.2 ms,均优于其他四种算法,达到速度与精度的良好平衡。This study proposes an optimized ERFNet lane detection algorithm to reduce the imbalance between the speed and accuracy of current lane detection algorithms based on semantic segmentation. First, an efficient core module is designed;introducing operations such as channel separation and channel reorganization, the number of model parameters and calculations are greatly reduced. Then, the down-sampling is adjusted to increase the single-branch down-sampling process, which improves the model parallelism while reducing information loss. Finally, a feature fusion module is introduced at the end of the encoder, and the receptive field is expanded using dilated convolution to extract differently-scaled lane features. We compare the proposed algorithm with four lane detection algorithms based on semantic segmentation on the CULane dataset. Results show that the comprehensive F1-measure of the proposed algorithm is 73. 9% when the intersection-over-union is 0. 5, and the inference time per image can reach 12. 2 ms,which is superior to the other four models and achieves a good balance between speed and accuracy.

关 键 词：机器视觉 语义分割 车道线检测 空洞卷积 特征融合 

分 类 号：U491.6[交通运输工程—交通运输规划与管理]