欢迎访问《汽车安全与节能学报》，
汽车安全与节能学报

JASE ›› 2019, Vol. 10 ›› Issue (3): 334-341.DOI: 10.3969/j.issn.1674-8484.2019.03.009

• 汽车安全 • 上一篇    下一篇

基于全卷积网络的车道区域分割算法

魏民祥，滕德成   

  1. （南京航空航天大学 能源与动力学院，南京 210016，中国）
  • 收稿日期:2019-03-19 出版日期:2019-09-30 发布日期:2019-10-01
  • 作者简介:第一作者：魏民祥 (1963—)，男( 汉)，山东，教授。E-mail: weimx@nuaa.edu.cn。 第二作者：滕德成 (1993—)，男( 汉)，江苏，硕士研究生。E-mail: tengdecheng@outlook.com。

Algorithm for lane region segmentation based on fullyconvolutional-network

WEI Minxiang, TENG Decheng   

  1. (College of Energy and Power Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China)
  • Received:2019-03-19 Online:2019-09-30 Published:2019-10-01

PDF

416

摘要：

       为智能车辆的轻量化、实时性检测，提高车道识别的准确度、鲁棒性，提出了一种利用全卷 积网络(FCN)实现车道区域分割的方法。采用一种对称结构的全卷积网络对车道区域作逐像素预测： 利用卷积、池化提取车道特征，利用池化索引辅助上采样，用卷积来恢复特征信息。在既定网络结构下， 比较3×3、5×5和7×7尺寸的卷积核对模型性能的影响。基于FCN-32s 和FCN-16s，分别设计混 叠结构的FCN和无混叠结构的FCN与本网络作测试对比。结果表明：该算法对车道分割准确、鲁 棒性强、实时处理能力优秀，分割效果优于传统FCN；在3种不同尺寸中，小尺寸(3×3)卷积核的 实时处理速率最高，达53 帧/s。因此，该算法适合自动驾驶道路感知任务。

关键词: 智能车辆 , 车道识别 , 实时检测 , 车道区域分割 , 深度学习 , 全卷积网络 (FCN) , 卷积核尺寸

Abstract:

 A method of implementing lane region segmentation was proposed by using fully convolutional network (FCN) for lightweight and real-time detection of intelligent vehicles with a high accuracy and robustness of lane recognition. A symmetry-structured fully convolutional network was used to predict the lane area pixel by pixel: the convolution and pooling were used to extract the lane features, and the up sampling wes aided by pooling indices, and convolution were used to recover the feature information. Under the established network structure, the effects of convolution kernels of 3×3, 5×5 and 7×7 sizes on the performance of the model were compared. The FCN with skip layers and the FCN without skip layers were compared with the proposed network based on FCN-32s and FCN-16s. The results show that the proposed algorithm is accurate, robust and accurate in real-time processing, and the segmentation is better than traditional FCNs. The small convolution kernel (3×3) method has the best real-time handling speed of 53 frames per second among the three different sizes. Therefore, the proposed algorithm is suitable for road perception for autonomous driving.

Key words: intelligent vehicles  , lane detection , real-time detection , lane region segmentation , deep learning , fully convolutional network (FCN) ,  size of convolution kernel