基于MobileNetV3的车道线检测算法分析  

Analysis of Lane Detection Algorithm Based on MobileNetV3

作　　者：朱冰冰 甘海云 林伟文 ZHU Bingbing;GAN Haiyun;LIN Weiwen(School of Automobile and Transportation,Tianjin University of Technology and Education,Tianjin 300222,China)

机构地区：[1]天津职业技术师范大学汽车与交通学院,天津300222

出　　处：《汽车实用技术》2022年第23期71-76,共6页Automobile Applied Technology

摘　　要：车道线检测是车辆辅助驾驶中的重要一环,为实现对车道线进行准确快速的检测,文章提出一种基于MobileNetV3网络的轻量型车道线检测算法。首先对MobileNetV3网络的深度可分离卷积模块进行改进,同时在其基础上加入空间注意力机制模块;然后将车道线表示为三阶多项式,利用优化的MobileNetV3网络对图像中车道线特征进行提取得到用来拟合三阶多项式的车道线参数;最后构建一种车道线回归模型,通过不断地对车道线参数进行修正以提高车道线检测精度。在Tusimple车道线数据集上的实际测试结果表明,提出的算法其图像帧处理速度为210 fps、检测准确度达到了83.35%,能够实时运行,且具有较高的检测精度。Lane detection is an important part of vehicle assisted driving. In order to realize detecting lane lines accurately and quickly, a lightweight lane detection algorithm based on MobileNetV3 network is proposed. Firstly, the deep separable convolution module of MobileNetV3 network is improved, and the spatial attention mechanism module is added on this basis. Then the lane lines are represented as third-order polynomials, and the lane lines features in the image are extracted by using the optimized MobileNetV3 network to obtain the lane lines parameters used to fit the third-order polynomials. Finally, a regression model of lane-line is constructed to improve the detection accuracy of lane lines by constantly revising the lane lines parameters. The experimental results on Tusimple lane dataset show that the proposed algorithm has a frame processing speed of 210 fps and a detection accuracy of 83.35 %. It can run in real time and has high detection accuracy.

关 键 词：车道线检测 深度卷积神经网络 多项式拟合 辅助驾驶 MobileNetV3网络 

分 类 号：U495[交通运输工程—交通运输规划与管理]