基于语义分割的车辆行驶车道定位方法

Vehicle driving lane location method based on semantic segmentation

提出一种基于语义分割的车辆行驶车道定位方法.首先采用“编码器-解码器”网络架构实现多车道语义分割,通过最大池化计算的池化索引来进行非线性上采样,消除上采样的学习需要;然后结合目标检测YOLO v2算法,判断行驶车辆所属车道的位置,从而进行车道定位.利用卡尔斯鲁厄理工学院和丰田美国技术研究院公布的数据集(KITTI)中城市道路(UM)的数据制作训练和测试数据库,并将其公开发布.该算法可以实现端到端训练,网络结构简单、速度快、内存需求低,每帧图像的执行速度在60 ms以内.

A vehicle driving lane positioning method was proposed based on semantic segmentation.Firstly,an“encoder-decoder”network architecture was designed to implement multi-lane semantic segmentation.The decoder uses the pooled index of the largest pooled calculation in the corresponding encoder to perform nonlinear upsampling,eliminating the need for upsampling learning.Then,combine with the target detection YOLO v2 algorithm,and determine the location of the lane to which the vehicle belongs,further realize lane positioning.This project produced a train and test database using urban marked(UM)data from the Karlsruhe Institute of Technology and Toyota dataset(KITTI)and published it.The lane detection algorithm is fast and simple,running at 60 ms each frame,which can be trained end-to-end.