Lane Detection: A Semantic Segmentation Approach

Abstract

Driver assistance technology allows people not to concentrate on driving a car, in which the lane is the main and important features to guide vehicles and keep them safe. Traditional lane detection algorithms usually require high quality input images and hand-crafted features, that are computationally expensive and sensitive to environment. In order to overcome these problems, this paper comes up with an end-to-end lane detection method based on semantic segmentation, which has an accuracy as other deep learning methods, but has fewer parameters. In this paper, it uses two CNN blocks to extract features, and designs a loss function to help train the network. It has been tested on tuSimple dataset and could detect lanes in most of the images in the dataset.