Abstract: This article proposes a lane line detection and fitting method based on the results of semantic segmentation. The detection of ground landmarks such as lane lines is a significant content of the environment perception of autonomous vehicles, which can provide the vehicle with information about the driving area. The vehicle-mounted monocular camera obtains the road images collected during the driving process of the vehicle and sends them to the convolutional neural network for semantic segmentation of lane lines. The segmented binary image containing only lane lines is perspective transformed to obtain a bird's-eye view, the effective lane line pixels are screened, the effective lane line points are polynomial fitting using the least square method, and the left and right lane line polynomial fitting coefficients are output. It can effectively solve the problems of poor environmental adaptability of traditional lane line detection algorithms, weak robustness, and inaccurate detection information for curved lane lines.

Key words: Lane line detection; Autopilot; Neural networks; Semantic segmentation

摘要： 车道线等地面标志物的检测是自动驾驶车辆环境感知的重要内容，能够为车辆提供可行驶 区域的信息。文章提出一种基于语义分割结果的车道线检测拟合方法。使用车载单目相机获取车 辆行驶过程中采集的道路图像，送入卷积神经网络进行车道线语义分割。将分割得到的仅含车道 线的二值图像进行透视变换得到鸟瞰图，筛选有效车道线像素点，对有效车道线点使用最小二乘 法进行多项式拟合，输出左右车道线多项式拟合系数，能够有效解决传统车道线检测算法的环境 适应性差，鲁棒性不强，对弯道车道线检测信息不够准确等问题。

关键词: 车道线检测；自动驾驶；神经网络；语义分割