Lane detection is a useful technique in modern autonomous vehicles systems, which assists vehicle to accurately localize itself according to detected road lines. Traditional methods leveraged edge detection and Hough transform based algorithms to plot lines along the detected lane. Noticeably, they did not take the informative feature road gradient into account. In addition, most previous deep learning-based algorithms consider lane detection as pixel-wise lane segmentation, where only fixed number of lanes can be detected. In order to solve these limitations, we propose a quality guided lane detection algorithm by modeling the sophisticated traffic context, where variable number of lanes can be satisfactorily handled. Specifically, we first leverage chessboard images for camera calibration to calculate correspondence between real world and image coordinate system. Subsequently, we capture image regions of interest that only contains lane information by leveraging the prior knowledge and image quality scores. Afterwards, we design an end-to-end two-stage CNN architecture for lane detection, where binary lane mask is utilized for lane matching. Comprehensive experiments have demonstrated that our proposed method can cope with variable number of lanes effectively.

车道检测在现代自动驾驶汽车系统中是一项有用的技术，可帮助车辆根据检测到的道路线准确定位自身。传统方法利用边缘检测和基于霍夫变换的算法来沿检测到的车道绘制线。值得注意的是，他们没有考虑信息量丰富的道路坡度。此外，大多数以前的基于深度学习的算法都将车道检测视为逐像素车道分割，其中只能检测到固定数量的车道。为了解决这些限制，我们通过对复杂的交通环境进行建模，提出了一种质量指导的车道检测算法，其中可以令人满意地处理可变数量的车道。特别，我们首先利用棋盘图像进行相机校准，以计算现实世界与图像坐标系之间的对应关系。随后，我们利用先验知识和图像质量得分来捕获仅包含车道信息的感兴趣图像区域。之后，我们设计了端到端的两阶段CNN架构用于车道检测，其中二进制车道掩码用于车道匹配。综合实验表明，我们提出的方法可以有效地应对可变数量的车道。