Lane detection is the most basic part of automatic driving, driver assistance and violation-detection systems. A lane detection method based on semantic discrimination is proposed to address the problems of diversity in illumination, occlusion and lack of prior knowledge about lane markings. Two symmetric ridge operators are designed to improve the precision of candidate-pixel extraction. The double-constrained random sample consensus method (DC-RANSAC) introduces colour and geometric constraints into the lane fitting to reduce the number of iterations and improve the accuracy of the hypothetical model. Classifiers are added to further validate the hypothetical model and identify its underlying semantics. The proposed method was evaluated using two different data sets with various scenarios, including unclear lane markings, dense traffic, occlusion of vehicles, complex shadows, road surface markings, poor lighting conditions, and unknown number of lane markings. The detailed evaluations show that the detection rate of the proposed method is comparable with that of existing state-of-the-art lane detection methods, whereas the precision rate is much higher. Moreover, the experiments prove the reliability of the proposed algorithm in lane marking semantic recognition.

中文翻译：

---

基于语义区分的鲁棒多车道检测方法

车道检测是自动驾驶，驾驶员辅助和违规检测系统的最基本部分。提出了一种基于语义歧视的车道检测方法，以解决照明的多样性，遮挡和缺乏关于车道标记的先验知识的问题。设计了两个对称的脊形运算符，以提高候选像素提取的精度。双约束随机样本共识方法（DC-RANSAC）将颜色和几何约束引入车道拟合，以减少迭代次数并提高假设模型的准确性。添加了分类器以进一步验证假设模型并识别其基本语义。所提出的方法是使用两种不同的数据集进行评估的，这些数据集具有各种场景，包括不清楚的车道标记，密集的交通，车辆阻塞，复杂的阴影，路面标记，不良的照明条件以及未知数量的车道标记。详细的评估表明，该方法的检测率与现有的最新车道检测方法相当，而准确率则更高。实验证明了该算法在车道标记语义识别中的可靠性。