面向多元场景结合GLNet的车道线检测算法

各种复杂环境下路面车道线的高效精确检测是自动驾驶领域中车道偏离预警系统的关键性技术之一。由于车辆实际运行环境的复杂性和路面车道线的多样性，现有方法在车道线检测的准确性和鲁棒性上仍需不断增强。提出一种面向多元场景结合GLNet的车道线检测算法。首先采用改进Gamma校正对待检测路面图像预处理，消减光照不均匀、夜晚等环境干扰，增强车道线纹理。然后为增强数据集的多样性，在LaneNet网络的基础上引入对抗生成网络DCGAN，构建GLNet网络模型。该模型采用编码-解码的网络结构提取车道线特征（车道蒙板和像素点），通过DBSCAN聚类算法将不同车道线划分为不同的实体，使用H-Net网络学习的视觉转换矩阵优化并拟合输出车道线。最后基于已训练好的GLNet权重模型对车道线进行精确提取，并在Tusimple数据集和自制数据集上测试验证。试验结果表明：该方法的检测准确率可达97.4%，相较于基于LaneNet网络的车道线检测算法明显提高；DCGAN网络的加入丰富了数据集类型，并提高了该模型的表征及分类能力；DBSCAN聚类算法的平均聚类时间约为0.016 s，相较于Meanshift算法运行效率更高。所提出的方法考虑了不规范、环境复杂等多种道路类型的车道线检测任务，提升了对复杂噪声与多元场景的处理能力，在车辆辅助驾驶领域具有较好的鲁棒性和适用性。

Lane Detection Algorithm Based on GLNet for Multiple Scenes

Accurate and efficient lane detection is one of the key technologies of lane departure warning systems in the field of automatic driving. Owing to the complexity of the vehicle running environment and diversity of road lane lines, the accuracy and robustness of existing lane detection methods still need to be enhanced. A lane detection algorithm based on GLNet for multiple scenes is proposed. First, improved Gamma correction was used to preprocess the road image, so as to eliminate the interference of uneven illumination, night, and other environmental variations. This could enhance the lane texture. Second, to enhance the diversity of the dataset, DCGAN was added to build a GLNet model based on LaneNet. In this model, the lane features (lane mask and pixels) were extracted using the network structure of coding and decoding. Different lanes were divided into different entities by using DBSCAN clustering. A visual transformation matrix based on H-Net learning was used to optimize and fit the lanes. Finally, the lane was accurately extracted based on the trained GLNet weight model, and the study was tested on the Tusimple and self-made datasets. The experimental results show that the average accuracy of this method can reach 97.4%, which is significantly improved compared with the LaneNet-based algorithm. Meanwhile, the addition of the DCGAN network enriches the dataset and improves the representation and classification ability of the model. The average time of the DBSCAN clustering algorithm is approximately 0.016 s, which is more efficient than Meanshift. The study considers the lane detection task of nonstandard and various roads with complex environments, improves the processing ability for complex noise and multiple scenes, and has good robustness and applicability in the field of vehicle-assisted driving.