随着人工智能技术的不断发展，无人驾驶技术已经成为当下社会发展的热门，车道线检测是无人驾驶技术的关键一环，但传统的基于视觉的车道线检测方法处理时间较长、过程繁琐、需要人为干预。基于深度学习的车道线检测可大大减少此类问题。本文设计了一个完成双任务的Enet网络，分别解决目标区域分割问题和不同车道区分问题，然后使用了一个可以预测转换矩阵的T-net网络来自动完成后处理时的透视变换，以实现端到端的检测。主要研究内容如下：

（1）改进基于Enet网络的车道图像数据前处理框架。该网络模型用于完成车道线的分割并为车道线像素赋予权重。该网络分别采用二值分割和语义分割的处理步骤，并将车道线检测问题看作一个实例分割问题，对语义分割和二值分割进行聚类以完成实例分割的目标。

（2）使用基于T-Net的图像后处理框架并优化其训练参数。传统的做法是把图片转换为鸟瞰图形式进行曲线拟合，在鸟瞰图模式下拟合的曲线可以通过矩阵转换到原始图片上，但是如果地平面发生变化传统的拟合方式就不再有效。为了解决这个问题，首先在曲线拟合之前进行图像视野转换，然后通过神经网络的训练输出转换系数。这种以图片作为输入，使用loss函数进行优化车道拟合问题后处理框架可以根据路面平面的变化进行自动适应，可以更好的拟合车道。

（3）为了完成复杂道路环境下的车道线检测实验，收集了10000张复杂环境下的车道线图像，并对其进行手工标注，将这10000张图像作为数据集的一部分（数据集主体为Tusimple数据集），以供网络模型的训练。

（4）与主流车道线检测网络模型deeplab v3和YOLO v3进行对比。实验验证表明，网络模型的最终训练准确率为99.8%，相对于deeplab v3和YOLO v3网络分别提升1.2%和0.8%，在测试集上的平均准确率为96.44%，相较于另外两种模型提升0.73%和0.45%。

With the continuous development of artificial intelligence technology, driverless technology has become a hot topic in the current social development. Lane line detection is a key link of driverless technology, but the traditional vision-based lane line detection method takes a long time to process, and the process is tedious, requiring human intervention. Deep learning based lane detection can greatly reduce such problems, and the detection effect is good. In this paper, a dual-task Enet network is designed to realize the lane line detection network, which solves the segmentation problem of target area and the problem of different lanes, and then a t-net network with predictive transformation matrix is used to automatically complete the perspective transformation during post-processing, so as to realize the end-to-end detection. The main research contents of this paper are as follows:

（1）Improved the lane image data pretreatment framework based on Enet network. The network model is used to segment the lane lines and assign weight to the lane pixels. In this network, binary segmentation and semantic segmentation are adopted, and the lane line detection problem is regarded as an instance segmentation problem.

（2）The image post-processing framework based on t-net is used. The traditional method is to convert the image into the form of aerial view and perform curve fitting based on it. In the mode of aerial view, the curve fitting can be converted to the original image by matrix, but the traditional fitting method is no longer effective if the ground plane changes. In order to solve this problem, the image field is converted before curve fitting, and then the conversion coefficient is trained by neural network. This post-processing frame can automatically adapt to the change of road surface and better fit the lane by taking the picture as the input and using the loss function to optimize the lane fitting.

（3）In order to complete the lane line detection experiment in complex road environment, this paper collected 10,000 lane line images in complex road environment and manually marked them. These 10,000 images were taken as part of the data set (the main data set was Tusimple data set) for the training of the network model.

（4）Compared with deeplab v3 and YOLO v3 lane line detection network model, the experimental verification shows that the final training accuracy of the network model used in this experiment is 99.8%, which is 1.2% and 0.8% higher than deeplab v3 and YOLO v3, respectively. The average accuracy of the test set is 96.44%, which is 0.73% and 0.45% higher than the other two models.