一种基于无人机影像的高精地图车道线检测与提取方法

高精度地图是实现自动驾驶技术必不可少的基础设施，车道线是高精度地图车道级路网的重要组成部分。以往高精度地图的车道线检测多基于车载摄像头数据，存在成像范围有限、需要透视变换和多次拼接造成的效率问题。基于无人机航拍影像，采用U-Net网络识别道路区域，过滤非道路区域噪声，通过HSL颜色变换和Sobel算子分别计算车道线颜色和边缘梯度特征，使用Otsu算法自动确定特征分割阈值获得二值化车道线特征图，通过局部最大值算法确定滑动窗口的初始位置，最后借助滑动窗口算法和多项式检测拟合车道线。实验结果表明，在保证一定检测精度的前提下，单条车道线检测长度超过了百米，道路检测效率达到25.2 m/s，对比于地面影像的检测算法具有明显的效率优势。

A high definition map lane line detection and extraction method based on UAV images

High definition map is an essential infrastructure to realize automatic driving technology, and lane line is an important part of lane level road network of high definition map. Currently, lane detection of high definition map is mostly based on the data of vehicle camera, which is low efficient due to limited imaging range and need for perspective transformation and multiple stitching. In this paper, based on UAV aerial images, U-Net network is used to identify road areas and filter noise in non-road areas. HSL color transform and Sobel operator are used to calculate lane color and edge gradient features respectively. Otsu algorithm is used to automatically determine feature segmentation threshold to obtain binary lane feature map. Local maximum algorithm is used to determine the initial position of sliding window. Finally, lane lines are fitted by sliding window algorithm and polynomial detection. The experimental results show that with certain detection accuracy, the detection length of a single lane line exceeds 100 m, and the road detection efficiency reaches 25.2 m/s. Compared with the lane line detection algorithms based on vehicle-mounted camera data, the proposed method is obviously more efficient.