PCLane: Accurate Lane Localization with LiDAR and Abstract Data

Accurate Lane localization is one of the fundamental tasks for autonomous driving and trajectory planning for precise control. Deep learning models are more suitable for accurate and precise lane detection and segmentation considering their scope for generalizations especially in complex urban scenarios like junctions and intersections. This paper presents a deep learning framework for LiDAR point cloud segmentation using abstract map data. To train such deep learning-based computer vision models for lane segmentation it is very important to have large datasets of annotated data. Point cloud data annotation is, especially, very expensive and difficult due to their sparseness and complexity when compared to camera images. This paper highlights a self-supervised annotation and training strategy of deep neural network-based model using abstract High Definition OpenDRIVE maps to detect and segment lanes from the lidar point cloud data. PointNet [1] based architecture was used as the backbone of the deep neural network model and the LiDAR point clouds were collected using Velodyne LiDAR sensor of the DLR Experimental Autonomous Driving Vehicle ViewCar2.