École Nationale Supérieure de
Techniques Avancées

Abstract

Road detection is a key issue to ensure a robust navigation for autonomous vehicles. Main sensor in such perception systems, the Velodyne LiDAR provides high resolution and omnidirectional information of the car's surroundings. Therefore, this project adresses the processing of 3D point clouds to produce fast and reliable segmentation and mapping of the navigable space. Named RoadSeg, the proposed solution uses a specially designed light-weight convolutional neural network to predict pixel-wise semantic segmentation from spherical LiDAR images. Thanks to an evidential framework, a fusion scheme aggregates labelled scans and produces dense local maps as 2D occupancy grids intended for navigation. Experiments conducted on an on purpose collected dataset as well as field tests show great performances while ensuring real-time processing.