Unified stereovision for ground, road, and obstacle detection

Abstract

This paper presents a method for road detection and obstacle detection entirely based on stereovision. The ground plane is estimated online by least square fitting of disparity data. This operation allows deleting road features for obstacle detection, estimating directly camera roll and pitch, and deriving some clues on road-surface image regions. A model-based algorithm employing only disparity information is demonstrated to be able to segment the whole road surface without knowledge of infrastructures and features like lane markings. This helps navigation in suburban and country-road environments, and recovery from critical failure of lane-markings trackers.