Off-road obstacle detection with robust parametric modeling of the ground stereo geometry

Abstract

Autonomous navigation in off-road environments presents many challenges in contrast to the more conventional, urban environments. Unstructured surroundings, non-uniform visual cues and lack of prior knowledge about the scene complicate the design of even basic functionalities such as obstacle detection. This paper presents a stereo vision based ground geometry modeling and obstacle detection algorithm that is well suited for cross-country navigation. Our mathematical analysis shows that the "ν-disparity" method is inadequate for accurate terrain modeling under vehicle pose variations; to compensate for this shortcoming, we propose a novel extension to the original algorithm. As the preliminary step of this extension, lateral gradient of the ground disparity is estimated using histogram analysis. This information is subsequently propagated to a modified "ν-disparity" algorithm that models the longitudinal terrain disparity variation. The effectiveness of this two-phase ground modeling technique for obstacle detection is demonstrated with empirical results.