Stereo based Obstacle Detection with Uncertainty in Rough Terrain

Autonomous robot vehicles that operate in off-road terrain should avoid obstacle hazards. In this paper we present a stereo vision based method that is able to cluster reconstructed terrain points into obstacles by evaluating their relative angles and distances. In our approach, constraints are enforced on these geometric properties by a set of pixel threshold values. Because these values are all computed during an initialisation step, only simple pixel threshold operations remain to be performed during the real-time obstacle detection. An advantage of this novel approach is that the distance uncertainties can be incorporated into the thresholds. Detected obstacle points are clustered into objects on the basis of their pixel connectivity. Objects with insufficient pixels, elevation and slope are rejected. Remaining non-obstacle pixels are regarded as ground surface points. They are used to update the orientation of the stereo camera relative to the ground surface. This prevents orientation errors during stereo reconstruction and the subsequent obstacle detection steps. Our results show the drawbacks of ignoring the uncertainties in the stereo distance estimates for obstacle detection. It leads to over-segmentation and increases the number of falsely detected obstacles. Because our method incorporates these uncertainties, it can detect more of the obstacle surface pixels at larger distances. This leads to significantly less false obstacle detections.