Dense stereo matching in restricted disparity space

Abstract

We propose a new domain-specific area-based stereo method that is applicable to various tasks including safe navigation and pedestrian detection. Incorporating a spatial restriction in the height direction of road scenes into the recursive computation of correlation measures, the proposed method efficiently provides dense disparity measures. A pair of cameras just has to be weakly calibrated due to the non-metric formulation. We also implement our stereo method on a newly developed onboard LSI, which has a dedicated architecture for practical automotive use and realizes more than BOM disparity estimations per second. Furthermore, we apply our method to a side collision warning system, which detects approaching objects such as pedestrians and bicycles. Through experiments we demonstrate the efficiency of our stereo method in road environments.