Sub-pixel disparity estimation in continuous space

Abstract

For depth-from-stereo vision applications such as driving assistance and object-size estimation, the accuracy of disparity estimation determines the precision of depth measurement. Conventional dense methods are hard to estimate disparity within a 0.1 pixel precision. In this paper, we present a novel object-based method to achieve robust and deep sub-pixel accurate disparity estimation. In our experimental system, it can provide depth measurement with less than 5% relative error within 150 m.