Local stereo matching using motion cue and modified census in video disparity estimation

In the human visual system, proximity, similarity, and motion are fundamental attributes that group visual objects together locally. The objects grouped by these attributes are most likely to have the same depth. In previous works, proximity and similarity have been considered in the computation of image disparity maps. However, they are insufficient for video disparity estimation because motion cues are very important for accurate depth estimation near edges of moving objects. We incorporate motion flow to compute each pixel's support weight, a measure directly affecting the accuracy of disparity maps in local methods. For robustness to image noise in flat areas, we propose a modified census transform with a noise buffer. The experimental results show that the proposed method produces more accurate disparity maps than current state-of-the-art methods, both on edges and in flat areas according to subjective and objective measures.