Block-matching disparity map estimation using controlled search range

In this paper, disparity map estimation algorithm targeting for advanced driver assistance systems (ADAS) is described. A disparity map estimation algorithm for ADAS needs to be both accurate and to be able to work in real time. These algorithms can be classified into two categories: global estimation and local estimation. While global estimation algorithms tend to output a less noisy disparity map, most of them cannot operate in real time. In this paper, disparities are estimated locally by block matching and information acquired by calculating a disparity value for a pixel is transferred to its neighborhood by restricting the search range. Thus, in this way, while great efficiency is achieved during the block matching process, the effect of outliers is also minimized. The proposed algorithm suggests a new disparity calculations order which makes it possible for a pixel to get support from three different neighborhood directions (i.e. left, right and bottom). Using Horizontal support (from both directions) along with vertical support further reduces the noise compared to using support from only one direction (bottom). Thus, an accurate disparity map can be calculated while low computational complexity is maintained. Experimental results are quantified using KITTI data-sets and the percentage of erroneous pixels in non-occluded areas (Out-Noc) is calculated as 8.86%, while run-time is estimated as 0.84 seconds on an i7-870 CPU using single thread C implementation.