SGM-based dense disparity estimation using adaptive Census transform

Generating an accurate and dense disparity image is one of the important requirements for many applications such as 3D video and stereo vision-based advanced driver assistance systems (ADAS). Depth estimation is the process of obtaining a depth map based on two or more reference images. Recently, several techniques that use semi-global optimization for estimating depth maps have been suggested. Although robustness against illumination changes is a vital factor in applications like ADAS, semi-global matching (SGM) based on mutual information achieves limited performance under illumination changes. In this paper, a modified SGM algorithm is proposed which is based on adaptive window patterns of census transform. The goal of the proposed method is to improve the quality of the estimated depth map while reducing the processing time and making it applicable for depth-based pedestrian detection techniques. To enhance the quality of the estimated depth map, a spatial Gaussian weighted averaging filter along with a color-aware filter is implemented. To demonstrate the efficiency of the proposed method, the Middlebury stereo dataset and the KITTI vision benchmark have been used on the experiments. Experimental results show that the proposed method reduces the percentage of bad pixels by 0.7-1.2% for the test sequences compared to the original SGM algorithm with reduced processing time.