Local Stereo Matching Algorithm Based on Pixel Difference Adjustment, Minimum Spanning Tree and Weighted Median Filter

Abstract

This paper proposed a new algorithm framework for stereo vision system. A four stage taxonomy is applied to obtain disparity map or depth map. The initial stage started with matching cost computation based on pixel difference adjustment strategies. In this stage, the proposed algorithm uses combination of Absolute Difference (AD) and Gradient Matching (GM) that focus on simple computation and radiometric distortion reduction. Next, the second stage continues with cost aggregation that apply image segmentation technique. Minimum spanning tree method is applied for object boundary preservation. During the third stage; disparity optimization takes a local approach by using Winner-Takes-All (WTA) strategy. This strategy normalizes the disparity values of each pixel of the image. Finally, the disparity refinement stage, noise smothering and reduction is applied on the final disparity with the aid of weighted median (WM) filter.