high performance computing on graphics processing units: hgpu.org

Stereo matching techniques are used to extract 3D information from 2D stereo pair of images. It can be classified into feature based approach, window (area) based approach, and optimization based approach. Feature based approach generally generates sparse disparity map with high accuracy and low execution time. Window based approach produces dense disparity map with low accuracy and low execution time. Optimization based approach generates dense disparity map with high accuracy and high execution time. Since the ultimate goal of stereo matching is to obtain dense disparity map with high accuracy and low execution time, we choose to select optimization based approach and implement it in parallel framework to overcome execution speed deficiency. There are several optimization methods including dynamic programming, energy minimization, and graph algorithms. We choose to use dynamic programming based on disparity space image (DSI) since it is most appropriate for parallel framework. In this paper, we propose a new parallel algorithm and framework for DSI construction, dynamic programming (DP), and disparity computation using Compute Unified Device Architecture (CUDA). We tested the method on several stereo pairs and found that the method shows remarkable speedup while preserving the quality at a reasonable level.