GPU-Accelerated Real-Time Stereo Matching

Abstract

A problem in the field of computer vision is the correspondence problem, the problem of finding pixels which correspond to each other in different images. A stereo matching algorithm is used to solve this kind of problem, and typically produces a disparity map, or a depth map. Current approaches are often too slow to be used in real-time, leading to the question of which algorithm is best for such purposes. This thesis explores which approach to stereo matching is most appropriate for real-time purposes. In addition, it is also explored what optimizations and approximations can be applied in order to improve performance. This was accomplished by implementing an Adaptive Support Weights based stereo matching algorithm in CUDA, and exploring various approximations and performance optimizations related to it. It is shown that Adaptive Support Weights is a good method for real-time use. This thesis’ most significant contribution is the performance optimizations presented, which significantly improve upon the performance of the algorithm compared to previous work.