Parallel Implementation in a GPU of the Calculation of Disparity Maps for Computer Vision

Abstract

The high computational cost in computing disparity maps has relegated their use in real-time computer vision tasks. This work shows the design of a parallel algorithm for the generation of disparity maps which decreases the redundant operations during the stereo matching process. The proposed algorithm also uses a median filter to decrease the noise present in the generated disparity map, achieving an acceleration of 1135x on a GPU. In addition to the implementation of the parallel algorithm, GPU resources were used to improve memory access time and code optimization.