Stereo Matching Performance Analysis of Cost Functions on the Graphic Processing Unit (GPU) for Pervasive Computing

摘要

Stereo imaging is a powerful technique for determining the distance to objects using a pairs of camera spaced apart. The extremely high computational requirements of stereo vision limit application to non realtime applications where high computing power is available. To overcome the limitation, we utilized the general strategy for parallelization of dense cost functions on Compute Unified Device Architecture (CUDA) with Graphic Processing Unit (GPU), especially for pervasive environment. The challenges of mapping a sequential stereo matching algorithm to a massively parallel thread environment are considered. Compared to the CPU counterpart, the processing speed of the stereo matching algorithm based on CUDA programming can be improved by about from 107-369 times.