Acceleration of Stereo-Matching on Multi-core CPU and GPU

Abstract

This paper presents an accelerated version of a dense stereo-correspondence algorithm for two different parallelism enabled architectures, multi-core CPU and GPU. The algorithm is part of the vision system developed for a binocular robot-head in the context of the CloPeMa research project. This research project focuses on the conception of a new clothes folding robot with real-time and high resolution requirements for the vision system. The performance analysis shows that the parallelised stereo-matching algorithm has been significantly accelerated, maintaining 12× and 176× speed-up respectively for multi-core CPU and GPU, compared with SISD (Single Instruction, Single Data) single-thread CPU. To analyse the origin of the speed-up and gain deeper understanding about the choice of the optimal hardware, the algorithm was broken into key sub-tasks and the performance was tested for four different hardware architectures.