Post-Processing Algorithms for Real-time Active Stereo Vision

This paper presents post-processing algorithms which work on a commercial FPGA for high quality stereo vision results. Specially, in this paper we focus the description of algorithms and their performance evaluation results instead of explaining hardware architecture. The proposed algorithms consists of five sub-blocks in cascade manner, which are consistency check, hole filling, variance check, weighted median filter, and joint bilateral filter. The experimental results show that error rates are 6.02% (6.57% with no using) for real world camera input(1280x720 intensity, maximum disparity 256), and 8.30% (13.7% with no using) for Middlebury stereo image set when we use the proposed post-processing.