Implementation of Real-Time Post-Processing for High-Quality Stereo Vision

We propose a novel post-processing algorithm and its very-large-scale integration architecture that simultaneously uses the passive and active stereo vision information to improve the reliability of the three-dimensional disparity in a hybrid stereo vision system. The proposed architecture consists of four steps - left-right consistency checking, semi-2D hole filling, a tiny adaptive variance checking, and a 2D weighted median filter. The experimental results show that the error rate of the proposed algorithm (5.77%) is less than that of a raw disparity (10.12%) for a real-world camera image having a $1,280{\times}720$ resolution and maximum disparity of 256. Moreover, for the famous Middlebury stereo image sets, the proposed algorithm's error rate (8.30%) is also less than that of the raw disparity (13.7%). The proposed architecture is implemented on a single commercial field-programmable gate array using only 13.01% of slice resources, which achieves a rate of 60 fps for $1,280{\times}720$ stereo images with a disparity range of 256.