A Real-time Design of Stereo Vision System using Disparity Calibration

This paper presents a real-time stereo vision system to extract disparity map with good error rate from stereo images. The error rate was reduced by the disparity calibration module that can correct errors in the initial disparity map using adaptive support-weights. Performance can be improved by overlapping the initial disparity and the disparity calibration stages. Because the adaptive support-weight algorithm includes nonlinear functions, it is impossible to reuse already processed data. In this study, we designed the system to reuse data by separating rows and columns, and as a result, performance is improved so that real-time processing is possible. To optimize the delay path, the exponential function is approximated with a step function and five stage pipeline is applied to the division operation. The proposed architecture was designed using VerilogHDL and synthesized with DBhitek 110nm standard cell library. The synthesized system used 33K gates and 4.9KB memory and showed a maximum frequency of 666MHz (57fps) and error rate of 1.88% for tsukuba image.