Disparity estimation method and processor architecture for real-time image processor

Abstract

Stereo vision is inferring 3D structure of a scene from two or more images taken from different viewpoints. In stereo vision, there are two primary problems, correspondence search and reconstruction. Correspondence search is also called stereo matching, and this is finding the corresponding pixels from two or more images. Correspondence search is an important problem, because correspondence search is computationally expensive and consumes much time among the whole processing of finding disparity map. The issues of correspondence search are how to reduce the complexity of computation and how to get disparity map which is similar to ground truth disparity map as possible. In this thesis, a new disparity estimation method is proposed and also the architecture of image processor for real-time disparity estimation is proposed. The proposed method is focusing on processing textureless regions. The proposed method uses edge values of images for classifying textureless regions. If pixel numbers that have larger edge value than edge threshold value in support window are less than 5% of window size, the support window region is classified into textureless region. If a region is textureless, window shape is changed from square window to cross window. The change of window shape gives the effect of global-like window and can check the textureless regions. The proposed processor architecture is well-designed for this proposed disparity estimation method. To reduce the latency, double buffer is used for storing next frame and internal memories are exploited to store absolute difference results of previous pixels. To improve the throughput, slice-by-slice processing is adopted too.