YUV luma/chroma quantization and sparse correspondence for real time video stereo matching

To realize the real time dense disparity map running at a video rate of 30 fps, the dynamic time warp algorithm (DTW) provides a robust method of stereo matching. This method requires to calculate the pixel-by-pixel similarity matrix that indicates the similarity from one pixel of the left image to a pixel of the right image in the raster profile. The size of the similarity matrix S is N² for the raster size N. Down sampling to reduce N defeats the purpose of disparity measurement because the spatial resolution is also reduced. A method to introduce sparse samples is proposed in this paper, so that the proposed method reduces N but not sacrificing the spatial resolution of stereo matching. The denoised raster waveform is coarsely quantized to produce a binary train at each of the quantization level. The run-length for the contiguous one´s is used as a feature to represent the raster waveform profile. The size of the sparse set is about N = 30 as opposed to the raster size N = 352 for the CIF size image. This idea is applied to the luma Y-image and the chroma UV-image in the YUV color space to take advantage of the concept of image segmentation by color in addition to the stereo matching normally performed in the gray scale.