Dense Stereo Disparity Maps - Real-time Video Implementation by the Sparse Feature Sampling
Kunio Takaya · 2011
To realize the real time dense stereo disparity map (DDM) running at a video rate of 30 fps, the dynamic time warp algorithm (DTW) is time wise a bottle neck despite its robustness for stereo matching. The DTW method requires to calculate a large similarity matrix S of the size N 2 for the raster size N, ifpixel-by-pixel matching is attempted. The computation time to calculate S is significant for real-time systems and embedded hand-held devices. Two methods, coarse quantization method and hump detection method, to reduce N by sparse feature sampling are proposed in this paper. Both proposed methods reduce N much below the raster size, and create a set of sparse samples without sacrificing the spatial resolution for stereo matching. The size of the sparse set is typically N =30and N =15for each respective method, compared with the raster size N = 320. Thus, the calculation time of DDM is dramatically improved by more than 100 times. By using the proposed methods, a real-time system was realized on the Windows platform. 1