A convolver-based real-time stereo machine (SAZAN)

For 3D reconstruction, polynocular stereo based on multiple image fusion is a promising method. We developed a convolver-based nine-eye stereo machine called SAZAN. It performs real-time acquisition of dense depth map at 20 MDPS (Million Depth-pixels Per Second). The reduction of matching ambiguities, which is the most crucial part in stereo matching, is effectively performed by filtering operations of 2D convolver LSI. Several new ideas and capabilities including a nonlinear data reduction of LoG outputs, an efficient geometric calibration and a subpixel disparity are also implemented in hardware. Considering the hardware size and the various factors that have an influence on the final processing quality, the computational performance is compared with existing stereo systems including the CMU stereo machine.