A New Miniaturized Embedded Stereo-Vision System (MESVS-I)

We have developed a fully integrated, miniaturized embedded stereo vision system (MESVS-I) which fits into a tiny package of 5 times 5 cm and consumes very low power (700 mA @ 3.3 V). The system consists of two small profile CMOS cameras, and a power efficient, dual-core embedded media processor, running at 600 MHz per core. The stereo-matching engine performs sub-sampling, rectification, pre-processing using rank transform, correlation-based matching using three levels of recursion, L/R consistency check and post-processing. We have proposed a novel and efficient post-processing algorithm that removes outliers due to low-texture regions and depth discontinuities by combining the contributions from the variance map of the rectified image, disparity map, and the variance map of the disparity map. To further enhance the performance of the system, we have implemented a two staged pipelined-processing scheme that takes advantage of the dual-core architecture of the embedded processor, thereby achieving a processing speed of around 10 fps for disparity maps.