Comparative Study of Local SAD and Dynamic Programming for Stereo Processing Using Dedicated Hardware

The processing results of two stereo accelerators implemented in reconfigurable hardware are presented. The first system implements a local method to find correspondences, the sum of absolute differences, while the second uses a global approach based on dynamic programming. The basic design principles of the two systems are presented and the systems are tested using a multitude of reference test benches. The resulting disparity maps are compared in terms of rms error and percentage of bad matches, using both textured and textureless image areas. A stereo head is developed and used with the accelerators, testing their ability in a real-world experiment of map reconstruction in real-time. It is shown that the DP-based accelerator produces the best results in almost all cases and has advantages over traditional hardware implementations based on local SAD correlation.