Abstract
High-speed acquisition of reliable three-dimensional (3-D) information plays an essential role in real-world intelligent systems such as intelligent robots and highly-safe intelligent vehicles. Stereo vision is a well-known method for 3-D instrumentation. A major issue of stereo vision is to establish reliable correspondence between images. This paper presents a reliable stereo-matching algorithm based on SAD (Sum of Absolute Differences) computation. Reliable correspondence can be established by selecting a desirable window size of the SAD computation based on the uniqueness of a minimum of the SAD graph. Moreover, to increase reliability in matching, edge information is efficiently used for selecting the window size. A pixel-serial and window-parallel architecture is also presented to achieve 100% utilization of processing elements. The performance of the VLSI processor is evaluated to be 1500 times higher than that of a general-purpose processor.
