A generalised convolver for computer vision

This paper describes the video convolution processor (VCP) which meets the stringent requirements of a computer vision algorithm for depth estimation from stereo image pairs. Use of the VCP allows the acceleration of this process to near real-time rates. In particular, the VCP is designed to perform a high precision image transformation to force image pairs to satisfy, an epipolar constraint (image rectification) and the high level of mathematical integrity required for this operation is not provided by commercial devices. Additionally, the generality and scalability of the VCP allows it to be used in a wide range of computer vision and image processing applications. A prototype device has been fabricated using a 2LM, 1 /spl mu/m, CMOS process and the device operates at a frequency of 30 MHz. Under this condition the VCP can ´rectify´, decimate, and filter a 256/spl times/256 image at 25 frames.s/sup -1/ forming a very versatile ´front-end´ to computer vision problems.