Real-Time Stereo Matching Using Transputer Arrays for Close-Range Applications

The Alvey MMI-137 project is concerned with the development of real-time 2.5D vision systems using reconfigurable transputer arrays. Over the last two years, a Close-Range Vision Cell (CRVC) based on active stereo using CClR frame-transfer CCD cameras and a texture light projection system has been desiprled and installed at THORN EM1 Central Research Laboratories. This CRVC has been developed for research into direct entry of surface shape into CAD systems for free form objects at a standoff range of 500-850mm with a rms accuracy of O.lmm in plan and range. CRVC images have been used to evaluate a variety of stereo matchers for their amuracy, reliability and sampling density. The CRVC is currently integrated into a special-purpose real-time image processing module (the THORN EM1 Modular Image Processing System, TEMIPS) which allows image data to be transferred to a Sun-3TM workstation in 0.5s and in the future to a PARSYSTM SN 1000 at CCl R video frame-rates. This PARSYSTM transputer array is predicted to be able to estimate depth values using an adaptive least-squares patch-fitting algorithm at rates of 50 patches/second/ transputer. Hence for a surface model to be produced every 5th pixel using 32 T800 transputers should take 6.5 seconds. 1. INTRODUCTION Coordinate Measuring Machines (CMM) have been in use for many years for the determination of surface shape of a wide variety of industrial objects from machined components to automobile clay models and airframes (Higashimoto, 1983). Typical tolerances required for these applications are on the order of 0. lmm. There are a number of significant drawbacks for such CMMs for the precise determination of the shape of surfaces of free form objects. These include their high cost, both in capital and in manpower, their potential to impact and damage the surface of an object, their very slow speed and their consequent inability to measure dynamically-changing structures. A number of alternative non-contact techniques have been investigated. These have tended to concentrate on systems employing active vision (see recent comprehensive review by Besl, 1988). However, most of these types of systems suffer from problems associated with trading-off accuracy against speed, specularity, non-extensibility to other problem domains and eye safety. An alternative approach is to use active stereo or multiple-image matching to extract a surface model. To provide sufficient features for image matching, unstructured light illumination may be used. Systems incorporating this approach include the PRISM system developed at MIT (Nishihara, 1984) …