Implementation and Performance of a Complex Vision System on a Systolic Array Machine

Abstract

Complex vision systems are usually quite slow, requiring tens of seconds or minutes of computer time for each image. As the complexity and experimental nature of the system increases, the speed is especially low, since all components of the system must be optimized if the system is to show good performance. The FIDO system, a stereo vision system for controlling a robot vehicle, has existed for a number of years and has been implemented on a number of different computers. These computers have ranged from a DEC K L l O to the current implementation on the Warp machine, a 100 Million Floating-point Operations Per Seconds (MFLOPS) systolic array machine. FIDO has shown an enormous range in speed; its ancestor took 15 minutes per step, while the Warp implementation takes less than 5 seconds per step. Moreover, while early versions of FIDO moved in slow, start-and-stop steps, FIDO now runs continuously at 100 mdsecond. We review the history of the FIDO system, discuss its implementation on different computers, and concentrate on its current Warp implementation.