A FAST PARALLEL PROCESSOR FOR ROBOTIC VISION

Abstract

Computer vision and digital image processing tasks require a large number of computations because of the vast amount of data to be processed and the difficulty of the tasks themselves. For example, a 256 by 256 gray level image, with 8 bits per pixel, is 64K bytes of image data. To process that amount of data using a serial computer would take seconds, if not minutes. Given that the real-time image data acquisition rate of 33.33 ms per frame, the computation speed is clearly inadequate. Moreover in many computer applications. images have to be captured and processed at intervals of less than one second as in the case of inspecting parts on a conveyor belt. This thesis describes a fast multi-processor based vision system called the Parallel Image Processor, or PIP. The PIP architecture which is of the master-slave configuration where a MC68000-based single board computer controls the multi-processor system consisting of eight TMS32010 Digital Signal Processors (DSPs). The program instructions downloaded into the common (RAM) memory of the TMS32010s are broadcast to the eight processors simultaneously in the Single Instruction Multiple Data (SIMD) mode. The image to be processed is then loaded into an image buffer which is segmented into eight segments, each to be processed by one TMS32010. The processed results are stored in another image buffer. Different processes can thus be implemented by downloading the respective programs into the program memory before the processing is initiated. To evaluate the PIP, some commonly used image processing and pattern recognition algorithms were tested. The advantages and limitations of the PIP and the ways to overcome the latter are discussed.