Uxbridge uB8 3PHUnited Kingdom
INTRODUCTIONReal-time computer vision applications are characterised by the need to provide a rapid response to information derived from a high-speed continuous data stream as indicated in Figure 1. It is the requirement for such high-performance, unattainable by sequential architectures, which can only be achieved with massively parallel processing. Figure 1: Real-time computer vision scenario schematic Sensors may include some form of visual, thermal, electromagnetic or nuclear radiation energy and applications include remote sensing, surveillance and tracking in aerospace, avionics and autonomous navigation and increasingly in biomedical. In particular, real-time computer vision systems appear very attractive to any form of robotic applications.Typical detection tasks range from image reconstruction or restoration (to compensate for sensor non-uniformities, ageing and eventual failure), through filtering (to improve signal-to-noise ratios and enhance particular image features) to correlation (to discriminate relevant image information from background clutter).Typical analysis tasks range from grouping data associated with meaningful objects, through specific transforms for feature extraction and edge or contour following (to establish feature connectivity) to object quantification (for length or area measurement or centre-of-mass coordination) and listing of relevant object properties.Typical decision tasks range from pattem matching (correlation of invariant object descriptors with model descriptions) and navigation of semantic networks through hypothesis analysis to interpretation of object-related information to facilitate decision-making and response determination. The paper describes the use of the ASP (Associative String Processor) a massively parallel processing architecture to implement representative tasks of integrated computer vision applications. 2 ASP ARCHITECTURE OVERVIEW Based on the encouraging results emerging from research into parallel computing technology at Brunel University and being developed by Aspex Microsystems Ltd., ASP (Associative String Processor) modules comprise highlyversatile parallel processing building-blocks for the simple construction of fault-tolerant massively parallel processors [l-21, which is capable of TeraOPS performance (Tera = and support of continuous data input. Indeed, according to application requirements, an appropriate combination of ASP modules would be plugged into the control bus of Data Communications Network of an appropriate modular ASP system, as indicated in Figure 2.Figure 2: Modular ASP systemBased on a fully-programmable, reconfigurable and faulttolerant homogeneous parallel processing architecture, ASP modules offer considerable application flexibility, maintaining computational efficiency over a particularly wide range of applications, due to
