A vision machine must support a wide range of complex operations on massive amounts of time-varying image data, and execute all of the operations at high speeds. The goal of this research is to design, implement, and evaluate dynamic, cooperative, parallel vision algorithms which execute simultaneously on two tightly-coupled multiprocessors: a linearly-connected, 8-processor Aspex PIPE, and a 10-processor, shared-memory Sequent Balance. This two-level organization corresponds to the two kinds of processing required in computer vision. In the first, low-level vision algorithms will be developed for the PIPE, detecting primitive "tokens" in a video-rate sequence of images. The second, high-level vision, requires more symbolic and geometric processing (e.g., in model-based object recognition), and will be developed for the Sequent. Research activities will include the development of methods for coordinating multiple levels of image analysis on the two machines, dynamic load- balancing techniques for high-level vision on the shared-memory multiprocessor, and parallel algorithms for multiresolution image analysis and model-based recognition. //
