Machine Vision for multitarget tracking and automatic recognition must support heterogeneous tasks in uncontrolled environments. Conventional uniform acuity machine vision imposes commercially unfeasible processing requirements for all but the simplest of applicatins. Multiresolution multisensor active vision features imaging sensors and processing with different acuities coupled with context sensitive gaze control, analogous to the foreal paradigm prevalent throughout vertebrate vision. Such multiresolution vision operates more efficiently in dynamiic scenarios than single sensor uniform acuity vision because resolution is treates as a dynamically allocatavle resource, simultaneously achiveing wider field-of-view, greater localized resolution, and faster me rates with significantly reduced processing requirement. The proposed program will develop commercially feasible multisensor multiresolution image fusion (MMIF) technology, in particular (1) a hierarchical parallel data structure and multiprocessor for MMIF, (2) parallel fusion algorithms for MMIF, (3) parallel algorithms for the subsequent processing of fused information (filtering, detection, recognition, and tracking), and (4) algorithms for gaze control and multisensor resource allocatin. The effort will exploit established hierarchical foveal machine vision techniques, and a multiresolution active vision facility at Amherst Systems. Evaluation criteria for MMIF techniques will include systolic speed, communications overhead the of information redundancy, computational complexity, and implementability with commercially available components.
Keywords: Sensor Fusion, Active Vision, Tracking, Recognition, Multiresolution, Intelligent Vehicle
