Visual targeting is an inherently multiresolution active vision process requiring wide field-of-view for detection, high frame rates for tracking, and high spatial resolution for identification/false alarm rejection judiciously allocated at regions of interest detectable with lower resolution. Conventional machine vision has uniform spatial resolution throughout the field-of-view, and thus generates predominantly irrelevant data that saturates signal and processing bandwidth, costrains frame size and rate, and limits targeting performance. Amherst Systems proposes the demonstration of hierarchical foveal machine vision (HFMV) for targeting applications, which exploits their active multiresolution nature to yield higher performance at lower cost and payload than uniform resolution solutions with the same processing throughput. Foveal active vision features imagers and processing with graded acuity coupled with context-sensitive gave control, analogous to that prevalent throughout vertebrate vision; its practical feasibility, scalability, exploitation of existing (predominantly rectilinear) technology, and extendibility to the infrared spectrum distinguishes HFMV from previous approaches to foveal machine vision. The proposed effort will develop a multitarget cueing system that integrates algorithms for multiresolution video processing (real-time moving target detection and tracking) and visual attention (real-time sensor configuration management), a PC hosted multiprocessor configured for multiresolution video processing, and a JPL visual domain reconfigurable multiresolution active pixel sensor (APS) imager whose prototype was selected and characterized in Phase I. Regions of interest will be tracked electronically without mechanical articulation through real-time sensor reconfiguration. The Phase II baseline effort will (1) mature the APS imager prototype by increasing the number of receptive fields and adding on-chip control logic to streamline the sensor interface and increase its I/O throughput, (2) integrate the sensor into a portable HFMV targeting system, and (3) demonstrate this biomimetic technology by having the system search for, detect, interrogate, and display with high resolution multiple ground vehicles in real-world outdoor settings. A Phase II Option effort will integrate the targeting system with one of several candidate customer test facilities
Keywords: targeting active vision active pixel sensors detection foveal tracking cueing multiresolution
