Accession Number : ADA332211
Title : Fusion of Sensors That Interact Dynamically for Exploratory Development of Robust, Fast Object Detection and Recognition
Descriptive Note : Final rept.
Corporate Author : AMHERST SYSTEMS INC BUFFALO NY
Personal Author(s) : Bandera, Cesar ; Peng, Jing
PDF Url : ADA332211
Report Date : 01 DEC 1996
Pagination or Media Count : 73
Abstract : Foveal vision simultaneously offers a wide field of view for better detection and central high acuity for better recognition; additionally, it is highly optimized and cost effective for time critical active vision applications. However, space variant data acquisition necessitates the development of gaze control techniques for demand driven allocation of resources to improve relevant information acquisition and the overall system performance. This report describes a commercially feasible, efficient reinforcement learning approach to gaze control for foveal machine vision. The report first lays a theoretical foundation for reinforcement learning. It then introduces particular reinforcement learning algorithms in conjunction with function approximation as an efficient learning control method for visual attention. The efficacy of the method is validated on a number of moderately complex target detection and active perception problems. By contrast, the substantial body of work on gaze control for active vision has not taken advantage of the power and flexibility of machine learning methods for visual attention. Finally, the report describes several experiments designed to evaluate the relative efficiency of various reinforcement learning algorithms and techniques for input generalization using both prediction and control problems. Computational results show that reinforcement learning with neuro function approximation can be successfully used to obtain achievable gaze control performance in commercially feasible foveal machine vision products.
Descriptors : *TARGET RECOGNITION, *ARTIFICIAL INTELLIGENCE, *COMPUTER VISION, *TARGET DETECTION, ALGORITHMS, NEURAL NETS, REAL TIME, LEARNING MACHINES, DATA ACQUISITION, PATTERN RECOGNITION, VISUAL PERCEPTION, MULTISENSORS, MULTISPECTRAL.
Subject Categories : Cybernetics
Target Direction, Range and Position Finding
Distribution Statement : APPROVED FOR PUBLIC RELEASE