Adaptive Agent Based System for Knowledge Discovery in Visual Information

Abstract

One of the main limitations of most current machine vision systems is a lack of flexibility to consider the wide variety of information provided by visual data. The research proposed here aims to improve this situation by the development of an adaptive visual system able to selectively combine information from different visual algorithms. The problem is cast as a knowledge discovery problem, where the two main steps are detection and characterization of relevant patterns. The algorithms will be able to perceive different attributes of the visual space such as color, depth, motion or specific shapes. The intended system should be able to adaptively select and combine the information provided by the algorithms according to the quality of the information given by each of them. The system proposed is based on an intelligent agent paradigm. Each visual module will be implemented as an agent that will be able to adapt its behavior according to the relevant task and environment constraints. The adaptation will be provided by a local self-evaluation function on each agent. Cooperation among the agents will be given by a probabilistic scheme that will integrate the evidential information provided by them. The proposed system aims to achieve two highly desirable attributes of an engineering system: robustness and efficiency. By combining the outputs of multiple vision modules the assumptions and constrains of each module will be factored out to result in a more robust system overall. Efficiency will be still kept through the on-line selection and specialization of the algorithms according to the relevant structures and conditions present at each time in the visual scene. The advantages of the approach proposed here will be demonstrated in two frequent problems faced by a mobile robot: dynamic target tracking and obstacle detection.