Principles and techniques for sensor data fusion

After a brief introduction, we review the background of the problem of fusion in machine vision. We then present fusion as part of the process of dynamic world modeling, and postulate a set of principles for the "fusion" of independent observations. These principles lead to techniques which permit perceptual fusion with qualitatively different forms of data, treating each source of information as constraints. For numerical information, these principles lead to specific well known tools such as various forms of Kalman filter and Mahalanobis distance. For symbolic information, these principles suggest representing objects and their relations as a conjunction of properties encoded as schema.