Task-domain vision : theory and architecture

Active vision, as a discipline, is relatively new in the short history of machine vision. It has, however, left an indelible mark in the minds of vision researchers with respect to the design of visual systems. It is the purpose of this research to explore the design of an active component in a purposive, or task-domain, framework and formulate a general theory of task-oriented design for a class of imaging devices. In particular, this theory will address three general areas that are problematic in the design of general vision modules. These are (1) Sensor Geometry, Calibration, and Representation. It is shown how application of task-domain constraints can reduce the complexity of these problems and in certain cases present closed-form solutions to optimal design and configuration problems.

In addition, a systemic view of algorithm design is presented. Using this methodology, a Bayesian network of algorithms is used to formulate possible solutions to a particular problem. Using purposive constraints, the most favorable solutions are propagated and the undesirable solutions are discarded. This architecture offers significant advantages over the traditional block-level system design in terms of accuracy and robustness.

Finally, a design philosophy of purposive vision is presented. This methodology is employed in the design of three active visual modules. The three applications are: (1) A parallel-axis stereo system for small object localization; (2) A convergent-axis stereo system with efficient dynamic reconfiguration capabilities; and (3) an active Bayesian network for synthesis of cranial structures in positron emission tomography. The versatility of this theory is evident in its range of applicability to vision problems.