Strategy for autonomous control of correlation-based trackers

A fully implemented vision based tracker must be able to identify an object in a variety of poses or distortions and estimate its position in a scene. After locking onto the object, continuous steady state tracking is required as the object gradually changes its position and orientation. The tracker must be able to recognize loss of track and take action to recover the object during break-lock and other transient conditions. While correlators can optimally recognize an object under theoretically ideal conditions, autonomous tracking of objects would require the development of a high level controller, or intelligent supervisor to deal will an uncontrolled visual environment. The supervisor would need to configure and control the analysis of the input environment, the detection procedure of the target, the trajectory estimation for maintaining lock on the target, and the camera orientation. In this paper we review the tracking problem. We then describe supervisor design based on configuring a suite of specific operations. Some of the operations include wide-area scan and prescreening operations using vector inner product composite filters; accurate detection and location with distortion-invariant composite filters to isolate a large data base of training views; filter banks distortion and which is distinguishable from characters associated with other objects. We discuss both hardware and algorithm considerations for the tracking problem. A general conclusion is that specific composite filter designs can be combined and configured to perform the tracking process.