Detection of Changes in Tracked Targets

In this paper we address the detection of maneuvers in tracked objects. Some previous work by a subset of the authors has compared maneuver detection via input estimation to that using exact modeling of innovation covariances, the latter with special attention to a maneuver's manifestation via its cross-correlation. Here we explore different approaches based on the optimal (asymptotically and under fairly strict conditions) Page “quickest” change detection procedure. We explore several variants, especially with the idea to robustness and to calcula-bility of the test's reliability in terms of the average number of samples between false alarms. To this end, we especially focus on CUSUMs that involve single-bit quantization. We also investigate the use of the normalized innovations-squared (NIS). A side-issue arises in that study: if one's “plant” process truly uses covariance Q, but one's Kalman filter is correct in every way except that it uses kQ, are the innovations still white? The answer, surprisingly, is: no. This means that a conservatively-designed tracker that deliberately inflates the process noise - although it may indeed work well - cannot be relied upon to have the expected independent innovations sequence.