I’ve been working on information fusion for target inference since my Ph.D. study. My work on performance evaluation of distributed tracking algorithms, in collaboration with the researchers from the Air Force Office of Scientific Research, has resulted in a superior distributed tracking system for track-to-track fusion. This was verified on a benchmark problem supplied by Wright-Patterson Air Force Lab. I also worked on the acquisition of low observable tracks especially the missile tracks at the exoatmospheric phase. My theoretical study on multiframe assignment with possibly unresolved measurements leads to a better way of tracking spawning targets, which is very important for early discrimination between warheads and other decoys of long range missiles. This work is supported by the US Missile Defense Agency (MDA) and the US Office of Naval Research (ONR) as part of MDA's Hercules Project.

I am interested in advanced fusion techniques that can handle data from disparate and degraded sensors and/or local trackers and associate local tracks from the same target on demand. I’ve been working on track association and fusion with additional features/attributes/classification outputs for the SIAP Project. Below are a few critical issues that have to be addressed.

(a) Practical Data Fusion Architecture

In view of the nature of the data available for the planned network centric architecture, which will consist mostly of track estimates, we will investigate the practical aspects of track-to-track fusion. Since the track estimation errors are crosscorrelated across sensors, the fusion algorithm will need, in order to be optimal, the track estimates, their covariance matrices as well as the crosscovariance matrices between the various track estimation errors. While the optimal track-to-track fusion is known to be suboptimal in comparison with the centralized estimator, its degree of suboptimality is relatively small. Results quantifying this will be presented.

(b) Crosscovariance Between Local Tracks

It is well known that, due to the common process noise, track estimation errors for the same target, based on data from different sensors, are correlated even though the sensor measurement errors are independent. Ignoring this phenomenon would make the fused estimates optimistic, i.e., their calculated covariances would be significantly smaller than their actual errors. Thus it is important to have not only the covariance matrices of the track estimates but also their crosscovariances. Since the exact calculation of the crosscovariances is believed not to be practical, an approach that allows a “table look-up” for the crosscorrelations will be developed. This should be also applicable to sensors that have different sampling rates.

(c) Data Fusion with Legacy Track Sources

For legacy track sources which use alpha-beta filters that do not calculate covariances, an approximate covariance matrix can be derived from knowledge of the originating track source, the track location and its track-quality index. Specifically, using the sensor measurement noise statistics, its sampling interval, the target motion uncertainty (its process noise statistics), the relative sensor-target geometry and the number of recent updates, then one can “replicate” its Kalman filter Riccati equation to obtain a covariance matrix that can be used in the fusion. While this approach is, obviously, subject to unavoidable inaccuracy, it is still believed to be superior to discarding such estimates, as it has been proposed in certain quarters (e.g., in SIAP).

(d) Data Fusion with Arbitrarily Delayed Measurements

The track-to-track fusion approach has the advantage that it can fuse track data regardless of whether it has latency. Track estimates that have latency have to be predicted to the current time and then they will be associated and fused with other current (“real-time”) track estimates.

------------------------------------------------------------------------------------------------------------------------------

People working in Data Fusion area having homepages (excluding ISL@UNO)

Yaakov Bar-Shalom at UConn, James Llinas at SUNYSB, K C Chang at GMU, A S Willsky at MIT, Moshe Kam at Drexel