MULTISENSOR ATR AND IDENTIFICATION OF FRIEND OR FOE USING MLANS

Automatic target recognition and a related problem of non-cooperative identification friend or foe often require fusing of multiple sensor i nformation into a unified battlefield picture. State of the art approa ches to this problem attempt to solve it in steps: first targets are d etected, then target tracks are estimated, these are used to correlate , or associate information between multiple sensors, the associated in formation is combined into the unified picture and targets are identif ied. A drawback of dividing the problem into smaller steps is that onl y partial information is utilized at every step. For example, detectio n of a target in clutter may not be possible on a single frame of a si ngle sensor, and the target motion information may have to be utilized requiring track estimation, so several steps have to be performed con currently. This paper describes such a concurrent solution of the mult iple aspects of this problem based on the MLANS neural network that ut ilizes internal world models. The internal models in MLANS encode a la rge number of neural weights in terms of relatively few model paramete rs so MLANS learning occurs with significantly fewer examples than req uired with unstructured neural networks. We describe the model-based n eural network paradigm, MLANS, present results on MLANS concurrently p erforming detection and tracking, adaptively estimating background pro perties, and learning and classifying similar U.S. and foreign militar y vehicles. The MLANS performance is compared to that of the multiple hypothesis tracker, the classical statistical quadratic classifier, an d the nearest neighbor classifier, demonstrating significant performan ce improvement.