The Use of Automated Critics to Improve the Fusion of Marginal Sensors for ATR and IFFN Applications

Abstract : A basic goal of multi-sensor data fusion is to increase the accuracy and reliability of inferences by combining data and information from multiple sources. In particular, applications such as automatic target recognition (ATR) and identification-friend-foe-neutral (IFFN) processing seek to characterize, classify, and ultimately identify targets of interest such as aircraft, tanks, or enemy units. Ideally, the use of multi-sensor data from non-commensurate sensors (viz., sensors observing fundamentally different physical phenomena) would improve the ability to identify targets by broadening the physical baseline of observation. For example, the use of a combination of acoustic, seismic, infra-red, and radar data has the potential to improve the ability to characterize ground-based targets. In addition, the use of a broad range of physical measurements improves the ability to counter an enemy's information warfare efforts. There are several circumstances, however, in which the fusion of multi-sensor data actually produces worse results (on average) than can be achieved by an individual sensor. That is, the fused results are less accurate and less reliable than those of the best individual contributing sensor. As one example, sensor data may be incorrectly weighted, due to a lack of knowledge of the dynamic sensor performance in realistic operating conditions. Another example most germane to the present work is the case where decisions from one or more of the contributing sensors have accuracy less than 50 percent. It is well-known that decision-level fusion schemes, such as voting techniques, produce unreliable results when the accuracy of the contributing sensors is less than 50 percent. Unfortunately, such relatively poor performance is not uncommon in applications such as IFFN and ATR. This is particularly true in anticipated information warfare conditions.