A fundamental task that arises in processing signals from multiple sensors and sensor arrays is that of multi-target detection, identification, and tracking in cluttered environments. Examples of such applications (discussed in Thrust area V) include the ARL battlefield acoustic sensor arrays testbed, the diagnostic monitoring of helicopter gearboxes using multiple acoustic sensors, and processing of radar returns from a variety of conventional and non-conventional Navy sensor systems.
Usually in these applications, multiple sensors provide complementary measurements to increase accuracy and resolve ambiguities. However, to perform their tasks reliably, these systems must overcome two basic obstacles: (1) Sounds from different sources must first be segregated and associated uniquely with specific targets; (2) The cues from each source must then be processed to extract appropriate tracking information. The auditory and other biological sensory systems perform the same tasks regularly in organizing their sensory environment by grouping sound sources based on binaural and monaural cues, and fusing acoustic, visual, and other cues at higher levels to form coherent views and resolve potential conflicts before final decisions are made.
There is a host of auditory and other biological mechanisms that enable complex organisms to perform these sensory fusion tasks rapidly and accurately [.Bregman 1978, Cooke editor 1992, Slaney 1995.]. Our objective is to develop mathematical abstractions and signal processing algorithms of these processes that will permit their application in various engineering tasks. Some of these model formulations will depend on psychoacoustical experiments that we describe in detail in Thrust area III(B,C). To focus our efforts, we shall examine in detail the specific example problem of simultaneous multi-target tracking with multi-sensory data in cluttered environments. Multi-target, multi-sensor target-tracking incorporates the dynamic on-line definition, selection, and evaluation of tracks of various targets, that is, the path of targets in 3D space. This problem has become increasingly significant to the Navy and Marines recently, due to the increased emphasis on the operation of the small group of Marines (e.g., the Sea Dragon concept) or a single ship platform using its own sensors as well as other sensors and databases of target identifiers.
In this multi-target tracking problem, the first objective is to develop algorithms by which sensors can provide estimates of localization, e.g. target range and bearing, and other sound attributes such as timbre and pitch. These measurements are derived using auditory pre-processing algorithms ( Thrust area I), combined with grouping principles [.cooke editor 1992, Bregman 1978.] ( Thrust area III(B)) and clustering algorithms such as the independent component and principal component analysis methods ( Thrust area II). The second objective is to develop dynamic signal processing algorithms that act on these derived sensory measurements to provide detailed information on the paths of targets in 3D space. Because of the motion of each target and its mechanical interaction with the environment (aerodynamics of airplanes and missiles, or hydrodynamics of ships), its state vector (position, velocity, and acceleration) will satisfy a differential equation, whereas its vector of digital identifiers can be described by a Hidden Markov Model (see Thrust area II(B)). Since we do not know these dynamics precisely, it is necessary to model these dynamics stochastically as one of the approaches pursued in the proposed effort. Examples of this fruitful approach can be found in our formulations of optimal nonlinear filtering for multi-target tracking problems [.Bar-Shalom 1992, Marcus 1981, Baras 1988.]. We propose to provide computationally feasible signal processing algorithms based on these stochastic partial differential equations (see Thrust area II(A,B)).