Fuse-before-Track in Large Sensor Networks

While multi-sensor systems hold great potential for surveillance performance, the technical challenges are significant, and include the need for effective calibration as well as a statistically-valid characterization of environmental uncertainties and contact measurement errors. Additionally, automatic tracking and fusion processing with active sensors must contend with high false contact rates and target fading effects. Issues in multisensor surveillance and numerous design approaches are discussed in [4, 9, 24]. In [12—13], we present model-based, simulationbased, and sea-trial tracking performance results with a track-oriented, modular multi-hypothesis tracking scheme. Of particular interest is the tradeoff between centralized and multi-stage processing: we have found that, when faced with significant target fading effects and for modest false contact rates, distributed processing can outperform centralized processing. This somewhat surprising result is based on the fundamental suboptimality of all tracking algorithms that must contend with measurement origin uncertainty. This explains the seeming contradiction with results in the nonlinear filtering and distributed detection literature, in particular the well-known optimality of centralized processing schemes [9, 25]. Ultimately, for sufficiently low-SNR target scenarios, effective real-time automatic tracking is extremely challenging regardless of the choice of data processing architecture. One approach is to relax the real-time requirement, and to leverage powerful batch processing techniques [3]. However, such schemes are not easily amenable to real-time surveillance requirements, and generally assume non-maneuvering targets. An alternative approach in challenging scenarios is to consider enlarging the surveillance network, possibly through bootstrapping approaches that include sub-band processing techniques [19], whereby a sensor is effectively “replaced” with a number of slightly-degraded sensors. The latter approach (enlarging the surveillance network) implicitly assumes that an increased number of like-performing, calibrated, and registered sensors are always to be preferred, i.e., more sensors are always better than fewer. Is this true in general or are there performance limits as the number of sensors becomes large? This is the issue that we address in this paper. We start by introducing in Section 2 a simple analytical model for tracker performance. We study tracker performance as a function of local detection threshold, number of sensors, and track management criteria. The model supports the conclusion that there are performance bounds on achievable performance in large sensor networks. Can we do better if we consider a more complex, multi-stage processing architecture? In Section 3, we describe the fuse-before-track (FbT) architecture and