Learning-based approaches to nonlinear multisensor fusion in target tracking

We consider a network of sensors wherein the state estimates are sent from the sensors to the fusion center to generate a global state estimate. Conventionally, state estimates are linearly combined to produce the global (fused) state estimate, but the use of nonlinear fusers in multisensor fusion for target tracking has been fairly unexplored. In this work, we compare several learning-based nonlinear fusers (namely, Artificial Neural Networks, Support Vector Regression, the Nadaraya-Watson estimator, and the Nearest Neighbor Projective Fuser) in system-level simulations under two different scenarios: one where the target is a ballistic target in the coast phase, and in the other the target is performing a maneuver. Results demonstrate that several of these learning-based fusers are able to outperform linear fusion. In addition, we propose a modification to one of the nonlinear fusers to incorporate additional information that we have about the input data, which appears to result in better generalization capabilities for the Artificial Neural Network Fuser and superior performance.