Neural Unscented Kalman Filter for Submarine Active Target Tracking

The motive of the work is to investigate the use of active sonar measurements by submarines to track a ship's trajectory underwater. An unscented Kalman filter (UKF) with neural network (NN) is applied to estimate the motion parameters of target. The model is approximated by the nonlinear state space NN, and its weights are then trained online through the UKF. As the process is stochastic, Monte-Carlo simulation is used and the outcomes are compared with that of UKF algorithm. As expected, the results with UKF are better than that of UKF with NN. The purpose of NN is to be used when the process uncertainties cannot be modeled exactly. NN need not be used when the analytical solution is possible, and solution is tractable.