Target tracking and sensor scheduling in sensor networks

A new dynamic filter switching algorithm to track people who randomly enter, exit, move and stop in a region of interest using a network of uniformly spaced, stationary acoustic sensors is presented. The existence of a new target is determined by jointly weighting the particles of a Track-Before-Detect particle filter and an Interacting Multiple Model particle filter that is used to track confirmed targets. The algorithm detects new targets as well as tracks targets with intermittent motion, as is shown by Monte-Carlo simulations.;In a network of biased sensors, the biases should be estimated otherwise the target would be lost. A stochastic dynamic bias estimation strategy using a Rao-Blackwellized Particle Filter is presented. The effects of bias on sensor scheduling using the Posterior Cramer Rao lower bound to estimate the predicted root mean square error are also studied. Using Monte-Carlo simulations it is shown that the bias should not be included in the Posterior Cramer Rao bound while performing sensor scheduling.;When tracking a target in a sensor network with constrained resources, one can realize significant reductions in target state estimate error by using non-myopic sensor scheduling strategies. Moreover, reductions in sensor communications and usage cost can also be realized by using non-myopic sensor scheduling. Integer non-linear programming has been successfully used to obtain myopic sensor schedules. In this report, its benefits are also extended to two non-myopic sensor scheduling scenarios: a distributed network consisting of bearing sensors which is called the Leader Node Scheduling problem and a centralized network consisting of acoustic sensors and a single fusion center which is called the Central Node Scheduling problem. The leader node problem is cast as an integer non-linear programming problem with the objective of minimizing the total sensor usage and communications cost over the entire planning horizon subject to tracking error constraints per time step. For the central node problem, an integer non-linear programming problem is formulated to minimize the total predicted tracking error over a planning horizon subject to sensor usage and start-up cost constraints Monte Carlo simulations are used to show that non-myopic sensor scheduling strategies provide improved performance for these scenarios compared to myopic sensor scheduling.