SCHEDULING MULTIPLE SENSORS USING PARTICLE FILTERS IN TARGET TRACKING Amit S. Chhetri, Darryl Morrell and Antonia Papandreou-Suppappola Department of Electrical Engineering, Arizona State University, Tempe, Arizona 85287, USA ABSTRACT A critical component of a multi-sensor system is sensor sche- duling to optimize system performance under constraints (e.g. power, bandwidth, and computation). In this paper, we apply particle filter sequential Monte Carlo methods to im- plement multiple sensor scheduling for target tracking. Un- der the constraint that only one sensor can be used at each time step, we select a sequence of sensor uses to minimize the predicted mean-square error in the target state estimate; the predicted mean-square error is approximated using the particle filter in conjunction with an extended Kalman filter approximation. Using Monte Carlo simulations, we demon- strate the improved performance of our scheduling approach over the non-scheduling case. 1. INTRODUCTION Sensor technology has become important for commerce, en- vironmental science, medicine, defense, etc. The last two decades have seen an unprecedented growth in the number and variety of sensors and methods to extract information from sensor measurements. Cost effective sensor use is im- perative for effective system performance. Sensor schedul- ing, which is the allocation of sensing resources over time, is thus becoming an essential component of sensor systems. The problem of scheduling sensors to optimize the ex- pected cost function over time is a stochastic control prob- lem; in principle, solutions to this problem can be computed using dynamic programming [1, 2]. In practice, computing optimal solutions may be prohibitively expensive, and sub- optimal greedy algorithms are used instead [3–5]. Some cost functions used in scheduling algorithms include sen- sor usage cost, accuracy of sensor state estimate, cost of re- sources, (weighted) squared error and desired estimate co- variance [6]. Recently, information theoretic cost functions that measure the difference between prior and posterior dis- tributions have been applied together with particle filter se- quential Monte Carlo techniques [3, 4, 7]. In this paper, we minimize the mean squared estimate error. This work was supported by a Raytheon contract with Integrated Sens- ing and Processing DARPA program. Specifically, we schedule infrared (IR) and radar sensors to obtain measurements, and track the target based on the measurements using a particle filter. The sensors are sched- uled using the particle filter and an extended Kalman fil- ter (EKF). The scheduling algorithm predicts multiple steps ahead (where each step corresponds to a time difference ) to find the sensor sequence that minimizes the estimated co- variance at those steps. Monte Carlo simulations show that the tracking performance improves significantly with sensor scheduling. 2. PROBLEM FORMULATION We consider a target moving in a 2-dimensional Cartesian space. Let be a real-valued random vector denoting the target state at time . The state vector is defined as where and are the positions in the and directions, and are the velocities in the and directions, and denotes matrix transpose. The target evolves according to a linear system driven by white Gaussian noise as (1) Here, models the state kinematics: is the time difference between measurements, and is a Gaussian random vector with covariance . This co- variance is obtained by converting a continuous time stochas- tic target model into an equivalent discrete time model [8]: