PARTICLE FILTERING METHOD FOR OBJECT TRACKING AND SENSOR

Abstract

Generally, there is no analytic solution to object tracking problems in non-linear non-Gaussian scenarios, which is a common type of problem nowadays. A particle filter is a numerical method that can be applied to any class of model regardless of linear and Gaussian assumptions as in the Kalman filter, and has the same benefits of constant memory requirement and real-time recursive estimation. In this report, a hidden Markov model is set up for state and observation evolution, and both the particle filter and the Kalman filter are developed and applied to generate tracking results. Our results show that in linear and Gaussian case, the performance of particle filtering is very close to the classic Kalman filter, which achieves the Cramer?Rao lower bound, while the particle filtering method can be applied much more extensively when linear and Gaussian assumptions are not justified in real problems. In both object tracking and other problems such as detection, sensor management is an issue, as there has to be a trade-off between performance and cost. As sensors are commonly utilized for multiple purposes, a generic performance measure based on mutual information is developed. An overall sensor cost is computed by summing up the one-time cost of installation and the life-time cost of operation. To make the information gain and the cost comparable, a bit-dollar exchange rate is defined to compute the monetary value of the information gain. By combining the monetary gain and the cost into a single objective, a sensor configuration strategy can be chosen among multiple options.