Bayesian Sensor Fusion Methods for Dynamic Object Tracking—A Comparative Study

Abstract

In this paper we study the problem of Bayesian sensor fusion for dynamic object tracking. The prospects of utilizing measurements from several sensors to infer about a system state are manyfold and they range from increased estimate accuracy to more reliable and robust estimates. Sensor measurements may be combined, or fused, at a variety of levels; from the raw data level to a state vector level, or at the decision level. In this paper we mainly focus on the Bayesian fusion at the likelihood and state vector level. We analyze two groups of data fusion methods: centralized independent likelihood fusion, where the sensors report only its measurement to the fusion center, and hierarchical fusion, where each sensor runs its own local estimate which is then communicated to the fusion center along with the corresponding uncertainty. We compare the prospects of utilizing both approaches, and present explicit solutions in the forms of extended information filter, unscented information filter, and particle filter. Furthermore, we also propose a solution for fusion of arbitrary filters and test it on a hierarchical fusion example of two of the aforementioned filters. Hence, the main contributions of this paper are systematic comparative study of Bayesian fusion methods, and a method for hierarchical fusion of arbitrary filters. The fusion methods are tested on a synthetic data generated by multiple Monte Carlo runs for tracking of a dynamic object with several sensors of different accuracies by analyzing the quadratic Rényi entropy and root-mean-square error.