Application of particle filters in a hierarchical data fusion system

In recent years, the particle filter has become commonly accepted as the preferred tool for single target tracking in highly non-linear and non-Gaussian environments. This paper investigates the issues that arise when particle filters are integrated into a hierarchical data fusion system, in which the sensor-level tracking is performed using particle filters, but central-level track fusion is performed using a Gaussian model. The context of the investigation is multistatic sonar tracking using a field of bistatic receiver nodes (sensors). Tracking performance of the hierarchical data fusion system with particle filter sensor level tracking is compared with the equivalent system using Kalman based filters for sensor level tracking. It is found that, while the particle filter possesses measurably better performance at the sensor level, much of this performance is lost if the sensor level tracks are fused using a Gaussian model.