A Recursive Distributed Linear Bias Model

Abstract

Associating sets of observations is fundamental to multi-sensor data fusion. If sensor bias is relevant to the system, algorithms that model the error structure accurately are needed for optimal performance. The global nearest pattern (GNP) framework assumes a linear relative bias exists in the state space of two sets of tracks. This model effectively forces linearization of sensor registration errors about the mean target location, treats the effect of any bias as identical across all tracks, and is capable only of estimating relative bias, not absolute bias per sensor. Furthermore, as we will show, the GNP assignment costs cannot consume the cross-correlated tracks that are the result of optimal fusion given this bias model. Here, a new linear bias model that uses per-track Jacobians of track error against an arbitrary set of bias parameters is described. This work presents a new, much more flexible distributed linear bias (DLB) model for track fusion and shows that this bias model supports recursive minimum mean squared error (MMSE) optimal fusion for an arbitrary number of sensors. In addition, this model can support track fusion for non-linear bias models where the bias is dispersed across separated targets.