Distributed sensor registration based on random finite set representations

In the investigation of multi-sensor fusion problems, it is commonly assumed that all the parameters necessary to transform the information from the sensors to a common frame are known. Imperfect knowledge of these registration parameters induce systematic biases which would inhibit the benefits of multisensor fusion. For example, they can result in deterioration of the target localisation accuracy. We consider the problem of sensor registration in a distributed setting in which the sensor platforms propagate random finite set distributions that describe the multiobject scene using Probability Hypothesis Density (PHD) filters, and, transmit these posteriors to their neighbours. To find the respective registration error between two platforms, we propose a recursive method whereby a posterior distribution on the registration error is updated upon receiving the incoming multiobject distribution. The update is performed in a Bayesian fashion using a likelihood that captures the similarity of the multiobject scene as determined by the two sensors. In particular, we utilise the Bhattacharyya distance of multi-object distributions as a measure of similarity. We demonstrate the proposed method through a simulation example.