A Comparison of JPDA and belief propagation for data association in SSA

Abstract

The process of initial orbit determination, or catalogue maintenance, using a set of unlabelled observations requires a method of choosing which observation was due to which object. Realities of imperfect sensors mean that the association must be made in the presence of missed detections, false alarms and previously undetected objects. Data association is not only essential to processing observations, it can also be one of the most signi cant computational bottlenecks. The constrained admissible region multiple hypothesis lter (CAR-MHF) is an algorithm for initial orbit determination using short-arc, optical (angles only), observations of space objects. CAR-MHF uses joint probabilistic data association (JPDA), a well-established approach to multi-target data association. A recent development in the target tracking literature is the use of graphical models to formulate data association problems. Using an approximate inference algorithm, belief propagation (BP), on the graphical model results in an algorithm that is both computationally e cient and accurate. This paper compares association performance on a set of deep-space objects with CAR-MHF using JPDA and BP. The results of the analysis show that by using the BP algorithm there are signi cant gains in computational load, with negligible loss in accuracy in the calculation of association probabilities.