A scalable method of tracking targets with dependent distributions

We develop a more accurate multiple target tracker by considering the dependence between target states caused by unknown measurement-to-target associations. Since maintaining a joint probability distribution over all the target states is infeasible except when the number of targets is small, we approximate by constructing disjoint trees over the set of targets and storing the joint distributions of adjacent targets in the same tree. These trees are adaptively chosen at each time step to minimize the information lost through the approximation. An existing data association algorithm is generalized to efficiently and exactly compute the measurement association probabilities for this case. This approach is shown to reduce the incidence of track swapping and improve tracking accuracy when identification measurements are occasionally available.