Short-term ambiguity assessment to augment tracking data association information

A tracking system performs both state estimation and data association. Most trackers provide as output tracks with kinematic uncertainty information but no measure concerning data association uncertainty. In scenarios with closely spaced objects, even the most advanced trackers will sometimes produce impure tracks. Functions that process tracks, however, often implicitly assume correct data association and may produce suboptimal results if the association uncertainty is ignored. This paper describes how to augment a multiple hypothesis tracking system with association uncertainty measures. We show how the assignment probabilities are computed from ranked association hypotheses. Based on simulations we illustrate the use of these "short-term" association uncertainty measures in identifying potentially false correlations. To improve tracking performance in the presence of ambiguity we propose to use entropy of the assignment problem to adjust the length of the sliding window. We then discuss an approach to maintain "long-term" association uncertainty, based on Bayesian networks, that accumulates short-term uncertainty information provided from the multiple hypothesis tracker.