Position Paper: Panel on Issues and Challenges in Performance Assessment of Multitarget Tracking Algorithms with Applications to Real-World Problems

摘要

Traditionally, performance analysis of multitarget data fusion algorithms has been based on measures of effectiveness (MOEs) that quantify 'local' aspects of performance: probability of target detection, average localization miss distance, probability of correct ID. etc. Because multitarget tracking algorithms are highly nonlinear, improving performance with respect to one local MOE can result in decreasing performance with respect to another-and thus conflicting MOEs. Because of such problems, attempts have been made to construct over-all ('global') MOEs via, for example, heuristic weighted averages of local MOEs. Unfortunately, this approach suffers from its own difficulties. The resulting single MOEs have tended to be arbitrary, non-intuitive, difficult to interpret, and sometimes not even invariant with respect to a change of units of measurement.This position paper argues that comprehensive multitarget tracking performance estimation must be based on the fundamental multitarget statistics of a given tracking problem, as described by the theory of finite-set statistics.5 Specifically, truly comprehensive MOE should be denned in terms of the multitarget Bayes posterior probability distribution that characterizes the problem. I briefly survey the most promising techniques: multi-target Kullback-Leibler MOEs and multitarget Czsisar in formation-theoretic MOEs. I also survey intermediately comprehensive MOEs: the multitarget miss distances.