Detecting track-loss for the Probabilistic Data Association Filter in the absence of truth data

Of significant interest in the practical application of data association algorithms to target tracking in cluttered environments is how to determine track-loss in the absence of truth data. An approach is laid out for the Probabilistic Data Association Filter (Bar-Shalom and Fortmann, 1988) where the predicted measurement innovations covariance is used to define nominal "tracking" and "track-lost" regimes, and the sample variance of the "effective" filter innovations (prediction errors) is the metric used to determine the regime in a two-class decision rule. A major advantage of this method is that confidence intervals can be placed on the probabilities of both correctly and incorrectly identifying the regimes. Theoretical approximations of the conditional mean and distribution of the sample innovations variance in the tracking and track-lost regimes are proposed for use in construction of the decision rule. Simulation results demonstrating the method are provided.