Probabilistic sampling method for track association

Abstract

A method for solving the track association problem updates and samples a marginal association likelihood conditioned upon existing track assignments and marginalized over possible sensor biases to assign tracks and build a full association hypothesis. The 'sampling' is repeated multiple times for a given seed track pair and for different seed track pairs to quickly generate hypotheses that approximate the solution space. The probabilistic track association includes a plurality of likely full association hypotheses and a soft association matrix that probabilistically reflects the likelihood of the track association for a pair of sensors. Efficacy can be enhanced by sensing and processing non-metric features (e.g. size, shape, color) to supplement the metric features (e.g. location, velocity). Efficiency may be enhanced by prioritizing the list of seed track pairs in order of decreasing likelihood, saving only full association hypotheses that are both unique and close to the current most likely hypothesis and terminating the search based on a staleness criteria. Probabilistic sampling may be used for such diverse applications as missile defense, autonomous vehicles and package handling.