Track Initiation for a Multiple Hypothesis Tracker with ML-PMHT

Abstract

The Multiple Hypothesis Tracker (MHT) is a pow-erful and widely used algorithm for simultaneously tracking multiple targets. As its name suggests, it reasons over multi-ple measurement-to-track assignments, or hypotheses, and at periodic intervals, outputs the optimal global hypothesis. In an ideal world, an MHT could keep track of all possible hypotheses; however, in the real world, the number of hypotheses rapidly becomes extremely large, so for computation and/or memory considerations, low-probability hypotheses must be continu-ously removed, or pruned. One reason for the large number of hypotheses is that in many MHT implementations, every received measurement is allowed to birth a new track. While this has a good chance of discovering actual new targets, it also definitely creates many false tracks that ultimately need to be pruned. We examine a new technique for track initiation - instead of simply creating a track from every new measurement, we use the Maximum Likelihood Probabilistic Multi-Hypothesis Tracker (ML-PMHT), a deterministic batch tracker to search for new tracks from recently received measurements. Only new tracks discovered by ML-PMHT will birth tracks considered by the MHT. This has the effect of reducing the high number of low-probability hypotheses that need to be reasoned over by the MHT, ultimately improving its performance.