Innovations in Multiple Hypothesis Tracking

Multiple Hypothesis Tracking (MHT) is an increasingly important element of the sensor-to-shooter computational chain. However, modern MHT methods are still limited in their ability to integrate multiple sources of information to reduce ambiguity, and in their ability to sustain multiple hypotheses over long periods of time. These limitations reduce the quality of products delivered by MHT systems. We propose that these limitations arise from two basic sources, and present a plan addressing each: 1. The standard hypothesis space representation, the track tree, conflates representational issues with computational issues, limiting the duration over which multiple hypotheses can be maintained. We propose the track graph, a more expressive representation of the basic hypothesis space; 2. Existing methods for extracting highly-ranked consistent pictures from an MHT hypothesis space do not exploit all the regularities in the space. We propose adapting new methods that offer a several order of magnitude increase in the size hypothesis space that can be considered in real time. The methods developed will provide a basis for long term track management, and will support inputs from multiple modalities and fusion levels. The proposed methods will be based in recent advances in representation and inference in structured probabilistic modeling.