Multi-Stage Multiple-Hypothesis Tracking

A broad overview of approaches to data fusion is provided in [1]. The most powerful current approach to real-time, scan-based data fusion is multi-hypothesis tracking (MHT), which was first introduced in the late 1970s [11] and made feasible in the mid-1980s with the track-oriented approach [9]. A number of enhancements to the basic approach have appeared over the years [1]. If contact measurement information is available at the tracker output, one can think of a multi-target tracker as a filter of sorts that discards spurious contacts and associates the remaining ones through track labeling. As such, tracking is a modular operator which, when applied to contact-level data, takes as input singleton (i.e. single-measurement) tracks. More generally, a mix of track-level and contact-level feeds may be provided to the tracker. Upstream track labels are preserved in downstream processing, except in cases where discrepancies are detected in downstream tracking. This tracker modularity allows for arbitrarily complex multi-stage data fusion architectures. This philosophy, combined with the necessary software modularity, is the basis for the multi-stage MHT approach that we consider in this paper. We find that in some applications multi-stage MHT processing outperforms single-stage MHT processing. In this paper, we introduce two multi-stage MHT architectures and compare these to single-stage, trackwhile-fuse processing. The first multi-stage architecture, track-break-fuse, is computationally efficient without sacrificing the tracking performance of track-while-fuse. The second architecture, track-before-fuse, provides further computational efficiency at the cost of some tracking performance. The track-while-fuse approach is intractable when the application requires deep hypothesis trees; conversely, both of the multi-stage MHT approaches that we introduce here identify a small set of relevant association hypotheses, enabling deep hypothesis trees. The paper is organized as follows. In Section 2, we provide a short introduction to standard (track-whilefuse) track-oriented MHT, following closely on the formalism introduced in [9]. The multi-stage MHT architectures of interest, track-break-fuse and track-beforefuse, are introduced in Section 3. In Section 4 we study track-break-fuse for a challenging, slowly-crossing targets problem. In Section 5 we study track-before-fuse for multi-sensor surveillance with complementary, multiscale sensors. Concluding remarks are in Section 6. Early results on the multi-stage processing introduced here are in [6] (track-break-fuse) and [3] (trackbefore-fuse). A related MHT approach to track-beforefuse is discussed in [4], which introduces group-tracking logic to enable deep hypothesis trees. Additionally, within the MHT framework, some techniques to hypothesis management do exist, including K-best assignment or hypothesis-clustering approaches [7, 10]. However,