An equivalence-class approach to multiple-hypothesis tracking

Abstract

This paper introduces an equivalence-class approach to multi-target tracking. The approach seeks to address a fundamental limitation in multiple-hypothesis tracking: its selection (albeit with some delay and after reasoning over multiple hypotheses) of a unique global hypothesis. For some problems, the resulting tracking solution does a poor job with respect to metrics of interest. We seek instead to identify a class of similar hypotheses that have a larger aggregate likelihood than the maximum likelihood solution and, more importantly, whose members provide an improved tracking solution. Correspondingly, we introduce the Equivalence-Class MHT (ECMHT) and show its performance benefits in two-target tracking scenarios with a network of synchronous sensors. 12

Track Coalescence and Repulsion: MHT, JPDA, and BP

“…In a crossing target scenario the corresponding estimated tracks have the tendency to "bounce" more often than they cross. A method for mitigating the detrimental effects of track repulsion has been presented in [34].…”

“…Methodological adaptations tailored to alleviate these detrimental effects have been introduced in [32]- [35]. Track coalescence can be avoided by using a special hypothesis pruning strategy [32], [33] and track repulsion by selecting a global hypothesis that is not the single MAP solution, but an element of a MAP equivalence class [34]. While these adaptations can reduce the effects of track coalescence and repulsion, they are unsuitable for large-scale tracking scenarios with an unknown and time-varying number of targets.…”