SCS Undergraduate Thesis Topics

Jeremy Maitin-Shepard Carlos Guestrin Multiple-Target Tracking Based on a Fully-General Data Association Model Using a Fourier-Domain Representation

Multiple-target tracking (inferring the position of each target over time, where each target has a unique identity) based on measurements that provide only imprecise information about the position or identity of the target that generated the measurement is a difficult problem due to the inherent combinatorial complexity that arises from the ambiguity regarding the association of measurements to targets; tackling this complexity depends on the use of approximation algorithms. Although existing sampling-based methods allow highly general observation and data association models, in general there is no reason to believe the probability mass can reasonably be concentrated on a small number of samples, and consequently accuracy guarantees for such methods depend on maintaining a number of samples that grows exponentially with the number of targets. Recent work by Kondor et al. and Huang et al. demonstrated the applicability of group-theoretic methods, specifically band-limited Fourier-domain representation of distributions over groups of permutations, to multiple-target tracking problems, but their methods depended on a restricted data association model based on the concept of tracks. Unlike sampling-based methods, this Fourier-domain representation can represent diffuse distributions, which are believed to likely occur in these tracking problems. This work extends the approach of Huang et al. to support a more general data association model while also modeling imprecision and uncertainty in the position information from measurements.

Close this window