Multi-object tracking using random finite sets

Abstract

The aim of multi-object tracking is the estimation of the number of objects and their individual states using a sequence of measurements. While state of the art algorithms use object individual single-object trackers, the multi-object Bayes filter models the multi-object state as well as the measurement process using random finite sets which naturally represent the uncertainty in the number of objects as well as in the state of the objects. Hence, a realization of a random finite set valued random variable represents the complete environment and facilitates the incorporation of object interactions. In the update of the multi-object Bayes filter, the multi-object likelihood function averages over all possible track to measurement associations which avoids error-prone association decisions.

In this thesis, the first real-time capable sequential Monte Carlo implementation of the multi-object Bayes filter and its application to real-world sensor data are presented. The proposed implementation of the multi-object Bayes filter is based on an approximation of the multi-object likelihood function which significantly reduces the computational complexity. Further, several methods to incorporate object interactions in the prediction step of the multi-object Bayes filter are proposed. Additionally, a novel multi-object tracking algorithm, the labeled multi-Bernoulli filter, is proposed in this thesis. The approximation of the multi-object posterior density using labeled multi-Bernoulli random finite sets results in an accurate and real-time capable tracking algorithm. The labeled multi-Bernoulli filter facilitates an implementation using Gaussian mixtures and is capable to track a large number of objects. The sequential Monte Carlo implementation of the multi-object Bayes filter and the labeled multi-Bernoulli filter are evaluated using simulated data as well as real-world sensor data. - e-ISBN 978-3-941543-13-3