Poisson Multi-Bernoulli Mixtures for Multiple Object Tracking
Doctoral thesis, 2022
Multi-object tracking (MOT) refers to the process of estimating object trajectories of interest based on sequences of noisy sensor measurements obtained from multiple sources. Nowadays, MOT has found applications in numerous areas, including, e.g., air traffic control, maritime navigation, remote sensing, intelligent video surveillance, and more recently environmental perception, which is a key enabling technology in automated vehicles. This thesis studies Poisson multi-Bernoulli mixture (PMBM) conjugate priors for MOT.
Finite Set Statistics provides an elegant Bayesian formulation of MOT based on random finite sets (RFSs), and a significant trend in RFSs-based MOT is the development of conjugate distributions in Bayesian probability theory, such as the PMBM distributions. Multi-object conjugate priors are of great interest as they provide families of distributions that are suitable to work with when seeking accurate approximations to the true posterior distributions.
Many RFS-based MOT approaches are only concerned with multi-object filtering without attempting to estimate object trajectories. An appealing approach to building trajectories is by computing the multi-object densities on sets of trajectories. This leads to the development of many multi-object filters based on sets of trajectories, e.g., the trajectory PMBM filters.
In this thesis, [Paper A] and [Paper B] consider the problem of point object tracking where an object generates at most one measurement per time scan. In [Paper A], a multi-scan implementation of trajectory PMBM filters via dual decomposition is presented. In [Paper B], a multi-trajectory particle smoother using backward simulation is presented for computing the multi-object posterior for sets of trajectories using a sequence of multi-object filtering densities and a multi-object dynamic model. [Paper C] and [Paper D] consider the problem of extended object tracking where an object may generate multiple measurements per time scan. In [Paper C], an extended object Poisson multi-Bernoulli (PMB) filter is presented, where the PMBM posterior density after the update step is approximated as a PMB. In [Paper D], a trajectory PMB filter for extended object tracking using belief propagation is presented, where the efficient PMB approximation is enabled by leveraging the PMBM conjugacy and the factor graph formulation.
Finite Set Statistics provides an elegant Bayesian formulation of MOT based on random finite sets (RFSs), and a significant trend in RFSs-based MOT is the development of conjugate distributions in Bayesian probability theory, such as the PMBM distributions. Multi-object conjugate priors are of great interest as they provide families of distributions that are suitable to work with when seeking accurate approximations to the true posterior distributions.
Many RFS-based MOT approaches are only concerned with multi-object filtering without attempting to estimate object trajectories. An appealing approach to building trajectories is by computing the multi-object densities on sets of trajectories. This leads to the development of many multi-object filters based on sets of trajectories, e.g., the trajectory PMBM filters.
In this thesis, [Paper A] and [Paper B] consider the problem of point object tracking where an object generates at most one measurement per time scan. In [Paper A], a multi-scan implementation of trajectory PMBM filters via dual decomposition is presented. In [Paper B], a multi-trajectory particle smoother using backward simulation is presented for computing the multi-object posterior for sets of trajectories using a sequence of multi-object filtering densities and a multi-object dynamic model. [Paper C] and [Paper D] consider the problem of extended object tracking where an object may generate multiple measurements per time scan. In [Paper C], an extended object Poisson multi-Bernoulli (PMB) filter is presented, where the PMBM posterior density after the update step is approximated as a PMB. In [Paper D], a trajectory PMB filter for extended object tracking using belief propagation is presented, where the efficient PMB approximation is enabled by leveraging the PMBM conjugacy and the factor graph formulation.
Bayesian filtering
random finite sets
multi- object tracking
trajectory estimation
extended object
conjugate prior
Author
Yuxuan Xia
Chalmers, Electrical Engineering, Signal Processing and Biomedical Engineering
Multiscan implementation of the trajectory poisson multi-Bernoulli mixture filter
Journal of Advances in Information Fusion,;Vol. 14(2019)p. 213-235
Journal article
Multiple Object Trajectory Estimation Using Backward Simulation
IEEE Transactions on Signal Processing,;Vol. 70(2022)p. 3249-3263
Journal article
Poisson Multi-Bernoulli Approximations for Multiple Extended Object Filtering
IEEE Transactions on Aerospace and Electronic Systems,;Vol. 58(2022)p. 890-906
Journal article
Trajectory PMB Filters for Extended Object Tracking Using Belief Propagation
IEEE Transactions on Aerospace and Electronic Systems,;Vol. In Press(2023)
Journal article
With the advances in object tracking techniques as well as sensing and computing technologies, there has been an explosion in the use of object tracking technology in numerous research venues as well as application areas, including air traffic control, maritime navigation, remote sensing, biomedical research, intelligent video surveillance, and more recently environmental perception, which is a key enabling technology in automated vehicles.
Object tracking refers to the problem of using noisy sensor measurements to determine the location, trajectory, and characteristics of objects of interest. This is a challenging problem as the tracker needs to deal with complex sources of uncertainty, such as measurement origin uncertainty, false alarms, misdetections, appearance of new objects, and disappearance of existing objects. Due to these challenges, we cannot be certain about the properties of interest and that we therefore would like to use a probability distribution to capture relevant information about the objects that we are interested in. For the choice of probability distribution, we would like the distribution to not only encapsulate the information contained in the measurements as much as possible but also can be efficiently computed, especially when applied in real-time applications.
In this thesis, we investigate Poisson multi-Bernoulli mixture distributions for Bayesian object tracking, with a focus on multi-object tracking based on sets of trajectories. In a nutshell, we aim at finding Poisson multi-Bernoulli mixture distributions and their approximations that are suitable to work with for different tracking applications. In particular, we are interested in developing efficient tracking algorithms to compute the probability distribution that can describe the object trajectories during a given time interval in a principled and straightforward manner. A promising application of such algorithms is to collect estimates that serve as ground truth, which is of utmost importance for the development and verification of both perception and control modules in automated vehicles.
Object tracking refers to the problem of using noisy sensor measurements to determine the location, trajectory, and characteristics of objects of interest. This is a challenging problem as the tracker needs to deal with complex sources of uncertainty, such as measurement origin uncertainty, false alarms, misdetections, appearance of new objects, and disappearance of existing objects. Due to these challenges, we cannot be certain about the properties of interest and that we therefore would like to use a probability distribution to capture relevant information about the objects that we are interested in. For the choice of probability distribution, we would like the distribution to not only encapsulate the information contained in the measurements as much as possible but also can be efficiently computed, especially when applied in real-time applications.
In this thesis, we investigate Poisson multi-Bernoulli mixture distributions for Bayesian object tracking, with a focus on multi-object tracking based on sets of trajectories. In a nutshell, we aim at finding Poisson multi-Bernoulli mixture distributions and their approximations that are suitable to work with for different tracking applications. In particular, we are interested in developing efficient tracking algorithms to compute the probability distribution that can describe the object trajectories during a given time interval in a principled and straightforward manner. A promising application of such algorithms is to collect estimates that serve as ground truth, which is of utmost importance for the development and verification of both perception and control modules in automated vehicles.
Deep multi-object tracking for ground truth trajectory estimation
VINNOVA (2017-05521), 2018-07-01 -- 2022-06-30.
Areas of Advance
Information and Communication Technology
Subject Categories
Control Engineering
Signal Processing
ISBN
978-91-7905-690-2
Doktorsavhandlingar vid Chalmers tekniska högskola. Ny serie: 5156
Publisher
Chalmers