Naslov
Multiple Moving Objects Tracking Based On Random Finite Sets And Lie Groups

Autori
Ćesić, Josip

Vrsta, podvrsta i kategorija rada
Ocjenski radovi, doktorska disertacija

Fakultet
Fakultet elektrotehnike i računarstva

Mjesto
Zagreb

Datum
29.09

Godina
2017

Stranica
140

Mentor
Petrović, Ivan

Ključne riječi
multiple moving objects tracking, Lie groups, directional statistics, concentrated Gaussian distribution, extended Kalman filter, extended information filter, random finite sets, probability hypothesis density, joint integrated probabilistic data association

Sažetak
Autonomous navigation of an agent strongly relies on the capability of tracking multiple moving objects using various on-board sensing technologies. In the thesis we first consider a type of application arising when multiple objects are tracked using a microphone array as a single on-board sensor system. Both objects and measurements state space in this application arise as directional value represented either as a vector belonging to a unit sphere or equivalently as an angle. The thesis presents a method for multiple moving objects tracking on the unit sphere based on the von Mises distribution defined directly on this space of interest, and probability hypothesis density filter based on random finite sets. The state of objects in the agent’s surrounding are typically determined with their position and orientation which evolve on a non-Euclidean geometry. The orientation of such object can be described using a special orthogonal group, while full pose, including translation vector and orientation information, can be given with a special Euclidean group employing either their ò or ç dimensional counterparts. The thesis further proposes several methods for estimating motion evolving on the special Euclidean group based on the extended Kalman filter on Lie groups, and accounting for the statistics of concentrated Gaussian distribution. It also describes approaches for performing full body human motion estimation using marker position measurements or inertial measurement units, accounting for the full kinematic chain of the body. As an alternative to the extended Kalman filter on Lie groups, the thesis proposes the estimation method relying on an information form for states evolving on matrix Lie groups. A trivial example of suitable application is when the number of measurements is larger than the size of the state space, while other examples include any filter constructed such that the information form can be exploited in terms of computational complexity. As an extension of the multiple moving objects tracking algorithm limited exclusively to the space of a unit circle, the thesis proposes two methods suitable for applications when states evolve on matrix Lie groups. The first one relies on joint integrated probabilistic data association filter modified such that it can operate with variables on matrix Lie groups, while the second one employs the probability hypothesis density filter on matrix Lie groups. In the thesis we propose an approach to reduction of mixture of concentrated Gaussian distributions, which is an essential part of the probability hypothesis density filter.

Izvorni jezik
Engleski

Znanstvena područja
Elektrotehnika, Računarstvo, Temeljne tehničke znanosti

POVEZANOST RADA