A probabilistic graphical model approach to multiple object tracking

Date
2018-03
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: Probabilistic graphical models (PGMs) provide a framework for efficient probabilistic inference using graphs that correspond to factorised representations of high-dimensional probability distributions. The problem of tracking objects from noisy measurements is inherently a probabilistic one and the use of PGMs to solve this problem is therefore appropriate. In this work, we investigate how PGMs can be used for tracking an unknown and varying number of targets in challenging scenarios. While many existing algorithms provide solutions to the multiple object tracking (MOT) problem, none of the established algorithms are framed as PGMs. In order to develop a graphical model for multiple object tracking, the connections between PGM theory and the Kalman filter algorithm, which is commonly used for single object tracking, are investigated. The PGM equivalent of the Kalman filter is used as a starting point for the development of the MOT PGM. The Kalman filter PGM is first expanded to allow a known and constant number of targets to be tracked, and Bayesian model selection is then used to allow the number of targets to be inferred automatically. In order to allow the model to track targets in the presence of false detections, a clutter classification model is developed and incorporated into the developed PGM. The efficiency of the model is improved through the use of an alternative model selection method. It is also shown that the tracking accuracy can be improved through more accurate Gaussian mixture approximations of the target state distributions. The developed model is compared to a state-of-the-art method and is tested by way of a large number of simulations. We conclude that the model is capable of consistently and accurately tracking targets and that it offers advantages over some existing methods. Finally, the model output is compared to an industrial application with real radar data as input. The outputs of the two models are largely similar and the test results therefore indicate that the developed model can be used for real-world applications. In order to create a general software resource for implementing the type of PGMs designed in this work, the University of Stellenbosch PGM library, EMDW was expanded. A large portion of the software created as part of this work is therefore not limited to the multiple object tracking problem, but useful for PGM inference in general.
AFRIKAANSE OPSOMMING: Grafiese waarskynlikheidsmodelle (GWM'e) maak van gefaktoriseerde voorstellings van hoëdimensionele waarskynlikheidsverdelings gebruik en bied 'n raamwerk vir die doeltreffende berekening van randverdelings. Om akkurate vooruitskattings van teikenposisies te maak aan die hand van metings waar ruis teenwoordig is, is inherent 'n waarskynlikheidsprobleem. GWM'e kan dus gebruik word om die probleem van teikenvolging effektief op te los. In hierdie werk ondersoek ons hoe GWM'e vir die volg van 'n onbekende en wisselende aantal teikens in uitdagende scenario's gebruik kan word. Alhoewel daar baie bestaande algoritmes is wat oplossings vir die probleem van multi-teikenvolging (MTV) bied, word geen van die gevestigde algoritmes as GWM'e voorgestel nie. Ten einde die grafiese MTV-model te ontwikkel, word die verband tussen GWM-teorie en die Kalman-filter-algoritme, wat algemeen gebruik word vir enkel-teikenvolging, ondersoek. Die GWM-ekwivalent van die Kalman-filter word as uitgangspunt vir die ontwikkeling van die MTV-GWM gebruik. Die Kalman-filter-GWM word eers uitgebrei om 'n bekende en konstante aantal teikens te volg, en Bayesiese model-seleksie word dan gebruik om die aantal teikens outomaties te bepaal. 'n Vals meting-klassifikasiemodel word ook ontwikkel om die model toe te laat om teikens in die teenwoordigheid van vals metings te volg. Die doeltreffendheid van die model word verbeter by wyse van 'n alternatiewe metode vir model-seleksie. Ons wys ook dat akkuraatheid deur 'n beter benadering tot die teikenverdelingsfunksies verbeter kan word. Die ontwikkelde model word met 'n ultramoderne metode vergelyk en ook by wyse van 'n groot aantal simulasies getoets. Ons kom tot die gevolgtrekking dat die model daartoe in staat is om teikens konsekwent en met hoë akkuraatheid te volg en dat die ontwikkelde model 'n paar voordele bo sommige bestaande metodes bied. Laastens word die model-afvoer vergelyk met dié van 'n industriële toepassing met regte radar data as toevoer. Die afvoere van die twee modelle is meestal soortgelyk, en die toetsresultate dui dus daarop dat die ontwikkelde model vir werklike toepassings gebruik kan word. Ten einde 'n algemene sagteware-hulpbron te skep vir die implementering van die tipe GWM'e wat in hierdie werk ontwerp is, is die Universiteit van Stellenbosch se GWM-biblioteek, EMDW, uitgebrei. 'n Groot gedeelte van die sagteware wat in hierdie werk ontwikkel is, is dus nuttig vir die gebruik van GWM'e in die algemeen en nie beperk tot die MTV-probleem nie.
Description
Thesis (MEng)--Stellenbosch University, 2018.
Keywords
Decision making with multiple objectives, Graphical modeling (Statistics) -- Probabilities, UCTD
Citation