Tracking with context

Du Toit, Carlo (2016-12-01)

Thesis (MSc)--Stellenbosch University, 2016

Thesis

ENGLISH ABSTRACT : Tracking in an unconstrained environment presents a difficult challenge. Abrupt object motion, appearance changes, non-rigid objects and occlusion are but a few of the trials faced. To overcome these challenges a tracking algorithm, often an unsupervised learning problem, should be fast and capable of on-the-fly modelling. A huge variability in the range of input observation data is to be expected. Generative tracking models are good with dealing with unsupervised learning, but not always trustworthy without good verification, which leads to drifting. In this thesis we investigate the effectiveness of integrating knowledge (context) into the tracking model and to provide verification to generative models, improving the drifting problem and trustworthiness of the model. To accomplish this we implement a model based on what we learn from other context aware implementations. Our model is context flexible, capable of integrating any existing object detector, providing the model with valuable knowledge. Experimentation shows it is capable of integrating with any target tracker, and provides valuable assistance in the form of verification. When the target undergoes aggressive appearance changes, gets fully occluded or even leave the field of view, our model is capable of tracking the target successfully until the main tracker can resume its task. Context is not only there to serve the main target tracker, but also to improve learning of the model itself. We use the model to minimise the possibility of a miss-match during training itself, providing increased certainty.

AFRIKAANSE OPSOMMING : Die spoor van ’n voorwerp in ’n onbeperkte omgewing het baie uitdagings. Skielike beweging, voorkoms verandering, nie-rigiede voorwerpe en okklusies is slegs ’n paar van die moontlike uitdagings. Om hierdie uitdagings te oorkom moet ’n sporingsalgoritme vinnig wees en oor die vermoë beskik om intyds te kan modelleer. Dikwels word dit as ’n leerprobleem sonder toesig hanteer. Groot variasie in die invoer waarnemingsdata is te verwagte. Generatiewe modelle is goed om te leer sonder toesig, maar nie altyd betroubaar sonder goeie verifikasie nie. Dit lei tot dryf. In hierdie tesis ondersoek ons die effektiwiteit om kennis (konteks) in die sporingsmodel te integreer. Verder bied dit verifikasie vir generatiewe modelle. Verifikasie verminder die kans vir dryf en verbeter die betroubaarheid van die model. Ons implementeer ’n model gebaseer op vroeëre kontekssporingsalgoritmes. Ons model is onafhanklik van die spesifieke konteks en in staat om met enige bestaande voorwerpherkenner te kan integreer, sodoende die model te verskaf met waardevolle kennis/konteks. Eksperimentele resultate toon aan dat ons model integreer met enige bestaande sporingsimplementasie en in staat is om waardevolle verifikasie vir sporingsalgoritmes te bied. Wanneer die teikenvoorwerp van voorkoms verander, volledige okklusie ondergaan of selfs die beeldraam verlaat, is ons model in staat om die teiken te volg totdat die sporingsimplementasie weer oorneem. Konteks is nie slegs daar om die sporingsimplementasie te help nie, maar ook om die leerproses van die model te verbeter. Ons gebruik die model om die moontlikheid van ’n herkenningsfout tydens die leerfase te verminder, en so dan die akkuraatheid te verhoog.

Please refer to this item in SUNScholar by using the following persistent URL: http://hdl.handle.net/10019.1/100271
This item appears in the following collections: