Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.676605
Title: Dynamic background modelling for foreground detection in surveillance video
Author: Varadarajan, Sriram
ISNI:       0000 0004 5373 0234
Awarding Body: Queen's University Belfast
Current Institution: Queen's University Belfast
Date of Award: 2015
Availability of Full Text:
Full text unavailable from EThOS.
Please contact the current institution’s library for further details.
Abstract:
Foreground detection is the primary step in the analysis of surveillance video. Robust foreground detectors are essential in intelligent surveillance systems as any higher level analysis of the scene requires good detection results. In many real-world situations, the detection process is complicated by the presence of dynamic elements in the scene background such as rippling water or moving trees that vary unpredictably over time. In this thesis, different methods are proposed for robust foreground detection in scenes containing dynamic background. First, foreground detection on moving transport platforms such as buses is investigated. This scenario is quite challenging because of the dynamically changing background in the window regions of the bus. Since this variation is influenced by the motion of the bus, a fusion of motion features using Optical Flow and a colour based background subtraction algorithm such as Mixture of Gaussians (MoG) is proposed to differentiate between the dynamic background region and the interesting foreground region. The second contribution of this thesis is a generalised modelling framework, called region based Mixture of Gaussians (RMoG) that takes into consideration neighbouring pixels while generating the model of the observed scene. This handles the spatial perturbations caused by either intrinsic uncertainties in the scene due to the background dynamics, or extrinsic disturbances such as jitter in the video due to an unsteady camera. Finally, a momentum term is introduced in online gradient based learning to address the issue of slow convergence. In the context of dynamic background modelling, faster convergence is particularly useful to handle the fast, dynamic variations in the pixel process. The asymptotic convergence of the on line gradient method with momentum type updates is proved and an expression is derived to show the O(1/kˆ2) convergence rate of the algorithm when the individual gradients are constrained by a growth condition.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.676605  DOI: Not available
Share: