Use this URL to cite or link to this record in EThOS:
Title: Data association for visual tracking with particle filters
Author: Jin, Yonggang
ISNI:       0000 0001 3590 7217
Awarding Body: University of Surrey
Current Institution: University of Surrey
Date of Award: 2008
Availability of Full Text:
Access from EThOS:
Access from Institution:
This thesis addresses the problem of tracking one or more objects in monocular video sequences for visual surveillance. Two probabilistic tracking frameworks are proposed. The first is a multi-cue-based tracking framework to fuse high-level object detection cues with low-level color and edge cues using particle filters, which is based on object-level data association. The second is a multi-object tracking framework to deal with tracking multiple objects with occlusions, which is based on feature-level data association. In the proposed multi-cue-based tracking framework using particle filters, an adaptive ICONDENSATION (AICONDENSATION) is presented to exploit object detection cues for adaptive importance sampling where the number of importance samples and the number of prior samples are adaptively changed. An adaptive detection fusion ICONDENSATION (AFCONDENSATION) is then proposed to directly fuse high- level object detection cues with low-level color and edge cues. Results on sequences with both simulated detections and real detections show the improved performance of the proposed AICONDENSATION and AFCONDENSATION in comparison with the ICONDENSATION. In the proposed multi-object tracking framework, a novel edge-based appearance model is first presented, where an object is modelled by a mixture of a non-parametric contour model and a non-parametric edge model using the kernel density estimation. Visual tracking using mixture models is then formulated as a Bayesian incomplete data problem where measurements in an image are associated with a generative model. The generative model is a mixture of mixture models including object models and a clutter model. Unobservable associations of measurements to densities in the generative model are regarded as missing data. A likelihood for tracking multiple objects jointly with an exclusion principle is also presented. Based on the formulation, multi-object tracking with the expectation-maximization algorithm (MOTEXATION) is presented and a variational particle filter (VPF) is then proposed to deal with the curse of dimensionality in multi-object tracking with particle filters. Results on challenging sequences with heavy occlusions demonstrate the performance of the proposed methods.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available