Use this URL to cite or link to this record in EThOS:
Title: Visual data association : tracking, re-identification and retrieval
Author: Zheng, Feng
ISNI:       0000 0004 6350 3641
Awarding Body: University of Sheffield
Current Institution: University of Sheffield
Date of Award: 2016
Availability of Full Text:
Access from EThOS:
Access from Institution:
As there is a rapid development of the information society, large amounts of multimedia data are generated, which are shared and transferred on various electronic devices and the Internet every minute. Hence, building intelligent systems capable of associating these visual data at diverse locations and different times is absolutely essential and will significantly facilitate understanding and identifying where an object came from and where it is going. Thus, the estimated traces of motions or changes increasingly make it feasible to implement advanced algorithms to real-world applications, including human-computer interaction, robotic navigation, security in surveillance, biological characteristics association and civil structure vibration detection. However, due to the inherent challenges, such as ambiguity, heterogeneity, noisy data, large-scale property and unknown variations, visual data association is currently far from being established. Therefore, this thesis focuses on the studies of associating visual data at diverse locations and different times for the tasks of tracking, re-identification and retrieval. More specifically, three situations including single camera, across multiple cameras and across multiple modalities have been investigated and four algorithms have been developed at different levels. Chapter 3 The first algorithm is to explore an ensemble system for robust object tracking, primarily considering the independence of classifier members. An empirical analysis is firstly given to show that object tracking is a non-i.i.d. sampling, under-sample and incomplete-dataset problem. Then, a set of independent classifiers trained sequentially on different small datasets is dynamically maintained to overcome the particular machine learning problem. Thus, for every challenge, an optimal classifier can be approximated in a subspace spanned by the selected competitive classifiers. Chapter 4 The second method is to improve the object tracking by exploiting a winner-take-all strategy to select the most suitable trackers. This topic naturally extends the concept of ensemble in the first topic to a more general idea: a multi-expert system, in which members come from different function spaces. Thus, the diversity of the system is more likely to be amplified. Based on a large public dataset, a prediction model of performance for different trackers on various challenges can be obtained off-line. Then, the learned structural regression model can be directly used to efficiently select the winner tracker online. Chapter 5 The third one is to learn cross-view identities for fast person re-identification, in a cross-camera setting, which significantly differs from the single-view object tracking in the first two topics. Two sets of discriminative hash functions for two different views are learned by simultaneously minimising their distance in the Hamming space, and maximising the cross-covariance and margin. Thus, similar binary codes can be found for images of the same person captured at different views by embedding the images into the Hamming space. Chapter 6 The fourth model is to develop a novel Hetero-manifold regularisation framework for efficient cross-modal retrieval. Compared with the first two settings, this is a more general and complex topic, in which the samples can be relaxed to the images captured in the very far distance or very long time, even to text, voice and other formats. Taking advantage of the hetero-manifold, the similarity between each pair of heterogeneous data could be naturally measured by three order random walks on this hetero-manifold. It is concluded that, by fully exploiting the algorithms for solving the problems in the three situations, an integrated trace for an object moving anywhere can be definitely discovered.
Supervisor: Ling, Shao Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available