Recent advances in surface performance capture have demonstrated highly realistic reconstruction of human motion sequences acquired in a multi-camera studio. Freeviewpoint rendering techniques allow video-realistic replay of captured motions with interactive viewpoint control. However, current approaches do not provide methods for real-time motion and appearance manipulation, hindering the reuse of captured data. Non-sequential temporal alignment techniques enable the conversion of an unstructured set of 3D videos into temporally consistent mesh sequences with shared topology and vertex correspondence, known as 4D videos. The work presented in this thesis aims to develop methods to exploit 4D video datasets of human motion. In particular, a framework for real-time interactive control of a 4D video character created by the combination of multiple captured sequences is investigated. Our goal is to provide methods for video-realistic character animation and rendering with the flexibility and real-time interactive control associated with conventional skeleton driven animation. An approach for parametric motion control of mesh sequences is introduced by blending multiple semantically related 4D video motions. This enables real-time interactive control using high-level parameters such as speed and direction for a walk motion. A hybrid non-linear mesh blending method is introduced. Our approach provides both realistic deformations and real-time performance, allowing parametric mesh spaces to be built at run time. A novel graph representation, referred to as 4D Parametric Motion Graph (4DPMG), is presented to en~apsulate multiple independent parametric motion spaces and transition between them whilst maintaining natural non-rigid surface motions. 4DPMGs provide online path optimisation for transitions between parametric spaces of 4D video motions with low-latency, enabling responsive interactive character control with a large range of motions. The final piece in the puzzle to enable video-realistic animations is provided by 4D Video Textures (4DVT), a new approach for free-viewpoint appearance manipulation that maintains the visual realism of the source video data. 4DVT enables video-realistic rendering of novel motions by combining multiple 4D video textures to synthesise plausible dynamic surface appearance. View-dependant optical flow is used for online alignment of parametric appearance from multiple views. The research presented in this thesis demonstrates for the first time video-realistic interactive character animation from 4D actor performance capture. This represents a new approach to animated character production which achieves video-quality rendering and does not require highly skilled manual authoring.

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.

This thesis proposes three novel solutions to the problem of detection of existence of secret messages embedded in motion vectors (MY) of a video. To this end, behaviours of MVs of natural videos are thoroughly examined in the beginning of the thesis. It has been demonstrated that MVs have strong spatial and correlation and the correlation strength is measured for the first time. Then, the following algorithms are developed considering this fact. Firstly, a novel flatness measure for video steganalysis targeting LSB based motion vector steganography has introduced. Secondly, a spatio-temporal rich model of motion vector planes as a part of a full steganalytic system against motion vector based steganography is proposed. Rich models, which have been used in image steganalysis, were developed to capture the natural correlation among image pixels. This idea is extended to motion vector based steganalysis. Lastly, a rich model based motion vector steganalysis benefiting from both temporal and spatial correlations of motion vectors is presented. The proposed method has a substantially superior detection accuracy than the previous methods, even the targeted ones. The improvement in detection accuracy lies in several novel approaches introduced in this thesis: Firstly, it is shown that there is a strong correlation for longer distances between not only spatially but also temporally neighbouring motion vectors. Therefore, temporal motion vector dependency is utilized along side the spatial dependency. Secondly, unlike the filters previously used, which were heuristically designed against a specific motion vector steganography, a diverse set of many filters which can capture aberrations introduced by various motion vector steganography methods is employed. The variety and also the number of the filter kernels are substantially more than that of previous ones. Also filters up to fifth order are employed whereas at most second order filters are used by previous methods. As a result of these novelties, the proposed system can capture various de-correlations in a wide spatio-temporal range and provide a better cover model.

This dissertation aims to raise the question of how individuals and groups become placed – or take up place – in the contemporary environment and to consider what forms the need for situatedness takes today, by examining the phenomena of the Flash Mob and Second Life. In a Flash Mob, an email activates a virtual community and converts it into a physical performance in the city, challenging a new cognition of place, where place is constituted by the event. On the other hand, Second Life takes the form of a digitally constructed world, which opens the possibility of a “virtual place” that enables users to establish connections not only with each other, but also with the [virtual] environment itself. The two case studies together question place in its materiality and its symbolism, and it is argued that they act as media to re-code “groundedness”. Thus we reach a paradoxical conclusion: although the contemporary world suggests a dynamic and more flexible existence on the earth, the need for “situatedness” and the demand for “well-grounded claims” remain stronger than ever. The structure of this research reflects a double set of conditions that, although not new, have intensified due to the emergence of new technologies: first, the expansion of the human body beyond its corporeal limits and second, the augmentation of the perceived world beyond the mere materiality of any kind of environment. Therefore the thesis studies how, on the one hand, bodies, communities and crowds transform within digitisation, and, on the other, how the world develops as a consequence of the digital reconstruction of grounds. It examines the way in which individuals detach from their “real-world groundedness” by forming bonds-connections to these digitised grounds, which display – as generators of endless possibilities – a kind of utopian openendedness. Finally, it explores the phenomenon of “virtualisation” to raise the question of whether the contemporary world is infused by information and thus augmented in terms of meanings, connections, and attachments, or is instead made of a series of projections, transforming reality into an idealised version of itself.

A consequence of the ‘data deluge’ is the exponential increase in digital video footage, while the ability to find relevant video clips diminishes. Traditional text based search engines are no longer optimal for searching, as they cannot provide a granular search of the content inside video footage. To be able to search the video in a content based manner, the content features of the video need to be extracted and modelled into a content model, which can then act as a searchable proxy for the video content. This thesis focuses on the extraction of syntactic and semantic content features and content modelling, using machine driven processes, with either little or no user interaction. Our abstract framework design extracts syntactic and semantic content features and compiles them into an integrated content model. The framework integrates a four plane strategy that consists of a pre-processing plane that removes redundant data and filters the media to improve the feature extraction properties of the media; a syntactic feature extraction plane that extracts low level syntactic feature and mid-level syntactic features that have semantic attributes; a semantic relationship analysis and linkage plane, where the spatial and temporal relationships of all the content features are defined, and finally a content modelling stage where the syntactic and semantic content features are integrated into a content model. Each of the four planes can be split into three layers namely, the content layer, where the content to be processed is stored; the application layer, where the content is converted into content descriptions, and the MPEG-7 layer, where content descriptions are serialised. Using MPEG-7 standards to produce the content model will provide wide-ranging interoperability, while facilitating granular multi-content type searches. The framework is aiming to ‘bridge’ the semantic gap, by integrating the syntactic and semantic content features from extraction through to modelling. The design of the framework has been implemented into a prototype called MAC-REALM, which has been tested and evaluated for its effectiveness to extract and model content features. Conclusions are drawn about the research output as a whole and whether they have met the objectives. Finally, future work is presented on how concept detection and crowd sourcing can be used with MAC-REALM.

The aim of this thesis is to develop a colour reproduction system for high-dynamic-range (HDR) imaging. Classical colour reproduction systems fail to reproduce HDR images because current characterisation methods and colour appearance models fail to cover the dynamic range of luminance present in HDR images. HDR tone-mapping algorithms have been developed to reproduce HDR images on low-dynamic-range media such as LCD displays. However, most of these models have only considered luminance compression from a photographic point of view and have not explicitly taken into account colour appearance. Motivated by the idea to bridge the gap between crossmedia colour reproduction and HDR imaging, this thesis investigates the fundamentals and the infrastructure of cross-media colour reproduction. It restructures cross-media colour reproduction with respect to HDR imaging, and develops a novel cross-media colour reproduction system for HDR imaging. First, our HDR characterisation method enables us to measure HDR radiance values to a high accuracy that rivals spectroradiometers. Second, our colour appearance model enables us to predict human colour perception under high luminance levels. We first built a high-luminance display in order to establish a controllable high-luminance viewing environment. We conducted a psychophysical experiment on this display device to measure perceptual colour attributes. A novel numerical model for colour appearance was derived from our experimental data, which covers the full working range of the human visual system. Our appearance model predicts colour and luminance attributes under high luminance levels. In particular, our model predicts perceived lightness and colourfulness to a significantly higher accuracy than other appearance models. Finally, a complete colour reproduction pipeline is proposed using our novel HDR characterisation and colour appearance models. Results indicate that our reproduction system outperforms other reproduction methods with statistical significance. Our colour reproduction system provides high-fidelity colour reproduction for HDR imaging, and successfully bridges the gap between cross-media colour reproduction and HDR imaging.

No description available.

No description available.

No description available.

Standard real-time 3D graphics rendering algorithms use brute force polygon rendering, with complexity linear in the number of polygons and little regard for limiting processing to data that contributes to the image. Modern hardware can now render smaller scenes to pixel levels of detail, relaxing surface connectivity requirements. Sub-linear scalability optimizations are typically self-contained, requiring specific data structures, without shared functions and data. A new point based rendering algorithm 'Canopy' is investigated that combines multiple typically sub-linear scalability solutions, using a small core of data structures. Specifically, locale management, hierarchical view volume culling, backface culling, occlusion culling, level of detail and depth ordering are addressed. To demonstrate versatility further, shadows and collision detection are examined. Polygon models are voxelized with interpolated attributes to provide points. A scene tree is constructed, based on a BSP tree of points, with compressed attributes. The scene tree is embedded in a compressed, partitioned, procedurally based scene graph architecture that mimics conventional systems with groups, instancing, inlines and basic read on demand rendering from backing store. Hierarchical scene tree refinement constructs an image tree image space equivalent, with object space scene node points projected, forming image node equivalents. An image graph of image nodes is maintained, describing image and object space occlusion relationships, hierarchically refined with front to back ordering to a specified threshold whilst occlusion culling with occluder fusion. Visible nodes at medium levels of detail are refined further to rasterization scales. Occlusion culling defines a set of visible nodes that can support caching for temporal coherence. Occlusion culling is approximate, possibly not suiting critical applications. Qualities and performance are tested against standard rendering. Although the algorithm has a 0(f) upper bound in the scene sizef, it is shown to practically scale sub-linearly. Scenes with several hundred billion polygons conventionally, are rendered at interactive frame rates with minimal graphics hardware support.