No description available.

No description available.

Recently, soft biometrics has emerged as a novel attribute-based person description for identification. It is likely that soft biometrics can be deployed where other biometrics cannot, and have stronger invariance properties than traditional vision-based biometrics, such as invariance to illumination and contrast. Previously, a variety of soft body and face biometrics have been used for identifying people and have increasingly garnered more research interest and are often considered as major cues for identity, especially in the absence of valid traditional hard biometrics, as in surveillance. Describing a person by their clothing properties is a natural task performed by people. As yet, clothing descriptions have attracted little attention for biometric purposes as it has been considered unlikely to be a potential cue to identity. There has been some usage of clothing attributes to augment biometric description, but a detailed description has yet to be used. In everyday life, several cases and incidents arise highlighting the usefulness and capability of information deduced from clothing regarding identity. Clothing is inherently more effective for short-term identification, since people can change clothes. This thesis introduces semantic clothing attributes as a new form of soft biometrics. The usability and efficacy of a novel set of proposed soft clothing traits is explored, showing how they can be exploited for human identification and re-identification purposes. Furthermore, the viability of these traits is investigated in correctly retrieving a subject of interest, given a verbal description of their clothing. The capability of clothing information is further examined in more realistic scenarios offering viewpoint invariant subject retrieval. Although clothing traits can be naturally described or compared by humans for operable and successful use, it is desirable to exploit computer-vision to enrich clothing descriptions with more objective and discriminative information. This allows automatic extraction and semantic description and comparison of visually detectable clothing traits in a manner similar to recognition by eyewitness statements. This thesis proposes further a novel set of automatic clothing attributes, described using small groups of high-level semantic labels, and automatically extracted using computer-vision techniques. In this way, we can explore the capability of clothing attributes inferred by human vis-a-vis those which are inferred automatically by computer-vision. Extended analysis of clothing information is conducted. Human identification and retrieval are achieved, evaluated, and compared using different proposed forms of soft clothing biometrics in addition and in isolation. The experimental results of identification and retrieval highlight clothing attributes as a potentially valuable addition to the field of soft biometrics.

The fusion of images from the visible and long-wave infrared (thermal) portions of the spectrum produces images that have improved face recognition performance under varying lighting conditions. This is because long-wave infrared images are the result of emitted, rather than reflected, light and are therefore less sensitive to changes in ambient light. Similarly, 3D and 2.5D images have also improved face recognition under varying pose and lighting. The opacity of glass to long-wave infrared light, however, means that the presence of eyeglasses in a face image reduces the recognition performance. This thesis presents the design and performance evaluation of a novel camera system which is capable of capturing spatially registered visible, near-infrared, long-wave infrared and 2.5D depth video images via a common optical path requiring no spatial registration between sensors beyond scaling for differences in sensor sizes. Experiments using a range of established face recognition methods and multi-class SVM classifiers show that the fused output from our camera system not only outperforms the single modality images for face recognition, but that the adaptive fusion methods used produce consistent increases in recognition accuracy under varying pose, lighting and with the presence of eyeglasses.

The research work presented in this thesis is directed at addressing the knowledge acquisition bottleneck frequently encountered in computer simulation. The central idea is to extract the required knowledge from video data and use this to drive a computer simulation instead of the more conventional approach of interviewing domain experts and somehow encapsulating this knowledge in a manner whereby it can be used in the context of computer simulation. More specifically the idea presented in this thesis is to extract object location information from video data and then to mine this information to identify Movement Patterns (MPs) and then to utalise these MPs in the context of computer simulation. To act as a focus for the work rodent behaviour simulation was considered. Partly because video data concerning rodent behaviour was relatively easy to obtain and partly because there is a genuine need to achieve a better understanding of rodent behaviour. This is especially the case in the context of crop damage. There are a variety of computer simulation frameworks. One that naturally lends itself to rodent simulation is Multi Agent Based Simulation (MABS) whereby the objects to be simulated (rodents) are encapsulated in terms of software agents. In more detail the work presented is directed at a number of research issues in the context of the above: (i) mechanisms to identify a moving object in video data and extracting associated location information, (ii) the mining of MPs from the extracted location information, (iii) the representation of MPs in such a way that they are compatible with computer simulation frameworks especially MABS frameworks and (iv) mechanisms where by MPs can be utilized and interacted with so as to drive a MABS. Overall two types of mechanisms are considered, Absolute and Relative. The operation of rodent MABSs, driven using the proposed MP concept, is fully illustrated in the context of different categories of scenarios. The evaluation of the proposed MP driven MABSs was conducted by comparing real world scenarios to parallel simulated scenarios. The results presented in the thesis demonstrated that the proposed mechanisms for extracting locations, and consequently mining MPs, from video data to drive a MABS provides a useful approach to effective computer simulation that will have wide ranging benefits.

Biometric authentication systems provide alternative solutions to traditional methods that are based on knowledge (e.g. password) or physical tokens (e.g., smart card). Many studies now focus on getting high accuracy rates for biometric verification. However,with advances in technology, biometric data (e.g. fingerprint, face, iris) can be captured/sniffed, duplicated, modified, and then resubmitted in the same or in other applications that utilize the same biometric features. Watermarking techniques can be used effectively to protect the genuine ownership of biometric data, either to accept or reject. This thesis presents a proposal for a suitable and viable combination of a face recognition algorithm and a watermarking technique, namely a Principal Component Analysis (PCA) and Discrete Cosine Transform (DCT) combination, that will ensure the authenticity of the data being transmitted in the face recognition system, which will then increase its level of security. The emphasis is on replay attack, which is recognizing and rejecting captured biometric data resubmitted into the system. The research begins with an analysis of biometric systems, with an emphasis on face recognition systems, and in particular with reference to the recorded threats on such systems. Biometric watermarking algorithms proposed by previous researchers within the face recognition environment are then studied, noting their proposed solutions to the said threats. This would then give a good idea towards a watermarking scheme to be proposed to enhance the security of face recognition systems, especially in terms of the authenticity of the data being transmitted. This proposed watermarking face recognition scheme is the main objective, which will be implemented in a PCA—DCT combination, followed by a check on all the 8 vulnerable positions where data may be captured and/or resubmitted. All the results produced are positive, apart from a few situations that will have to be left for future work. Non degradation of the individual PCA and DCT systems due to the combination is also checked and experimented on, again with positive results. Finally, the robustness of the watermarking scheme is experimented on to evaluate its resilience against attacks. The contributions from this research constitute a meaningful solution step to security problems associated with biometric techniques. The outcome of the research should also stimulate further research by opening up more research gaps in the area of combining biometric and watermarking techniques.

The ever increasing demand for higher levels of autonomy for robots and vehicles means there is an ever greater need for such systems to be aware of their surroundings. Whilst solutions already exist for creating 3D scene maps, many are based on active scanning devices such as laser scanners and depth cameras that are either expensive, unwieldy, or do not function well under certain environmental conditions. As a result passive cameras are a favoured sensor due their low cost, small size, and ability to work in a range of lighting conditions. In this work we address some of the remaining research challenges within the problem of 3D mapping around a moving platform. We utilise prior work in dense stereo imaging, Stereo Visual Odometry (SVO) and extend Structure from Motion (SfM) to create a pipeline optimised for on vehicle sensing. Using forward facing stereo cameras, we use state of the art SVO and dense stereo techniques to map the scene in front of the vehicle. With significant amounts of prior research in dense stereo, we addressed the issue of selecting an appropriate method by creating a novel evaluation technique. Visual 3D mapping of dynamic scenes from a moving platform result in duplicated scene objects. We extend the prior work on mapping by introducing a generalized dynamic object removal process. Unlike other approaches that rely on computationally expensive segmentation or detection, our method utilises existing data from the mapping stage and the findings from our dense stereo evaluation. We introduce a new SfM approach that exploits our platform motion to create a novel dense mapping process that exceeds the 3D data generation rate of state of the art alternatives. Finally, we combine dense stereo, SVO, and our SfM approach to automatically align point clouds from non-overlapping views to create a rotational and scale consistent global 3D model.

Human faces are one of the most frequently captured objects in both videos and photographs due to their fundamental role in communication and social interactions. The variability of this facial imagery makes it difficult to automate the understanding of scenes containing faces under unconstrained conditions. For faces, recovering accurate dense 3D facial shape from images and videos enables much richer understanding of the human face and its interaction with the scene. In this thesis, we seek to extend the work in the area of dense 3D facial shape recovery under challenging unconstrained conditions. There are a wealth of ways to recover 3D shape from images and videos, all of which make specific assumptions about the relationship between the individual images and the construction of the scene. Given this broad selection of methods available, we examine three different scenarios for dense 3D facial shape recovery: i) recovery from a single image, ii) recovery from an unconstrained image collection without any explicit 3D shape priors and iii) recovery from a video sequence. We focus on these three cases and show how facial priors can be introduced to tackle the dense 3D facial surface recovery problem. We propose to investigate the use of shading constraints for dense shape recovery from unconstrained images. Given the challenging nature of these images, the introduction of priors greatly improves performance over the generic shape-from-shading literature. However, the introduction of explicit priors comes with a further problem, that of correspondence. That is, recovering the relationship between pixels in the image and the structure of our model. For this reason, we also investigate the importance of finding dense correspondences between facial images. We show that it is possible to recover plausible dense 3D facial surfaces under a variety of different input conditions.

This thesis describes a novel method for the layout of undirected graphs. It works by identifying certain patterns within the graph and drawing these in a consistent manner. For graphs to be useful and of benefit to a user, the result must clear and easy to understand. This process attempts to draw graphs in such a manner. Firstly, a background of graph problems and graph drawing is introduced, before the benefits of patterns are explained. Following this, there is an in-depth discussion of a number of existing graph drawing techniques, perceptual theories and methods for subgraph isomorphism. This pattern-based method is then explained in great detail. Firstly, the patterns required are defined and examples given. Then, there is an explanation of the methodology involved in identifying these patterns within a graph. Following on from this, the order in which patterns are drawn based on their connection types to those already drawn is detailed, before a detailed description of each drawing method. Evaluation of this method follows, starting with analysis mainly based on three real world data sources. This is in the form of side-by-side comparisons of graphs drawn with this method and a force-directed method. Following this, a metric based evaluation compares the two methods on edge crossings and occlusion, while also detailing some pattern based metrics. Further evaluation continues in the form of an empirical study. The methodology of this study is detailed before results are displayed. Analysis of these results follows, with conclusions drawn. Finally, potential further work is detailed and possible implementations discussed. All study materials and results are provided in the Appendix for those who wish to repeat the study or analysis.

This thesis addresses problems in dissimilarity (proximity) learning, particularly focusing on identifying the sources and rectifying the non-Euclidean dissimilarity in pattern recog- nition. We aim to develop a framework for analyzing the non-Euclidean dissimilarity by combining the methods from differential geometry and manifold learning theory. The algorithms are applied to objects represented by the dissimilarity measures. In Chapter 3 we describe how to reveal the origins of the non-Euclidean behaviors of the dissimilarity matrix for the purpose of rectifying the dissimilarities. We com- mence by developing a new measure which gauges the extent to which individual data give rises to departures from metricity in a set of dissimilarity data. This allows us to as- sess whether the non-Euclidean artifacts in a dataset can be attributed to individual objects or are distributed uniformly. The second novel contribution of Chapter 3 is to provide sim- ple empirical tests that can be used to determine the sources of the negative dissimilarity eigenvalues. We consider three sources of the negative dissimilarity eigenvalues, namely a) that the data resides on a manifold, b) that the objects may be extended and c) that there is Gaussian noise. We experiment with the algorithms on a set of public dissimilarities used in various applications available from the EU SIMBAD project. In Chapter 4, we propose a framework for rectifying the dissimilarities using Ricci flow on the manifolds so that the non-Euclidean artifacts are eliminated, as the second main contribution of this thesis. We consider the objects of interest to be represented by points on a manifold consisting of local patches with constant curvatures, and the given dissimilarities to be the geodesic distances on the manifold between these points. In dif- ferential geometry, Ricci flow changes the metric of a Riemannian manifold according to the curvature of the manifold. We seek to flatten the curved manifold so that a corrected set of Euclidean distances are obtained. We achieve this by deforming the manifold usingRicci flow. In the first technique, we consider each edge as a local patch and apply Ricci flow independently to flatten each patch. In this way, the local structure of the manifold is ignored, as Ricci flow is applied independently on each edge. To overcome this prob- lem, we propose a second technique, where add a curvature regularization process before evolving the manifold. Specifically we use the heat kernel to smooth out the curvatures on the edges. The results show both improved numerical stability and lower classification error in the embedded space. To reduce the reliance on the piecewise embedding and its effects on individual edges, we extend the previous two techniques and develop a third means of correcting non- Euclidean dissimilarity data as the first contribution of Chapter 5. This is done by using a tangent space reprojection to inflate the local hyperspherical patches and align the local patches with the shortest edge-connected path. These three Ricci-flow-based techniques proposed through this thesis are investigated as a means of correcting the dissimilarities so that the the non-Euclidean artefacts are eliminated. We experiment on two datasets represented by dissimilarities, namely the CoilYork and the Chickenpieces datasets. In the framework for correcting the non-Euclidean dissimilarities using the Ricci flow process, estimating the curvatures of the embedded manifold is an important component prior deforming the manifold. The second contribution of Chapter 5 is the investigation of the effects of the piecewise embedding methods (the kernel embedding and the Isomap embedding) on the curvatures computation and the introduction of a new way of com- puting the curvatures from a set of dissimilarities. We consider each local patch on a hypersphere, and deduce the enclosed volume of the points in terms of the curvature. We estimate the curvature by fitting the volume. We illustrate the utility of this method for estimating curvatures on the artificial dataset (2-sphere dataset).