Search

The recent years have witnessed the prevalence and abundance of vision sensors in various applications such as security surveillance, healthcare and Ambient Assisted Living (AAL) among others. This is so as to realize intelligent environments which are capable of detecting users’ actions and gestures so that the needed services can be provided automatically and instantly to maximize user comfort and safety as well as to minimize energy. However, it is very challenging to automatically detect important events and human behaviour from vision sensors and summarize them in real time. This is due to the massive data sizes related to video analysis applications and the high level of uncertainties associated with the real world unstructured environments occupied by various users. Machine vision based systems can help detect and summarize important information which cannot be detected by any other sensor; for example, how much water a candidate drank and whether or not they had something to eat. However, conventional non-fuzzy based methods are not robust enough to recognize the various complex types of behaviour in AAL applications. Fuzzy logic system (FLS) is an established field of research to robustly handle uncertainties in complicated real-world problems. In this thesis, we will present a general recognition and classification framework based on fuzzy logic systems which allows for behaviour recognition and event summarisation using 2D/3D video sensors in AAL applications. I started by investigating the use of 2D CCTV camera based system where I proposed and developed novel IT2FLS-based methods for silhouette extraction and 2D behaviour recognition which outperform the traditional on the publicly available Weizmann human action dataset. I will also present a novel system based on 3D RGB-D vision sensors and Interval Type-2 Fuzzy Logic based Systems (IT2FLSs) ) generated by the Big Bang Big Crunch (BB-BC) algorithm for the real time automatic detection and summarization of important events and human behaviour. I will present several real world experiments which were conducted for AAL related behaviour with various users. It will be shown that the proposed BB-BC IT2FLSs outperforms its Type-1 FLSs (T1FLSs) counterpart as well as other conventional non-fuzzy methods, and that performance improvement rises when the number of subjects increases. It will be shown that by utilizing the recognized output activity together with relevant event descriptions (such as video data, timestamp, location and user identification) detailed events are efficiently summarized and stored in our back-end SQL event database, which provides services including event searching, activity retrieval and high-definition video playback to the front-end user interfaces.

In human communication, direct speech (e.g., Mary said: ‘I’m hungry’) coincides with vivid paralinguistic demonstrations of the reported speech acts whereas indirect speech (e.g., Mary said [that] she was hungry) provides mere descriptions of what was said. Hence, direct speech is usually more vivid and perceptually engaging than indirect speech. This thesis explores how this vividness distinction between the two reporting styles underlies language comprehension. Using functional magnetic resonance imaging (fMRI), we found that in both silent reading and listening, direct speech elicited higher brain activity in the voice-selective areas of the auditory cortex than indirect speech, consistent with the intuition of an ‘inner voice’ experience during comprehension of direct speech. In the follow-up behavioural investigations, we demonstrated that this ‘inner voice’ experience could be characterised in terms of modulations of speaking rate, reflected in both behavioural articulation (oral reading) and eye-movement patterns (silent reading). Moreover, we observed context-concordant modulations of pitch and loudness in oral reading but not straightforwardly in silent reading. Finally, we obtained preliminary results which show that in addition to reported speakers’ voices, their facial expressions may also be encoded in silent reading of direct speech but not indirect speech. The results show that individuals are more likely to mentally simulate or imagine reported speakers’ voices and perhaps also their facial expressions during comprehension of direct as opposed to indirect speech, indicating a more vivid representation of the former. The findings are in line with the demonstration hypothesis of direct speech (Clark & Gerrig, 1990) and the embodied theories of language comprehension (e.g., Barsalou, 1999; Zwaan, 2004), suggesting that sensory experiences with pragmatically distinct reporting styles underlie language comprehension.

By creating antibodies against antigens, B-cells, also named B-lymphocytes, play an important role in the immune system to fight against foreign invasion to the host body. Within the antigen specific to a certain B-cell antibody, the sections recognized and bound by antibody are called B-cell epitopes. As antigenic determinants, B-cell epitope identification is of vital importance in many immunological processes, such as vaccine design, immunodiagnostic tests, and antibody production. Towards this goal, biologists and immunologists have applied a variety of methods to identify B-cell epitopes through both experiments and bioinformatic predictions. There are two kinds of B-cell epitopes: linear (continuous) epitopes and conformational (discontinuous) epitopes. The methodologies and difficulties of bioinformatic predictions for the two categories are quite different. Due to more challenges of conformational B-cell epitope prediction, currently most of prediction tools aim to linear B-cell epitope. The importance of B-cell epitopes has driven the development of faster and more precise tools in the past thirty years. Unfortunately, the limited success of these existing methods cannot match expectation because the achieved specificity and sensitivity leave room to be desired. In this dissertation, we developed new linear B-cell epitope tool SVMTriP with a sensitivity of 80.1% and a precision of 55.2%, which is higher than other tools such as BCPred and AAP (Chapter Two). We also developed new conformational B-cell prediction tool EPSVR and a meta server EPmeta based on Support Vector Regression (Chapter Three). Comparing to other conformational B-cell prediction tools such as DiscoTope, EPSVR shows a better prediction with AUC (Area Under receiver operating characteristic Curve) of 0.597. In addition, we are working on the tool SVMKER to predict epitopic residues on antigen (Chapter Four). To our knowledge, SVMKER is the first epitopic residue prediction tool just using protein sequence as input. These online tools will provide more choices for the identification of protein epitope by bioinformatic methodology. Adviser: Chi Zhang

FePt films with the L10 phase have potential applications for magnetic recording and permanent magnets due to its high magnetocrystalline anisotropy energy density. Heat treatment of n multilayer films is one approach to form the L10 FePt phase through a solid state reaction. This thesis has studied the diffusion and reaction of n multilayer films to form the L10 FePt phase and has used this understanding to construct exchange spring magnets. The process-structure-property relations of n multilayer films were systematically examined. The transmission electron microscopy (TEM) study of the annealed multilayers indicates that the Pt layer grows at the expense of Fe during annealing, forming a disordered fcc FePt phase by the interdiffusion of Fe into Pt. This thickening of the fcc Pt layer can be attributed to the higher solubilities of Fe into fcc Pt, as compared to the converse. For the range of film thickness studied, a continuous L10 FePt product layer that then thickens with further annealing is not found. Instead, the initial L10 FePt grains are distributed mainly on the grain boundaries within the fcc FePt layer and at the Fe/Pt interfaces and further transformation of the sample to the ordered L10 FePt phase proceeds coupled with the growth of the initial L10 FePt grains. A comprehensive study of annealed n films is provided concerning the phase fraction, grain size, nucleation/grain density, interdiffusivity, long-range order parameter, and texture, as well as magnetic properties. A method based on hollow cone dark field TEM is introduced to measure the volume fraction, grain size, and density of ordered L10 FePt phase grains in the annealed films, and low-angle X-ray diffraction is used to measure the effective Fe-Pt interdiffusivity. The process-structure-properties relations of two groups of samples with varying substrate temperature and periodicity are reported. The results demonstrate that the processing parameters (substrate temperature, periodicity) have a strong influence on the structure (effective interdiffusivity, L10 phase volume fraction, grain size, and density) and magnetic properties. The correlation of these parameters suggests that the annealed n multilayer films have limited nuclei, and the subsequent growth of L10 phase is very important to the extent of ordered phase formed. A correlation between the grain size of fcc FePt phase, grain size of the L10 FePt phase, the L10 FePt phase fraction, and magnetic properties strongly suggests that the phase transformation of fccL10 is highly dependent on the grain size of the parent fcc FePt phase. A selective phase growth model is proposed to explain the phenomena observed. An investigation of the influence of total film thickness on the phase formation of the L10 FePt phase in n multilayer films and a comparison of this to that of FePt co-deposited alloy films is also conducted. A general trend of greater L10 phase formation in thicker films was observed in both types of films. It was further found that the thickness dependence of the structure and of the magnetic properties in n multilayer films is much stronger than that in FePt alloy films. This is related to the greater chemical energy contained in n films than FePt alloy films, which is helpful for the L10 FePt phase growth. However, the initial nucleation temperature of n multilayers and co-deposited alloy films was found to be similar. An investigation of L10 FePt-based exchange spring magnets is presented based on our understanding of the L10 formation in n multilayer films. It is known that exchange coupling is an interfacial magnetic interaction and it was experimentally shown that this interaction is limited to within several nanometers of the interface. A higher degree of order of the hard phase is shown to increase the length scale slightly. Two approaches can be used to construct the magnets. For samples with composition close to stoichiometric L10 FePt, the achievement of higher energy product is limited by the average saturation magnetization, and therefore, a lower annealing temperature is beneficial to increase the energy product, allowing a larger fraction of disordered phase. For samples with higher Fe concentration, the (BH)max is limited by the low coercivity of annealed sample, and a higher annealing temperature is beneficial to increase the energy product.