Search

The analysis and understanding of abnormal behaviours in human crowds is a challenging task in pattern recognition and computer vision. First of all, the semantic definition of the term "crowd" is ambiguous. Secondly, the taxonomy of crowd behaviours is usually rudimentary and intrinsically complicated. How to identify and construct effective features for crowd behaviour classification is a prominent challenge. Thirdly, the acquisition of suitable video for crowd analysis is another critical problem. In order to address those issues, a categorization model for abnormal behaviour types is defined according to the state-of-the-art. In the novel taxonomy of crowd behaviour, eight types of crowd behaviours are defined based on the key visual patterns. An enhanced social force-based model is proposed to achieve the visual realism in crowd simulation, hence to generate customizable videos for crowd analysis. The proposed model consists of a long-term behavior control model based on A-star path finding algorithm and a short-term interaction handling model based on the enhanced social force. The proposed simulation approach produced all the crowd behaviours in the new taxonomy for the training and testing of the detection procedure. On the aspect of feature engineering, an innovative signature is devised for assisting the segmentation of crowd in both low and high density. The signature is modelled with derived features from Grey-Level Co-occurrence Matrix. Another major breakthrough is an effective approach for efficiently extracting spatial temporal information based on the information entropy theory and Gabor background subtraction. The extraction approach is capable of obtaining the texture with most motion information, which could help the detection approach to achieve the real-time processing. Overall, these contributions have supported the crucial components in a pipeline of abnormal crowd behaviour detecting process. This process is consisted of crowd behaviour taxonomy, crowd video generation, crowd segmentation and crowd abnormal behaviour detection. Experiments for each component show promising results, and proved the accessibility of the proposed approaches.

Magnetic nano-materials have recently attracted much attention since they may have potentially many new exciting applications in a wide range of areas. The research described in this thesis has particularly focused on design, synthesis, assembling, testing and characterisation of some novel magnetic nano-materials for potential applications. Thus, a number of new magnetic nanoparticles, in particular core-shell metal, metal oxide or alloy particles encapsulated in ultra-thin porous materials including FeOx, FePt, and Fe3Pt in C, Si02 and Ti02 coatings have been prepared using various bottom-up nanochemistry synthetic methods including micro-emulsion, controlled decomposition, polyol reduction process, hydrothermal and slow evaporation techniques. Also, in this thesis, a number of chemical treatments for the surface engineering of the magnetic particles are discussed. It is important to stress that this thesis describes some first exploratory research on the uses of magnetic nanoparticles as magnetic separable nano-carriers for carrying solvent droplets, bio-molecules and catalytically active species (oxidation) for applications. The studies of two dimensional arrays of coated magnetic nanoparticles as ultra high density surfaces for magnetic memory storage and three dimensional arrays as colloid supercrystals for protein separation, etc are also described in this thesis. The thesis is divided into seven chapters. The first chapter reviews the latest literature in the area of magnetic nano-materials. The second chapter describes the uses of these nanomaterials to assist partition coefficient and protein binding constant measurements in collaboration with my sponsor, AstraZeneca. The third, fourth and fifth chapters reveal the design, synthesis and the architecture for one, two and three dimensional assembled magnetic nano-materials for applications. The sixth chapter describes analytical methodologies and instruments used for the characterisation of these particles. The final chapter summarises the main contributions from my research during the PhD period.

Learning node embedding for graphs has been proved essential for a wide range of applications, from recommendation to community search. However, most existing approaches mainly focus on either the graph structure information or the attribute (feature) information, and thus cannot make full use of the information of the graph data. In this thesis, we study four typical problems on node embedding in graphs, and utilize both structural and attribute information of the original graph to learn representative node embeddings to solve various tasks. Firstly, we investigate the problem of inductive link prediction on attributed graphs. Many real-world applications require inductive prediction for new nodes having only attribute information. It is more challenging since the new nodes do not have structure information and cannot be seen during the model training. To solve this problem, we propose a model called DEAL, which is versatile in the sense that it works for both inductive and transductive link prediction. Extensive experiments on several benchmark datasets show that our proposed model significantly outperforms existing inductive link prediction methods, and also outperforms the state-of-the-art methods on transductive link prediction. Secondly, we focus on the data imbalance problem of link prediction. Positive Unlabeled (PU) learning is a good choice to address this issue since only positive links are available. Unfortunately, the unknown class prior and data imbalance of graphs impede the use of PU learning. To deal with these issues, we propose a novel model-agnostic PU learning algorithm for Graph Neural Network (GNN)-based link prediction via PU-AUC optimization. The proposed method is free of class prior estimation and able to handle the data imbalance. Moreover, we propose an accelerated method to reduce the operational complexity of PU-AUC optimization from quadratic to linear. Extensive experiments back up our theoretical analysis and validate that the proposed method is capable of boosting the performance of the state-of-the-art GNN-based link prediction models. Thirdly, we focus on generating representative node embedding on dynamic graphs. Most of the existing graph embedding methods assume that the graph is static, where nodes and edges are fixed. However, in practice, real-world graphs (e.g., social networks, citation networks, and transportation networks) are often dynamic, meaning that both the topology and node attributes of the graph evolve over time. As a result, these methods ignore the evolution history and cannot perform well in dynamic graphs. Last but not least, we study the graph keyword search problem which aims to find subtrees or subgraphs containing all query keywords ranked according to some criteria. Existing studies all assume that the graphs have complete information. However, real-world graphs may usually contain some missing information (such as edges or keywords), thus making the problem much more challenging. To solve this, we propose a novel model named KS-GNN based on the graph neural network and the auto-encoder. By considering the latent relationships and the frequency of different keywords, KS-GNN alleviates the effect of missing information and is able to learn low-dimensional representative node embeddings that preserve both graph structure and keyword features. The experiments on four real-world datasets show that our model consistently achieves better performance than state-of-the-art baseline methods in graphs having missing information.

Human and environment influence each other dynamically and are co- adapting. Not only a specific stimulus influences people, but stimuli change may also trigger state transition. This dissertation treats dynamic stimuli with changing valence or intensity as probes to detect emotional responses and self- regulation, measuring electroencephalograph (EEG) and psychophysiological signals (ECG, skin conductance, eye tracking, respiration, etc.). I hypothesize that dynamic stimuli are more salient than static stimuli for emotional processes because they are unpredictable and consume more control resources. For ex- ample, intermittent noise may have a more negative impact than static noise. I designed four studies with different dynamic stimuli to understand how they influence emotion and the role of self-regulation. Study 1 shows that the deterioration of attention inhibition after dynamic emotional processing may reflect the additional cognitive effort required to process dynamic shifts in affective stimuli. I also found evidence this relationship is exacerbated by chronic stress. Study 2 further examined how flexible self- regulation is related to flexible neural activity change during affective stimuli transitions. Study 3 examined cortical and cardiovascular regulation to neural responses in an ambient sound environment. This study implicated that environmental context can trigger different self-regulation mechanisms both cortically and in the autonomic nervous system. Study 4 examined the effect of dynamic changes in brightness on acute stress recovery. Contrary to study 1, though dynamic sunlight captured more of people’s attention than static sunlight, it reduced beta power, which was associated with more relaxation. These studies suggest dynamic stimuli with transitions have more impact on people, and this may happen via shifts in executive functioning (task-switching, inhibition, and working memory). Furthermore, the impacts of self-regulation vary according to context.