Pocket Switched Networks (PSNs) offer a means of infrastructureless inter-human communication by utilising Delay and Disruption Tolerant Networking (DTN) technology. However, creating PSNs involves solving challenges which were not encountered in the Deep Space Internet for which DTN technology was originally intended.End-to-end communication over multiple hops in PSNs is a product of short range opportunistic wireless communication between personal mobile wireless devices carried by humans. Opportunistic data delivery in PSNs is far less predictable than in the Deep Space Internet because human movement patterns are harder to predict than the orbital motion of satellites. Furthermore, PSNs require some scheme for efficient neighbour discovery in order to save energy and because mobile devices in PSNs may be unaware of when their next encounter will take place.This thesis offers novel solutions for neighbour discovery and opportunistic data delivery in PSNs that make practical use of dynamic inter-human encounter patterns.The first contribution is a novel neighbour discovery algorithm for PSNs called PISTONS which relies on a new inter-probe time calculation (IPC) and the bursty encounter patterns of humans to set the time between neighbour discovery scans. The IPC equations and PISTONS also give participants the ability to easily specify their required level of connectivity and energy saving with a single variable.This thesis also contains novel distributed spatio-temporal clustering and opportunistic data delivery algorithms for PSNs which can be used to deliver data over multiple hops. The spatio-temporal clustering algorimths are also used to analyse the social networks and transient groups which are formed when humans interact.

Epileptic seizures are characterised by an onset of abnormal brain activity that evolves in space and time, which ultimately returns to normal background activity. For different types of seizures, the abnormal activity can be vastly different both in duration, electrographic morphology and spatial extent. Mechanistic understanding of the different seizure dynamics (spatially, as well as temporally) is crucial for the advancement and improvement of clinical treatment. To gain a deeper mechanistic insight into different seizure dynamics, mathematical models of brain processes were developed in this thesis. These models are used to explain electrographic seizure dynamics in their temporal, as well as their spatio-temporal evolution. Our studies show that the temporal evolution of seizure dynamics can be understood in terms of prototypic waveforms, which in turn can be represented in terms of three neural population processes. Such a minimal framework lends itself to a detailed phase space analysis, which elucidates seizure waveforms and seizure transitions as topological properties of the phase space. Based on the phase space considerations we show how during spike-wave seizures, single-pulse stimuli can have more complex effects than previously thought. In terms of the spatio-temporal dynamics of seizures, mechanisms for focal seizure onset and propagation are investigated in a model cortical sheet of coupled, discretised columns. The coupling followed nearest-neighbour, as well as realistic mesoscopic cortical connectivities. Different possible causes (e.g. spatial heterogeneities) of seizure generation, as well as different seizure spreading patterns (via different networks) have been investigated. We conclude that focal seizure onset can be due to global (e.g. whole-brain level) causes, global conditions & local triggers, and local (e.g. cortical column level) causes. Clinically relevant predictions from this work include the suggestion of a specific stimulation protocol in spike-wave seizures that incorporates phase space information; and the suggestion of using microscopic cortical incisions to disrupt the integrity of abnormal cortical tissue in order to prevent focal seizure onset. In conclusion, multi-scale computational modelling of seizure dynamics is proposed as an important tool to link theoretical understanding, experimental results, and patient-specific clinical data.

Philosophiae Doctor - PhD / This research investigated the application of recurrent neural networks (RNNs) for recognition of facial expressions based on facial action coding system (FACS). Support vector machines (SVMs) were used to validate the results obtained by RNNs. In this approach, instead of recognizing whole facial expressions, the focus was on the recognition of action units (AUs) that are defined in FACS. Recurrent neural networks are capable of gaining knowledge from temporal data while SVMs, which are time invariant, are known to be very good classifiers. Thus, the research consists of four important components: comparison of the use of image sequences against single static images, benchmarking feature selection and network optimization approaches, study of inter-AU correlations by implementing multiple output RNNs, and study of difference images as an approach for performance improvement. In the comparative studies, image sequences were classified using a combination of Gabor filters and RNNs, while single static images were classified using Gabor filters and SVMs. Sets of 11 FACS AUs were classified by both approaches, where a single RNN/SVM classifier was used for classifying each AU. Results indicated that classifying FACS AUs using image sequences yielded better results than using static images. The average recognition rate (RR) and false alarm rate (FAR) using image sequences was 82.75% and 7.61%, respectively, while the classification using single static images yielded a RR and FAR of 79.47% and 9.22%, respectively. The better performance by the use of image sequences can be at- tributed to RNNs ability, as stated above, to extract knowledge from time-series data. Subsequent research then investigated benchmarking dimensionality reduction, feature selection and network optimization techniques, in order to improve the performance provided by the use of image sequences. Results showed that an optimized network, using weight decay, gave best RR and FAR of 85.38% and 6.24%, respectively. The next study was of the inter-AU correlations existing in the Cohn-Kanade database and their effect on classification models. To accomplish this, a model was developed for the classification of a set of AUs by a single multiple output RNN. Results indicated that high inter-AU correlations do in fact aid classification models to gain more knowledge and, thus, perform better. However, this was limited to AUs that start and reach apex at almost the same time. This suggests the need for availability of a larger database of AUs, which could provide both individual and AU combinations for further investigation. The final part of this research investigated use of difference images to track the motion of image pixels. Difference images provide both noise and feature reduction, an aspect that was studied. Results showed that the use of difference image sequences provided the best results, with RR and FAR of 87.95% and 3.45%, respectively, which is shown to be significant when compared to use of normal image sequences classified using RNNs. In conclusion, the research demonstrates that use of RNNs for classification of image sequences is a new and improved paradigm for facial expression recognition.

Sur les dernières décennies, progresse sur la compréhension de la sismicité de clustering dans le temps, taille et l'espace ont été chassés par deux approches parallèles. De l'on étudie la main sur la mécanique des failles dans un milieu élastique soutiennent pour la contrainte statique de déclenchement à dominent dans le champ proche, c'est à dire à une distance de moins de 10 longueur de faute. De l'autre part, les propriétés de champ moyen du déclenchement sont reproduits en utilisant des effets en cascade dans modèles de processus point. Dans cette étude, nous essayons de concilier ces approches en soulignant l'importance de la faille de style sur les propriétés moyennes de la sismicité. En commençant par l'étude du taux de sismicité déclenchée par la Ville Muzaffarabad, au Cachemire, 2005 Mw = 7.6, Ms = 7,7 tremblement de terre, qui apparaît comme supérieure à la moyenne dans l'analyse des séquences répliques dans la ceinture de collision Inde-Asie, nous décidons de la productivité évènement grève de glissement sont en moyenne 4 fois plus petit que la poussée des failles de productivité. En utilisant le catalogue sismique global, nous étendre ce résultat comme tous les paramètres de la loi d'Omori (P, K, α, N (t)) étant dépendant failles styles. Dans le K ETAS modèle solide, N et de faibles valeurs de α sont entraînés par branchement ratio élevé (n). Comme conséquences de la relative n haute - La valeur la poussée des événements, il prédit aussi une faible p - La valeur pour l'événement de la poussée que comparer à bordereau de grève et des manifestations normales de la PN> PS S> PT que nous observons. Dans l'état de taux et cadre de la friction cela implique un changement dans les habitudes hétérogénéité stress. Nous ne résolvent pas tout changement dans le taux robustes foreshocks, p ' - La valeur, alors que notre analyse nous permettra d'étendre B ˚ droit aths dans le temps, d'espace et mécanisme focal. Pour des failles inverses, l'ampleur différence et la distance entre le choc principal de la plus grande réplique sont un peu moins que pour les défauts de glissement de grève. La distribution des intervalles de temps entre mainshocks et leurs le plus grand répliques est conforme au droit d'Omori, mais avec un taux un peu plus rapide de la carie que pour les répliques en général. Ceci implique que la plus grande réplique est plus susceptible de survient plus tôt que plus tard dans une séquence donnée de répliques. Par ailleurs, cette constatation plaide en faveur allant au-delà du modèle de point de branchement, avec des implications pour les prévisions à court terme. Aussi nous résoudre la dépendance univoque de p - La valeur du droit à l'ampleur Omori choc principal pour la réplique dans les 10 jours après la survenance choc principal, cette dépendance étant perdus lors de l'utilisation des séquences en cascade complète. Nous trouvons ce seuil correspond également le temps à un changement dans les schémas de diffusion, tous ces changements se synchroniser avec l'apparition de la plus grande réplique. En conséquence, nos résultats convergent vers le rôle clé du secondaire répliques sur la mécanique de la taille, le temps et l'espace modèle de processus en cascade. / On the last decades, progresses on the understanding of clustering seismicity in time, size and space have been driven by two parallel approaches. From the one hand studies on the mechanics of faulting in an elastic medium argue for the static stress triggering to dominate in the near field, i.e within distance less than 10 fault length. From the other hand, mean field properties of the triggering are reproduced using cascading effects in point process models. In this study we try to reconcile these approaches by emphasizing the importance of faulting style on average properties of seismicity. Starting with the study of the seismicity rate triggered by the Muzaffarabad, Kashmir, 2005 Mw = 7.6, Ms = 7.7 earthquake, which appears as above the average when analysing the aftershocks sequences in the India-Asia collision belt, we resolve the strike slip event productivity to be on average 4 times smaller than the thrust faulting productivity. Using global earthquake catalog, we further extend this result as all the parameters of the Omori law (p, K, α, N (t)) being dependent on faulting styles. Within the ETAS model strong K, N and low α values are driven by high branching ratio (n). As consequences of the relative high n − value of the thrust events, it also predicts a lower p − value for thrust event as compare to strike slip and normal events as the pN > pS S > pT we observe. Within rate and state friction framework it implies a change in stress heterogeneity patterns. We do not resolve any robust changes in foreshocks rate, p′ − value, whereas our analysis allow us to extend B˚aths law in time, space and focal mechanism. For reverse faults, both the magnitude difference and the distance from the mainshock to the largest aftershock are somewhat less than for strike slip faults. The distribution of time intervals between mainshocks and their largest aftershocks is consistent with Omori's law but with a somewhat faster rate of decay than for aftershocks in general. This implies that the largest aftershock is more likely to occurs earlier than later in a given sequence of aftershocks. Moreover, this finding argues for going beyond the branching point model, with implications for short term forecasts. Also we resolve unambiguous dependency of p − value of Omori law to mainshock magnitude for the aftershock within 10 days after the mainshock occurrence, this dependency being lost when using complete cascade sequences. We find this time threshold also corresponds to a change in diffusion patterns, all these changes synchronize with the occurrence of the largest aftershock. Accordingly, our results converge toward the key role of the secondary aftershocks on the mechanics of size, time and space pattern of cascading processes.

Many physical quantities around us vary across space or space-time. An example of a spatial quantity is provided by the temperature across Sweden on a given day and as an example of a spatio-temporal quantity we observe the counts of the corona virus cases across the globe. Spatial and spatio-temporal data enable opportunities to answer many important questions. For example, what the weather would be like tomorrow or where the highest risk for occurrence of a disease is in the next few days? Answering questions such as these requires formulating and learning statistical models. One of the challenges with spatial and spatio-temporal data is that the size of data can be extremely large which makes learning a model computationally costly. There are several means of overcoming this problem by means of matrix manipulations and approximations. In paper I, we propose a solution to this problem where the model islearned in a streaming fashion, i.e., as the data arrives point by point. This also allows for efficient updating of the learned model based on newly arriving data which is very pertinent to spatio-temporal data. Another interesting problem in the spatial context is to study the causal effect that an exposure variable has on a response variable. For instance, policy makers might be interested in knowing whether increasing the number of police in a district has the desired effect of reducing crimes there. The challenge here is that of spatial confounding. A spatial map of the number of police against the spatial map of the number of crimes in different districts might show a clear association between these two quantities. However, there might be a third unobserved confounding variable that makes both quantities small and large together. In paper II, we propose a solution for estimating causal effects in the presence of such a confounding variable. Another common type of spatial data is point or event data, i.e., the occurrence of events across space. The event could for example be a reported disease or crime and one may be interested in predicting the counts of the event in a given region. A fundamental challenge here is to quantify the uncertainty in the predicted counts in a model in a robust manner. In paper III, we propose a regularized criterion for learning a predictive model of counts of events across spatial regions.The regularization ensures tighter prediction intervals around the predicted counts and have valid coverage irrespective of the degree of model misspecification.

The presence of acoustic noise in audio recordings is an ongoing issue that plagues many applications. This ambient background noise is difficult to reduce due to its unpredictable nature. Many single channel noise reduction techniques exist but are limited in that they may distort the desired speech signal due to overlapping spectral content of the speech and noise. It is therefore of interest to investigate the use of multichannel noise reduction algorithms to further attenuate noise while attempting to preserve the speech signal of interest. Specifically, this thesis looks to investigate the use of microphone arrays in conjunction with multichannel noise reduction algorithms to aid aiding in speaker identification. Recording a speaker in the presence of acoustic background noise ultimately limits the performance and confidence of speaker identification algorithms. In situations where it is impossible to control the noise environment where the speech sample is taken, noise reduction algorithms must be developed and applied to clean the speech signal in order to give speaker identification software a chance at a positive identification. Due to the limitations of single channel techniques, it is of interest to see if spatial information provided by microphone arrays can be exploited to aid in speaker identification. This thesis provides an exploration of several time domain multichannel noise reduction techniques including delay sum beamforming, multi-channel Wiener filtering, and Spatial-Temporal Prediction filtering. Each algorithm is prototyped and filter performance is evaluated using various simulations and experiments. A three-dimensional noise model is developed to simulate and compare the performance of the above methods and experimental results of three data collections are presented and analyzed. The algorithms are compared and recommendations are given for the use of each technique. Finally, ideas for future work are discussed to improve performance and implementation of these multichannel algorithms. Possible applications for this technology include audio surveillance, identity verification, video chatting, conference calling and sound source localization.

Cette thèse est consacrée à une analyse de l'activité d'éclairs des systèmes orageux en Afrique équatoriale (10°E - 35°E ; 15°S - 10°N) sur la période de temps 2005-2013. Tout d'abord, les données fournies par le réseau global de détection d'éclairs WWLLN (World Wide Lightning Location Network) ont été comparées à celles obtenues par le capteur optique spatial LIS (Lightning Imaging Sensor) afin d'estimer l'efficacité de détection relative du WWLLN. Ensuite, elles ont permis d'établir une climatologie régionale à haute résolution de l'activité d'éclairs. Enfin, elles ont été associées à des données sur les caractéristiques nuageuses et météorologiques pour des études de cas d'orages dans différentes situations, afin d'examiner les corrélations entre activité d'éclairs, activité orageuse, caractéristiques nuageuses et conditions météorologiques. La méthode adaptée pour estimer l'efficacité de détection du WWLLN dans la zone d'étude a permis d'obtenir des valeurs compatibles avec celles trouvées dans d'autres régions du monde, et de mettre en évidence une variabilité spatio-temporelle qui aide à l'interprétation des changements affectant plusieurs paramètres de l'activité d'éclairs. La climatologie réalisée dévoile des caractéristiques originales de l'évolution temporelle et de la distribution spatiale de l'activité d'éclairs, notamment celles d'un maximum très aigu dans l'Est de la République Démocratique du Congo. Ainsi, la localisation, les dimensions, la forme, la persistance saisonnière et l'environnement de ce maximum ont été précisés. La distribution zonale des éclairs montre une forte proportion dans la bande tropicale sud, liée au maximum principal mais aussi à une forte activité étalée longitudinalement et constituant un large maximum secondaire où l'activité orageuse est plus variable spatialement d'une année à l'autre, temporellement d'une saison à l'autre, et où le cycle diurne est moins marqué.[...] / This thesis is devoted to an analysis of the lightning activity of storm systems in Equatorial Africa (10°E-35°E; 15°S-10°N) over the period 2005-2013. Firstly, data from the World Wide Lightning Location Network (WWLLN) were compared with those from the Lightning Imaging Sensor (LIS) to estimate the relative detection efficiency of the WWLLN. Then, they established a high-resolution regional climatology of lightning activity. Finally, they were combined with data on cloud and meteorological characteristics to carry out thunderstorm case studies in different situations in order to examine the correlations between lightning activity, storm activity, cloud characteristics and meteorological conditions. The appropriate method introduced for estimating the WWLLN detection efficiency in the study area provides values ??consistent with those found in other regions of the world. Its spatial and temporal variability helps to interpret changes affecting several parameters of lightning activity. The climatology realized reveals original characteristics of the temporal evolution and the spatial distribution of the lightning activity, in particular those of a very sharp maximum in the Eastern Democratic Republic of Congo. Thus, the location, the dimensions, the shape, the seasonal persistence and the environment of this maximum have been specified. The zonal distribution of lightning shows a high proportion in the southern tropical band, linked to the principal maximum but also to a high activity spread out longitudinally and constituting a large secondary maximum where the storm activity is more spatially variable from one year to another, temporally from one season to another, and where the diurnal cycle is less marked. [...]

As a core problem in video analysis, action recognition is of great significance for many higher-level tasks, both in research and industrial applications. With more and more video data being produced and shared daily, effective automatic action recognition methods are needed. Although, many deep-learning methods have been proposed to solve the problem, recent research reveals that single-stream, RGB-based networks are always outperformed by two-stream networks using both RGB and optical flow as inputs. This dependence on optical flow, which indicates a deficiency in learning motion, is present not only in 2D networks but also in 3D networks. This is somewhat surprising since 3D networks are explicitly designed for spatio-temporal learning. In this thesis, we assume that this deficiency is caused by difficulties associated with learning from videos exhibiting strong temporal variations, such as sudden motion, occlusions, acceleration, or deceleration. Temporal variations occur commonly in real-world videos and force a neural network to account for them, but often are not useful for recognizing actions at coarse granularity. We propose a Dynamic Equilibrium Module (DEM) for spatio-temporal learning through adaptive Eulerian motion manipulation. The proposed module can be inserted into existing networks with separate spatial and temporal convolutions, like the R(2+1)D model, to effectively handle temporal video variations and learn more robust spatio-temporal features. We demonstrate performance gains due to the use of DEM in the R(2+1)D model on miniKinetics, UCF-101, and HMDB-51 datasets.

As cities grow, efficient public transport systems are becoming increasingly important. To offer a more efficient service, public transport providers use systems that predict arrival times of buses, trains and similar vehicles, and present this information to the general public. The accuracy and reliability of these predictions are paramount, since many people depend on them, and erroneous predictions reflect badly on the public transport provider. When public transport vehicles move throughout the cities, they create motion patterns, which describe how their positions change over time. This thesis proposes a way of modeling their motion patterns using Gaussian processes, and investigates whether it is possible to predict the arrival times of public transport buses in Linköping based on their motion patterns. The results are evaluated by comparing the accuracy of the model with a simple baseline model and a recurrent neural network (RNN), and the results show that the proposed model achieves superior performance to that of an RNN trained on the same amounts of data, with excellent explainability and quantifiable uncertainty. However, an RNN is capable of training on much more data than the proposed model in the same amount of time, so in a scenario with large amounts of data the RNN outperforms the proposed model.

Congestion is a major problem in most metropolitan areas and given the increasingrate of urbanization it is likely to be an even more serious problem in the rapidlyexpanding mega cities. One possible method to combat congestion is to provide in-telligent traffic management systems that can in a timely manner inform drivers aboutcurrent or predicted traffic congestions that are relevant to them on their journeys. Thedetection of traffic congestion and the determination of whom to send in advance no-tifications about the detected congestions is the objective of the present research. Byadopting a grid based discretization of space, the proposed system extracts and main-tains traffic flow statistics and mobility statistics from the grid based recent trajectoriesof moving objects, and captures periodical spatio-temporal changes in the traffic flowsand movements by managing statistics for relevant temporal domain projections, i.e.,hour-of-day and day-of-week. Then, the proposed system identifies a directional con-gestion as a cell and its immediate neighbor, where the speed and flow of the objectsthat have moved from the neighbor to the cell significantly deviates from the histori-cal speed and flow statistics. Subsequently, based on one of two notification criteria,namely, Mobility Statistic Criterion (MSC) and Linear Movement Criterion (LMC),the system decides which objects are likely to be affected by the identified conges-tions and sends out notifications to the corresponding objects such that the numberof false negative (missed) and false positive (unnecessary) notifications is minimized.The thesis discusses the design and DBMS-based implementation of the proposedsystem. Empirical evaluations on realistically simulated trajectory data assess the ac-curacy of the methods and test the scalability of the system for varying input sizes andparameter settings. The accuracy assessment results show that the MSC based systemachieves an optimal performance with a true positive notification rate of 0.67 and afalse positive notification rate of 0.05 when min prob equals to 0.35, which is superiorto the performance of the LMC based system. The execution time of- and the spaceused by the system scales linearly with the input size (number of concurrently movingvehicles) and the methods mutually dependent parameters (grid resolution r and RTlength l) that jointly define a spatio-temporal resolution. Within the area of a large  city (40km by 40km), assuming a 60km/h average vehicle speed, the system, runningon a commodity personal computer, can manage the described congestion detectionand three-minute-ahead notification tasks within real-time requirements for 2000 and20000 concurrently moving vehicles for spatio-temporal resolutions (r=100m, l=19)and (r=2km, l=3), respectively.