A significant number of action recognition research efforts use spatio-temporal interest point detectors for feature extraction. Although the extracted features provide useful information for recognizing actions, a significant number of them contain irrelevant motion and background clutter. In many cases, the extracted features are included as is in the classification pipeline, and sophisticated noise removal techniques are subsequently used to alleviate their effect on classification. We introduce a new action database, created from the Weizmann database, that reveals a significant weakness in systems based on popular cuboid descriptors. Experiments show that introducing complex backgrounds, stationary or dynamic, into the video causes a significant degradation in recognition performance. Moreover, this degradation cannot be fixed by fine-tuning the system or selecting better interest points. Instead, we show that the problem lies at the descriptor level and must be addressed by modifying descriptors.

L'acquisition et la modélisation de connaissances ont été abordés jusqu'à présent selon deux approches principales : les êtres humains (experts) à l'aide des méthodologies de l'Ingénierie des Connaissances et le Knowledge Management, et les données à l'aide des techniques relevant de la découverte de connaissances à partir du contenu de bases de données (fouille de données). Cette thèse porte sur la conception d'un processus d'apprentissage conjoint par l'être humain et la machine combinant une approche de modélisation des connaissances de type Ingénierie des Connaissances (TOM4D, Timed Observation Modelling for Diagnosis) et une approche d'apprentissage automatique fondée sur un processus de découverte de connaissances à partir de données datées (TOM4L, Timed Observation Mining for Learning). Ces deux approches étant fondées sur la Théorie des Observations Datées, les modèles produits sont représentés dans le même formalisme ce qui permet leur comparaison et leur combinaison. Le mémoire propose également une méthode d'abstraction, inspiée des travaux de Newell sur le "Knowledge Level'' et fondée sur le paradigme d'observation datée, qui a pour but de traiter le problème de la différence de niveau d'abstraction inhérent entre le discours d'un expert et les données mesurées sur un système par un processus d'abstractions successives. Les travaux présentés dans ce mémoire ayant été menés en collaboration avec le CSTB de Sophia Antipolis (Centre Scientifique et Technique du Bâtiment), ils sont appliqués à la modélisation de l'activité humaine dans le cadre de l'aide aux personnes âgées maintenues à domicile. / Knowledge acquisition has been traditionally approached from a primarily people-driven perspective, through Knowledge Engineering and Management, or from a primarily data-driven approach, through Knowledge Discovery in Databases, rather than from an integral standpoint. This thesis proposes then a human-machine learning approach that combines a Knowledge Engineering modelling approach called TOM4D (Timed Observation Modelling For Diagnosis) with a process of Knowledge Discovery in Databases based on an automatic data mining technique called TOM4L (Timed Observation Mining For Learning). The combination and comparison between models obtained through TOM4D and those ones obtained through TOM4L is possible, owing to that TOM4D and TOM4L are based on the Theory of Timed Observations and share the same representation formalism. Consequently, a learning process nourished with experts' knowledge and knowledge discovered in data is defined in the present work. In addition, this dissertation puts forward a theoretical framework of abstraction levels, in line with the mentioned theory and inspired by the Newell's Knowledge Level work, in order to reduce the broad gap of semantic content that exists between data, relative to an observed process, in a database and what can be inferred in a higher level; that is, in the experts' discursive level. Thus, the human-machine learning approach along with the notion of abstraction levels are then applied to the modelling of human behaviour in smart environments. In particular, the modelling of elderly people's behaviour at home in the GerHome Project of the CSTB (Centre Scientifique et Technique du Bâtiment) of Sophia Antipolis, France.

Cette thèse s’inscrit dans le contexte de la reconnaissance des activités à partir de séquences vidéo qui est une des préoccupations majeures dans le domaine de la vision par ordinateur. Les domaines d'application pour ces systèmes de vision sont nombreux notamment la vidéo surveillance, la recherche et l'indexation automatique de vidéos ou encore l'assistance aux personnes âgées. Cette tâche reste problématique étant donnée les grandes variations dans la manière de réaliser les activités, l'apparence de la personne et les variations des conditions d'acquisition des activités. L'objectif principal de ce travail de thèse est de proposer une méthode de reconnaissance efficace par rapport aux différents facteurs de variabilité. Les représentations basées sur les points d'intérêt ont montré leur efficacité dans les travaux d'art; elles ont été généralement couplées avec des méthodes de classification globales vue que ses primitives sont temporellement et spatialement désordonnées. Les travaux les plus récents atteignent des performances élevées en modélisant le contexte spatio-temporel des points d'intérêts par exemple certains travaux encodent le voisinage des points d'intérêt à plusieurs échelles. Nous proposons une méthode de reconnaissance des activités qui modélise explicitement l'aspect séquentiel des activités tout en exploitant la robustesse des points d'intérêts dans les conditions réelles. Nous commençons par l'extractivité des points d'intérêt dont a montré leur robustesse par rapport à l'identité de la personne par une étude tensorielle. Ces primitives sont ensuite représentées en tant qu'une séquence de sac de mots (BOW) locaux: la séquence vidéo est segmentée temporellement en utilisant la technique de fenêtre glissante et chacun des segments ainsi obtenu est représenté par BOW des points d'intérêt lui appartenant. Le premier niveau de notre système de classification séquentiel hybride consiste à appliquer les séparateurs à vaste marge (SVM) en tant que classifieur de bas niveau afin de convertir les BOWs locaux en des vecteurs de probabilités des classes d'activité. Les séquences de vecteurs de probabilité ainsi obtenues sot utilisées comme l'entrées de classifieur séquentiel conditionnel champ aléatoire caché (HCRF). Ce dernier permet de classifier d'une manière discriminante les séries temporelles tout en modélisant leurs structures internes via les états cachés. Nous avons évalué notre approche sur des bases publiques ayant des caractéristiques diverses. Les résultats atteints semblent être intéressant par rapport à celles des travaux de l'état de l'art. De plus, nous avons montré que l'utilisation de classifieur de bas niveau permet d'améliorer la performance de système de reconnaissance vue que le classifieur séquentiel HCRF traite directement des informations sémantiques des BOWs locaux, à savoir la probabilité de chacune des activités relativement au segment en question. De plus, les vecteurs de probabilités ont une dimension faible ce qui contribue à éviter le problème de sur apprentissage qui peut intervenir si la dimension de vecteur de caractéristique est plus importante que le nombre des données; ce qui le cas lorsqu'on utilise les BOWs qui sont généralement de dimension élevée. L'estimation les paramètres du HCRF dans un espace de dimension réduite permet aussi de réduire le temps d'entrainement / Human activity recognition (HAR) from video sequences is one of the major active research areas of computer vision. There are numerous application HAR systems, including video-surveillance, search and automatic indexing of videos, and the assistance of frail elderly. This task remains a challenge because of the huge variations in the way of performing activities, in the appearance of the person and in the variation of the acquisition conditions. The main objective of this thesis is to develop an efficient HAR method that is robust to different sources of variability. Approaches based on interest points have shown excellent state-of-the-art performance over the past years. They are generally related to global classification methods as these primitives are temporally and spatially disordered. More recent studies have achieved a high performance by modeling the spatial and temporal context of interest points by encoding, for instance, the neighborhood of the interest points over several scales. In this thesis, we propose a method of activity recognition based on a hybrid model Support Vector Machine - Hidden Conditional Random Field (SVM-HCRF) that models the sequential aspect of activities while exploiting the robustness of interest points in real conditions. We first extract the interest points and show their robustness with respect to the person's identity by a multilinear tensor analysis. These primitives are then represented as a sequence of local "Bags of Words" (BOW): The video is temporally fragmented using the sliding window technique and each of the segments thus obtained is represented by the BOW of interest points belonging to it. The first layer of our hybrid sequential classification system is a Support Vector Machine that converts each local BOW extracted from the video sequence into a vector of activity classes’ probabilities. The sequence of probability vectors thus obtained is used as input of the HCRF. The latter permits a discriminative classification of time series while modeling their internal structures via the hidden states. We have evaluated our approach on various human activity datasets. The results achieved are competitive with those of the current state of art. We have demonstrated, in fact, that the use of a low-level classifier (SVM) improves the performance of the recognition system since the sequential classifier HCRF directly exploits the semantic information from local BOWs, namely the probability of each activity relatively to the current local segment, rather than mere raw information from interest points. Furthermore, the probability vectors have a low-dimension which prevents significantly the risk of overfitting that can occur if the feature vector dimension is relatively high with respect to the training data size; this is precisely the case when using BOWs that generally have a very high dimension. The estimation of the HCRF parameters in a low dimension allows also to significantly reduce the duration of the HCRF training phase

The assessment of the direct and indirect requirements for energy is known as embodied energy analysis. For buildings, the direct energy includes that used primarily on site, while the indirect energy includes primarily the energy required for the manufacture of building materials. This thesis is concerned with the completeness and reliability of embodied energy analysis methods. Previous methods tend to address either one of these issues, but not both at the same time. Industry-based methods are incomplete. National statistical methods, while comprehensive, are a ‘black box’ and are subject to errors. A new hybrid embodied energy analysis method is derived to optimise the benefits of previous methods while minimising their flaws. In industry-based studies, known as ‘process analyses’, the energy embodied in a product is traced laboriously upstream by examining the inputs to each preceding process towards raw materials. Process analyses can be significantly incomplete, due to increasing complexity. The other major embodied energy analysis method, ‘input-output analysis’, comprises the use of national statistics. While the input-output framework is comprehensive, many inherent assumptions make the results unreliable. Hybrid analysis methods involve the combination of the two major embodied energy analysis methods discussed above, either based on process analysis or input-output analysis. The intention in both hybrid analysis methods is to reduce errors associated with the two major methods on which they are based. However, the problems inherent to each of the original methods tend to remain, to some degree, in the associated hybrid versions. Process-based hybrid analyses tend to be incomplete, due to the exclusions associated with the process analysis framework. However, input-output-based hybrid analyses tend to be unreliable because the substitution of process analysis data into the input-output framework causes unwanted indirect effects. A key deficiency in previous input-output-based hybrid analysis methods is that the input-output model is a ‘black box’, since important flows of goods and services with respect to the embodied energy of a sector cannot be readily identified. A new input-output-based hybrid analysis method was therefore developed, requiring the decomposition of the input-output model into mutually exclusive components (ie, ‘direct energy paths’). A direct energy path represents a discrete energy requirement, possibly occurring one or more transactions upstream from the process under consideration. For example, the energy required directly to manufacture the steel used in the construction of a building would represent a direct energy path of one non-energy transaction in length. A direct energy path comprises a ‘product quantity’ (for example, the total tonnes of cement used) and a ‘direct energy intensity’ (for example, the energy required directly for cement manufacture, per tonne). The input-output model was decomposed into direct energy paths for the ‘residential building construction’ sector. It was shown that 592 direct energy paths were required to describe 90% of the overall total energy intensity for ‘residential building construction’. By extracting direct energy paths using yet smaller threshold values, they were shown to be mutually exclusive. Consequently, the modification of direct energy paths using process analysis data does not cause unwanted indirect effects. A non-standard individual residential building was then selected to demonstrate the benefits of the new input-output-based hybrid analysis method in cases where the products of a sector may not be similar. Particular direct energy paths were modified with case specific process analysis data. Product quantities and direct energy intensities were derived and used to modify some of the direct energy paths. The intention of this demonstration was to determine whether 90% of the total embodied energy calculated for the building could comprise the process analysis data normally collected for the building. However, it was found that only 51% of the total comprised normally collected process analysis. The integration of process analysis data with 90% of the direct energy paths by value was unsuccessful because: • typically only one of the direct energy path components was modified using process analysis data (ie, either the product quantity or the direct energy intensity); • of the complexity of the paths derived for ‘residential building construction’; and • of the lack of reliable and consistent process analysis data from industry, for both product quantities and direct energy intensities. While the input-output model used was the best available for Australia, many errors were likely to be carried through to the direct energy paths for ‘residential building construction’. Consequently, both the value and relative importance of the direct energy paths for ‘residential building construction’ were generally found to be a poor model for the demonstration building. This was expected. Nevertheless, in the absence of better data from industry, the input-output data is likely to remain the most appropriate for completing the framework of embodied energy analyses of many types of products—even in non-standard cases. ‘Residential building construction’ was one of the 22 most complex Australian economic sectors (ie, comprising those requiring between 592 and 3215 direct energy paths to describe 90% of their total energy intensities). Consequently, for the other 87 non-energy sectors of the Australian economy, the input-output-based hybrid analysis method is likely to produce more reliable results than those calculated for the demonstration building using the direct energy paths for ‘residential building construction’. For more complex sectors than ‘residential building construction’, the new input-output-based hybrid analysis method derived here allows available process analysis data to be integrated with the input-output data in a comprehensive framework. The proportion of the result comprising the more reliable process analysis data can be calculated and used as a measure of the reliability of the result for that product or part of the product being analysed (for example, a building material or component). To ensure that future applications of the new input-output-based hybrid analysis method produce reliable results, new sources of process analysis data are required, including for such processes as services (for example, ‘banking’) and processes involving the transformation of basic materials into complex products (for example, steel and copper into an electric motor). However, even considering the limitations of the demonstration described above, the new input-output-based hybrid analysis method developed achieved the aim of the thesis: to develop a new embodied energy analysis method that allows reliable process analysis data to be integrated into the comprehensive, yet unreliable, input-output framework. Plain language summary Embodied energy analysis comprises the assessment of the direct and indirect energy requirements associated with a process. For example, the construction of a building requires the manufacture of steel structural members, and thus indirectly requires the energy used directly and indirectly in their manufacture. Embodied energy is an important measure of ecological sustainability because energy is used in virtually every human activity and many of these activities are interrelated. This thesis is concerned with the relationship between the completeness of embodied energy analysis methods and their reliability. However, previous industry-based methods, while reliable, are incomplete. Previous national statistical methods, while comprehensive, are a ‘black box’ subject to errors. A new method is derived, involving the decomposition of the comprehensive national statistical model into components that can be modified discretely using the more reliable industry data, and is demonstrated for an individual building. The demonstration failed to integrate enough industry data into the national statistical model, due to the unexpected complexity of the national statistical data and the lack of available industry data regarding energy and non-energy product requirements. These unique findings highlight the flaws in previous methods. Reliable process analysis and input-output data are required, particularly for those processes that were unable to be examined in the demonstration of the new embodied energy analysis method. This includes the energy requirements of services sectors, such as banking, and processes involving the transformation of basic materials into complex products, such as refrigerators. The application of the new method to less complex products, such as individual building materials or components, is likely to be more successful than to the residential building demonstration.

The continuous progression of urbanization has resulted in an increasing number of people living in cities or towns. In parallel, advancements in technologies, such as the Internet, telecommunications, and transportation, have allowed for better connectivity among people. This has engendered drastic changes in urban systems during the recent decades. From a social geographic perspective, the changes in urban systems are primarily characterized by intensive contacts among people and their interactions with the surrounding urban environment, which further leads to subsequent challenging problems such as traffic jams, environmental pollution, urban sprawl, etc. These problems have been reported to be heterogeneous and non-deterministic. Hence, to cope with them, massive amounts of geographic data are required to create new knowledge on urban systems. Due to the thriving of Volunteer Geographic Information (VGI) in recent years, this thesis presents knowledge on urban systems based on extensive VGI datasets from three sources: highway dataset from the OpenStreetMap (OSM) project, photo location dataset from the Flickr website, and GPS tracking datasets from volunteers, taxicabs, and air flights. The knowledge primarily relates to two issues of urban systems: the urban space and the corresponding human dynamics. In accordance, on one hand, urban space acts as a carrier for associated geographic activities and knowledge of it benefits our understanding of current social and economic problems in urban systems. On the other hand, human dynamics reflect human behavior in urban space, which leads to complex mobility or activity patterns. Its investigation allows a derivation of the underlying driving force that is very instructive to urban planning, traffic management, and infectious disease control. Therefore, to fully understand the two issues, this thesis conducts a thorough investigation from multiple aspects. The first issue is investigated from four aspects. First, at the city level, the controversial topic of city size regularity is investigated in terms of natural cities, and the conclusion is that Zipf’s law holds stably for all US cities. Second, at the sub-city level, the size distribution of spatial units within different cities in terms of the clusters formed by street nodes, photo locations, and taxi static points are explored, and the result shows a remarkable scaling property of these spatial units. Third, enlightened by the scaling property of the urban space at the city or sub-city level, this thesis devises a novel tool that can demarcate the cities into three categories: compact cities, normal cities, and sprawling cities. The tool is then applied to cities in both the US and three European countries. In the last, another representation of urban space is taken into account, namely the transportation network. The findings report that the US airport network displays the properties of scale-free, small-world, and disassortative mixing and that the individual natural airports show heterogeneous patterns that are probably subject to geographic constraints and socioeconomic factors. The second issue is examined from four perspectives. First, at the city level, the movement flow contributed by agents using two types of behavior is investigated through an agent-based simulation, and the result conjectures that the human mobility behavior is mainly shaped by the underlying street network. Second, at the country level, this thesis reports that the human travel length by air can be approximated well by an exponential distribution, and subsequent simulations indicate that human mobility behavior is largely constrained by the underlying airport network. Third, at the regional level, the length that humans travel by car is demonstrated to agree well with a power law with exponential cutoff distribution, and subsequent simulation further reproduces this levy flight characteristic. Based on the simulation, human mobility behavior is again revealed to be primarily shaped by the underlying hierarchical spatial structure. Finally, taxicab static points are adopted to explore human activity patterns, which can be characterized as the regularities in space and time, the heterogeneity and predictability in space. From a complex system perspective, this thesis presents the knowledge discovered in urban systems using massive volumes of geographic data. Together with new knowledge from empirical findings, the development of methods, and the design of theoretic models, this thesis also shares the research community with geographic data generated from extensive VGI datasets and the corresponding source codes. Moreover, this study is aligned with a paradigm shift in that it analyzes large-size datasets using high processing power as opposed to analyzing small-size datasets with low processing power. / <p>QC 20121113</p>

Comprendre la genèse d’une situation de conduite requiert d’analyser les choixfaits par le conducteur au volant de son véhicule pendant l’activité de conduite, dans sacomplexité naturelle et dans sa dynamique située. Le LESCOT a développé le modèleCOSMODRIVE, fournissant un cadre conceptuel pour la simulation cognitive de l’activitéde conduite automobile. Pour exploiter ce modèle en simulation, il est nécessairede produire les connaissances liées à la situation de conduite sous forme d’un automatepar exemple. La conception d’un tel automate nécessite d’une part de disposer de donnéesissues de la conduite réelle, enregistrées sur un véhicule instrumenté et d’autrepart d’une expertise humaine pour les interpréter.Pour accompagner ce processus d’ingénierie des connaissances issues de l’analysed’activité, ce travail de thèse propose une méthode de découverte interactive deconnaissances à partir de traces d’activité. Les données de conduite automobile sontconsidérées comme des M-Traces, associant une sémantique explicite aux données,exploitées en tant que connaissances dans un Système à Base de Traces (SBT). Le SBTpermet de filtrer, transformer, reformuler et abstraire les séquences qui serviront à alimenterla synthèse de modèles automates de l’activité de conduite. Nous reprenons destechniques de fouille de workflow permettant de construire des automates (réseaux dePetri) à partir de logs. Ces techniques nécessitent des données complètes ou statistiquementreprésentatives. Or les données collectées à bord d’un véhicule en situationde conduite sont par nature des cas uniques, puisqu’aucune situation ne sera jamaisreproductible à l’identique, certaines situations particulièrement intéressantes pouvanten outre être très rarement observées. La gageure est alors de procéder à une forme degénéralisation sous la forme de modèle, à partir d’un nombre de cas limités, mais jugéspertinents, représentatifs, ou particulièrement révélateurs par des experts du domaine.Pour compléter la modélisation de telles situations, nous proposons donc de rendreinteractifs les algorithmes de synthèse de réseau de Petri à partir de traces, afin depermettre à des experts-analystes de guider ces algorithmes et de favoriser ainsi la découvertede connaissances pertinentes pour leur domaine d’expertise. Nous montreronscomment rendre interactifs l’algorithme α et l’algorithme α+ et comment généralisercette approche à d’autres algorithmes.Nous montrons comment l’utilisation d’un SBT et de la découverte interactived’automates impacte le cycle général de découverte de connaissances. Une méthodologieest proposée pour construire des modèles automates de l’activité de conduiteautomobile.Une étude de cas illustre la méthodologie en partant de données réelles de conduiteet en allant jusqu’à la construction de modèles avec un prototype logiciel développédans le cadre de cette thèse / Driving is a dynamic and complex activity. Understanding the origin of a driving situationrequires the analysis of the driver’s choices made while he/she drives. In addition,a driving situation has to be studied in its natural complexity and evolution. LESCOThas developed a model called COSMODRIVE, which provides a conceptual frameworkfor the cognitive simulation of the activity of car driving. In order to run themodel for a simulation, it is necessary to gather knowledge related to the driving situation,for example in the form of an automaton. The conception of such an automatonrequires : 1) the use of real data recorded in an instrumented car, and, 2) the use of humanexpertise to interpret these data. These data are considered in this thesis as activitytraces.The purpose of this thesis is to assist the Knowledge Engineering process of activityanalysis. The present thesis proposes a method to interactively discover knowledgefrom activity traces. For this purpose, data from car driving are considered as M-traces– which associate an explicit semantic to these data. This semantic is then used asknowledge in a Trace Based System. In a Trace Based System, M-traces can be filtered,transformed, reformulated, and abstracted. The resulting traces are then used as inputsin the production of an automaton model of the activity of driving. In this thesis,Workflow Mining techniques have been used to build automata (Petri nets) from logs.These techniques require complete or statistically representative data sets. However,data collected from instrumented vehicles are intrinsically unique, as no two drivingsituations will ever be identical. In addition, situations of particular interest, such ascritical situations, are rarely observed in instrumented vehicle studies. The challenge isthen to produce a model which is a form of generalisation from a limited set of cases,which have been judged by domain experts as being relevant and representative of whatactually happens.In the current thesis, algorithms synthesising Petri nets from traces have been madeinteractive, in order to achieve the modelling of such driving situations. This thenmakes it possible for experts to guide the algorithms and therefore to support the discoveryof knowledge relevant to the experts. The process involved in making the α-algorithm and the α+-algorithm interactive is discussed in the thesis in a way that canbe generalised to other algorithms.In addition, the current thesis illustrates how the use of a Trace Based System andthe interactive discovery of automata impacts the global cycle of Knowledge Discovery.A methodology is also proposed to build automaton models of the activity of cardriving. Finally, a case study is presented to illustrate how the proposed methodologycan be applied to real driving data in order to construct models with the softwaredeveloped in this thesis

Avec la mise en place des politiques de conservation des espèces sauvages, l'extension de l'urbanisation et l'accroissement des populations humaines, le contact homme-faune a considérablement augmenté au cours de ces dernières décennies. Par conséquent, le nombre de maladies d'origines zoonotiques a explosé avec six apparitions d'agents infectieux par an, dont 75% sont susceptibles d’être transmises par un vecteur. La plupart de ces maladies n'ayant pas encore de vaccins, les principales méthodes d'évitement sont basées sur les stratégies de lutte anti-vectorielle adaptées à l'écologie et au comportement alimentaire des vecteurs. Au Gabon, particulièrement dans les parcs nationaux, nous avons identifié six espèces de glossines (Glossina palpalis palpalis, G. fuscipes fuscipes, G. fusca congolense, G. pallicera newsteadi, G. caliginea et G tabaniformis) vivant principalement en milieux forestiers, six espèces de stomoxes (Stomoxys calcitrans, S. inornatus, S. niger niger, S. niger bilineatus, S. omega omega et S. transvittatus) inféodées aux milieux ouverts types forêt secondaire, savane et villages. Nous avons également identifié six espèces de tabanides (Ancala sp., Atylotus sp., Chrysops sp., Haematopota sp., Tabanus par et T. taeniola), mais leur distribution n'était pas claire dans les milieux prospectés. Par ailleurs, nous constatons que ces mouches hématophages ont un régime alimentaire très diversifié, comprenant les mammifères terrestres et aquatiques, les reptiles et les oiseaux. Elles se nourrissent à 86% sur la faune, contre seulement 14% sur l'homme. Cependant, dans les milieux anthropisés les repas sanguins d'origine humaine sont très importants, notamment dans les villages (100%) et autour des camps de recherche implantés dans les parcs (24%). Ainsi en l'absence de faune dans le milieu, ces mouches hématophages se nourrissent sur l'homme. Comme 75% des maladies émergentes chez l'homme proviennent de la faune sauvage et que près de ¾ d'entre elles circulent via le sang, elles sont donc susceptibles d’être détectées dans les repas sanguins de mouches hématophages. Cette technique d'échantillonnage non-invasif de la faune sauvage semble être un bon moyen d'identifier les agents infectieux à ADN (plasmodiums et trypanosomes), mais reste encore imprécise pour les agents infectieux à ARN (arbovirus). / The contact between human and wild fauna has considerably increased during these last decades due to the increase of human population size but also to conservation policies. As a consequence, the number of zoonotic diseases soared with a mean of six new infectious diseases per year, 75% of whom being vectorially transmitted. The way to avoid the human contamination by these emergent diseases is based on the efficient vector control resulting from a deep knowledge of the ecology and the feeding behavior of the different vector species. During our work, we have identified and characterized the ecology of 6 tsetse species (Glossina palpalis palpalis, G. fuscipes fuscipes, G. fusca congolense, G. pallicera newsteadi, G. caliginea and G. tabaniformis) that live in forests and 6 stomoxe species (Stomoxys calcitrans, S. inornatus, S. niger niger, S. niger bilineatus, S. omega omega and S. transvittatus) that live in and around (anthropized places) conservation areas. We have also identified 6 tabanid species (Ancala sp., Atylotus sp., Chrysops sp., Haematopota sp., Tabanus par and T. taeniola). The feeding ecology of the tsetse species have been studied through the determination of host extracted from blood meals in the insect caught with molecular techniques. These hematophagous insects had a diversified diet that was constituted of diverse mammal species but also reptiles and birds. The food intake results mostly from wild fauna (86%) and more rarely from humans (14%). However, in anthropised habitats (villages and research’s camps within the parks), the blood intakes from human origin were important, in particular in the villages (100%), suggesting that without wild fauna the flies shift on human host. In the last part of our work, we tried to identify pathogens in the blood samples extracted from the tsetse species in order to test whether these species could be used as living sampling syringe of the wild fauna. This new proposed non-invasive sampling techniques allowed to detect the DNA of various infectious agents (plasmodiums and trypanosomes), but failed to detect the RNA of viruses (arbovirus) suggesting that this approach could be useful but need to be improved.

AMI Meeting Corpus (AMI) -databasen används för att undersöka igenkännande av gruppaktivitet. AMI Meeting Corpus (AMI) -databasen ger forskare fjärrstyrda möten och naturliga möten i en kontorsmiljö; mötescenario i ett fyra personers stort kontorsrum. För attuppnågruppaktivitetsigenkänninganvändesbildsekvenserfrånvideosoch2-dimensionella audiospektrogram från AMI-databasen. Bildsekvenserna är RGB-färgade bilder och ljudspektrogram har en färgkanal. Bildsekvenserna producerades i batcher så att temporala funktioner kunde utvärderas tillsammans med ljudspektrogrammen. Det har visats att inkludering av temporala funktioner både under modellträning och sedan förutsäga beteende hos en aktivitet ökar valideringsnoggrannheten jämfört med modeller som endast använder rumsfunktioner[1]. Deep learning arkitekturer har implementerats för att känna igen olika mänskliga aktiviteter i AMI-kontorsmiljön med hjälp av extraherade data från the AMI-databas.Neurala nätverks modellerna byggdes med hjälp av KerasAPI tillsammans med TensorFlow biblioteket. Det finns olika typer av neurala nätverksarkitekturer. Arkitekturerna som undersöktes i detta projektet var Residual Neural Network, Visual GeometryGroup 16, Inception V3 och RCNN (LSTM). ImageNet-vikter har använts för att initialisera vikterna för Neurala nätverk basmodeller. ImageNet-vikterna tillhandahålls av Keras API och är optimerade för varje basmodell [2]. Basmodellerna använder ImageNet-vikter när de extraherar funktioner från inmatningsdata. Funktionsextraktionen med hjälp av ImageNet-vikter eller slumpmässiga vikter tillsammans med basmodellerna visade lovande resultat. Både Deep Learning användningen av täta skikt och LSTM spatio-temporala sekvens predikering implementerades framgångsrikt. / The Augmented Multi-party Interaction(AMI) Meeting Corpus database is used to investigate group activity recognition in an office environment. The AMI Meeting Corpus database provides researchers with remote controlled meetings and natural meetings in an office environment; meeting scenario in a four person sized office room. To achieve the group activity recognition video frames and 2-dimensional audio spectrograms were extracted from the AMI database. The video frames were RGB colored images and audio spectrograms had one color channel. The video frames were produced in batches so that temporal features could be evaluated together with the audio spectrogrames. It has been shown that including temporal features both during model training and then predicting the behavior of an activity increases the validation accuracy compared to models that only use spatial features [1]. Deep learning architectures have been implemented to recognize different human activities in the AMI office environment using the extracted data from the AMI database.The Neural Network models were built using the Keras API together with TensorFlow library. There are different types of Neural Network architectures. The architecture types that were investigated in this project were Residual Neural Network, Visual Geometry Group 16, Inception V3 and RCNN(Recurrent Neural Network). ImageNet weights have been used to initialize the weights for the Neural Network base models. ImageNet weights were provided by Keras API and was optimized for each base model[2]. The base models uses ImageNet weights when extracting features from the input data.The feature extraction using ImageNet weights or random weights together with the base models showed promising results. Both the Deep Learning using dense layers and the LSTM spatio-temporal sequence prediction were implemented successfully.

Human activity data is typically represented as time series data, and RNNs, often with LSTM cells, are commonly used for recognition in this field. However, RNNs and LSTM-RNNs are often too resource-intensive for real-time applications on resource constrained devices, making them unsuitable. This thesis project is carried out at Wrlds AB, Stockholm. At Wrlds, all machine learning is run in the cloud, but they have been attempting to run their AI algorithms on their embedded devices. The main task of this project was to investigate alternative network structures to minimize the size of the networks to be used on human activity data. This thesis investigates the use of Fast GRNN, a deep learning algorithm developed by Microsoft researchers, to classify human activity on resource-constrained devices. The FastGRNN algorithm was compared to state-of-the-art RNNs, LSTM, GRU, and Simple RNN in terms of accuracy, classification time, memory usage, and energy consumption. This research is limited to implementing the FastRNN algorithm on Nordic SoCs using their SDK and TensorFlow Lite Micro. The result of this thesis shows that the proposed network has similar performance as LSTM networks in terms of accuracy while being both considerably smaller and faster, making it a promising solution for human activity recognition on embedded devices with limited computational resources and merits further investigation. / Rörelse igenkännings analys är oftast representerat av tidsseriedata där ett RNN modell meden LSTM arkitektur är oftast den självklara vägen att ta. Dock så är denna arkitektur väldigt resurskrävande för applikationer i realtid och gör att det uppstår problem med resursbegränsad hårdvara. Detta examensarbete är utfört i samarbete med Wrlds Technologies AB. På Wrlds så körs deras maskin inlärningsmodeller på molnet och lokalt på mobiltelefoner. Wrlds har nu påbörjat en resa för att kunna köra modeller direkt på små inbyggda system. Examensarbete kommer att utvärdera en FastGRNN som är en NN-arkitektur utvecklad av Microsoft i syfte att användas på resurs begränsad hårdvara. FastGRNN algoritmen jämfördes med andra högkvalitativa arkitekturer som RNNs, LSTM, GRU och en simpel RNN. Träffsäkerhet, klassifikationstid, minnesanvändning samt energikonsumtion användes för att jämföra dom olika varianterna. Detta arbete kommer bara att utvärdera en FastGRNN algoritm på en Nordic SoCs och kommer att användas deras SDK samt Tensorflow Lite Micro. Resultatet från detta examensarbete visar att det utvärderade nätverket har liknande prestanda som ett LSTM nätverk men också att nätverket är betydligt mindre i storlek och därmed snabbare. Detta betyder att ett FastGRNN visar lovande resultat för användningen av rörelseigenkänning på inbyggda system med begränsad prestanda kapacitet.

Applications for sensor-based human activity recognition use the latest algorithms for the detection and classification of human everyday activities, both for online and offline use cases. The insights generated by those algorithms can in a next step be used within a wide broad of applications such as safety, fitness tracking, localization, personalized health advice and improved child and elderly care.In order for an algorithm to be performant, a significant amount of annotated data from a specific target audience is required. However, a satisfying data collection process is cost and labor intensive. This also may be unfeasible for specific target groups as aging effects motion patterns and behaviors. One main challenge in this application area lies in the ability to identify relevant changes over time while being able to reuse previously annotated user data. The accurate detection of those user-specific patterns and movement behaviors therefore requires individual and adaptive classification models for human activities.The goal of this degree work is to compare several supervised classifier performances when trained and tested on a newly iterative user-specific human activity classification approach as described in this report. A qualitative and quantitative data collection process was applied. The tree-based classification algorithms Decision Tree, Random Forest as well as XGBoost were tested on custom based datasets divided into three groups. The datasets contained labeled motion data of 21 volunteers from wrist worn sensors.Computed across all datasets, the average performance measured in recall increased by 5.2% (using a simulated leave-one-subject-out cross evaluation) for algorithms trained via the described approach compared to a random non-iterative approach. / Sensorbaserad aktivitetsigenkänning använder sig av det senaste algoritmerna för detektion och klassificering av mänskliga vardagliga aktiviteter, både i uppoch frånkopplat läge. De insikter som genereras av algoritmerna kan i ett nästa steg användas inom en mängd nya applikationer inom områden så som säkerhet, träningmonitorering, platsangivelser, personifierade hälsoråd samt inom barnoch äldreomsorgen.För att en algoritm skall uppnå hög prestanda krävs en inte obetydlig mängd annoterad data, som med fördel härrör från den avsedda målgruppen. Dock är datainsamlingsprocessen kostnadsoch arbetsintensiv. Den kan dessutom även vara orimlig att genomföra för vissa specifika målgrupper, då åldrandet påverkar rörelsemönster och beteenden. En av de största utmaningarna inom detta område är att hitta de relevanta förändringar som sker över tid, samtidigt som man vill återanvända tidigare annoterad data. För att kunna skapa en korrekt bild av det individuella rörelsemönstret behövs därför individuella och adaptiva klassificeringsmodeller.Målet med detta examensarbete är att jämföra flera olika övervakade klassificerares (eng. supervised classifiers) prestanda när dem tränats med hjälp av ett iterativt användarspecifikt aktivitetsklassificeringsmetod, som beskrivs i denna rapport. En kvalitativ och kvantitativ datainsamlingsprocess tillämpades. Trädbaserade klassificeringsalgoritmerna Decision Tree, Random Forest samt XGBoost testades utifrån specifikt skapade dataset baserade på 21 volontärer, som delades in i tre grupper. Data är baserad på rörelsedata från armbandssensorer.Beräknat över samtlig data, ökade den genomsnittliga sensitiviteten med 5.2% (simulerad korsvalidering genom utelämna-en-individ) för algoritmer tränade via beskrivna metoden jämfört med slumpvis icke-iterativ träning.