The mission of the research presented in this thesis is to give computers the power to sense and react to human activities. Without the ability to sense the surroundings and understand what humans are doing, computers will not be able to provide active, timely, appropriate, and considerate services to the humans. To accomplish this mission, the work stands on the shoulders of two giants: Machine learning and ubiquitous computing. Because of the ubiquity of sensor-enabled mobile and wearable devices, there has been an emerging opportunity to sense, learn, and infer human activities from the sensor data by leveraging state-of-the-art machine learning algorithms. While having shown promising results in human activity recognition, most existing approaches using supervised or semi-supervised learning have two fundamental problems. Firstly, most existing approaches require a large set of labeled sensor data for every target class, which requires a costly effort from human annotators. Secondly, an unseen new activity cannot be recognized if no training samples of that activity are available in the dataset. In light of these problems, a new approach in this area is proposed in our research. This thesis presents our novel approach to address the problem of human activity recognition when few or no training samples of the target activities are available. The main hypothesis is that the problem can be solved by the proposed NuActiv activity recognition framework, which consists of modeling the hierarchical and sequential structure of human activities, as well as bringing humans in the loop of model training. By injecting human knowledge about the hierarchical nature of human activities, a semantic attribute representation and a two-layer attribute-based learning approach are designed. To model the sequential structure, a probabilistic graphical model is further proposed to take into account the temporal dependency of activities and attributes. Finally, an active learning algorithm is developed to reinforce the recognition accuracy using minimal user feedback. The hypothesis and approaches presented in this thesis are validated by two case studies and real-world experiments on exercise activities and daily life activities. Experimental results show that the NuActiv framework can effectively recognize unseen new activities even without any training data, with up to 70-80% precision and recall rate. It also outperforms supervised learning with limited labeled data for the new classes. The results significantly advance the state of the art in human activity recognition, and represent a promising step towards bridging the gap between computers and humans.

Durant les situations de crise, telles que les catastrophes, le besoin de recherche d’informations (RI) pertinentes partagées dans les microblogs en temps réel est inévitable. Cependant, le grand volume et la variété des flux d’informations partagées en temps réel dans de telles situations compliquent cette tâche. Contrairement aux approches existantes de RI basées sur l’analyse du contenu, nous proposons de nous attaquer à ce problème en nous basant sur les approches centrées utilisateurs tout en levant un certain nombre de verrous méthodologiques et technologiques inhérents : 1) à la collection des données partagées par les utilisateurs à évaluer, 2) à la modélisation de leurs comportements, 3) à l’analyse des comportements, et 4) à la prédiction et le suivi des utilisateurs primordiaux en temps réel. Dans ce contexte, nous détaillons les approches proposées dans cette thèse afin de prédire les utilisateurs primordiaux qui sont susceptibles de partager les informations pertinentes et exclusives ciblées et de permettre aux intervenants d’urgence d’accéder aux informations requises quel que soit le format (i.e. texte, image, vidéo, lien hypertexte) et en temps réel. Ces approches sont centrées sur trois principaux aspects. Nous avons tout d’abord étudié l’efficacité de différentes catégories de mesures issues de la littérature et proposées dans cette thèse pour représenter le comportement des utilisateurs. En nous basant sur les mesures pertinentes résultant de cette étude, nous concevons des nouvelles caractéristiques permettant de mettre en évidence la qualité des informations partagées par les utilisateurs selon leurs comportements. Le deuxième aspect consiste à proposer une approche de modélisation du comportement de chaque utilisateur en nous basant sur les critères suivants : 1) la modélisation des utilisateurs selon l’évolution de l’évènement, 2) la modélisation de l’évolution des activités des utilisateurs au fil du temps à travers une représentation sensible au temps, 3) la sélection des caractéristiques les plus discriminantes pour chaque phase de l’évènement. En se basant sur cette approche de modélisation, nous entraînons différents modèles de prédiction qui apprennent à différencier les comportements des utilisateurs primordiaux de ceux qui ne le sont pas durant les situations de crise. Les algorithmes SVM et MOG-HMMs ont été utilisés durant la phase d’apprentissage. La pertinence et l’efficacité des modèles de prédiction appris ont été validées à l’aide des données collectées par notre système multi-agents MASIR durant deux inondations qui ont eu lieu en France et des vérités terrain appropriées à ces collections. / During crisis events such as disasters, the need of real-time information retrieval (IR) from microblogs remains inevitable. However, the huge amount and the variety of the shared information in real time during such events over-complicate this task. Unlike existing IR approaches based on content analysis, we propose to tackle this problem by using user-centricIR approaches with solving the wide spectrum of methodological and technological barriers inherent to : 1) the collection of the evaluated users data, 2) the modeling of user behavior, 3) the analysis of user behavior, and 4) the prediction and tracking of prominent users in real time. In this context, we detail the different proposed approaches in this dissertation leading to the prediction of prominent users who are susceptible to share the targeted relevant and exclusive information on one hand and enabling emergency responders to have a real-time access to the required information in all formats (i.e. text, image, video, links) on the other hand. These approaches focus on three key aspects of prominent users identification. Firstly, we have studied the efficiency of state-of-the-art and new proposed raw features for characterizing user behavior during crisis events. Based on the selected features, we have designed several engineered features qualifying user activities by considering both their on-topic and off-topic shared information. Secondly, we have proposed a phase-aware user modeling approach taking into account the user behavior change according to the event evolution over time. This user modeling approach comprises the following new novel aspects (1) Modeling microblog users behavior evolution by considering the different event phases (2) Characterizing users activity over time through a temporal sequence representation (3) Time-series-based selection of the most discriminative features characterizing users at each event phase. Thirdly, based on this proposed user modeling approach, we train various prediction models to learn to differentiate between prominent and non-prominent users behavior during crisis event. The learning task has been performed using SVM and MoG-HMMs supervised machine learning algorithms. The efficiency and efficacy of these prediction models have been validated thanks to the data collections extracted by our multi-agents system MASIR during two flooding events who have occured in France and the different ground-truths related to these collections.