Constellation, le dépôt institutionnel de l'Université du Québec à Chicoutimi

Résumé

At present time, aging of the population is one of the main challenges of the 21st century. The current situation is leading to an increased number of people afflicted with cognitive disorders such as Alzheimer's disease. This group of people suffers from a progressive decline in their abilities to perform what are called the activities of the daily living (ADLs).The consequence of this reality is the urgent need for more home assistance services, as these people desire to continue living independently at home. To address this important issue, Smart Home laboratories such as LIARA, DOMUS and MavHome perform research in order to propose technological solutions for assistance provision to residents of the Smart Home. Assisting people in carrying out their ADLs, increasing quality of life and optimizing spent energy are some of the goals in Smart Home design. Technically speaking, a Smart Home is an ambient environment which, through its embedded sensors, captures data resulting from the observation of activities carried out in this environment. This data is then analyzed by artificial intelligence techniques in order to provide information about home state normality and needed assistance. In the end, the system aims to intervene by providing guidance through its actuators. In this context, activity recognition becomes a key element in order to be able to provide adequate information services at the right moment.

This thesis aims to contribute to this important challenge relating to activity recognition in the Smart Home designed for cognitive assistance. This contribution follows in the footsteps of temporal data mining and activity recognition approaches, and proposes a new way to automatically recognize and memorize ADLs from low-level sensors. From a formal point of view, the originality of the thesis relies on the proposition of a new unsupervised temporal datamining model for activity recognition addressing the problem of current temporal approaches based on Allen's framework. This new model incorporates some applications of fuzzy logic in order to take into account the uncertainty present in the realization of daily living activities by the resident. More specifically, we propose an extension of the fuzzy clustering technique in order to cluster the observations based on the degree of similarity between observations, so that activities are modeled and recognized. Moreover, anomaly recognition, decision making for assistance provision and judgment for simultaneous activities are some of the applicative contributions of this thesis. From a practical and experimental standpoint, the contribution of this research has been validated in order to evaluate how it would perform in a realistic context. To achieve this, we used MATLAB software as a simulation platform to test the proposed model. We then performed a series of tests which took the form of several case studies relating to common activities of daily living, in order to show the functionality and efficiency of the proposed temporal data-mining approach for real-life cases. This was especially relevant to the activity recognition application. We obtained very promising results which have been analyzed and compared to existing approaches. Finally, most parts of the contribution presented in this thesis have been published in documents ensuing from reputed international conferences (Springer LNCS proceedings [7], AAAI symposium and workshops [8, 9], MAICS [10], IEEE [11]) and a recognized journal (Springer Journal of Ambient Intelligence and Humanized Computing [12, 13]). This clearly constitutes recognition showing the potential of the proposed contribution.