Detection of life habits evolution of frail people in a smart dwelling

To face the increase of the number of frail people due to global population ageing, innovative solutions are explored to ensure to people staying in at home a satisfying quality of health.Ambient Assisted Living (AAL) scientific field aims at exploiting smart home technologies to ease ageing at home and to offer satisfying health and living conditions.In particular, numerous existing works exploit wearable sensors in order to monitor their vital signs: temperature, heart rate, blood pressure, etc.These sensors offer relevant information to medical staff to assess the health status of an individual.However, some health troubles such as physical decline or cognitive impairments trigger first behavior changes, which trigger in a second step alterations of vital signs.This situation is particularly complex for medical staff as observing vital signs alone consequently leads to late and complex diagnostic of the original disease.As these diseases typically impair the elderly, this problem is particularly critical.Therefore, this thesis proposes an approach for smart home inhabitant behavior monitoring.Behavior refers to the way the inhabitant carries his everyday tasks.The objective is to detect behavioral deviations and to inform medical staff in order to help them in their prognosis and diagnosis.This methodology is enabled thanks to recent existing works in activity recognition, allowing to know which activity the inhabitant is carrying according to the sensors he triggers.Human behavior is extremely rich.An extensive literature review is proposed, concerning both the medical and AAL scientific field, in order to identify the health trouble and symptoms of interest for medical staff, and the way they impact patient behavior.Two behavioral features were identified as relevant due to their wide coverage: activity ordering, and activity duration.Moreover, the behavior of an individual might be impacted in two different manners: behavioral anomalies which correspond to brutal behavioral changes due to accident or sudden disease, and long-term deviations which are slow and progressive changes of behavior mainly due to degenerative troubles.The work presented in this thesis aims at detecting these two types of behavioral deviations regarding the two identified features.It focus on a single smart home inhabitant, and consider binary information only.This way, any sensor type can be used, including the most respectful of life privacy.The contributions of this thesis can be discomposed into three parts.In order to detect behavior deviations, model-based approach is proposed.Therefore, the first contribution is a Stochastic Timed Automaton (STA) model which represents the usual life habits of the inhabitant after a training phase.In a second step, this model is exploited in order to detect anomalies within inhabitant's behavior during a monitoring phase.Lastly, the model is used to detect long-term deviations through data forecasting in order to detect potential degenerative troubles.For each of these contributions, two case studies are proposed.The first one is based on artificial data generated from a real smart home in order to test challenging scenarios, while the second scenario is proposed to assess the relevancy of the proposed approach on a real scenario.Finally, as the handle data are particularly sensitive, a reflection about the potential negative ethical impacts and a method to evaluate their seriousness is proposed in an appendix, along with consideration to decrease their severity.