Current projections show that the world's population is aging. This elderly population is prone to frailty with increased risks in their daily lives. Thus, aging in the best possible conditions and maintaining frail subjects at home has become a challenge, almost a necessity. Numerous projects have been carried out around the world, with architectures including connected objects with different types of sensors and different detection techniques to assist the elderly in their daily lives. Frailty is a geriatric syndrome characterized by weakness, emaciation and low physical activity. The objective of this thesis is to identify the daily risks related to frailty and to remedy them. We propose a wearable device to prevent at-risk situations for frail individuals in their homes. The proliferation of wearables has enabled the collection of a huge amount of physiological, motor and environmental data to assess the condition and risks faced by humans in different environments. Thus, instrumenting a frail person to be able to accompany them in their daily life is becoming an increasingly affordable task.
Different semantic risk situations have been reported such as fraud risk and domestic accident risk. The difficulty of real world studies with frail subjects is that semantic risk situations are both complex and "complementary". For our research we set up a device worn by healthy volunteers recording data using a wearable kit including IoT devices. Using this device we recorded a dataset with simulated risk situations in a real environment. After conclusive tests, performed on a lifelog dataset for the detection of semantic risk situations, we were interested in the detection of risk situations on time series.
First, we tried LSTM neural networks as a basic algorithm and then we tested neural networks with added attention.

Visioconférence PhD, here.