R ́eseaux de capteurs UWB pour la reconnaissance d’activit ́es humaines bas ́ee sur l’IoT et l’intelligence artficielle

Abstract

Post-stroke care involves intensive physical therapy over several months, creating a significant financial burden on both patients and society, as well as a heavy reliance on professionals and medical facilities. Home rehabilitation is a promising and cost-effective

approach, allowing long-term monitoring with minimal supervision. The smart environ- ment provides an encouraging solution to address this concern by assisting patients at home through the integration of sensors and artificial intelligence. At the core of this ad-vancement is activity recognition, a complex task aimed at identifying gestures and actions performed by patients. The challenge is linked to the nature of activities to be recognized,available sensors, and the impact on privacy. Among the sensor choices for human activity recognition, Impulse Radio Ultra-Wideband (IR-UWB) have garnered significant interest, capable of striking a favorable balance between precision and privacy preservation. The research covered in this thesis aims to establish an activity recognition system that en- compasses Hand Gestures (HG) and Human Actions (HA) to assist post-stroke patients in their home therapy. Initially, a deep learning model was proposed, combining a Con-volutional Neural Network (CNN) and a Long Short-Term Memory (LSTM) network for recognizing HG and HA from IR-UWB data. This model has the advantage of a parallel and hybrid architecture, maintaining an optimized structure with a minimal number of trainable parameters. Subsequently, we proposed a new classification model specifically designed for multi-sensor systems, called Multi-Input Multi-Output Convolutional Extra Trees (MIMO-CxT). This model excels in merging data from various sources with min-imal preprocessing, standing out for its high performance and reinforcing the accuracy, reliability, and robustness of recognition systems.