Falls are a major health concern for the elderly, especially in care facilities, where delayed response can lead to severe consequences. This work presents a proof-of-concept wearable fall detection system based on a low-power smartwatch and a Bluetooth Mesh communication protocol. The wearable device continuously monitors movement using a triaxial accelerometer and runs a lightweight machine learning model for real-time classification of falls and daily activities. Data was collected from seven participants using both the PineTime smartwatch and a smartphone, resulting in a small labeled dataset. To improve model performance, techniques such as window slicing and overlapping strides were applied. A simple multilayer perceptron model was trained and deployed on the Nordic nRF52832 microcontroller. Despite limited input modalities (accelerometer only), the model achieved 92\% točnost on the test set from the PineTime dataset and 78\% točnost on the smartphone dataset. A Bluetooth Mesh network was also implemented using the PineTime watch, ESP32, nRF52832-DK, and a smartphone. The network successfully demonstrated provisioning, group communication, and message relaying. This system showcases the feasibility of integrating fall detection and wireless communication on resource-constrained devices, forming a basis for future development and clinical validation.
|
