Fall-detection on a wearable micro controller using machine learning algorithms

Abstract

Wearables providing fall detection can provide faster emergency services for elderly, yet privacy concerns limit acceptance of this technology. In this work, we evaluate a machine learning algorithm, called Bosnai, for embedded edge devices to detect falls. The prototype is Arduino based and can be integrated into fabrics for clothes, belts, or other accessories. The fall detection is performed offline on the device. We used data from public datasets of movement and fall events to train a tree-based machine learning model. We evaluated different combinations of prepossessed parameters as input features for the learning algorithm. The learned model is transferred to the microcontroller and can classify the sensor data offline but in real-time. We evaluate the performance of our device by performing intensive test runs with the prototype. The microcontroller is extremely limited in terms of memory capacity and computing performance, which only allows a limited number of features for learning. For this reason, it is especially important to preprocess the raw accelerator data and select the right features for training and inference. Our results show that the best performance (approx. 94.2 % accuracy) is achieved when we choose absolute acceleration and variance as features, with a sampling rate of 20 Hz and a recording window of 3s, as this system is the most robust against external interference.