Fall Detection by Ambient Sensors on Years-Long Simulation Data
Kai Tanaka, Mineichi Kudo, Keigo Kimura · 2024
Falls in indoors are a risk that can happen for many elderly people. Especially for elderly persons living alone, it is important to detect falls in real time for keeping their healthy and independent living. In this line, many studies have been made, mainly using cameras or wearable sensors. Compared with those sensors, ambient motion sensors have less impact on residents' privacy, and are also advantageous in battery consumption and setting cost. Unfortunately, traditional fall detection methods with ambient sensors are often limited in experimental situation, observation duration and resident variety. To cope with this problem, we improved a sensor data simulator so as to make falls occur in a large variety in when, where, and who, and produced sensor data for nine years including 26 falls. On the basis of this data, we develop a house-scale, long-term fall detection system using ceiling motion sensors in a virtual smart home. It achieved sensitivity of 0.92 with one false alarm per 216.60 hours.