Exploring accuracy-cost tradeoff in in-home living activity recognition based on power consumptions and user positions

Advanced context-aware services at home such as elderly monitoring requires highly accurate living activity recognition in a home environment. Existing studies on living activity recognition suffer from high deployment and maintenance costs, privacy intrusion due to utilization of cameras and microphones, and few recognizable activities or low recognition accuracy. In this paper, to solve these problems, we propose a new living activity recognition method. Our method utilizes only power meters attached to appliances and a positioning sensor attached to a resident of a home to mitigate privacy intrusion. We target 10 different living activities which cover most of our daily lives at home and construct activity recognition models based on machineleaning. To accurately recognize the activities from the sensor data by power meters and position sensor, we explore the best combination of time window width for samples of training/test data, features, and machine-learning algorithms. Furthermore, we thoroughly investigate the tradeoff between the sensor data granularity and the consequent recognition accuracy. Through experiments using sensor data collected by four participants in our smart home, the proposed method achieved 97.8 % average F-measure of recognizing 10 target activities with the finest sensor data granularity (position estimation error ≤ 0.1m, 16 power meters) and 86.9 % F-measure with room-level position accuracy and one power meter for each of four rooms.