Multi-person Daily Activity Recognition with Non-contact Sensors based on Activity Co-occurrence
Tomokazu Matsui, Shinya Misaki, Yuma Sato, Manato Fujimoto, Hirohiko Suwa, Keiichi Yasumoto · 2021
Recognizing the activities of daily living (ADLs) for a multi-person residence is necessary for the provision of smart services, such as health care and behavior change support in “real-world” home environments. Especially, activity recognition by non-contact sensors, such as infrared sensors, is required for monitoring elderly people who are reluctant to wear devices or install cameras. However, an ADL recognition system for a multi-person environment using non-contact sensors has not been proposed due to the difficulty of identifying individual residents. In this study, we analyzed the data collected from ordinary homes with multiple residents with our ADL sensing system, consisting of only non-contact sensors, and developed a method for recognizing ADLs in multi-person environments. In a multi-person environment, multiple residents often simultaneously perform the same activity, such as sleeping, but they also frequently perform different mutually related activities at the same time, such as one person eating while another person is cooking. Hence, to recognize multi-person ADLs, we used co-occurrence relationships between the multi-person ADLs in each home as knowledge, and developed a machine learning model that used the co-occurrence of each resident’s activity history as a feature. To evaluate the proposed method, we conducted a one-month ADL sensing experiment in four ordinary homes under conditions of multi-person and non-scripted living. The results of the analysis with a deep neural network showed that the method of considering co-occurrence of multi-person ADLs improved the recognition accuracy by 5%, giving a weighted F-measure of 66.7%.