Investigating recognition accuracy improvement by adding user's acceleration data to location and power consumption-based in-home activity recognition system

Recently, there are many studies on automatic recognition of activities of daily living (ADL) to provide various ser-vices such as elderly monitoring, intelligent concierge, and health support. In particular, real-time ADL recognition is essential to realize an intelligent concierge service since the service needs to know user's current or next activity for supporting it. We have been studying real-time ADL recognition using only user's position data and appliances' power consumption data which are considered to include less privacy information than audio and visual data. In the study, we found that some activities such as reading and operating smartphone that happen in similar conditions cannot be classified with only position and power data. In this paper, we propose a new method that adds the acceleration data from wearable devices for classifying activities happening in similar conditions with higher accuracy. In the proposed method, we use the acceleration data from a smart watch and a smartphone worn by user's arm and waist, respectively, in addition to user's position data and appliances' power consumption data, and construct a machine learning model for recognizing 15 types of target activities. We evaluated the recognition accuracy of 3 methods: our previous method (using only position data and power consumption data); the proposed method using the mean value and the standard deviation of the acceleration norm; and the pro-posed method using the ratio of the activity topics. We collected the sensor data in our smart home facility for 12 days, and applied the proposed method to these sensor data. As a result, the proposed method could recognize the activities with 57 % which is 12 % improvement from our previous method without acceleration data.