Robust Methods for Sensor-based Activity Recognition

Sozo Inoue,
International Symposium on Applied Engineering and Sciences (SAES2016)
(Not Available)
(Not Available)
In this paper, based on our work, we introduce two methods to solve the problems which appear in human activity recognition with sensors.
Firstly, we propose an approach to improve mobile activity recognition, given a training dataset with inaccurate segments, in which the beginning and ending timestamps of homogeneous and continuous activities have inaccurate boundaries due to human errors. We evaluate the proposed approach for three datasets, including simulation data and real activity data, two machine-learning algorithms, and various inaccuracies, and show that the proposed approach outperforms the naive methods.
Secondly, we propose an approach to improve the accuracy of home activity estimation using device-free sensors in the home. This is achieved by transferring existing training data to a new household considering the differences between households. We proposed the method to compose supervised transfer between labeled data and unsupervised transfer between labeled unlabeled data for each scenario. To evaluate in realistic settings, we developed the system which consists of an application for use on tablet terminals, which continuously collect light and any optional sensor data, and a Web-based server system that stores the sensor data, estimates activities, provides visualization on users' Web browsers, and enables users to edit the activity labels. Using the system, we gathered subjects from open called households during a period of approximately four months, and obtained approximately 11,745 activity inputs, approximately 7.14GB of sensor data, and power consumption data of 237,280 hours from 35 households. As a result of evaluation, our method outperformed naive methods.

Data Files

No records to display