Deep Metric Learning for Sensor-based Human Activity Recognition

Abstract

Although human activity recognition using wearable sensors has become a useful technology, activity recognition using acceleration sensor data is still under development. We obtained ideas from the image field and verified the method of introducing deep metric learning into sensor-based activity recognition. As verification of this method, we confirmed the effect on the estimation accuracy and the effect of the visualization of the feature representation. In addition, we proposed three methods for verification and searched for suitable methods by comparing them. There was a difference in the estimation accuracy and the visualization result with the proposed method. We also confirmed that significant features can be obtained for activity recognition when a suitable method is used.