Human Activity Recognition Using a Single Wrist IMU Sensor via Deep Learning Convolutional and Recurrent Neural Nets

Abstract

Recognizing and recording human activities using a smart sensor device is an essential technology for smart living. The recorded activities (i.e., life logs) could be used as valuable information to support smart life, lifecare, and healthcare services. For sensing human activities , smart sensors are required and most smart devices such as smart phones, smart bands, and smart watches incorporate Inertial Measurement Units (IMUs) which could be utilized for this purpose. However, implementing a robust Human Activity Recognition (HAR) system with high recognition accuracy using only a single sensor (i.e., no multiple sensors) is still a technical challenge. In this paper, we propose novel deep learning-based HAR systems with a single wrist IMU sensor. We used time series activity data from only one IMU sensor at a wrist to build two deep learning algorithm-based HAR systems: one is based on Convolutional Neural Nets (CNN) and the other Recurrent Neural Nets (RNN). Our two HAR systems are evaluated by 5-fold cross-validation tests to compare the performance of both systems. Five primary daily activities including standing, walking, running, walking downstairs, and walking upstairs were recognized. Our results show that the CNN-based HAR system achieved an average accuracy of 95.43% and the RNN-based HAR system an accuracy of 96.95%.