Comparative Study of Time Series-based Human Activity Recognition using Convolutional Neural Networks

Human Activity Recognition (HAR) is an emerging active research area which applies both digital signal processing methods and machine learning algorithms to predict motion activity based on ubiquitously available sensor measurements. Inertial sensors are commonly used in HAR systems. Deep learning (DL) techniques, especially using Convolutional Neural Network (CNN), have been widely used due to their structure flexibility and high classification accuracy. In this paper, we study different CNN structures and test their generalization ability under various conditions. The aim is to find the best CNN model for HAR systems in obtaining higher prediction rates using short time segment of inertial measurements data (1 s) compared to longer time segments (2.5 s) that are used in current literature. The results from this study show that using a simple 1-D shallow CNN fused with standard global statistical features of the input time signal provide the highest precision and recall. The study also shows that enhancing the architecture to multimodal CNN improves the recognition rate of some activities. The system was tested on two publicly available datasets, namely UCI HAR and m-Health. The proposed architecture achieved recognition accuracy of 97.12% using subject-based cross validation of UCI HAR dataset. Moreover, the generalization ability of the UCI HAR based system was verified by testing on m-Health dataset. This cross-dataset evaluation showed an overall accuracy of 93.09%.