A New Motion Data Structuring for Human Activity Recognition Using Convolutional Neural Network

Human activity recognition is an essential field to study for healthcare services and smooth human-machine interaction. For human motion data collection, wearable sensors are the common devices utilized. In this paper, a new double-channel motion data structuring method for classification is proposed for the data collected from the inertial sensor of a smartphone. It was shaped as a virtual image in a way to extract deep features of the temporal and spatial dependencies among motion signals. The time-series raw data underwent Fourier and wavelet transformations as alternative forms of input data. The virtual images from the three types of representations were made to fit a Convolutional Neural Network model for classification. The proposed model was evaluated using our dataset and other public datasets, where it performed well with all datasets. The trained model showed excellent results when tested on a computer and smartphone for real-time recognition. An exclusive iOS application with data handling and real-time recognition functions was developed for the smartphone. The model has the attributes of lower computational cost and better accuracy, which makes it a sound model for a practical purpose.