Human Activity Recognition Based on Gramian Angular Field and Convolutional Neural Network

EasyChair Preprint 5347

Abstract

Human Activity Recognition (HAR) is the use of sensor data to predict human behaviors. With the advancement of the Internet of Things and the development of microelectromechanical systems, HAR is becoming increasingly commonplace in daily life, such as smart phones and smart bracelets with built-in sensors that can detect body movements and states, which can predict the user's activities in real time. However, the data collected by the sensors are time-series, the feature values are difficult to extract. If we directly use deep learning to obtain features, it will not be able to preserve the characteristics of the time series in the data. The Gramian Angular Field (GAF) converts the cartesian coordinates of the original sensor data into polar coordinates to maintain the correlation and continuity of the time series data, and the triaxial data are combined into two dimensions as the data input. The classifier apply Convolutional Neural Network (CNN), which has excellent performance in image classification, to automatically extract the eigenvalues from the image data. In this study, we use the Actitracker dataset and our proposed method improves the accuracy rate by 5.8% over the CNN model.

Keyphrases: Convolutional Neural Network, Gramian Angular Field, Human Activity Recognition, time series

@booklet{EasyChair:5347,
  author    = {Wan-Yi Hsieh and Jui-Chung Hung},
  title     = {Human Activity Recognition Based on Gramian Angular Field and Convolutional Neural Network},
  howpublished = {EasyChair Preprint 5347},
  year      = {EasyChair, 2021}}