Deep Learning for Human Activity/Action Recognition Based Sensor and Smartphone

EasyChair Preprint 8478

Abstract

In recent years, the field of sensor-based Human Action Recognition (HAR) has become one of the most developed research areas, benefiting from the evolution and availability of electronic devices in our daily lives, and from the exponential evolution of artificial intelligence (AI) as well. In this context, its powerful advances are constantly exploited by researchers to develop useful methods for obtaining the best performance needed to solve existing challenges, thus, putting its contributions in favor of medical applications, taking the example of the intelligent monitoring field which is characterized by automatic, continuous, remote, and real-time monitoring of the actions of the elderly, pregnant women at home, and even hospitalized patients suffering from all kinds of mental and behavioral disorders caused by neurodegenerative diseases such as Alzheimer's, Parkinson's disease and various types of addiction. The opportunities, as well as the advantages offered by HAR, are then largely exploited in medicine and other fields such as robotics (human-computer interaction), sports discipline, and many others. Thanks to the different architectures of recurrent neural networks, deep learning (DL) has proven its effectiveness and robustness in the fields of AI and computer vision to the extent of being equal to or even exceeding human skills in terms of particular tasks such as speech recognition, language translation, pattern recognition, and in particular HAR that are going to be covered in this article. Our goal is to examine and compare the performance of some pivotal approaches in the field of human activity recognition based on smartphone sensors. We analyzed both LSTM and GRU models. In order to ensure equality of treatment and to obtain a reliable comparison, the implementation of these architectures is carried out on several sets of data. The experimental results obtained relate to the regular indicative measurements of the HAR domain.

Keyphrases: Accelerometer and gyroscope sensor, Convolutional Neural Network (CNN), Deep Learning (DL), Gated Recurrent Unit (GRU), Human activity/action Recognition (HAR), Long Short-Term Memory (LSTM), Recurrent Neural Networks (RNN), Sensors, Smartphone, time series

@booklet{EasyChair:8478,
  author    = {Youssef Errafik and Adil Kenzi and Younes Dhassi},
  title     = {Deep Learning for Human Activity/Action Recognition Based Sensor and Smartphone},
  howpublished = {EasyChair Preprint 8478},
  year      = {EasyChair, 2022}}