Time-Series Deep-Learning Classifier for Human Activity Recognition Based On Smartphone Built-in Sensors

Abstract

Human Activity Recognition (HAR) is gaining more interest in recent years due to its growing role in many human-related sectors such as the health sector especially with elderly people and motion restricted patients. In recent years, there has been great progress in identifying human activity using various machine learning approaches. However, traditional methods of feature extraction are the most challenging in the feature selection process. Deep learning is a promising approach in the human activity recognition research area and has overcome the feature selection problem. However, several challenges are still open to research issues such as classification performance. This paper describes how to identify specific types of human physical activities using the accelerator and gyroscope data generated by the smartphone user. A deep convolutional neural network architecture has been proposed to perform HAR efficiently and effectively the system has been trained and tested over a dataset generated with the aid of 50 volunteers with four activities (walking, running, walking up-down stairs finally sitting-standing on the chair) events in real-world conditions. We chose four classes, each of which performs well, get to know our range of activities achieving 99% for validation and 99.8% for testing overall accuracy