Human Activity Recognition By Optimizing Neural Network With Stochastic Gradient Descent On Sensor Data
Abstract
Human Activity Recognition (HAR) is a challenging area of research and it is an important one. People perform various activities in their daily life. Analysis of these activities is mainly beneficial for elderly support and healthcare. In this paper, a simulation environment has been modeled for the purpose of human activity recognition. A user independent machine learning based approach for human activity recognition is presented. The use of stochastic gradient descent for optimizing the neural network is being proposed while testing the x, y and z-axis data that is obtained from accelerometer and gyroscope sensor. The accuracy of the proposed approach is being evaluated on UCI dataset that consist of pre labeled data from multiple users. The experiment has been conducted to precisely recognize six activities of daily life, and the proposed approach has been compared to various machine learning methods. The results showcase that the proposed approach outperforms the currently available methods in terms of accuracy and other performance parameters, without any manual feature engineering.