Deep Learning Methodologies for Human Activity Recognition

Abstract

Human activity recognition (HAR) is a field that has shown great attention in recent years. The main reasons are the high demand in several application domains, and the HAR process makes use of the time-series sensor data to deduce activities. In this paper, three main deep learning methodologies are proposed based on RNN architecture. The three methodologies are based on the long-term short memory (LSTM), Bi-directional long short-term memory (Bi-LSTM), and gated recurrent unit (GRU). The proposed methodologies are capable of classifying six main movements with acceptable performance. Data were collected from 30 subjects with 6 main activities obtained from them. Five main classifiers are applied to test the performance of these methodologies, and these classifiers are the random trees, random forests, k-nearest neighbor, artificial neural network (ANN), and support vector machine (SVM). The highest accuracy was achieved using BiLSTM based on ANN classifier reaching an accuracy of 95.2155%. Several performance measurements are provided to test the methodologies' recognition capability. A comparison with other related works is done to exploit how the proposed methods are capable of providing a reasonable accuracy for HAR.