1D Convolution approach to human activity recognition using sensor data and comparison with machine learning algorithms

Abstract

Human Activity Recognition (HAR) has emerged as a major player in this era of cutting-edge technological advancement. A key role that HAR plays is its ability to remotely monitor people. The objective of this paper is to classify human activities based on the data retrieved from smartphone sensors (accelerometer and gyroscope). The human activities that will be classified are namely; sitting, standing, climbing up and down the staircase, walking and laying down. To perform HAR from the data obtained, machine learning models are formed and fine-tuned in order to achieve the best results. The classic Machine Learning algorithms that have been put to use are Logistic regression, Linear and Kernel SVM, Decision Tree and Random Forest. Furthermore, a feed-forward Deep Neural Network and a 1D Convolutional Neural Network are proposed and compared with the aforementioned machine learning algorithms. Evaluation of a particular model has been carried out based on its Recall score, Precision and F1 score. Based on the results obtained from the evaluation process, it was found that the SVM and the proposed 1D convolution neural network were the best-performing models.