Human Activity Recognition Based on Smartphone Sensor Data Using Principal Component Analysis and Linear Multiclass Support Vector Machine

Abstract

Human activity recognition (HAR) requires the categorization of sequences of accelerometer data acquired by dedicated equipment or smartphones, in order to recognize discrete motions. Since there is no straightforward method to match accelerometer data to known motions, the challenge is complicated by the high number of observations made per second and the temporal structure of the observations. This paper presents a machine learning-based HAR model for classifying the primary actions such as sitting, standing, lying, walking, walking upstairs, and downstairs. The proposed model reads the data from accelerometer and gyroscope of the smartphone. This data is sent to the principal component analysis (PCA) for dimensionality reduction. PCA reduces the dimension in the input data by retaining the important features and removing the redundancy. The dimensionality reduced data is sent to linear multiclass support vector machine (SVM) for classification. SVM is first trained to identify the best hyperplane for the classification of the data which is then used to classify data in real-time. The proposed model obtained an accuracy of 98.85 during testing.