Ensemble One-vs-One SVM Classifier for Smartphone Accelerometer Activity Recognition
Quentin Bouchut, Kofi Appiah, Ahmad Lotfi, Patrick Dickinson · 2018
A recognition framework to identify six full body motion from smartphone sensory data is proposed. The proposed system relies on accelerometer, gyroscope and magnetometer data to classify user activities into six groups (sitting, standing, lying down, walking, walking up stairs and walking downstairs). The proposed solution is an improvement of a one-verse-one SVM classifier with an ensemble of different learning methods each trained to discriminate a single activity against another. The improvement presented here doesn't only focus on accuracy but also potential embedded implementation capable of performing real-time classification with mobile data from the cloud. The presented one-versus-one approach, based on a linear kernel achieved 97.50 percent accuracy on a public dataset; second best to 98.57 percent reported in literature which uses a polynomial kernel.