AbstractsEngineering

A machine learning framework for automatic human activity classification from wearable sensors

by Edmond Mitchell




Institution: Dublin City University
Department: School of Electronic Engineering; Dublin City University. INSIGHT Centre for Data Analytics
Year: 2015
Keywords: Machine learning; Signal processing; Wearable sensors
Record ID: 1182359
Full text PDF: http://doras.dcu.ie/20336/


Abstract

Wearable sensors are becoming increasingly common and they permit the capture of physiological data during exercise, recuperation and everyday activities. This work investigated and advanced the current state-of-the-art in machine learning technology for the automatic classification of captured physiological data from wearable sensors. The overall goal of the work presented here is to research and investigate every aspect of the technology and methods involved in this field and to create a framework of technology that can be utilised on low-cost platforms across a wide range of activities. Both rudimentary and advanced techniques were compared, including those that allowed for both real-time processing on an android platform and highly accurate postprocessing on a desktop computer. State-of-the-art feature extraction methods such as Fourier and Wavelet analysis were also researched to ascertain how well they could extract discriminative physiological information. Various classifiers were investigated in terms of their ability to work with different feature extraction methods. Consequently, complex classification fusion models were created to increase the overall accuracy of the activity recognition process. Genetic algorithms were also employed to optimise classifier parameter selection in the multidimensional search space. Large annotated sporting activity datasets were created for a range of sports that allowed different classification models to be compared. This allowed for a machine learning framework to be constructed that could potentially create accurate models when applied to any unknown dataset. This framework was also successfully applied to medical and everyday-activity datasets confirming that the approach could be deployed in different application settings.