Radar based discrete and continuous activity recognition for assisted living

Shrestha, Aman (2021) Radar based discrete and continuous activity recognition for assisted living. PhD thesis, University of Glasgow.

Full text available as:
[img]
Preview
PDF
Download (12MB) | Preview

Abstract

In an era of digital transformation, there is an appetite for automating the monitoring process of motions and actions by individuals who are part of a society increasingly getting older on average. ”Activity recognition” is where sensors use motion information from participants who are wearing a wearable sensor or are in the field of view of a remote sensor which, coupled with machine learning algorithms, can automatically identify the movement or action the person is undertaking. Radar is a nascent sensor for this application, having been proposed in literature as an effective privacy-compliant sensor that can track movements of the body effectively. The methods of recording movements are separated into two types where ’Discrete’ movements provide an overview of a single activity within a fixed interval of time, while ’Continuous’ activities present sequences of activities performed in a series with variable duration and uncertain transitions, making these a challenging and yet much more realistic classification problem. In this thesis, first an overview of the technology of continuous wave (CW) and frequency modulated continuous wave (FMCW) radars and the machine learning algorithms and classification concepts is provided. Following this, state of the art for activity recognition with radar is presented and the key papers and significant works are discussed. The remaining chapters of this thesis discuss the research topics where contributions were made. This is commenced through analysing the effect of the physiology of the subject under test, to show that age can have an effect on the radar readings on the target. This is followed by porting existing radar recognition technologies and presenting novel use of radar based gait recognition to detect lameness in animals. Reverting to the human-centric application, improvements to activity recognition on humans and its accuracy was demonstrated by utilising features from different domains with feature selection and using different sensing technologies cooperatively. Finally, using a Bi-long short term memory (LSTM) based network, improved recognition of continuous activities and activity transitions without human-dependent feature extraction was demonstrated. Accuracy rate of 97% was achieved through sensor fusion and feature selection for discrete activities and for continuous activities, the Bi-LSTM achieved 92% accuracy with a sole radar sensor.

Item Type: Thesis (PhD)
Qualification Level: Doctoral
Subjects: T Technology > TA Engineering (General). Civil engineering (General)
Colleges/Schools: College of Science and Engineering > School of Engineering
Supervisor's Name: Le Kernec, Dr. Julien and Fioranelli, Dr. Francesco and Imran, Prof. Muhammad
Date of Award: 2021
Depositing User: Theses Team
Unique ID: glathesis:2021-82482
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 05 Oct 2021 14:57
Last Modified: 05 Oct 2021 14:57
Thesis DOI: 10.5525/gla.thesis.82482
URI: http://theses.gla.ac.uk/id/eprint/82482
Related URLs:

Actions (login required)

View Item View Item