Towards Robust and Deployable Gesture and Activity Recognisers

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Sähkötekniikan korkeakoulu | Master's thesis

Date

2021-12-13

Department

Major/Subject

Autonomous Systems

Mcode

ELEC3055

Degree programme

Master's Programme in ICT Innovation

Language

en

Pages

62+2

Series

Abstract

Smartphones and wearables have become an extension of one's self, with gestures providing quick access to command execution, and activity tracking helping users log their daily life. Recent research in gesture recognition points towards common events like a user re-wearing or readjusting their smartwatch deteriorate recognition accuracy significantly. Further, the available state-of-the-art deep learning models for gesture or activity recognition are too large and computationally heavy to be deployed and run continuously in the background. This problem of engineering robust yet deployable gesture recognisers for use in wearables is open-ended. This thesis provides a review of known approaches in machine learning and human activity recognition (HAR) for addressing model robustness. This thesis also proposes variations of convolution based models for use with raw or spectrogram sensor data. Finally, a cross-validation based evaluation approach for quantifying individual and situational-variabilities is used to demonstrate that with an application-oriented design, models can be made two orders of magnitude smaller while improving on both recognition accuracy and robustness.

Description

Supervisor

Oulasvirta, Antti

Thesis advisor

Aslan, Ilhan
Kannala, Juho

Keywords

gesture recognition, deep learning, wearables, robustness

Other note

Citation