Towards biometric footstep recognition: Dimensionality reduction-inspired approaches to pressure-based gait recognition
Loading...
Date
2024-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of New Brunswick
Abstract
This thesis explores the use of a family of principal component analysis (PCA)-based approaches to extract discriminative features for pressure-based gait recognition. Two datasets are used; 1) an openly available CASIA-D dataset comprised of barefoot samples from 88 subjects to examine the performance of various pre-features, dimensionality reduction, and deep learning-based approaches, and 2) a custom 7-subject UNB-collected dataset of footprints and non-footprints to explore the feasibility of a PCA-based footprint detection system. Even with relatively few training samples per participant, strong performances were found for a variety of PCA-based approaches, especially when combined with additional feature selection approaches. It was found that the deep-learned PCANet+ combined with Minimum Redundancy Maximum Relevance was the best performing combination for a biometric verification system, achieving a 96.21% accuracy. Furthermore, PCA-based approaches inspired by the concept of eigenfaces were found to effectively discriminate incomplete and non-foot images.