Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.816927
Title: Privacy aware human action recognition : an exploration of temporal salience modelling and neuromorphic vision sensing
Author: Al-Obaidi, Salah Mahdi
ISNI:       0000 0004 9356 4576
Awarding Body: University of Sheffield
Current Institution: University of Sheffield
Date of Award: 2020
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
Solving the issue of privacy in the application of vision-based home monitoring has emerged as a significant demand. The state-of-the-art studies contain advanced privacy protection by filtering/covering the most sensitive content, which is the identity in this scenario. However, going beyond privacy remains a challenge for the machine to explore the obfuscated data, i.e., utility. Thanks for the usefulness of exploring the human visual system to solve the problem of visual data. Nowadays, a high level of visual abstraction can be obtained from the visual scene by constructing saliency maps that highlight the most useful content in the scene and attenuate others. One way of maintaining privacy with keeping useful information about the action is by discovering the most significant region and removing the redundancy. Another solution to address the privacy is motivated by the new visual sensor technology, i.e., neuromorphic vision sensor. In this thesis, we first introduce a novel method for vision-based privacy preservation. Particularly, we propose a new temporal salience-based anonymisation method to preserve privacy with maintaining the usefulness of the anonymity domain-based data. This anonymisation method has achieved a high level of privacy compared to the current work. The second contribution involves the development of a new descriptor for human action recognition (HAR) based on exploring the anonymity domain of the temporal salience method. The proposed descriptor tests the utility of the anonymised data without referring to RGB intensities of the original data. The extracted features using our proposed descriptor have shown an improvement with accuracies of the human actions, outperforming the existing methods. The proposed method has shown improvements by 3.04%, 3.14%, 0.83%, 3.67%, and 16.71% for DHA, KTH, UIUC1, UCF sports, and HMDB51 datasets, respectively, compared to state-of-the-art methods. The third contribution focuses on proposing a new method to deal with the new neuromorphic vision domain, which has come up to the application, since the issue of privacy has been already solved by the sensor itself. The output of this new domain is exploited by further exploring the local and global details of the log intensity changes. The empirical evaluation shows that exploring the neuromorphic domain provides useful details that have demonstrated increasing accuracy rates for E-KTH, E-UCF11 and E-HMDB5 by 0.54%, 19.42% and 25.61%, respectively.
Supervisor: Abhayaratne, Charith Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.816927  DOI: Not available
Share: