Privacy protection preserving the utility of visual surveillance

Due to some tragic events such as crime, bank robberies and terrorist attacks, an unparalleled surge in video surveillance cameras has occurred in recent years. In consequence, our daily life is overseen everywhere (e.g. on the street, in stations, in shops and in the workplace). For example, on average, people living in London can be caught on cameras more than 300 times a day. At the same time, automatic processing technology and quality of sensors have advanced significantly, which has even enabled automatic detection, tracking and identification of individuals. With the proliferation of video surveillance systems and the progress in automatic recognition, privacy protection is now becoming a significant concern. Video surveillance is intrusive because it allows the observation of certain information that is considered as private (i.e., identity or some characteristics such as age, race, gender). Nowadays, some processing technologies are able to limit the intrusiveness of surveillance by automatically hiding private information. Indeed, a number of studies have recently begun to focus on the protection of personal privacy, especially the identity, but they are seldom taken care of crucial criteria required for the surveillance (e.g. reversibility, the utility of the surveillance, compliant with the compression standards). For a long time, privacy and safety of people were viewed as mutually exclusive factors. One could not preserve the utility of visual surveillance and protect the privacy at the same time, both are very significant, though. While privacy is essential to freedom, surveillance is a major actor for our safety. The purpose of this thesis is to find technological solutions to the issue of privacy protection of individuals while preserving the utility of the surveillance (i.e. leaving the understanding of the scene and enabling the re-identification of a person in case of an incident). Indeed, an ideal surveillance system should protect the personal privacy of individuals while still providing a high level of the utility of visual surveillance. Existing methods have issues to manage this trade off, usually, the increase in utility of surveillance brings about a significant decrease in personal privacy. To address privacy concerns regarding digital image or video surveillance cameras, we propose one main concept: using the most important information to preserve the ability to recognize actions while protecting individual identities by encrypting and hiding their original information in the least important information of the data. This fulfils a better trade off between privacy and safety of people compared to the existing methods in that domain. We integrate this idea, first, in the spatial/pixel domain, and, then, in the transform/frequency domain to be compliant with the classical compression standards such as JPEG and H.264/AVC. Indeed, nowadays almost all images and videos are compressed. Moreover, the data that we encrypt is reversible with a secret key being available for authorized people only.