Adding Privacy Constraints to Video-Based Applications

Remote accessibility to video cameras, face recognition software, and searchable image and video databases enable automated video surveillance. The same technologies and features enabling advanced functionalities facilitate as well the misuse of a surveillance system and can lead to the infringement of the privacy of an individual. Examples of misuse are voyeurism and the unauthorized collection of data on activities or behaviours of an individual. To overcome this problem, we propose a system architecture for privacypreserving video-based applications which uses image and video segmentation. Image and video segmentation are inherently ill-posed problems. In this sense, no unique segmentation of a scene exists: a priori knowledge about the nature of the problem is required in most situations. To this end, we propose a segmentation framework that embeds such a priori knowledge in the form of privacy constraints. The proposed framework is composed of two levels of semantic segmentation and subsequent description. The first level extracts video objects whereas the second extracts parts of a scene that can identify an individual, such as faces and license plates. The rationale behind this framework is that the data that one needs to hide for preserving the privacy of individuals and the data one needs to process to understand a scene may be divided into two non-overlapping sets. These two sets can be stored and processed separately thus enabling the use of automated surveillance in public places while protecting privacy.