Protection of visual privacy in videos acquired with RGB cameras for active and assisted living applications

Active and assisted living technologies are much needed, but some aspects of them cause user rejection due to concerns on privacy. This is even more concerning to users when visual information is used, processed, and transmitted. To respond to these concerns, and maximise user acceptance, visual privacy protection measures have to be put in place. In the past, human detection and object segmentation in video were constrained by technological limitations, and could only run with specific hardware and sensors. This paper introduces a proposal for an RGB-only based visual privacy preservation filter, which capitalises on ‘deep learning’-based segmentation and pose detectors. A background update scheme is presented, which limits leakage of sensitive information when detection fails. Dilation of the mask can further prevent information leakage, but a trade-off is necessary to correctly update background information. This is achieved via a specific study which is also presented. A comparative study is performed to determine the best configuration for privacy preservation. Results show that union of dilated masks from different deep networks achieves the best overall result.