Visual privacy protection using false colors

2017
Çiftçi, Serdar
Visual privacy protection (VPP) in video surveillance is an important problem, which is likely to be even more important with the rapid expansion of video surveillance. Although many VPP algorithms exist, none of them simultaneously meets all the desired characteristics of a good privacy protection algorithm. The chief limitation of existing VPP methods is that they require regions of interest (ROI) to be automatically detected or manually marked, both of which are difficult to achieve and prone to errors. In this thesis, a false color based visual privacy protection method is proposed, which protects visual privacy while maintaining intelligibility. The proposed solution alleviates the ROI requirement as the solution can be applied on full frames without severely hampering intelligibility. The proposed method also aims to strike a balance among other desired characteristics of a good VPP algorithm. As another chief benefit, the proposed method is reversible making it possible to recover the unprotected information in case of necessity. The effectiveness of the proposed method is investigated through a large number of objective and subjective experiments using existing low dynamic range (LDR), high dynamic range (HDR) image datasets, as well as novel video dataset created as part of this work.