Smooth and Consistent Video Anonymization Using Triangular Inpainting and Optical Flow

Surveillance cameras have been deployed extensively in big cities, such as London and Shanghai. To protect people’s privacy and avoid fully exposed, it is necessary to remove sensitive facial information in the surveillance footage. In this thesis, we study the anonymization of CCTV footage with face inpainting. In more detail, we employ deep neural networks to generate faces and replace the original faces in the video. Particularly, a masking method called triangular inpainting is employed to produce videos where the original faces are removed. Furthermore, we adopted an object detection method Optical Flow to ensure the smooth movement and transition of the computer generated face when masked on the original face. The thesis also tries to keep the age and gender of the generated faces to the original subjects as close as possible. To ensure that each human visible in videos is masked with a unique face throughout the whole video, we index the original face and the inpainted face for a one-to-one mapping. The designed system has been tested via extensive experiments. The results show that the human subjects are anonymized efficiently. The inpainted faces can also maintain the uniqueness and the smoothness in the video with the age and gender preserved