De-identifying people in videos using neural art

We propose a computer vision-based de-identification pipeline that enables automated segmentation of humans in videos and effective protection of their identities. Due to the ubiquity of video surveillance, many jurisdictions implement strict regulations for the protection of personal data in publicly collected video sequences, requiring the data to be de-identified. However, soft biometric and non-biometric features like clothing, hair color, personal items, skin marks, etc., are often overlooked in the process. Assuming a surveillance scenario, we combine GMM-based background subtraction with an improved version of the GrabCut algorithm to find and segment pedestrians. We use the responses of a deep neural network to de-identify soft and non-biometric features through style mixing with images of other pedestrians. Our method produces de-identified versions of the input frames while preserving the naturalness and utility of the de-identified data.