Privacy-Preserving of Human Identification in CCTV Data using a Novel Deep Learning-Based Method

With the rapid development of Information Tech-nologies, the number of surveillance cameras has increased, leading to a decrease in the rate of violence and crimes. Despite the advantages, it has some drawbacks, such as the lack of privacy-preserving in human identification when video data is shared. This paper focuses on the privacy-preserving issue and proposes a novel deep learning-based solution. The proposed approach uses the recent state-of-the-art models such as RetinaFace and TinaFace for human face detection from the input videos and other Computer Vision tools such as OpenCV for framing the input video and connecting them again for restoring the initial form of the video in the output. Moreover, computer vision tools such as Blur and Gaussian Blurring are used for anonymizing faces. Our proposed approach allows the closed-circuit television (CCTV) data to be shared for public use, where human identification is perfectly preserved. Experimental results show the effectiveness of our proposed method by outperforming the state-of-the-art methods in constrained conditions. Furthermore, we have created a face dataset from the input CCTV videos, where the face detection tools have failed. The created dataset is annotated with the five face landmarks and can be used for the face detection task.