Deep Learning based Human Detection in Privacy-Preserved Surveillance Videos

Visual surveillance systems have been improving rapidly over the recent past, becoming more capable and pervasive with incorporation of artificial intelligence. At the same time such surveillance systems are exposing the public to new privacy and security threats. There have been an increasing number of reports of blatant abuse of surveillance technologies. To counteract this, data privacy regulations (e.g. GDPR in Europe) have provided guidelines for data collection and data processing. However, there is still a need for a private and secure method of model training for advanced machine learning and deep learning algorithms. To this end, in this paper we propose a privacy-preserved method for visual surveillance. We first develop a dataset of privacy preserved videos. The data in these videos is masked using Gaussian Mixture Model (GMM) and selective encryption. We then train high-performance object detection models on the generated dataset. The proposed method utilizes state-of-art object detection deep learning models (viz. YOLOv4 and YOLOv5) to perform human/object detection in masked videos. The results are encouraging, and are pointers to the viability of the use of modern day deep learning models for object detection in privacy-preserved videos.