Segmentation-guided privacy preservation in visual surveillance monitoring

Abstract

[en] Video surveillance has become a necessity to ensure safety and security. Today, with the advancement of technology, video surveillance has become more accessible and widely available. Furthermore, it can be useful in an enormous amount of applications and situations. For instance, it can be useful in ensuring public safety by preventing vandalism, robbery, and shoplifting. The same applies to more intimate situations, like home monitoring to detect unusual behavior of residents or in similar situations like hospitals and assisted living facilities. Thus, cameras are installed in public places like malls, metro stations, and on-roads for traffic control, as well as in sensitive settings like hospitals, embassies, and private homes. Video surveillance has always been as- sociated with the loss of privacy. Therefore, we developed a real-time visualization of privacy-protected video surveillance data by applying a segmentation mask to protect privacy while still being able to identify existing risk behaviors. This replaces existing privacy safeguards such as blanking, masking, pixelation, blurring, and scrambling. As we want to protect human personal data that are visual such as appearance, physical information, clothing, skin, eye and hair color, and facial gestures. Our main aim of this work is to analyze and compare the most successful deep-learning-based state-of-the-art approaches for semantic segmentation. In this study, we perform an efficiency-accuracy comparison to determine which segmentation methods yield accurate segmentation results while performing at the speed and execution required for real-life application scenarios. Furthermore, we also provide a modified dataset made from a combination of three existing datasets, COCO_stuff164K, PASCAL VOC 2012, and ADE20K, to make our comparison fair and generate privacyprotecting human segmentation masks.