Pedestrian detection and segmentation in videos captured by cameras exhibiting large, uncontrolled motion

点击翻译按钮获取中文摘要

Video surveillance is increasingly being used to record police-citizenry interactions to increase officer safety and to curb police brutality. Many police services, both in Canada and internationally, have started pilot projects where officers will be equipped with body worn cameras that will record their surroundings as they respond to events. Body worn cameras raise serious privacy concerns. Police departments, for example, often have to release the videos in response to a freedom of information request. The police departments are required by law to hide or mask the identities of non-actors before these videos are released. This is to preserve the privacy of innocent individuals seen in the videos. Existing video redaction or masking tools are time consuming and laborious. These are designed to deal with videos captured from stationary cameras. Body worn cameras undergo large, uncontrolled motions, and we need new tools and theory to develop automated tools for detecting and masking individuals seen in videos captured by body worn cameras. In this project, Dr. Qureshi's Visual Computing Lab at the University of Ontario Institute of Technology will work with Xiris to investigate pedestrian detection and segmentation methods suitable for videos exhibiting exaggerated motions. The theory and methods developed as a part of this project will not only enable us to develop automated video redaction tools, these also constitute the core functionality of privacy-preserving video surveillance systems of the future.