Privacy Preserving Object Detection in COSMOS Smart Intersection
Video
Team Information
Team Members
Alex Angus, MS Candidate, Columbia Engineering
Faculty Advisor: Zoran Kostić, Professor of Professional Practice, Department of Electrical Engineering, Columbia Engineering
Abstract
Project COSMOS relies heavily on large image datasets that are gathered from street-level cameras. In the process of collecting real-time images and videos of public spaces, cameras also inadvertently capture sensitive information such as faces and license plates. To avoid the compromise of large amounts of private information in the follow-up research involving our image datasets, we generate a pipeline to systematically blur the faces of pedestrians and the license plates of vehicles in street level intersection videos. We train various YOLOv4 object detection models, using the Darknet framework, for the detection of pedestrians and vehicles as well as for the detection of faces and license plates. We train the models with our own custom intersection video dataset annotated Summer 2021. Ultimately, we are able to automatically blur 99% of visible faces and license plates in any given 1st floor intersection video.
Team Lead Contact
Alex Angus: ala2197@columbia.edu
