Respectful Cameras: Privacy Enforcement through Marker Recognition and Tracking

As more large-scale camera systems are deployed around the world, the need for video privacy is becoming increasingly vital. State of the art face and people tracking systems are not yet sufficiently robust for this domain, due to changing lighting conditions, occlusions, and the need to be realtime. This has motivated us to instead track visual markers worn by individuals who wish to have their privacy retained. We build a color-tracker which projects the typical RGB color space into a 9 dimensional space of RGB, HSV, and LAB. We then use AdaBoost, a statistical classifier, to learn a model for classifying color which corresponds to the marker versus color that does not. For AdaBoost, we use decision stump weak-learners. We use a connected component method for clustering pixels into groups that correspond to different markers. We present a new Probabilistic AdaBoost method, which provides a probability that a pixel corresponds to a marker, as opposed to the conventional version which has no notion of confidence in the classification. We then use Particle Filtering to track the markers over time. We use the Probabilistic AdaBoost for our Observation Model, and a transition model which explicitly models velocity and orientation of the markers. We present preliminary results from recorded data illustrating the performance of our system under changing lighting conditions, nearby but never crossing people, multiple people which occlude one another, and similarly colored objects as markers in the scene.