Balancing Privacy and Utility with Pattern Based Activity Detection

Abstract

The diffusion of surveillance cameras often leads to conflicts between utility, that is, the benefits of preserving the information the camera records, and privacy, that is, the ability for the people being observed to conceal information they want to protect. For example, a camera monitoring an office kitchen may be useful in identifying a food thief, but might unintentionally reveal the PIN someone enters on a mobile phone. We design a video processing system that detects private activities in surveillance video and filters them out of the recording with minimal disruption of video quality. At the core of our system is the light-weight computation of a fixed-size feature that describes the spatio-temporal aspects of human activities that extend over variable amounts of time and space. Converting events of variable length and extent to a fixed-size descriptor makes it possible to use off-the-shelf classifiers to recognize and localize activities to be protected from recording. Comparisons of our descriptor with several alternatives show improved performance with less computation. We contribute two new video datasets recorded with a kitchen security camera, and we carry out a pilot user study to show that PIN theft is a valid concern.