Activity Analysis in Privacy-Protected Video

The conflict between security and privacy has become a growing concern for the public, especially with the recent employment of large-scale video analysis technologies for surveillance purposes. Although researchers have developed many obscuring algorithms to protect the identities of subjects in the surveillance videos, they have rarely addressed the problem of how we can protect a subject’s privacy while preserving enough visibility to allow both human observers and machines to analyze the activity. In this paper, we aim to address problems of both activity analysis and privacy protection in video. To minimize the possibility of revealing identities of concerned subjects in video records, we propose an approach using a novel representation called the edge motion history (EMH). We show that the EMH can generate privacy-protected video that allows a human to analyze activities of subjects in video without using appearance information of the subjects and environment. We demonstrate that the EMH generated video can also be used for automatic action recognition/classification algorithms. Our experimental results in the KTH human action collection yield comparable accuracy to several state-of-the-art action recognition approaches. Besides benefits of activity analysis in privacy-protected video, the EMH can greatly reduce amounts of video data needed to be stored or transmitted over a network.