ARefine: Efficiently Storing Surveillance Videos via Accuracy-aware Refining Mechanism

Benefiting from the automatic video analysis technology, high-definition surveillance cameras have been increasingly deployed all over the city. The deployment of these cameras facilitate social management dramatically, but the huge amount of recorded video data puts tremendous pressure on the existing storage system. This paper proposes ARefine, an efficient accuracy-aware refining mechanism that adaptively compresses surveillance videos based on video content and the prediction accuracy of video analytics. ARefine first explores the potential content redundancy inherent in varied video slices from the perspective of video analysis. And then, ARefine designs a content-sensitive video refining method to condense video content dramatically without compromising the prediction accuracy. Experimental results show that ARefine significantly reduces the storage cost by 43.0%-75.0%, with a slight and controllable reduction in prediction accuracy (5%).