In the era of smart and connected communities, a video surveillance system, which usually involves tens and thousands of video cameras, has increasingly become a prominent component for the public safety. In the current practice, when the video surveillance system has a failure, the operation and maintenance team usually spends a lot of time to identify and locate the failure, which cannot
... [Show full abstract] guarantee real-time in a large-scale video surveillance system. Meanwhile, the video data with a failure wastes amount of storage space in the cloud. The emergence of edge computing is very promising in the preprocessing for source video data at an edge camera, and video surveillance systems are one of the popular applications for edge computing. In this paper, we propose VU, a V ideo U sefulness model for large-scale video surveillance systems, and explore its application, such as early failure detection and storage saving. The VU model evaluates the usefulness of video data in a real-time fashion and notifies failures to end-users on the fly.
This paper has three contributions: (1) a comprehensive video usefulness model has been proposed. To the best of our knowledge, this is the first work aiming to quality the video usefulness in a real application; (2) real-time failure detection algorithms based on edge computing and cloud computing are proposed to efficiently improve the mean time to repair (i.e., MTTR); (3) effective storage and bandwidth saving schemes for large-scale video surveillance systems are proposed and implemented.
Results from a university-wide surveillance system consisting of 2,960 cameras show that failures of video data in different domains are accurately detected by VU model. MTTR is largely shortened by the fast detection algorithm in real time. The video data with the worst degree of VU is mostly discarded to reduce overload in the network and save storage space in the cloud.