According to a January 2013 report published by the National Public Radio, the Chinese government has installed more than 20 million cameras across the country. Similarly, a recent BBC report stated, there is one camera for every 14 people in London. Other large cities including New York, Paris, and Tokyo are also deploying cameras in large numbers so it is reasonable to claim that networked cameras are everywhere. These cameras are deployed for a wide variety of commercial and security reasons. But the problem is with so many incoming video streams these systems are costly to operate (think on-going network, storage, and computation costs) and they are not smart. Digital video is stored and analyzed after an unexpected event occurs. We want to change this.
In discussions with local government officials we have found that their idea of smart cities includes using cameras to make their streets safer for cars, bicyclist and pedestrians. They want to do this by estimating their speeds and trajectory to avoid potential collisions and accidents. Law enforcement and homeland security officers want to use these cameras to help with “amber alerts”, detecting suspicious behavior and raising alarms to help prevent potential crimes. With the rising popularity of Internet of Things (IoT) surveillance videos from factory floors, traffic and police, and retail shops, cameras represent the largest and most challenging of the “things” in terms of data volume, (vision) processing algorithms, response latencies, and security sensitivities. Consequently, we believe that large-scale video analytics is a grand challenge for the research community representing an important and exciting frontier for big data systems. Furthermore, anecdotally we have heard that the overall market value for video analytic services is in billions of dollars.
Unlike text or numeric processing, videos require high bandwidth (e.g., up to 5 Mbps for HD streams), need fast CPUs and GPUs, richer query semantics, and tight security guarantees. Our goal is to build and deploy a highly efficient distributed video analytic (DVA) system. DVA will lead to new research on (1) building a scalable, reliable and secure systems framework for capturing and processing video data from geographically distributed cameras; (2) efficient computer vision algorithms for detecting objects, performing analytics and issuing alerts on streaming video; and (3) efficient monitoring and management of computational and storage resources over a hybrid cloud computing infra-structure by reducing data movement, balancing loads over multiple cloud instances, and enhancing data-level parallelism. Our proposed DVA infrastructure will handle both real-time streaming and archival data in a uniform manner, so that the same analysis programs can be reused. We will carefully build a system that allows new vision algorithms, analytics, and systems capabilities to be integrated easily by the research community.
We are engaging with the City of Bellevue to use their existing city-wide cameras for traffic safety planning (link). We believe video analytics is a rich business opportunity for Azure. By enabling a number of different surveillance scenarios, both in industry and in government, DVA can position Azure to go after this opportunity. Consequently, we intend to deploy DVA as an Azure service as well as a distributed solution directly installed by customers.