What happens when machines can see?
The latest advances in Machine Learning enable a new class of intelligent video monitoring for commercial and corporate buildings. Surveillance video has historically been used only for after-the-fact forensic evidence or real-time viewing. If instead security cameras notify people of events that warrant attention, automate workflows, and make it fast to share information across organizations, then new levels of security, efficiency, and public safety are possible.
The tech industry is racing to make machines understand what’s happening in the real world. The unexpected success of Amazon Echo has accelerated R&D to change the way we interact with computers, and the potential of self-driving cars has accelerated R&D to change the way computers understand the real world. The key technical insight behind this innovation is Machine Learning — where deep layers of Neural Networks learn on their own by observing lots of training examples. The Machines learn without supervision. That’s what’s so significant about this new wave of intelligence, because so many kinds of problems can be solved so quickly.
Cloud computing has simultaneously connected this Machine Learning to existing cameras, DVRs, and NVRs. So old devices get a “new brain in the cloud” while new devices begin thinking on their own in collaboration with the cloud. As an example, Camio is a new smart video monitoring Software-as-a-Service that continuously filters, ranks and labels video events in real-time. Its interest-based video compression and natural language search are far beyond anything seen before in video surveillance. Neural Networks not only recognize what’s happening at any given moment but also — and more important — learn which events are interesting for each individual camera. Two very different problems — 1) labeling objects in the video and 2) learning which events are interesting — are solved by using these latest advances in Machine Learning.
Even before we reach the level of interactive ambient computing shown in the classic movie “2001: A Space Odyssey” (where astronauts plea “open the pod bay doors, HAL”), it’s already clear that adding intelligence to security cameras reduces operational costs. Even the mundane requirement to archive security video for long periods of time becomes cheaper and simpler when AI powers the storage decisions. By running Machine Learning on the local network — so that video streams are analyzed before uploading them to the cloud — Camio reduces storage by a factor of 10 by ranking the importance of each video event in real-time to vary its decisions about resolution, frame rate, storage, and bandwidth for each individual video event. The connection to the cloud then enables an unlimited amount of video to be stored for any length of time — without any need to manage on-site servers, disc failures, and storage upgrades. So one of the first implications of having machines that see is that they save us money.
Safety is often the primary motivation for video surveillance. But most security video is archived without review until after something bad happens. Machines are particularly good at being attentive and fast in noticing and notifying people. When machines can see, events in the real world programmatically trigger actions like dispatching security guards for [people approaching side alley 2am to 6am], or recording in a spreadsheet every time a person has entered a secure room. Plus, any forensic review takes seconds rather than days when machines have already indexed all the video — in real-time — for fast search for people, objects, colors, direction of movement, zones, and time. That makes it easy to see that the man-in-black at 7:49am is the same man-in-black in the side lot at 9:26am across two different cameras. Security video is encrypted for, and controlled by, each building owner. However, the video can be shared instantly (and selectively, on an opt-in basis) with neighboring buildings and government agencies to help investigate and respond to criminal and terrorist activity.
Computer Vision and Machine Learning make it practical to marshal a coordinated responses to security concerns — even across cameras in different buildings.
But the security problems with video surveillance have been particularly bad. The Mirai virus took down the Internet with a DDoS attack made possible by insecure video surveillance equipment on networks with open inbound ports for remote viewing and configuration. The infected equipment simply overwhelmed DNS servers with requests. And the infections could be much worse. With smart video monitoring, local networks are closed and secure; the intelligence to manage video and notifications using encrypted communication with the cloud means that there’s no need for open ports for inbound network requests. In Camio’s case, the Machine Learning operates on video that’s stored redundantly across multi-regional data centers that are at least 100 miles apart. So even in the case of natural disasters or large scale cyber-attacks, important video evidence remains maximally accessible.
Whether it’s called Artificial Intelligence, Cognitive Computing, Deep Learning, or Ambient Computing, the convergence of technologies enabling machines to understand the real world in real-time changes the way we secure and manage buildings. Now that cameras can learn, all those building entrances, loading docks, server rooms, reception desks, and garages can talk to us — and to the machines. Open the door, HAL!
This article originally appeared as the Realcomm Advisory on March 8, 2017.
