Page 34 - Realcomm13-Final-LR

BUSINESS SOLUTIONS

What Happens When Machines Can See?

Carter Maslan
CEO
Camio

he latest advances in Machine Learning enable a new astronauts plead, “open the pod bay doors, HAL”), it’s already clear
class of intelligent video monitoring for commercial and that adding intelligence to security cameras reduces operational
T corporate buildings. Surveillance video has historically costs. Even the mundane requirement to archive security video
been used only for after-the-fact forensic evidence or real-time for long periods of time becomes cheaper and simpler when
viewing. If instead security cameras AI powers the storage decisions. By
notify people of events that warrant at- running Machine Learning on the local
tention, automate workflows and make network—so that video streams are
it fast to share information across orga- analyzed before uploading them to the
nizations, then new levels of security, cloud—storage is reduced by a factor
efficiency, and public safety are possible. of 10 by ranking the importance of
The tech industry is racing to make each video event in real-time to vary
machines understand what’s happen- its decisions about resolution, frame
ing in the real world. The unexpected rate, storage, and bandwidth for each
success of Amazon Echo has accelerated individual video event. The connection
R&D to change the way we interact to the cloud then enables an unlimited
with computers, and the potential of “Even the mundane requirement amount of video to be stored for any
self-driving cars has accelerated R&D to length of time without any need to
change the way computers understand to archive security video for long manage on-site servers, disc failures, and
the real world. The key technical insight periods of time becomes cheaper storage upgrades. So, one of the first
behind this innovation is Machine and simpler when AI powers the implications of having machines that
Learning—where deep layers of Neural see is that they save us money.
Networks learn on their own by ob- storage decisions.” Safety is often the primary motivation
serving lots of training examples. The for video surveillance. But most security
Machines learn without supervision. That is what’s so significant video is archived without review until after something bad hap-
about this new wave of intelligence, because so many kinds of pens. Machines are particularly good at being attentive and fast
problems can be solved so quickly. in noticing and notifying people. When machines can see, events
Cloud computing has simultaneously connected this Machine in the real world programmatically trigger actions like dispatching
Learning to existing cameras, DVRs, and NVRs. Old devices get security guards for potential security breaches, or recording on a
a “new brain in the cloud” while new devices begin thinking on spreadsheet every time a person has entered a secure room. Plus,
their own in collaboration with the cloud. New smart video mon- any forensic review takes seconds rather than days when machines
itoring Software-as-a-Service options continuously filter, rank have already indexed all the video—in real-time—for fast search
and label video events in real-time. Their interest-based video for people, objects, colors, direction of movement, zones, and time.
compression and natural language search are far beyond any- That makes it easy to see that the man-in-black at 7:49 a.m. is the
thing seen before in video surveillance. Neural Networks not only same man-in-black in the side lot at 9:26 a.m. across two different
recognize what’s happening at any given moment but also—and cameras. Security video is encrypted for, and controlled by, each
more importantly—learn which events are interesting for each building owner. However, the video can be shared instantly (and
individual camera. Two very different problems—1) labeling ob- selectively, on an opt-in basis) with neighboring buildings and
jects in the video, and 2) learning which events are interesting— government agencies to help investigate and respond to criminal
are solved by using these latest advances in Machine Learning. and terrorist activity. Computer Vision and Machine Learning make
Even before we reach the level of interactive ambient com- it practical to marshal a coordinated response to security con-
puting shown in the classic movie—A Space Odyssey (where cerns—even across cameras in different buildings.

32 Realcomm