Dealing with “noise” pollution

Surveillance video analytics promise to automate much of the mind-numbing monitoring security staff must perform. The technology is starting to catch up with the hype, but the false positive rate is still about 65 per cent, according to Frost & Sullivan. Although it isn’t ready for prime time yet, video analytics technology can nevertheless deliver real value in some security scenarios today when it’s combined with human judgement.

January 8, 2008
By Rosie Lombardi

The technology has improved significantly in the past two years, says
Dilip Sarangan, analyst at Frost & Sullivan, a Menlo Park, CA-based
consultancy. “Last year, the false alarm rate was about 85 per cent, so
it’s coming down – but it’s not at a level that will drive mass
adoption,” he says.

Attitudes have also evolved as awareness about the benefits within the
limitations of the technology builds. “People assumed this could be
used to replace staff, but it doesn’t work that way,” says Sarangan.
While there are limits to the number of cameras a human can reasonably
monitor, there are also limits on a computer’s ability to discern real
threats from benign changes in the environment. As a consequence, video
analytics technology needs to deliver a high rate of alerts and alarms
for human decision-making. Nevertheless, the technology increases human
effectiveness by automating video monitoring to a certain degree, so
fewer staff can manage an increasing number of cameras.

How it works 

Video analytics is essentially intelligent motion detection, says
Sarangan. Dumb motion detection, which is fairly mature technology
that’s often free with many devices, simply measures the number of
pixels that change on screen. “Some systems issue alerts if, say 20 per
cent of the pixels change – but the issue is filtering out video
“noise” to increase accuracy,” says David Ngau, product manager at
Morristown, NJ-based Honeywell International Inc., noting systems need
to be programmed with complex algorithms to distinguish unimportant
pixel changes. The noise issue is a stumbling block the industry has
recently started to overcome, he says.

Gregory Dack, security analyst at the City of Ottawa says he explored
video analytics last year. “It worked fine indoors, but outside – snow,
waving branches, and deer set off false alarms,” he says. “There’s
promise in this area but vendors still have a way to go.”

Although the algorithms to filter out this type of noise are growing
more robust, a related issue is classifying an object once a
significant change in pixels is detected. “This depends on the camera
view: people two feet away can be picked up, but they’re smaller 10
feet away, so this becomes harder,” says Ngau. Vendors use different
methods to differentiate and classify shapes and objects, which may be
more suited for particular applications: indoors or outdoors,
discerning people or vehicles, and so on.

What systems can’t do today is provide specifics, says Dvir Doron,
vice-president at Denton, Tex-based Ioimage. “They can detect if an
intruder jumped over a fence but not a specific intruder – to do that,
you would need facial recognition and a database of characteristics,”
he says. “The industry is working on improving classification, but
today only generic stuff is available.”

The accuracy rate is higher if environmental factors are controlled,
says Steve Langford, director at Ottawa-based March Networks. Dedicated
cameras need to be placed in optimal locations with the right lighting
and other conditions to allow video analytics to deliver maximum value.

“Constraining environmental factors may be seen as a limitation, but
unless you do that you will get an unacceptable level of accuracy,”
says Langford. For example, the technology is more accurate in
detecting people in a known, limited area such as a lobby than an
airport. “To make it work, security staff must put some constraints in
place, so they must decide if there’s any ROI left if they do that.”

Other types of detection technologies are also being developed and
combined with analytics to increase accuracy in some high-end
applications, says Doron. Thermal cameras can detect vehicles or
intruders approaching in particular terrains or waterscapes. Human
radar that uses radio frequency (RF) signals to detect intruders is
also being developed.

Another promising area lies in pan-tilt-zoom (PTZ) cameras that work
with analytics, says Wes Fernley, a Brantford, Ont.-based consultant
and founder of videoanalytics.net. “These cameras are smart enough to
automatically zoom in and follow a particular motion once it’s
detected,” he says. “It’s still not reliable technology and is easily
confused in a busy parking lot during the day, but it could be helpful
late at night.”

There are many video analytics capabilities that can work in controlled
environments, says Sarangan. “Nothing is pure hype – but it’s not easy
getting the technology to work.”

Where it might work

Organizations that need to deal with security problems that can’t be
solved by traditional means are prime candidates for the technology,
says Ngau. For example, an airport with a 20-mile perimeter may not
have enough manpower to secure it all. But placing video
analytics-enabled cameras throughout its entire length would be
expensive and counter-productive due to the high rate of alerts.
Security managers need to decide which high-risk areas warrant extra
coverage. “If five miles are particularly risky, that’s where the
cameras should be placed,” he says. “We would never say these cameras
should be placed in every situation.”

Sarangan notes early adopters of the technology are big enterprises and
high-risk environments such as government, airports, and
transportation. There is also uptake in the retail sector, where
security needs are merging with operational concerns. “Since the
cameras are already there, analytics can make better use of the
information being collected for marketing or other purposes, and this
boosts the ROI.”

Langford points out there are two types of video analytics available
that can help create a stronger business case for its implementation in
retail and financial services organizations. “There’s real-time
analytics that detect a vehicle going the wrong way in a parking lot so
security staff can react quickly,” he explains. “But there’s another
historical class that monitors the environment over days or months,
such as people counting, traffic pattern analysis, service levels and
lengths of queues. A retailer, for example, might use people counting
to compare traffic around Christmas displays this year versus last
year. These are typically rendered as reports and not necessarily as
video data, and the point is to optimize the business. “

Large enterprises with big budgets aren’t the only organizations eyeing
the technology. Recent trends show uptake is starting to occur in small
and medium-sized (SMB) businesses as the technology grows more
accessible, says Doron. “Up to 2005, it’s been associated with the
high-end market but in the past two years, we’ve seen a shift to
mid-range businesses, and the focus there is saving money, not
terrorism.”  He cites a recent example of a BMW car dealership that
replaced dozens of passive infrared sensors with just four video
analytics-enabled cameras to prevent vandalism damages of about $1
million annually.

In Canada, many SMB businesses are adopting the technology, says Karen
Letain, president of Ottawa-based CMI Inc., a systems integrator and
the Canadian reseller for Ioimage. “Business owners looking to secure
their premises without having someone constantly on watch, and even
museums are using it,” she says. Many major vendors’ solutions are
designed for the high-end market and are built around proprietary
cameras that compel organizations to rip and replace existing
infrastructure, but more accessible solutions are emerging for the SMB
sector that will work with existing equipment, she adds. 

There are two fundamental architectures in video analytics design, each
with an associated set of pros and cons, explains Fernley. One category
is software-driven and typically runs in a centralized fashion on the
network via video servers. The other approach embeds analytics in
hardware such as cameras, DVRs and encoders at the edge of the network.

The centralized approach is more suited for huge installations with
lots of cameras, says Sarangan. “It becomes expensive to run analytics
at the edge in this scenario,” he says. And software-driven analytics
can eat up bandwidth, storage and create other network issues.

In the hardware-driven, edge approach, all the analytics are handled at
the device level, says Fernley. “So all the video isn’t going to one
source on the network to get filtered. Firmware only sends video to the
recording station when it detects something significant instead of
sending a constant stream.”

Fernley believes the hardware-driven approach has some advantages.
“From my testing, software is great but you can only get a few cameras
running on one PC typically,” he says. “If you’re running analytics on
multiple cameras, it’s probably best to get products with built-in
analytics in the firmware.”

As the IP-based video market grows and matures, he believes analytics
will be become commonplace and will likely be bundled into cameras as
an integral part of their operations. He also believes new businesses
will spring up to offer remote hosting, monitoring and notification
services. “In the future, people will hook up cameras to the Web, and
that’s it – they’ll outsource this piece of physical security,” he
says. “With CCTV cameras, you had to pay someone to monitor 10 cameras,
so it was more cost-effective to just hire someone to do it in-house.
But with analytics, costs are lower and you don’t need staff watching
so many cameras.”

Rosie Lombardi is a Toronto-based freelance writer.