How Intelligent Are Analytics
SOMETIMES, EXPECTATIONS ARE SET SO high that they become impossible to meet. Lance Armstrong can’t win every Tour de France. Not every sequel is as good as the original movie. And intelligent video analytics—which are algorithms that run on computers and analyze video to pinpoint behavior, patterns, and other types of data—can only achieve so much.
Analytics first made a splash on the security scene in the years following the 9-11 terror attacks as a way to help companies spot significant events that might otherwise go unnoticed amid hundreds of cameras and thousands of hours of surveillance video. Their various programs now aim to serve as everything from perimeter tripwires to warnings of objects left behind to crowd control systems.
The technology holds great promise, but the video analytics market has not grown the way analysts expected, says IMS Research’s Niall Jenkins. IMS now projects that there will be 450,000 channels of analytics embedded in cameras, encoders, DVRs, and NVRs by 2013, which is significantly lower than projections from three years ago. A channel is any camera or encoder that has analytics performed on its data.
John Honovich of the Web site IP Video Market Info points out that “while the total number of cameras sold each year is debated, it seems to be at least 10 million and perhaps triple that amount. Even assuming no growth in annual camera sales, IMS’s projection would mean that, at best, 5 percent of cameras would be using video analytics in 2013.”
Why so few? Many industry experts point to excessive hype over the capabilities of intelligent video as a factor behind the revised growth expectations and lowering of public confidence in the technology. “The economy has impacted everyone to a certain degree, but video analytics has been hurt more significantly than other segments because of its overstated claims,” says Honovich.
It’s a phenomenon that has been seen before, according to Marc Holtenhoff, CEO of Aimetis Corp., a video management and analytics company. With new technology, early adopters take it on, people get excited, then they get disillusioned when it does not live up to the hype.
The good news is that the industry appears to have learned its lesson. An effort is underway on the part of analytics companies and other manufacturers and integrators to provide realistic expectations to end users for what these tools can truly do and how they can best assist in security and other business operations. Jenkins has seen this play out at the video-analytics conferences that IMS holds. There “has been a shift towards a more realistic selling of the technology” on the part of the manufacturers, he says. For example, John Whiteman, president, Americas, for analytics company ioimage LLC, (recently acquired by DvTel), says that the company conducts continuing education and certification of integrators and business partners to help “reset” industry expectations and ensure that the analytics are being used appropriately.
Bob Banerjee, product marketing manager of IP video products at Bosch Security Systems, Inc., echoes Jenkins’s view of the changes occurring in the industry. A few years back, he notes, there was an emphasis on functionality of the analytics, such as whether a company was offering trip wire or left-behind object detection. That has changed.
“The focus is on trying to combat the bad reputation that video analytics has with people that deployed early versions where they focused on functionality rather than ease of use and reliability,” he says.
Manufacturers are also trying to better educate users about the proper implementation of analytics. Jacob Loghry, a systems engineer with Adesta LLC, says that one of the mistakes end users were making with analytics earlier in the decade was that they expected them to be a blanket security tool, a catch-all.
Customers didn’t understand that analytics might not be useful in all settings and that simply implementing them throughout one’s security system would not be an effective use of resources. Loghry sees that changing now.
For example, if the goal is to ensure that there are no loiterers near a building, analytics efforts might be concentrated at the back or sides, where the problem occurs, rather than all over, such as in the front, where people legitimately wait and gather and where it would be harder for the analytics to avoid false alarms.
It’s been a struggle to drill this point home with end users, says Loghry. But he stresses the importance of setting realistic and specific goals for the use of the technology within a security system, and he emphasizes the importance of having a security policy to accompany the system.
Analytics at the Edge
One major development in analytics over the last few years is the move to put the software at the “edge,” or on the camera or encoder in the field, rather than back on the server at the central station. There are several benefits to doing analytics at the edge, according to experts. One is that it frees up bandwidth because the video is analyzed before being sent back to the server. Another is that the image being analyzed will be uncompressed, so it will be a much higher quality. The better the image quality, the more information the analytic algorithm has to work with and the more likely that the analytics will be successful. It would also require less processing power on the back end.
According to Fredrik Nilsson, general manager of Axis Communications, “One issue that video analytics had early on is that if you wanted to have video analytics, you needed to have an additional server. And that additional server needs to sit in the server room. And in the server room, you already have 20 servers, and all of a sudden, you have too much heat and too many servers, and it was too much to manage and too expensive. So, we thought that putting the processing power out close on the camera would mean that you have a very scalable system…. You didn’t have to add any servers.”
Still, the growth in this area has been modest as well. Honovich says few of the largest camera manufacturers are offering edge analytics embedded in cameras. “There’re many reasons: one is an issue of performance—will it even work well enough? And two, it substantially increases the cost of the camera. So I think the combination of those two things is going to drag in terms of wide scale adoption,” he says.
But many of the industry experts interviewed for this article cited edge analytics as an important trend. Carolyn C. Ramsey, director of program management at Honeywell, says it’s “important today but will be very important in about three years when the IP market is much more saturated, and it’s more an expected way of doing business.”
Putting the functionality at the edge may not work for all analytics, however. Some types might require a level of processing power found only back on the server, says Perry Levine, senior director of business development for security products at Siemens Industry, Inc., who cites geospatial analysis and classification as an example of an analytic that might have to stay on a central processor for now. Additionally, Banerjee points out that currently many of the good analytics come from companies whose software must still be run on a central server.
The overall ability to use analytics on different platforms, such as at the cameras, encoder, or DVR, is quite helpful to end users, says Ramsey. She cites the utilities sector as an example. “Some of our utilities customers have giant stations which are very well lit, have a lot of available power, and they have a lot of space, and they have space that’s clean, so they could use analytics on just about any platform. But then they may have small substations where they don’t have a clean space. The equipment might have to be outdoors, it might have to be incredibly rugged, and there may not be any lighting. So now they need the same analytics algorithms but on a very robust, ruggedized platform that supplies its own power or its own light. And we’re seeing increasing ability for the industry to deliver solutions that match the disparate needs of customers that have multiple locations with diverse physical characteristics,” says Ramsey.
One impediment to edge analytics and interoperability of systems using analytics is the lack of metadata standards for cameras and analytics, says Levine. Not all software can be used with all cameras, so an integrator might be restricted in options when trying to install software. (This is an issue with various other types of security applications as well). That could change. Many of the experts interviewed for this article expressed optimism that ONVIF (the Open Industry Video Interface Forum, a global industry standards group) or PSIA (the Physical Security Interoperability Alliance) would address analytics standards. For now, however, the groups are focused on more general IP camera interoperability standards and standards for video.
In the absence of that type of broader effort, at least one company, Axis, has taken the initiative to do something on its own, establishing what it calls the Open Applications platform. It is a significant development in analytics and in the analytics-at-the-edge movement, says Honovich. Unveiled in the fall of 2009, the platform allows analytics companies to load their analytics applications onto Axis cameras. It’s similar to the Apple iPhone system of downloading applications onto the phones.
A company “can load, rewrite any application they want,” explains Nilsson. “We won’t guarantee the results you get out of it, but we will guarantee that the camera is not affected by the function of having it running there.”
Nilsson adds that the analytics companies will still sell software licenses for use of the analytics. He says the first companies using the platform are the ones that worked with Axis to develop the program, but it is open to any company that wants to write an application for it.
Another positive trend is that analytics are better than they used to be; the algorithms have improved. That does not mean, however, that the software can just be installed and forgotten. The technology needs to be set up appropriately for the environment, but as mentioned, there’s a much greater recognition of that. Prices are coming down as well. For example, Cisco is among several companies that now offer high definition IP cameras that include an analytics package at no extra cost.
One reason prices have come down is that installation has become easier, which reduces labor costs as does the fact that a single camera can now run multiple types of analytics.
Anyone considering analytics should be aware of the potential for false alarms. For some applications, it is an acceptable cost of doing business. For others, it might be too much of an inconvenience.
Experts say that analytics systems are between 95 and 97 percent accurate, although with tweaking, some can get analytics systems above 99 percent accuracy rates. Nilsson says that a 95 or 97 percent accuracy rate in an airport where 100,000 people pass through each day means constant disruptive false alarms. But in a parking garage where operators are using license plate recognition, it might be acceptable to stop a few cars out of every hundred.
A lot of the false alarms are caused by the same environmental factors that inhibit traditional camera surveillance. One example would be the amount of available light. An analytic in the daytime is going to work differently from an analytic in the same location in the nighttime.
Some factors are seasonal. Outdoor camera and analytics systems may need fine tuning for different times of the year, says Levine. For example, in the winter when the trees are bare, there might be a different chance of false alarms than in the spring. Levine recommends maintaining a service agreement so that updates and tune-ups can be done throughout the year.
These types of variables should be taken into consideration during installation as well as in assessing performance during any trial period.
Setup and Testing
Sometimes, analytics can provide too many events for security managers to deal with, even if they aren’t false alarms. That problem arises when the rules that the security team sets for the analytics aren’t specific enough or aren’t appropriate for that installation.
“Rules” refer to the types and characteristics of alerts that analytic software is programmed to send out. For example, if a tripwire analytic is set up in a certain section of the tarmac at an airport, a rule might prescribe that any time anything moves into that section, security receives an alert. If that type of rule is applied where constant motion is par for the course, then there will be a lot of meaningless alerts.
“They’re not necessarily false alarms, but they’re just too many alarms for anybody to really make sense out of,” says Loghry. “They are what we call nuisance alarms.” He adds that when there are too many alarms, the system becomes unusable. It often comes down to trial and error to refine the rules enough in these situations, so end users must be aware that it won’t be perfect right out of the box.
With that need to tweak the system in mind, Banerjee says that one of the differences in the way that vendors and manufacturers are approaching analytics these days is the time spent on the installation and pilot testing of the product. Steve Collen of Cisco says the biggest tip he would provide to end users is to conduct a pilot test prior to full implementation. And Banerjee notes that installers are receptive to doing this for a week or a month to prove that the technology can work in that environment.
Types of Analytics
Many analytics are being applied on an everyday basis. Some of the most popular ones cited by those interviewed for this article are perimeter intrusion detection and people counters. Others include license plate recognition and detection of objects left behind.
Applications can be found in both the government and commercial sectors. “Everybody wants to look at it, no matter whether it’s a bank, a school, a government organization. So there is universal demand for it and universal interest in analytics,” says Collen. However, IMS’s Jenkins says there has not been a “killer application” for analytics thus far. “It works well in a lot of installations and a lot of circumstances, but there’s yet to be this one market where you have to have it,” he says.
In some cases, analytics are applied primarily to catch something in real time. In other cases, they are used to identify past events or to pull out relevant moments from archived data. For example, some companies are now offering forensic search capabilities that would make it easy to search video data files to find all the red cars that passed through an area during a certain time period. Among the companies offering forensic search are Bosch, Agent Vi, and IBM. Many of the companies are also offering statistical analysis for the data as well.
Breaking the Rules
While some analytics have the ability to learn what is a false alarm and improve their performance over time, they have nonetheless generally been rules-based. One company that has taken a different approach is BRS Labs, which has a technology called AISight. Instead of relying on rules set by the end user, the behavioral recognition software analyzes a scene and formulates what constitutes normal patterns for that location. It can then sound an alert when it detects something that does not fit the pattern.
“We connected the video acquisition side (which is the cameras and the computer involved in maintaining the camera imagery) to machine learning and machine intelligence,” says BRS CEO Ray Davis. “What that allowed us to do,” he explains, “was take out the middle man, which is the human programmer, so to speak. So now we no longer have to tell the software what to look for, because it can look through every angle of every camera and observe patterns and understand what is normal—trajectory, speed, what they pick up, what they leave behind, all those things.”
The benefit is that it does not require the end user or integrator to program rules in for the system. However, Davis says AISight does provide an option to click on a certain activity and direct the system to provide an alert for all similar activities, which can be used in situations where end users want to catch something that may not be abnormal to the situation, such as if they want to catch a certain car pulling up to a house that has many cars pulling up to it.
One downside to BRS’s approach is that it could generate a period of very high false alarm rates, and Davis says this technology does require time at the beginning of the installation for the technology to learn the patterns while the system is in a “suppression mode” of not sending out all of the alerts to the end user. That suppression mode could take anywhere from a few hours to a few weeks.
Davis says AISight has been employed in five “pilot” projects in various locations, such as a hotel in Mumbai, and the company is now developing reseller channels. He adds that in the Mumbai location, they now see only about one false alert every five days, and that the ratio of false alerts to valid alerts is one to one hundred.
Honovich has been skeptical of what the company says the product can achieve, saying it would be tough to deliver on some promises (for example, the ability to tell the difference between human and animals based on height or certain nighttime detections). However, Davis defends the company’s claims, stating that the best way for customers to know if it will work for them is by testing it themselves.
Return on Investment
As with any expenditure, security managers may be asked to justify analytics in terms of return on investment. Loghry and Ramsey say that on the loss prevention side, it’s fairly easy to show the financial return. It gets a little trickier on the physical protection and business intelligence protection side.
One way to ensure or enhance ROI is to find complementary nonsecurity uses for the analytic. For example, Ramsey cites a car sales lot that might be using analytics at night to prevent thefts. It might also want to set up the analytics for daytime alerts when someone has been looking at a car for a certain amount of time on the theory that someone who is checking out the car for awhile is a serious customer. This multipurpose use is called business optimization.
In some applications, the ROI is evident because it is essentially a force reduction tool, says Loghry. Additionally, it can reduce equipment costs, Loghry says, because it eliminates the need for a giant wall of video monitors.
The ROI is just one of the factors that will affect how widely analytics are adopted in the future. Among the other factors, as noted, are the need to streamline and standardize installation and the need for manufacturers and integrators to educate end users on the realistic expectations of the technology.
Laura Spadanuta is an associate editor at Security Management.