The potential to reliably detect intruders, criminals and terrorists by analyzing CCTV video without human intervention: That’s the power of video analytics.
by James Careless
At the Miami Parking Authority facilities, Infinova cameras capture images of the car, the license plate and the face of the driver, going both in and out of the garage. It is a specialized form of software that can decide whether or not something is amiss with what the surveillance camera is seeing—Is that a person trying to climb the fence?—and decide whether or not to alert the system’s human monitors. In doing so, video analytics technology eliminates the mind-numbing aspect of human surveillance; namely having to watch numerous camera feeds in real-time, when nothing is occurring with most of them.
“Today’s intrusion detection systems often have hundreds of cameras; some have thousands,” says Steve Vinsik, vice president for critical infrastructure protection with Unisys Federal Systems. “A human can only watch so many video feeds at one time, and it is extremely fatiguing to the viewer when nothing is happening in most of them. Video analytics takes this job away, allowing the human to only turn their attention to cameras where something is detected to be happening. This improves security, and makes the CCTV monitoring job more effective and much easier to bear.”
This said, video analytics has its limits. It can substantially shoulder the most mundane aspects of CCTV surveillance, but the final decisions still must be made by human operators. And no system is perfect: Even with the best of software and hardware, video analytics systems do come up with “false positives” and “missed detections.”
“It is always a trade-off between missed detections and false positives,” says Moti Shabtai, executive vice president of Sales for NICE Systems in the Americas. “If you lower the threshold at which the system decides to send an alert, when an anomaly is detected in the camera’s field of vision, you lower the rate of missed detections but raise the potential for false positives. This is because the images are being analyzed using mathematical algorithms, with the human operator setting the line between ignoring or alerting. Move the line, and you change the conditions. How you set the line depends on whether missed detections are more important than false positives in your particular deployment.”
A video analytics system is not a form of “artificial intelligence.” The software application does not make reasoned decisions as to whether or not to send out alerts. Instead, the system registers what is being detected on each camera’s CCD, and compares that against a preset baseline. Ultimately, the decision to send an alert (or not) is a matter of pixels: A person moving within the camera’s field of view changes the way light is being received on the CCD, and that is what the software notes.
Pedestrian detection in an Axis system. “Video analytics is a rules-based system,” says John Whiteman, vice president of Strategic Programs for DVTel. “What we do is create algorithm-based rules that tell the software what to do when something changes in its field of view that meet the rule requirements. For instance, we may draw a “tripwire” in the camera’s field of view in a specific sensitive area. If something in the image crosses this tripwire, an alert is sent to the operator. The same can be done to detect people going the wrong way in secured airport areas, leaving cars unattended where they’re not supposed to—as was the case with the Oklahoma City bombing—or climbing fences. It’s all about the rules we design, and how well the software can analyze the video input to apply them.”
There are two ways of embedding video analytics in a CCTV system; each with their own benefits and limitations. One method is put the actual detection software at the “edge” of the network; that is, within the cameras themselves. This is ideal from a system scalability point of view and today many of the basic analytic applications such as motion detection, audio detection and tampering alarms are included as embedded options in most high quality cameras. The benefit of this approach is a minimal need for network bandwidth: Cameras only use the network to alert the monitoring station and stream a video feed when something is happening. Another benefit is that the video stream is analyzed before it’s compressed, which utilizes the best possible image and reduces overall processing power of the system. The limitation is that the cameras themselves can only contain so much processing power today, thus limiting the degree of analysis that they can execute.
“The solution to this limitation is to have all the video sent to a central processing location, when more powerful servers can run more complex analysis,” says John Bartolac, manager of government programs and government business development for Axis Communications. “The downside is that a much larger network is needed to handle all of the camera feeds coming into the central point. This is why the camerabased model—running the analytics at the edge—is becoming more popular. And, using Moore’s Law, the processing power is quickly increasing and within a few years more or less all analytics could run within the camera.”
According to Bartolac, there are three features needed to effectively run analytics at the edge: superior image quality, processing power, and an advanced algorithm. The first two features are determined by the camera manufacturer, while the third is often provided by an application development partner who specializes in video analytics and monitoring software.
THE BIG 5 APPS
Although video analytics are deployed across a wide range of situations, there are basically five applications that describe how this technology is used. These are object detection,object recognition, tracking, behavior analysis, and metadata analysis.
“Object detection is simply detecting that an object is present within the viewing area,” says Vinsik. “Object recognition goes one step further, by telling the system what that object may be;such as a person, car or animal. Tracking allows you to find out where the intruding object camefrom, based on correlation of its visual pattern across cameras in the CCTV system. Somesystems can break this down to the color of shirt the person is wearing, which helps the humanoperators track the person in a real-time crowd situation.”
Behavior analysis attempts to predict what the object is likely to do next, based on what theyhave been detected doing to date, he adds. “Metadata analysis is the highest level: That’s whereyou compare the video analytics’ data against data in other systems, to look for known patternsand outcomes. This is where the system can really shine; for instance by comparing a white vanin the current shot, and finding out that it was seen in the same area three times before.”
The power of metadata video analysis is that it allows a security unit to compile and then screenfor known and/or suspected threats. “It’s like having a Dewey Decimal library-style collection ofarchival profiles to choose from and watch for,” says Mark Wilson, vice president of marketing forInfinova. “This adds an extra level of power to video analytics and its usefulness.”
Line control in a NICE system. MANAGING EXPECTATIONS
Given how much capability that video analytics offers, it is easy to expect it to offer everything. “But this is not the kind of stuff you see on ‘CSI’ on TV,” warns DVTel’s John Whiteman. “You can’t access the system through a browser search window, type in ‘1998 Chevy Van with dented right fender and Grateful Dead sticker’ and expect it to spit back clips.”
In the real world, what you get out of video analytics is based on what you start with. A system of cameras with on-board processing cannot be expected to do facial recognition. And, upgrading the cameras’ resolution from SD to HD doesn’t necessarily solve the problem: “More resolution means more processing is required, which means more powerful hardware and software, potentially bigger bandwidth demands on the network, and higher costs,” Wilson says. “There is a point at which the data desired exceeds the capacity of your system and budget.”
For video analytics vendors, installers and users, it is critically important to manage the expectations attached to a system from the outset. “If people know what they can and cannot do with their video analytics software right from Day One, then issues of disappointment are vastly reduced,” says Moti Shabtai. “Disappointment matters, because it leads to human monitors doubting the system and dismissing real alerts as false positive, with dire results.”