Seeing Limitations
Expecting another video analytics as a feature might deliver less than promised
- By Oliver Vellacott
- Jul 16, 2007
ANALYTICS can detect suspicious movement from people walking along a street, detect terrorists walking around a hillside a mile away or pick an offender out from a sea of faces. These are just some of the misconceptions about current analytics. Was there ever a product so over-promised and under-delivered? The reality is analytics technology is still in its infancy.
Expectations are at the heart of the issue—being realistic with end users about what can be achieved. The fundamental problem is that humans do these tasks without even thinking. People read license plates and recognize faces subconsciously. It may have taken humans years of learning during childhood to acquire these skills, but now, these tasks are taken for granted. Computers, on the other hand, lack even the basics of visual intelligence. The technology can perform some video analytic functions reliably, but often, only by severely constraining the application. Qualification is everything, and meeting end-user expectations is vital.
Analytics Today
So what can be done today? License plate recognition technology has been around for years and is well-proven. However, still it is not 100 percent accurate. Face recognition is notoriously difficult to perform reliably and is extremely easy to fool by using disguises. For face recognition to work with any degree of accuracy, an excellent headshot of the subject is required. But there are, however, some bread-and-butter analytics functions that can perform well.
Motion detection is the most simple and basic form of analytics. Many manufacturers support it, but few systems achieve sufficiently low false alarm rates to be useable. A system that generates anything in excess of 20 false alarms in a night becomes ineffective, because all alarms quickly become ignored or the motion detection gets switched off to prevent the operator getting drowned in alarms. Aesop unwittingly foresaw the reality of most security systems when he wrote "The Boy Who Cried Wolf."
Congestion detection has evolved from basic motion detection. When the density of humans or cars reaches a certain level, an alarm is triggered. Counter flow looks for objects that move against the flow and is valuable in applications like airport security. Virtual tripwire also is a refinement of motion detection because it triggers an alarm when someone or something breaks a line drawn in the image. This is useful in large areas with no-go zones, such as in factories.
People are allowed to happily move in free areas, but the system alarms as soon as movement is detected outside the defined zone. With all these analytics applications, camera positioning, lens selection and lighting are critical. Just changing the camera position improves analytic performance by an order of magnitude. For example, in counter-flow detection, the algorithm has an easier job if the camera is positioned pointing down, so it sees the area in plain view. If it is looking from a perspective view with human bodies occluding one another, the process of tracking people becomes much harder, and expectations can once again become unrealistic.
There are hundreds of companies touting analytics software. Manufacturers of IP video solutions are approached by analytics providers each week, asking to integrate an analytic product into their IP video management platform. Why so many? Any small software company can develop a suite of video analytics by buying a frame grabber, a powerful PC and writing software. These are then sold as separate, standalone systems that sit next to the main CCTV system. Video is split from the matrix and fed into the analytics system. This delivers only limited benefits because it’s not integrated into the operation of the main CCTV system.
Analog technology just can’t support integrated analytics. DVRs do to a greater extent, but the technology still remains in islands and is not fully integrated.
IP Intelligence
IP-based video management systems provide an ideal platform for powerful analytics to be completely integrated into the system, making it a core and integral part of operation. Leading IP video solutions support analytics that can be performed in two fundamental modes: live to detect events during occurrence and post-processing to test various scenarios on recorded footage.
The optimum place to locate live analytics is obviously at the camera, as it is the only truly scalable solution and doesn't use up network bandwidth. Central, real-time processing will eventually run out of steam, but every camera can have dedicated processing. For example, a camera with built-in analytics can monitor scene activity and transmit only on specified events, such as a person moving the wrong way through airport security. This cuts unnecessary video traffic on the network, reducing bandwidth requirements—something that cannot be done with traditional analog technology.
The optimum place to locate post-processing analytics is obviously on a central server so recorded video can be searched many times with different parameters. One of the biggest time wasters for operators used to be fast forwarding and rewinding through VCR tape. This is improved with DVRs, but most still feature digital fast forward or rewind. Analytics offer the potential to further transform the task by searching large amounts of recorded video for possible events for operator validation. Computers can accomplish one things—locating possible events—and humans do what they are good at—verifying those events.
Setting expectations is everything. It’s important not to believe all the hype about what analytics can deliver. In 30 years time, it will probably be possible, but today it's all about making sure that what is achievable is done well.