Smart surveillance using video analytics or ‘AI’ on camera views has increased in sophistication and capacity over the past few years, and with proper use can contribute hugely to expanding effective security coverage and detection.
If one looks at sales pitches, web sites and demos, and discussions in the industry, it seems that the capability of video analytics has reached almost optimal levels. With the use of the ‘AI’ term and adverts talking about intelligent systems and learning, it seems from the marketing narratives we can almost hand over much of control room security to AI based systems. So, I was somewhat surprised to find when visiting a site earlier this year that the setup of the analytics was a major weak point in an otherwise well-constructed security infrastructure and accompanying personnel. More to the point, the capabilities of an international top line system were far less capable than I expected.
My disappointment in delivery was primarily around the combined functions of movement detection, electronic trip wires or line crossing and bounded boxes where movement rules can be applied, and object classification. These are all key analytics techniques used in perimeter and area protection along with other security functions.
The problems I viewed were around accuracy of detection and extremely high volumes of false alarms, the importance of which I wrote about a couple of years ago (https://www.securitysa.com/9445a). In fact, I had a flashback to about 25 years ago when I remember a movement detection system at a mine being trialled and then turned off because it was generating about four alarms per second and turned out to be more of a distraction than anything.
Yet only a few months ago in 2023, I was watching a control room commonly experience about one alarm per second, despite all this modern security technology and software systems, excellent cameras, and a fairly well designed and functional control room. The vast majority of these alarms were also false alarms, which overload the operator and make the real alarm condition much more difficult to anticipate and detect. With an alarm per second occurring on a regular basis, multiple activations sometimes occurring almost simultaneously, and a spot monitor refreshing so quickly that operators didn’t have a chance to properly view the alarm conditions, analytics were hamstringing the operators of a perfectly capable security system.
Perception vs reality
This experience, which made me question why the video analytics of a sophisticated new system on a site wasn’t working well, caused me to speak to a number of technical specialists in other operations where I was working, as well as the supplier of security cameras and VMS on the site. I found that these problems were not just common, but widespread, showing industry performance just didn’t meet the claims that you read about, where you get a packaged solution that will ‘solve all your problems’. I identified a number of reasons for the shortfall in performance compared to expectations.
1. The demonstrations we often see with techniques like movement detection, electronic trip wires, and object classification are based on near perfect conditions for the analytics to work. The more they are exposed to potentially difficult conditions, the worse they get. This type of problem gets compounded when one is attempting to pick up things at the extremes of camera ranges – not unusual as people tend to maximise the distance of camera ranges to avoid the costs of having more cameras.
2. Video analytics are not one size fits all. A technique that works well in one place may totally fail in another, and even what works on one camera on a site will not work with another camera on the same site and similar conditions.
3. The nature of the background, activity level in the foreground, uniqueness of the object you are looking to detect, and its potential variable nature can significantly affect detection success.
4. Smart analytics needs a lot of smart people input. Like any tool, how the analytics are set up are going to deliver variable results – these are not plug and go processes and interfaces. The better the setup, they better they will work. Not only this, but one needs to actively work at eliminating false alarms due to anything from trees to live creatures.
5. Object classification of these ‘intelligent’ systems can be rudimentary at best. Speaking to a technical specialist at one site, he was having to spend appreciable time getting the system distinguishing between cars and people, something I thought would easily be done with modern analytics. Distinguishing between a buck, a rabbit, a lizard, and a human are done remarkably poorly, especially when these are not in the camera foreground. The further the camera viewing range is pushed, the more difficult it is to make these kinds of distinctions. Movement does make event detection easier, but object classification or recognition is still lacking.
6. It is possible to teach these ‘intelligent’ systems something specific to look for and customise this amazingly well. But somebody has to do the teaching and it takes huge amounts of time for very specific purposes – you are likely to have to resort to camera specific learning and define specific targets such as certain PPE. Machine learning needs a lot of practice. A lot of practice with real behaviours and not just some simulated vectors. Once done, the one size fits all issue may mean that you can’t simply transfer that knowledge to another camera. Even under different situations such as changes in season, the learning effect in the algorithms may become invalid.
7. Using blank (or black) screen technology with analytics where a screen activates when there is an alarm creates a constant flow of false alarms with only a few positives, creating a cry wolf effect and reducing the control room operator’s attention and concern over the ‘alarm’. Where I’ve seen operators with a video wall showing all cameras (at a viewable size), operators acclimatise themselves to some degree to what causes triggers on which cameras, how these get displayed on screen with boxes or highlights, and what is unusual. This enables them to get used to the pattern of false alarms and focus on potential events and overcome the shortfalls of the technology they are working with. However, continued alerts, which are mainly false still cause a major distraction from detection of potential key event alarms.
8. Natural factors such as trees and other vegetation such as bushes and grass, especially exaggerated by wind, still create huge numbers of false alarms despite the ‘intelligence’ of analytics.
9. There is a distinct trade-off between moving analytics further away from the critical protection point or fence line to give early warning, and the number of environmental variables that are going to negatively impact on system performance beyond your control.
Much of the success in setting up and managing video analytics is how much time and attention is given to defining the rules, parameters, and conditions under which the analytics must operate. Smart analytics need smart people to set them up with an understanding of the risk profile of the area and the impact that it makes. Rushed jobs are not going to deliver, as thoroughness is one of the important success factors.
Most analytics systems allow a fair amount of customisation. Looking at issues such as triggering time, thresholds of when something is recognised, and the combination of movement and vector signals are all ways in which recognition can be improved. These can be applied to the nature and direction of movement, which is already a potentially potent tool. If you are using a camera on a long fence line, defining different zones like foreground, middle ground and background, and customising the analytics for these different zones, can give improved detection and lower false alarms at some expense of use for middle and background conditions.
Developing your own parameters through the learning capabilities of the analytics is also a way of getting more targeted effect and getting better efficiency, but be prepared to spend a lot of time on it – teaching both normal and event conditions that have to be realistic.
Finally, maintenance and ongoing refinement are essential to a properly functioning video analytics system. Identifying causes of false alarms, keeping areas under surveillance free of environmental triggers, such as branches and other moving vegetation, telling the system what to ignore, and defining areas on screen that should not be affected, are all ways of optimising the analytics system and getting it to deliver.
The consequences of poor setup of analytics are typically born by the control operators, not the management or technology providers, until a major failure hits the site. Often poor implementation is hidden by a wealth of alerts that are flashed up on screens to show the video analytics system is ‘working’, when instead it is a facade of poorly implemented alerts that have no relevance to the actual task of detection. False alarms are still the biggest deterrent to effective analytics, and it is as much of a human issue as a technical issue. Your smart surveillance, if it really is smart, is only as smart as the way you set it up and implement it.
Dr Craig Donald is a human factors specialist in security and CCTV. He is a director of Leaderware which provides instruments for the selection of CCTV operators, X-ray screeners and other security personnel in major operations around the world. He also runs CCTV Surveillance Skills and Body Language, and Advanced Surveillance Body Language courses for CCTV operators, supervisors and managers internationally, and consults on CCTV management. He can be contacted on
|Tel:||+27 11 787 7811|
|Fax:||+27 11 886 6815|
|Articles:||More information and articles about Leaderware|
© Technews Publishing (Pty) Ltd. | All Rights Reserved.