How do we balance the use of human CCTV control room surveillance operators and AI systems as societies and organisations rush to implement AI? At this stage, in a ‘shootout’ for effectiveness of crime detection on a single monitor showing one camera view, I would bet on the success of a skilled and trained human operator over an AI to detect theft activity almost every time over an hour.
The human IQ has several competitive advantages, including the essential capacity to view things in context. Aspects of the scene can be weighed up against each other to make sense of the scene and ascribe intention and purpose. Having said that, not everyone does this effectively, and it is a particular focus of my advanced surveillance training. Operators trained in crime behaviour detection are capable of spotting small or subtle crime behaviour indicators that are beyond the capabilities of working AI systems. The camouflaging of behaviour in places like crowded settings is also beyond AI’s ability to recognise and comprehend.
Yet it makes a lot of sense for AI to be used to offset human limitations and the cost of human resources to staff security systems. Always on, never tiring, and capable of monitoring hundreds of cameras simultaneously, this means organisations are considering balancing or even replacing natural or human surveillance with AI. Covering long perimeters of a major building, mine site, or factory for intrusion attempts is done more effectively with automated detection than with operators regularly sweeping the areas with cameras.
Automatically supplying information from facial recognition and ANPR, linked to networked personal profile information, is also expected to expand the information base for informed decision-making by operators. Finding the balance between this IQ versus AI approach will define the nature of security over the next 10 years.
Understanding AI beyond chat
When we discuss AI, our impressions are often shaped by interactions with systems like ChatGPT and Copilot. The creation of imaginary scenes and people, faking the appearance and voices of individuals, and the fabrication of what appears real, but has no basis in fact, is becoming increasingly prevalent in everyday media. These almost miraculous capabilities have been transferred to an expectation of AI in the security area, without much foundation.
Modern AI security perimeter systems are based on video analytics that date back several years, and movement detection that dates back even further. Although the AI capacity of algorithms providing analysis and interpretation of images has increased along with the processing power of technology, systems for areas such as perimeter protection are still largely based on pixel analysis of 2D video and images. Using pixels, we can detect changes in movement, shape, colour, and even temperature, but this is still far from the capacity of large-scale AI systems like ChatGPT and its competitors.
While mainstream AI systems may suffer from so-called ’hallucinations’, the false alarm rates involved in the detection of perimeter AI may amount to hundreds, thousands, or even tens of thousands in sites with long perimeters to protect. Currently, we accept that AI generates these false alarms. because these systems are being ‘nursemaided’ by human operators who are having to bear with these efforts, and often are supposed to verify and try and teach AIs.
As part of this care and development function, however, human operators are being distracted by excessive false alarms and pulled away from their key tasks, sometimes for several hours per shift. Despite the ‘it can learn’ marketing claims that AI product promoters make, I have found in my experiences that they do this very poorly. AI providers have shifted responsibility for learning to their clients, when such technology should largely have the capacity for accurate detection built in. I have seen multiple examples of cats triggering perimeter AI detection systems, which seems to be one of the more simple things that they should get right. You may need to consider investing in a specialised false alarm handling package.
Narrow range and restricted flexibility
Security AI systems require tightly defined parameters to function effectively. The greater the variability of what is being viewed and categorised, the higher the false alarm rate. In a sense, this kind of surveillance is akin to a railway – it works best within a narrow range and its flexibility is restricted. Machine learning needs extensive repetition to learn the characteristics or shape of even one type of object.
With perimeter protection, for example, the further the detection fields move away from the perimeter fence, the greater the variability of conditions and potential for false alarms. This is especially true when you have little control over conditions outside the perimeter, such as vegetation, weather conditions, insects, and animals.
Conversely, human surveillance is more akin to driving a car, allowing for easier and more flexible movement, and even stepping out to investigate specific issues with the use of PTZ cameras or digital zoom capabilities. The AI security systems are also not tailor-built for a client – they are typically off-the-shelf solutions that may have a range of capabilities tweaked for a client. Humans bring to the control room a knowledge of the site, the people and movements, expected behaviour and patterns of behaviour, and a knowledge of a range of policies and procedures that govern behaviour which make it easier to pick up unusual or incident circumstances. A human operator can generally adjust to even unique or unexpected conditions.
When reading about the introduction or marketing of AI systems, we see little reporting on the accuracy rates. Yet different products within the same categories have significantly different accuracy rates, for example, in facial recognition software. Different AI capabilities also vary in their accuracy rates. Part of this may be that what goes into the system determines what the AI can do with it. Even with generic large AI systems, it is often possible to obtain incorrect or inappropriate results.
Pixel-based analysis is more prone to this issue because it must deal with overlapping objects within the environment and distinguish between them. Firearm detection is also affected by this, and there have been some high-profile incidents recently where the lack of detection and activation has been blamed on the shooter’s position and location relative to cameras. More recently, a high-profile case of mistaken identification of a firearm in Baltimore resulted in multiple police vehicles responding to a teen’s empty bag of chips as a possible firearm.
A group of teens at a local high school waiting for a lift suddenly found themselves confronted by police, and the teen with the chips packet found himself cuffed and searched. The incident prompted a statement on social media from Baltimore County Councilman Izzy Patoka that “No child in our school system should be accosted by police for eating a bag of Doritos,” and calling on the school district “to review procedures around its AI-powered weapon detection system.”
Protocols and procedures, and context
The Baltimore incident reinforced the importance of having standardised protocols and procedures to deal with such events, covering the relevant parties and their involvement in any AI notification. There are questions when things like when firearm detection becomes effective relative to the incident – it is likely that the crisis has already begun by the time the firearm can be recognised easily by identifying someone pointing a gun. Ironically, in these situations, AI detection of behaviours like having hands up, lying on the floor, or groups of people running in the direction of an exit may be more effective in detecting firearm-related incidents.
Importantly, the Baltimore incident prompted a response from the company providing the service, which expressed regret over the incident and emphasised that its system was designed to identify possible threats and elevate them to human review. The need for handling consequences, from both accurate and false alarms, is something that needs to be built into the human side of the control room from an operations, public relations, credibility, and legal point of view.
The importance of human verification as part of the process is also highlighted, particularly in applications involving AI, such as perimeter protection, firearm detection, and facial recognition. Currently, AI providers probably pay less attention to this human function of using AI than they should.
Humans provide context in a way that many AI systems cannot. In areas such as perimeter protection, AI systems should be interacting with each other; however, most of these detection systems operate in parallel. Providing some integration context for the AI is likely to have a strong benefit in managing operational risk. For instance, if a system is triggered by motion detection and it is based on movement in the direction of the perimeter fence, one may give it a 30% risk rating.
Combined with object detection, which reports the shape of a person, may increase the risk rating to 50%, line crossing by the object over an electronic trip wire or boundary may boost risk to 85%, and the addition of a fence activation may heighten the risk of an incident happening to something like 95%. These are all working within certain tolerances, as movement may be triggered by a tree blowing in the wind, the object may be a cat rather than a human, line crossing may be triggered by something other than a person, as is the possibility of a fence alarm.
However, when considered in the context of all of them being activated, the context enhances the validity of the total risk notification. Unfortunately, integrating the alarms from all these AI inputs in real-world environments is sorely lacking, never mind the integrated linking of CCTV AI detection to other systems, such as voice monitoring, access control, ANPR, and facial recognition databases. At this stage, AI can provide detection using specific behaviours like crouching, crawling, lying on the ground, hands up, and wearing of PPE equipment, such as hard hats, but these remain basic and need to be easily identifiable relative to their surroundings. Human perspectives in reviewing the combination of events and CCTV footage remain critical to the effective operation of these systems.
The evolving balance between IQ and AI
So, what is the evolving balance of human operator insight and surveillance relative to system AI capabilities likely to look like?
• There are a range of CCTV surveillance areas which remain the preserve of human operators. Places in which there are people, activity, or behaviour are all done much better by people rather than AI. Human operators have a far greater ability to examine elements that can predict incidents, evaluate threat levels, identify groups of suspects, and anticipate how crisis situations will evolve. AI is likely to be used in these scenarios to augment surveillance and other information as it steadily improves and gets new capabilities.
• The nursemaid role of operators in resolving AI issues is likely to continue, although it is expected to decline over the next few years. AI, by itself, in areas such as perimeter security and object identification, is just too inefficient in providing accurate results, and without human intervention, it will compromise real threat detection and effective responses. Operators will still need to train the AI, but its success will be limited. Simply, AI suppliers will have to get their own capacity right, and specialised false alarm screener products have already shown the way. The danger in the interim is that human operators will be pulled away from core functions in order to fulfil this nursemaid role.
• Human verification will be more acknowledged across AI functions in the short term, as AI failures will highlight the response, reputational, and potential legal costs of using it without verification and structured reporting processes.
• Human operators will still largely be responsible for dispatching responses due to their capacity to verify footage and examine the situation and its components more closely. However, as the integration of multiple AI notifications provides better predictions, we will see more automated response preparation, even if this involves preparation to be picked up by an operator. AI will also provide operators with current information from available data sources, enabling them to make more informed decisions by updating live data inputs.
• AI will take over more extended functions where its strengths lie, like perimeter protection, movement in restricted areas, and simple behaviour identification. There is just too much appeal in being able to cover such broad areas without continual human coverage, but this is dependent on the development of greater accuracy and integration, as detailed below.
• AI integration of multiple detection products into a reconciled live risk analysis based on multiple criteria will increase confidence and acceptance. This, in turn, will make human verification easier, as part of the context evaluation is already being done.
We can assume that AI will become increasingly sophisticated and effective as algorithms evolve to better suit specific circumstances. Increased observation from multiple technologies is also likely to enhance the accuracy and predictability of AI findings, including more subtle signs of movement and behaviour, much like you are profiled on the internet and by instruments such as Alexa.
For the foreseeable future, people will still be more competent decision-makers. However, even now, we have situations where some operators working in control rooms do not have the insights we would want, and the gap and the need to fill it is only going to get worse. If people do not have the skills to detect crimes, evaluate situations and understand context, can they truly provide an effective role themselves, never mind confirmation of AI-sourced notifications?
Place the wrong kind of person in the human verification role, and you have a crisis waiting to happen. It places even more emphasis on what the key human contribution should be, where operators should be trained in recognising crime behaviour, situational awareness, and behavioural indicators such as types of body movements, positioning, posture, and behavioural and crime scene indicators. This is one of their key strengths, relative to AI, in delivering results from a CCTV system.
About Craig Donald
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
| Email: | sales@leaderware.com |
| www: | www.leaderware.com |
| Articles: | More information and articles about Leaderware |
© Technews Publishing (Pty) Ltd. | All Rights Reserved.
printer friendly version