My focus on CCTV surveillance since the early days of CCTV implementation has been on behavioural analysis and crime behaviour detection, detecting the ‘body language’ of crime incidents. I have consulted with and trained organisations internationally, including police, city and town centres, customs, airports, precious minerals, casinos, residential estates, banks, shopping centres, universities, factories, and retail stores, on threats and how to detect them. All places where there is high importance in the protection of human lives or a high value of property.
The exposure has also given me a wide view of the impact of AI across different industries. I have seen the use of facial recognition to verify aircraft boarding in the US and identification of suspects entering one of the major international malls, use of AI to recognise behaviour under firearm threat, all sorts of perimeter analytics, successful number plate recognition while in a police vehicle, heat mapping of crowd disturbances, detection of unusual movement, searching of video for people matching certain clothing colour descriptions, and even vehicle tracking.
The introduction of analytics, especially AI, has been seen previously as a way to enhance surveillance’s capacity to detect incidents. Interestingly, the initial marketing strategy for video analytics was that it would remove the need for CCTV operators. Some discussions on AI and marketing have also featured replacing operator activities, but there has also been an emphasis on extending or augmenting control room capacity along with operators. So, has AI reduced the need for these traditional surveillance behavioural analysis activities?
In the last few weeks, I’ve been involved in training surveillance across a range of operations, including casinos, estates, entertainment and business complexes, port operations, tourist destinations, and even drone operations. After some reflection and discussions with colleagues, my feeling is that, despite the increased use and benefits of AI, the role of behavioural recognition and the evaluation of situational threats by CCTV surveillance operators is just as important as it ever was.
The way things are displayed in the control room, and the type of surveillance activities may have changed or are being supplemented. AI, particularly in cases such as perimeter protection and basic behaviour identification, like crouching, has enabled control rooms to effectively monitor far larger areas and longer perimeters. It means that, in some operations, one of the central control room tasks is responding to AI-generated alarms based on movement, shape, thermal signatures, line crossings, etc. Large display screens displaying many camera views have an alarm condition typically shown with a coloured outline around the triggered alarm condition. Alternatively, camera views may pop up on display screens when the alarm condition is activated. At first glance, this situation may seem similar to responding to a simple alert like a flashing light, but in fact, if done properly, the evaluation of the alarm condition may be far more complex.
Human knowledge understanding the environment
Effective handling of AI-generated alarms requires that surveillance operators understand the context in which the alarm occurs, especially given the propensity of AI systems to give high false alarm rates. I have spent a lot of time discussing these issues with operators responsible for monitoring residential estates recently, and I have been emphasising the importance of having a knowledge of the area being viewed by the camera, and what they should be looking for. Operators usually get to know likely causes of alarms even before inspecting the camera view due to their familiarity with likely triggers and the areas being monitored.
These may be social activities, such as someone jogging every morning on a standard route along the fence line, community activities outside the fence line, like using throughfares on the way to work, people gathering firewood or picking up litter, dog walkers, or even wildlife or insect presence. However, operators also need to be sensitive to behaviour that indicates a potential or actual breach of the perimeter, both when an alarm is activated and at other times.
This may include prior reconnaissance of the area, preparation for a breach, carrying of articles that may assist in breaching, testing of responses and response time by deliberate activation of alarms, evaluation of weak points or ease of access, risk factors such as physical features of the surrounding ground or fence itself that may assist access, and even the conditions within the property that may show a preferred breach point and hiding area within the property once inside. There is a need for an understanding of the area’s social dynamics and the kinds of movement and common activities that provide the basis for benchmarking behaviour and detecting inconsistencies. This allows the operators to view proactively and even predict potential issues, and resolve alarm conditions much more quickly.
There remains a wide range of common control room surveillance situations where AI simply does not work or is still far behind people in capacity. Detecting pickpockets in a crowd through movement, differences in flow, positioning, posture, and even exit or escape behaviour is far better done by a skilled surveillance operator. Picking up shoplifters is also done more effectively by people, and surveillance operators at shopping malls can sometimes detect them even before they enter the shops. Pushing a trolley out of a retail area with hangers still in the clothing is a relatively good indicator of someone shoplifting, which current AI would find it almost impossible to detect unless specially tuned to recognise this.
Spotting a suspect who is going to steal a phone or bag in a library is relatively simple with crime behaviour analysis. Similarly, surveillance of cheat moves or scams in a casino, bag thefts at malls or airports, or criminals in a car park is the preserve of human operators, as is monitoring searches at exit points to some operations. Identification of syndicate involvement in the enactment of a crime in almost any situation is something AI would find extremely difficult to do, given that much of this involves covert communication and subtle positioning, but a trained human operator would find it far simpler to achieve successfully.
AI can provide an indicator of things like ATM theft, where suspects are crowding somebody drawing money, but a human operator will detect the person hanging around or outside the ATM and suspicious queuing behaviour far more easily and earlier. Sometimes people’s behaviour may be a far better indicator than technical AI detection. For example, people holding their hands up, lying down, or groups running for an exit are far more recognisable signs of weapon detection than AI use of pixel analysis for detecting people holding firearms where these are at angles that are difficult to analyse, or there are background distractors which hamper AI recognition.
Activity heat mapping of more active group activity or concentrations is something that AI can potentially provide across a wide area, but a human control room operator is still needed to confirm the nature and reasons for the behaviour.
It comes down to training and experience
CCTV operators who have been properly trained in crime behaviour analysis can evaluate the potential for a crime to happen, even before it occurs. CCTV AI, on the other hand, typically focuses on a specific condition reflected in pixel display on screen to raise an alert. We can expect AI to get increasingly sophisticated, however.
I was having a discussion with someone recently who was interviewing for a major UK football club about using algorithmic analysis of video movement of matches to identify players and plays that went into setting up goal scoring well before the goal happened, with the aim of identifying players and moves that were responsible for a successful build up and were an essential part of the team in that kind of role.
Currently, however, we need to be careful not to simplify the task of operators interacting with AI systems. The use of AI may reduce the headcount of people or expand the operational coverage, but increase the cognitive demands and work complexity of those remaining. If surveillance operators in control rooms do not understand the social and crime dynamics of the areas they are viewing and the crime behaviours they are supposed to be looking for, delivery may fall well short of expectations.
Most AI systems do not have the intuitive understanding required and machine learning requires huge amounts of learning material. These behavioural surveillance skills are as essential now as they have ever been. This can be even more so in some situations where quick evaluation of context is needed, and training in these kinds of behavioural recognition skills is a vital part of service delivery of the CCTV control room.
About Dr 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