Facial recognition has become one of the defining technologies of modern surveillance. AI systems can identify individuals from images and video with a level of accuracy that would have seemed remarkable only a decade ago. Despite these advances, public safety remains one of the most difficult challenges for AI. Preeti Patel, Professor and Co-Director of the AI and Data Science Research Group, London Metropolitan University, explores.
Whether at festivals, sporting events, transport hubs, public demonstrations or large-scale religious pilgrimages, risks often emerge not because of who is present, but because of how people behave collectively.
This raises an interesting question. What if we've been focusing on the wrong problem? What if public safety is not always about identifying individuals, but about understanding behaviour? As discussions around AI surveillance continue to evolve, there is growing interest in approaches that can improve public safety while reducing reliance on personal identification.
From watching to understanding
The evolution of surveillance technology can be viewed in stages.
The first stage was largely passive. CCTV systems provided evidence after something had happened. They were useful, but fundamentally reactive.
The second introduced AI-driven analytics. Cameras became capable of counting people, monitoring occupancy levels and detecting congestion. Today, crowd analytics systems are already being used in transport hubs and large venues to understand how people move through spaces. For example, companies such as CrowdVision have developed systems that analyse passenger flow and crowd density in airports, helping operators identify bottlenecks and improve operational efficiency.
These systems answer questions such as:
• How many people are present?
• Where are they moving?
• Which areas are becoming congested?
Useful as these capabilities are, they only tell part of the story.
The next challenge is understanding behaviour.
The rise of behavioural AI
Recent developments in computer vision and machine learning are making it possible to analyse not just where people are, but how they are behaving.
Researchers increasingly refer to this as behavioural AI: systems designed to identify patterns, interactions and behavioural changes within groups.
Think of a busy railway station. Most of the time, crowd movement follows predictable patterns. People enter, leave, queue and disperse in familiar ways.
But sometimes behaviour changes.
People suddenly stop moving. Groups begin clustering unexpectedly. Movement becomes erratic. Individuals start moving against the flow of the crowd. Density increases rapidly in a particular location.
Human observers are often surprisingly good at recognising these signals. Experienced event managers and security professionals frequently describe situations where they sensed that something was wrong before an incident actually occurred.
The ambition of behavioural AI is not to replace human judgement, but to augment it by identifying patterns that may otherwise go unnoticed.
This is reflected in wider research trends. Studies in crowd analysis have increasingly moved beyond simple counting towards behaviour understanding, anomaly detection and predictive modelling.
Several developments in AI are helping to accelerate this shift. Modern computer vision systems no longer analyse images frame by frame, but model behaviour over time using temporal and transformer-based architectures. Vision-language models (VLMs), such as GPT-4o and Gemini, have also demonstrated an increasing ability to interpret complex visual scenes and relationships, suggesting future systems may become better at understanding context rather than simply recognising objects.
The technology is advancing rapidly. But recognising patterns in a video stream is one thing; identifying meaningful indicators of risk is another. Can AI really distinguish between normal crowd behaviour and the early signs of congestion, panic or aggression?
Can it work in practice?
This is the question that underpins our current research.
The work sits at the intersection of computer vision, behavioural modelling and privacy-preserving AI. Most biometric systems rely on what are often called ‘hard biometrics’ - characteristics such as faces, fingerprints or iris patterns that are directly linked to identity. Our research focuses on behavioural characteristics, sometimes referred to as ‘soft biometrics’, including movement trajectories, posture, group interactions and changes in collective behaviour over time.
These behavioural representations are then used to train deep learning models that learn patterns associated with different crowd activities such as gathering, dispersing, queueing or exhibiting unusual movement patterns. By analysing sequences of behaviour rather than individual images, the aim is to determine whether emerging risks can be identified before they become obvious to human observers.
For you
Be part of something bigger, join BCS, The Chartered Institute for IT.
A particular focus of the project is determining whether useful behavioural information can be retained while reducing the potential for identity disclosure. In other words, can AI learn enough about what people are doing without needing to know who they are?
The research remains at an early stage. The goal is not to predict specific incidents with certainty, but to investigate whether behavioural signals can provide earlier warnings of elevated risk.
The challenges ahead
As with many emerging AI technologies, the opportunities are accompanied by important challenges.
One challenge is context. Human behaviour varies enormously between environments;. systems that perform well in one setting may not generalise to another.
False positives are another concern. Excited celebrations, spontaneous gatherings or unusual but harmless behaviour may be incorrectly interpreted as signs of risk.
Data remains a major challenge. Real crowd incidents are rare, sensitive and difficult to capture, meaning that building reliable datasets may be just as important as developing the AI itself. As a result, researchers are increasingly exploring self-supervised learning to identify unusual crowd behaviour.
There are also broader questions around transparency and accountability. If an AI system flags a crowd as potentially high risk, how should that decision be interpreted? What evidence supports the assessment? How should human operators respond?
These issues are becoming increasingly important as regulators develop guidance for the responsible use of AI. The Guidance on AI and data protection from the UK Information Commissioner's Office, and the NIST AI Risk Management Framework emphasise the importance of governance and risk management when deploying AI systems in real-world environments.
Looking ahead
Looking further ahead, behavioural AI is unlikely to rely on video alone. Future systems may combine video, audio, sensor feeds, location data and environmental information to build a more complete picture of what is happening in a space.
In this sense, the future may be less about smarter cameras and more about multimodal AI which builds a richer understanding of how people, spaces and events interact in real time.
Whether this leads to safer public spaces remains an open question. The technology is still evolving, and many ethical and governance challenges remain. However, the direction of travel is becoming clearer.
As AI moves beyond simple recognition towards understanding behaviour and context, the focus may shift from identifying individuals to understanding situations.
With thanks to Dr Bilal Hassan and Mr Sonjoy Das for their contributions to our current research project, which informed this article.
Further reading on the topic of artificial intelligence
