GOTCHA: How AI and Digital Twins Can Strengthen Stadium Safety
As stadium operations become increasingly complex, the challenge is no longer simply having enough cameras in place. The real question is whether the available infrastructure can deliver the right footage, at the right time, and in the right quality to support effective decision-making.
That was the central theme of the GOTCHA project, presented by Professor Steven Verstockt of Ghent University. The initiative explores how AI video analytics, 3D digital twins and smarter camera control can help stadiums improve situational awareness, accelerate incident analysis and make better use of existing CCTV infrastructure.
From Camera Networks to Intelligent Safety Systems
Many stadiums already operate extensive camera networks, yet these systems often remain limited in practice. Operators must monitor large numbers of feeds, incidents are highly time-sensitive, and camera quality is not always sufficient to produce identification-grade evidence. As highlighted in the presentation, the result can be missed events, late detection and time-consuming investigations.
The GOTCHA project addresses this by moving beyond passive surveillance. Instead of treating cameras as isolated devices, the approach connects them into an intelligent ecosystem that combines video feeds with spatial context and automated analytics.
In simple terms, the project aims to help clubs answer critical questions more quickly: What happened? Where did it happen? Which cameras captured it best? And can the footage actually be used for identification?
A Practical AI Approach Rooted in Stadium Operations
A key strength of the presentation was its pragmatic view of AI. Rather than presenting artificial intelligence as a solution in itself, Steven Verstockt stressed that AI is valuable because it saves time. It can support difficult tasks such as searching for relevant clips, generating faster insights and acting as an extra pair of eyes for operators.
At the same time, the presentation also underlined that AI should not be treated as infallible. Human validation remains essential. Questions around synchronisation, model accuracy and consistency still need to be addressed, and human oversight must remain part of the workflow. As noted in the presentation, validation, evaluation and a human-in-the-loop approach are necessary.
This balance between automation and human judgement is particularly important in safety operations, where evidence quality and decision-making reliability are critical.
Why Data Quality Matters More Than Algorithms Alone
One of the clearest messages from the session was that successful AI deployment depends first and foremost on the quality of the input data.
The presentation argued that around 80 per cent of the challenge lies in data preparation and collection: getting the right data, in the right place, at the right time, and in the right format. AI is only one part of the puzzle. Poor camera positioning, insufficient image quality or inconsistent coverage will limit the usefulness of any downstream analytics.
This is highly relevant for stadium operators. Many venues invest significantly in CCTV infrastructure, yet may not know whether every seat is actually covered on matchday or whether image quality is sufficient for later identification. GOTCHA responds to that gap by introducing objective ways to measure both coverage and image quality.
Defining What Good Camera Coverage Really Means
A major part of the project focuses on the lack of objective camera requirements in stadium safety.
The presentation refers to the DORI framework, a recognised CCTV standard that classifies what footage can realistically achieve, from detection and observation to recognition and identification. The issue, however, is that many stadiums may have cameras in place without knowing whether they truly reach these thresholds in practice.
The project raises operationally important questions:
- Is continuous full coverage in a recognition zone required?
- Do high-risk areas need regular close-ups in identification quality?
- Should every person be identifiable within a specific timeframe?
These are not merely technical questions. They go to the heart of what stadium safety systems are expected to deliver. The presentation also stressed that pixels per metre alone are not enough. Quantity does not equal quality, and image clarity can also be degraded by compression, networking issues, weather conditions or poor camera configuration.
Using a 3D Digital Twin to Understand Coverage
To tackle these issues, the GOTCHA project uses a 3D digital twin of the stadium. By calibrating and mapping each camera onto that model, clubs can generate an objective coverage map showing which areas of the venue are visible and at what level of quality.
This brings several immediate benefits. It allows operators to identify blind spots, confirm whether every seat is covered, and understand how camera performance changes over time. It also reveals a reality that may otherwise go unnoticed: cameras that appear correctly positioned on paper may in practice have shifted, been re-angled for events, or no longer provide the intended view.
In this way, the digital twin becomes more than a visualisation tool. It acts as a decision-support layer that links physical infrastructure with safety operations.
Smarter PTZ Control for Faster Incident Analysis
Another major building block of the project is the use of smart PTZ control.
Rather than leaving pan-tilt-zoom cameras idle when they are not being manually operated, the system enables them to run through preset scanning patterns and create a time-indexed logbook of close-up footage. When an incident occurs, operators can then link the time and 3D location of the event to the PTZ log and quickly retrieve the most relevant high-resolution camera views.
This has significant operational value. According to the project demonstration, finding all relevant camera footage could be reduced to around five minutes, whereas a manual search might otherwise take several hours.
For safety teams, this means faster analysis, quicker evidence gathering and less operator overload during post-incident reviews.
From Smoke Detection to Event-Based Camera Steering
The project also explores how automatic detection models can improve responsiveness in real time.
One example presented was smoke and flare detection, where text-prompt-based analytics can identify an incident on a static camera feed and automatically steer a PTZ camera towards that location. This creates a more dynamic response workflow, allowing the camera system to react to events as they unfold rather than relying solely on manual intervention.
This is especially relevant in the context of pyrotechnics, which were highlighted in the wider safety discussions. The ability to detect such incidents early and immediately direct enhanced camera coverage towards them could strengthen both situational awareness and follow-up investigation.
Addressing Thrown Objects Through Synthetic Data
A particularly innovative part of the project concerns the detection of thrown objects, another recurring stadium safety issue.
Because real annotated datasets for such incidents are often limited or inaccessible, the project experimented with game-engine-based synthetic data generation. By using real stadium footage and overlaying artificial objects within a simulated environment, the team created training data for models that can learn to detect bottles or other objects landing on the pitch.
This is a notable step because it shows how clubs and researchers can train safety-related models even when real incident footage is scarce, sensitive or difficult to share. It also opens up possibilities for simulating other scenarios, including artificial flares or orchestrated incidents, to improve model robustness.
A Multi-Purpose Use of Existing Infrastructure
Although the project is primarily focused on safety, the broader message is that intelligent stadium systems can create value beyond incident management.
By combining AI analytics with spatial mapping and camera control, the same infrastructure could also support:
- fan flow and queuing analysis
- operational logistics
- media and storytelling applications
- wider venue intelligence use cases
This is an important point for stadium operators weighing investment decisions. The goal is not simply to add more technology, but to extract greater value from the systems already in place.
Key Takeaways for Stadium Operators
- AI can add real value to stadium safety, especially by reducing time spent on clip retrieval and incident analysis
- Human oversight remains essential, as AI outputs still require validation and quality control
- Camera coverage and image quality should be measured objectively, not assumed
- 3D digital twins can help clubs understand blind spots, coverage gaps and identification capability
- Smart PTZ workflows can turn existing camera infrastructure into a more proactive operational tool
- Synthetic datasets offer a practical way to train models for difficult incident types such as thrown objects
