Xavier Oliva
Senior Key Account Manager, Lanaccess
Telecommunications Engineer
Lecturer, Pompeu Fabra University, Barcelona
20 October 2025
Share
Table of contents
Artificial intelligence has become a core component of modern video surveillance solutions.
While model training remains a resource-intensive, centralised process in data centres,
inference – the practical application of AI – needs to take place closer to the edge (in cameras
and video recorders) for reasons of connectivity, bandwidth and security.
This article outlines key hardware design principles, architectural decisions and use cases for
metadata exploitation, and concludes with a legal note on facial recognition.
From Cloud AI to Edge AI
In recent years, investment in data centres for generative AI and language models have
increased. It is important to distinguish between two phases:
- Training: Creating and fine-tuning models; requires large computing clusters and is carried out centrally.
- Inference: Operational use of AI to respond to new inputs; in video surveillance, it is optimal to execute it close to the data source.
Why Edge AI in CCTV?
- Isolation: Many security systems operate without an internet connection.
- Bandwidth: Video is data-intensive; sending raw streams to a data centre for
inference is inefficient and costly.
Specific Features of CCTV Systems
Unlike other AI domains, video surveillance works with continuous video streams under
demanding conditions. This requires technological decisions based on reliability, efficiency
and cybersecurity from hardware design through to data exploitation.
Video Recorder Design Principles:
- Energy Efficiency and Thermal Management: More power ≠ better if it
compromises operating temperature. Proper sizing minimises overheating and
facilitates heat dissipation. - Reliability and Lifespan: Avoid thermal bottlenecks and fragile mechanical
components such as fans, which are frequent points of failure. In real-world
environments (cabinets, vandal-proof enclosures, dusty or vibrating spaces), fans
significantly reduce durability and increase failure rates. - Built-in Cybersecurity: If the system itself is a security device, it must be secure by
design. It must be protected against cyberattacks and always available to respond to a
security incident.
Real-world cases exist where clients encountered camera model changes introducing fans
without prior notice. Although seemingly minor, this can have serious consequences: in
installations with vandal-proof enclosures or confined spaces, ventilation is insufficient,
causing overheating and accelerating component wear, leading to premature failure.
Where Should Intelligence Reside?
In Cameras: First Line
Wherever possible, algorithms should run directly in the camera itself: this is closest to the
video source and avoids unnecessary data transfer. Object detection, classification, attribute
generation, area entry or line crossing are examples of such analytics.
In Video Recorders: Second Line
The video recorder acts as an aggregator and reinforcement layer:
Management of alerts and metadata generated by the camera’s AI.
Combined storage of video and metadata to enable efficient searches.
Additional inference when algorithms not supported by the camera are required or when custom algorithms are needed for specific use cases. The AI processor can be more powerful than those in cameras, as recorders are usually installed in more protected environments.
Hardware Strategies for AI in Recorders
- Single, high-performance CPU (main CPU also running AI): simple to integrate, but increases power consumption and heat, reducing reliability.
- Dedicated AI coprocessor (additional NPU/GPU/ASIC): separates functions, activates as needed and allows capacity to scale without overloading the system. It can be integrated into the NVR or housed in an adjacent module working in tandem.
Separating the AI from the main CPU therefore reduces thermal risk and makes scaling by
project easier.
Exploitation: From Event to Insight
Operational Alerts
AI helps operators in the control room prioritise alerts such as intrusions, loitering,
abandoned objects or line crossings. The value lies in accuracy and minimising false alerts.
Efficient Searches with Metadata
AI generates structured metadata (object type: person/vehicle/bicycle/car, colour, attributes). Storing this alongside video allows near-instant searches without reanalysing
hours of footage. Examples:
- “Red van between 10:00 and 11:00 at the north entrance.”
- “Person with a backpack seen in the reception area today.”
Connectors to LLMs allow natural language queries when connectivity exists: “Show me all incidents involving bicycles this morning.”
Statistics and BI
Beyond security, metadata feeds business intelligence: footfall analysis, peak activity periods, object type distribution, seasonality… This informs decisions like increasing staff in
critical areas or redesigning access flows.
Legal Framework: Facial Recognition
Facial recognition raises legitimate concerns: mass surveillance versus utility in crime prevention, particularly in cases of terrorism or locating abducted persons. Under the current legislative framework, based on the EU AI Act and GDPR, operational deployment requires a specific national law that enables and regulates its use with appropriate safeguards. Until such a law exists, its use in real-time applications is practically prohibited.
For other AI applications in security, there is ongoing debate, with differences across European countries based on the interpretations of national data protection authorities.
Contact our team to learn how we can integrate artificial intelligence into your video
surveillance systems without changing your existing security infrastructure.
