Real-Time Smart Spaces: Vision Language Models, VLMs.

April 02, 2026

The city is a system of high semantic density, in which every event acquires meaning only in relation to its context, its timing and the interaction between multiple actors. In this scenario, analysing urban video streams does not simply mean recognising objects, but extracting the dynamism of the observed reality in real time. It is precisely by working on real-world cases that it becomes clear that the real turning point lies in the shift towards a computational vision capable of reasoning.



Shared intelligence

The architectural and computational breakthrough represented by Vision Language Models (VLMs) marks precisely this transition. These models do not merely extract visual features, but integrate linguistic representations that enable observations to be structured logically, introducing elements such as causality, internal consistency and uncertainty management. Highly specialised VLMs, each trained on a specific, optimised semantic domain. It is an approach which, when tested on real-world infrastructure, enables greater accuracy to be achieved by reducing the scope for ambiguity.



Reasoning in the Flow

A key factor is the temporal dimension. Events in an urban setting cannot be captured in a single frame, but emerge as patterns distributed over time. For this reason, the ability to analyse video sequences, rather than static images, is a fundamental requirement for any system of architecture that aims to operate reliably in the real world.

Added to this is a further layer of complexity: the need to ensure that inferences are robust and reliable. Multimodal models are notoriously prone to ‘hallucination’, i.e. the generation of information not supported by the observed data. In an urban context, it therefore becomes necessary to introduce control mechanisms such as logical consistency constraints, explicit confidence estimates and temporal anchoring of observations—elements that play a central role when architectures are required to operate in real time.

Finally, there is a strategic dimension that goes beyond mere technical performance: that of data sovereignty. In applications involving public space and analytics, the ability to keep the entire processing cycle within controlled infrastructures is not merely an architectural choice, but ensures regulatory compliance, security and trust. It is an industrial methodological heritage that K2K® has built and consolidated over time, translating these principles from theory into complex and highly variable operational environments.

It is therefore not merely an incremental improvement on existing technologies, but a far more profound paradigm shift. Artificial intelligence applied to cities is moving away from a logic of passive observation towards a form of structured interpretation of reality, in which efficiency, specialisation and accountability become inseparable elements.

In this context, thanks to the extraordinary growth in computing power, unimaginable just a few years ago, the key transition lies in designing architectures capable of adapting intelligently to the domain in which they operate.

Marco De Paoli