From asphalt to data: how RENEW's Edge-Cloud architecture works

Continuously monitoring a city's road surfaces means collecting enormous amounts of data every day: high-frequency images, vibrations, position. Sending everything to a central server would be costly, slow and disrespectful of privacy. That is why the first design decision of RENEW — made by Key4 together with the Department of Computer Science at the University of Bari — was to move beyond the "everything in the Cloud" model.

Why the Cloud alone is not enough

In a cloud-centric architecture the vehicle is just a recorder: it acquires and transmits. On a fleet of urban vehicles this approach runs into three concrete limitations: the bandwidth required to transfer continuous video streams, the latency that delays alerts, and the handling of images that may capture people and licence plates.

RENEW reverses the perspective: every vehicle becomes an intelligent node in the network. Time-critical processing happens on board; the Cloud handles the heavier analysis, aggregation over time, and decision support.

Four layers, one chain

The architecture is organised across four layers:

The distress_event: only what matters travels

The key concept of the architecture is the distress_event: when the Edge node detects a possible defect, it generates a georeferenced event containing only the information relevant to maintenance — starting with position and type of detected distress.

This event-driven approach drastically reduces the bandwidth required. The severity estimate is always accompanied by confidence and uncertainty values, because the manager needs to know how much to trust each report.

The Data Collection Device: a small data centre on two wheels

The DCD was designed by Key4 and IAT to work in real conditions, on a vehicle exposed to vibrations, rain and dust:

Why the Jetson Orin NX

The choice of computing module came from a comparative test campaign across different GPU, embedded and data-centre platforms, using RF-DETR object detection models. On the Jetson Orin NX 16 GB, inference times ranged from approximately 9 to 40 milliseconds per image, depending on model size and power mode — values compatible with real-time processing. The previous generation, Jetson Xavier NX, reached only 65 to 93 milliseconds.

The result is a balanced trade-off between performance, power consumption, size and cost — essential for a device intended for fleet-wide deployment.

Built for the road: network resilience

An urban vehicle passes through tunnels, blind spots and areas with weak coverage. The DCD handles these situations with local buffering and store-and-forward logic: events are stored on board and transmitted as soon as the connection allows, according to priority and retention rules. If a component stops responding, the firmware switches to degraded mode instead of stopping.

Modern software, from Edge to Cloud

Both on the vehicle and in the Cloud, services are containerised and orchestrated with K3s, a lightweight Kubernetes distribution. This allows AI models and services to be updated uniformly across the entire fleet, and the Cloud infrastructure to scale from a single node to multiple nodes as vehicles and territories served grow.

In the Cloud, a georeferenced data warehouse stores every event with its provenance: which device generated it, where, when and with which version of the AI model. This is the foundation for building a reliable history of deterioration and for verifying decisions over time.

From technology to decision

The final layer is the one the manager sees. The Application Layer was designed starting from user profiles: 11 use cases including network condition, deterioration trends, alerts, maintenance plans, reports and notifications. The goal is that the technical office of a municipality should not have to interpret raw data, but receive clear indications of where and when to act.

Next steps

The design of the architecture and the Data Collection Device is complete. Functional and integration testing of the prototypes, implementation of the software layers and model training are currently underway. The final phase of the project, by January 2027, will involve validating the entire chain in an urban pilot case.

For Key4, this is the application to public infrastructure of expertise built up over years of work on IIoT, Edge Computing and MES systems for industry.

Project "RENEW – Sensor-based and data-driven framework for safe and efficient PMS" – CUP B67H24001820007 – Practice code XF4UNN8. PR Puglia ERDF ESF+ 2021-2027 (ERDF fund) – Action 1.5 "Interventions for the strengthening of the regional innovation system and support for collaboration between companies and research structures" – "Reti" Call. Project co-funded by the European Union.