The VCD4.2 system combines several state-of-the-art technologies into flexible and scalable architecture. It uses AI models for object detection and tracking to better understand what is happening on the road, while also running time-series models to forecast air quality levels. At the same time, it supports federated learning, allowing multiple devices to collaboratively improve models without sharing raw data, which is important for privacy and scalability. Since all these capabilities need to run on resource-limited devices, techniques like model compression are used to keep performance high while reducing computational load.
ITML plays an important role in making these technologies work efficiently at the edge. One of its main contributions is the development of a machine learning pipeline for air quality forecasting. This includes everything from data preparation to model training and evaluation, delivered as a reusable software library. A key part of this work has been optimizing the models so they can run on edge devices, where ITML identified ONNX compression as the best way to reduce resource usage without losing accuracy.
In addition, ITML has worked on integrating federated learning into these pipelines, enabling distributed training across devices. By adapting existing frameworks and selecting suitable strategies like FedProx, ITML ensures that models can be trained reliably even in heterogeneous environments. The team has also added features to improve performance and support privacy, making the solution more robust for real-world use.
Another important contribution is in monitoring the health of the RSPU devices themselves. ITML has developed anomaly detection models that analyze system metrics such as CPU, GPU, and memory usage to detect unusual behavior. After testing different approaches, the team focused on models that offer a good balance between accuracy and efficiency, and continues to refine them using real data from the project.
As the project moves forward, ITML is also involved in integrating these components into the full system and validating their performance in realistic conditions. This includes benchmarking how well the models perform on edge devices and ensuring that all parts of the system work smoothly together.
Overall, VC4 and VCD4.2 show how edge AI can support smarter and more efficient mobility systems. Through its work on model optimization, federated learning, and system monitoring, ITML contributes to building solutions that are not only advanced but also practical and ready for real-world deployment.
Blog signed by: ITML team