D4._AI-driven Analysis Onsite for Monitoring Combined with Advanced Representations v1

This deliverable presents the first technical outcomes of WP4 Intelligent Edge Computing and Monitoring for Safer and Well-maintained Roads during the initial reporting period (M03-M18). It summarises the development of real-time on-site computing solutions that integrate advanced AI-driven analysis, behavioural understanding, environmental monitoring and 3D scene representation. The work conducted in this period establishes the foundation for the intelligent monitoring components that will be integrated into the iDriving pilots and digital ecosystem in subsequent phases. Overview of Objectives The primary objective of D4.1 is to deliver preliminary versions of the core WP4 modules that enable real-time processing and interpretation of road infrastructure, road-use behaviour and environmental conditions. These components support: • On-site visual analysis through embedded AI models. • Behavioural detection of unsafe or non-compliant road-user actions. • Automated identification of road defects and maintenance needs. • High-resolution environmental monitoring relevant to road safety. • AI-based 3D reconstruction and representation for improved situational understanding. These achievements provide the first operational capabilities needed to support proactive safety measures, predictive maintenance and dynamic updates of digital twins. Key Technological Advancements and Innovations Several technological innovations have been delivered in this first period: • Real-time edge analytics pipeline integrating detection, tracking and video stabilisation, optimised through quantisation and parallel processing to run efficiently on Jetson and GPU platforms. • Behavioural analysis framework capable of transforming low-level detections into interpretable traffic violations, including zebra-crossing misuse, improper lane usage and failure-to-yield, fully integrated into the Kafka streaming ecosystem. • Automated road defect detection using YOLOv11 models, enhanced with data augmentation and explainability (EigenCAM), achieving near–real-time performance and robust detection across diverse conditions.
Environmental monitoring workflow with containerised WRF/WRFDA systems, real-time data ingestion pipelines and configuration of highresolution domains for Thessaloniki and Karlovac. • Advanced 3D scene reconstruction using 3D Gaussian Splatting (Splatfacto), validated on benchmark and real-world UAV datasets, and enabling semantic enrichment from defect detection outputs. Together, these technologies deliver a multi-layered, AI-enhanced monitoring system capable of supporting safety-critical decision-making. Relevance to overall Project Goals The results achieved in D4.1 directly contribute to iDriving’s mission of improving road safety and infrastructure quality across urban and rural environments. The components developed in WP4 strengthen: • Early detection of risks, including unsafe behaviour, surface degradation and hazardous environmental conditions. • Efficient and proactive maintenance, supported by automated defect identification and integration with the digital twin developed in WP5. • Smart V2I-enabled safety services, thanks to real-time on-site analytics that reduce latency and bandwidth demand. • Pilot readiness, as the v1 modules provide the technical basis for integration, validation and refinement in WP6. Overall, this deliverable marks a significant step toward the deployment of intelligent, data-driven and safety-oriented infrastructure management within the iDriving ecosystem