Edge-Native Federated Lidar Analytics for Robust Guardrail Monitoring in Rural Environments

Artificial intelligence can improve the monitoring and maintenance of roadside safety infrastructure in rural environments, where connectivity, compute, and energy constraints limit cloud-centric approaches. We present a LiDAR-based, edgenative framework for guardrail presence detection, observabilityaware typology inference, and metric deformation estimation. Inference runs fully on resource-constrained edge nodes, while learning is confined to an upstream semantic disambiguation module enabling parameter-efficient, privacy-preserving federated adaptation orchestrated by the Flexible Learning and Evaluation Toolchain (FLeeT) platform. Downstream reasoning, geometric modeling, and deformation analysis are deterministic, enforcing conservative operation under partial observability. A lightweight control plane trades accuracy for energy by adapting inference frequency, processing depth, and participation in federated rounds. Experiments on off-road LiDAR data show repeatable deformation indicators with conservative reporting, supporting scalable deployment across the edge-cloud continuum.