Infrastructure-based localization across scales

RWTH Publications (RWTH Aachen)

Abstract

Road infrastructure faces increasing traffic volumes, aging structures, and limited maintenance resources, calling for improved traffic management, road safety, and resource allocation. Digital twins of road systems address this challenge by providing real-time virtual representations that combine live data with predictive models for traffic management, incident detection, and infrastructure health monitoring. To enable this integration, digital twins require an accurate machine-readable representation of all road users within a designated area. Infrastructure-based localization is a key enabler of such digital representations, providing an area-centric view of road users that is independent of any individual vehicle. This dissertation investigates an infrastructure-based localization system for Intelligent Transportation Systems (ITS) and its integration into a digital twin of the road system. The objective is to estimate, in real time, the dynamic state, physical dimensions, class, and existence probability of all relevant road users within a fixed area of interest, under the computational and power constraints of roadside embedded hardware. The objective is achieved through a localization pipeline that fuses three complementary data sources: in-road sensing based on a sensitive surface layer embedded in the pavement, roadside lidar sensors, and vehicle self-state estimates transmitted to the infrastructure. The pipeline performs object-level fusion, merging uncertainty-aware object lists into a single global object list in a common map frame. For in-road sensing, a three-stage approach detects wheel contact patches based on load measurements, tracks wheels over time, and assembles them into vehicle hypotheses. To the best of our knowledge, this approach represents the first end-to-end implementation of vehicle detection and tracking based solely on in-road load sensors. On the roadside, a model-based lidar tracker is developed that relies on explicit geometric and motion models together with probabilistic filtering, rather than black-box learning methods, enabling deployment on low-power embedded processors. All components are implemented within a unified software framework with standardized interfaces. The proposed system is evaluated at three scales: (i) a simulated urban intersection, (ii) a small-scale laboratory testbed, and (iii) a full-scale deployment in a real parking lot with high-precision reference localization. Across all three scales, the fused global object list achieves real-world equivalent position accuracies on the order of 0.3 m root-mean-square error and existence estimation precision and recall above 96%. The results lead to two main conclusions. First, model-based infrastructure localization pipelines with uncertainty representation provide a strong foundation for digital twins of road systems while meeting real-time constraints on embedded roadside platforms. Second, small-scale laboratory testing reproduces the behaviour and performance of full-scale deployments, revealing integration issues and performance trends prior to costly real-world testing.

Authors 1

1. Simon Schäfer corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen