Framework and Multi-modal Dataset for Roadwork Zone Detection and Geo-localization

Autonomous vehicles rely heavily on high-definition (HD) maps for navigation. However, these maps are static and often fail to reflect transient changes like temporary roadwork zones. This gap in situational awareness poses significant safety risks and hinders efficient route planning. While previous research has focused on detecting roadwork objects, there has been a lack of publicly available, multi-modal datasets that support both detection and precise global geo-localization of these zones.

The authors propose the Roadwork Zone Detection and Geo-localization (RZDG) dataset, which provides a comprehensive resource for autonomous driving research. The dataset includes both real-world data collected in Ingolstadt, Germany, and simulated data generated via the CARLA simulator. It features multi-modal sensor inputs—including camera, LiDAR, and GPS/IMU—and provides annotations for image semantic segmentation, 3D object detection, and object geo-localization. By including both real and simulated environments, the dataset allows researchers to evaluate model performance across diverse conditions and configurations.

To address the challenge of geo-localization, the authors introduce a tracker-based RZDG pipeline. This framework extends the AB3DMOT approach to transform object positions from a local vehicle-centric coordinate system into a global coordinate system. The pipeline consists of three main stages: 3D object detection, multi-frame tracking to form object tracklets, and coordinate transformation using GPS/IMU data. This approach effectively filters out redundant detections and provides a unique, accurate global position for each roadwork element.

By providing a standardized benchmark and a functional pipeline, this work enables the integration of transient roadwork information into Local Dynamic Maps (LDMs). This capability is essential for enhancing the safety and reliability of autonomous driving systems in complex, changing urban environments. The inclusion of simulated data also lowers the barrier to entry for developing and testing new perception algorithms.