GeoLaneRep: Behavior-Grounded Lane Representation Learning for Traffic Digital Twins
GeoLaneRep has introduced a cutting-edge framework designed to enhance lane representations within traffic digital twins, focusing on dynamic functional semantics rather than static geometry. This innovative approach integrates lane geometry, vehicle trajectories, and operational attributes into a unified semantic model using data from multiple cameras. The framework utilizes advanced training methodologies, including contrastive cross-camera alignment and temporal anomaly detection. Evaluations conducted with 16 roadside cameras and 132 lanes showed remarkable accuracy, registering a lateral-rank error of 0.004, an edge-role F1 score of 1.000 for zero-shot cross-camera matching, and an AUROC of 0.991, aimed at improving traffic control systems.
Key facts
- GeoLaneRep is a behavior-grounded lane representation learning framework.
- It jointly encodes static lane geometry, observed vehicle trajectories, and operational descriptors.
- The encoder uses contrastive cross-camera alignment, auxiliary role supervision, and temporal anomaly detection.
- Tested on 16 roadside cameras and 132 lanes.
- Achieves lateral-rank error of 0.004.
- Achieves edge-role F1 of 1.000 in zero-shot cross-camera matching.
- Achieves AUROC of 0.991.
- Published on arXiv with ID 2605.01901.
Entities
Institutions
- arXiv