Geo-ORBIT: A Federated Digital Twin Framework for Scene-Adaptive Lane Geometry Detection

Paper | Rei Tamaru, Pei Li, and Bin Ran | University of Wisconsin–Madison

---

Geo-ORBIT (Geometrical Operational Roadway Blueprint with Integrated Twin) is a unified framework that integrates real-time lane detection, federated learning, and digital twin synchronization. It is designed to support active traffic management, infrastructure monitoring, and real-time scenario testing without relying on centralized data collection.

At the core of Geo-ORBIT is FedMeta-GeoLane, a federated meta-learning-based lane detection model that adapts to scene-specific geometry using only vehicle trajectory data. By preserving privacy and reducing bandwidth, this system enables scalable deployment across diverse roadside camera environments.

Figure: Qualitative comparison across multiple locations. The camera at Park is treated as an unseen location for Meta-GeoLane and FedMeta-GeoLane. Blue lines represent trajectory contours, and each lane is colored accordingly in the same lane group.

System Architecture

Geo-ORBIT is composed of three modular and interconnected processes:

• Detection Process
  Roadside cameras capture traffic video, from which vehicle trajectories are extracted and projected to GPS space.

• Service Process
  The FedMeta-GeoLane model infers lane geometries from trajectories using adaptive parameters, refined through meta-learning and weak supervision (e.g., OpenStreetMap).

• Simulation Process
  Detected lanes are synchronized with SUMO and CARLA to create a high-fidelity, real-time Digital Twin that supports traffic flow rendering and scenario replay.

Installation

Geo-ORBIT works with Python 3.10+ and Pytorch 2.5.1+.

Clone the repository

git clone https://github.com/raynbowy23/FedMeta-GeoLane.git
cd FedMeta-GeoLane

Create uv environment with Python >= 3.10.

uv init
uv sync
# Linux
source .venv/bin/activate

Install dependencies

pip install -r requirements.txt

Set up SUMO

sudo apt-get install sumo sumo-tools sumo-doc
export SUMO_HOME="/usr/share/sumo"

Prepare data directories

mkdir -p dataset/511video dataset/511calibration
mkdir -p results logs

How to Use

Quickstart

Run the complete federated learning pipeline:

bash run.sh

This executes the defualt configuration with federated meta-learning on historical data.

Basic Usage

1. Prepare camera calibration data.

• Place GPS calibration points in dataset/511calibration.
• Format: camera_name.csv with columns: pixel_x, pixel_y, latitude, longitude.

2. Add camera locations

• List camera names in dataset/camera_location_list.txt.
• One camera name per line.

3. Map Data Selection

• Extract corresponding OpenStreetMap data using python osmWebWizard.py in ./LaneDetection/osm_extraction. Alternatively, download it online. (e.g. https://www.openstreetmap.org/#map=17/43.034678/-89.426753)
• Change extracted folder name to camera_name. (e.g. US12_Park)
• Extract osm.net.xml.gz.
• Run netconvert -s osm.net.xml --plain-output-prefix osm, and convert to plainXML osm.nod.xml and osm.edg.xml.

4. Map Data Preprocess

• Open osm.net.xml in local SUMO.
• Trim it to have only target road (Remove unnecessary part).

5. Run Lane Detection

python main.py --T 60 --is_save --model federated

Advanced Configuration

Federated Learning

python main.py --model federated --T 60 --is_save --skip_continuous_learning --use_historical_data

Note: include --skip_continuous_learning and --use_historical_data if you want to skip video detection part for test, which reduce a lot of time.

Meta Learning (Training on Single Camera)

python main.py --model meta --T 60 --is_save --skip_continuous_learning --use_historical_data

Baseline (Fixed Parameters)

python main.py --model federated --T 60 --is_save --skip_continuous_learning --use_historical_data

Key Parameters

• --T: Time interval for data collection (seconds)
• --model: Learning approach (federated, meta, baseline)
• --is_save: Save intermediate results and visualizations
• --use_historical_data: Use pre-processed trajectory data
• --skip_continuous_learning: Skip real-time detection
• --lambda_thres: Vehicle count threshold for learning cycles

Output Structure

results/
└── 511video/
    └── model_name/
        ├── camera_location/
        |   ├── figures/                            # Visualizations
        |   ├── preprocess/                         # Processed data
        |   ├── sumo/                               # SUMO network files
        |   ├── pixel/                              # Mid opeartion visualization
        |   └── federated_trajectory_clustering.csv # Only for federated learning
        └── training_results/                       # Final results

Simulation Integration

Look at the detailed process to create digital twin integration.

SUMO Integration

1. Generate SUMO network

bash convert_sumo2xodr.sh camera_name

2. Run trajectory synchroniztaion

python OpenDriveConversion/det2sumo_sync.py \
    --camera_loc camera_name \
    --dataset_path ./dataset/

CARLA Integration

1. Start CARLA server

# In CARLA directory
make launch
# OR
./CarlaUE4.sh

2. Load generated map

python OpenDriveConversion/openDrive2Carla.py \
    --map_file results/camera_name/sumo/camera_name

3. Run co-simulation

python OpenDriveConversion/run_synchronization.py osm.sumocfg --sumo-gui

(Short Summary) FedMeta-GeoLane: Federated Meta-Learning Lane Detection

FedMeta-GeoLane treats each roadside camera deployment as a unique task. A shared meta-learner predicts optimal detection parameters using context features like vehicle speed and trajectory distribution. Key highlights include:

• Black-box meta-learning: No need for gradient flow through detection pipeline
• Federated optimization: Local training with privacy-preserving aggregation
• Scene adaptation: Immediate configuration for unseen locations

Figure: Overview of Knowledge-Based Lane Detection Algorithm. (a) Video detection and trajectory projection to GPS coordinates. (b) Lane center estimation using histogram analysis. (c) Lane-based trajectory clustering with KMeans. (d) Lane geometry estimation and boundary generation.

Compared to baseline and centralized models, FedMeta-GeoLane reduces geometric error by over 50% in unseen locations while achieving a 98% reduction in communication cost.

Performance Summary

Lane Detection Accuracy

Table: Validation Loss Component Comparisons of Each Model on Seen and Unseen Locations

Model	Consistency Loss (m) ↓	Geometry Loss (m) ↓	Centerline Error (m) ↓	Lane Count Error ↓	Total Loss ↓
Seen
Baseline	5.45	15.12	6.78	5.00	77.84
Meta-GeoLane	7.04	11.76	4.73	2.67	12.16
FedMeta-GeoLane	0.0	2.65	3.16	2.67	6.94
Unseen
Meta-GeoLane	18.51	105.35	34.60	12.00	69.61
FedMeta-GeoLane	0.0	12.82	21.39	12.00	32.38

Figure: Qualitative comparison across multiple locations. The camera at Park is treated as an unseen location for Meta-GeoLane and FedMeta-GeoLane. Blue lines represent trajectory contours, and each lane is colored accordingly in the same lane group.

Transmission Cost Analysis

Table: Bit Per Second Performance Comparison for All Clients

Parameters	Baseline	Meta	Federated Meta
Model size (MB)	0	0	0.2
Clients	4	4	4
Rounds	1	20	20
Model Upload (MB)	0	0	0.01
File Upload (MB)	427.3	427.3	5.6
Download (MB)	0	0	0.018
BPS (Mbps)	3418	3418	47.2

Digital Twin Integration

Geo-ORBIT connects real-world observations to virtual testbeds using a synchronized SUMO–CARLA pipeline:

• GPS-aligned trajectories enable accurate replay in simulation
• Supports scene-level validation, vehicle re-routing, and visual analytics
• (Will be implemented) Extendable to multi-scenario environments with dynamic overlays (e.g., vegetation, accidents, road closures)

Figure: Digital twin synchronization with SUMO and CARLA at multiple locations employing real-time vehicle trajectory.

Citation

If you use this work in your research, please cite:

@article{tamaru2025geo,
  title={Geo-ORBIT: A Federated Digital Twin Framework for Scene-Adaptive Lane Geometry Detection},
  author={Tamaru, Rei and Li, Pei and Ran, Bin},
  journal={arXiv preprint arxiv:2507.08743},
  year={2025}
}