- Jupyter Notebook 66.8%
- Python 33%
- Shell 0.1%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
- Add Numerical_Simulation: batch offline simulation - Add Numerical_Simulation_Realtime: online real-time simulation with two-stage design - Add UAV_Realtime: real-world UAV experiment with YOLO-TS and Kalman filter - Reorganize UAV experiment folder structure - Update README.md with new documentation - Add multi-YOLO version detection results (yolo8-12) - Remove deprecated files and clean up |
||
| dataset | ||
| Numerical_Simulation | ||
| Numerical_Simulation_Realtime | ||
| UAV_Realtime | ||
| .gitignore | ||
| README.md | ||
PoseFree-GeoLocator
Time Series for Flying Object Geographic Coordinate Recognition without Attitude Measurement and Reference Objects
This repository provides the codebase, dataset, and documentation for our work on localizing UAVs in 3D space using only time-series visual data — without relying on attitude sensors or spatial references.
📁 Repository Structure
PoseFree-GeoLocator/
├── Numerical_Simulation/ # Batch offline simulation
├── Numerical_Simulation_Realtime/ # Online real-time simulation
├── UAV_Realtime/
│ ├── UAV_3D_recognition/
│ │ ├── UAV_Realtime/ # Real-world online processing
│ │ └── calibration_in/ # Camera calibration files
│ └── YOLOv12/ # YOLOv12 detection model
└── dataset/
├── detections/ # YOLO detection results
└── trajectory/ # Ground-truth GPS trajectories
🧪 Numerical_Simulation (Batch Offline)
Validates the algorithm correctness using synthetic UAV trajectories in an ideal environment.
📁 Numerical_Simulation/
| File | Description |
|---|---|
main.py / main.ipynb |
Main pipeline: trajectory generation, 3D reconstruction |
core.py |
Triangulation, similarity transformation, error simulation |
geotracker.py |
GeoTracker class for camera pose estimation |
visualization.py |
3D plotting utilities for scene and results |
generate_sim_data.py |
Generate synthetic detection CSV |
analysis.ipynb |
Error sensitivity analysis (camera position, detection, scale) |
Pipeline: All frames → Batch initialization → Triangulation → Similarity transform → 3D trajectory
⚡ Numerical_Simulation_Realtime (Online Real-time)
Validates the real-time online algorithm using simulated frame-by-frame arrival.
📁 Numerical_Simulation_Realtime/
| File | Description |
|---|---|
main.py / main.ipynb |
Two-stage online pipeline demo |
geotracker.py |
GeoTracker with Stage1/Stage2 separation |
stream_feeder.py |
Simulates frame stream with FPS pacing |
benchmark_timing_sim.py |
Per-frame latency benchmarking |
plot_dual_phase_scene.py |
Visualize dual-phase scene |
analysis.ipynb |
Stage2 result analysis and error simulation |
Two-Stage Design:
- Stage 1 (Warmup): Batch initialization using first N frames (e.g., 300 frames)
- Stage 2 (Online): Strictly causal processing — each frame uses only current and past data
Key Metrics: Per-frame latency < 33ms @ 30 FPS → Real-time capable
🛩️ UAV Experiment (Real-World Validation)
Real-world validation using multi-camera video recordings and YOLO-based UAV detection.
🔹 UAV_Realtime — Online 3D Reconstruction
📁 UAV_Realtime/UAV_3D_recognition/UAV_Realtime/
| File | Description |
|---|---|
main.sh |
Main entry script |
1merge_detections.py |
Merge multi-camera YOLO detections |
2main.py |
Two-stage online geo-tracking pipeline |
3analysis.py |
Evaluate reconstruction vs. RTK ground truth |
geo_tracker.py |
GeoTracker with YOLO-TS post-processing & Kalman filter |
geo_tracker_ba.py |
Bundle Adjustment optimization |
core.py |
Geometry functions (triangulation, PnP, etc.) |
transformation.py |
Coordinate transformations |
visualization.py |
Trajectory comparison plots |
config_yolo*.json |
Configuration for different YOLO versions (8-12) |
batch_test.sh |
Batch testing across YOLO versions |
benchmark_timing.py |
Real-world timing benchmark |
Pipeline:
- YOLO detection → YOLO-TS temporal smoothing → Kalman filter alignment
- Stage 1: Initialize camera poses from warmup frames
- Stage 2: Online PnP-based 3D localization per frame
- Compare with RTK GPS ground truth
🔹 YOLOv12 Detection
📁 UAV_Realtime/YOLOv12/
- Forked and adapted YOLOv12 architecture for UAV detection
- Over 48,000 labeled UAV images under diverse conditions (twilight, backlight, buildings)
- Training scripts:
train.ipynb,run_train.sh - Inference:
predict.ipynb
📊 Dataset
🔹 dataset/detections/
YOLOv12 detection results for UAV flights recorded by three cameras.
Format (YOLO normalized):
<class_id> <x_center> <y_center>
🔹 dataset/trajectory/
Ground-truth 3D trajectories collected via RTK GPS.
Format (one line per frame):
X Y Z
📁 Video Dataset (Baidu Netdisk)
The full UAV detection dataset and video recordings are hosted on Baidu Netdisk:
This includes:
- Raw synchronized video sequences (3 views × 3 trajectories)
- YOLOv12 detection dataset: training/test images and annotations
💡 If the link becomes invalid, please contact the author for alternative access.
🚀 Quick Start
Numerical Simulation (Offline)
cd Numerical_Simulation
python main.py
Numerical Simulation (Real-time)
cd Numerical_Simulation_Realtime
python generate_sim_data.py --output analysis/simulated_detections.csv --gt-output analysis/ground_truth.csv
python main.py --csv analysis/simulated_detections.csv --gt-path analysis/ground_truth.csv
UAV Real-time Experiment
cd "UAV_Realtime/UAV_3D_recognition/UAV_Realtime"
bash main.sh
📄 Citation
If you use this dataset or code in your research, please cite:
@article{Yi2025PoseFreeGeoLocator,
title = {Time Series for Flying Object Geographic Coordinate Recognition without Attitude Measurement and Reference Objects},
author = {Junfan Yi, Ke-ke Shang and Michael Small},
journal = {In submission},
year = {2025}
}
🛡 License
This dataset and code are released for non-commercial academic use only.
For commercial use or redistribution, please contact the author.