- Title: uDopplerTag: CNN-Based Drone Recognition via Cooperative Micro-Doppler Tagging
- ArXiv: https://arxiv.org/abs/2601.08042
- DOI: https://doi.org/10.48550/arXiv.2601.08042
- Authors: O. Yerushalimov, D. Vovchuk, A. Glam, P. Ginzburg
- Published: 2026-01-12 (arXiv v1)
- Module state: ALMOST (phase-1 code + planning in progress)
- Verification state: VERIFIED-WITH-RISKS
- Public code repo from paper authors: MISSING
- Public dataset release from paper: MISSING
- Independent citations signal (OpenAlex): cited_by_count = 0 (as checked on 2026-04-10)
| Model | Size | Source | Path on Server | Status |
|---|---|---|---|---|
| YOLO26 base (planned) | TBD | Internal ANIMA standard | /Volumes/AIFlowDev/RobotFlowLabs/datasets/shared/models/yolo26/ | MISSING |
| uDopplerTag CNN (this module) | Small (<10 MB est.) | Trained in-module | /Volumes/AIFlowDev/RobotFlowLabs/datasets/wave10_staging/project_radar/checkpoints/ | PLANNED |
| Dataset | Size | Split | Source | Expected Path | Status |
|---|---|---|---|---|---|
| Internal Mega UAV | ~1.8M samples (internal target) | train/val/test (to define) | Internal | /Volumes/AIFlowDev/RobotFlowLabs/datasets/wave10_staging/mega_uav/ | MISSING/UNKNOWN |
| VisDrone | ~10K images (paper-external adaptation data) | standard splits | https://github.com/VisDrone/VisDrone-Dataset | /Volumes/AIFlowDev/RobotFlowLabs/datasets/wave10_staging/visdrone | FOUND |
| UAVDT | ~80K frames | train/test | http://uavdt.uavbenchmark.com/ | /Volumes/AIFlowDev/RobotFlowLabs/datasets/wave10_staging/uavdt | MISSING |
| DroneVehicle | varies | train/val/test | public benchmark | /Volumes/AIFlowDev/RobotFlowLabs/datasets/wave10_staging/dronevehicle | MISSING |
| SeaDronesSee | varies | train/val/test | public benchmark | /Volumes/AIFlowDev/RobotFlowLabs/datasets/wave10_staging/seadronessee | MISSING |
| Paper indoor tag data (43 codes) | private | indoor folds | paper proprietary | N/A | UNAVAILABLE |
| Paper outdoor tag data (7 codes) | private | outdoor folds | paper proprietary | N/A | UNAVAILABLE |
| Parameter | Value | Paper Evidence |
|---|---|---|
| conv blocks | 3 blocks (32/64/128 filters, 3x3) | Section IV + Appendix 1 |
| activation | ReLU | Section IV + Appendix 1 |
| pooling | MaxPool 2x2 after each conv block | Section IV + Appendix 1 |
| dropout | p = 0.5 | Section IV + Appendix 1 |
| FC hidden | 128 | Section IV + Appendix 1 |
| optimizer | Adam | Section IV |
| learning rate | 0.001 | Section IV |
| loss | Cross-entropy | Section IV |
| batch size | 32 | Section IV |
| split (outdoor) | 80/10/10 | Section IV |
| Scenario | Metric | Paper Value | Our Target |
|---|---|---|---|
| Indoor 43-class synthetic-SNR | Accuracy at >9 dB | ~99% | >=95% on surrogate/internal equivalent |
| Indoor 43-class synthetic-SNR | Accuracy at 0 dB | low/unreliable (~10%) | Similar degradation trend |
| Outdoor 7-class selected codes | Useful classification down to 7 dB | Reported robust | Match trend on internal field-like data |
| Environment | Role |
|---|---|
| Mac Studio (Apple Silicon) | MLX development + smoke training |
| Vast.ai RTX 4090 | CUDA training + larger runs |
- Dataset checker:
bash scripts/download_data.sh --check - Download gate (disabled by default): set
RADAR_DOWNLOAD_GATE=OPENbefore--download
- The paper's own radar data is not public; this project implements a faithful method replica and adapts to available/internal data.
- Any metric claims against the paper remain provisional until internal radar-equivalent data is assembled.