Table of Contents
⚠ Disclaimer: This entry may be incomplete, out of date, or inaccurate. It is AI-maintained on a best-effort basis. Do not rely on it as a sole source — verify claims independently using the sources listed below.
Summary
Drone Tracking Datasets is an open collection of five multi-view drone-flight recordings — captured with 4 to 7 consumer-grade cameras per sequence — paired with high-accuracy 3D ground-truth trajectories from a real-time RTK positioning system (Fixposition). Unlike the individual open-source trackers elsewhere in this section, this is a benchmarking resource, not a detection/tracking tool itself: it exists to let researchers develop and validate multi-camera 3D trajectory-estimation methods against known-accurate ground truth, including the harder case of cameras that are unsynchronized and have unknown relative poses. The authors’ companion repository, mvus, implements a pipeline that reconstructs 3D object trajectories and camera poses from exactly this kind of unsynchronized, uncalibrated multi-camera setup.
Key Facts
- Author: CenekAlbl (GitHub)
- Type: Open dataset (+ companion reconstruction pipeline
mvus) - Status: Well-adopted relative to other entries in this section — 213 stars, 55 forks on the dataset repo; 89 stars, 18 forks, 200 commits on
mvus - Ground truth source: Real-time RTK positioning (Fixposition)
- Associated paper: “3D trajectory reconstruction from unsynchronized, uncalibrated camera networks” (arXiv:2003.04784) — describes reconstructing an airborne object’s 3D trajectory from cameras that may be unsynchronized, exhibit rolling-shutter distortion, and have unknown viewpoints
Dataset Contents
| Dataset | Cameras | Duration | Notes |
|---|---|---|---|
| 1 | 4 | ~2 min | Easy — short, slow flight |
| 2 | 4 | ~2.5 min | Easy — longer, faster flight |
| 3 | 6 | ~9 min | Medium difficulty, various velocities |
| 4 | 7 | ~7 min | High difficulty — fast motion, challenging conditions |
| 5 | 6 | ~10 min | Winter dataset — 3 drones simultaneously, snow-covered terrain |
Ground truth and supporting data vary by set: all five include 3D trajectories and per-camera calibration; datasets 3–5 add temporal synchronization data; datasets 3 and 5 include camera GPS locations; datasets 1–4 include 2D pixel-space labels; dataset 5 (the multi-drone winter set) additionally includes 3D orientation.
Why This Matters for This Section
This section’s steering priorities call out multi-sensor fusion and multi-camera systems that discriminate drones reliably, but nearly every vendor and open-source project here reports its own accuracy figures on its own test data — there is no shared, independently-collected ground truth to check claims against. This dataset is the one resource in this section built specifically to fill that gap: RTK ground truth is materially more trustworthy than GPS-log or manual-inspection “truth” data (the comparison method used in, e.g., TRAMIS’s own preliminary evaluation), and the multi-camera, multi-drone winter set (dataset 5) stresses exactly the close-proximity, adverse-condition scenario this section’s editorial focus prioritizes.
Limitations
- Dataset, not a tool: Using this requires separate detection/tracking/reconstruction code — either the authors’ own
mvuspipeline or a different one adapted to consume the provided calibration/sync data. - Small sample: Five sequences total, several only 2–2.5 minutes long — useful for method validation, not a large-scale training corpus.
- Consumer camera baseline: Captured with consumer-grade cameras rather than the industrial/purpose-built sensors used in this section’s commercial hardware entries — a reasonable proxy for DIY/open-source builds, less so for evaluating high-end fixed-site systems.
- License: A license file is present in the repository, but its specific terms were not confirmed for this entry — check directly before reuse.