Multi-Modal UAV Detection, Classification and Tracking Algorithm -- Technical Report for CVPR 2024 UG2 Challenge
Paper • 2405.16464 • Published
Datasets:
Search is not available for this dataset
image image |
|---|
End of preview. Expand in Data Studio
MMAUD
A Comprehensive Multi-Modal Anti-UAV Dataset for Modern Miniature Drone Threats
Camera · LiDAR · mmWave radar · Audio · Leica ground truth
MMAUD brings complementary sensing modalities together for drone detection, classification, tracking and trajectory estimation. This repository provides the recordings, extracted data, calibration resources and practical guides in one public download location.
📄 Paper · PDF · 🌐 Project · 💻 GitHub · 📦 Browse data
Overview · Versions · Download · Quick start · Baselines · Calibration · Citation
Camera images, LiDAR and radar point clouds, and audio from the original MMAUD project. Figure source.
Dataset at a glance
✅ Upload complete and verified — approximately 367.86 GB, with original filenames and public downloads.
| Resource | Included in this release |
|---|---|
| Sensor recordings | 12 ROS1 bags with bz2 chunk compression |
| Ground truth and calibration recordings | 13 bags |
| Extracted V1 data | 76,786 individual files: 33,200 PNG images + 43,586 NumPy arrays |
| Other source files | 43 files: 28 WAV, 10 extraction scripts, 3 CSV, noise.bin and the original point-cloud ZIP |
| Calibration resources | 20 files, including camera parameters, reports, CAD/STL and the full 7.24 GB calibration archive |
| Getting started | 11 baseline code entries, a YOLOv5 starting point, 18 common-issue answers and an index of 69 GitHub issues |
All 76,829 supplementary source files are included. Images and NumPy arrays can be browsed and downloaded individually, preserving their sequence, modality and timestamp filenames. Sizes on this page use decimal GB.
Dataset versions
| Version | Role | Scenario difficulty | Typical localization error |
|---|---|---|---|
| V1 | The official dataset contribution of the ICRA 2024 paper | Relatively easier | 0.3–1 m |
| V2 / V3 | CVPR 2025 workshop challenge extensions | More challenging | 1–10 m is common |
These error ranges are practical guidance from the dataset maintainer, rather than scores for a single benchmark method. Results vary with the method, sequence and evaluation setup.
The source collection groups V2 and V3 under V2_V3/. The maintainer's filename clarification assigns s1 to V2, s2 to V3, and m3e to V2. Original filenames and directory grouping are preserved.
Download
Choose the format that fits your workflow: ROS bags for replaying sensor streams, PNG/NPY files for direct processing, or WAV files for audio experiments.
| Data | Download / browse | Contents |
|---|---|---|
| V1 sensor recordings | V1/rosbag | Avata, M300, Mavic2, Mavic3, Pham4 |
| V1 ground truth | V1/ground_truth | Five corresponding GT bags |
| V1 extracted files | V1/filesystem_format | Images, LiDAR, radar and ground-truth arrays |
| V2 / V3 sensor recordings | V2_V3/rosbag | m300s1, m300s2, m30s1, m30s2, m3e, p4s1, p4s2 |
| V2 / V3 ground truth | V2_V3/ground_truth | Seven GT bags and calibration.bag |
| Extracted audio | V2_V3/audio | Four WAV files per sequence; 28 in total |
| Original extraction scripts | V1 scripts · V2 / V3 scripts | Ten source scripts; usage notes in Common Issues |
| Supplementary files | supplementary | Interpolated CSVs, noise.bin and original 点云.zip |
| Calibration and sensor geometry | calibration | Camera YAMLs, reports, raw calibration archive and CAD |
Large recording: the 54.74 GB m3e.bag exceeds the Hub's 50 GB per-file limit, so it is provided as three ordered byte parts and a .bag.parts.json manifest. The reassembly script restores the original bag byte for byte and checks its SHA-256. Follow the download and reassembly guide.
See File layout and source coverage for per-sequence counts and source-to-repository mapping. The separate challenge train.zip, val.zip and 59-sequence test.zip are outside this source collection; their references are in Common Issues.
Quick start
Download a selected sequence
Install the Hugging Face Hub client:
pip install -U huggingface_hub
This example downloads the extracted V1 Mavic3 sequence into a local MMAUD/ folder:
import os
os.environ["HF_HUB_DISABLE_XET"] = "1"
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="Yuan-Shenghai/MMAUD",
repo_type="dataset",
allow_patterns="V1/filesystem_format/Mavic3/*",
local_dir="MMAUD",
)
Use V2_V3/audio/* for audio or V1/rosbag/* for the V1 sensor bags. A single file can also be downloaded from its page using the Download button. The full Mavic3 extracted sequence is approximately 13.04 GB.
Inspect and replay a ROS bag
The sensor recordings use ROS1 bag format with bz2 chunk compression. With ROS Noetic installed, inspect a downloaded bag, decompress it for playback, then replay it:
rosbag info Mavic3.bag
rosbag decompress Mavic3.bag
rosbag play Mavic3.bag
Allow space for decompression: the five V1 sensor recordings occupy approximately 77.88 GB compressed and 298.29 GB uncompressed.
Sensor setup
The collection platform combines two cameras, two 3D LiDAR sensors, mmWave radar and four audio streams, with Leica tracking for ground truth.
Original sensing rig, Leica tracking and example detection, tracking and classification tasks. Figure source. Sensor geometry and transform usage are documented in the calibration guide.
Baseline code
💻 Start with a baseline. The table collects all 11 methods from the MMAUD GitHub baseline list. The baseline guide covers installation, input preparation and training/evaluation entry points. Paper references and BibTeX are listed after the MMAUD citation.
| Method | Sensor input | Code |
|---|---|---|
| ResNet | Images | ResNet comparison model |
| Darknet | Images | Darknet comparison model |
| VorasNet / DroneChase | Audio | VoraNet comparison model |
| AV-PED | Audio + images | AV-PED comparison model |
| Audionet | Audio | AudioNet comparison model |
| AV-FDTI | Audio + images | AV-FDTI repository |
| TAME | Audio | TAME repository |
| AAUTE | Audio | AAUTE repository |
| UnLiDAR | LiDAR | LiDAR drone detection pipeline |
| MMUDCT | LiDAR + images | Multi-Modal-UAV repository |
| AV-DTEC | Audio + images; LiDAR-derived supervision | AV-DETC repository |
For 2D image detection, use YOLOv5 with the annotation subset linked by the MMAUD project. The baseline guide also links the related AV-PedAware tutorial.
These entries link to the authors' code and released resources. Check each method's preprocessing and data split before comparing results; this repository's release verification covers files and documentation, not a new training reproduction.
Calibration and coordinate frames
🧭 Calibration overview · Transforms and ROS TF · Initial radar-to-camera matrix
The calibration collection includes official left/right fisheye YAML parameters, TXT/PDF reports, CAD drawings, individual STL parts and the complete fisheye calibration archive, including the original calibration bags. The cameras use the omni + radtan model at 1280 × 960.
Use the supplied calibration and CAD geometry as an initial calibration, then refine the inter-sensor transforms on your calibration data. The CAD-derived radar-to-left-camera matrix states its axis and origin assumptions; camera intrinsics, installation geometry and cross-sensor transforms are documented separately.
LiDAR, camera and Leica data share a common time reference while retaining their own spatial frames. V1 needs trajectory-based spatial alignment to Leica; V2 includes an added calibration recording. Use time-paired trajectories to estimate the spatial rotation and translation for V1. See the version-specific alignment guide for the workflow and V1 export-axis conventions.
Verified point/cluster annotations are not included. Ground-truth search regions can support candidate annotation generation; see Common Issues for the current annotation scope.
Verification and documentation
🔎 Every release file has a recorded size and SHA-256. All filesystem files were compared individually with public repository metadata. One sample from each of the 25 sequence/modality folders was also downloaded in full anonymously and hashed. The calibration ZIP passed a full-member CRC check.
| Resource | What it provides |
|---|---|
| SHA256SUMS.txt | Release-wide checksums, excluding the checksum file itself |
| RELEASE_VERIFICATION.json | Counts, byte totals, verification methods and completion record |
| Source coverage | Mapping from source files to public paths, with explicit exclusions |
| Filesystem checksums | Per-file hashes for all 76,786 extracted PNG/NPY files |
| Common Issues | 18 practical answers on downloads, timing, calibration, annotations and use |
| Complete issue index | 69 GitHub issues and linked discussions, reviewed on 23 October 2026 |
| Baseline guide | Code links, configurations and preparation steps |
| File layout | Directory structure, sequence counts and source coverage |
License and citation
The dataset is released under CC BY-NC-SA 4.0, following the original MMAUD dataset documentation.
📚 Please cite the ICRA 2024 paper when using MMAUD:
Shenghai Yuan, Yizhuo Yang, Thien Hoang Nguyen, Thien-Minh Nguyen, Jianfei Yang, Fen Liu, Jianping Li, Han Wang, and Lihua Xie. MMAUD: A Comprehensive Multi-Modal Anti-UAV Dataset for Modern Miniature Drone Threats. 2024 IEEE International Conference on Robotics and Automation (ICRA), pp. 2745–2751, 2024.
IEEE / DOI · arXiv · Paper PDF · Download BibTeX
@inproceedings{yuan2024MMAUD,
author = {Yuan, Shenghai and Yang, Yizhuo and Nguyen, Thien Hoang and Nguyen, Thien-Minh and Yang, Jianfei and Liu, Fen and Li, Jianping and Wang, Han and Xie, Lihua},
title = {{MMAUD}: A Comprehensive Multi-Modal Anti-{UAV} Dataset for Modern Miniature Drone Threats},
booktitle = {2024 IEEE International Conference on Robotics and Automation (ICRA)},
year = {2024},
pages = {2745--2751},
doi = {10.1109/ICRA57147.2024.10610957}
}
Baseline references
📖 Cite MMAUD and the baseline papers used in your experiments. The 11 methods above map to 10 papers: AV-PED and AudioNet share the AV-PedAware reference. The Darknet comparison uses Darknet-53, described in YOLOv3. When using the comparison implementations in AV-FDTI, also cite the AV-FDTI paper.
⬇️ Download MMAUD + all baseline BibTeX · Citation sources
📋 Expand BibTeX for all baseline papers
@inproceedings{he2016resnet,
author = {He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian},
title = {Deep Residual Learning for Image Recognition},
booktitle = {2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2016},
pages = {770--778},
doi = {10.1109/cvpr.2016.90},
url = {https://doi.org/10.1109/cvpr.2016.90}
}
@misc{redmon2018yolov3,
author = {Redmon, Joseph and Farhadi, Ali},
title = {{YOLOv3}: An Incremental Improvement},
year = {2018},
eprint = {1804.02767},
archivePrefix= {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/1804.02767}
}
@inproceedings{vora2023dronechase,
author = {Vora, Neel and Wu, Yi and Liu, Jian and Nguyen, Phuc},
title = {{DroneChase}: A Mobile and Automated Cross-Modality System for Continuous Drone Tracking},
booktitle = {Proceedings of the Ninth Workshop on Micro Aerial Vehicle Networks, Systems, and Applications},
year = {2023},
pages = {8--14},
doi = {10.1145/3597060.3597237},
url = {https://doi.org/10.1145/3597060.3597237}
}
@inproceedings{yang2023avpedaware,
author = {Yang, Yizhuo and Yuan, Shenghai and Cao, Muqing and Yang, Jianfei and Xie, Lihua},
title = {{AV-PedAware}: Self-Supervised Audio-Visual Fusion for Dynamic Pedestrian Awareness},
booktitle = {2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
year = {2023},
pages = {1871--1877},
doi = {10.1109/iros55552.2023.10342257},
url = {https://doi.org/10.1109/iros55552.2023.10342257}
}
@article{yang2024avfdti,
author = {Yang, Yizhuo and Yuan, Shenghai and Yang, Jianfei and Nguyen, Thien Hoang and Cao, Muqing and Nguyen, Thien-Minh and Wang, Han and Xie, Lihua},
title = {{AV-FDTI}: Audio-visual fusion for drone threat identification},
journal = {Journal of Automation and Intelligence},
year = {2024},
volume = {3},
number = {3},
pages = {144--151},
doi = {10.1016/j.jai.2024.06.002},
url = {https://doi.org/10.1016/j.jai.2024.06.002}
}
@inproceedings{xiao2025tame,
author = {Xiao, Zhenyuan and Hu, Huanran and Xu, Guili and He, Junwei},
title = {{TAME}: Temporal Audio-based Mamba for Enhanced Drone Trajectory Estimation and Classification},
booktitle = {ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
year = {2025},
pages = {1--5},
doi = {10.1109/icassp49660.2025.10888776},
url = {https://doi.org/10.1109/icassp49660.2025.10888776}
}
@inproceedings{lei2025aaute,
author = {Lei, Allen H.-X. and Deng, Tianchen and Wang, Han and Yang, Jianfei and Yuan, Shenghai},
title = {Audio Array-Based {3D} {UAV} Trajectory Estimation with {LiDAR} Pseudo-Labeling},
booktitle = {ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
year = {2025},
pages = {1--5},
doi = {10.1109/icassp49660.2025.10887734},
url = {https://doi.org/10.1109/icassp49660.2025.10887734}
}
@inproceedings{liang2025unlidar,
author = {Liang, Hanfang and Yang, Yizhuo and Hu, Jinming and Yang, Jianfei and Liu, Fen and Yuan, Shenghai},
title = {Unsupervised {UAV} {3D} Trajectories Estimation with Sparse Point Clouds},
booktitle = {ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
year = {2025},
pages = {1--5},
doi = {10.1109/icassp49660.2025.10890359},
url = {https://doi.org/10.1109/icassp49660.2025.10890359}
}
@misc{deng2024mmudct,
author = {Deng, Tianchen and Zhou, Yi and Wu, Wenhua and Li, Mingrui and Huang, Jingwei and Liu, Shuhong and Song, Yanzeng and Zuo, Hao and Wang, Yanbo and Yue, Yutao and Wang, Hesheng and Chen, Weidong},
title = {Multi-Modal {UAV} Detection, Classification and Tracking Algorithm -- Technical Report for {CVPR} 2024 {UG2} Challenge},
year = {2024},
eprint = {2405.16464},
archivePrefix= {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2405.16464}
}
@article{xiao2026avdtec,
author = {Xiao, Zhenyuan and Yuan, Shenghai and Xu, Guili and Zeng, Xianglong and Hu, Huanran and He, Junwei and Yang, Yizhuo},
title = {{AV-DTEC}: Self-Supervised Audio-Visual Fusion for Drone {3-D} Trajectory Estimation and Classification},
journal = {IEEE Sensors Journal},
year = {2026},
volume = {26},
number = {10},
pages = {15912--15924},
doi = {10.1109/jsen.2026.3680898},
url = {https://doi.org/10.1109/jsen.2026.3680898}
}
- Downloads last month
- 1,987