DroneRFc-MM: Anti-UAV Multi-modal Measured Dataset

It contains data from six types of sensors: a pan-tilt-zoom (PTZ) camera, a wide-angle camera, a radio-frequency (RF) antenna, a LiDAR sensor, a millimeter-wave radar, and a microphone array. The dataset covers more than 30 minutes of flight data from six consumer-grade drone models. It provides fine-grained annotations for drone model, position, attitude, and velocity, and can support tasks such as target detection, drone model recognition, and motion-direction reasoning.

Easy-to-use sample-processing code is also provided.Comparison with Existing Multimodal Counter-UAV DatasetsAnti-UAVData acquisition devices: Visible-light camera and infrared cameraDrone models: DJI Phantom 4, DJI Spreading Wings S1000, and DJI Tello2. MMAUDData acquisition devices: Stereo camera, LiDAR, millimeter-wave radar, and microphone arrayDrone models: DJI Mavic 2, DJI Mavic 3, DJI Phantom 4, DJI Avata, and DJI M3003.

DroneRFc-MMData acquisition devices: PTZ camera, wide-angle camera, RF antenna, LiDAR, millimeter-wave radar, and microphone arrayDrone models: DJI Mini 2 SE, DJI Mini 3, DJI Mavic Air 2S, DJI Air 3, DJI Mavic 3, and DJI AvataData Acquisition EquipmentsPTZ cameraHikvision DS-2DC6423IW2560 x 1440 resolution and 23x optical zoom2. Wide-angle cameraHikvision DS-2CD33462560 x 1440 resolution, 180-degree horizontal field of view, and 98-degree vertical field of view3.

Omnidirectional RF antennaEttus VERT2450Used with an NI USRP-2955; 100 MHz sampling rate; 2.45 GHz center frequency; 50 ms continuous sampling duration per acquisition4. LiDARRoboSense EM4512 channels; 25.92 million points per second; 120-degree horizontal field of view; 27-degree vertical field of view5. Millimeter-wave radarArbe Phoenix77-81 GHz operating band; point-cloud mode; 100-degree horizontal field of view; 30-degree vertical field of view6.

Condenser microphoneAihua AWA14411Four-channel array; 25.6 kHz sampling rate; 6.5-137 dB dynamic rangeTested UAVsDJI Mavic 3Recording identifier: A1Recording periods: 14:00-14:03 and 15:27-15:302. DJI AvataRecording identifier: B1Recording period: 15:13-15:193. DJI Mini 2 SERecording identifier: C1Recording period: 14:36-14:424.

DJI Mini 3Recording identifier: E1Recording periods: 14:11-14:14 and 15:42-15:455. DJI Air 3Recording identifier: F1Recording period: 14:49-14:546. DJI Air 2SRecording identifier: G1Recording periods: 13:49-13:52 and 15:34-15:37Drone Flight TrajectoryEach drone was first flown to a position approximately 20 meters in front of the sensor array, after which data acquisition was started on all devices.

The drone was then flown along a rectangular loop. During this phase, the drone remained approximately 20-60 meters from the equipment, moved across a horizontal range of about 20 meters, and maintained an approximately constant altitude. This phase lasted about 180 seconds.The drone then returned to a position approximately 20-30 meters in front of the equipment and performed repeated vertical movements for about 120 seconds.

Finally, data acquisition was stopped on all devices and the drone was returned to its takeoff point.Dataset Directory StructureThe root directory contains three top-level directories./raw_data stores the original sensor recordings and flight logs./samples stores samples and annotations that can be used directly for model training and evaluation./directionQA stores the test code for the demonstration task developed from this dataset, as described in Section III of the associated paper.The /samples directory contains seven second-level directories, including /PTZ_samples and /fisheye_samples.

These directories store sample and annotation files extracted from the raw data recorded by each sensor. Each second-level directory contains six third-level directories, such as /A1 and /B1, which distinguish the six drone models.Sample files within each third-level directory are named according to their time-segment index. For example, samples/PTZ_samples/A1/23.mp4 is the PTZ-camera video recorded during time segment 23 for drone identifier A1, which corresponds to the DJI Mavic 3.The DJI Avata, identified as B1, does not allow its communication band to be manually set to either 2.4 GHz or 5.8 GHz using the remote controller.

Consequently, RF signals were not collected for this drone model.Custom Sample Segmentation ToolThe dataset supports custom sample generation to accommodate different sample-duration requirements. First, run /samples/labels/clip_labels.py to regenerate indexed annotation files, such as A1.csv, in /samples/labels. The annotation files contain the index, start time, and end time for each sample.The command-line argument --drone specifies the list of drone models, --sample-count specifies the number of samples for each modality, and --sample-duration specifies the duration of each sample.After generating the annotation files, run the relevant clip_XXX.py script in each modality directory under /samples to regenerate the sample files for that modality.Dataset PaperTaosong Yu, Qianqian Yang, Zhuo Hu, Mingkai Li, Jiajun Wu, Yufan Su, Junyu Pan, and Zhiguo Shi. "DroneRFc-MM: Anti-UAV Multi-modal Measured Dataset." Journal of Electronics & Information Technology, under review.