Several popular drone datasets on GitHub

With the explosive growth of the low-altitude economy, the widespread adoption of drones has brought increasingly prominent safety risks, making counter-drone technology a fiercely contested domain in the security field. However, confronted with detection challenges such as “extremely small targets, ultra-fast flight speeds, and complex backgrounds,” algorithm performance is often bottlenecked by data scarcity. In the era of deep learning, high-quality data is the fuel that powers progress. To address the resource-hunting difficulties faced by developers worldwide, this article systematically reviews mainstream publicly available drone detection datasets globally, encompassing multimodal resources from ground-to-air visible light (RGB) and infrared thermal imaging to radio frequency signals. It provides an in-depth analysis of their applicable scenarios and core characteristics, helping you shatter data silos and rapidly build high-precision drone detection systems.

VisDrone-Dataset

VisDrone is a large-scale multi-scene drone vision benchmark, offering over 2.6 million finely annotated boxes with rich attributes to advance drone-based vision. This is not a drone image dataset, please note

Quantity Information:

  • Video clips: 288
  • Total frames: 261,908 frames
  • Static images: 10,209
  • Annotated bounding boxes: Over 2.6 million
  • Covered cities: 14 Chinese cities

Core Advantages:

  1. Large scale with fine-grained annotations: Massive data volume, providing over 2.6 million meticulously annotated target bounding boxes.

  2. High diversity: Covers various drone models, diverse scenarios (urban and rural), and a wide range of weather and lighting conditions.

  3. Comprehensive coverage: Includes rich target categories (pedestrians, vehicles, bicycles, etc.) and scenes with varying densities (from sparse to crowded).

  4. Rich attributes: Additionally provides key attributes such as scene visibility, object categories, and occlusion, enhancing the practical value of the data.

  5. Application-oriented: Specifically designed to advance the integration of computer vision and drone technology, supporting a variety of visual tasks.

DroneDetectionThesis

This is a multimodal drone detection dataset designed to advance the development and evaluation of counter-drone systems. It integrates infrared (IR), visible light, and audio data, enabling comprehensive sensor fusion for robust drone identification in complex environments.
Quantity Information:

  • Video clips: 650 (365 IR + 285 visible)
  • Audio clips: 90
  • Total annotated images (extracted from videos): 203,328
  • Video labels: Airplane, Bird, Drone, Helicopter
  • Audio labels: Drone, Helicopter, Background

Core Advantages:

  1. Multimodal data: Integrates infrared (IR), visible light, and audio modalities, ideal for training and evaluating comprehensive drone detection sensors and systems.

  2. Rich labeling: Provides detailed video and audio annotations with diverse target categories (including drones, helicopters, airplanes, birds, and background).

  3. Large-scale image extraction: Yields over 200,000 annotated images when frames are extracted from videos, enabling robust model training.

  4. Fully open access: Free to download, use, and modify; includes annotation files in .mat format (generated via MATLAB Video Labeler).

  5. User-friendly resources: Accompanied by a detailed video description file (“Video_dataset_description.xlsx”) and a script with instructions and examples (“Create_a_dataset_from_videos_and_labels.m”) for seamless dataset creation and utilization.

DroneAudioDataset

A specialized drone audio dataset featuring indoor-recorded propeller noise, augmented with diverse environmental sounds (ESC-50), white noise (Speech Commands), and custom silence clips. Organized for binary/multiclass classification, it supports deep learning-based drone detection. Tied to the paper Audio Based Drone Detection and Identification using Deep Learning by Sara Al-Emadi.

Quantity Information:

  • Audio clips: Drone propeller noise (indoor recordings) + artificially augmented with random noise
  • Noise categories:
    • “Unknown” clips sourced from ESC-50 (environmental sounds)
    • White noise from Speech Commands dataset
    • Custom-created silence clips for class balance
  • Organized in binary and multiclass folder structures

Core Advantages:

  1. Specialized drone audio focus: Captures real indoor drone propeller sounds, enhanced via controlled augmentation for robust deep learning training.

  2. Rich noise diversity: Integrates open-source environmental noise (ESC-50), speech-derived white noise, and synthetic silence to simulate real-world acoustic variability.

  3. Research-backed: Directly tied to the conference paper “Audio Based Drone Detection and Identification using Deep Learning” by Sara Al-Emadi.

  4. Open and extensible: Leverages established public datasets (ESC-50, Speech Commands) and includes custom silence augmentation for balanced classification.

  5. Ready for acoustic AI: Structured for binary/multiclass drone detection tasks, ideal for developing noise-robust audio-based counter-drone systems.

dasmehdix/drone-dataset

Patch Antenna, also called Microstrip Antenna, is a compact, planar antenna widely used in C-UAS detection and portable drone jammers.

Key Uses in C-UAS:

  • Detection: Integrates into handheld drone detectors and wearable RF scanners to monitor 2.4GHz/5.8GHz ISM bands.
  • Jamming: Arrays of high-power patches form lightweight jamming panels in anti-drone guns and vehicle-mounted systems.

Advantages:

  • Ultra-low profile (<1cm thick)
  • Lightweight, conformal to surfaces
  • Easy to array for sector coverage
  • Cost-effective mass production

Disadvantages:

  • Narrow bandwidth (~5–10%)
  • Low gain (5–8 dBi) vs. Yagi/LPDA
  • Sensitive to ground plane size
  • Poor efficiency at <1GHz

Best for compact multi-band C-UAS sensors and urban drone defense.

In summary, GitHub hosts a rich variety of drone-related datasets, spanning multiple modalities: aerial imagery/video datasets (e.g., VisDrone, UAVDT), flight trajectory datasets (for prediction and anomaly detection), high-resolution drone image datasets (e.g., DroneVehicle, UAV123), and audio datasets (e.g., indoor-recorded propeller noise with augmentation). Each modality drives distinct technical approaches: image data relies on object detection and tracking; trajectory data enables behavioral analysis and intent inference—a key non-traditional counter-drone technique; audio datasets focus on acoustic feature extraction and deep classification, ideal for silent or nighttime detection. Developers can selectively integrate these resources to build multimodal, high-precision anti-drone systems tailored to real-world security needs.