LRDDv3: High-Resolution Long-Range Drone Detection Dataset with Range Information and Thermal Data

Technical Summary: LRDDv3 – High-Resolution Long-Range Drone Detection Dataset

Problem Statement
The rapid proliferation of Unmanned Aerial Vehicles (UAVs) for recreational and commercial purposes necessitates robust detection systems capable of identifying drones at long ranges to ensure safe airspace integration. While various datasets exist for drone detection, the authors identify a critical gap: a lack of high-quality, high-resolution data that reflects the complex, long-range operational contexts of real-world scenarios. Existing datasets often suffer from limited resolution, static camera views, insufficient range metadata, or a lack of diversity in lighting, weather, and background conditions. Furthermore, few datasets combine RGB and thermal imaging with precise range annotations, which are essential for training algorithms to detect small, distant targets under varying environmental challenges.

Methodology and Dataset Construction
To address these limitations, the authors introduce LRDDv3, a comprehensive dataset constructed through controlled flight operations over 8 months across 17 data collection days. The methodology emphasizes diversity and realism:

• Data Acquisition: The dataset comprises 102,532 RGB images and 29,630 thermal (IR) images. These were sampled at 5 FPS from 128 distinct video clips (totaling 5.7 hours of footage) to prioritize background and scene diversity over frame volume.
• Hardware Configuration: Data was captured using a "drone-to-drone" approach.
  • RGB Collection: A DJI Mini 3 acted as the camera drone (recording 4K video at 30 FPS) while another DJI Mini 3 served as the target.
  • Thermal Collection: An Autel Robotics EVO II Dual 640T V3 served as the camera drone, capturing synchronized 4K RGB and 640x512 IR footage, with a DJI Mini 3 as the target.
• Environmental Diversity: Flights occurred in urban, rural, forested, and open field environments. Conditions varied across times of day (dawn, noon, dusk), weather (clear, rainy, snowy), and lighting (sunny, cloudy, overcast).
• Range and Metadata: A key feature is the inclusion of precise range information. Using GPS coordinates and altitude logs from both drones, the authors calculated the 3D distance between the camera and target using the Haversine formula for horizontal distance and altitude differences for vertical distance. This metadata covers distances from 0 to 200 meters, with many instances where the drone occupies as few as 12 pixels in the frame.
• Annotation Process: The dataset includes 93,652 drone annotations, 6,031 bird annotations, and 722 airplane annotations. The labeling pipeline utilized a three-step process: automated pre-labeling via YOLOv11 and SAHI, manual correction using LabelImg, and final verification. Thermal annotations were generated by rescaling and cropping the corresponding RGB annotations.

Key Contributions
The paper positions LRDDv3 as a significant advancement over existing resources through several distinct features:

1. Scale and Resolution: It offers a large-scale collection of 4K resolution RGB images, a rarity in drone detection datasets, alongside paired 640x512 IR images.
2. Comprehensive Range Information: Unlike most prior works, LRDDv3 provides explicit range metadata for the target drone, enabling research into distance-dependent detection performance.
3. Multi-Modal and Environmental Diversity: It uniquely combines RGB and thermal data with dynamic camera movement (air-to-air), diverse weather conditions, and complex backgrounds (including occlusion and blending).
4. Realistic Sampling: By sampling at 5 FPS rather than using full video frame rates, the dataset avoids artificial inflation and ensures a wider variety of backgrounds and lighting scenarios per image.

Results and Benchmarking
The authors evaluated the utility of LRDDv3 through two primary benchmarking experiments:

• Cross-Dataset Generalization: Models trained on LRDDv3 were tested against the Det-Fly dataset (a benchmark for small targets). The YOLOv11m model trained on LRDDv3 achieved an mAP@50 of 0.485 and mAP@50-95 of 0.261, outperforming models trained on the DUT Anti-UAV and previous LRDD versions, and performing comparably to the Drone-vs-Bird dataset. This suggests LRDDv3 provides sufficient diversity to generalize well to other detection scenarios.
• Resolution Impact: The authors investigated the impact of input resolution by training YOLOv11m models on LRDDv3 at 640x640 versus 1920x1920. The higher resolution model showed a substantial performance increase, achieving an mAP@50 of 0.822 compared to 0.543 for the lower resolution model. This demonstrates that the high-resolution nature of the dataset is effectively leveraged by models capable of processing larger input sizes.

Significance
The paper claims that LRDDv3 represents a significant step forward in enabling reliable drone detection systems for shared airspace. By consolidating high resolution, thermal imaging, precise range data, and diverse real-world conditions into a single resource, the dataset addresses the specific challenges of long-range detection where targets are small and environmental conditions are variable. The authors assert that LRDDv3 serves as a comprehensive benchmark for developing next-generation detection algorithms that must operate robustly across the full spectrum of detection ranges and environments, bridging the gap between controlled laboratory datasets and real-world operational requirements.