Improving Small Drone Detection Through Multi-Scale Processing and Data Augmentation

TL;DR

This work tackles small drone detection in cluttered environments by building a YOLOv11m-based pipeline that uses multi-scale input processing (full frame plus overlapping segments), extensive copy-paste data augmentation, and a temporal post-processing stage to reduce missed detections. The method, trained on a mix of the DDS and public UAV datasets with a two-class setup (drone and bird) and evaluated with MAP$_{50}$, demonstrates substantial gains over single-scale processing, achieving first place in the 8th WOSDETC Drone-vs-Bird Grand Challenge. Key contributions include a practical multi-scale aggregation strategy, aggressive data augmentation to bolster small-object representations, and a frame-consistency approach to stabilize detections across sequences. The approach offers a robust solution for real-world UAV surveillance, albeit with notable computational overhead that motivates future optimization and integration of trackers or faster architectures.

Abstract

Detecting small drones, often indistinguishable from birds, is crucial for modern surveillance. This work introduces a drone detection methodology built upon the medium-sized YOLOv11 object detection model. To enhance its performance on small targets, we implemented a multi-scale approach in which the input image is processed both as a whole and in segmented parts, with subsequent prediction aggregation. We also utilized a copy-paste data augmentation technique to enrich the training dataset with diverse drone and bird examples. Finally, we implemented a post-processing technique that leverages frame-to-frame consistency to mitigate missed detections. The proposed approach attained first place in the 8th WOSDETC Drone-vs-Bird Detection Grand Challenge, held at the 2025 International Joint Conference on Neural Networks (IJCNN), showcasing its capability to detect drones in complex environments effectively.

Figures (6)

• Figure 1: The similarity in size and appearance often makes it challenging to differentiate a drone (blue box) from birds (red box), especially at a distance.
• Figure 2: Representative images from the dds dataset. Each image contains at least one drone, but bounding boxes were intentionally omitted to emphasize the challenge of visually identifying them in certain scenarios.
• Figure 3: Samples from the three public datasets integrated into our training set.
• Figure 4: Overview of the proposed approach. First, the medium-sized YOLOv11 model (YOLO11m) yolov11 is applied to both the full input frame and its segmented regions (simulating a zoom effect). Detections across the original frame and segments are then aggregated, with redundant bounding boxes removed via nms. Lastly, temporal consistency and robustness to missing detections are achieved by tracking drones across a temporal window and applying linear interpolation.
• Figure 5: Demonstration of the applied copy-paste data augmentation technique: original images (left) and their augmented counterparts (right), where patches of drones and birds were inserted to enhance variability in the training set.