ABSTRACT
Estimating the 3D pose of drones is vital for anti-drone systems, yet drone keypoint detection faces challenges due to the high visual similarity of propellers and the diversity of drone poses. We propose DroneKey, a framework combining transformer-based 2D keypoint detection and geometric 3D pose estimation with a gated key-representation and pose-adaptive Mahalanobis loss. Our model achieves 99.68% AP (OKS) for keypoint detection and accurate 3D pose estimation with an MAE-angle of 10.62°, RMSE of 0.221m, and MAE-absolute of 0.076m, all at real-time speed (44 FPS).
OVERVIEW
DroneKey addresses the core difficulties of drone keypoint detection by leveraging a transformer encoder to capture spatial relationships among propellers and introducing gated key-representations that integrate intermediate and compact features. A novel pose-adaptive Mahalanobis loss stabilizes learning across extreme poses, while datasets 2DroneKey (2D keypoints) and 3DronePose (3D poses) provide training and evaluation resources. As shown in Fig. 2, the pipeline integrates CNN-based feature extraction, transformer encoding, gated summation for keypoint prediction, and PnP-based 3D pose estimation.
RESULTS
We conducted ablation studies on encoder layers, embedding dimensions, and loss functions to determine optimal configurations. The analysis (Tables II–IV, Fig. 7–8) shows that a 6-layer transformer with an embedding dimension of 1024 provided the best trade-off between accuracy and parameter size, reaching APkp = 99.7%. Gate weight analysis further revealed that not only the final but also intermediate layers significantly contribute to keypoint learning. Moreover, the proposed pose-adaptive Mahalanobis loss with exponential decay (α=5, D=10) outperformed MSE and Gaussian losses, achieving SRkp90 = 99.8% and SRkp95 = 98.5%, confirming stable learning under extreme drone poses.
We compared DroneKey against state-of-the-art object detectors (YOLOv8, DETR), heatmap-based detectors (ViTPose, TokenPose), and coordinate-based methods (SimCC). As shown in Table V and Fig. 9, existing methods struggled with high visual similarity among propellers and produced errors in keypoint order or missed detections under cluttered backgrounds. In contrast, DroneKey achieved APkp = 99.68%, SRkp90 = 99.8%, and SRkp95 = 98.5%, significantly outperforming prior approaches. Visualizations confirm that our model robustly detects and correctly orders propeller keypoints even under complex scenes.
For 3D pose estimation, DroneKey uses predicted 2D keypoints with a PnP solver to reconstruct the drone’s 6DoF pose. On the 3DronePose dataset (Sequences 11–13), our method achieved an MAE-angle of 10.62°, RMSE of 0.221m, and MAE-absolute of 0.076m, demonstrating high reliability across linear, nonlinear, and rotational motion patterns. As shown in Table VI and Fig. 10, the predicted trajectories closely match the ground truth in both location and orientation, even in challenging scenarios with rotations and complex backgrounds.
🖇️ Citation
@inproceedings{ Hwang2025DroneKey,
author = {Seo-Bin Hwang and Yeong-Jun Cho},
title = {DroneKey: Drone 3D Pose Estimation in Image Sequences using Gated Key-representation and Pose-adaptive Learning},
booktitle = {Proceedings of the 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
year = {2025},
address = {Hangzhou, China},
month = {Oct.},
publisher = {IEEE},
doi = {10.1109/IROS60139.2025.11247556}, }
