An equivariant convolution-based feature detection algorithm for matching unmanned aerial vehicles images with large rotations
Jiachuang Zhang, Hongguang Li, Song WANG, Rong Li, Xinjun Li · Chinese Journal of Aeronautics · 2026
Image matching is critical for Unmanned Aerial Vehicle (UAV) mapping and geolocation. However, traditional methods struggle with weak textures, and most deep learning models often lack rotation equivariance, causing performance degradation under large-angle changes. To address this, this paper proposes the Equivariant Steerable Feature network (ES-Feat), a lightweight, rotation-equivariant matching network for UAV scenarios. By integrating the E(2)-Equivariant Steerable CNN (ESCNN) framework, ES-Feat enables explicit rotational feature extraction. This is reinforced by a tightly coupled keypoint prediction head for feature consistency, alongside a global offset prediction head and loss function to correct systematic downsampling offsets and improve localization robustness. The proposed method is evaluated using the Area Under the Curve (AUC) for pose estimation accuracy and Frames Per Second (FPS) for processing speed. Experimental results demonstrate that on the standard MegaDepth-1500, the AUC@5° of ES-Feat_l exceeds that of XFeat by 7.6 %. Furthermore, because existing public datasets lack sufficient large-angle rotational samples to fully evaluate UAV flight scenarios, a fixed-wing UAV was used to collect real-flight images to construct a novel Aircraft_Rot dataset. On this challenging dataset, ES-Feat_l achieves an AUC@5° of 51.5 %, representing a 42.6 % improvement over XFeat and significantly outperforming LoFTR and the traditional ORB algorithm. Moreover, the most lightweight variant, ES-Feat_s, surpasses ALIKE and SuperPoint in matching speed, achieving a feature extraction frame rate of 4.39 FPS on the RK3588 edge device. These results demonstrate the efficiency and practical value of ES-Feat for robust image matching on resource-constrained UAV platforms. Data is available at https://github.com/ZHjiuang/ES-Feat .