基于主导方向校正的无人机航拍图像匹配算法研究

Research on UAV Aerial Image Matching Algorithm Based on Dominant Direction Correction

摘要: 图像匹配是导航定位、地图构建、三维场景重建等计算机视觉下游任务的核心支撑，但大角度旋转等几何变换常导致其匹配性能显著退化。当前主流深度学习匹配模型，由于特征表示环节缺乏显式旋转不变性设计，在未知的大角度旋转场景中的匹配精度大幅下降。针对这一局限，本文提出一种基于主导方向校正的图像匹配算法IMA-DDC (Image Matching Algorithm Based on Dominant Direction Correction)，该算法具有较强的旋转不变性，且模型推理实时性较高。在特征点匹配阶段，根据初始的特征匹配对，构建能够表示旋转特性的主导方向；同时在匹配推理环节嵌入旋转一致性约束，优化SuperGlue的匹配过程。实验结果显示，在构建的无人机航拍数据集上，本文提出的IMA-DDC方法在四种地形场景中均取得了最高的定位精度，且与对比方法存在显著性能优势。如在城镇地形中，相较SuperGlue和EfficientLoFtr，IMA-DDC分别提升5.98%和2.72%。该方法有效强化了现有匹配模型的旋转不变性，拓展了其在复杂几何变换场景中的应用可靠性。

Abstract: Image matching is the core support for downstream computer vision tasks such as navigation and positioning, map construction, and 3D scene reconstruction. However, geometric transformations like large-angle rotation often lead to significant degradation of its matching performance. Current mainstream deep learning matching models, due to the lack of explicit rotation invariance design in the feature representation stage, suffer a sharp drop in matching accuracy in unknown large-angle rotation scenarios. To address this limitation, this paper proposes an image matching algorithm based on dominant direction correction, namely IMA-DDC (Image Matching Algorithm Based on Dominant Direction Correction). This algorithm possesses strong rotation invariance and high real-time performance in model inference. In the feature point matching stage, a dominant direction that can characterize rotational properties is constructed based on initial feature matching pairs; meanwhile, rotation consistency constraints are embedded into the matching inference process to optimize the matching procedure of SuperGlue. Experimental results show that on the constructed UAV aerial image dataset, the proposed IMA-DDC method achieves the highest positioning accuracy in all four terrain scenarios and exhibits significant performance advantages over the comparison methods. For example, in urban terrain, compared with SuperGlue and EfficientLoFtr, IMA-DDC improves the accuracy by 5.98% and 2.72% respectively. This method effectively enhances the rotation invariance of existing matching models and expands their application reliability in complex geometric transformation scenarios.