关键词: 无人机定位, 建筑物语义, 图卷积网络, 语义嵌入, 对比学习

Abstract:

UAV geographic-matching localization in GNSS-denied urban environments is a challenging problem in the navigation field. Existing image-matching-based methods are sensitive to texture variations and rely heavily on large-scale image databases. To address these issues, this paper proposes a building semantics-driven localization method. Taking pre-extracted building vector outlines from UAV orthophotos as input, the method formulates localization as a similarity search in a semantic embedding space. The proposed approach consists of two phases: offline database construction and query localization. In the offline phase, geometric and spatial-contextual semantic features are extracted from building outlines to construct semantic graphs; a specially designed semantic embedding model then learns discriminative embeddings, which are stored together with geographic coordinates to establish a reference library. In the query phase, a semantic graph is constructed from building outlines in the query image, and the UAV position is estimated through embedding retrieval, spatial clustering, and geometric consistency estimation. Experiments on both homologous and heterogeneous datasets demonstrate that the proposed method achieves median localization errors of 0.74 and 4.13 m, with success rates (error<20 m) of 98.30% and 77.00%, respectively, and per-query processing times of 0.938 and 0.957 s, outperforming nine competing methods including GCN, GraphSAGE, and SGFormer. Cross-city generalization experiments further validate its strong generalization capability in unseen urban environments.

Key words: UAV localization, building semantics, graph convolutional networks, contrastive learning, semantic embedding