林定位：在野外定位不用全球衛星定位系統

Abstract

無人機（UAV）在各種應用領域中發揮著重要作用，然而，其對全球導航衛星系統（GNSS）的依賴，在城市峽谷、茂密森林和室內環境等GNSS受限場景中面臨嚴峻挑戰。本研究探討了一種基於深度學習的無依賴GNSS定位方法，透過影像檢索與特徵匹配技術來提升無人機的自主定位能力。基於Deep Visual Geo-localization Benchmark框架 [Berton, Gabriele, et al. 2022]，我們使用了一種結合ResNet [He, Zhang et al. 2016] 與NetVLAD [Arandjelovic, Gronat et al. 2016]進行全局影像檢索，並使用SuperGlue [Sarlin et al. 2020] 進行精細特徵匹配的方法。我們的方法在Visual Terrain Relative Navigation（VTRN）數據集（ALTO數據集的子集）[Cisneros, Ivan, et al. 2022] 上進行評估，該數據集涵蓋多種環境條件下的真實無人機影像。實驗結果顯示，我們的模型在Top-1檢索準確率達67.5%，在Top-20檢索準確率超過93%，提升了無人機在GNSS受限環境中的定位性能。然而，特徵匹配與位姿估計仍面臨挑戰，特別是在有極端視角與光照變化，或是缺乏特徵的地形等情境下。因此，未來研究將專注於提升多尺度特徵表示能力，結合混合式定位技術，並進一步優化模型於實際場景中的部署。本研究希望提升無人機自主導航與無依賴GNSS定位技術，促進其在各類受限環境中的應用發展。

Unmanned Aerial Vehicles (UAVs) play a crucial role in various applications, yet their reliance on Global Navigation Satellite System (GNSS) signals poses significant challenges in GNSS-denied environments such as urban canyons, dense forests, and indoor spaces. This research explores GNSS-free localization by leveraging deep learning-based image retrieval and feature matching techniques. Building upon the Deep Visual Geo-localization Benchmark framework [Berton, Gabriele, et al. 2022], we propose an approach integrating ResNet [He, Zhang et al. 2016] with NetVLAD [Arandjelovic, Gronat et al. 2016] for global image retrieval and SuperGlue [Sarlin et al. 2020] for fine-grained feature matching. Our method is evaluated on the Visual Terrain Relative Navigation (VTRN) dataset, a subset of the ALTO dataset [Cisneros, Ivan, et al. 2022], which provides real-world UAV imagery under diverse environmental conditions. Experimental results demonstrate that our model achieves 67.5% top-1 retrieval accuracy and over 93% accuracy within the top-20 matches, significantly improving UAV localization performance in GNSS-denied scenarios. However, challenges remain in feature matching and pose estimation, particularly under extreme viewpoint variations, lighting changes and terrain that lack of features. Future work will focus on enhancing robustness through multi-scale feature representation, hybrid localization techniques, and real-world deployment optimizations. Our research contributes to the advancement of autonomous UAV navigation, enabling more reliable and efficient operations in GPS-denied environments.