應用輕量級視覺里程計之自主無人機於無 GPS 環境: 實機飛行驗證

Abstract

自主無人機（UAV）在複雜的室內或無 GPS 環境中運作時，亟需具備穩定且準確的定位能力。傳統視覺里程計（VO）系統依賴手工(hand-crafted)的特徵偵測與追蹤方法，但在低紋理或動態場景中，往往面臨定位效能下降的問題。本論文提出 XVINS，一個專為無人機設計的輕量化視覺里程計框架，結合深度學習特徵（XFeat），以提升前端追蹤流程的穩定性及準確性。本方法採用混合式設計：首先利用傳統光流法進行高效率特徵追蹤，當傳統追蹤無法維持足夠特徵時，則透過XFeat 深度特徵進行補充。經由模擬資料集與實際室內飛行場景的實驗結果顯示，所提出的 XVINS 系統在軌跡穩定性與環境適應性方面，均顯著優於傳統視覺里程計方法。為進一步確保自主飛行的穩定性與可靠性，本研究亦整合光流計，並系統性地調整控制參數，以提升在複雜環境下的飛行控制能力。透過模擬資料集與實際室內飛行場景的實驗結果顯示，所提出的 XVINS 系統結合上述強化措施後，在軌跡穩定性與面對環境挑戰的適應性方面，均顯著優於傳統視覺里程計方法。

Autonomous unmanned aerial vehicles (UAVs) require robust and accurate localization to operate in challenging indoor or GPS-denied environments. Traditional visual odometry (VO) systems, which rely on hand-crafted feature detection and tracking, often struggle in low-texture or dynamic scenes, leading to degraded localization performance. In this thesis, we present XVINS, a lightweight visual odometry framework for UAVs that integrates deep learning-based features (XFeat) to enhance both the robustness and stability of the front-end tracking pipeline. Our method employs a hybrid approach: conventional optical flow is initially used for efficient feature tracking, while deep features from XFeat are selectively introduced to supplement the system when conventional tracking is insufficient. To further ensure stable and reliable autonomous flight, we also incorporate an onboard optical flow sensor and systematically tune control parameters to improve flight control under challenging conditions. Experimental results in both simulated datasets and real-world indoor flight scenarios demonstrate that the proposed XVINS system, together with these additional enhancements, significantly outperforms traditional VO in terms of trajectory stability and resilience to environmental challenges.