| 摘要: | 室內定位 (Indoor Localization, IL) 是建構智慧家庭、智慧辦公室、智慧教室及智慧工廠等智慧室內環境的基礎技術。許多室內應用，例如室內擴增實境 (Indoor Augmented Reality, IAR)，需依賴精準的室內定位以確保優良的使用者體驗與系統效能。本研究體認到室內定位是建構高效IAR系統的關鍵促成組件，因此提出「室內定位擴增實境框架 (Indoor Location-Based Augmented Reality Framework, ILARF)」。此框架為一可擴展架構，將室內定位系統作為核心模組整合至IAR應用中。對ILARF的探討顯示，利用低功耗藍牙 (Bluetooth Low Energy, BLE) 和Wi-Fi等無線設備訊號的射頻定位方法，具有高度發展潛力。這些方法提供多項實務優勢，包含低成本、易於部署及相當不錯的定位準確度。然而，研究過程亦突顯出欲藉此類方法達成高精度定位所面臨的重大挑戰，包括動態變化的嘈雜室內環境、不穩定的訊號強度、多路徑傳播及訊號干擾。為應對這些挑戰，本研究進一步聚焦於運用先進計算技術，特別是機器學習 (Machine Learning, ML) 與量子機器學習 (Quantum Machine Learning, QML)，以提升室內定位的準確性。
為此，本研究開發並評估了三種不同的計算方法學。第一種是以特徵為核心的方法，稱為指紋特徵提取 (Fingerprint Feature Extraction, FPFE)，其採用自動編碼器 (Autoencoders, AE) 與主成分分析 (Principal Component Analysis, PCA) 從原始的接收訊號強度指示 (RSSI) 數據中提取穩健特徵，繼而透過基於閔可夫斯基距離 (Minkowski distance) 的匹配演算法進行位置估算，在一個於單一房間內收集的數據集上，展現了實現次米級定位準確度的能力，其平均絕對誤差 (Mean Absolute Error, MAE) 為0.68公尺。第二種是端到端的隨機森林迴歸 (RFR-IL) 模型，此模型在一個涵蓋多個房間與走廊大範圍區域且具挑戰性的公開基準數據集上，達成了約0.6公尺的MAE，從而建立了一個強大的新古典基準。最後，為探索可能超越傳統計算限制的解決方案，第三種方法引入了一個名為「量子隨機森林室內定位 (QRF-IL)」的量子啟發式模型。在相同的基準數據集上，其表現超越了先前最知名的基於訊號傳播模型的古典方法21\%，最終達成了2.3公尺的MAE。所有三種方法均透過多個系統原型與基準數據集進行驗證。總體而言，這些方法共同為提升當前最先進的室內定位技術提供了一份清晰的技術路線圖。
;Indoor Localization (IL) is a foundational technology for constructing smart indoor environments, including smart homes, smart offices, smart classrooms, and smart factories. Many indoor applications, such as Indoor Augmented Reality (IAR), rely on accurate IL to ensure good user experiences and system performance. Recognizing that IL is a critical enabling component for building effective IAR systems, this research proposes the Indoor Location-Based Augmented Reality Framework (ILARF)—an extensible architecture that integrates IL systems as core modules within IAR applications. The investigation of ILARF demonstrates that radio-based IL methods utilizing signals from wireless equipment, such as Bluetooth Low Energy (BLE) and Wi-Fi devices, are particularly promising. These methods offer several practical advantages, including low cost, ease of deployment, and reasonably good IL accuracy. However, the investigation also highlights significant challenges in achieving highly accurate IL with such methods. These challenges include dynamically changing and noisy indoor environments, unstable signal strength, multipath propagation, and signal interference. In response to these challenges, this research further focuses on improving IL accuracy with advanced computational techniques, particularly Machine Learning (ML) and Quantum Machine Learning (QML).
To this end, this research develops and evaluates three distinct computational methodologies for IL. The first, Fingerprint Feature Extraction (FPFE), which employs Autoencoders (AE) and Principal Component Analysis (PCA) to derive robust features from raw Received Signal Strength Indicator (RSSI) data, followed by a Minkowski distance-based matching algorithm for location estimation, demonstrated the ability to achieve sub-meter localization accuracy, namely 0.68 meters of the Mean Absolute Error (MAE) on a dataset collected in a single room. The second, an end-to-end Random Forest Regression (RFR-IL) model, established a powerful new classical baseline by achieving an MAE of approximately 0.6 meters on a challenging public benchmark dataset collected in a large area of multiple rooms and corridors. Finally, to explore solutions beyond classical limits, a novel Quantum Random Forest for Indoor Localization (QRF-IL) was introduced, which outperformed the best-known classical signal-propagation-based method on the same benchmark by 21\%, achieving a final MAE of 2.3 meters. All three methods are validated through multiple system prototypes and benchmark datasets. Together, they provide a roadmap of techniques that advance the state of the art in IL. |
