駕駛安全至今仍然是一個重要問題，尤其是在準確識別駕駛者手部位置和分類車輛運動以預防事故發生。本論文聚焦於兩項關鍵研究任務。第一項任務是探索如何利用單一智慧手錶捕捉雙手資訊的潛力，其困難點在於智慧手錶只能有效檢測佩戴它的那隻手。我們的目標是實現單一智慧手錶的手部位置準確檢測，達到與雙智慧手錶相當的表現。第二項任務是解決從模擬數據和真實數據中提取一致性特徵的難題，以建立適用於真實駕駛的車輛運動分類模型。雖然真實數據更能反映行為分析，但其收集過程具有一定危險性；相較之下，模擬數據更安全，但與真實數據之間存在差距。我們的目標是開發一種基於模擬且更安全的車輛運動分類方法，並縮小模擬數據與真實數據之間的差距。為實現這兩個目標，一個關鍵步驟是發現精確的特徵提取方法。我們提出的方法利用希爾伯特-黃轉換（HHT）從單一智慧手錶提取雙手的特徵，並結合主成分分析（PCA）和線性判別分析（LDA）從模擬數據中提取一致性特徵，用於車輛運動分類。本研究的兩大貢獻為：提升手部位置檢測準確率，讓單一智慧手錶達到98.29%的準確率，超過雙智慧手錶的97%；以及提出一種創新、安全且基於模擬模型的數據收集解決方案，並應用於真實數據，將車輛運動分類準確率相比原始模擬數據提升了29.74%。;Driving safety remains a critical issue, especially in terms of accurately identifying hand positions and classifying car movements for accident prevention. This thesis presents two essential research assignments. First is motivated by the potential of a single smartwatch to capture both hands which is challenging because smartwatches are only effective on the hand wearing it. The objective is to achieve competitive hand position detection accuracy with a single smartwatch, comparable to the performance of dual smartwatch. The second is to address the challenge of extracting consistent features from simulation and real data to build a car movement classification model applicable to real-world driving. Real data is more relevant for behavior analysis, but it is dangerous to collect, whereas simulation data, though safer, introduces a gap with real-world consistency. The objective is to develop a safer, simulation-based approach for car movement classification. An important phase to achieve both objectives is to discover precise feature extraction. Our proposed method leverages Hilbert-Huang Transform (HHT) to extract features from both hands using a single smartwatch and uses PCA and LDA to derive consistent features from simulation data for car movement classification. The two contributions are: improvement in hand position detection accuracy, achieving 98.29% with one smartwatch, surpassing the 97% accuracy achieved with two smartwatches; a novel solution for safer data collection using simulation models applied to real data, improving car movement classification accuracy by 29.74% over raw simulation data.
