超過百分之十的交通事故都起因於不專心的駕駛。當駕駛人分心於其他事情時，雙手仍須握在方向盤上，否則就是不安全的駕駛。根據Setyan的論文，將智慧手錶中三軸加速器的震動幅度資訊套用 empirical decomposition 和 Hilbert-Huang 轉換，不安全的駕駛行為就可以被偵測出來。據文中所述，以上的方法可以在每個駕駛人資料集的70%中達到97%的平均單一模型正確率，以及93.4%的grouping model正確率。在此研究中，我們提出一個提升 grouping model 正確率的方法去計算每個group的權重，而非選擇單一個 group 並以 leave one person test method 驗證。同時，我們也建立了一個新的獨立模型來訓練少量的資料。實驗成果顯示，在同樣的測試方法下我們提出的方法達到了93.6%的平均正確率；在2%的訓練資料中達到80%的平均單一模型正確率；在5%的訓練資料中達到了90%的平均單一模型正確率。 關鍵字: 駕駛者行為辨識、分心駕駛、智慧手錶、有限資料、希爾伯特-黃轉換 ;More than 10 percent of car accidents cause by a distracted driver, Distraction occurs when driver divert their attention from the driving task, the driver should hold the steering wheel with both two hands during the driving otherwise, it will be classified as unsafe driving behavior. Based on Setyan’s thesis, he can detect the driver′s unsafe driving behavior by extracting vibration information from a smartwatch accelerometer sensor using empirical decomposition and Hilbert-Huang transform, and he claims using 70% of each driver dataset he can obtain 97% individual model accuracy in average, and 93.4% average in grouping model. In this research we propose an approach to improve the group model accuracy by calculate weight every group instead of selecting one group and validate the model with leave one person test method, we also a create new individual model to handle the small data training. Our experiment result shows an average accuracy up to 93.6% in group model with the same testing method, 80% on average for the individual model with only 2% of data training, and 90% on average for the individual model with 5% data training. Keywords: driver’s hand detection, distracted driving, smartwatch, limited data, Hilbert-Huang transform.
