日常生活動作之動作分類與穩定性評估系統開發

Abstract

日常生活動作佔了生活的大部分時間，而日常生活動作的穩定性對於生活品質影響甚鉅，過去日常生活動作的穩定性相關研究大多針對單一動作的動作參數比較，少有針對範圍較廣且包含許多種動作的居家環境的動作穩定性研究。本研究提出即時性動作分類系統整合多種動作穩定性評估方式，針對自由度較高的連續性試驗進行即時動作分類，並依照受試者當下所執行的動作種類進行相對應的穩定性評估。由於動作控制能力因人而異，因此本研究提出相對穩定性，將每位受試者自身的正常動作視為相對穩定性評估的基準，透過正規化來量化成可比較的正規化參數。 本研究第一階段實驗旨在建構即時動作分類演算法。三十位受試者分別執行日常生活動作，利用三軸加速規附著於受試者的第七節頸椎處、下背薦骨處、慣用手的手腕處與慣用腳的大腿前側，並擷取實驗過程中附著肢段的三軸加速度。經過滑動窗口設定(窗口長度0.5秒、滑動長度0.25秒)、動作特徵參數選取(跑步：垂直方向加速度標準差、下樓梯：三軸疊加加速度的最小值、上樓梯：前進方向加速度的平均值)、演算法選用以及閾值設定後，整合多種動作穩定性評估方法(靜止站立：95%信賴區間晃動半徑、行走：三軸疊加加速度的RMS值、上下樓梯：前後、左右方向加速度的變異數)，並將演算法透過程式撰寫實現，供後續驗證。 本研究第二階段實驗旨在驗證本研究所開發之系統，五位受試者配戴三軸加速規於下背薦骨處與大腿前側，並執行連續日常生活動作兩次，其中第二次時增加外在動作干擾條件(濕滑地面、非平整地面…等)。透過與實驗錄影檔案交叉比較，動作分類演算法的分類準確率在受試者執行連續正常動作之下呈現94%的分類準確率。而透過比較受試者執行連續正常動作時，穩定性正規化參數的表現，證明了整合動作分類演算法的有效性。而在干擾條件下的正規化參數表現也符合過去文獻所觀察到的趨勢。 本研究成功建構出適用於生活環境中，且針對多種動作進行即時性分類與穩定性評估的系統雛形。未來可用於提早預警動作控制出現障礙之個人，以及時進行醫療或訓練的介入，降低跌倒或意外傷害的風險。

Activities of daily living (ADLs) are major components of our life. However, instability situations, such as loss of balance or falling, might jeopardize the completion of these fundamental activities and ultimately decrease the quality of life. Therefore, providing a real-time stability assessment methods, which is suitable for multiple activities in real-life situations, is extremely crucial. Previous research studies focused on assessing stability during individual activity, while none of them successfully provided a comprehensive manner to assess multiple activities in real-life situation. In this study, integrating real-time activity classification algorithm with multiple activity stability assessment methods was proposed. Since the movement control ability varies with each individual, relative stability, comparing current activity with personal normal activity, was also proposed to quantify activity stability. In the first-stage experiment, activity classification algorithm construction, thirty subjects performed nine ADLs with four tri-axial accelerometers attached to spine (C7), sacrum, dominant wrist and dominant thigh. The acceleration data were divided into consecutive sliding window for further analysis. Activity features were extracted and the process of setting individualized thresholds was established. Finally, multiple activity stability assessment methods were integrated into the developed algorithm for further evaluation. In the second-stage experiment, system evaluation, five subjects performed two series of continuous activities with two tri-axial accelerometers attached on sacrum and front side of dominant thigh. First series was performed without any disturbance, whereas second series was performed with disturbances, such as slippery surface and obstacle crossing. Activity classification algorithm provided 94% accuracy on classifying normal activities in first series. The validity of integrating activity classification algorithm with multiple assessment methods was also been authenticated by comparing the results from first series with and without classification algorithm. Finally, the relative stability results during second series also met the results of previous studies. In this study, we successfully developed a real-time activity classification and stability assessment system, which can be applied under the conditions of daily living and multiple activities.