應用在上肢外骨骼復健之代償動作偵測系統

Abstract

臨床上證明，機器人介入的中風復健療程可有效改善中風倖存者的上肢運動恢復，同時可減少治療師的負擔。然而，在沒有治療師監督的情況下，患者可能會因為肌肉無力、姿勢調整異常或是關節間協調性的不足，而使用其他的肌肉或是關節去完成復健動作，此補償方式就稱為代償動作，也會導致不良的復健效果。 本研究擬開發一種基於視覺的室內復健姿勢監控系統，用於自動偵測患者在進行復健動作時有無代償動作發生，並給予及時的視覺、語音回饋，提示患者糾正錯誤的姿勢。除使用彩色深度攝影機的骨架追蹤技術來獲取人體骨架的關節位置資訊外，並透過移除異常資料以及關節位置正規化來抑制骨架追蹤產生的雜訊和降低受試者間的差異性。所有的關節位置都會經過座標轉換到統一的座標系統，可達到視覺不變的效果。在模型訓練階段，則使用極端梯度提升算法來分類三種代償動作及一種無代償動作。 本研究提出的系統及模型皆建立在上肢外骨骼復健機器人NTUH-II。所採用的驗證方法是排除受試者的交叉驗證，且以接收者操作特徵曲線和混淆矩陣來比較不同操作變因。實驗結果顯示本研究訓練出來的模型的分類準確率優於相關的研究，並藉由額外設計的移動視窗投票方法，使得即時偵測的表現更好，使該系統能直接應用於復健療程中。

It has been clinically proven that the robot-mediated stroke rehabilitation can effectively improve the upper limb motor recovery and alleviate the burden of therapists at the same time. However, without the supervision of therapists, the patient might use unaffected muscles or joints to complete the rehabilitation therapy due to muscle weakness, abnormal posture adjustment, or loss of inter-joint coordination. This kind of compensation is called compensatory movements, which can lead to bad effects. In this thesis, we develop a skeleton-based indoor rehabilitation posture monitoring system which is used to automatically detect whether the patient’s compensatory movement occurs during the rehabilitation exercise, and to give real-time visual and auditory feedback to prompt the patient to correct the abnormal posture. The skeletal tracking technique from Kinect v2 device is used to obtain the patient’s joint position information. To suppress the tracking noise and reduce variations among subjects, outlier data removal is utilized and a 3D joint position normalization algorithm is designed. All the joint positions are transformed into a unified coordinate system that can keep view-invariant. In the classifier model training phase, we implement the Extreme Gradient Boosting (XGBoost) classifier to classify the compensatory movement. The system and model proposed in this research are established on the upper limb exoskeleton rehabilitation robot NTUH-II. The leave-one-subject-out cross-validation is used to validate our model and receiver characteristic curve (ROC), area under the ROC curve (AUC), and Confusion matrix are used to compare with different independent variables. The results show that the classification accuracy of the proposed method is better than the relevant studies. Besides, through the design of the window voting method, the real-time performance of compensatory movement detection is well, so that the system can be applied to the rehabilitation therapies.