基於慣性傳感器的圖卷積網路應用於下肢運動軌跡估測

Abstract

近年來，動作捕捉系統受到大量的發展與關注，因為其應用場域廣泛，像是動畫的製作、體育和醫療上的應用。在復健領域中，也時常透過動作捕捉系統，擷取病患參與復健時的動作資訊，作為量化患者復健前後變化的依據，提供更客觀的數據供醫師參考。 由於以光學原理為主的動作捕捉系統，需要在固定場域架設相機，故多半只適用於實驗室內。而以慣性傳感器為系統的主軸，雖然它不受應用場域限制，但他相較於光學原理的方法，容易產生量測結果較不準確，若要提高準確度，需要更高規格硬體與更多數量的傳感器以還原更全面性姿態，成本也會增加。 於相關文獻中，多半使用不同傳感器融合技術與演算法，結合運動學模型，及人體關節限制來改善慣性感測器追蹤不準確的問題，但不同做法與應用，難以套用在不同系統架構上。最後，我們分析了當前最先進的人體姿態追蹤期刊後我們發現，目前大多數的研究會因傳感器飄移問題，而造成追蹤不準確。 此外，傳感器會因佩帶時間過長，受到訊號雜訊、陀螺儀估測值不準確的影響，而產生飄移問題。基於上述原因，本論文提出了一種深度學習模型架構，透過慣性傳感器蒐集到的有效資訊，預測在未來動作序列3D空間各關節點的相對位置，藉由模型生成的資訊與原始訊號做對照，以縮減在只依賴傳感器而產生的追蹤誤差。並在最終實驗，以走路的下肢運動軌跡應用，作為驗證，證明本論文提出的方法，其表現可以減緩傳感器飄移問題。於未來展望，為驗證此模型與系統之效能，將本研究方法應用於實際復健場域中。

In recent years, the motion capture system has received a lot of attention because of its wide application, such as movie animation, sport and medical applications. In the field of rehabilitation, motion capture systems are often used to collect the motion information of the patient when he/she is performing rehabilitation tasks. It can be used as a method to quantify patient rehabilitation effectiveness, and also provide more objective data for physician as reference. However, the current commercial motion capture systems are not widely used. Since they are based on principle of optics and need to be set up in fixed region, they are mostly suitable for use in the laboratory. Besides the optical systems, there is another system which adopts inertial sensors as its core. Although it is not limited by the fields of application, it is less accurate in terms of the produced data than the optical motion capture system. If one wants to improve the accuracy of the inertial sensors, higher specification and more sensors are required to reconstruct more comprehensive postures. However, doing so will also increase the costs to users. In the related literature of using inertial sensors, most of them use different sensor fusion technologies and algorithms, which together with kinematic models and human joint constraints, to increase the accuracy of assessed tracking data. After a comprehensive survey of the literature in human posture tracking, we also found that most of the current researches suffer from sensor drift problem, which unfortunately causes inaccurate tracking. In addition, the sensor measurements will be affected by signal noise, bias and inaccurate gyroscope estimation during long-term wearing, which may cause drift problems. Based on the aforementioned reason, this thesis proposes a deep learning architecture that uses effective information collected by inertial sensors to predict the relative position of each joint of human’s lower limbs in 3D space of the future motion sequence. The motion prediction result from the deep learning model is used to alleviate the tracking error caused by reliance only on inertial sensor. In the final experiment, the application of walking lower limb motion trajectory is used as a verification to prove that the method proposed can perform more accurately. In the future work, to verify the effectiveness of the proposed model and system for rehabilitation is required to be validated through clinical studies.