以範例為基礎利用Kinect感應裝置輔助之上身序列作評估

Abstract

基於影像資料的人體動作辨識技術，因其廣泛的應用一直是一個重要的研究領域。近來，隨著商用深度相機，如Microsoft Kinect，的發展，人體姿勢分析的難度大幅降低，更加刺激該領域的蓬勃發展。越來越多讓人眼睛為之一亮的應用相繼問世。然而，絕大多數的研究致力於提升動作分析的準確度和運算速度而非詳細的動作評估方式，人體姿勢預測和了解動作表現的好壞仍舊存在很大的鴻溝。而這樣的動作評估技術對於運動教學和健康照護的應用尤為重要。 在本篇論文中，我們提出了一個運動教學系統，藉由Microsoft Kinect捕捉使用者的色彩及深度資訊，以OpenNI人體骨骼追蹤的資訊作為主要輸入，計算更能代表人體的座標系統將關節位置轉換成球狀座標，建立穩定而具鑑別度的姿勢特徵描述，降低人體比例和視角變化的影響。並利用以範例為基礎的人體動作分析技術，以非線性時間彎曲校正的方法辨認使用者的動作，尋找資料庫中相應的範例片段作為學習的對象。藉由非線性時間彎曲校正的方式，即使在有局部速度變化的情況下，依然可找出對應的動作片段，計算動作間的細微差距。最後我們利用系統分析的資訊，呈現給使用者更有效且更有組織的動作評估結果，幫助使用者改善動作表現。

Vision-based human action analysis is an important research area because of its wide-ranging applications. Nowadays, with the aid of commercial depth camera, such as Microsoft Kinect, the difficulty of the posture estimation process is greatly decreased and this has even spurred the research progress. More and more amazing applications are developed. However, most of the research works focus on improving the estimation rate and the computation speed rather than the detailed motion evaluation and there is still much to do to bridge the estimation of human posture and the understanding of action performance. Moreover, this movement assessment technique is especially important in the healthcare and motion instruction applications. In this work, we proposed a motion instruction system, using Microsoft Kinect to capture the color and depth information of the user and then taking OpenNI skeleton tracking model as system input. We use this joint-matched skeleton model to calculate human body's object coordinate system and describe the joint positions in spherical coordinate to construct a more robust and discriminative pose descriptor, lowering the effect of anthropometric and viewpoint transformations. Furthermore, we also use the exemplar-based action recognition technology, applying non-linear time-warping approaches to recognize the actions performed by the users and then use this action sequence to find the similar segments in the database as the learning model. This non-linear time warping approaches could find the correspondent pairs under certain degrees of time variation, defining a similarity measurement and calculating the subtle difference among different executions. Finally, we utilize these information, present the motion assessment result to the users in a more effective and organized way, hoping to help them improve their action performance.