手部平面運動切割

Abstract

本論文的主要目的在於將手部運動切割成數個有意義的動作片段，以期望能將連續的運動表示成動作片段序列，如此一來，即可提供一個的基礎給後續的分析與處理（動作辨識、動作重編輯等等）。

本中利用關節旋轉軸與肢節垂直與否，將肢節在空間中的運動軌跡分為共面與共圓錐兩種基本運動模式，而手部運動即可視為此兩種運動模式搭配不同的旋轉角度的組合情形。

藉由計算肢節運動軌跡在空間中形成的平面，可求出該段運動過程中關節的旋轉軸為何。依照旋轉軸是否發生改變來進行動作切割，即可得到數個動作片段，其運動過程中旋轉軸固定不變。接著，再對片段中的角度變化，以二階的AR模型來做進一步的切割。最後，即可將動作切割成數個“同旋轉軸與同角度變化模式”的動作片段。

在文章中會說明如何計算運動平面及如何利用旋轉軸變化及角度變化來對動作進行切割。最後用實驗驗證文中提出的方法。

The main purpose of this thesis is to segment a video clip of planar hand motion into motion subsequences each with some unique semantic meaning. Further analysis or processing on these motion segments such as motion recognition, motion editing can be conducted.

In this thesis, the hand motion segments in 3D space are classified into two types: co-planar and co-axial, depending on whether the rotation axis is perpendicular to the trajectory of each hand joint and whether the hand joints are coplanar. Any hand motion sequence can be treated as a combination of these two types of motion with different motion parameters.

By finding the plane of the hand joint trajectory, the rotation axis of the planar hand motion can be obtained. Then, the projection of the 3D joint position trajectory onto the motion plane is computed. Then, perform the second AR model analysis on the projected trajectory to find any further possible segmentation of planar joint motion. In this way, planar hand motion clip is segmented into sub-segments of the same type of AR model.

The proposed method is mathematically formulated. Experimental simulations applied to various clips of hand motion are shown to obtain the segments of hand motions each described by an AR model.