利用逆向動力學修正立體視覺人體動作原始資料

Abstract

本論文的主要目的，是為了在不貼感應標誌(Marker)的情況下，以逆向動力學修正由立體視覺所得到較粗略的人體動作原始資料，使其肢節長度與內部角度更為正確可靠。 本論文使用兩部攝影機分別以不同的視角拍攝表演者的動作，並得到兩串立體影像序列，透過影像處理方式得到人體各關節粗略三維座標位置後，以肢節長度正確定義的人體模型為初解，使用逆向動力學修正人體動作原始資料。 文中對本論文所定的人體模型架構做了完整的介紹，將人體模型定義為以腳跟為Root的樹狀架構，其中包含了十一個關節點，共二十一個關節的自由度，再以四個End-effectors在人體動作原始資料中的位置為逆向動力學的修正目標。並且說明了各種逆向動力學方法的理論根據，Jacobian的定義、pseudo-inverse方法、Damped Least Square方法與Selectively Damped Least Square方法。 實驗中展示了以不同的初解來源進行逆向動力學修正的結果，以不同的逆向動力學解法求解的結果。修正完各個畫格後，以關鍵畫格取代完整畫格的成果。最後以IP為初解使用的DLS逆向動力學方法通常是在效率與正確性上較好的選擇。

The main purpose of this thesis is to refine a sequence of stereo vision-based human motion raw data with a goal that the true limb lengths and proper joint angles can be satisfied.

Two CCD cameras are used to gain a video of human motion from two different view points. 3D joint positions are extracted from these two image sequences through techniques of image-processing and computer vision.

The thesis introduces a hierarchic human body structure, which is a tree structure using the heel as root. It includes twenty-one degrees of freedom in eleven joints. 4 end-effectors are used in inverse kinematics refinement. The overview of Jacobian, pseudo-inverse method, Damped Least Square Method is given.

In the experiments, results of refinement using different inverse kinematics methods are obtained. The analysis and comparisons of the results are given. A typical segment of a smooth motion data is used to show the similarity between the frame-by-frame refined stick figure motion and the key-frame interpolated motion. At the end, the damped least square method using the IP-based solution as the initial solution is generally top-choice solution in terms of efficiency and precision.