無校正視覺伺服系統之影像Jacobian矩陣估測研究

Abstract

本文建立一個以影像為基礎的視覺回授系統，此系統使用不需要校正的相機，估測並更新影像Jacobian矩陣，進而控制機器手臂到達所需的位置。本文將視覺回授系統分成兩大部分。第一部份是影像特徵的擷取，我們把電腦視覺的平均位移法，應用在特徵擷取上，以提高特徵追蹤的準確率。平均位移法是以空間中的色彩分佈為特徵，並且以相似函數來鎖定追蹤的對象，在複雜的環境，也能即時準確的擷取特徵。第二部分是視覺回授的控制，我們利用簡單的演算法，估測相機與機器手臂末端的關係。相較於其他文獻，此方法易於實現。 在模擬的部分，本文透過仿真的相機與機器手臂，可以有效的模擬。而實作的部份，本文僅使用一台電腦，來完成所有的軟體開發，包括對機器手臂的傳輸與雙相機的處理，以解決全域與區域的Jacobian問題。最後，在實驗的部份，我們選擇使用球具的球類運動，加入了運動偵測的演算法，展示機器手臂連續擊球的任務，並將結果加以分析與討論。

This thesis proposes an image-based visual servo system, which can estimate and update Image Jacobian Matrix to control a robot arm without camera calibration. There are two parts in the visual servo system. The first one is image feature extraction. We apply the Mean-Shift algorithm in order to improve the performance of feature tracking. Mean-Shift algorithm, which takes the color distribution as a model, is based on the similarity measure function to decide a tracking candidate. This algorithm works well in a real-time and complicated environment. The second part is the control of a manipulator to track visual objects. An easy-to-implement algorithm is proposed to find the relationship between the camera and the manipulator. It can be carried out easily as compared with other methods. We use the camera and the manipulator emulator to simulate the system effectively. In the implementation, we use only one computer to develop all the software, including the communication to the manipulator and the processing of two cameras' images to solve the local and the global Image Jacobian matrices. Ball-hitting experiments such as the juggling task are presented to analyze the real-time performance of the proposed algorithms.