基於曲率變化之關鍵零件定位物件姿態估計

Abstract

物件姿態估計技術在工業與虛擬技術等三維相關任務中被廣泛的運用，一直以來也是許多機構研究的重要課題之一，其中包含了機器手臂抓取任務、精密儀器的零件組裝及虛擬擴增實境的人與場景互動等等，根據不同的使用環境下所面臨的挑戰的重要性程度也會有所不同，如遮蔽、光影變化和速度等等。然而在傳統上，由於深度相機的價格較為昂貴且取得不易，因此有較多研究較著重在彩色影像的探討上，經典的方法如模板比對等。 但隨著近年來深度相機技術的蓬勃發展，使得深度相機的成本顯著降低，這也讓越來越多的應用開始導入了深度資訊，使得機器從原本的二維彩色世界開始邁向了三維的世界。本論文提出一套六自由度物件姿態估計系統，系統的主要發想來自於由下而上的人體姿態估計方法，該系統首先基於一3D物體曲率的大小自動化地將其劃分出數個關鍵部件，接著通過各關鍵部件內的點彼此之間互相投票的機制，來找出位在每個關鍵部件中最核心的那個點，最後採用隨機抽樣一致演算法來進行姿態估計；此外，為了要更快速地找出關鍵部件，在本論文中也採用了二維辨識技術來進行加速，該方法不但能更快速地找出關鍵部件，並且也能考慮在曲率變化較大的區域之色彩資訊。在本論文中有測試在不同的情況下如遮蔽、光影變化、執行時間和多尺度下等實驗。

Object pose estimation is widely used in industry, and 3D virtual technology and also has been one of the most critical studies by many institutions. Including robotic grasping, precision parts assembly and virtual reality, etc. The proportion of challenges faced in different user environments will also be different, such as occlusion, light and shadow changes, and speed. Traditionally, because depth cameras are expensive, they focus more on the discussion of color images. In recent years, more and more machines have begun to add depth information. We proposed an 6DoF object pose estimation method that combines depth information and color images. Ideas come from the bottom-up human pose estimation. The method can automatically divide an object into several key-component. Use Gaussian curvature to find the region of interest and group them. Then, finding the key regions through the voting mechanism in the component and localizing the center points in each area. In addition, for the requirement of speed, apply real-time two-dimensional detection techniques to optimize the method. Finally, using RANSAC to estimate the pose of the object. Some experiments such as occlusion, light and shadow changes, execution time and multi-scale show in this thesis.