在本篇論文中,我們提出一個系統,此系統於多個物件下的深度圖中自動定位物體並得到其相對應的姿態估測,並可以將之拓展至機器人之應用上。所提出的物體定位演算法,最主要是根據關鍵點的擷取,並結合FPFH特徵描述法及RANSAC演算法尋找正確的對應關係。 關鍵點的偵測主要是從針對二維影像之Harris偵測方法延伸而來,此方法於立體空間中利用網格的關係建立點與點的資訊,並進一步以一個點以及其周圍的點為資訊應用並實作為三維Harris偵測方法,接著以FPFH特徵描述法描述關鍵點中各個關鍵點所代表的特徵,計算並找到相似的對應點集合,最後利用幾何RANSAC演算法,結合幾何的特性從相似的對應點集合中選取正確的組合。在我們所提出的系統中,結合了關鍵點偵測以及RANSAC演算法以偵測物體,並利用ICP演算法修正物體定位的結果,並利用正確的對應點計算剛體轉換並得到姿態估測。 在我們的實驗,我們藉由立體的物件模型以及模擬生成的深度資料中來評估我們所提出的方法在姿態估測的誤差,最後並展示其在真實資料的實驗結果。
In this thesis, we propose a system for automatic object detection and pose estimation from a single depth map containing multiple objects for robot applications. The proposed object detection algorithm is based on matching the keypoints extracted from the depth image by using the proposed geometry-based RANSAC algorithm with the FPFH descriptor. The keypoint detection method used in this work is extended from the 2D Harris corner detector to the 3D Harris corner detector. Then, similar corresponding points with FPFH feature are extracted based on their distance. The proposed geometry-based RANSAC algorithm integrates the characteristics of the geometry to choose the inliers from similar corresponding points. In the proposed system, we combine the keypoint detection and the geometry-based RANSAC algorithm to detect the objects, followed by the ICP algorithm to refine the 3D object alignment. We exploit the corresponding points to calculate the rigid transformation for pose estimation. In the experimental results, simulated and real world depth data are shown to demonstrate the accuracy of pose estimation by using the proposed system.