基於CAD模型之物體姿態辨識及其於機械臂隨機堆疊抓取之應用

Abstract

本論文的主要目的是針對Kinect RGB-D感測器開發基於CAD模型之物體姿態辨識及其於機械臂隨機堆疊抓取之應用。本論文運用投票機制(voting scheme)進行待抓物件之6-DOF姿態辨識，使能在一組不同種類雜亂堆疊工件堆裡辨識個別工件及其6-DOF姿態。在資料庫建立方面，本論文結合物件的3D CAD Model及Virtual Camera快速地建立姿態辨識所需的資料庫，然後利用Voxel Grid Filter大幅度減少物件的3D點雲數量以降低姿態辨識所花費的時間。在機械臂抓取物件方面，本論文利用Outlier Filter剔除錯誤的姿態及被遮蔽但正確辨識的姿態，使機械臂總是可以抓取工件堆最上層的工件以提高抓取成功率。透過分析物件與機械臂yaw旋轉角度的相對姿態關係，機械臂可以順利抓取任意姿態的工件。本論文以Kuka 6-DOF機械臂進行實驗，驗證所提出的方法在一連串多物件隨機堆疊抓取實驗當中，所設計的系統抓取隨機姿態工件的成功率為89.7%，且可成功抓取工件直到箱子內沒有工件為止。

In this thesis, we propose a CAD-based 6-DOF pose estimation design for random bin-picking of multiple objects using a Kinect RGB-D sensor. A voting-scheme was adopted for 6-DOF pose estimation as well as recognition of a set of 6-DOF poses of different types of objects in the bin. We combine 3D CAD model of objects with a virtual camera to generate point cloud database for pose estimation. Voxel grid filter is suggested to decrease the number of 3D point cloud of object for reducing time of pose estimation. Furthermore, we use an outlier filter to filter out bad matching poses and occluded ones, so that the robot arm always picks up the upper object in the bin to increase pick up success rate. The yaw rotation angle of object relative to robot arm is calculated, so that robot arm can pick up different types and random poses of objects. A series of experiments of practical random bin-picking of multiple objects reveals that our proposed system can pick up random pose of objects in the bin with a success rate of 89.7%, and the objects in the bin are all picked up until there is no object left.