應用於自動化生產及分揀之物件姿態估測系統
No Thumbnail Available
Date
2020
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
近幾年來，產業為了提升生產效率，大量使用自動化生產設備取代人力，透過電腦視覺與機器運動控制的整合搭配，已大幅增加自動化生產的效率。受惠於GPU計算平台的普及，不論機器學習或是深度學習技術紛紛出現於各種應用場景之中，以往使用電腦視覺方法不能或是難以解決的問題，透過引進深度學習都有出色的表現。本文主要研究內容可分為三部分：第一部分利用輝達(Nvidia)所提出之基於深度學習單攝影機物件姿態估測演算法(Deep Object Pose Estimation, DOPE)，其中包含產生物件的立體模型，再匯入Unreal Engine遊戲引擎並搭配輝達深度學習資料集合成器(Nvidia Deep learning Dataset Synthesizer, NDDS)套件，產生訓練數據，用來對神經網路進行權重訓練，完成後即可用來對物件姿態進行估測；第二部分使用加拿大Kinova公司所生產之Jaco 2四軸機械手臂並透過機器人作業系統(Robot Operating System, ROS)完成物件夾取功能；第三部分運用PyQt設計一圖形使用者介面(Graphical User Interface, GUI)整合前兩部分，讓使用者透過單一介面即可獲得物件估測和手臂執行資訊，也可透過其進行參數調整。模擬於生產線上應用，用以輔助加工與分類之程序，達成自動化生產製造之目的。
In this thesis, we propose an object pose estimation system for pick and place automation. There are three parts of the system. In the first part, we use a single camera to develop an object pose estimation system based on Deep Object Pose Estimation (DOPE). Then we use a Kinova JACO 2 4-DoF robotic arm to perform object picking through Robot Operating System (ROS). Finally, a Graphical User Interface (GUI) is designed to integrate the pose estimation system and robotic arm, so that users can obtain object estimation and arm execution information through a user friendly interface. To validate the viability, the proposed system is applied to a simplified production line automation environment. The experiment results show the robotic arm is able to properly pick objects base on the pose estimated, thus processing and sorting automation is achieved.
In this thesis, we propose an object pose estimation system for pick and place automation. There are three parts of the system. In the first part, we use a single camera to develop an object pose estimation system based on Deep Object Pose Estimation (DOPE). Then we use a Kinova JACO 2 4-DoF robotic arm to perform object picking through Robot Operating System (ROS). Finally, a Graphical User Interface (GUI) is designed to integrate the pose estimation system and robotic arm, so that users can obtain object estimation and arm execution information through a user friendly interface. To validate the viability, the proposed system is applied to a simplified production line automation environment. The experiment results show the robotic arm is able to properly pick objects base on the pose estimated, thus processing and sorting automation is achieved.
Description
Keywords
深度學習, 機器人作業系統, 物件姿態估測, 資料集生成, 機械手臂, 圖形使用者介面, deep learning, ROS, object pose estimate, synthetic data, robotic arm, GUI