應用實例分割與點雲配準於平面放置物件之機械手臂分揀系統開發;Development of a Robotic Arm Sorting System for Planar-Placed Objects by Applying Instance Segmentation and Point Cloud Registration

NCUIR > Engineering College > Graduate Institute of Mechanical Engineering > Electronic Thesis & Dissertation > Item 987654321/99500

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/99500

Title:	應用實例分割與點雲配準於平面放置物件之機械手臂分揀系統開發;Development of a Robotic Arm Sorting System for Planar-Placed Objects by Applying Instance Segmentation and Point Cloud Registration
Authors:	周韋安;Zhou, Wei-An
Contributors:	機械工程學系
Keywords:	物件分揀;機械手臂;深度學習;點雲配準;Object Sorting;Robot Operating System 2
Date:	2026-01-30
Issue Date:	2026-03-06 19:08:54 (UTC+8)
Publisher:	國立中央大學
Abstract:	本研究針對醫療器械經清洗消毒後之回收處理流程，開發一套自動化分揀系統，旨在改善傳統人工作業仰賴人力判讀、易疲勞且效率受限之問題。在系統架構方面，硬體採用TM5-900協作型手臂搭配Ensenso N35立體相機，建立眼在手架構；針對扁平且金屬表面易滑動之器械特性，設計整合V型結構指夾與真空吸嘴的複合式末端執行器。軟體基於ROS2平台，整合YOLOv11實例分割與FastAPI後端介面，建構模組化的自動化控制系統。在技術方法上，提出「由粗至細」的配準策略。首先利用YOLOv11識別類別並生成遮罩以擷取目標點雲，接著利用FPFH(Fast Point Feature Histograms)特徵結合RANSAC(Random Sample Consensus)進行粗配準，最後透過點對平面ICP(Iterative Closest Point)演算法進行精細姿態修正，以計算器械在空間中的三維坐標。實驗結果顯示，針對高反光金屬器械，本系統之點雲配準平均適配度為0.969，均方根誤差控制在1.894 mm內。在隨機分類揀放實驗中，總共20個物件平均成功率為96.62%，平均處理耗時13.73秒。研究結果證實，本系統在定位準確度與作業效率上，已具備應用於醫療後勤自動化流程之可行性。 ;This study develops an automated pick-and-place system for the recycling process of medical instruments after cleaning and sterilization, aiming to address the efficiency limitations and fatigue issues associated with traditional manual sorting. Regarding the system architecture, the hardware integrates a TM5-900 collaborative robot with an Ensenso N35 stereo camera in an Eye-in-Hand configuration. A hybrid end-effector combining a V-shaped gripper and a vacuum nozzle was designed to handle flat instruments with slippery metal surfaces. The software is built on the ROS2 platform, integrating YOLOv11 instance segmentation and a FastAPI backend to establish a modular automated control system. For the technical methodology, a "Coarse-to-Fine" registration strategy is proposed. First, YOLOv11 is used to identify classes and generate masks to extract target point clouds. Then, FPFH(Fast Point Feature Histograms) features combined with RANSAC(Random Sample Consensus) are employed for coarse registration, followed by the Point-to-Plane ICP(Iterative Closest Point) algorithm for fine pose correction to calculate the 3D coordinates of the instruments. Experimental results indicate that for highly reflective metal instruments, the system achieved an average point cloud registration fitness of 0.969, with a Root Mean Square Error (RMSE) maintained within 1.894 mm. In random sorting and pick-and-place experiments involving 20 objects, the average success rate was 96.62%, with an average processing time of 13.73 seconds. These findings confirm that the system possesses the necessary positioning accuracy and operational efficiency for feasible application in automated medical logistics processes.
Appears in Collections:	[Graduate Institute of Mechanical Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	128	View/Open

社群 sharing

Loading...