結合AI影像技術與自製末端執行器之餐具分揀系統

NCU Institutional Repository > 資訊電機學院 > 電機工程研究所 > 博碩士論文 > Item 987654321/99407

請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/99407

題名:	結合AI影像技術與自製末端執行器之餐具分揀系統
作者:	莊惟婷;Chuang, Wei-Ting
貢獻者:	電機工程學系
關鍵詞:	餐具分揀;機械手臂;YOLOv8分割模型;朝向估測;物件辨識;自動化;cutlery sorting;robotic arm;YOLOv8 segmentation model;orientation estimation;object recognition;automation
日期:	2025-11-20
上傳時間:	2026-03-06 18:55:21 (UTC+8)
出版者:	國立中央大學
摘要:	本論文旨在設計與實現一套全自動化金屬餐具(刀、叉、湯匙)分揀系統，使用者僅需將餐具倒置於待分類區域，系統可藉由機械手臂控制，自動將所有餐具分揀完成。金屬餐具廣泛應用於餐廳、航空餐飲與大型宴會場域，大量的餐具使用後清洗完畢，往往依靠人工進行分揀收拾作業，不但耗時也耗力。若是利用傳統影像辨識與機械手臂抓取來工作，金屬餐具因具備高反光、顏色相近和隨機堆疊等特性，實務上也面臨相當挑戰。因此，本論文提出結合影像辨識、末端執行器設計與機械手臂控制流程的完整系統架構，以解決現有的高人工成本與以上實務面對的問題。本系統涵蓋三大部分：於影像辨識方面，考量到實際場景中的餐具複雜堆疊、角度隨機與光照多變，結合資料擴增技術建立的金屬餐具資料集，訓練出最佳模型進行即時物件辨識與分割；於餐具定位方面，採用主成分分析方法(Principal Component Analysis, PCA)作為朝向角度演算基礎，並建立抓取點估算流程，能有效定位餐具並保持分揀過程的穩定性與成功率；在硬體方面，設計一個具備電磁彈性觸發機制的末端執行器模組，整合電磁鐵、微動開關與彈簧等組件，成功補償深度攝影機於反光表面上的量測誤差，大幅提升操作精度與系統整體效能。實驗結果中，YOLOv8x-seg模型為最佳，在遮罩(Mask) mAP0.5-0.95的指標分數為0.865，於測試集的實驗驗證中，餐具朝向分析與抓取點演算法之準確率達99%。在光照條件700 lux實驗場域下，整體系統分揀成功率為96.19%。基於上述分揀策略，本研究亦延伸設計自動化餐具擺盤系統，成功率達90%。綜合而言，本論文成功建立一套從影像辨識、抓取到分類與擺盤的完整自動化流程，不僅有效降低人力負擔，亦展現了智慧餐飲自動化的應用潛力，對後續相關研究與實務部署具有重要的參考價值。 ;This thesis aims to design and implement a fully automated system for sorting metallic cutlery (knives, forks, and spoons), where users place cutlery in a designated area for the robotic arm to handle automatically. Metallic cutlery is widely used in restaurants, airline catering, and large-scale banquet venues. After extensive washing, the sorting and packing process typically relies on manual labor, which is both time-consuming and labor-intensive. Utilizing traditional image recognition and robotic arm grasping for this task also presents significant challenges due to the high reflectivity, similar colors, and random stacking characteristics of metallic cutlery. Therefore, this study proposes a comprehensive system architecture that integrates image recognition, end-effector design, and robotic arm control, thereby overcoming the problems of high labor costs and insufficient depth sensing accuracy. The proposed system consists of three main components. For image recognition, a dedicated dataset of metallic cutlery was established, with data augmentation to account for random orientations, complex stacking, and variable lighting conditions. The dataset was then used to train an optimized model capable of real-time object recognition and segmentation. For cutlery localization, the system used Principal Component Analysis (PCA) to estimate orientation, followed by a grasp-point estimation procedure to ensure accurate localization and stable sorting. Regarding hardware design, we developed an end-effector module with an electromagnetic elastic triggering mechanism. This module integrates electromagnets, micro-switches, and springs to compensate for depth camera measurement errors on reflective surfaces, thereby improving overall precision and system performance. Experimental results demonstrated that the YOLOv8x-seg model achieved the best performance, with a mask mAP0.5–0.95 of 0.865. On the test set, the cutlery orientation and grasp-point estimation algorithm achieved 99% accuracy. With lighting below 700 lux, the overall sorting success rate reached 96.19%. Furthermore, based on the proposed sorting strategy, an extended automated cutlery placement system was developed, achieving a 90% success rate. In summary, this thesis establishes a complete automated workflow that covers image recognition, grasping, classification, and placement. The system not only effectively reduces manual labor but also demonstrates the practical potential of intelligent automation in the food service industry.
顯示於類別:	[電機工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	45	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....