碩士論文(2022)
電腦視覺為基之人工作業績效評量方法與技術研發
Development of Computer Vision-based Manual Performance Evaluation Method and Technology
關鍵字 Keywords
摘要
製造為國家經濟命脈，世界各國也極為重視製造業的發展。在生產作業中，業者透過規劃(P)、執行(D)、監控(C)、評量與行動(A)來提升生產績效。物聯網的普及與人工智慧的發展，使產業進入智慧製造時代，透過即時資料蒐集與分析，掌握生產現況與趨勢以及偵測異常，也提供績效評量與持續改善之依據。然而，傳統中小企業因資金與技術能力不足，仍需仰賴人工的作業。我國產業仍以傳統與中小企業為主，在此工業4.0的時代，如何為人工作業之監控與績效評量提供人工智慧的解決方案，仍是當前業界與學界的課題。
針對中小企業人工作業之智慧化監控與評量之需求，本研究提出一「電腦視覺為基之人工作業績效評量方法」，並開發「人員辨識」、「動作評量」與「品質檢驗」技術，提供一套人工作業監控與評量之人工智慧解決方案。
本研究所提「人員辨識」技術，可解決戴口罩之場域的人員辨識，在1張照片多張人臉的辨識準確率達93%以上，並可實際應用於即時之人員辨識；在動作評量方面，「動作評量」技術能夠區分合格與不合格的動作，準確率高達98%；在產出品質檢驗方面，能夠有效分辨出合格的產出與不合格的產出，辨識準確率達75%以上，在某些測試場景中，甚至高達100%。實驗證實，本研究所提之人工作業績效評量方法與技術之可行性與有效性。
Abstract
In the manufacturing process, humans still play an indispensable role. For example, in the tasks where machines cannot fulfill the current needs and rely on humans, and in quality inspection through sensory inspection (Vigneau et al., 2018). Therefore, even in the modern era of 2022, although many industries have moved into the realm of Industry 4.0, human still plays a significant role in the production process.
However, human is not cold machine that can operate for long periods of time and whose output status can be understood through parameter settings. Therefore, it is very important to understand the performance of the personnel. The traditional performance measurement method is to calculate the total output, effective output, and ineffective output of the workers mainly by "manual" calculation, which is very unreliable and costly.
In view of this, this study hopes to help traditional SMEs reduce human intervention and cost by using artificial intelligence, data science, and computer vision in the “manual performance evaluation" section, and proposes a manual performance evaluation method, which is subdivided into "personnel identification model", "action evaluation model", and "output quality inspection model" to help SMEs grasp more comprehensive information when making decisions.
In the proposed method, the “personnel identification model” can solve the recognition of personnel in the field of wearing masks, and the recognition accuracy of multiple faces in one photo is over 93%, and can be applied to real-time personnel recognition. In the “action evaluation model”, it is able to distinguish between qualified and unqualified actions with an accuracy rate of 98%. In the “output quality inspection model”, it can effectively identify the qualified output and the unqualified output with an accuracy rate of more than 75% and even up to 100% in some testing scenarios; and the analysis results obtained from the above model can be used for manual performance evaluation to analyze the performance of personnel. The experiments and results show that the method is feasible and effective.