應用影像辨識分析挖土機之行為模式及其產值

Applying Image Recognition to Analyzing Excavator Activity Recognition and Productivity

吳勻棠(Yun-Tang Wu)

指導教授 ： 詹瀅潔

國立臺灣大學/工學院/土木工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202102313

摘要

營造產業占我國GDP 2.5%，為台灣經濟上不可或缺的一部分，各項工程都必須在工期內完成，控管營造產業的施工進度還有效率成為是否能在工期內完工的因素之一。然而，目前的工程控管嚴重依賴人工管理，這不僅會消耗人力更會造成更多不必要支出。
其中，挖土機為工地現場最常見的土方設備，挖土機在工地現場的使用時機多屬於工程早期-營建工程之基礎開挖，也常用於道路工程、老舊房屋之拆除等等。近年來，人工智慧普遍用在各個產業，營造業也不例外，此項科技不僅能減少人力成本，也能準確幫助營造業控管工程進度。其中，影像辨識技術即為此類工具之一，工程現場的監視器可取代現場人力。本研究透過影像辨識技術對於挖土機進行行為模式偵測，利用網路上提供的現場工地照片作為訓練資料，框選出挖土機的上部結構以及挖斗進行訓練。之後便進一步利用影片中框選出來的定界框之座標以及幀數差加以判斷挖土機的行為模式，分別為:右轉、左轉、挖掘、前進、後退、閒置六項行為模式，爾後更可以利用上述已判斷的行為模式計算出挖土機的產量，此模型對於行為模式的準確度中閒置達96%、移動90%、工作30%，產能則達85%的準確率，以此研究幫助營造業控管挖土機之產能。

關鍵字

挖土機 ； 行為分析 ； Yolov3 ； 產值

並列摘要

The construction industry accounts for 2.5% of Taiwan’s GDP and is an indispensable part of Taiwan’s economy. All projects must be completed within the construction period. Controlling the construction progress and efficiency of the construction industry has become one of the factors that can be completed within the construction period. However, the current engineering control and management rely heavily on manual management, which not only consumes manpower but also causes more unnecessary expenditures.
Excavators are the most commonly used earthmoving equipment on the construction site. The use of the excavator mostly belongs to the early stage of the construction project, and it is also often used for road construction, demolition of old houses, and so on. In recent years, artificial intelligence has been widely used in various industries, and the construction industry is no exception. This technology can not only reduce labor costs, but also accurately help the construction industry control the progress of projects. Among them, image recognition technology is one of the tools. On-site monitors could replace manpower. This study uses image recognition technology to detect the behavior of the excavator, and uses the site photos provided by ACID as training data. Label the main body of the excavator and the bucket for training. After that, the coordinates of the bounding box and the number of frames are further used to judge the behavior of the excavator, which are: turn right, turn left, dig,forward, back, and idle. Furthermore, use behavior to calculate the productivity of excavator. This model has an accuracy of 96% for idle, 90% for move, and 30% for work. An accuracy of 85% for the productivity. This study helps the construction industry to control the productivity of the excavator.

並列關鍵字

Excavator ； Motion Analysis ； Yolov3 ； Productivity

參考文獻

[1] “台灣經濟研究院各業產值概況, ” 2021. [線上].Available:https://www.cier.edu.tw/cef/P2.pdf

[2] “營造業經濟概況,” 09 12 2020. [線上].Available: https://www.cpami.gov.tw/filesys/file/statistic8/1100305-1.pdf

[3] 林進南, “我國營造業景氣之現況與展望 (今日合庫) ,” 04 01 2018. [線上]. Available:https://www.tier.org.tw/achievements/pec3010.aspx?GUID=a83eeac2-986f-4dcc-b2b6-7b24734fbcb9

[4] “台經院產經資料庫, ” 2019. [線上]. Available: https://tie.tier.org.tw/db/industry_definition/class_2.aspx?code=IND11-12

[5] J. S. Bohn, J. Teizer, "Benefits and Barriers of Construction Project Monitoring Using High-Resolution Automated Cameras, " Journal of Construction Engineering and Management, p. 35, 2010.