營建生產作業行為自動辨識系統

MOTION-SENSING IDENTIFICATION SYSTEM FOR CONSTRUCTION OPERATION

楊廷曄(Ting-Yeh Yang) ； 薛憲徽(Sian-Jhen Syue) ； 曾仁杰(Ren-Jye Dzeng)

《中國土木水利工程學刊》 32卷1期 (2020/03) Pp. 75-90

https://doi.org/10.6652/JoCICHE.202003_32(1).0007

摘要

生產效率評估有助於營造廠商評估勞力成本以及規劃作業工期，現行有現場工作抽樣、DEA資料包絡分析等方法可使用，然而這些方法皆屬事後分析（非直接即時量測生產過程），亦需仰賴人為判斷，以及有限可供人工處理之採樣頻率。本研究以深度攝影技術補捉工人之骨骼節點，並建立演算法及系統，根據工人姿態辨識其是否有在生產。當生產作業（如模板）己知時，並可進一步自動辨識其子項作業類別（如模板組立、釘模板）。研究並針對常見施工作業，包含鋼筋綁紮、模板組立、搬運作業、讀圖溝通、砌磚作業、磁磚鋪貼六項生產作業，評估系統之辨識準確率。結果顯示，在已知受測者之作業類別（如已知進行模板組立中）時，在已知狀態下，系統針對上述各作業之生產行為之辨識準確率依序為92.23%、80.19%、90.82%、90.65%、62.24%和94.40%。

關鍵字

營建自動化 ； 作業行為辨識 ； 深度攝影 ； 體感姿態感測

並列摘要

Productivity assessment helps contractors estimate labor cost and activity duration. Some methods such as work sampling or Data Envelope Analysis can be used to assess productivity. However, they are post-analyzed based on recorded video of construction activities instead of real-time assessment. Their base upon human judgment also limits the feasible sampling rate of the video. This research uses depth cameras to capture joints of human skeleton and builds a system to automatically determine whether a subject's posture is a productive or nonproductive in a real time fashion. When the target activity (e.g., formwork) is known, the system may further categorize the subject's posture into the associated sub-activities (e.g., formwork assembly, formwork nailing). Experiments, which targeted on common construction activities including rebar assembly, formwork assembly, moving materials, reading blueprints, laying bricks, and tiling, were conducted to evaluate the identification accuracy. The results show that accuracies are 92.23%, 80.19%, 90.82%, 90.65%, 62.24%, and 94.40%, respectively.

並列關鍵字

construction automation ； work posture analysis ； depth camera ； motion sensing