營建工程常由於開放且動態不確定的施工環境以及緊迫的施工進度要求,導致勞工安全經常被忽略。儘管政府部門非常注重此一議題,但很難完全地排除施工意外發生。其中,營造重大職災死亡災害發生之類型,以「墜落與滾落」為最常見。因此,對於預防營建職災發生應首重墜落事故之預防。本研究以更快速區域卷積式(Faster R-CNN)深度學習類神經網路為基礎,結合簡易之星狀圖形辨識標記,以建構有效之工地開口安全防護設施狀態辨識方法。利用營建工地常見之攝影機所取得的即時影像資料之視覺辨識分析,達到開口安全防護設施安全無虞之目的,作為降低墜落事故之目標。本研究經標記圖形選擇、資料收集、訓練參數分析及網路訓練後,以實際工地影像資料進行測試,經訓練參數調整後,包括電梯直井、地板開口及施工圍籬等三類目標之最終誤報率介於2.1%至8.4%間。此一結果證明本研究所提出之方法,具有監控安全設施危害狀態之實務應用潛力。
The construction managers tend to overlook the safety issue due to the dynamic and open site environment and the tight schedule. Despite the tremendous efforts spent by the Government labor safety agencies; the construction accidents are hardly eliminated completely. Falls from height due to temporary construction openness has been the most common accidents on site. As a result, this research aims at developing a fall prevention method by integrating a deep learning neural networks-based computer visualization technique and a specially designed tag to identify the openness status of fall protection facilities on site. The Faster Region based Convolutional Neural Networks (Faster R-CNN) is adopted and a specially designed Star-shape tag is selected for implementation of the proposed method. Such a method can determine the openness status of the safety facility images obtained from a regular Closed-Circuit Television (CCTV) camera set up on site. After tuning with various network parameters, the field-testing results show that the proposed method can achieve False Positive Rate with 2.1% ~ 8.4% for three types of safety facility openings, i.e., elevator shaft openness, floor openness, and safety fence openness. Such an outstanding performance implies profound potentials for practical application in monitoring the status of safety protection facilities on construction sites.