運用深度學習技術之影像語意分割於施工中電梯直井墜落風險監控

APPLICATION OF IMAGE SEMANTIC SEGMENTATION USING THE DEEP LEARNING TECHNIQUE IN MONITORING THE FALL RISK OF CONSTRUCTION WORKERS IN A BUILDING ELEVATOR SHAFT

營建施工意外一直高居世界各國產業職災之首，究其原因在於營建工地具有高度開放與動態特性。尤其是電梯直井在未安裝設備之前，因安全防護不足所導致之高空墜落事故，常是造成建築工地死亡事故的主要原因。雖然法令規範已要求裝設安全防護設施及個人防護裝備，理論上應有充足的安全保障；然而由於施工人員的不安全行為導致防護措施失效而造成之事故仍時有發生。為加強輔助職安人員電梯直井之安全管理，本研究提出以語意分割深度學習技術為基礎之人工智慧電腦視覺辨識方法，開發建築工地電梯直井安全監控輔助系統；自動警示因不當行為導致潛在安全危害，並降低職安管理人員之工作負擔。本研究建構完成之系統可應用於建築工地現場之各電梯直井安全區域監控，透過實驗室訓練測試結果發現，所提出方法不但訓練之召回率及精確率皆高於95%，甚至在工地實測之正確率與純淨度亦高於95％，可見其在產業實務之應用潛力。故本研究之成果應可作為營建施工安全管理之輔助方法，以達到即時且準確管控建築工地電梯直井施工安全並降低墜落事故發生之風險。

Construction accidents are the most significant contributor to occupational disasters among all industries worldwide. This is due to the highly open and dynamic characteristics of construction sites. The unprotected elevator shaft before the installation of mechanical equipment is especially risky for construction workers. Although the safety regulations require the employer to provide sufficient safety behavior facilities and the workers to wear personal protection devices, accidents still occur due to the unsafe practices of the workers. In order to improve the situation, this paper presents an image semantic segmentation method using the Deep Learning (DL) technique for monitoring the fall risk of construction workers near the building elevator shaft. Based on both the laboratory results and in-situ testing, it has been found that the Recall and Precision during the training process in laboratory and the Cleanness and Correctness obtained on site surpassed 95% high performance criteria. It has been concluded that the proposed method provides construction safety personnel an effective tool for monitoring the risks and preventing accidental falls for the construction workers.