基於深度學習於營建施工現場危害警示模型之建立
龍立偉 龍立偉,王裕仁 王裕仁
DOI: https://doi.org/10.53106/101632122023060124002
2023-06-01
建築學報
Abstract:近期因營建現場工安意外頻傳,2023年2月高雄市一處建築工地,施工人員在進行吊掛作業時,因吊車傾斜翻覆,吊臂壓到一名李姓勞工,造成職業災害;同年4月高雄市發生李姓工人正在焊接鋼筋時,遭吊車不慎碰撞,而發生工安意外;5月新北市吳姓磁磚公司老闆疑似誤觸操作堆高機,導致整個人懸空卡在機器上方,而發生憾事;因此施工人員與機具於現場作業時的安全距離與位置,常因為營建職場環境之複雜,而存有許多潛在風險,造成長期面臨職業災害的問題,包含重大公安事等發生。 在整個施工階段中,大多營造廠皆採用相機或攝影機等影像裝置來記錄現場的施工進度狀況,因影像資料較大、處理時間長,所以多數處於被動使用的狀態,最後僅於存查而未臻理想,所以本研究提出透過深度學習之影像辨識技術,利用卷積神經網路(CNN)具有強力的特徵提取功能,藉由濾波器(Filters)從輪廓、邊緣線條到局部特徵自動進行提取,再以全連接(Fully Connected)方式,將特徵資料輸入至整個神經網路進行訓練的一套工程現場施工危害警示系統,其對於施工影像辨識處理效果佳,達到即時監測施工現場的影像數據,並運用安全評估區塊模組快速識別潛在的危害情況,並發出警示信號,提醒工地管理人員立即採取適當的措施,確保職業安全。本研究將與營建公司合作,蒐集正在施工的建案影像和場地分析,驗證危害警示模型的可行性,並評估其對職業安全管理的實際效果,協助工地管理人員結合科技進步,運用更為有效的管理工具保障施工人員安全,減少工安意外發生,同時提高營建業的生產效率和競爭力。Recently, there has been a high frequency of construction site occupational accidents. In February 2023, at a construction site in Kaohsiung City, a crane tilted and overturned during lifting operations, causing the crane arm to press against nearby construction workers, resulting in occupational injuries. In April, a worker was welding steel bars and was accidentally struck by a crane, leading to a work safety accident. In May, the owner of a tile company in New Taipei City was suspected of mistakenly operating a forklift, causing the person to be suspended above the machine and resulting in a tragic incident. Therefore, the safety distance and positioning between construction workers and machinery during on-site operations are often compromised due to the complexity of the construction workplace environment, posing numerous latent risks and leading to long-term occupational hazard issues, including severe public safety incidents. In the construction industry, cameras or video devices are commonly used to record the progress of construction activities throughout the entire construction phase. However, due to the large size of image data and time-consuming processing, these devices are often passively utilized, resulting in suboptimal outcomes where the data is merely stored for reference purposes. To address this issue, this study proposes a construction hazard warning system that leverages deep learning techniques, specifically convolutional neural networks (CNNs), known for their powerful feature extraction capabilities. By automatically extracting features from contours, edges, and local characteristics through filters and subsequently inputting the extracted features into a fully connected neural network, a practical framework for on-site construction hazard recognition is established. The system enables real-time monitoring of construction sites by processing and analyzing image data. It utilizes a safety assessment module to identify potential hazards and issue warning signals quickly, alerting site managers to take immediate actions to ensure occupational safety. In collaboration with construction companies, this research collects construction project images. It conducts site analysis to validate the feasibility of the hazard warning model and evaluate its practical effectiveness in occupational safety management. By integrating technological advancements, this research aims to assist site managers in adopting more efficient management tools, safeguarding the safety of construction personnel, reducing work-related accidents, and enhancing productivity and competitiveness in the construction industry.