摘要
建筑物结构外观多样和施工现场存在遮挡、杂乱背景等因素导致工地区域识别难度大,传统的识别方法效率低、效果易受到主观因素的影响。因此,提出了一种基于像素级分割算法的深度学习模型,对施工工地的各个区域进行智能识别,以提高工地管理的自动化水平。通过引入先进的DeepLabv3+网络对无人机航拍图像中各个工地区域的分布进行识别,通过改进DeepLabv3+网络的主干网络和优化器以提升对工地区域的识别效果,结果显示最佳的识别正确率、Io U和F1-Score分别是88%、0.85和0.88。基于深度学习的施工工地区域识别技术可以实现对施工工地的实时监控和管理,提高工地管理的智能化水平,从而提升施工效率、降低安全风险,并减少资源浪费。
Identifying construction site areas is challenging due to the diverse appearance of building structures,occlusion,and chaotic backgrounds at the site.Traditional methods are inefficient and prone to subjective influence.A deep learning model based on pixel-level segmentation is proposed for intelligent identification of various site areas,aiming to improve the automation level of site management.The advanced DeepLabv3+network is introduced to recognize the distribution of site areas in UAV aerial images,with improvements made to the backbone network and optimizer to enhance recognition performance.Results show the best recognition accuracy,IoU,and F1-Score are 88%,0.85,and 0.88,respectively.The deep learning-based construction site identification technique enables real-time monitoring and management,enhancing site management intelligence,improving construction efficiency,reducing safety risks,and minimizing resource waste.
作者
梁嘉韵
温喜廉
杨智诚
陈广浩
杨永民
LIANG Jiayun;WEN Xilian;YANG Zhicheng;CHEN Guanghao;YANG Yongmin(Guangzhou Pearl River Construction Development Co.,Ltd.,Guangzhou 510075,Guangdong,China;School of Civil Engineering,Guangzhou University,Guangzhou 510006,Guangdong,China;College of Urban and Rural Construction,Zhongkai University of Agriculture and Engineering,Guangzhou 510225,Guangdong,China;Guangdong Lingnan Township Green Building Industrialization Engineering Technology Research Center,Guangzhou 510225,Guangdong,China)
出处
《建筑施工》
2024年第12期1946-1950,共5页
Building Construction
基金
广东省住房和城乡建设厅科技创新计划项目(2023-K1-463769)
广州市科技计划项目(2023A04J0647)。
关键词
深度学习
工地区域自动识别
语义分割
智能识别
工地管理
deep learning
automatic recognition of construction site regions
semantic segmentation
intelligent recognition
site management
