建筑室内施工形象进度智能跟踪方法研究
雷 蕾
摘 要
建筑施工进度的跟踪与监测对保障项目进度和成本目标实现至关重要， 室内工程因施工参与方多、施工流水作业衔接混乱使其相较于室外进度跟踪变得更加复杂。如何准确有效地采集施工现场信息和及时检索、 分析收集的信息仍是室内进度跟踪的一个亟待解决的问题。现有的室内施工信息采集仍依赖于传统的人工巡检和记录，导致进度信息收集、处理和分析效率低，难以满足复杂室内施工进度信息的有效传递与管理需求。 因此， 本文提出基于动态信息反馈理论的室内施工形象进度智能跟踪方法。 主要研究内容如下：
（1） 室内施工形象进度智能跟踪框架的构建。结合动态信息反馈理论，分析室内形象进度信息化管理内涵与意义，明确信息采集、识别和分析是信息化动态管理的重要支撑技术。从不同阶段施工进度信息构成和进度管理流程，阐述施工进度信息特点以及基于信息化进度管理的必要性。结合室内施工形象进度跟踪现实需求，分析现有形象进度信息化管理中关键技术的挑战与难点，提出面向点云信息采集、识别和分析的室内施工形象进度动态跟踪和监测方法。
（2） 室内弱纹理场景施工形象进度信息采集方法。阐述现有基于计算机视觉的快速建模方法的适用范围和室内应用的局限性，针对室内施工弱纹理场景因特征提取难、数据冗余和不利光照条件干扰，导致难以高效、准确获取室内施工形象进度信息问题，构建了室内弱纹理场景施工进度信息采集模型。面向室内小尺度建筑材料等，提出了基于分层优化的深度相机室内小尺度目标快速建模方法；面向室内大尺度二次结构等，提出了基于视觉引导的移动终端大尺度目标快速建模方法， 由此，实现了高效、准确的室内弱纹理场景施工形象进度信息的采集，为室内施工形象进度的智能跟踪提供了数据基础。
（3） 室内施工形象进度信息识别方法。面向室内进度关键要素识别需求，结合深度学习对施工场景图像的分类与识别优势，针对室内施工场景相互遮挡多、纹理相似度高、边缘识别难等问题，构建基于点云的室内形象进度跟踪要素语义分割和识别模型。以 RandLA-Net 模型为基础， 本文提出了融合粗分割和精分割的 RF-RandLA-Net模型。 相较于前者该方法平均类别精度和平均交并比提升明显。 由此，实现基于点云的施工形象进度跟踪要素识别，为后续室内实景点云施工形象进度动态跟踪与分析提供依据。
（4） 室内施工形象进度信息分析方法。考虑到室内安装工程变更多，进度敏感性高，构建基于室内实景点云的施工形象进度信息动态分析与跟踪框架。为识别间隔时间内获取的室内施工形象进度跟踪要素的点云差异，提出基于点云块特征的自动配准方法，相较于 FPFH 算法在召回率和精确率方面提升明显。最后，基于施工形象进度的信息采集、信息识别和信息分析方法，实现进度智能跟踪与推送系统原型设计与开发。
本文结合计算机视觉和人工智能技术应用于室内施工形象进度管理，从施工工序层面实现室内形象进度的跟踪和监测，为施工进度精细化管理提供新思路。
关键词：室内施工；形象进度跟踪；信息模型；计算机视觉；深度学习；点云数据
Abstract
Tracking and monitoring of construction progress are vital for achieving project schedule and cost objectives. However, indoor progress tracking poses greater complexities compared to outdoor projects, primarily due to the involvement of numerous construction participants and the intricate interconnection of construction flow operations. The accurate and effective collection of on-site construction information, along with its timely retrieval and analysis, remains a pressing challenge in indoor progress tracking. The existing methods for indoor construction information collection still rely on traditional manual inspection and recording, resulting in low efficiency in progress information collection, processing, and analysis. Consequently, it difficult to meet the needs of effectively transmitting and managing complex indoor construction progress information. To address this issue, this thesis proposes an intelligent tracking method that based on dynamic information feedback theory for indoor construction virtual progress. The main research contents are as follows:
(1) Construction of an intelligent tracking framework for indoor construction visual progress. Combining with the dynamic information feedback theory, the connotation and significance of informatization management of indoor visual progress are analyzed, and it is clarified that information collection, recognition, and analysis are important supporting technologies for dynamic information management. Based on the construction progress information composition and management process at different stages, the characteristics of construction progress information and the necessity of informatization progress management are elaborated. Considering the practical needs of indoor construction visual progress tracking, the challenges and difficulties of key technologies in existing visual progress informatization management are analyzed, and a method for dynamic, real-time tracking and monitoring of indoor construction visual progress based on information collection, recognition, and analysis is proposed.
(2) Method for indoor construction visual progress information collection. This study examines the applicability and limitations of existing computer vision-based 3D reconstruction methods, specifically focusing on the challenges of indoor construction environment with weak textures. These challenges include difficulties in feature extraction, data redundancy, and interference from unfavorable lighting conditions, which hinder the efficient and accurate acquisition of indoor construction progress information. To address this issue, a framework for collecting visual progress information in indoor textureless construction environments is developed. For small-scale indoor materials, a hierarchical optimization-based method using depth cameras is proposed for rapid 3D reconstruction. For large-scale secondary structures in indoor environments, a visual-guided approach utilizing mobile devices is introduced for rapid 3D reconstruction. This approach enables efficient and accurate collection of indoor visual progress information in textureless construction environments and provides a solid foundation for intelligent tracking of indoor construction visual progress.
(3) Method for indoor construction visual progress information recognition. Focusing on the requirements of key elements in indoor construction progress, leveraging the advantages of deep learning in construction scene image classification and recognition. It tackles challenges such as mutual occlusion, high texture similarity, and difficult edge recognition in indoor construction scenes. A point cloud-based semantic segmentation and recognition model is proposed for tracking elements of indoor visual progress. Building upon the RandLA-Net model, a refined and fused segmentation model called RF-RandLANet is proposed. Compared to the former, the RandLA-Net method significantly improves indicator of mean accuracy and mean intersection over union. As a result, the proposed method enables the recognition of construction visual progress elements from point clouds, providing a foundation for subsequent dynamic construction progress tracking and analysis based on indoor as-built point cloud.
(4) Method for indoor visual progress information analysis. Considering the high sensitivity of indoor installation engineering changes and progress, a dynamic analysis and tracking framework for construction progress information based on indoor real-scene point cloud is constructed. To identify the point cloud differences of construction progress tracking elements obtained in interval time, an automatic registration method based on point cloud patches’ features representions is proposed, which improves the indicator of the recall rate and precision rate compared to the FPFH algorithm. Finally, based on the information collection, identification, and analysis of construction progress, an intelligent progress tracking and notification system prototype is designed and development.
This thesis focuses on the application of computer vision and artificial intelligence technology in the management of indoor construction progress. It aims to achieve the tracking and monitoring of indoor progress at the construction process level and provides new ideas for lean construction progress management.
Key words: Interior construction, Visual progress monitoring, Information model, Computer vision, Deep learning, Point cloud data
国家数字建造技术
创新中心
