Model-driven visual data analytics for monitoring work-in-progress on construction sites

Adherence to project schedules and budgets is the most highly valued performance metric among project owners. Despite its significance, most projects struggle to keep track of accurate as-built status. More than 53% of typical construction projects are behind schedule, and more than 66% do not meet their budget requirements. A few major factors accounting for such delays and cost overruns include 1) inconsistency among contractors, subcontractors and owners in terms of how much a construction project is faring at any given date, 2) flawed performance management due to lack of infrequent reporting of actual performance to project teams, and 3) planners' missed connections to most up-to-date construction progress information. To address these inefficiencies and contribute to the National Research Council (NRC)'s goal of improving efficiency in the construction industry, this dissertation proposes a construction progress monitoring framework that streamlines the utilization of existing large collections of site photographs – captured with consumer grade cameras, commodity smartphones as well as Unmanned Aerial Vehicles (UAV) - together with Building Information Modeling (BIM) for automated detection, analysis, and visualization of progress deviations at the operation level. To do so, several computer vision algorithms are developed to automatically create 4D as-built point clouds using the collected images (with or without using BIM as a priori) and to automatically analyze their deviations with as-planned 4D BIM models based on both geometry and visual appearance features. A reasoning mechanism based on formalized sequences of construction activities and inter-dependency of BIM elements is also presented which alleviates problems associated with limited visibility in visual capture processes, as well as lack of details in as-planned representations. To enhance practicality of the proposed framework, a crowdsourcing method is also proposed to enhance the accuracy and completeness of the visual data that is necessary to train the underlying machine learning algorithm used for visual data analytics. The proposed methods are validated using a large range of real-world construction datasets under normal and challenging conditions. The results of applying these computer vision methods show that appearance based recognition of construction materials and their comparison to BIM (with or without using geometry) outperforms the state of the art geometrical based method. Feedbacks from industry practitioners also show that reasoning based methods are acceptable for inferring progress for incomplete site datasets. The framework provides an easy and quick solution for project-level monitoring and provides project teams with a mechanism for better understanding of how a project compares with others in terms of cost, schedule, and labor hours. It also enhanced communication by providing real-time project information, improving onsite decision-making and work-sequencing, and fostering collaborative partnerships.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Civil Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2016

Author dc:creator

• Han, Kook In

Contributors dc:contributor

• Golparvar-Fard, Mani
• Liu, Liang
• Hoiem, Derek
• El-Rayes, Khaled
• Haas, Carl

• Construction Progress Monitoring
• Computer Vision
• Sequencing Knowledge
• Building Information Modeling
• Material Classification
• Big Visual Data

Statement dc:rights

• Copyright 2016 Kook In Han

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/92910

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F92910