Three-Dimensional Reconstruction and Modeling Using Low-Precision Vision Sensors for Automation and Robotics Applications in Construction

Abstract

Automation and robotics in construction (ARC) has the potential to assist in the performance of several mundane, repetitive, or dangerous construction tasks autonomously or under the supervision of human workers, and perform effective site and resource monitoring to stimulate productivity growth and facilitate safety management. When using ARC technologies, three-dimensional (3D) reconstruction is a primary requirement for perceiving and modeling the environment to generate 3D workplace models for various applications. Previous work in ARC has predominantly utilized 3D data captured from high-fidelity and expensive laser scanners for data collection and processing while paying little attention of 3D reconstruction and modeling using low-precision vision sensors, particularly for indoor ARC applications. This dissertation explores 3D reconstruction and modeling for ARC applications using low-precision vision sensors for both outdoor and indoor applications. First, to handle occlusion for cluttered environments, a joint point cloud completion and surface relation inference framework using red-green-blue and depth (RGB-D) sensors (e.g., Microsoft® Kinect) is proposed to obtain complete 3D models and the surface relations. Then, to explore the integration of prior domain knowledge, a user-guided dimensional analysis method using RGB-D sensors is designed to interactively obtain dimensional information for indoor building environments. In order to allow deployed ARC systems to be aware of or monitor humans in the environment, a real-time human tracking method using a single RGB-D sensor is designed to track specific individuals under various illumination conditions in work environments. Finally, this research also investigates the utilization of aerially collected video images for modeling ongoing excavations and automated geotechnical hazards detection and monitoring. The efficacy of the researched methods has been evaluated and validated through several experiments. Specifically, the joint point cloud completion and surface relation inference method is demonstrated to be able to recover all surface connectivity relations, double the point cloud size by adding points of which more than 87% are correct, and thus create high-quality complete 3D models of the work environment. The user-guided dimensional analysis method can provide legitimate user guidance for obtaining dimensions of interest. The average relative errors for the example scenes are less than 7% while the absolute errors less than 36mm. The designed human worker tracking method can successfully track a specific individual in real-time with high detection accuracy. The excavation slope stability monitoring framework allows convenient data collection and efficient data processing for real-time job site monitoring. The designed geotechnical hazard detection and mapping methods enable automated identification of landslides using only aerial video images collected using drones.