PITA Fiscal Year 2011 Projects

Tools for Monitoring and Analyzing Structural Steel, Pipes, and Other Thin Objects in Process Plants, Construction Sites and Infrastructure

Lead University: Carnegie Mellon University
PI: Daniel Huber, Robotics Institute
Co-PI: Burcu Akinci, Civil and Environmental Engineering
PA Industry: Quantapoint, Inc. Three dimensional (3D) sensors, such as laser scanners, are often used to capture the raw, as-is conditions of a facility. Models derived from the as-is conditions support a variety of purposes, including quality assurance of construction progress and preparation for rapid replacement of components in process plants. Our goal is to develop new tools and techniques to facilitate the semi-automated modeling and analysis of long, thin structures, such as steel beams, columns, and pipes, which are the predominant structures within process plants and structural steel construction projects. We will focus on two aspects of this problem: 1) Understanding the effect of mixed pixel artifacts on thin objects; and 2) Developing algorithms to model and analyze the structure and connectivity of thin objects in process plants.

Thin objects present a special challenge for laser scanners, primarily due to mixed pixel artifacts. Although mixed pixels can occur at all depth boundaries, the effect can dominate in the case of thin objects. To address this, we will develop a novel, learning-based mixed pixel detection algorithm, with a focus on the effects on long, thin objects. Our previous research has shown that state-of-the-art mixed pixel detection algorithms are not very accurate and that the mixed pixel effect can cause dramatic modeling errors for thin objects.

Existing tools and algorithms for modeling and analyzing thin structures are limited and require significant user interaction. We will address the lack of relevant tools by using techniques from the computer vision community to develop robust tools for semi-automated modeling and analysis of thin objects within process plants. We will first establish surface connectivity among points in the data, and then use a proven eigen-analysis technique to identify long, thin objects. Finally, we will conduct automated analysis on the resulting objects, such as checking the plumbness of a column.