Automated segmentation, detection and fitting of piping elements from terrestrial LIDAR data

Yun-Ting Su, Purdue University

Abstract

Since the invention of light detection and ranging (LIDAR) in the early 1960s, it has been adopted for use in numerous applications, from topographical mapping with airborne LIDAR platforms to surveying of urban sites with terrestrial LIDAR systems. Static terrestrial LIDAR has become an especially effective tool for surveying, in some cases replacing traditional techniques such as electronic total stations and GPS methods. Current state-of-the-art LIDAR scanners have very fine spatial resolution, generating precise 3D point cloud data with millimeter accuracy. Therefore, LIDAR data can provide 3D details of a scene with an unprecedented level of details. However, automated exploitation of LIDAR data is challenging, due to the non-uniform spatial sampling of the point clouds as well as to the massive volumes of data, which may range from a few million points to hundreds of millions of points depending on the size and complexity of the scene being scanned. This dissertation focuses on addressing these challenges to automatically exploit large LIDAR point clouds of piping systems in industrial sites, such as chemical plants, oil refineries, and steel mills. A complete processing chain is proposed in this work, using raw LIDAR point clouds as input and generating cylinder parameter estimates for pipe segments as the output, which could then be used to produce computer aided design (CAD) models of pipes. The processing chain consists of three stages: (1) segmentation of LIDAR point clouds, (2) detection and identification of piping elements, and (3) cylinder fitting and parameter estimation. The final output of the cylinder fitting stage gives the estimated orientation, position, and radius of each detected pipe element. A robust octree-based split and merge segmentation algorithm is proposed in this dissertation that can efficiently process LIDAR data. Following octree decomposition of the point cloud, graph theory analysis is used during the splitting process to separate points within each octant into components based on spatial connectivity. A series of connectivity criteria (proximity, orientation, and curvature) are developed for the merging process, which exploits contextual information to effectively merge cylindrical segments into complete pipes and planar segments into complete walls. Furthermore, by conducting surface fitting of segments and analyzing their principal curvatures, the proposed segmentation approach is capable of detecting and identifying the piping segments. A novel cylinder fitting technique is proposed to accurately estimate the cylinder parameters for each detected piping segment from the terrestrial LIDAR point cloud. Specifically, the orientation, radius, and position of each piping element must be robustly estimated in the presence of noise. An original formulation has been developed to estimate the cylinder axis orientation using gradient descent optimization of an angular distance cost function. The cost function is based on the concept that surface normals of points in a cylinder point cloud are perpendicular to the cylinder axis. The key contribution of this algorithm is its capability to accurately estimate the cylinder orientation in the presence of noise without requiring a good initial starting point. After estimation of the cylinder's axis orientation, the radius and position are then estimated in the 2D space formed from the projection of the 3D cylinder point cloud onto the plane perpendicular to the cylinder's axis. With these high quality approximations, a least squares estimation in 3D is made for the final cylinder parameters. Following cylinder fitting, the estimated parameters of each detected piping segment are used to generate a CAD model of the piping system. The algorithms and techniques in this dissertation form a complete processing chain that can automatically exploit large LIDAR point cloud of piping systems and generate CAD models.

Degree

Ph.D.

Advisors

Bethel, Purdue University.

Subject Area

Civil engineering|Remote sensing

COinS