Tomohiro Mizoguchi

Segmentation of scanned mesh into analytic surfaces based on robust curvature estimation and region growing

For effective application of laser or X-ray CT scanned mesh models in design, analysis, and inspection etc, it is preferable that they are segmented into desirable regions as a pre-processing. Engineering parts are commonly covered with analytic surfaces, such as planes, cylinders, spheres, cones, and tori. Therefore, the portions of the parts boundary where each can be represented by a type of analytic surface have to be extracted as regions from the mesh model. In this paper, we propose a new mesh segmentation method for this purpose. We use the mesh curvature estimation with sharp edge recognition, and the non-iterative region growing to extract the regions. The proposed mesh curvature estimation is robust for measurement noise. Moreover, our proposed region growing enables to find more accurate boundaries of underlying surfaces, and to classify extracted analytic surfaces into higher-level classes of surfaces: fillet surface, linear extrusion surface and surface of revolution than those in the existing methods.

Quasi-Optimal Mesh Segmentation Via Region Growing/Merging

Recently meshes of engineering objects are easily acquired by 3D laser or high energy X-ray CT scanning systems, and these meshes are widely used in product developments. To effectively use scanned meshes in engineering applications, such as inspection, CAD model reconstruction, and convergent-type CAE, we need to segment meshes and extract desirable regions and their approximating surfaces as preprocessing. Engineering objects are commonly represented as a set of analytic surfaces, such as planes, cylinders, spheres, cones, and tori. Therefore, the mesh surface of engineering objects needs to be approximated as a set of analytic surfaces. Moreover, a mesh surface should be approximated with a minimum number of analytic surfaces and their approximating error should be minimized as a result of segmentation. We call the segmentation that satisfies these two conditions the optimal segmentation as proposed in [1]. However, optimal segmentation algorithms need a long calculation time. Today’s high energy X-ray CT scanning systems generate large meshes with millions of triangles from objects including hundreds of regions. Thus, computationally expensive algorithms, such as [1], cannot be directly applied to these large and complex meshes from the aspect of efficiency. In this paper we propose an efficient new quasi-optimal mesh segmentation algorithm via region growing and region merging. First, our algorithm robustly and accurately estimates mesh principal curvatures using the local surface fitting by two-pass algorithm. Second, it uses the curvatures to appropriately create seed regions, and then it quickly grows each seed region and extracts grown regions and their approximating analytic surfaces from a whole mesh. Finally, our region merging algorithm efficiently merges extracted regions in order to minimize the number of regions while keeping the user specified tolerances of the surface fitting, and it results in quasi-optimal segmentation. We demonstrate the performance of our algorithm with scanned meshes acquired from real engineering objects by 3D laser and X-ray CT scanning systems.Copyright

Manhattan-World Assumption for As-Built Modeling Industrial Plant

As industrial plants such as chemical and power plants continue to age, their CAD models are increasingly required for model-based planning and simulation. However, in the case of old plants, the original CAD models rarely exist, and hand-drawings do not precisely match the present states of the plants due to repeated remodeling. It is therefore becoming a common approach to reconstruct CAD models from the point cloud of such plants captured by terrestrial laser scanning and use these models for the above purposes. Such a reconstruction process is usually called “as-built modeling”. However, existing methods for as-built modeling come with such problems as the need for many human operations and computational cost. In this paper, we propose an automatic and efficient method for as-built modeling industrial plants using Manhattan-world assumption which states that there exist three dominant axes orthogonal to each other in artificial buildings and the internal parts are arranged so that they are parallel or orthogonal to one of them. In the case of industrial plants, it is reasonable to consider that long pipes and shaped steels are arranged so that they follow this assumption. In addition, plant parts are supposed to be designed as long linear sweep surfaces on CAD system or hand drawings. Our method can automatically recognize such sweep parts and their cross sectional shape which follow the assumption, as well as efficiently recognize them even from a large point cloud which may contain as many as one hundred million points in a few minutes. We demonstrate the effectiveness of our proposed method from various experiments on real scanned data.

Lidar-based individual tree species classification using convolutional neural network

Terrestrial lidar is commonly used for detailed documentation in the field of forest inventory investigation. Recent improvements of point cloud processing techniques enabled efficient and precise computation of an individual tree shape parameters, such as breast-height diameter, height, and volume. However, tree species are manually specified by skilled workers to date. Previous works for automatic tree species classification mainly focused on aerial or satellite images, and few works have been reported for classification techniques using ground-based sensor data. Several candidate sensors can be considered for classification, such as RGB or multi/hyper spectral cameras. Above all candidates, we use terrestrial lidar because it can obtain high resolution point cloud in the dark forest. We selected bark texture for the classification criteria, since they clearly represent unique characteristics of each tree and do not change their appearance under seasonable variation and aged deterioration. In this paper, we propose a new method for automatic individual tree species classification based on terrestrial lidar using Convolutional Neural Network (CNN). The key component is the creation step of a depth image which well describe the characteristics of each species from a point cloud. We focus on Japanese cedar and cypress which cover the large part of domestic forest. Our experimental results demonstrate the effectiveness of our proposed method.

Decomposing scanned assembly meshes based on periodicity recognition and its application to kinematic simulation modeling

Along with the recent growth of industrial X-ray computerized tomography (CT) scanning systems, it is now possible to non-destructively acquire the entire meshes of assemblies. This technology has the potential to realize an advanced inspection process of an assembly, such as estimation of their assembly errors or examinations of their dynamic behaviors in motion using a model reflecting real assembled situations. However, to realize the process, it is necessary to accurately decompose the mesh and to extract a set of partial meshes, each of which corresponds to a single part, from the entire meshes of assemblies measured from the CT scans. Moreover, it is required to create models that are ready for dynamic behavior simulations. In this paper, we focus on CT scanned meshes of gear assemblies as examples, and propose beneficial methods for establishing such advanced inspections. We first propose a method that accurately decomposes the mesh into partial meshes, each of which corresponds to a single gear, using periodicity recognitions. The key idea is first to accurately recognize the periodicity of each gear, then to extract sets of topologically connected mesh elements where periodicities are valid, and finally to interpolate points in plausible ways from an engineering viewpoint to the area where surface meshes are not generated, especially the contact area between parts in the CT scanning process. We also propose a method for creating kinematic simulation models which can be used for a gear teeth contact evaluation using extracted partial meshes and their periodicities. Such an evaluation of teeth contacts is one of the most important functions in kinematic simulations of gear assemblies for predicting the power transmission efficiency, noise and vibration. The characteristics of the proposed method is that (1) it can robustly and accurately recognize periodicities from noisy scanned meshes, (2) it can estimate the plausible boundaries of neighboring parts without any previous knowledge from single-material CT scanned meshes, and (3) it can efficiently extract partial meshes from large scanned meshes containing millions of triangles in a few minutes. We demonstrate the effectiveness of our method on a variety of artificial and real CT scanned meshes.

Quantitative damage assessment of concrete structures based on 3D laser scanning

Recently scaling damages of concrete structures have become apparent and it is now investigated to develop a method for quantitatively assessing the damages. In this paper, we propose a beneficial method for this purpose based on long-distance 3D laser scanning. In our proposed method, original shapes of the concrete structures before to be damaged are estimated on their laser scanned data by region growing, the distances to the estimated shapes are computed at each scanned point, and then their surface concavity and convexity are quantitatively evaluated based on the computed distance. We demonstrate the effectiveness of our proposed method from various experiments on real scanned data of concrete structures with scaling damages.

Euclidean Symmetry Detection From Scanned Meshes Based on a Combination of ICP and Region Growing Algorithms

Recently, meshes of engineering objects have been easily acquired by 3D laser or high-energy industrial X-ray CT scanning systems and they are widely used in product developments. For the effective use of scanned meshes in inspection, re-design, and simulation of the objects, it is important to reconstruct CAD models from the meshes. Engineering objects often exhibit Euclidean symmetries for their functionalities. Therefore, it is essential to detect such symmetries when reconstructing CAD models with compact data representations which are similar to the ones already defined in CAD systems. However, existing methods for reconstructing CAD models have not focused on detecting such symmetries. In this paper, we propose a new method that detects partial or global Euclidean symmetries, including translation, rotation, and reflection, from scanned meshes of engineering objects based on the combination of the ICP and the region growing algorithms. Our method can robustly and efficiently extract pairs of symmetric regions and their transformations under which the pair can be closely matched to each other. We demonstrate the effectiveness of the proposed method from experiments on various scanned meshes.Copyright

Parts identification and motion estimation on CT scanned assembly meshes

Along with the recent improvements of the industrial X-ray CT scanning systems, it is now possible to non-destructively acquire the entire meshes of mechanical assemblies. This technology has the potential to realize an advanced inspection of assemblies, such as examining assembling errors or dynamic behaviors in motion using the meshes reflecting really-assembled situations. However, to realize such advance inspections, it is required to identify each part and to estimate their motions in the meshes.