CHAPTER 2 LITERATURE REVIEW

Transcription

1 CHAPTER 2 LITERATURE REVIEW Reverse engineering technology refers to a series of process that generates a CAD model of an existing part from measuring the part surface and processing the point data. This technology can be extremely useful when a part having complex surfaces needs to be duplicated. Reverse engineering typically starts with measuring an existing object so that a surface or solid model can be deduced in order to exploit the advantages of CAD/CAM technologies. There are several application areas of reverse engineering. It is often necessary to produce a copy of a part, when no original drawings or documentation are available. In other cases we may want to re-engineer an existing part, when analysis and modifications are required to construct a new improved product. In areas where aesthetic design is particularly important such as in the automobile industry, real-scale wood or clay models are needed because stylists often rely more on evaluating real 3D objects than on viewing projections of objects on high resolution 2D screens at reduced scale. Another important area of application is to generate custom fits to human surfaces, for mating parts such as helmets, space suits or prostheses. Reverse engineering is a rapidly evolving discipline, which covers a multitude of activities. The reverse engineering procedure consists of the following basic phases: 1. Data capture 2. Preprocessing (mainly for noise filtering) 3. Segmentation and data reduction 4. CAD model creation Given a CAD model of a part, traditional method such as CNC machining can be used to produce the parts. However, it is sometimes impossible for CNC machines to machine complicated surfaces both rapidly and cost-effectively. Over the past a few years, rapid prototyping machines have been widely used in industry. Hence while integrating Reverse Engineering and Rapid prototyping process for a rapid development of any product following are the major steps: 1. Data capture 2. Preprocessing (mainly for noise filtering) 3. Segmentation and data reduction 4. STL file generation 29

2 In capturing the surface data of a part, either a contact type measuring device or a non-contact type device has been used. The contact type devices show better accuracy but these are generally very slow in point data acquisition. Recent advancement in laser scanning technology provided for a non-contact type measuring device that can scan parts with very high speed and good accuracy. The se machines enabled us to capture the surface data of a part with complex and freeform surfaces. Now, in return, another problem has been created due to the enormous amount of point data that is obtained in the data acquisition step. This data too requires huge storage space and increase computational time significantly. Segmentation has been an essential part in the process of surface modeling from scanned point data. It is the process of partitioning a point cloud into meaningful regions or extracting important features from the point data. The majority of point data segmentation methods can be classified into three categories: edge-detection methods, region-growing methods and hybrid methods. Tamas Varady, Ralph R Martin and Jordon Cox [1] while contributing to the advancement in the state of the art in Reverse Engineering studied pros and cons of various data acquisition techniques with related problems of boundary representation model construction. Specific issues addressed by them include characterization of geometric models and related surface representations, segmentation and surface fitting for simple and free form shapes, multiple view combination and creating consistent and accurate B-rep models. On the key areas of segmentation they emphasized that there are two basic different approaches to segmentation that may be considered, namely edge-based and face-based methods. The first works by trying to find boundaries in the point data representing edges between surfaces. If sharp edges are being sought, one must try to find places where surface normals estimated from the point data change direction suddenly, while if smooth (tangent-continuous) edges are also possible, it is required to look for places where surface curvatures or other higher derivatives have discontinuity. This technique thus basically attempts to find edge curves in the data, and infers the surfaces from the implicit segmentation provided by the edge curves. The second technique goes in the opposite order, and tries to infer connected regions of points with similar properties as belonging to the same surface (e.g. groups of points all having the same normal belong to the same plane), with edges then being derived by intersection or other computations from the surfaces. They highlighted that edge-based techniques suffer from the following problems. In this technique the number of points used for 30

3 segmenting the data is small, i.e. only points in the vicinity of the edges are used. Finding smooth edges, which are tangent continuous, or have even higher continuity, is very unreliable, as computation of derivatives from noisy point data is error prone. On the other hand, if smoothing is applied to the data first to reduce errors, this distorts the estimates of the required derivatives. Thus sharp edges are replaced by blends of small radius, which may complicate the edge-finding process; also the positions of features may be moved by noise filtering. On the other hand face-based techniques have the following advantages they observed. They work on a larger number of points, in principle using all available data. Deciding which points belong to which surface is a natural by-product of such methods, whereas with edge-based methods, it may not be entirely clear to which surface a given point belongs even after finding a set of edges. Typically, face based segmentation also provides the best-fit surface to the points as a by-product. Overall, the professionals believe that face-based rather than edge-based segmentation is preferable. They concluded that current commercial software systems often only allow simple point cloud processing and single surface fitting with interactive help; the production of complete B-rep models is only possible for very simple objects or polyhedral approximations. They further observed that users wish to automatically process a wide range of objects, possibly from a variety of data capture devices with differing characteristics, to produce models in a variety of representations and accuracies and relate to systems exist which can perform the simple operation of 3D copying, the goals of extracting higher level information which can be edited and analyzed are still some way of and relate to key research areas, which still need further work before general-purpose reverse engineering becomes widely available include: Improving data capture and calibration, coping with noise, merging views, coping with gaps in the data, reliable segmentation, fair surface fitting, recognizing natural or human-intended structure of the geometry of the object, and finally ensuring that consistent models are built. Fan et al. [2] used local surface curvature properties to identify significant boundaries in the data range. In order to avoid the loss of localization, scale-space tracking, which convolves the entire data with Gaussian masks having different values of the spread parameter, was employed. Y. H. Chen and C. Y. Liu [3] present a method for compound freeform surface segmentation and reconstruction. In their proposed method, a cloud of measurement data are collected through a coordinate measuring machine (CMM). The set of measurement data is then sliced along, at most, three orthogonal directions. On each slicing plane, measurement data is fitted by a 2D NURBS spline. Now, maximum 31

4 curvature points on each NURBS spline are obtained. These points represent the boundary of the digitised object. Their method thus simplifies, three-dimensional segmentation is simplified to a twodimensional problem. Highlighting the limitations they pointed out that at current implementation, slicing planes can be applied only to two directions: slicing planes that are parallel to the x-z-plane; and slicing planes that are parallel to the y-z-plane. This is mainly because of the fact that measurement points projected on the x-y-plane can be arranged in a matrix form. If the interval between consecutive slicing planes is set at the measurement increment, it is guaranteed to have all the measurement points on slicing planes. In the z-direction, however, it is difficult to define slicing planes because every measurement point may have a different z- value. It is, therefore, not practical to arrange the measurement points in equal z-intervals. Segmentation in the past has been applied to single range maps, which are capable of modelling only the part of an object that is visible from the camera viewpoint. In the field of reverse engineering, there is a great need for the segmentation of complete wrap-around object models. A primary requirement for the acquisition and merging of multiple range maps is accurate registration of the range data. Since repositioning the camera for the various views tends to introduce errors into the data, a correction is usually applied to the transformation that relates the viewing frame of reference with the global frame. The correction may be determined using one of the following approaches: 1. A set of reference spheres is included with the objects. At least one sphere is always visible, and the data are shifted so that the sphere centres appear to coincide in each of the views, or 2. Matching planar regions in overlapping range maps are identified, and the range data are shifted to reduce the registration error These methods yield a wrap-around model consisting of data points from registered range maps, but the shortcomings of the model are immediately evident: a. The map contains regions of overlapping data, b. The mosaic of range data points that forms the global model does not have the ordered row-and-column structure of a single range map, and c. The model consists of unconnected data points-the connectivity of points in adjacent views is not readily established. On the segmentation feature of a wrap around model using on a active contour M.J.Milroy [4] suggested a new technique by interpolating 32

5 and merging multiple range maps which requires accurate registration of the range data. Milroy et al. used a semi-automatic edge-based approach for orthogonal cross-section (OCS) models. Surface differential properties were estimated at each point in the model, and the curvature extremes were identified as possible edge points. Then an energy-minimizing active contour was used to link the edge points. The technique suggested by them is applied to telephone handset and a water timer. Yang and Lee [5] identified edge points as the curvature extremes by estimating the surface curvature. After detecting edge points, an edgeneighborhood chain-coding algorithm was used for the formation of boundary curves. The original point set was then divided into subsets by a scan line algorithm. The region-growing methods, on the other hand, proceed with segmentation by detecting continuous surfaces that have homogeneity or similar geometrical properties. Hoffman and Jain [6] segmented the range image into many surface patches and classified these patches as planar, convex or concave shapes based on a non-parametric statistical test for trend, curvature values and eigen value analysis. Computer vision systems attempt to recover useful information about the three-dimensional world from huge image arrays of sensed values. Since direct interpretation of large amounts of raw data by computer is difficult, it is often convenient to partition (segment) image arrays into low-level entities (groups of pixels with similar properties) that can be compared to higher-level entities derived from representations of world knowledge. Segmentation through variable - order surface fitting suggested by Paul [7] and others through light on solving the segmentation problem through a mechanism for partitioning the image array into low-level entities based on a model of the underlying image structure. Using a piecewise-smooth surface model for image data that possesses surface coherence properties, the authors have developed an algorithm that simultaneously segments a large class of images into regions of arbitrary shape and approximates image data with bivariate functions so that it is possible to compute a complete, noiseless image reconstruction based on the extracted functions and regions. Surface curvature sign labeling provides initial coarse image segmentation based on mean and Gaussian Surface Curvature, which is refined by an iterative region growing method based on variable-order surface fitting. Experimental results show the algorithm's performance on six range images and three intensity images. A large number of published works are devoted to the range image segmentation problem. Milgrim and Bjorklund [8] proposed to adjust a planar surface for each point according to a window of size 5 X 5 by using least square 33

6 minimization. Based on the normal vectors and the fitting-error of planar surfaces, they connect the points which satisfy a "constraint of planarity", to build the regions of the final segmentation. Henderson [9] developed an algorithm also based on "planarity test" that creates convex surfaces starting from initial point-sets, each one containing three non-collinear points. Sethi et al. [10] consider spheres and ellipsoids in addition to planar surfaces, cylinders and cones. Sapidis et al. [11] proposed an algorithm based on the approximation of regions with polynomial surfaces of degree not exceeding four. An initial partition is computed by an elegant method of point classification according to eight fundamental forms obtained by using the sign of the Gaussian and mean curvatures. Seed regions are then computed by erosion of connected regions whose points belong to the same fundamental form. Seed regions are then grown iteratively until the final partition is achieved. The advantage of this method is the efficient use of differential-geometric quantities and the classification of points using fundamental forms. The weakness is the considerable calculation-time necessary to grow the seeds. In the iterative process, each seed region requires several evaluations of least square polynomials and of the corresponding error to determine the adequate degree of the polynomial surface approximating it, and this takes a significant time. Also, controlling the overlapping of the grown regions seems a difficult task. An extension of this algorithm is developed, where the shape of the output regions is constrained to be rectangular, thus leading to improved B-spline surface fits. Relatively recent algorithms have been developed by Leonardis et al. [12], [13], where data aggregation is performed via model recovery in terms of variable-order bivariate polynomials using iterative regression. Model recovery is initiated independently at regularly placed seed regions in the image. All recovered models are candidate to be used as the "segmentation-defining model". Selection of the most appropriate model is treated as a quadratic boolean problem. Taylor et al. [14] use an approach of divide and merge, where the criteria of homogeneity are based on the angles which indicate the directions of the normals to the surface and the depth value at a point. A novel method for range image segmentation based on triangulation and region growing edges and critical points suggested by Dongming Zhao [15] and others is based on an integration of edge and region information. The algorithm consisted of three steps: edge and critical point detection, triangulation, and region growing. The edge detection method studied by them is variant on morphological residues and implemented as a top - hat transformation. Edge curves are closed ones. Critical points are detected on two-dimensional (2-D) curves, and each critical point corresponds to a 3-D coordinate on a range surface. By 34

7 connecting the points in pairs using straight lines based on a set of rules, a number of approximate facets of range surfaces are obtained. These facets are the surface patches with similar geometric characteristics. Experiment reveal that the method is efficient for segmentation of range images and contain polyhedral objects. This segmentation method uses edge information of a range image to constitute a 3-D SSG, which is an approximation of a range surface. Through computing normals of the approximated planes, region grouping and final segmentation are completed. This method is not computationally complicated. The method is not sensitive to noise, compared to curvature computation methods that directly compute local curvature on range image surfaces, because the normals for region growing are based on larger triangles. The SSG is a description of the surface structure of an object and also presents a data set that can be used to establish a surface model for CAD-based vision and object recognition. Most researchers have tried to develop segmentation methods by exactly fitting curves or surfaces to find edge points or curves. These surface- or curve-fitting tasks take a long time and, furthermore, it is difficult to extract the exact edge points because the scan data are made up of discrete points and edge points are not always included in the scan data. Woo et al [16] in their method of segmentation for point cloud data proposed the process of generating a surface model from point cloud data, a segmentation that extracts the edges and partitions the threedimensional (3D) point data is necessary and plays an important role in fitting surface patches and applying the scan data to the manufacturing process. In their research, a new method for segmenting the point cloud data is proposed. They proposed an algorithm, which uses the octree-based 3Dgrid method to handle a large amount of unordered sets of point data. The final 3D-grids are constructed through a refinement process and iterative subdivisioning of cells using the normal values of points. Their 3D-grid method enables us to extract edge-neighborhood points while considering the geometric shape of a part. They applied this method to two quadric models and the results are discussed. A new and efficient algorithm for the decomposition of 3D arbitrary triangle meshes and particularly optimized triangulated CAD meshes. The algorithm is based on the curvature tensor field analysis and presents two distinct complementary steps: a region based segmentation, which is an improvement of that presented by Lavoue et al. [17]. Constant curvature region decomposition of 3D-meshes by a mixed approach vertex-triangle, 35

8 and which decomposes the object into near constant curvature patches, and a boundary rectification based on curvature tensor directions, which corrects boundaries by suppressing their discontinuities. Experiments conducted on various models including both CAD and natural objects, show satisfactory results. Resulting segmented patches, by virtue of their properties (homogeneous curvature, clean boundaries) are particularly adapted to computer graphics tasks like parametric or subdivision surface fitting in an adaptive compression objective. Garland et al. [18] present a face clustering of which aim is to approximate an object with planar elements; this algorithm is especially adapted for radiosity or simplification. Benko and Varady [19] consider a mesh decomposition approach specifically adapted for reverse engineering by applying a hierarchy of tests to recognize conventional engineering objects (extrusions, surfaces of revolution,). Several approaches use discrete curvature analysis combined with the Watershed algorithm described by Serra [20] in the 2D image segmentation field. In visualizing or analyzing a scanned object, the meshed models, especially using triangular patches, have been used. In the case of triangulated models, data reduction is performed by reducing the number of triangles based on application-dependent criteria; then, with the remaining nodes, triangulation is performed repeatedly. Data reduction methods based on meshed models can be divided into two categories: one for manipulating the triangulated point data and the other for using levelof-detail (LOD) methods. Data reduction, therefore, has become an important issue in reverse engineering. Many data reduction methods were proposed in the area of image processing, but they were mostly designed for dealing with the meshed point data. Only a few methods were developed that could be directly applied to the point data generated from measurement devices but they were also limited to specific types, such as 2D or planar sets of point data. Hamann [21] proposes a data reduction scheme for the triangulated surface. He removes a triangle based on the curvatures at the three vertices. A user can specify a percentage of triangles to be removed, or in the case of bivariate functions, an error tolerance. His research has smooth surface fitting in mind. Gieng et al. [22] classify mesh simplification algorithms into three types: removing vertices, removing edges and removing faces. After having reviewed relative works, they propose an algorithm to produce a hierarchy of triangle meshes that can be used to blend different levels of detail in smooth fashion through a set of triangle-collapse operations. They assign a weight to each triangle of the original mesh, which depends on the absolute curvature, a shape measure, a topological measure, an error 36

9 measure and the triangular area. The later four factors are weighted by users. The triangles with the smallest weight are collapsed first. The criterion used when collapsing triangles is a certain percentage of triangles to be collapsed in a single step. Fischer and Park [23] generated multilevel-of-detail models for design and manufacturing in reverse engineering. The multilevel 3D meshes are constructed using a 2D structure: the quadtree structure and the transition between the levels can be achieved in real-time. The multilevel structure preserves the geometric and topological behavior at each level of details for a given tolerance. Through this method, the reconstructed geometric model is represented by hierarchical levels of details [24]. For reducing the laser scanned data, conventional sampling methods such as uniform sampling, chordal deviation sampling and space sampling have been widely used due to their simplicity and fast computation time. But, these sampling methods are order-sensitive [25]. The data reduction method proposed in this paper uses a grid method and related research is described below. Martin, Stroud and Marshall [26] proposed a data reduction method using uniform grid s in their EU Copernicus project. Their method uses uniform sized grids with a 'median filtering' approach. The procedure starts with choosing a grid structure, and input data points are assigned to the corresponding grid. For the points assigned to a given grid, a point in the median location is selected to represent data points belonging to that grid. This approach intends to overcome the problem of z-axis inaccuracy resulting from laser scanning. However, it only uses uniform grid s without any consideration of part shape, vc Lee, Woo, and Suk [27] proposed improved grid methods, which effectively reduce point data utilizing part geometry information. The onedirection non-uniform grid method extracts significant curvature changing points and non-uniform grids are generated along a direction based on these points. If a grid is too long, it is divided by the user-defined maximum grid length. After generating non- uniform grids, a representative point is selected within each grid based on median filtering. The bi-directional non-uniform grid method uses point normal values as part geometry information. A planar set of point data is gridded as the same manner in the uniform grid method and each grid is sub-divided based on point normal values. During subdivision of grids, the quadtree is used; that is, a grid is divided into four sub-grids, if necessary. With the non-uniform grids generated after sub-division, a representative point is extracted from each 37

10 grid. These grid methods, however, only work for two dimensional laser scanned data and they require merging of point data after data reduction if a complete 3Dpoint data model is needed. Reverse engineering is a methodology for constructing computeraided design (CAD) models of physical parts by digitizing an existing part, creating a computer model and then using it to manufacture the component. When a digitized part is to be manufactured by means of rapid prototyping machines, it is not necessary to construct the CAD model of a digitized part. This can be achieved by constructing an STL file after preprocessing, segmentation and data reduction. These STL files which are produced by 3D modelling systems contain triangular facet representation of surface and have become standard data inputs of rapid prototyping and manufacturing systems. In rapid prototyping technology, physical objects are produced layer by layer, each layer a 2D cross-section of the 3D mesh in STL format. Chen and Wang [28] proposed a genetic algorithm for optimized retriangulation. They showed an optimized STL file generation method for reducing laser scanned data. Data reduction is performed by decreasing the number of triangles in an STL file using normal vectors of triangles. After removing triangles, re-triangulation is performed as in other reduction methods [21, 22, 29, 30, 31, 32]. Hoppe et al. [33] use an energy function to represent the trade-off between geometric fit and compact representation. A user desired parameter is used to control the trade-off between geometric fit and compact representation. A large value indicates that a sparse representation is strongly preferred, but at the expense of degrading the fit. A merging algorithm based on edge collapse is proposed for automatically computing approximations of a given triangulated object at different levels of detail [34]. Edges are queued according to their cost functions, which indicate the error caused by edge collapse. Approximation levels are controlled by prescribing geometric tolerances. The fact that a given set of data points have many possible results of triangulation suggests that we might look for a best triangulation. To get best triangulation for different applications, different triangulation methods have been developed under different optimality criteria [35-38]. Of these, Delaunay triangulation is the most widely used. Numerous algorithms for constructing Delaunay triangulation from a set of data points in 2D and 3D spaces have been studied, developed and used in many areas. But slivers may appear in 3D Delaunay triangulation [39]. This problem may not be acceptable for some applications. 38

11 As a result of their global optimization property [40,41], Genetic Algorithms (GAs) have been widely used in various fields such as state space search, nonlinear optimization, machine learning, travelling salesman problems, etc. [42]. In recent years, preliminary study on GAs in triangulation has been reported. Absaloms and Tomikawa propose a GA to triangulate two adjacent contour data from a digitized geographical map [43]. They claim that GA based triangulation is relatively technique independent and can be implemented by parallel processing. Quadrilateral patch [44] is used to reconstruct free-form surface for dividing curvature deviation region and boundary detection. Piegl and Tiller [45] introduce an algorithm to impose the compatibility of crosssectional curve during NURBS skinning and eliminates the problem of too many control points while specifying tolerances in approximation. Werner et al. [46] presents the complete procedure of the free-form surface in reverse engineering. An integrated solution for measuring and geometric modeling of free-form surfaces is introduced, and the machining errors in CNC milling machine was compensated by the comparison of the error between the machined surface and the theoretically defined surface. Park and Kim [47] interpolate a rough surface from scattered 3D point data and refine it repeatedly to get the desired approximation accuracy. A variety of rapid prototyping technologies have emerged [48-50]. They include stereolithography (SLA), selective laser sintering (SLS), fused deposition manufacturing (FDM), laminated object manufacturing (LOM), ballistic particle manufacturing (BPM), and three-dimensional printing (3D Printing). These technologies are capable of directly generating physical objects from a CAD model. They have an important common feature: physical parts are produced by adding materials layer by layer. This is in contrast to traditional machining methods that make physical parts by removing the material. In rapid prototyping, STL file format has become the de facto standard [51]. An STL file, which is composed of triangular facets and their associated unit normals, can be generated from either a solid model or a set of digitized data. Most recent commercial CAD /CAM software systems are capable of generating STL files directly from a solid model. The majority of triangulation methods are based on a known surface model [52, 53]. Triangulation of scattered data in 3D space often has an objective of constructing smooth surfaces [47, 54]. Literature about optimized STL file generation from digitized data cannot be found. When digitizing a part, in order not to miss any detail of its geometry, a large number of measurement points are normally collected. If all points are used in STL file generation, the file can easily become huge enough so that the whole rapid prototyping process will be slowed down significantly. In fact, keeping lots of data points in planar or nearly planar region is rather unsophisticated. Removal of data in these regions would 39

12 not affect the accuracy for rapid prototyping machines when the deviation between the surface model and point cloud data is less than the layer thickness of the RP machine. Hence the validation of STL file is also essential before submitting it to the RP machine. 40