reverse engineering of automotive parts applying laser scanning

Reverse Engineering of Automotive Parts Applying Laser Scanning and Structured

Light Techniques

Project in Lieu of Thesis presented for the

Masters of Science Degree

The University of Tennessee, Knoxville

Ngozi Sherry Ali May 2005

ACKNOWLEDGEMENTS

I would like to first give thanks to Dr. M. Abidi for allowing me the opportunity to join the IRIS lab as a graduate research assistant. His support and patients has allowed me to complete one of my many life goals. Thanks for giving me the chance to earn a degree of higher education and to reach heights that I knew I could reach. Saying “thank you” will never be enough to repay you for all you have done. Again, I thank you for everything and I am very appreciative. I would like to thank Dr. Koschan for advising me on all my work and projects. I value the guidance that was giving to me. Thanks for being the best advisor a student could ask for. I would also like to thank Dr. Page for always being there when I needed. There was never a time when I was in need that you did not offer you assistance. The IRIS Lab has been my family since I first arrived in August of 2003, I thank all that is apart of the Lab. In closing, I would like to thank Vicki Courtney-Smith and Kim Cate for always making sure I had all the little things that I couldn’t get myself. Special thanks go to Rangan and Chris Kammerud for helping me with my research and class work. Thanks to my close friends Sharon Sparks and Syreeta Dickerson for helping me through my first year of school with the support and encouragement that they gave me. Lastly, thanks to my family for always supporting me and my endeavors. You all are the driving force in my life and career. Without your love, none of this would matter.

i

ABSTRACT

The automotive industry has an increasing need for the remanufacturing of spare parts through reverse engineering. In this project we will review the techniques of laser scanning and structured lighting for the reverse engineering of small automotive parts. Laser Range Scanning is the use of a CCD camera that captures the profile of the laser as it passes on an object. Structured light is the projection of a light pattern on an object also with the use of several cameras to obtain a profile of the object. The objective of the project is to be able to generate part-to-CAD and CAD-to-part reconstruction of the original for future usage. These newly created 3D models will be added to the IRIS 3D Part Database.

ii

TABLE OF CONTENTS

1 INTRODUCTION..................................................................................................... 1

2 ACQUISITION CLASSIFICATIONS AND RELATED WORKS ..................... 5

2.1 Contact Data Acquisition Techniques ................................................................ 6

2.2 Reverse Engineering Applying Contact Techniques .......................................... 7

2.3 Non-Contact Data Acquisition Techniques ........................................................ 8

2.4 Reverse Engineering Applying Non-Contact and Hybrid Techniques............. 10

2.5 General Constraints of Data Acquisition Techniques....................................... 12

3 SURFACE RECONSTRUCTION ........................................................................ 15

3.1 Single Views and Data Segmentation............................................................... 15

3.2 Multiple View Integration and Registration ..................................................... 17

3.3 Post-processing Registered Images................................................................... 19

4. SYSTEM DESCRIPTIONS AND SETUP ........................................................... 21

4.1 IVP Range Scanning Profiling System............................................................. 21

4.2 Genex 3D FaceCam Profiling System .............................................................. 25

5 RESULTS AND DISCUSSIONS........................................................................... 28

5.1 Part Identification.............................................................................................. 28

5.1 IVP Laser Range Data Results.......................................................................... 31

5.2 Genex Structured Light Data Results ............................................................... 36

5.3 Comparison: IVP Laser Range Scanner and Genex 3D FaceCam Data.......... 45

6 CONCLUSION ....................................................................................................... 49

REFERENCES................................................................................................................ 51

iii

LIST OF FIGURES

Figure 1.1: Flowchart for basic transformation phases of reverse engineering................. 2 Figure 1.2: The sequence of steps required for the reconstruction of a model from

multiple overlapping scans. ........................................................................................ 4 Figure 2.1: Classification of data acquisition techniques used in contact and non-contact

approaches for reverse engineering systems............................................................... 5 Figure 3.1: (a) Original image of the water pump, (b) Sequenced single view range

images of the bottom surface of the water pump generated using our laser range scanner ...................................................................................................................... 16

Figure 3.2: Textured single view range images generated from the structured lighting system, (a) left view of the ramp, (b) front view of the ramp, (c) back/left view of the ramp, (d) front/ right view of the ramp. .............................................................. 17

Figure 3.3: Point cloud information of the side view of the water pump, (a) and (b) side view range images, (c) registered range image of the two side views...................... 19

Figure 4.1: Principle of a laser triangulation system. Arrangement for the IVP Smart Vision Camera and Laser.......................................................................................... 22

Figure 4.2: Equipment setup for IVP Range Scanning System, (a) front view of the IVP Range Scanner, (b) side angled view of the top of the IVP Range Scanner............. 23

Figure 4.3: Calibration grid for the IVP Range Scanner ................................................. 24 Figure 4.4: IVP Range Scanner User Interface................................................................ 24 Figure 4.5: Structured light grid pattern projected on the ramp with a neutral tan

background................................................................................................................ 25 Figure 4.6: Genex 3D FaceCam System.......................................................................... 26 Figure 4.7: Distance specification for data collection ..................................................... 27 Figure 4.8: Genex 3D FaceCam User Interface screen that displays the left, center and

right camera photos of the object.............................................................................. 27 Figure 5.1: Photos of the ramp part (a) front and side views, (b) back and top views, (c)

ramp measurements .................................................................................................. 29 Figure 5.2: Photos of the water pump part (a) top and side views (b) bottom view........ 30 Figure 5.3: Photos of the pulley arm, (a) is the top view (b) is the bottom view ............ 30 Figure 5.4: CAD model images of the ramp (a) left/front view, (b) right side view, (c)

bottom view, (d) back view ...................................................................................... 31 Figure 5.5: 3D Ramp CAD models, (a) left/front side view, (b) back/left side view, (c)

bottom view, (d) right/top/ front side view............................................................... 32 Figure 5.6: Reconstructed bottom surface of the water pump, (a) Point cloud

information, (b) Solid mesh model, (c) smooth reconstructed mesh model............. 33 Figure 5.7: Reconstructed views of the water pump, (a) reconstructed top view (b) side

view reconstruction................................................................................................... 34 Figure 5.8: Cleaned point cloud data of the side views of the water pump taken with the

IVP Range Scanner, (a) right view, (b) left view, (c) right view and (d) left view .. 35

iv

Figure 5.9: Point cloud model of the water pump (a) CAD model showing height variations (b) top view, (c) back view ...................................................................... 35

Figure 5.10: Pulley Arm (a) photo of the side view profile of the pulley arm, (b) point cloud CAD model of the bottom side, (b) point cloud CAD model of the top side . 36

Figure 5.11: Textured ramp images using the Genex System, (a) back view, (b) right/ back side view, (c) left view and (d) front view ....................................................... 37

Figure 5.12: Genex System reconstructed views of the water pump, (a) right view, (b) back view, (c) front view and (d) left view............................................................... 38

Figure 5.13: Complete final 3D model of the ramp (a) front view (b) back view (c) side view (d) bottom view................................................................................................ 39

Figure 5.14: Water pump placed in box for neutral background..................................... 39 Figure 5.15: Textured water pump images using the Genex System, (a) top view, (b)

view of the bottom surface, (c) side view and (d) top view...................................... 40 Figure 5.16: Genex System reconstructed bottom surface of the water pump, (a) point

cloud data, (b) solid mesh model, unsmooth surface textured and (d) smooth textured surface......................................................................................................... 41

Figure 5.17: Genex System reconstructed top surface of the water pump, (a) surface reconstruction with missing data and (b) complete surface reconstruction with filled holes .......................................................................................................................... 41

Figure 5.18: Genex System reconstructed views of the water pump, (a) right view, (b) left view, (c) right/angled view, (d) left/angled view, (e) front/angled view and (f) top/left angled view................................................................................................... 43

Figure 5.19: (a) and (b) Water pump photo image, (c – f) Final 3D CAD model views of the water pump.......................................................................................................... 44

Figure 5.20: Standard deviation calculations of the IVP ramp overlapped with the Genex ramp. ......................................................................................................................... 45

Figure 5.21: Standard deviation of the IVP ramp overlapped with the Genex ramp model................................................................................................................................... 46

Figure 5.22: Simulated ramp. This picture shows the measurements of the ramp. ......... 47

v

LIST OF TABLES

Table 1…………………………………………………………………………………..47 Table 2…………………………………………………………………………………..48

Chapter 1: Introduction 1

1 INTRODUCTION

Engineering is a growing field that continues to evolve to suit the rapid changes of the 21st century. Engineering fields are constantly improving upon current designs and methods to make life simple and easier. When referring to technology, simple and easy can be directly related to fast and accurate. Simple meaning that, you do not use up valuable time in assembly or doing a specific task. Easy meaning how many times you will have to do the process or task. When we think of engineering we think of the general meaning of designing a product from a blue print or plan. Engineering [36] is described as “the application of scientific and mathematical principles to practical ends such as the design, manufacture, and operation of efficient and economical structures, machines, processes, and systems”. This type of engineering is more commonly known as Forward Engineering. An emerging engineering concept is utilizing forward engineering in a reverse way. This method is more commonly referred to as Reverse Engineering. Reverse engineering is the opposite of forward engineering. It takes an existing product, and creates a CAD model, for modification or reproduction to the design aspect of the product. It can also be defined as the process or duplicating an existing component by capturing the components physical dimensions. Reverse engineering is usually undertaken in order to redesign the system for better maintainability or to produce a copy of a system without access to the design from which it was originally produced. With this knowledge, computer vision applications have been tailor to compete in the area of reverse engineering. Computer vision is a computer process concerned with artificial intelligence and image processing of real world images. Typically, computer vision requires a combination of low-level image processing to enhance the image quality (e.g. remove noise, increase contrast) and higher level pattern recognition and image understanding to recognize features present in the image. Three-dimensional (3D) computer vision uses two-dimensional (2D), images to generate a 3D model of a scene or object. There has been a mandatory need for 3D reconstruction of scenes and objects by the manufacturing industry, medical industry, military branches and research facilities. Manufacturing industry utilizes reverse engineering for its fast rapid prototyping abilities and accuracy associated with the production of new parts. This fast prototyping is done


through the use of CAD model designs for inspection purposes. Military branches also utilize reverse engineering to perform inspection task that are associated with safety. One of our laboratory’s current focuses is reverse engineering or 3D reconstruction of objects and scenes from real world data. The goal of reverse engineering an object is to successfully generate a 3D CAD model of an object that can be used for future modeling of parts where there exists no CAD model. We want to generate clean, smooth 3D models, which are free of noise and holes. This requires a strong, robust image acquisition system that can acquire data with a high level of accuracy in a sufficient time frame. Our system uses range and intensity images of objects as input. The output is transformed data that is represented as 3D reconstructions of geometric primitives. There are several building blocks or steps, which determine the process of building a complete 3D model from range and intensity data. These steps, listed in Figure 1.1 show the format of how range image data is acquired, transformed and generated. This flowchart can be characterized as a generic basic principle for reverse engineering. The steps shown often overlap during the process of each stage.

Figure 1.1: Flowchart for basic transformation phases of reverse engineering

There are many different approaches to acquiring 3D data of objects of various structural shapes. All 3D-based machine vision systems ultimately acquire and operate on image data. Acquisition can be based on collecting the Z-axis data using linear area, laser radar laser scanning techniques, point detectors, or other approaches. These systems incorporate the computer power to manage, process and analyze the data acquired. In addition to these tasks, make decisions relating the data to the application without

1. Data Capture

2. Data Segmentation

3. 3D CAD Model


operator intervention. This characterizes what is meant by the term ‘‘3D-based machine vision”.

Traditional processes for reverse engineering of objects and structures from 3D datasets have been initial data (e.g. triangulated models) and parametric surface (e.g. quadratic surface) driven. These approaches have been successful for simple parts, but have resulted in reconstructions that have errors when dealing with more complex structures. Typical errors arise from noisy data or missing data from the surface of the part. Other errors can also consist of incorrect relative positions of the object. Industries are looking for a method to improve upon these errors and migrated toward a fast efficient way of modeling parts for inspection purposes. Traditional practices use CMMs, which are coordinate measuring machines that have a touch probe to model the surface for inspection. Today’s industries are moving toward the improvement of better accuracy and faster inspection time. This can be improved through the integration of laser range scanning. When implementing a non-contact measurement solution, the end user has a large array of commercial systems to select from. To emphasis the use of this project we have chosen two different commercial systems with two different approaches to modeling 3D objects using vision based technology. Our approaches will be using data acquisition systems that are fairly robust to noise and yield high accuracy measurements. The two different techniques consist of using laser lighting and structured lighting techniques. The first system is the IVP Range Laser Scanning System and the second is the Genex 3D FaceCam System.

Although both systems primary focus is reconstruction of real world objects and scene, we will investigate the limitation of both systems. Although the structured lighting system is not designed for reverse engineering use, we will compare the modeling aspects of this system for reverse engineering of automotive parts to the laser range system. Figure 1.2 describes the data flow of our approach applied to both systems. While the approaches are similar, the steps taken may vary for each system. Figure 1.2 is a more detailed description of Figure 1.1. The blue area best describes the data capturing section, while the yellow and orange highlights the data pre and post-processing steps and the final outcome is a 3D CAD model. In the data segmentation stage several steps are taken to generate noise free, smooth models of the part. In data reduction, data such as the noise, outlier or erroneous background information is eliminated. Outliers are false data points that are captured during acquisition. Surface smoothing and multi-view registration are included in data integration. Surface smoothing is an additional feature to eliminate noisy data and make the surface of the object more uniform in texture. This can be performed both before and/or after several views of the part are merged. After all the steps are complete a final 3D CAD model is generated. Data segmentation and integration are discussed in depth in section 3.


Figure 1.2: The sequence of steps required for the reconstruction of a model from multiple overlapping scans.

Section 2 of this paper, will discuss the related merits and methods of reverse engineering and techniques. Section 3 will discuss surface reconstruction for 3D models. Following, is section 4 that discusses the applications and technique of our implemented systems. Finally, we will conclude this paper with the implementation of our procedure for the experimental results. A comparison and summary of the limitations will be expressed in the conclusion section of the paper.

3. Data Integration

1. Data Capture

2. Data Segmentation

Pre-processing

Data Reduction

Next Best View Plan

Noise Filtering

Acquisition of Range Images

Post-processing

Multi-view Registration

Hole Filling

Surface Smoothing

4. 3D CAD Model

Chapter 2: Acquisition Classifications and Related Works 5

2 ACQUISITION CLASSIFICATIONS AND RELATED WORKS

An important part to reverse engineering is data acquisition. Data acquisition systems are constrained by physical considerations to acquire data from a limited region of an objects’ surface. Therefore, multiple scans of the surface must be taken to completely measure a part. After reviewing the most important measuring techniques, the related merits and difficulties associated with these methods are discussed. Figure 2.1 classifies the types of application used for acquiring 3D data into contact and non-contact methods.

Figure 2.1: Classification of data acquisition techniques used in contact and non-contact approaches for reverse engineering systems.

Data Acquisition Methods

Non-contact Methods Tactile Contact Methods

Magnetic Acoustic Optical

Laser Triangulation

CMMs Robotic Arms

Stereo Analysis

Time-of-Flight Structured Lighting

Interferometers


2.1 Contact Data Acquisition Techniques

There are many different methods for acquiring shape data, as shown in Figure 2.1. Tactile methods represent a popular approach to shape capture. The two most commonly known forms are Coordinate Measuring Machines (CMMs) and mechanical or robotic arms with a touch probe sensing device. CMMs are often used when high precision is required. It is considered a contact type method that is NC-driven and can program sampling of points for predefined features efficiently [16]. These machines can be programmed to follow paths along a surface and collect very accurate, nearly noise-free data. Sampling basis is the only way a part can be inspected using a CMM. The ability for obtaining large amounts of point data from the parts surface quickly for complete inspection needs to be the number one quality of an inspection device. This needs to be done in conjunction with the idea that the part has free-forming surfaces. A 3-axis milling machine is an example of a mechanical or robotic arm. These machines can be fitted with a touch probe, as mentioned before, and used as a tactile measuring system. However, it is not very effective for concave surfaces. There are many different other robotic devices which are used because of their ability to have less noise and have a desirable accuracy, but like the CMM, they are the slowest method for data acquisition.

There are disadvantages when using a CMM or robotic arm to model surfaces of parts. The disadvantages of CMMs having contact to the surface of an object can damage the object. The reason being is if the surface texture is soft, holes can be inflicted on the surface. CMMs also show difficulties in measuring parts with free form surfaces. The part might have indentions that are too small. Flexibility of parts makes it very difficult to contact the surface with a touch probe without creating an indentation that detracts from the accuracy of the measurements. For CMM, geometric complexity increases the number of points required for accurate measurements. The time needed to capture points one by one can range from days or sometimes weeks for complicated parts. There are also external factors that affect the accuracy of a CMM. The main ones are temperature, vibration and humidity [1].

Xiong [33] gives an in depth discussion of measurement and profile error in tactile measurement. Sahoo and Menq [26] use tactile systems for sensing complex sculptured surfaces. Butler [4] provides a comparison of tactile methods and their performance.


2.2 Reverse Engineering Applying Contact Techniques

Reverse engineering is a growing industrial market for manufacturing and development. Various individuals and groups have developed new techniques, which have been improvements to the current existing techniques available. The first technique and method we visit is that of Thompson at el., [28]. Their research describes a prototype of a reverse engineering system which uses manufacturing features as geometric primitives for mechanical parts. Their method is geared toward reverse engineering of mechanical parts. They have identified a feature based approach that they state produces highly accurate models, even when the original 3D sensor data has substantial errors. In the feature based approach, the registration of two different point clouds is performed by matching three points to three points, three spheres to three spheres, or three planes to three planes [28].

Their main innovation was to use the features to fit scanned data, rather than using triangulated meshes or parametric surfaces patches. This research claims to have advantages over the current practice of ordinary CMMs. It states the resulting models can be directly imported into featured-based CAD systems without loss of the semantics and topological information inherent in featured-based representations. It assumes that the models are already generated into CAD model formation. The results from Thompson et al., [28] for quantitatively evaluating the accuracy of the models using the feature-based modeling approach, parts were used only if they had the original CAD model. The part was machined out of aluminum using a 3-axis NC mill. A non-contact digitizer measured the surface points. New CAD models were generated using the REFAB (Reverse Engineering – FeAture-Based) system. The system designs a model composed of mechanical features from a set of 3D surface points that could be defined by users. The geometric differences between the original and new generated model was computed. Further details and results from their work can be read in reference [28]. The second technique we visit is that of Yu Zhang [35]. This research focuses on the engineering application of reverse engineering. The system employed is built with a coordinate measurement machine and CAD/CAM software. By scanning the physical object, the measurement data is acquired. The basic principles of reverse engineering was applied to the design and manufacturing of the die of a diesel engine. The process described in the paper is the object digitization and CAD model reconstruction to NC machining. The die’s geometric shape is measured and data is acquired using a CMM in conjunction with KUM measurement software that has a linear scan mode. The number of points measured is determined automatically by the CMM according to the curvature change of


the surface. This is measured on the tactile point. The CMM that is used can measure about 1600 points for each scanned curve. Usually a machining tracing process results in a structured point of sequences with a large number of points and a line structure. The result of Zhang’s [35] system is a self-developed program to realize the transformation of the format of the measured data information from the CMM and KUM. First, the format of the measured data is transformed into an acceptable format for the software used. Then the data is filtered out and processed in a visualized way. The processed data is directly used for the creation of the die CAD model. After the CAD model of the die is complete, the NC machining process planning can generate the location for cutting the manufacturing application. The die is finally machined by the NC machine tool using the created CAD model.

2.3 Non-Contact Data Acquisition Techniques

Non-contact methods use light, sound or magnetic fields to acquire shape from objects. In the case of contact and non-contact, an appropriate analysis must be performed to determine the positions of the points on the objects surface. Each method has strengths and weaknesses that require the data acquisition system to be carefully selected for the shape capture functionality desired. Optical methods of shape capture are probably the broadest and growing in popularity over contact methods. This is because they have relatively fast acquisition rates. There are five important categories of optical methods: laser triangulation, time-of-flight, interferometers, structured lighting and stereo analysis. This section will discuss the various principles of each method.

Laser Triangulation is a method, which uses location and angles between light sources and photo sensing devices to deduce position. A high-energy light source is focused and projected at a pre-specified angle at the surface of interest. A photosensitive device, usually a video camera, senses the reflection of the surface and then by using geometric triangulation from the known angle and distances, the position of a surface point relative to a reference plane can be calculated. The light source and the camera can be mounted on a traveling platform which then produces multiple scans of the surface. These scans are therefore relative measurements of the surface of interest. Various different high-energy light sources are used, but lasers are the most common. Triangulation can acquire data at very fast rates. The accuracy is determined by the resolution of the photosensitive device and the distance between the surface and the scanner. Motavalli et al., [19] presents a reverse engineering strategy using laser triangulation. Moss et al., [22] present a detailed discussion of a classic laser triangulation system used to capture shape data from facial surfaces. The use of laser triangulation on a


coordinate measuring machine is presented by Modjarred [20]. These references give a broad survey of methods, approaches and limitations of triangulation.

Measuring distance by sensing time-of-flight of the light beams emitted is the way a ranging system works. Practical methods are usually based on lasers and pulsating beams. For example, in laser range finders, the time-of-flight is used to determine the distance traveled, and in stereo analysis the relative locations of landmarks in multiple images are related to position. Interferometer methods measure the distance in terms of wavelengths using interference patterns. This can be a very accurate method of measurement since visible light has a wavelength of the order of hundreds of nanometers, while most reverse engineering applications distances are in the centimeter to meter range. In principle, other parts of the electromagnetic spectrum could also be used. In practice, a high-energy light source is used to provide both a beam of monochromatic light to probe the object and a reference beam for comparison with the reflected light. Moring et al., [21] describe a range finder based on time-of-flight calculations. The article presents some information on accuracy and performance. Jarvis [14] presents an in-depth article on time-of- flight range finders giving detailed results and analysis. Structured lighting involves projecting patterns of light upon a surface of interest and capturing an image of the resulting pattern as reflected by the surface. The image must then be analyzed to determine coordinates of data points on the surface. A popular method of structured lighting is shadow Moire, where an interference pattern is projected onto a surface producing lighted contour lines. These contour lines are captured in an image and are analyzed to determine distances between the lines. This distance is proportional to the height of the surface at the point of interest and so the coordinates of surface points can be deduced. Structured lighting can acquire large amounts of data with a single image frame, but the analysis to determine positions of data can be rather complex. Will and Pennington [30] use grids projected onto the surface of objects to determine point locations. Wang and Aggarwal [32] use a similar approach but use stripes of light and multiple images. The final optical shape capture method of interest is stereo image analysis. This is similar to structured lighting methods in that frames are analyzed to determine coordinate data. However, the analysis does not rely on projected patterns. Instead, typically, stereo pairs are used to provide enough information to determine height and coordinate position. This method is often referred to as a passive method since no structured lighting is used. Active methods are distinguished from passive methods in that artificial light is used in the acquisition of data. Correlation of image pairs and landmarks within the images are big difficulties with this method and this is why active methods are preferred. Another stereo image analysis approach deals with lighting models, where an image is compared to a 3D model. The model is modified until the shaded images match the real images of the object of interest. Finally, intensity patterns within images can be used to determine coordinate information.


The final types of data acquisition methods we will examine are acoustic, where sound is reflected from a surface, magnetic, where a magnetic field touches the surface and a hybrid of both contact and non-contact. Acoustic methods have been used for decades for distance measuring. Sonar is used extensively for this purpose. Automatic focus cameras often use acoustic methods to determine range. The method is essentially the same as time-of-flight, where a sound source is reflected off a surface and then distance between the source and surface is determined knowing the speed of sound. Acoustic interference or noise is often a problem as well as determining focused point locations. Dynamic imaging is used extensively in ultra-sound devices where a transducer can sweep a cross-section through an object to capture material data internal to an object.

Magnetic field measurement involves sensing the strength of a magnetic field source. Magnetic touch probes are used which usually sense the location and orientation of a stylus within the field. A trigger allows the user to only record specific point data once the stylus is positioned at a point of interest. Magnetic resonance is used in similar applications to ultra-sound when internal material properties are to be measured. MRI (magnetic resonance) activates atoms in the material to be measured and then measures the response. Hybrid modeling systems are a combination of contact and non-contact systems. They can also be a combination of NC coding and laser scanning techniques. The first type usually consists of the coordinate measuring machine and integrated laser based technology. The second may consist of some other form of a non-contact technique such as software and laser-based technology integrated as one system. Hybrid based applications will be discussed in the next section. To conclude this section, all measuring methods must interact with the surface or internal material using some phenomenon, either light, sound, magnetism or physical surface contact. The speed with which the phenomenon operates as well as the speed of the sensor device determines the speed of the data acquisition. The sensor type selected also determines the amount of analysis needed to compute the measured data and the accuracy.

2.4 Reverse Engineering Applying Non-Contact and Hybrid Techniques

The first non-contact techniques that we explore are that of Fan and Tsai [13]. They present a measurement system that includes the combination of two CCD cameras, a line laser and a three-axis motion stage. They formed an optical non-contact scanning setup that works with the mathematical method of direct shape error analysis for engineering purposes. The profile measurement of free-form objects can be analyzed. Matching the images of the free-form surfaces with sufficient efficiency and accuracy is the final result.


Fan and Tsai research adopted the bicubic uniform B-spline interpolation approach for the shape error analysis method. This method was used to describe the first set of measurement point and to generate reconstructed multiple patches of the surface. The shape error analysis main function is to sum the squared nearest distances [13]. Based on this principle they have developed an algorithm called the direct method or DSEAM (Direct shape error analysis method). This algorithm is an adopted variation of the shape error algorithm, and the DFPM (Davidon-Fletcher-Powell Method) algorithm. Refer to Fan and Tsai [13] for detailed information on the DFPM algorithm. The B-spline surface construction and the DFPM algorithm are the foundation of their algorithm. They developed a computer program to analysis the shape error with respect to the surface that is referenced. The results of their approach are of that used on a free-form surface and a car rear-view mirror case. They report the rigid body transformation from the optimal shape error results and the optimal parameters using the DSEAM. They have reported a reduction in the shape error from their technique compared to the initial shape error of the objects. Further reading on their results can be viewed in reference [13]. The first hybrid-based technique reviewed is that explored by Jim Clark [8]. The technique focuses on modeling complex and free-form shapes of mechanical objects by comparing contact and non-contact methods for digitizing the surface. A hybrid- triangulation based hand held system integrated with a coordinate measuring machine is used for this approach. The effects of ambient lighting are discussed for non-contact systems. Whether or not the system can measure the ambient lighting depends on the projected color of light on the object. Clark summaries by writing that if a system projects laser light then the unwanted frequencies can be filtered out. If the system projects white light, then no particular frequencies can be blocked out. This is because it might be carrying the information required to measure the object. Therefore, white light area based systems will be limited in their ability to measure ambient lighting verses laser based systems. The results from Clark are of that modeled using a water pump. The water pump was scanned using both a contact and non-contact system. The results were compared based on the surface quality and the point cloud data obtained. He demonstrates that non-contact techniques in conjunction with advanced surfacing and inspection software yield sufficient results for the mechanical design process. Further reading of his work can be viewed in reference [8]. Chow et al., [7] developed an integrated laser-based reverse engineering and CAM machining system called RECSI (Reverse Engineering and CAM System Integration). They evaluate the feasibility of using concurrent engineering and reverse engineering methods with the data from laser scanning to remanufacture complex geometrical parts.


The first phase of their research demonstrates that laser scanning and CAD model reconstruction can duplicate aircraft structural components accurately and efficiently within a given tolerance. The second phase is the actual development of the system. The system utilizes NC coding generated from the software. The goal of the system is to show that an integrated reverse engineering and CAM machining system can make the remanufacturing process more automatic and efficient. Chow et al., [7] results are that of the comparison between the original parts and the duplicated parts. The samples were performed to evaluate the accuracy and efficiency of their concurrent reverse engineering system. They reported their findings of the errors of the overall integrated system were close to the calculated errors in the results of the reverse engineering feasibility study. The comparison table of the results and the time required to complete each step can be view in their paper [7]. To summarize this section, Fan and Tsai [13] implemented a non-contact system that utilizes CCD cameras and laser triangulation for reverse engineering. Clark [8] implemented a non-contact system that works in conjunction with surfacing and inspection software. He also discussed some of the issues regarding the implementation of such systems for manufacturing purposes. Chow et al., [7], developed and implemented a process planning system that interfaces with a tightly coupled CAD modeling system and CAM tooling path. They demonstrate the accuracy and efficiency of their laser-based reverse engineering system.

2.5 General Constraints of Data Acquisition Techniques

There are many practical problems with acquiring useable data, the major ones being:

a. Calibration b. Accuracy c. Accessibility d. Occlusion e. Fixture (placement) f. Multiple views g. Noise and incomplete data h. Statistical distributions of parts i. Surface finish

Calibration is an essential part of setting up and operating a position-measuring device. Systematic sensing errors can occur through lens distortions, non-linear electronics in cameras, and similar sources. Any sensing must be calibrated so as to, first, accurately determine parameters such as camera points and orientations, and second, to model and allow for as accurately as possible systematic sources of error. Most of the papers cited


present some discussion of accuracy ranges for the various types of scanners, but all methods of data acquisition require accurate calibration. Optical scanners' accuracies typically depend largely on the resolution of the video system used. Distance from the measured surface and accuracy of the moving parts of the scanning system all contribute to the overall measurement error.

Accessibility is the issue of scanning data that is not easily acquired due to the configuration or topology of the part. This usually requires multiple scans but can also make some data impossible to acquire with certain methods. Through holes are typical examples of inaccessible surfaces. Occlusion is the blocking of the scanning medium due to shadowing or obstruction. This is primarily a problem with optical scanners. However, acoustic and magnetic scanners may also have this problem. Multiple scanning devices are one approach to obviate this problem. As well as self-occlusion, occlusion may also arise due to fixtures-typically parts must be clamped before scanning. The geometry of the fixture used, becomes a part of the scan data. Elimination of fixture data is difficult and often requires multiple views. Also central gravity of the part makes most surfaces of the object difficult to scan. Multiple views introduce errors in acquired data because of registration problems (see more details later). Noise elimination in data samples is a difficult issue. Noise can be introduced in a multitude of ways, from extraneous vibrations, specular reflections, etc. There are many different filtering approaches that can be used. An important question is whether to eliminate the noise before, after, or during the model building stage. There are times when the noise should not be eliminated at all. Noise filtering, though, is often an unavoidable step in reverse engineering, but note, that this also destroys the "sharpness" of the data i.e. typically sharp edges disappear and are replaced by smooth blends, which in some cases may be desirable, but in other cases may lead to serious problems in identifying features. A similar problem is restoration of missing data. This is partly necessary due to the above-mentioned inaccessibility and occlusion problems. Moreover, because of the nature of optical and even tactile scanning, the data close to sharp edges is also fairly unreliable. Finally there are situations where only parts of a certain surface can be measured, there are missing parts or parts obscured by other elements, but we need to reconstruct the whole surface from just the visible parts. Further ideas on surface extensions, intersections and patching holes are given in the last part of the paper. Statistical distribution of parts deals with the fact that any given part, which is scanned, only represents one sample in a distributed population. When reverse engineering methods attempt to reproduce a given shape, the tolerance distribution of the scanned part must be considered. This gives rise to multiple part scans and the averaging of the


resulting data. However, it may be somewhat impractical to attempt to sample many parts from a population, and indeed, often only one is available. The final issue we bring up is surface finish of the part being measured. Smoothness and material coatings can dramatically affect the data acquisition process. Tactile or optical methods will produce more noise with a rough surface than a smooth one. Reflective coatings also can affect optical methods. Imagine an ideal scanner: the object is 'floating' in 3D spaces, so it is accessible from all directions. The data is captured in one coordinate system with high accuracy, with no need for noise filtering and registration. Possibly, the measurement is adaptive, i.e. more points are collected at highly curved surface portions, etc. Unfortunately, such a device does not exist at present. But, despite the practical problems discussed, it is possible to obtain large amounts of surface data in reasonably short periods of time even today using the methods described. Once the measured data is acquired, the process of recognition and model building can begin. The imperfect nature of the data, particularly inaccuracy and incompleteness, however, makes these steps fairly difficult as will be seen in the following sections.

Chapter 3: Surface Reconstruction

15

3 SURFACE RECONSTRUCTION

Obtaining a surface representation of objects and scenes has always been one of the most challenging and fundamental problems of 3D computer vision. In this section we will discuss the procedures to generating a successful 3D model of an object from single and multiple views. We will also discuss various surface reconstruction algorithms that have been developed and employed.

3.1 Single Views and Data Segmentation

An iterative process for 3D reconstructions of surfaces in static environments is defined by the following steps (see also Figure 1.2, which depicts this process):

1. Acquiring range images of the part 2. Pre-processing acquired data (data segmentation) 3. Data post-processing (data integration) 4. Final 3D CAD model

Range image acquisition is the first step of the process. In image acquisition, the CCD camera captures the scene. The result is a grey scale-image, which shows the intersection between the laser plane and the object, which is a line. The image acquisition process yields a number of selected range images. The range images that are generated are of the view angles that the user has positioned the part for capture. Figure 3.1 shows a sequence of range images obtained from the IVP Range Scanner. The black area in between the blades of the water pump show occluded areas. These are depth range images. The purposes of taking multiple views are to eliminate the missing data from the water pump. The structured lighting system that we use for our project generates a different set of images when performing the image acquisition process. The images obtained from this system are color range images. These images also show the texture of the part or object that is being scanned. Although both system produces different types of images, they are all part of the acquisition step of the process. Figure 3.2 below is an example of the type of images that are generated from the structured lighting system used in this project. The first scans from both systems are 2D representations of the real object. Although both the systems generate 2D views of the part, they are still considered single views. They are single views because one view cannot complete the reconstruction of the object.


16

(a) (b)

(c) (d)

(e) (f)

Figure 3.1: (a) Original image of the water pump, (b) Sequenced single view range images of the bottom surface of the water pump generated using our laser range scanner


17

(a) (b)

(c) (d)

Figure 3.2: Textured single view range images generated from the structured lighting system, (a) left view of the ramp, (b) front view of the ramp, (c) back/left view of the ramp, (d) front/ right view of the ramp.

Pre-processing the range images is the step proceeding data collection. Pre-processing is more commonly referred to as cleaning of the collected data. Pre-processing is applied to the single views individual before they are integrated and registered together. As mentioned before pre-processing includes reducing erroneous data, filtering noise and filling holes that may have occurred as a result of occlusions. The additional steps of the process are described in the next few sections.

3.2 Multiple View Integration and Registration

For all objects and parts to be scanned, we require a geometric model of the whole object’s surface, as stated above. Ideally, we would have the part “floating” in 3D space


18

(in a fixed position with a fixed orientation). This would be so that the scanner could move around the object from all sides to capture in a single coordinate system. In practice, the object will have to rest on some surface, so part of it is inaccessible to the scanner. Furthermore, if the scanner is fixed in position n, at any one time, it will be able to capture data from an even more limited region of the object’s surface. Thus, generally, it will be necessary to combine multiple views taken with the object placed in different orientations in front of the scanner. The more scans that are taken of the part, the longer the whole scanning process will take. For this reason, it is important to decide how many scans will be taken of the part or object. This can be determined based on the size and material make up the part. This is the second step of the post-processing stage for 3D reconstructions. The individual range images must be aligned, or registered, into a common coordinate system so that they can be integrated into a single 3D model. In high-end systems, registration may be performed by accurate tracking. For instance, the scanner may be attached to a coordinate measurement machine that tracks its position and orientation with a high degree of accuracy. Passive mechanical arms as well as robots have been used. Automatic feature matching for computing the initial alignments is an active area of research (recent work includes [3, 6, 9-12, 24, 27, 35]). The most general formulation of the problem that makes no assumptions on type of features (in the range and/or associated intensity images) and initial approximate registration is extremely hard to solve. Approximate position and orientation of the scanner can be tracked with fairly inexpensive hardware in most situations, and can be used as a starting point to avoid searching a large parameter space. The goal of collecting multiple views of range images is to take these sets of range images and register the images. We want to take these sets of registered range images of the entire surface of the object and from these images produce a corresponding set of parametric surface patches. Each range image is a dense sampling of the 3D geometry of the surface from a particular viewpoint. When reconstructing objects we want to have overlapping views of the object. The overlapping views should consist of the same area of the object being scanned. The main purpose for overlapping the different views is to omit occlusion in the object by matching various similar features on the object. Also, some views may have more details or some may have low resolution. In our system, the registration is a fairly easy process. Special software is use to match similar features and points on the different surfaces scanned. After the points have been selected for matching, they are matched based on the distance computed using the ICP (Iterative Closet Point) algorithm in the software. Figure 3.3 is an example of how our software registers two sets of range image view. Three or more points are matched based on similar corresponding features and feature location. This is commonly referred to as feature matching or extraction. Figure 3.3 a, contains one single view of the water pump and Figure 3.3 b is another single view. Figure 3.3 c is the registration of both views overlapping. The overlapping red region shows how they are aligned in the same shell.


19

(a) (b)

(c)

Figure 3.3: Point cloud information of the side view of the water pump, (a) and (b) side view range images, (c) registered range image of the two side views.

The registration of the views may take several tries to achieve the optimal aligning of the images that is desired. After the two desired views are registered to each other at the ideal feature locations, they can be merged together using various post-processing option offered in commercial software or by generated programs.

3.3 Post-processing Registered Images

After all the views of the object have been obtained, they are ready for post-processing. As mentioned above in Figure 1.2, post-processing of range images includes surface


20

smoothing and multiple view registration. This is the step before completing the 3D reconstruction of an object or part. Post-processing operations are often necessary to adapt the model resulting from scan integration to the application at hand. In other words, when performing the post-processing step, the registration process is made permanent. There are two techniques that are commonly used for this process. The first is surface-based methods and the second is volumetric methods. Surface-based methods create the surface by locally parameterization the surface and connecting each point to its neighbors by local operations [2]. The partial connectivity that is implicit in the range images is made use of in other methods. For example, Turk and Levoy [29] zippering approach works by triangulating all the range scans individually. Redundant overlapping triangles are then eroded for removal from the partial meshes. The intersecting regions are locally re-triangulated and then trimmed to create one seamless surface. The vertex positions are then readjusted to reduce error. Volumetric-based methods are useful for very large datasets. In volumetric methods, line of sight error compensation is done by computing a scalar field that approximates the signed distance to the true surface [2]. This is based on a weighted average of distances from sample points on the individual range image scans. Volumetric methods are also well suited to producing watertight models. The range images are used to carve out a spatial volume then, object definition can be obtained without holes in the surface. Solid modeling evolution from a series of range images can be demonstrated by Reed and Allen [23]. With the data obtained from each range image, they carve away the solid that lies between the scanner and each sampled data point. In our system, we use special software to perform the post-processing of the individual views. Rapidform2004 is 3D modeling software that we use to generate our complete 3D CAD models of our parts. We are allotted three options, surface merging, volumetric merging and point cloud merging. The surface merging option merges shells of range views that have been aligned by the registration process into one united shell. Overlapping shell regions between the two separate shells are removed and neighboring boundaries are connected together with newly added polygons. One of the built in functions in the software is for merging the different range image views into one united shell. The Volumetric merging option merges multiple shells into a single shell by allocating their geometry information to a reference voxel model with a volumetric method. A voxel is word created from two words (vector and pixel) to describe the 3D space of a pixel-based image. It is a volume element of a rectangular shape of the subject being imaged. Volume based merging is useful when surface based meshing creates poor merge results. This polygon-merging tool helps you to merge scanned data with many holes and messy boundary, or bad aligned data.

Chapter 4: System Descriptions and Setup 21

4. SYSTEM DESCRIPTIONS AND SETUP

There are many profiling systems that can be used to capture data of objects for reconstruction. As mentioned in section 2, the different options fall under the category of contact and non-contact. To emphasize the objective of this project and the flexibility of non-contact options, we have chosen two different profiling systems for our project. Both profiling systems will be discussed in the next two sections.

4.1 IVP Range Scanning Profiling System

There are a number of 3D laser scanners commercially available. Depending on the specific needs for design, a particular type is chosen. The vast majority of 3D non-contact systems employ triangulation. Triangulation systems are often classified as being either active or passive [8]. Passive methods such as stereo or photogrammetric systems use only cameras. Industrial settings however use systems that are active. In that they project some form of illumination onto the object and measure the position of the illumination on the object [8]. This is done using some form of camera or light sensing electronics. A typical triangulation sensor diagram is shown below in Figure 4.1. A typical triangulation scheme projects a point or line (sheet) of laser light on an object, and observes the intersection of the object and laser through electronic cameras. This description describes a static system that only measures points where the laser line and object meet. Too fully measure an object in 3D spaces, the sensor is moved in the x, y and z directions. This is to fully cover the area of the object. By sending laser beams radiated from the surface and received by CCD cameras, the 3D laser-scanning device can acquire the surface information of the part. Consisting of a beam projector radiating the laser beam, the laser scanner and a CCD camera sense the reflected beam from the surface. Laser stripe and point type are normally how laser beams can be categorized. A laser-type scanner radiates a line of laser beams, called a stripe, onto the surface so that several points can be acquired at once. In contrast to the stripe, a point type laser scanner obtains only one point at a time. Based on the configuration of the machine, laser-scanning devices can also be classified on those bases [15]. The first profiling system that we employ to reconstruct a 3D model of an object is the IVP Range Scanning System. This system uses a laser stripe for acquisition. The IVP Ranger SC386 is a laser triangulation scanner for range profiling using the MAPP family


Smart Vision Cameras. Figure 4.2 is the IVP systems equipment and configuration used for data acquisition. Figure 4.2 b shows the Smart Camera and motor in more detail.

Figure 4.1: Principle of a laser triangulation system. Arrangement for the IVP Smart Vision Camera and Laser

The IVP Laser Scanner consists of a thin laser light and a camera that are used to obtain the profile of objects. A thin laser light is projected onto the object and the CCD sensor from the camera detects the scan line (the peak of the reflected laser light). The profiles are displayed as a set of range images. The camera and laser are fixed on a stable structure that moves in a horizontal direction. Figure 4.1 shows the angle placement of the camera to the laser light for the IVP Ranger System. The speed of the laser is controlled through the software that is associated with the system. The black box houses the motor for the system. The speed can be adjusted depending on the quality of scanning that is to be achieved. Because the laser and camera are fixed on the same belt, they move at the same speed. For data to be acquired using the IVP Range Scanner, the system must be correctly calibrated before every successful set of scanned data. Calibration of the IVP Range Scanner involves identifying the correct world coordinate data for the system so that the measurements of the scanned object match both in real world data and transformed data. To calibrate the IVP a calibration grid is used to number all the coordinate data points. There are forty total points that must me identified, ranging from 0 to 39. Figure 4.3 is an example of the calibration grid used to calibrate the IVP Range Scanner. In the

Range distance

Objects

Sensor (CCD

Camera) Field of

View

Baseline distance


calibration grid, all the black dots must turn blue to be recognized by the sensor. Figure 4.4 is the user interface for the IVP Range Scanner. In this user interface window, the camera image of the object is displayed along with the object profile and range image that is acquired.

(a)

(b)

Figure 4.2: Equipment setup for IVP Range Scanning System, (a) front view of the IVP Range Scanner, (b) side angled view of the top of the IVP Range Scanner

Camera

Motor


Figure 4.3: Calibration grid for the IVP Range Scanner

Figure 4.4: IVP Range Scanner User Interface

During calibration of the IVP Range Scanner the lights are turned off so that the calibration grid can be viewed by the camera source. Also the lights are turned out to obtain the correct object profile of the calibration grid. The object profile can be seen as


the white line on the calibration grid in Figure 4.3. The goal of the object profile during calibration is to make sure that the entire object will be in the field-of-view of the camera and laser during data acquisition. The field-of-view refers to the measured distance between the lasers’ light, the camera and object, (refer to Figure 4.2).

4.2 Genex 3D FaceCam Profiling System

As previously mentioned in section 2, structured lighting is the projection of a light pattern (plane, grid, or more complex shape) at a known angle onto an object [25]. Scanning the object with the light pattern constructs 3D information of the shape of the object. This is the basic principle behind depth perception for machines. In most cases, structured lighting can be described as active triangulation. Active triangulation is a simple technique to achieve depth information with the help of structured lighting to scan a scene with a laser plane and to detect the location of the reflected stripes. The distortion along the detected profile is used to compute the depth information. In our acquisition system, the stripe pattern is projected by multiple stripes at once onto the scene. In order to distinguish between the stripes, they are coded (Coded Light Approach) with different brightness or different colors. The light grid has a rainbow color effect with the colors red, green and blue repeating. Figure 4.5 shows an example of a structured lighting grid projected onto our metallic ramp object.

Figure 4.5: Structured light grid pattern projected on the ramp with a neutral tan background

The Coded Light Approach (CLA) is an absolute measurement method of direct codification [5]. This method requires only a small number of images to obtain a full depth-image. This can be achieved with the sequence of projections using a grid of


vertical lines (light or dark). Direct codification is usually constrained to neutral color objects or not highly saturated colors. All the lines are numbered from left to right. In order to achieve a pattern where each pixel coordinate can be directly obtained, it is ideal to use a large range of color values or reduce the range and introduce periodicity in the pattern. There are two common forms of coded light approaches, coding based on grey levels and coding based on color.

In our second acquisition system for this project we use the Genex 3D FaceCam System for reverse engineering purposes. Our purpose for selecting this system is to explore the accuracy and limitation of the machine for reverse engineering of automotive parts. In this system, three regular cameras, a digital camera and a single projector are used verses one single camera and single projector as mentioned earlier. The use of three cameras yields three separate images in the results, (see Figure 4.6 below). A right, left and centered image are obtain from the different cameras lens because of their view position in the system set up. Figure 4.6a is the Genex 3D FaceCam 500 System used in this project. In this system configuration, the projector is located under the center lens. The digital camera is located on top of the center lens.

Figure 4.6: Genex 3D FaceCam System

In our system, there is a specified distance of how far the object can be away from the Genex 3D FaceCam System when acquiring data. The total distance is 85 cm, with an allotted distance of ±15 cm. Figure 4.7 shows the allotted distance when acquiring data from the Genex 3D FaceCam System. When scanning, the background should be a neutral color from the object. This is so that when eliminating the background information, the two can be easily distinguished. The tan background can be seen clearly in Figure 4.8. Also in Figure 4.8, the coded light pattern projected onto the ramp is seen. This snapshot was taken directly from the user


interface screen of the Genex 3D FaceCam system after the left side of the ramp data was collected. Three different view angles are captured with the three cameras to generate a complete model of the left side of the image, as mentioned previously in this section.

Figure 4.7: Distance specification for data collection

Figure 4.8: Genex 3D FaceCam User Interface screen that displays the left, center and right camera photos of the object

Object placement

Camera 3

Camera 1

Camera 2

85 cm Back ground

Chapter 5: Results and Discussions 28

5 RESULTS AND DISCUSSIONS

In this section, we present results of our complete modeling process: data acquisition, data segmentation (pre-processing and post processing) and a 3D CAD model. We will show a comparison of the actual data to the original data. We will also show a comparison of the data results obtained from both systems.

5.1 Part Identification

Our research is geared toward reconstruction of automotive parts. For our part selection, there was no predetermined criterion for selecting the parts. The main purpose was to select parts that were ideal for our system setup and that could be easily rotated. For this project, there are three parts (objects) that have been identified, a water pump, a ramp and an arm pulley. All the selected parts are composed mostly of metal. The first part is shown below in Figure 5.1.

The ramp object in Figure 5.1 is not classified as an automotive part. This specific ramp was specially designed for this project. It was designed on the basis of obtaining ground truth information for the structured lighting system used in our experiments. The main reason for using a ramp shape is that it should be easy to measure and compare the real world dimensions to the CAD coordinate measurements after the complete 3D model has been generated. The measurements of the ramp are 4 inches on the base for length and the height. The sloped side measures 3.5 inches, while the two shorter sides measure 1.4 inches. The weight of the ramp is about 48 ounces (1,587 grams). The corner angles on the ramp are 90 degrees for the back and bottom surface. With the current position of the ramp according to the photo below (Figure 5.1), the side with the holes is a perfect 45 degree angle. If the part is flipped to have the holes on top, then the angle is slightly higher at 47 degrees. This difference in angles does not pose any major difference in obtaining a 3D model of the part.


(a) (b)

(c)

Figure 5.1: Photos of the ramp part (a) front and side views, (b) back and top views, (c) ramp measurements

The water pump in Figure 5.2 was selected because of its complexity in shape and size measurements. Due to the water pump having a non-symmetrical shape, there is no defined center of gravity. The center of gravity refers to if the positioning of the water pump placed on the opposed side, that it would not be balanced. The water pump top base screw has a black tented color. The size of the water pump is 10 inches in length, 5.5 inches wide on the largest area and 2.5 inches wide on the lower base. The total height of the water pump is about 4 inches on the larger half and 1.5 inches on the flat end.

4 in.

1.4 in.

1.5 in.

4 in.

3.5 in.

BackFront


(a) (b)

Figure 5.2: Photos of the water pump part (a) top and side views (b) bottom view

Figure 5.3 is the pulley arm that was selected as the third object to reconstruct. The pulley has a black circular ring located on the top end of the part. It also contains circular holes on the object that may cause pose a challenge in modeling due to occlusion. The pulley arms measures about 11.5 inches in length. The width contains three different measurements. The pulley’s black circle ring measures 1.5 inches while the center measures 2.5 inches and the lower base measures 1.5 inches. The height measurement is 2.8 inches from the ground or working surface. All measurements are taken by hand.

(a) (b)

Figure 5.3: Photos of the pulley arm, (a) is the top view (b) is the bottom view


In the next section we will display our results from both the IVP Range Scanner and the Genex 3D FaceCam System. The results will be separated into three preliminary sections. Each section is a different set of scans, of the part, taken at different times. Each different set of scans were taken with different varying conditions. These conditions include the ambient lighting from the room and outside lighting from the windows and doors.

5.1 IVP Laser Range Data Results

The first set of data results we will discuss are the results from the ramp using the IVP Range Scanner. The ramp posed a challenge while scanning because of the surface finish. The ramp has a highly reflective surface that caused areas of the ramp to be occluded while collecting the data at different orientations. Figure 5.4, a through d, are some CAD model examples of the ramp. These different images are different view angles taken of the ramp.

(a) (b)

(c) (d)

Figure 5.4: CAD model images of the ramp (a) left/front view, (b) right side view, (c) bottom view, (d) back view

In the CAD model images of Figure 5.4, the various views represent different angles of the ramp. Figure 5.4 a, is the right side of the ramp. This figure also shows part of the front view. Figure 5.4 b is the left side of the ramp. It also shows the other half of the front view from a different angle. Figure 5.4 c is the bottom view of the ramp and d is the


back view of the ramp showing the holes. Some of the details of the holes have been lost due to smoothing the surface of the ramp. More complete CAD models can be seen Figure 5.5. In these CAD images, the holes have been filled and the ramp’s surface has been smooth a second time. We then obtain a complete 3D CAD model of the ramp.

(a) (b)

(c) (d)

Figure 5.5: 3D Ramp CAD models, (a) left/front side view, (b) back/left side view, (c) bottom view, (d) right/top/ front side view

The second set of data collected using the IVP Range Scanner was of the water pump. The water pump image can be seen in Section 2.1, Figure 2.1. The lighting factor for this system does not affect the data as much as it affects the Genex System data. The reason is because a filter can be placed on the camera lens to filter out any unnecessary light. For our experiments, we did not use the filter because the light source in our room environment was not a major issue. Figure 5.6 is the first attempt at reconstructing the bottom of the water pump. Figure 5.6 a, is a merge of the different views to obtained this CAD model. After merging the views using the volumetric merge technique, overlapping areas created holes in the model. The


holes were filled to have a more complete water tight model. Part a, of Figure 5.6, is the solid mesh model and part b shows the point cloud information after the holes were filled in part a. The point cloud information usually shows the distance (spaces) between each point cloud, while the solid modeling technique shows a smooth continues image or view. The range images used to reconstruct the bottom surface can be seen in Figure 3.2 above. Figure 5.6 c is the solid mesh model after the holes filling was applied.

(a) (b)

(c)

Figure 5.6: Reconstructed bottom surface of the water pump, (a) Point cloud information, (b) Solid mesh model, (c) smooth reconstructed mesh model

Figure 5.7 a, is a reconstruction of the top view of the water pump. It is displayed in varying colors to show the depth information of the water pump. The blue represents the highest part of the water pump while the orange color represents the surface that is closest to the ground. The green color represents the medium height level of the water pump. This image can be compared to the original photo in Figure 5.2. Figure 5.7 b, is the depth information or rotated side view of the top angle. This view gives a more vivid description of the depth of the water pump. This top view does not have the side views of the water pump merged to it.


(a)

(b)

Figure 5.7: Reconstructed views of the water pump, (a) reconstructed top view (b) side view reconstruction

In Figure 5.8, these images depict the attempts of modeling the right and left side of the water pump. The areas that have occluded data must be taken at a different angle to the side views completely. Figure 5.8 a, and Figure b are the CAD model point cloud data information. Figure 5.8 part a as well as b show the height information in varying colors. Part c and d of Figure 5.8 are the second attempts at modeling the side views. The missing data can be seen clearly in these views. These views will be merged together to complete the side profile of the water pump. Figure 5.9, shows the complete 3D CAD model of the water pump obtained using the IVP Range scanner. In this figure, the front complete view of the water pump is shown. Again, the varying colors show the height changes in the water pump. The pulley arm results are displayed in Figure 5.10 b and a below. In this figure, the original photo image of the pulley arm is shown as well as the point cloud information. Figure 5.10 a, is the bottom side of the pulley arm and Figure 5.10 b is the top side. The color variations in the CAD model images are the height relative to the laser light. To complete this model, additional view must be merged to the current views to have a successful 3D model of the pulley arm. This will eliminate all the holes and occluded areas of the pulley arm.


(a) (b)

(c) (d)

Figure 5.8: Cleaned point cloud data of the side views of the water pump taken with the IVP Range Scanner, (a) right view, (b) left view, (c) right view and (d) left view

(a)

(b)

(c)

Figure 5.9: Point cloud model of the water pump (a) CAD model showing height variations (b) top view, (c) back view


(a) (b) (c)

Figure 5.10: Pulley Arm (a) photo of the side view profile of the pulley arm, (b) point cloud CAD model of the bottom side, (b) point cloud CAD model of the top side

5.2 Genex Structured Light Data Results

To collect data using the Genex 3D FaceCam System, as mentioned before, there must be a neutral back ground to contrast against the object being scanned. For the first set of data collection, we used a box as the neutral back ground. The box (Figure 5.14) also served the purpose of controlling ambient light that was present from the window and room lighting. Figure 5.14 shows an example of the water pump placed in the box to control the light and still maintain a neutral background. Controlling the ambient light was important in the case of the ramp because, the part was highly reflective. We wanted to eliminate as much of the light being reflected off the part back into the camera lens. This would minimize the occlusion in the data sets. Figure 5.11 are a few textured range images collected from the Genex System. The different images show different views and orientations of the ramp. In Figure 5.11, parts a, and b, the ramp is position on a small black object to ensure that the detailed edges of the ramp are captured when collecting the data. In part c and d, the black regions on the ramp are areas that were not captured due to positioning of the object and lighting factors.


(a) (b)

(c) (d)

Figure 5.11: Textured ramp images using the Genex System, (a) back view, (b) right/ back side view, (c) left view and (d) front view

In the images that were captured of the ramp, there is data missing in sections of the ramp. This is because these parts of the ramp reflected the light that was projected onto the object by the cameras during data capture. More views of the ramp must be taken at different angles to fill in the missing data. Examples of the missing data can be seen clearly in the CAD models of the ramp displayed in Figure 5.12. Figure 5.12 are the first attempts of reconstructing the ramp using the Genex System. In this reconstruction some of the edge detail is missing (occluded) due to the edges being reflective. To solve this problem, more scans are taken of the ramp and merged to the already acquired range images. Figure 5.12 a, is the right side of the ramp. Figure 5.12 b, is the back view of the water pump. Part c and d of Figure 5.12 are the reconstructed front and left views. All the views can be compared to the original photo image in Figure 5.1 of section 5.


(a) (b)

(c) (d)

Figure 5.12: Genex System reconstructed views of the water pump, (a) right view, (b) back view, (c) front view and (d) left view

Figure 5.13 are the complete 3D CAD models of the ramp obtained from the Genex System. This is the second attempt at modeling the ramp. Compared to Figure 5.12, these CAD models images show the edge details that were missing and also have a smoother finish. Figure 5.13 a, is the front CAD model view. Figure 5.13 b is the back CAD model view. In Figure 5.13 b, you can see that there is some miss alignment of the views causing the unleveled edge detail. With proper smoothing, this unevenness can be fixed.


(a) (b)

(c) (d)

Figure 5.13: Complete final 3D model of the ramp (a) front view (b) back view (c) side view (d) bottom view

Figure 5.14: Water pump placed in box for neutral background


Figure 5.15 shows a few of the textured water pump range images that were captured using the Genex 3D System. The images contrast has been enhanced for better viewing of the details captured. These images also show some of the background from the box that was captured. The excess background does appear in the raw data files from the Genex 3D System. It is easily cleaned away using some data reduction techniques in the software.

(a) (b)

(c) (d)

Figure 5.15: Textured water pump images using the Genex System, (a) top view, (b) view of the bottom surface, (c) side view and (d) top view

Figure 5.16 are the CAD models of the water pump. These CAD models displayed are the first attempts at reconstruction of the bottom surface of the water pump. The reconstruction efforts yielded successful results as far as most of the detail of the water pump. The edges of the water pump are not refined or smooth. To obtain the detail of the holes along the outer brim of the water pump, more views can be merged to the current model. Figure 5.16 a, is the point cloud CAD model of the bottom surface. Figure 5.16, b through d are more detailed CAD models of the bottom surface. The blades of the bottom surface can be seen clearly and the edges are more distinct as you view the images starting from b and ending at d. Also Figure 5.16 d has a smoother surface finish compared to b and c.


(a) (b)

(c) (d)

Figure 5.16: Genex System reconstructed bottom surface of the water pump, (a) point cloud data, (b) solid mesh model, unsmooth surface textured and (d) smooth textured surface

Figure 5.17 displays the first and second attempts of reconstructing the top surface of water pump. In the first attempt, there are occlusions in the center of the pump. The occluded data is recovered with more range scans, seen in the second attempt of Figure 5.17 b. Some hole filling was also performed on the CAD model. The hole filling creates a more complete and smooth surface in the final model. However, with any smoothing technique that is used, some of the detail of the CAD model will be lost. The loss of detail can be seen on the outer brim on the holes of the model in Figure 5.17 b. Nevertheless, a good percentage of the detail, on the ridges, is still maintained.

(a) (b)

Figure 5.17: Genex System reconstructed top surface of the water pump, (a) surface reconstruction with missing data and (b) complete surface reconstruction with filled holes


Figure 5.18 below, are some CAD models of the side views of the water pump using the Genex System. Part a, is the right side and part b is the left side. A better angle of the views can be seen in parts c and d of the image. These images show more detail of the sides of the water pump. The views of the right and left side are used to complete the top and bottom surfaces of the water pump. When merged together, they show a more complete top and side view. As the CAD images are viewed from Figure 5.18 a, to f, it becomes more distinct of how the views are merged together to create a complete model. Figure 5.18 a, is the right side view, and b is the left side view.

Figure 5.18 c is another right side view that is angled to show how much of the detail is captured when the part is repositioned at a new orientation. Figure 5.18 d is a merging of two different views to complete the left side of the water pump. Once the left side is complete, the right side CAD model can be merged onto the existing model. The final merging can be seen below, in Figure 5.19. The first two images in Figure 5.19 are the original photos of the water pump. The photos are placed there for comparison purposes to the 3D model in c through f. Figure 5.19 shows four different views of the water pump. The views show the front, right side, left side and back of the water pump. This 3D CAD model shows all the details of the original water pump. As compared to the original water pump image, the details of the side profile of the CAD model still contains some unsmooth surfaces. By applying the smoothing technique again to the model, the unsmooth surfaces will even out. There is however a drawback to smooth the surface several times. This drawback is that the smaller details of the edges and raised surfaces may also be smooth. This extra smoothing results in the loss of important details. The next section is a conclusion of this paper.


(a) (b)

(c) (d)

(d) (f)

Figure 5.18: Genex System reconstructed views of the water pump, (a) right view, (b) left view, (c) right/angled view, (d) left/angled view, (e) front/angled view and (f) top/left angled view


(a) (b)

(c) (d)

(e) (f)

Figure 5.19: (a) and (b) Water pump photo image, (c – f) Final 3D CAD model views of the water pump.


5.3 Comparison: IVP Laser Range Scanner and Genex 3D FaceCam Data

In this section, a comparison of the IVP Range Scanner and the Genex 3D FaceCam is done. We first compare the ramp models generated from both systems. Figure 5.20 are rotated examples of the standard deviation calculated from both ramp models. In this figure, the IVP ramp and the Genex ramp are overlapped on top of each other to compare the surface differences. The blue regions represent the surfaces that are touching.

Figure 5.20: Standard deviation calculations of the IVP ramp overlapped with the Genex ramp.

a b

c d

e


Meaning that, in the overlapping surfaces the blue region represents a distance of 0 mm between the two surfaces. The position 0.0 means zero positioning of the coordinate distance. The red region represents a maximum deviation of 3.91693 mm. The standard deviation is calculated by registering the two models into one shell and then using the standard deviation option in the software to perform the calculations. The distance values shown here may be a little different from the ones shown in real time when the deviation is calculated. This is because the picking of the points are manual and leave error for not always being precise. The average of the deviation between the surfaces is calculated to be about .9 mm distance apart shown in Figure 5.21.

Figure 5.21: Standard deviation of the IVP ramp overlapped with the Genex ramp model

Table 1 below shows the derived measurements of the ramp compared to actual measurements. The derived measurements are from the 3D CAD model that was generated from the data collected using both systems. The measurements are off from the actual measurements. This is because the points used to determine the measurements are chosen manually. Figure 5.22 is to show the comparison with the measurements of the sides used in the table.


Figure 5.22: Simulated ramp. This picture shows the measurements of the ramp.

Table 1 Actual Ramp Measurements

IVP Laser Range Scanner Results

Genex 3D FaceCam Results

width: 3 in (76.2 mm)

2.95 in (74.93 mm)

2.93 in (74.422 mm)

length: 4 in (101.6 mm)

3.62 in (91.948 mm)

3.6 in (91.44 mm)

height: 4 in (101.6 mm)

3.86 in (98.044 mm)

3.78 in (96.012 mm)

1.5 in. (38.1 mm)

1.41 in. (35.814 mm)

1.3 in. (33.02 mm)

Table 2 is a comparison of both the IVP Range Scanner and the Genex 3D FaceCam system. This comparisons table was generated based on observations of the systems performance, and overall flexibility to acquire data. The Genex system has the advantage of fewer views when the object is of high complexity. Fewer views means, there is less occlusion in the data. Take the water pump for example (Figure 5.2), because the side of the water pump is small in area, the Genex system is able to capture the side and either some area of the top and bottom of the water pump. This makes it easier when overlapping the views to recreate the object. The IVP Range Scanner requires fixing the object at particular angles to capture the same profile. This could require placing the object on a smaller object that is able to absorb the laser light or be hidden from the laser. The flexibility of both systems is the same in comparison, in that they can both be easily moved. The contrast is that the IVP Range Scanner requires some assembly of the system. This is primarily because the current setup is built in house according to our specifications. Both systems are user friendly as far as setup and data collection is required. The contrast in the systems is that when setting up the IVP Range Scanner to

1.5 in

3 in

4 in

4 in


acquire data, the system must be calibrated before any data is collected. As mentioned previously, in the process of calibrating the IVP, the world coordinate points have to be input manually to make sure that the data is transformed properly into the 3D software. Also the data must have some y-scaling applied to it to make sure that the acquired data measurements are similar to the real data of the water pump. After the scans of the object have been collected, the IVP creates a shadow affect of the object. This is because of the laser light reflecting the object. The Genex 3D FaceCam does not have this problem. This is because the Genex system captures data from the front of the object as opposed to on an angle like the IVP Range Scanner. However, what does affect the Genex system is the ambient lighting that is in the room. The ambient lighting (if there is too much present) can saturate the object causing distortion to the data acquired. Both systems pose the same problem when registration of different views comes into play. The way to over come this problem is the next best view of capturing the object from certain angles where there is overlap of the surfaces. The noise in the images from both systems is not a big factor when collecting data. The noise for the IVP can be reduced as long as the speed of the scan is sufficient to have a smooth surface. The achieve a smooth surface, the speed should be slowed as to avoid ridged jerks of the laser as it scans the profile of the object. The Genex system has the option of reducing noise when the post-processing is performed on the collected data. Table 2

IVP Range Scanner Genex System 1 Creates a shadow effect when scanning 1 Creates a distortion from the ambient

lighting 2 Difficult in position of objects

(complexity in shape) 2 Difficult in position of objects with

current system set up 3 Needs more overlapping views 3 Needs fewer views to complete 3D

object 4 Calibration of the system and y scaling

after scans 4 No calibrating or y scaling

5 Manual input of the world coordinate information

5 No manual conversion of the world coordinate data

6 If calibration is not correct, then needs to be redone

6 Overall system is user friendly

7 Portability of the system is not optional 7 Portability is manageable

8 Difficulty registration when not symmetrical

8 Difficulty registration when not symmetrical

9 Less noise in previous scans with new setup

9 Overall system has better resolution

Chapter 6: Conclusion 49

6 CONCLUSION

Reverse engineering of geometric models and parts for CAD use is a rapidly evolving discipline in which interest is currently high. This is due in part to the recent commercial availabity of active non-contact systems that can produce some level of sufficient accuracy for many applications. In this project, we have successfully made several achievements to the reconstruction efforts of modeling our selected parts using two different systems, with two different techniques. As we revisit the data acquisition systems that were chosen for this project, we will discuss some of the benefits of using both systems. We also highlight some of our achievements in generating the 3D CAD models using both systems. Our achievements are:

• Literature survey on reverse engineering using CMM, Lasers and Structured Lighting Applications.

• Successfully generating a 3D model of the ramp using the Genex 3D FaceCam System.

• Successfully modeling the water pump using the Genex System. • Successfully modeling the top and bottom surfaces of the water pump using the

IVP Range Scanner. The IVP Range Scanner and the Genex System have different advantages to the systems that make them optimal choices for data acquisition task. The Genex System is more user friendly verses the IVP Range Scanner because there is no calibration involved in the set up process. There are however, necessary steps that must be followed to assure that the Genex Systems components are not affected when operating the system. Following the correct steps when turning the system on and off, guarantees the longevity of the equipment. Using a laser-based triangulation system and a structured lighting system has many different benefits that make both systems ideal for 3D reconstruction. A triangulation based system does not have the problem of ambient lighting affecting acquisition of data. The IVP Range Scanner has the advantage that ambient lighting is filtered better through the use of filters that can be attached to the lens. For the parts used in our experiments, the use of the filter was not required. The structured lighting system is affected by the

Chapter 6: Conclusion 50

amount of ambient lighting and natural lighting sources. The ambient lighting however was controlled through the use of additional backgrounds placed around the part. The importance of obtaining multiple single view profiles was also discussed. If multiple views of all the selected parts are acquired, the 3D CAD model can continue to be improved upon. The models will show to have less occlusion or missing data and have data that is free of noise and abnormal surfaces due to surface reflectance or object positioning. For the water pump, position was a challenge. With the use of smaller objects, different angles could be captured. Using the laser scanning technique to model parts and the structured light technique both proved to have successful and promising results. Even though the two techniques modeled the parts using different methods, the post-processing of the data was similar. After multiple scans were acquired, the water pump and the ramp had CAD models generated from the data. Both systems produced results that are less noisy and have smoother surface textures. However because of the structural setup of the IVP Range Scanner, more scans are required to produce a complete 3D model of the part. The reason is because of the flexibility of relocation and adjusting the IVP Systems components is far too tedious. The Genex System can be relocated at the discretion of the user. Also because the system is flexible in location, the ambient lighting can also be better controlled. This plays an important factor when scanning objects that have high reflective surfaces.

References 51

REFERENCES

[1] R. Bardell, V. Balendran, and K. Sivayoganathan, “Accuracy Analysis of 3D Data Collection and Free-Form Modeling Methods”, Journal of Materials Processing Technology, 133, pp. 26-33, 2003.

[2] F. Bernardini and H. Rushmeier, “The 3D Model Acquisition Pipeline”,

Computer Graphics Forum, Vol. 21, No. 2, pp. 149-172, 2002. [3] P. J. Besl and N. D. McKay, “A Method for Registration of 3D Shapes”, IEEE

Transactions on Pattern Analysis and Machine Intelligence, Vol. 14, No. 2, pp. 239–256, 1992.

[4] C. Butler, “Investigation Into the Performance of Probes on Coordinate Measuring Machines”, Industrial Metrology, Vol. 2, No. 1, pp 59-70, 1991. [5] S. Y. Chen and Y. F. Li, “Self Recalibration of a Structured Light Vision System from a Single View”, IEEE International Conference on Robotics and Automation, pp. 2539-2544, 2002. [6] Y. Chen and G. G. Medioni, “Object Modeling by Registration of Multiple Range Images”, Image and Vision Computing, Vol. 10, No. 3, pp.145–155, 1992. [7] J. Chow, T. Xu, S. M. Lee and K. Kengskool, “Development of an Integrated Laser-Based Reverse Engineering and Machining System”, International Journal of Advance Manufacturing Technology, Vol. 19, pp. 186-191, 2000. [8] J. Clark, “Implementing Non-Contact Digitization Techniques within the Mechanical Design Process”, Sensor review, Vol. 20, No. 3, pp. 195-201, 2000. [9] G. M. Cortelazzo, C. Doretto and L. Lucchese, “Freeform Textured Surfaces Registration by a Frequency Domain Technique”, Proceedings of the International Conference on Image Processing, ICIP ’98, pp. 813–817, 1998. [10] C. Dorai, J. Weng and A. K. Jain, “Optimal Registration of Object Views using Range Data”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 19, No. 10, pp. 1131–1138, 1997.

References 52

[11] C. Dorai, G. Wang, A. K. Jain and C. Mercer, “Registration and Integration of

Multiple Object Views for 3D Model Construction”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 20, No. 1, pp. 83–89, 1998.

[12] D. W. Eggert, A. W. Fitzgibbon and R. B. Fisher, “Simultaneous Registration of Multiple Range Views for use in Reverse Engineering of CAD Models, Computer Vision and Image Understanding, Vol. 69, No. 3, pp. 253– 272, 1998. [13] K. Fan and T. Tsai, “Optimal Shape Error Analysis of the Matching Image for a

Free-Form Surface”, Robotics and Computer Integrated Manufacturing, Vol. 17, pp. 215-222, 2001.

[14] R. A. Jarvis, A Laser Time-of-Flight Range Scanner for Robotic Vision, IEEE PAMI, Vol. 5, No. 5, pp 505-512, 1983. [15] K. H. Lee and H. Park, “Automated Inspection of Free-Form Shape Parts by

Laser Scanning”, Robotics and Computer Integrated Manufacturing 16, pp. 201-210, 2000.

[16] K. H. Lee, H. Woo and T. Suk, "Data Reduction Methods for Reverse Engineering", The International Journal of Advanced Manufacturing Technology, pp.735-743, 2001. [17] Y. F. Li and S. Y. Chen, “Automatic Recalibration of an Active Structured Light Vision System”, IEEE Transaction on Robotics ad Automation, Vol. 19, No. 2, pp. 259- 268, 2003. [18] H. Maas, "Robust Automatic Surface Reconstruction with Structured Light," International Archives of Photogrammetry and Remote Sensing, vol. 29, part B5, pp. 709-713, 1992. [19] S. Motavalli and B. Bidanda, “A Part Image Reconstruction System for Reverse Engineering of Design Modifications”, Journal of Manufacturing Systems, Vol. 10, No. 5, pp 383-395, 1991. [20] A. Modjarred, “Non-Contact Measurement Using a Laser Scanning Probe”, Proceedings of the SPIE, Vol. 1012, In-Process Optical Measurements, pp. 229-239, 1988. [21] I. Moring, T. Heikkinen, R. Myllyla, “Acquisition of Three- Dimensional Image Data by a Scanning Laser Range Finder”, Journal of Optical Engineering, Vol. 28, No. 8, pp 897-902, 1989.

References 53

[22] J. P. Moss, A. D. Linney, S. R. Grindrod, C. A. Mosse, “A Laser Scanning System for the Measurement of Facial Surface Morphology”, Journal of Optics and Lasers in Engineering, Vol. 10, pp 179-190, 1989. [23] M. Reed and P. Allen, “3D Modeling from Range Imagery: an Incremental

Method with Planning Component”, Image and Vision Computing, Vo. 17, pp. 99-111, 1999.

[24] G. Roth, “Registering Two Overlapping Range Images”, In Proceeding of the 2nd

International Conference on 3D Digital Imaging and Modeling, Ottawa, Canada: pp. 191–200, 1999.

[25] D. Scharstein and R. Szeliski, "High-Accuracy Stereo Depth Maps Using Structured Light," IEEE Computer Vision and Pattern Recognition, Vol. 1, pp. 195-202, 2003. [26] K. C. Sahoo, C. H. Menq, “Localization of 3-D Objects Having Complex Sculptured Surfaces Using Tactile Sensing and Surface Description”, Journal of Engineering for Industry, Vol. 113, pp 85-92, 1991. [27] M. Soucy and D. Laurendeau, “A General Surface Approach to the Integration of a Set of Range Views”, IEEE Transaction on Pattern Analysis and Machine Intelligence, Vol. 17, No. 4, pp. 1344-358, 1995. [28] W. B. Thompson, J. C. Owens and H. J. de St. Germain, “Featured-based Reverse

Engineering of Mechanical Parts”, IEEE Trans. on Robotics and Automation, Vol. 15, No. 1, pp. 57-66, 1999.

[29] G. Turk and M. Levoy, “Zippered Polygon Meshes from Range Images”, In A. Glassner (ed), Proceedings of SIGGRAPH, Computer Graphics Proceeding Annual Conference Series, pp. 311-318, 1994. [30] P. W. Will and K. S. Pennington, “Grid Coding: A Novel Technique for Image Processing”, Proceedings of the IEEE, Vol. 60, No. 6, pp 669-680, 1972. [31] J. Wang and M. Oliveria, “A Hole-Filling Strategy for Reconstruction of Smooth

Surfaces in Range Images”, Proceedings-16th-Brazilian-Symposium-on-Computer-Graphics-and-Image-Processing-SIBGRAPI-2003, pp. 11-18, 2003.

[32] Y. F. Wang and J. K. Aggarwal, “3D Object Description From Stripe Coding and

Multiple Views”, Proceedings of the 5th Scandinavian Conference on Image Analysis, Vol. 60, No. 6, pp 669-680, 1987.

References 54

[33] Y. L. Xiong, “Computer Aided Measurement of Profile Error of Complex Surfaces and Curves: Theory and Algorithm”, International Journal of Machine Tools and Manufacturing, Vol. 30, No. 3, pp 339-357, 1990. [34] D. Zhang and M. Hebert, “Harmonic Maps and Their Applications in Surface Matching”, In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR ’99), pp. 524–530, 1999. [35] Y. Zhang, “Research into the engineering application of reverse engineering

technology”, Journal of Materials Processing Technology, Vol. 139, pp. 472-475, 2003.

[36] Z. Zhang, “Iterative Point Matching for Registration of Free-Form Curves and Surfaces”, International Journal of Computer Vision, 13(2):119–152, 1994. [37] The American Heritage Dictionary of English Language, Fourth Edition, Houghton Mifflin Company, 2000.

reverse engineering of automotive parts applying laser scanning

Documents