Robotics and Computer Integrated Manufacturing 18 (2002) 53–67

A reverse engineering system for rapid manufacturing

of complex objects
L. Li, N. Schemenauer, X. Peng, Y. Zeng, P. Gu*
Department of Mechanical and Manufacturing Engineering, The University of Calgary, 2500 University Drive, Calgary, Alberta, Canada T2N 1N4
Received 1 July 2000; received in revised form 1 April 2001; accepted 1 August 2001

Abstract

This paper presents a reverse engineering system for rapid modeling and manufacturing of products with complex surfaces. The
system consists of three main components: a 3D optical digitizing system, a surface reconstruction software and a rapid prototyping
machine. The unique features of the 3D optical digitizing system include the use of white-light source, and cost-effective and quick
image acquisition. The surface reconstruction process consists of three major steps: (1) range view registration by an iterative closed-
form solution, (2) range surface integration by reconstructing an implicit function to update the volumetric grid, and (3) iso-surface
extraction by the Marching Cubes algorithm. The modeling software exports models in STL format, which are used as input to an
FDM 2000 machine to manufacture products. The examples are included to illustrate the systems and the methods. r 2002 Elsevier
Science Ltd. All rights reserved.

Keywords: Reverse engineering; Optical digitizers; Surface reconstruction; Rapid prototyping

1. Introduction example application is where CAD models are unavail-

Growing global competition requires manufacturers
to deliver more competitive products with better quality
and lower prices. One of the most important and
challenging tasks faced by manufacturing industry is
substantial reduction of product development time. The
traditional methods for making patterns and moulds
with complex surfaces are inefficient and error-prone.
For example, in medical applications such as orthotics
and prosthetics, because precise production and fitting
of such shapes are needed to satisfy medical standards
and treatment requirements, and also because of
product complexity with free-form surfaces, it is quite
often that many modifications are needed. Therefore,
reverse engineering and rapid prototyping of objects
with complex surfaces have received significant attention
from both research and industrial communities [1].
Reverse engineering is the process of engineering
backward to build a CAD model geometrically identical
to an existing product. Subsequently, CAD models are
used for manufacturing or other applications. An
*Corresponding author.
E-mail address: gu@enme.ucalgary.ca (P. Gu).
able, unusable, or insufficient for existing parts that
must be duplicated or modified. There are many
practical applications ranging from tool and die making
to biomedical device design and manufacturing.
Fig. 1 shows the reverse engineering system for rapid
manufacturing of complex objects. An object is digitized
by the 3D object digitizer and data is then fed into
surface reconstruction software, which outputs a
smooth surface model. The model is manufactured by
a rapid prototyping machine. Two of the key technol-
ogies in this reverse engineering system are the
3D optical digitizer and the surface reconstruction
algorithms.
Object digitizers can be classified into two broad
categories: contact and non-contact. Among non-
contact optical digitizers, laser triangulation is the most
commonly used technology. Drawbacks of laser trian-
gulation scanner include high cost and slow data
acquisition.
New scanning technologies are emerging that have
potential to further enhance the quality of 3D object
scans. One such development is non-contact scanning
systems employing white light rather than a single laser
tracking line. This system generates dense volumetric

0736-5845/02/$ - see front matter r 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 7 3 6 - 5 8 4 5 ( 0 1 ) 0 0 0 2 6 - 6
54 L. Li et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 53–67
Fig. 1. Schematic diagram of the reverse engineering system.
scans. Detailed scans can be generated quickly, and the
wide scanning volume reduces the importance of the
angle at which the tracking light hits the target object
[2].
Post-scanning data processing is an important step for
any successful application of these scanning systems.
The objective of post-scanning data processing is to
reconstruct surface models from 3D point clouds.
Strategies for surface reconstruction have proceeded in
two main directions: reconstruction from unorganized
points and reconstruction that exploits the underlying
structure of the acquired data. The latter is called
surface reconstruction using structured data. These two
strategies can be further divided according to whether
they operate by reconstructing parametric surfaces or by
reconstructing implicit surfaces. The algorithm em-
ployed here falls into structured data reconstruction
using implicit function methods that use samples of a
continuous function to combine structured data. Pre-
vious relevant work includes [3–6]. These methods
mainly apply to laser scanning technologies. In this
paper, the existing algorithms were adapted for building
of 3D models from point clouds generated by the 3D
optical digitizing system.
To create a single model from multiple range images,
two fundamental issues must be addressed: registration
and integration. Registration refers to computing a rigid
transformation that brings the points of one range
image into alignment with the overlapping portion of
another range image. Integration is the process of
creating a single surface obtained from the multiple
range images [6].
For a real world 3D object, no single range image
suffices to describe the entire object surface. Multiple
views must be digitized in order to obtain the complete
geometric information of the object. In order to extract
a 3D model from multi-view range images, the range
images must be aligned in one global coordinate system.
One of such methods that can achieve this is to capture
the range images using a digitizer mounted on a
precision motion device such as a coordinate measuring
machine (CMM). This method is costly and incon-
venient for many real world applications.
An alternative solution is to use a registration
algorithm to align the range images. Two main
approaches are used for the registration: (1) match
‘‘created’’ features of the images, such as the method
that was presented in [7]; or (2) minimize the distance
between all points on the overlapped surface represented
by the two images. Examples of the latter approach
include [6,8–10]. The minimal distance method was used
in this work.
After registration is completed, the final surface can
be extracted directly from range surfaces by the
following two methods: (1) geometric details of all scans
are averaged and the vertex positions of range surface
 L. Li et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 53–67 55

meshes are refined; or (2) volumetric grid points are
updated according to each range surface and iso-
surfaces are extracted from the volumetric grid. The
latter method was preferred in this research because
it allows the extraction of the iso-surface at any stage
of the registration without considering point-to-point
correspondence of two meshes that share a common
surface of the object.
The final step in the reverse engineering process is to
manufacture the object based on the geometric model
generated from the previous procedures. In this re-
search, the manufacturing process is carried out using
an FDM rapid prototyping machine.
This paper is organized as follows: the next section
introduces the 3D optical digitizing system; the third
section describes 3D data pre-processing; an algorithm
for registration of a set of range views is described in the
fourth section; the volumetric method and its imple-
mentation are presented in the following section; for
demonstration and analysis, a few experiments and
associated analysis are also provided. Conclusions are
given in the final section of the paper.
automatic fringe pattern analysis, and it has higher
accuracy when a fringe analysis algorithm is utilized
because there are no higher harmonic spatial frequencies
except for plus and minus fundamental frequency
components in the spatial frequency domain. Second,
a spatial and parallel phase-shift algorithm was used,
rather than using temporal phase-shift algorithm to
calculate the phase distribution. Such a parallel property
in phase calculation would be more preferable to the
case where the digitalization of live subjects such as
human or animal is required. Finally, a robust phase
unwrapping algorithm based on constructing a hybrid
mask was applied to this system for reconstructing the
3D range image of object surface.
The photograph of the system is shown in Fig. 2. This
optical digitizing technology does not require any

2. 3D optical digitizing system

The 3D optical digitizing system presented in this

paper is a computer aided non-contact optical system
that measures both the 3D geometric shape and the
reflectance of a given object surface. The principle of this
digitizing system is based on the digital Moire! topo-
graphic technique [11]. First, a sinusoidal fringe pattern
was used as a structured light illumination. This is
Fig. 3. Pre-processing of range data.
particularly desirable from the point of view of

Fig. 2. Photograph of the 3D optical digitizing system.

56 L. Li et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 53–67
scanning mechanisms as do conventional laser scanners.
Also, a single-view object surface can be captured in a
parallel mode. To capture a complete aðx; yÞ image of an
object, several approaches can be used. For example, the
object is mounted on a turntable. By rotating the table
with different angles, multi-view surface acquisition was
carried out. In the recording step, structured light with
sinusoidal fringes is projected onto the object surface.
Due to the depth variation of the surface the structured
light is phase-modulated, leading to a deformed spatial
carrier fringe pattern in which the topographic informa-
tion of the object surface has been encoded. Assuming
that a 2D gray-scale image of the object surface is
represented by a function, the intensity distribution of
deformed spatial carrier fringe pattern can be described
as
Iðx; yÞ ¼ aðx; yÞ þ bðx; yÞ cos½2pf0 x þ fðx; yÞ;
where aðx; yÞ provides the background intensity dis-
tribution (the texture image of the object surface) of the
fringe pattern; bðx; yÞ is a local contrast variation; f0 is a
spatial carrier frequency generated by projected grating;
and fðx; yÞ is a phase distribution to be determined.
Both geometric shape and the texture information
are recorded in Eq. (1), where aðx; yÞ represents the
texture information and fðx; yÞ involves topographic
information.
The combination of imaging optics and CCD photo
sensor array is used to capture the deformed grating that
is then digitized and stored in frame memory for later
processing. The process of information decoding for the
geometric shape of object surface extracting fðx; yÞ from
Eq. (1) is quite similar to that we reported previously
[12]. However, in this approach the laser interferometer-
based projector was replaced by a white-light (halogen
lamp) projector. The retrieved phase function can be
expressed as
 Pp  the range images. This process results in piecewise linear
Ik ðx; yÞsin dk
fðx; yÞ ¼ tan1 Ppk¼1 ; ð2Þ
k¼1 Ik ðx; yÞcos dk
where dk ¼ 2pk=p is the constant phase shift and p the
spatial phase-shifted interferograms that are obtained
digitally from only one recorded fringe pattern. This
spatial parallelism is an essential feature of the spatial
carrier techniques.
The relationship between the phase distribution and
the range image of a 3D object surface is determined by
optical sensitivity that depends on the geometry of
system design and system parameters. Such a relation
can be simply expressed as
fðx; yÞ ¼ Kzðx; yÞ;
where K represents the fringe sensitivity depending on
the optical geometry; fðx; yÞ is the unwrapped phase
distribution; zðx; yÞ represents the height variation of
object surface, that is, a 3D image or a range image.
Whenever the system configuration is fixed, K becomes
a constant. The fringe sensitivity K can be determined
under the first order approximation and is given by
2pL cos y
K¼ ; ð4Þ
P0 L0
where P0 ¼ ðL0 PÞ=ðdcos yÞ stands for the spatial
period of grating projected onto the reference plane; d is
the distance between the exit pupil of projection optics
and the center of the grating; P represents the spatial
period of projection fringes; y depicts the angle between
the imaging-optics axis and the projection-optics axis;
and L0 delineates the distance from entrance pupil of
both imaging and projection optics to the reference
plane. Therefore, the phase distribution of an object
surface can be linearly mapped to a range image, i.e., a
3D range image of a test surface being reconstructed.
ð1Þ
The concavity or convexity of the surface is deter-
mined directly from the unwrapped phase function. The
data acquisition, phase function calculation and phase
unwrapping have been conveniently automated and
computerized, and the whole process is carried out by a
software package dedicated to this digitizing system.
When geometric parameters of an optical configuration
are determined, the fringe sensitivity is determined
accordingly. Thus, the unwrapped phase function, f 
ðx; yÞ; is proportional to the surface shape function z 
ðx; yÞ: An example given in Section 8 illustrates the
process of the data acquisition of the 3D optical
digitizing system.
3. 3D data pre-processing
The pre-processing of range images begins by
converting the range images into range surfaces by
connecting the nearest neighbor points with triangles of
range surfaces, as shown in Fig. 3. Each point in the
range image is a potential vertex in the triangle mesh. To
remove background data from the data set, a filter-value
(threshold) is set and data from the object being scanned
are included in the range surfaces. Also, triangles with
an edge longer than the pre-set distance threshold are
discarded, thus to avoid inclusion of questionable or
noise data. The object shown in Fig. 3 was processed
and the result is shown in Fig. 4.
After the piecewise linear range surfaces are created,
the normals of the triangles on the range surfaces are
calculated. By averaging the normals of triangles that
are incident to a vertex, normals to vertices on the range
ð3Þ
surface are assigned.
Next, a weight is defined for each vertex on the range
surface. Weights are intended to represent variations in
certainty across the range surface. The definition of
weights should be specific to the range scanning
 L. Li et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 53–67
Fig. 4. Initial triangle mesh of an object.
technology. For this scanner, the position of the sensor
is fixed for every range view. For a sample vertex on the
range surface, the weight is defined as
nij ~
~ ns
wij ¼ ; ð5Þ
ds
where nij is the unit normal of surface at the point, as
shown in Fig. 4; ns is the unit normal from the vertex to
the sensor; and ds is the distance between the vertex and
sensor. The dot product between each vertex normal
and the unit vector from the vertex to sensor reflects the
influence of viewing direction on certainty. The smaller
the angle between these two vectors the higher the cer-
tainty. ds reflects the influence of distance on certainty;
the bigger the distance the lower the certainty.
Range data at the boundaries typically have greater
uncertainty and therefore have less weight. In practice, it
can be determined whether a vertex is on boundary by
counting the number of triangles that are incident to the
vertex. The weight of a point other than the vertices of
a triangle is calculated by linearly interpolating the
weights of the intersecting triangles’ vertices. With the
above information, the range images can be registered
with one another and integrated into a single ‘‘iso-
surface’’.
4. Registration algorithm
After a range image is converted to a triangle mesh
with a weight at each vertex, the next step is to align the
range surfaces from different views for integration.
57
Horn [9] proves a closed-form solution for finding the
rigid transformation that minimizes the least-squared
distance between point pairs. Besl and McKay [8] used a
method that they called as the ‘‘iterated closest-point
algorithm’’. This method finds the nearest points on one
surface to a collection of points on another surface and
then transforms one surface so as to minimize the
collective distance through an iterative process. This
method cannot be used directly to register range images
because it requires that every point on one surface have
a corresponding point on the other surface. Turk and
Levoy [6] adapt the method to register the range images
from a laser scanner.
Based on these algorithms and methods, an algorithm
was devised to register range images from the scanning
system. The algorithm consists of two steps: (1) find the
overlapping portion on two range surfaces; and (2)
transform one range surface to align with the other one.
For example, with two different range views meshes A
and B are formed, as shown in Fig. 5. To find the
overlapping portion of two range surfaces, it is first to
find the nearest point of mesh A to each vertex of mesh
B: This is done by casting a perpendicular ray from a
point Bi in mesh B to triangles on mesh A; until the
intersection is found, which is located inside a triangle of
mesh A; then Bi and the intersection Ai ; constitute a
pair; otherwise skip the triangle. The number of point
pairs is denoted by Np : By defining a distance threshold,
it defines the distance between the vertex of mesh B and
each vertex of a triangle of mesh A: If any of them is
greater than the distance threshold, the triangle is
skipped. This operation discards the pairs of points
58 L. Li et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 53–67

Fig. 5. Point correspondence for registration.

that are too far apart. Finally, the overlapping portion is and Mckay [8]. The calculation is carried out by the
represented by the point pairs. The triangles in one mesh following steps. P
are stored in a BSP tree to speed up this operation. First the 3  3 cross-covariance matrix, , is calcu-
Once all point pairs are found, the weighed least- lated as
square distance between pairs of points can be defined as N
X 1 X p
follows: ¼ wi ½ðBi  Bc ÞðAi  Ac ÞT 
Np Np i¼1
X 2 3
E¼ wi jjAi  RðBi  Bc Þ  T  Bc jj2 ; ð6Þ Sxx Sxy Sxz
i¼1 6 7
¼ 4 Syx Syy Syz 5: ð9Þ
where Bi and Ai are pairs of 3D points and Bc is the
Szx Szy Szz
centroid of Bi : The centroid is defined by
N N From the matrix in the above equation, the following
1 X p X p

Bc ¼ Bi wBi = wBi : ð7Þ matrix is constructed:

Np i¼1 i¼1

2 3
Sxx þ Syy þ Szz Syz  Szy Szx  Sxz Sxy  Syx
6 Syz  Szy Sxx  Syy  Szz Sxy þ Syx Szx þ Sxz 7
6 7
Q¼6 7: ð10Þ
4 Szx  Sxz Sxy þ Syx Sxx þ Syy  Szz Syz  Szy 5
Sxy  Syx Szx þ Sxz Syz þ Szy Sxx  Syy  Szz

Centroid Ac is calculated similarly. wi is the product Unit eigenvector P ¼ ½q0 q1 q2 q3 T corresponding to

of the weight of Bi ; wBi ; i ¼ ð1; y; nÞ; and the weight of the maximum eigenvalue of matrix Q is selected as the
Ai ; wAi : T is the translation vector and R is the rotation optimal rotation.
matrix. According to Horn [9], T is just the difference Finally, the rotation matrix is constructed as
between the centroid of points Ai and the centroid of 2 3
q20 þ q21  q22  q23 2ðq1 q2  q0 q3 Þ 2ðq1 q3 þ q0 q2 Þ
points Bi ; thus it is calculated as 6 7
R ¼ 4 2ðq1 q2 þ q0 q3 Þ q20  q21 þ q22  q23 2ðq2 q3  q0 q1 Þ 5:
T ¼ Ac  Bc : ð8Þ 2ðq1 q3  q0 q2 Þ 2ðq2 q3 þ q0 q1 Þ q20  q21  q22 þ q23
R is found by constructing a cross-covariance matrix ð11Þ
between centroid-adjusted pairs of points. The final
rotation is determined by a unit quaternion that is the The above procedure is iterated until the convergence
eigenvector corresponding to the largest eigenvalue of a is achieved. Then the second range surface and its sensor
matrix constructed from the elements of this cross- position are transformed to align with the first one.
covariance matrix. The details of the mathematical In this present system, there is a motion platform that
proofs can be found in the references Horn [9] and Besl provides an approximate value for transformation of the
 L. Li et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 53–67
Fig. 6. Two sphere patches after registration.
Fig. 7. Surface patches from a vase before and after registration.
model around an axis. These data were used to conduct
an initial registration of the range surfaces. Next, the
distances between all points on the overlapped surfaces
represented by the two images are minimized using the
algorithm described above. The combination of hard-
ware and software registration worked effectively for the
experiments. For example, Fig. 6 shows the sphere
patches after registration. Fig. 7 shows surfaces of a
vase before and after registration.
5. Volumetric method for surface integration
After the range surfaces have been aligned using
registration algorithm, the volumetric method proposed
by Curless and Levoy [3] was employed to integrate the
range surfaces into a single ‘‘iso-surface’’. Modifications
of the original volumetric method were made to apply it
to this system.
5.1. Concepts of volumetric method
Volumetric integration employs a continuous implicit
function, DðxÞ; represented by samples. The function is
the weighed signed distance of each point x to the
nearest range surface along the line of sight to the
sensor. The rules to compute the implicit function are
59
expressed as an incremental calculation [3]:
Wi ðxÞDi ðxÞ þ wiþ1 ðxÞdiþ1 ðxÞ
Diþ1 ðxÞ ¼ ; ð12Þ
Wi ðxÞ þ wiþ1 ðxÞ
Wiþ1 ðxÞ ¼ Wi ðxÞ þ wiþ1 ðxÞ; ð13Þ
where Di ðxÞ and Wiþ1 ðxÞ are the cumulative signed
distance and weight functions after the integration of the
ith range image; diþ1 and wiþ1 are signed distance and
weight function of i þ 1th range image. The signed
distance contribution is computed by casting a ray from
the sensor through each voxel near the range surface and
then intersecting it with the triangle mesh, as shown in
Fig. 8. The weight is computed by linearly interpolating
the weights stored at the intersection triangle’s vertices.
Having determined the signed distance and weight
Eqs. (12) and (13) are applied to each voxel.
5.2. Implementation of volumetric method
For each range surface the volumetric grid is updated.
For each voxel of the volumetric grid, a line can be
defined by connecting the sensor position and the voxel.
Intersections between this line and all triangles in the
range surface must be found in order to calculate DðxÞ
and WðxÞ of the voxel. A simple implementation would
iterate through all of the triangles in the range surface to
find the intersections. However, this is computationally
expensive and costly as the number of range surfaces
and the grid resolution increase. When updating the grid
for a surface the time complexity of this algorithm is
OðndmÞ3 where n is the number of triangles in a range
surface and m is the number of grid points along one
axis (assuming a cubic grid). A better voxel update
algorithm is required to make this integration method
practical.
The improved voxel update algorithm proposed in
this research is outlined in Fig. 9 and later in Fig. 10. A
voxel plane is selected to be perpendicular to axes X, Y
or Z according to the sensor position of each range
surface. For each voxel plane, triangles are projected
onto that plane. Only voxels inside the project triangles
need to be updated. As the plane is orthogonal, these
voxels are easily found. For typical situations, this
algorithm has a time complexity of OðndmÞ: For high
resolutions this is a significant improvement over the
simple algorithm mentioned above.
As shown in Fig. 11, although the distance and the
weight functions should extend indefinitely in either
direction, to prevent surfaces on opposite sides of the
object from interfering with each other, the weight
function is forced to taper-off behind the surface. As
mentioned in Ref. [3], there is a trade-off involved in
choosing where the weight function tapers off. It should
extend far enough behind the surface to ensure that all
distance ramps will contribute in the vicinity of the final
60 L. Li et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 53–67

Fig. 8. Updating of a voxel.

zero crossing. But it should also be as narrow as possible
to avoid influencing surfaces on the other side. Similarly,
the signed distance and weight functions should not be
extended too far in front of the surface. Restricting the
functions to the vicinity of the surface yields a more
compact representation and reduces the computational
expense of updating the volume. When employing the
distance cut-off idea for updating the volumetric grid, it
is found that the distance is dependant on the resolution
of the range image.
5.3. Iso-surface extraction using marching cubes
algorithm
To extract the surface from the volumetric grid, a
Marching Cubes algorithm [13] with a lookup-table to
resolve ambiguous cases [14] was used. There are two
main steps in the Marching Cubes algorithm. The first
step is to decide how to define the section or sections of
surface that divide up an individual cube. The second
step involves forming the correct facets from the
positions where the iso-surface intersects the edges of
the grid cell.
Each corner of a unit cube is classified into either
being below or above the iso-value. This gives 256
possible configurations of corner classifications to a unit
cube. Two of these are trivial. When all the values of the
8 corners are above or below the iso-value, it does not
contribute to the iso-surface. For all other configura-
tions it is necessary to determine where, along each cube
edge, the iso-surface crosses, and use these edge
intersection points to create one or more triangular
L. Li et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 53–67 61

Fig. 9. Locating voxels for updating.

Fig. 10. Improved algorithm to update voxels.

62 L. Li et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 53–67

Fig. 11. Surface boundary when updating voxels.

Fig. 12. Marching cubes algorithm.

patches for the iso-surface, and facets from adjacent grid
cells connect together correctly.
Two tables are used to look-up, which triangles need
to be created for each cube. First, an 8-bit cube index is
calculated based on the vertex values. This cube index is
used as a key for the edge table to find which edges
intersect the iso-surface. The edge table returns a 12-bit
number with each bit corresponds to one edge and a 1
indicating the edge has an intersection. The intersection
point on each intersected edge is calculated by linearly
interpolating the vertex values. These intersection points
and the triangle table are used to create facets for the
iso-surface inside the cube. The Marching Cubes
algorithm is shown in Fig. 12. Surface patches can be
created for a single voxel. This process can be applied to
the entire volume. Then all these triangle patches
constitute an iso-surface.
Fig. 13 shows the 2D grid to illustrate that some of
the boundary grids would be updated whereas others
would not. Thus, the surface extraction procedure is
restricted to skip boundary voxels. This is reasonable
because range surfaces come from multi-views and when
they have sufficient areas of overlapping with the cube
size being small enough.
6. Rapid prototyping of models
After the 3D surface representation of an object is
generated, the next step is to create a model using rapid
prototyping. This process comprises two steps: (1)
convert the model into STL format (the standard model
format for rapid prototyping machines); and (2) slice the
model for rapid prototyping operations. This system has
 L. Li et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 53–67
Fig. 13. Illustration of boundary voxels skipping.
Fig. 14. A part of an STL file.
a module, which converts the 3D surface models into
STL files. The STL file consists of normals and vertices
coordinates of every triangle on the surface. Fig. 14
shows part of such a file.
A Strategy’s FDM 2000 was used for manufacturing
parts. The FDM 2000 uses the Fused Deposition
Modeling technology to turn computer-aided design
(CAD) models that can be used for design reviews,
manufacturability studies, investment casting patterns
and many other applications [15]. The STL file is then
imported to the QuickSlice software included with the
machine to generate tool paths in SML(Stratasys
Modeling Language). The SML contains the FDM
machine instructions required to create the part.
7. Examples and analysis
In this section, three examples are presented to
illustrate the above systems and methods. The first
example shows the reverse engineering process using the
63
3D optical digitizing system, data processing techniques
and rapid prototyping processes. A person’s face was
digitized using the 3D optical digitizing system. The
deformed fringe pattern modulated by the height
variation of the face is illustrated in Fig. 15(a). The
reflectance (texture image) of the surface extracted is
shown in Fig. 15(b), and Fig. 15(c) is the shape
reconstruction of a human face from fringe pattern.
Fig. 15(d) shows 3D point cloud that is the output of the
digitizing system. The 3D point cloud was further
processed into a range surface model as shown in
Fig. 15(e). After surface model was exported in STL
format, it was then input into Stratasys QuickSlice, and
sliced. The sliced model is shown in Fig. 15(f) with
supports and base automatically formed by the software
QuickSlice. In this example, only one view of the object
was taken to represent the front view of the face and
therefore registration and integration of range surfaces
from different views were not required. The prototype
model was made and is shown in Fig. 15(g).
Second example shows the difference between an
initial range surface and the extracted iso-surface using
the volumetric implicit function and the Marching
Cubes Algorithm. Shown in Fig. 16 are: (a) the input
range surfaces (b) the iso-surface extracted, and (c) the
difference distribution. The difference was calculated by
comparing the z values of the vertices on the original
range surface with the z values of the correspondence
points on the iso-surfaces. The difference values were
multiplied by a factor of two in order to improve the
image. Some of the characteristics of the extraction
64 L. Li et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 53–67

Fig. 15. The process of the reverse engineering system.

method are summarized: (1) There was no vertex or
triangle correspondence between the two surfaces. The
vertices of the iso-surface were interpolation of original
data according to volumetric implicit function. The
density of the triangles on the iso-surface varied
according to the slope of the surface at the given unit
cube size. These features were very useful when
extracting 3D surface models from range surfaces of
different range views. (2) Details at the surface edge
might be lost after iso-surface extraction. For multiple
range surfaces, overlaps between range surfaces would
help to eliminate this effect. (3) The maximum relative
error was 3.6%, which appeared around the boundary
shown in Fig. 16(c). Bigger overlaps could help to
compensate the error of extraction.
The last example shows the process of building a
model from the range surfaces with different views via
registration and integration. The object vase was first
digitized using the 3D optical digitizing system in five
different views. The basic mechanical registration
 L. Li et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 53–67
Fig. 16. Comparison of initial range surface and iso-surface. (Contours appearing on the iso-surface correspond to small triangles created when the
iso-surface clips the edges and corners of voxel cubes.)
indicated that the rotational angles between the first
four views were 901 about x-axis; and the rotation
between the first and the fifth views was about 1001
around y-axis. This information was used as an input to
the integration software, shown in Figs. 17(a) and (b).
Fig. 17(c) shows the result of the registration. It is clear
from the figure that the range surfaces of different views
are in alignment with each other. Then the range
surfaces were cited to update the volumetric grid, from
which the final iso-surface model of vase was extracted,
as is shown in Fig. 17(d). This model contains some
holes around the shoulder of the vase because the range
surfaces did not overlap sufficiently in all locations. If
more range surfaces are used, the holes would be
eliminated. For example, the integrated surface shown
in Fig. 17(e) is more complete because 25 range surfaces
were integrated rather than five range surfaces of
Fig. 17(d).
65
8. Concluding remarks
This paper presented a reverse engineering system for
rapid modeling and manufacturing of objects with
complex surfaces. The system consists of three major
parts: the 3D optical digitizing system, the surface
reconstruction software and an FDM rapid prototyping
machine. The optical digitizer utilized a white-light
source for image acquisition that makes this technology
cost-effective, fast in image acquisition and portable for
various applications. The specially designed software
system processed the image data cloud acquired by the
digitizer which includes the 3D data pre-processing by
creating triangle meshes from range images. The vertices
of newly created range surfaces were also associated
with weights. The range surfaces were then registered by
minimizing the weighed least-square distance between
the points on the overlapping portion of two surfaces.
66 L. Li et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 53–67

Fig. 17. Surface reconstruction of multiple range surfaces of a vase.

 L. Li et al. / Robotics and Computer Integrated Manufacturing 18 (2002) 53–67
The volumetric method aimed to find the volume that
contained all the range surfaces, and divide the volume
into voxel grid according to the size of unit cubes. Then
the iso-surfaces were extracted from the volumetric grid,
which were a collection of points inside the volumes to
represent the iso-surface. The STL was created from the
iso-surface models and exported to an FDM 2000
machine, which produced prototypes. Currently, the
entire system is being tested for various applications.
Acknowledgements
The authors wish to acknowledge the Natural
Sciences and Engineering Research Council of Canada
(NSERC) for providing the support to this research
through a Strategic Grant STR192769.
References
[1] Yan X, Gu P. A review of rapid prototyping technologies and
systems. Comput Aided Des 1996;28(4):307–18.
[2] Aronson RB. Forward thinkers take to reverse engineering.
Manufacturing Engineering Dearborn 1996;117:34 –5.
[3] Curless B, Levoy M. A Volumetric method for building complex
models from range images. Proceedings of SIGGRAPH 1996,
New Orleans, LA, 5–9 August. New York: ACM Press, 1996.
p. 303–312.
67
[4] Grosso E, Sanini G, Frigato C. Extraction of 3D information and
volumetric uncertainty from multiple stereo images. Proceedings
of the Eighth European Conference on Artificial Intelligence,
August, 1988. p. 683–688.
[5] Hilton A, Toddart AJ, Illingwroth J, Windeatt T. Reliable surface
reconstruction from multiple range images. Proceedings of the
Forth European Conference on Computer Vision, vol. I, April
1996. p. 117–126.
[6] Turk G, Levoy M. Zippered Polygon Meshes from Range Images.
Proceedings of SIGGRAPH’94, Orlando, FL, 24–29 July.
New York: ACM Press, 1994. p. 311–318.
[7] Nobuhio W, Toriyama H, Tanaka HT, Kishio F. Reconstruction
of an Object Shape from Multiple Incomplete Range Data Sets
Using Convex Hulls. Computer Graphics International’93,
Lausanne, Switzerland, June 21–25, 1993. p. 193–203.
[8] Besl PJ, Mckay ND. A Method for registration of 3-D shapes.
IEEE Trans Pattern Anal Machine Intelligence 1992;14(2):
239–56.
[9] Horn BKP. Closed-form solution of absolute orientation using
unit quaternion. Opt Soc Am 1987;4(4):629–42.
[10] Huang X, Gu P. CAD-model based inspection of sculptured
surfaces with Datums. Int J Prod Res 1998;36(5):1351–76.
[11] Peng X, Zhu SM, Su CJ, Mithchell Tseng. Model-based digital
Moir!e topography. Optik, 1996;110(4):184–90.
[12] Peng X, Gao Z, Zhu SM. Surface contouring by a new type of
digital Moir!e technique. Optik 1995;100(2):63–7.
[13] Lorensen WE, Cline HE. Marching cubes: a high resolution
3D surface construction algorithm. Comput Graphics (SIG-
GRAPH’87 Proc) 1987;21(7):163–9.
[14] Montani C, Scateni R, Scopigno R. A modified look-up table for
implicit disambiguation of marching cubes. Visual Comput
1994;10(6):353–5.
[15] Stratasys. FDM 2000 manual. Release 2.0, 1998.