a multi-sensor approach to automating co-ordinate measuring machine-based reverse engineering

A multi-sensor approach to automating co-ordinate measuringmachine-based reverse engineering

V.H. Chana,*, C. Bradleyb, G.W. Vickersb

aDepartment of Mechanical Engineering, Ryerson Polytechnic University, 350 Victoria Street, Toronto, Ont., Canada M5B 2K3bDepartment of Mechanical Engineering, University of Victoria, Victoria, BC, Canada V8W 3P6

Received 6 December 1999; received in revised form 26 June 2000; accepted 1 December 2000

Abstract

Co-ordinate measuring machines (CMMs) have been widely used in industry for reverse engineering. In an ideal reverse

engineering system, the object would be automatically digitised, the data would be segmented into constituent surface patches,

and an accurate solid model would be generated. However, at present, the time-consuming manual digitisation of many

objects discourages the implementation of reverse engineering. As a major step toward attaining this goal, a charged coupled

device (CCD) camera and a CMM touch probe digitiser are used together in a novel reverse engineering process. Neural

network based stereo image processing is used to locate the object in the CMM work space, and to generate the CMM touch

probe path. Precise data is then collected using the CMM touch probe, from which a CAD/CAM model can be constructed.

# 2001 Elsevier Science B.V. All rights reserved.

Keywords: Machine vision; Reverse engineering; CMM programming; CAD/CAM; Neural networks

1. Introduction

The process of creating a geometric CAD solid

model from an existing physical part is often referred

to as geometric reverse engineering. Using 3-D data

collected by a tactile probe that is mounted on a

machine tool or a co-ordinate measuring machine

(CMM), a CAD model can be created and employed

in many subsequent manufacturing processes. Reverse

engineering applications can range from the creation

of drawings for an aircraft turbine engine air intake [1]

to the digitisation of a telephone handset for the

creation of a plastic injection mold [2]. An in-depth

review of reverse engineering is discussed in a paper

by Varady et al. [3].

Reverse engineering methods currently require that

the CMM operator, through the use of a pendant, to

manually guide a digitising sensor to gather hundreds,

or possibly thousands, of data points on the object to

accurately de®ne surface contours. This is a slow

process, requiring the use of an expensive piece of

equipment for what can be from hours to days.

Although research has been done in the past on

directing a CMM from CAD data [4] and using a

machine vision data to calculate the orientation of a

part on the CMM [5,6]. Research in directing a CMM

without prior knowledge of the part shape has not been

found. In this paper, a solution to this problem is

presented that utilises stereo vision to generate the

necessary spatial information to plan and control the

Computers in Industry 44 (2001) 105±115

* Corresponding author. Tel.: �1-416-979-5303;

fax: �1-416-979-5265.

E-mail address: [email protected] (V.H. Chan).

0166-3615/01/$ ± see front matter # 2001 Elsevier Science B.V. All rights reserved.

PII: S 0 1 6 6 - 3 6 1 5 ( 0 0 ) 0 0 0 8 7 - 7

movement of a tactile touch probe mounted on a

computer numerically controlled CMM.

1.1. Proposed digitisation methodology

In this work, two sensors are used to digitise an object

tobereverseengineered.First, achargedcoupled device

(CCD) camera is used as a low-level 3-D sensor, to

determine the spatial location of a part on the CMM

bed and to recognise individual surface patches on the

object's surface. Precise 3-D information is then

collected using a CMM touch probe. By using both

the CCD camera and the CMM touch probe together,

the following two main advantages are realised.

� The tool path for the CMM touch probe can be

programmed off-line, saving valuable machine time

on the CMM, and providing a more efficient use of

the CMM system.

� Different surfaces and types can be pre-determined

before digitisation by the touch probe, thus, allow-

ing the use of specialised touch probe routines.

A pair of images must be taken for each side of the

object for which surface data is required. Typically,

stereo images from the top, front, back, left and right

sides are taken with the CCD camera. Surface patch

location is achieved by matching the segmented

patches between the stereo image pairs. A Kohonen

neural network base segmentation algorithm is used to

segment the CCD images. Fig. 1 graphically illustrates

the two sensor reverse engineering process.

1.2. System description

A Renishaw touch probe system is mounted on the

z-axis arm of a gantry style CMM end effector for the

collection of 3-D data. The system components are

outlined in Table 1.

A silicon graphics indy (SGI) work station is used

for video image capturing and processing. A personal

computer connected to the CMM controller to allow

for programmed armature movements of the CMM. A

photograph of the CMM touch probe and the mounted

CCD camera is shown in Fig. 2 and Fig. 3 illustrates

the interconnection of the major components in the

system.

2. Applying stereo vision for object localisation

The CMM is employed to provide an accurate and

repeatable platform from which the stereo image pair

can be gathered. This allows the CCD camera to be in

Fig. 1. Two sensor reverse engineering process.

Table 1

Equipment speci®cations

Equipment Specifications

Mitutoyo BHN710 CMM Working volume: 700 mm� 1000 mm� 600 mm, encoder resolution: 0.5 mm, controller: CMMC 35

Renishaw touch trigger probe PH8 probe head, TP2 touch probe maximum repeatability at stylus 0.35 mm

Black and white CCD camera NEC model TI-324A, 380,000 pixels, Computer 8.5 mm f/1.3 close focus lens

PC 386 based PC, linked to CMM controller

Work station Silicon graphics indy, VINO video capture board

106 V.H. Chan et al. / Computers in Industry 44 (2001) 105±115

any of the standard orthographic views (top, front,

back, left and right elevations). Object illumination is

provided by regular ambient room lighting; no spe-

cialised lighting equipment is necessary. A similar

strategy to Marapane et. al. [7] employing region

correlation is used to achieve stereo correspondence.

In this work, however, the stereo pair of CCD images

are ®rst segmented into surface patches using a neural

network based algorithm. The patches are then

labelled with parameters that denote their speci®c

properties. Corresponding patches in each stereo pair

are matched in order to calculate the surface area and

location of each patch.

2.1. Neural network image segmentation

The potential variability amongst object shape, size

and location dictates the use of a robust segmentation

method. Recent research suggests the use of a neural

network based segmentation algorithm. Many of these

vision algorithms, such as those that were developed

for robot guidance [8,9], or object location [10] used

neural networks to match previously learned patterns.

This is a situation that usually does not exist in reverse

engineering. However, a few works have used neural

networks on images of non-constrained boundaries,

such as often found in the medical ®elds. Worth and

Kennedy [11] employed a four-layer neural network to

segment between grey and white matter from a brain

scan image. A fuzzy Hop®eld neural network is used

by Lin et al. [12] to classify tumours, lesions and

abnormalities in medical images. Ahmed and Farag

[13] used a self-organising feature map to automati-

cally label structures in the human brain from mag-

netic resonance images. A similar approach by Koh

et al. [14] used a multilayer self-organising feature

map to successfully segment range images using

varying degrees of higher resolution segmentation.

The network used in this research is based on the

Kohonen self organising map (SOM) network as

described in Rao and Rao [15]. Unlike the medical

applications used above, each layer on the SOM will

be used to represent different areas on the object and

not to segment by feature type.Layerone isused to input

one original image, whereas the layers above each

represent a possible output (winning patch). In this

`̀ winners-take-all'' strategy, the number of patches

(information not generally known in reverse engineer-

ing without user input) need not be known beforehand.

A competitive learning network, the SOM, consists

of n layers of two-dimensional arrays of neurons. Each

neuron is connected to its immediate neighbour on

its own layer and to (nÿ 1) neurons on the layers

Fig. 2. Picture of CMM touch probe and CCD camera.

Fig. 3. Schematic of experimental equipment.

V.H. Chan et al. / Computers in Industry 44 (2001) 105±115 107

immediately below and above its location, as shown in

Fig. 4. The input neurons �x; y; 0� includes the CCD

image on the bottom layer. The (nÿ 1) neurons

located above each input neuron are locked in a

competition to be the winning neuron for the input.

Each winning neuron excites (strengthens) connectors

in a neighbourhood on its own layer but inhibits the

neurons on other layers from being declared winners

for that speci®c input neuron. Once the network is

initialised, the learning of the network is self-organis-

ing. Iteration is complete when the output converges,

i.e. no new or different winning neurons are declared.

For each image pixel �x; y; 0� location, either the

strongest neuron for that pixel location �x; y; k� is

declared the winner or no winner is declared (where

k represents the patch layer). The maximum number of

patches that can be found is limited to the number of

layers on which the neural network is built. Ten layers

are used in this work, one layer for the input image and

the remaining nine layers for the output. Each neuron

has 24 excitatory connectors arranged in a square

neighbourhood around itself (see Fig. 4). Excitatory

connectors gain strength only if they are attached to a

winning neuron. The rate at which the excitatory

connectors are strengthened, dx/dt, is given below

(Eq. (1)):

dx�x;y�;�i;j�dt

��Ax�x;y�;�i;j��Bsgrey�Cymomentum�Zi; j (1)

After each iteration dt, the strength of the connector

between central pixel �x; y� and neighbouring pixel

�i; j� is increased by dx. Constant A in the ®rst term of

Eq. (1) represents the learning rate. The second term

increases the strength of the connectors in areas where

input neurons (image pixels) are of similar intensity.

The value of s is dependent on the difference in grey

level intensity of neighbouring pixels. The last term

provides additional momentum for the growth of large

patches (bigger patches have higher y values and, thus,

can eliminate smaller patches). There are (nÿ 2)

inhibitory connections for each pixel location �x; y�.Initially, the inhibitory connectors are all set equal to

one. The rate at which the inhibitory connectors

change is (Eq. (2))

dg�x;y;k�dg

� �Dg�x;y;k�; for Z�x;y;k� � 0;

dg�x;y;k�dg

� ÿDg�x;y;k�; for Z�x;y;k� � 0 (2)

The inhibitory connectors in Eq. (2) change by a value

of dg for each iteration of dt. The constant D represents

the learning rate and g is the strength of the inhibitory

connector at (t ÿ 1).

The strength of the connectors determines which

neuron �x; y; k� will be the winner for that particular

pixel �x; y� location, as expressed in Eq. (3):

Zvalue�x; y; k�

�X5

m�1

X5

n�1

�x�x;y�;�xÿ2�m;yÿ2�m� ÿ g�x;y;k�Z�x;y;k�� (3)

Employing the `̀ winners-take-all'' strategy, the neu-

ron at position �x; y; k� for k � 1 to 9, with the max-

imum value that is larger than other values for that

location, is declared the winner (1) while all other

neurons on the other layers at pixel location �x; y; k�are losers (0). The connectors are iterated until the

values converge (i.e. no more winning neurons are

declared).

To determine the appropriate values for constants A,

B, C and D, each was ®rst set to unity. Starting at

constant A, different values were tested and the effects

observed. Though examination of various images, the

values for A � 1:0, B � 2:0, C � 10 and D � 0:1gave the best results.

In summary, each iteration of the algorithm consist

of three steps. First, the excitatory connectors,

Fig. 4. Kohonen neural network con®guration.


consisting of the links between a pixel and its closest

24 neighbours, are updated. Second, the inhibitory

connectors, linking each pixel at location �x; y� to its

counterpart in the layers above and below, are recal-

culated. Finally, using the new values for the excita-

tory and inhibitory connectors, the strength of each

neuron is calculated. A ¯owchart on the neural net-

work algorithm is presented in Fig. 5.

2.2. Initialisation of the network

Rapid growth of patches is facilitated by selecting

initial points for the neuron layer in areas where the

grey levels are constant, allowing the initial points to

strengthen their connectors quicker. A raster scanning

method, similar to that described by Wahl [16] is

utilised in this work to ®nd patches of comparable

grey value. The initialisation algorithm starts by scan-

ning through the image, row by row, assigning a label

every time a pixel above a background threshold is

encountered. However, before a new label is assigned,

surrounding pixels are checked for similar grey levels

(i.e. within a certain set range). If they are similar, then

the same label is assigned. A histogram of the labels is

kept so that the top nine labels are assigned layers on

the neural network. The position of the initial point is

determined by calculating the centroid of its labelled

patch.

2.3. Edge enhancement of the stereo images

The neural network segmentation program is depen-

dent on a grey image gradient to de®ne the boundaries

of a surface patch. However, for curved surfaces that

are smoothly blended into plane surfaces, an unclear

edge results. In previous work, Chan et al. [17], found

that the neural network patches tended to pass over

these blended boundaries if an edge was not clearly

visible in the stereo image. To address this problem,

image enhancement routines are applied to the stereo

images to better de®ne partially exposed edges.

Edges in the images are found by running an edge-

linking algorithm on existing edges that are enhanced

by the use of several common image processing

routines. A 3� 3 linear smoothing mask is ®rst used

to reduce the effects of noise in the image before the

application of a 5� 5 Laplacian of Gaussian mask to

highlight the edges. To thin the resulting thick edge

lines, a non-maxima suppression algorithm as

described by Canny [18] is used. The effect of these

®rst three image processing algorithms on an object

can be seen in Fig. 6a±c. An edge following routine is

used to link pixels into lines; those lines consisting of

more than ®ve pixels are extended to the image

boundaries, as shown in Fig. 6d. The enhanced edges

are then subtracted from the original stereo images.

The resulting image is shown in Fig. 6e. These

enhanced edges provide enough of a barrier to prevent

the neural network patches from bleeding into other

surfaces.

3. Extraction of the surface location

The Kohonen neural network uses a `̀ winners-take-

all'' strategy. Therefore, the number of layers remain-

ing with winning neurons represents the actual num-

ber of distinct surface patches, as can be seen by the

stereo images. Depth information is extracted by

matching parameterised patches for the left and right

images. Each surface patch is described by ®ve para-

meters: (i) mean grey (Eq. (4)), (ii) number of pixels

(Eq. (5)), (iii) width of the neuron (Eq. (6)), (iv) length

Fig. 5. Neural network segmentation iteration cycle.


(Eq. (7)) and (v) co-ordinates of the centroid (Eq. (8)).

Gmean;k �Pm

x�0

Pny�0Z�x;y;k�Z�x;y;0�Psum;k

; for layer k (4)

Psum;k �Xm

x�0

Xn

y�0

Z�x;y;k�; for layer k (5)

width�max�x�ÿmin�x� of Z�x;y;k� for layer k (6)

length�max�y�ÿmax�y� of Z�x;y;k� for layer k (7)

centroid i �Xm

x�0

Xn

y�0

Z�x;y;k�xPsum;k

; for layer k;

centroid j �Xm

x�0

Xn

y�0

Z�x;y;k�yPsum;k

; for layer k (8)

It was found that a maximum of 10% error allowed

between the matched parameters allowed for correct

results from the stereo image pairs. Displacement of

the CCD camera along the cartesian axis of the CMM

ensures that the co-ordinates of the centroids will lie

on the epipolar line on both images.

The distance from the camera lens to the surface of

the matched pairs is calculated as shown in Fig. 7. The

co-ordinates of the camera is readily available from

the CMM controller.

Using the centroid of the patches as the base dis-

placement between the image pairs, the distance of the

patch from the camera can be calculated as

z � bf

x0l ÿ x0r(9)

4. Location of concavities on the patch surface

One important use for a CMM is to locate con-

cavities on an object being reverse engineered. Often,

Fig. 6. Image edge enhancements.

Fig. 7. Calculation of stereo depth.


concavities are used as bearing surfaces or for locating

pins. In this work, holes are found by searching for

voids inside the patch boundaries previously de®ned

through the neural network segmentation as shown in

Fig. 8. A ¯owchart of the concavity-®nding algorithm

can be seen in Fig. 9.

The location and radius of the holes found is derived

from the co-ordinate information already calculated

for each patch. It is assumed that the top of the hole

begins at the same height as the patch surface. A

speci®c-hole measuring routine is then incorporated

into the CMM probe tool path.

5. Testing of the algorithm

The algorithm was assessed with four different test

objects. The ®rst two objects are without holes, to test

the ability to ®nd individual patches. To begin with,

the L-shaped blended surface discussed in Section 2.3

was used to test the algorithm ability to work with

smoothed surfaces (Fig. 10).

The neural network algorithm identi®ed the top

surface; 37 iterations for the right and 35 iterations

for the left were required to complete the segmenta-

tion. This difference in the number of iterations is due

to the pixel grey level variance created by shadows

from the CMM end effector. The number of iterations

and patches found for the other sides of the object are

shown in Table 2.

At this stage in the reverse engineering process, the

algorithm has generated the following information.

� The number of surfaces on the object.

� The number of holes in the object.

Fig. 8. Raster scan for potential concavities.

Fig. 9. Surface hole location algorithm.

Fig. 10. Picture of L-shaped object with blended surfaces.


� The centroid (Xc, Yc, Zc) of each patch and the

location of each hole.

� The length and width of each patch and the radii of

each hole.

� The orientation of the patches and holes relative to

the CCD camera position at which the stereo image

was taken.

Thus, the object and its constituent features have been

described by the stereo vision system; this data are used

by the CMM control system to govern the location,

orientation and movement of the touch probe.

The algorithm was then tested if it could differenti-

ate between similar surfaces with slightly dissimilar

surface heights. A three surface test object was con-

structed and is shown in Fig. 11. The algorithm

performance is shown in Table 3.

Again, the algorithm correctly found the number of

patches for each view. It can also be noted that the

higher the actual patches on the object, the greater the

number of iterations required to segment the image.

Fig. 12 shows the test object comprised of a planar

surface with three through holes that was used to test

the algorithm's ability to ®nd holes. A stereo image

pair was taken from a view directly above the object.

The results are presented in Table 4.

The co-ordinates and radii of the holes calculated

based on data generated for the location on the patch.To

further test the algorithm, a more complex object, as

shown in Fig. 13, was selected. This L-bracket has three

top surfaces, one is inclined and another has a through

hole. The results of the segmentation and hole-®nding

algorithms are presented in Table 5.

Table 2

L-shaped object with blended surfaces

View Iterations Number of

holes found

Number of

patches foundRight

image

Left

image

Top 37 35 0 2

Right side 55 58 0 3

Left side 19 21 0 1

Front 41 40 0 3

Back 18 15 0 1

Fig. 11. Picture of three-level test object.

Table 3

Three-level test object


holes found

Number of

patches foundRight

image

Left

image

Top 21 18 0 3


Left side 14 14 0 1

Front 11 13 0 1

Back 13 12 0 1

Table 4

Simple test object with through holes


holes found

Number of

patches foundRight

image

Left

image

Top 44 40 3 1


Left side 92 81 0 1

Front 14 22 0 1

Back 14 14 0 1

Table 5

L-shaped test object with through hole


holes found

Number of

patches foundRight

image

Left

image

Top 22 32 1 4


Left side 24 30 0 2

Front 21 21 0 2

Back 23 14 0 1


As indicated in Table 5, the algorithm found one

more patch than actually existed. This problem

occurred due to the large upright portion of the test

object casting a large shadow area with even surface

intensity.

Comparing the results from Tables 2±5, a number of

interesting observations can be made. It is seen that the

number of iterations between right and left images is

usually not the same. This is a result of shadows cast

by the armature of the CMM gantry, causing a change

of the grey level gradient in the images. Because of the

placement of the object on the CMM deck, the right

side of the object is usually in shadow, thus, resulting

in less image contrast. This lack of image contrast

required longer iteration times to segment the image.

The larger patch sizes on the surface of the simple test

object increased the number of iteration cycles

required for the algorithm to cover the surface. With

the exception of the images from the top view of the

complex test object (Fig. 13), the algorithms correctly

identi®ed the number of patches and holes on all the

surfaces of both test objects. Due to the height of the

L-bracket, shadows from the vertical part of the

bracket cast strong shadows onto the top surface, thus,

misrepresenting the actual number of surfaces from

that view.

With the size and location of each patch known, the

path for a CMM touch probe can be programmed.

Each is directed to measure each patch in a grid like

manner. The relative locations between each patch is

considered to avoid collision of the CMM touch probe

stylus. A travelling salesman algorithm is used to

optimise the path of the touch probe. A sample tool

path code for the top surface of the simple test object is

presented in Fig. 14.

6. Conclusions and recommendations

Two primary aspects of reverse engineering are

accomplished by the incorporation of a CCD camera

into the CMM system: tool path planning for the touch

Fig. 12. Picture of simple test object with multiple holes.

Fig. 13. Picture of L-shaped test object with hole.


probe and the identi®cation of separate surfaces on the

object. Although use of neural networks for machine

vision is quite established, the use segmenting of

images with non-constrained boundaries for applica-

tion in reverse engineering has proved promising. The

Kohonen neural network segmentation of the stereo

images proved to provide several advantages over

traditional image processing methods.

� The program is adaptable to segment objects of

different shapes and sizes.

� The number of surface patches that make up the

object need not be known beforehand.

� The number of holes on each surface need not be

known beforehand.

However, their were problems with segmenting

objects with large shadow areas, such as those shown

with the 4th test object in Fig. 13. The grey level

re¯ectance on each separate surface must be fairly

even, and a distinct and continuous grey level gradient

outlining the surface patches must be visible. Thus, it

is a balance between directional and diffuse lighting to

produce the required shadows. Considering that the

experiments were carried out under ambient room

lighting, the results were quite encouraging.

One potential method of overcoming this problem,

which is common to all passive machine vision

methods, is to use structured lighting. Future work

will concentrate on the application of a multi-level

SOM on a true 3-D data derived from structured

lighting.

References

[1] S. Raab, Co-ordinate measurements accelerate reverse

engineering, Machine Design 66 (22) (1994) 50±53.

[2] M. Milroy, D.J. Weir, G.W. Vickers, C. Bradley, Reverse

engineering employing a 3-D laser scanner: a case study,

International Journal of Advanced Manufacturing 12 (2)

(1996) 111.

[3] T. Varady, R.R. Martin, J. Cox, Reverse engineering of

geometric models Ð an introduction, Computer-Aided De-

sign 29 (4) (1997) 255±268.

[4] H.T. Yau, C.H. Menq, Automated CMM path planning for

Fig. 14. CMM tool path code for top surface of simple object.


dimensional inspection of dies and molds having complex

surfaces, International Journal of Machine Tool Manufac-

turers 35 (6) (1995) 861±876.

[5] Y. Takeuchi, H. Shimizu, I. Mukai, Automatic measurement

of three-dimensional co-ordinate measuring machine by

means of CAD and image data, Annals of the CIRP 39 (1)

(1990) 565±568.

[6] M.W. Cho, M.K. Kim, K. Kim, Flexible inspection system

based on a vision guided co-ordinate measuring machine,

International Journal of Production Research 33 (5) (1995)

1433±1488.

[7] S.B. Marapane, M.M. Trivedi, Region-based stereo analysis

for robotic applications, IEEE transactions on systems, Man

and Cybernetics 19 (6) (1989) 1447±1464.

[8] S.S. Christensen, A.W. Andersen, T.M. Jorgensen, C.

Liisberg, Visual guidance of a pig evisceration robot using

neural networks, Pattern Recognition Letters 17 (1996) 345±

355.

[9] S. Baek, D.S. Park, J. Cho, Y.B. Lee, A robot endeffector

tracking system based on feedforward neural networks,

Robotics and Autonomous Systems 28 (1999) 43±52.

[10] M.B. Lynch, C.H. Dagli, Stereoscopic neuro-vision for three-

dimensional object recognition, Mathematical and Computer

Modelling 21 (1/2) (1995) 185±215.

[11] A.J. Worth, D.N. Kennedy, Segmentation of magnetic

resonance brain images using analogue constraint satisfaction

neural networks, Image and Vision Computing 12 (6) (1994)

345±354.

[12] J.S. Lin, K.S. Cheng, C.W. Mao, Multispectral magnetic

resonance images segmentation using fuzzy Hop®eld neural

network, International Journal of Bio-Medical Computing 42

(1996) 205±214.

[13] M.N. Ahmed, A.A. Farag, Two-stage neural network for

volume segmnetation of medical images, Pattern Recognition

Letters 18 (1997) 1143±1151.

[14] J. Koh, M. Suk, S. Bhandorkar, A multilayer self-organising

feature map for range image segmentation, Neural Networks

8 (1) (1995) 67±86.

[15] V.B. Rao, H.V. Rao, Neural Networks and Fuzzy Logic, 2nd

Edition, MIS Press, New York, 1995.

[16] F.M. Wahl, Digital Image Signal Processing, Artech House,

Boston, 1987.

[17] V. Chan, C. Bradley, G.W. Vickers, Automation of laser

scanning for reverse engineering, in: Proceedings of the SPIE

conference Ð Rapid Product Development Technologies,

Boston, 18±22 November 1996.

[18] J. Canny, A computational approach to edge detection, IEEE

Transactions on Pattern Analysis and Machine Intelligence

PAMI-8 (6) (1986) 679±698.

Dr. Vincent H Chan is an assistant pro-

fessor in the Department of Mechanical

Engineering at Ryerson Polytechnic Uni-

versity, in Toronto, Canada. He received

his BASc degree from the University of

Waterloo, MSc from Queen's University

at Kingston, and PhD from the University

of Victoria. He has worked as a design

engineer in the automotive industry and

is presently performing research in the

areas of machine vision and advanced

manufacturing.

Dr. Colin Bradley is an associate pro-

fessor in the Department of Mechanical

Engineering at the University of Victoria,

BC, Canada. He completed his BASc

degree from the University of British

Columbia, MSc from Heriot-Watt Uni-

versity and PhD from the University of

Victoria. His research work has centered

on industrial machine vision, computer-

aided manufacturing and advanced

industrial sensors.

Geoffrey W Vickers is a professor of

mechanical engineering at the University

of Victoria. He has a strong background

in computer-aided design and manufac-

ture. His research on the de®nition and

numerically controlled (CNC) machining

of curved surfaces, which was published

by Ellis Horwood Ltd., has found appli-

cation in mould and die making, in pro-

peller manufacture, in ship building, and

in the marine industry. The integrated

surface fairing and manufacturing program, marketed by Albacore

Research Ltd., has been installed in over a hundred shipyards world-

wide and has been used to produce a range of production vessels.

Some of Professor Vickers' patented inventions are the centrifuge

reverse osmosis desalination unit, the cavitation liquid-jet cleaning

lance, and the ice-blasting device.


a multi-sensor approach to automating co-ordinate measuring machine-based reverse engineering

Documents