2019 5th International Conference on Control, Automation and Robotics

Extraction of Weld Seam in 3D Point Clouds for Real Time Welding Using 5 DOF
Robotic Arm

Vrushali Patil∗ , Indraneel Patil†, V.Kalaichelvi‡, R. Karthikeyan§

∗ †‡Department of Electrical and Electronics Engineering
§Department of Mechanical Engineering
BITS Pilani, Dubai Campus, Dubai, UAE
e-mail: [email protected]∗ , [email protected]†
[email protected]‡, [email protected]§

Abstract—Welding involves the union of two workpiece

halves along a common edge of interface, also called as the
z Proposed methodology can be used for finding 3D pose
weld seam. In this paper a novel algorithm has been of weld seam, seam tracking and measuring dimensions
proposed which is independent of the shape of the workpiece of the weld seam.
to classify and extract weld seams from 3D point clouds. The
z To ensure that no restrictions are imposed on the 3D
real time tracking of the weld seam is also demonstrated on a
5 axis robotic manipulator. The aim was to eliminate the need shape of the weld seam or the geometrical shape of the
for manual computation of the robot trajectory and also welding workpiece.
ensure that the overall pipeline is computationally efficient. z To ensure real time tracking so that pipeline also works
The two workpiece halves have been clustered into different
point clouds and the edge of interface on either of the
for a dynamically moving work piece.
halves has been specified using the centroid shift method. The z To test the pipeline and compile experimental results
precision recall metrics of our algorithm were very close to using a 5 DOF industrial robotic manipulator and a
the ideal values on any shape of the weld seam. Singularity
Kinect stereo camera.
avoidance, trajectory planning, kinematics of the robotic
manipulator is beyond the scope of this paper. The paper is divided into six sections. The second section
discusses the novel proposed PCL methodology for automatic
Keywords-robotic manipulator; point cloud; weld seam; extraction of the weld seam. The third section consists of the
artificial neural networks description of the hardware experimental set up. The fourth
section consists of tracking the extracted weld seam using a
I. INTRODUCTION robotic manipulator in real time. The fifth section compiles all
Assuming the various advantages of robotic welding our results, evaluates the performance of both the algorithm
which have been introduced in brief in the related works and the motion planning and validates our claims. The final
section, in this paper the primary focus is on 3D weld section summarizes the findings, contributions and the future
seams applicable to any common robotic machining task. scope of this work.
One of the main problems here is extracting the weld seam II. RELATED WORK
autonomously and accurately irrespective of the shape of
the welding workpiece. It was observed that all previous With the ongoing advancement in optimal control,
solutions required manual input from the user to highlight trajectory planning and modern ways of sensor data processing
the weld seam which is then later sent to the manipulator robotic manipulators have become extremely mainstream in
as a 3D trajectory. Our work eliminates the need for any all types of industry machining applications like robotic
human interference in this problem and enables a real time welding. There are many advantages of using robotic
solution to extract and track a weld seam by a 5 DOF machining cells over other conventional machines like the
manipulator. We have mainly used the Point Cloud Library Computer Numerical Control machine especially while
to implement our algorithm, ROS robotics framework for considering the much wider working area which allows it to
data acquisition and communication and the Matlab accommodate much bigger workpieces and the flexible
Neural Networks Toolbox for transformation of coordinate kinematics of robotic arms, are often capable of machining
axes. parts with intricate details and complex shapes [1]. Despite
The objectives and contributions of this paper are: these obvious wins for robotic machining some common
criticism about their performance and efficiency state the
z To propose a PCL pipeline which automatically following facts: Low positioning accuracy during complex
extracts the weld seam from a 3D point cloud so that tasks resulting in far from robust operation More prone to
the manipulator trajectory does not have to be disturbances from reaction forces on the end-effector tool
manually specified using a CAD model. Possibility of resonance due to machining process vibrations

978-1-7281-3326-3/19/$31.00 ©2019 IEEE 727

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on June 02,2021 at 08:31:27 UTC from IEEE Xplore. Restrictions apply.
 Hence for robotic machining cells to take over structured different types of edges in a variety of workpiece shapes. We
tasks, it is a given that real time closed loop control and have successfully taken up this challenge and found a solution
advanced sensors are required to overcome the above using a novel approach in our paper.
shortcomings. 2D computer vision is proven as an
extremely useful tool to assist robotic machining tasks like III. METHODOLOGY
welding in the following literature 2, 3, 4 and 5. Organized point cloud of the workspace is the starting
However, in the recent times researchers have commonly point and the end goal of this PCL pipeline is to extract
preferred stereo cameras over normal cameras while a weld line from the noisy point cloud which can be later
implementing such complex tasks, due to the advancement sent to the robotic manipulator as a 3D trajectory for tracing
in research in 3D computer vision using the Point Cloud the weld seam. It is assumed that the welding workpiece is
Library (PCL) [6], emergence of low cost devices like the kept inside the workspace of the manipulator. Pipeline is
Kinect and since point clouds capture more data than described here with a cylindrical welding workpiece with a
images including the depth perception. Use of Stereo radially cut welding seam which is shown in Fig. 1. The entire
cameras in an ordered task like welding also eliminates the pipeline can be divided into two branches. Due to limitations
need for a fixed welding work-table, allows to place no in computation power, the input cloud has been downsampled
restriction on the shape of the workpiece or the weld to unorganized point cloud in Branch 1. Branch 2 describes
line and gives freedom to use the entire workspace of the rest of the organized point cloud processing to extract the
robotic manipulator [7]. Since the aim of the paper is to weld seam. The algorithm flowchart is shown in Fig. 2.
automatically segment the weld seam using 3D computer
vision, the subject of edge segmentation in complex point
clouds gains special importance. It usually serves as the
foundation for solving higher level problems such as
object recognition, interaction analysis and scene
understanding [8]. Edge based segmentation is quite (a) Straight Cut (b) Curved Cut
compli- cated since there are many different types of edges
in any point cloud just like any 2D image. Depth
discontinuities in the 3D data produce two related types of
edges: occluding and occluded, areas where surface
normals change rapidly lead to high curvature edges, RGB
edges are obtained by backprojecting edges on the 2D (c) Inclined Cut (d) Cross-section
photometric data onto the 3D structure[9]. The most
revolutionary papers in this area are edge detection by Figure 1. Workpieces used for testing.
analyzing the geometric properties of each query point’s
neighborhood, and then combining Random Sample
Consensous (RANSAC) and angular gap metric [10], edge
feature detection by computing a Gauss map clustering on
local neighborhoods and then an iterative feature selection
process [11], contour extraction by setting a contour score
for each individual point as predicted with a binary
classifier using a set of features extracted from the point’s
neighborhood and then selecting an optimal set of
contours from the candidates
[12] and more recently edge detection using the fast
nearest neighbor search on organized point clouds [13].
Most of the above methods are either restricted to
organized point clouds or are unsuitable for real time
application due to the high time delay. Hence real time
operation of the robotic manipulator is not possible. For
Unorganized point clouds most recent and according to us
the only technique feasible for real time robotic application
and also includes online trajectory generation is based on
classifying points as edge or non-edge based on evaluating
degree of shift in centroid from their initial position [14].
But an important aspect which has not been addressed in
this paper is how the weld seam is extracted from all the
boundary edges of the welding workpiece. In the
experimental results sections, the authors mention that the
panel contains only straight line seams but in all practical Figure 2. Algorithm flowchart.
applications, the weld seam must be extracted from all

728

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on June 02,2021 at 08:31:27 UTC from IEEE Xplore. Restrictions apply.
 A. Branch 1 estimated using these inlier points in the cloud.
1) Voxel Downsampling: Aim of this step is to reduce z Then all the other points are iterated and added to those
the number of points in the point cloud so that the points which fit this model to the inliers list.
processing is faster and pseudo real time. PCL
z The model is re-estimated from the final list of inliers
VoxelGrid class creates a 3D voxel grid in the form of
once some satisfactory threshold of inlier points has
boxes over the input point cloud data and then in each
been reached [16].
box all the points present are replaced by their
centroid. Unluckily organized point cloud dataset is
the one in which data is organized into a fixed
number of rows and columns and hence once the
cloud has been downsampled, it is converted into an
unorganized cloud making any type of computation Figure 4. Segmented workpiece.
all the more challenging. The input cloud and the
3) Euclidean Clustering: After segmenting the
downsampled cloud are shown in Fig. 3(a) and Fig. workpiece, the two halves that are to be welded together are
3(b) respectively. clustered individually and pushed to two separate point
clouds. In this step weld seam formed from the interface
edge of two halves is converted into two separate occluding
edges. The focus is now on the individual occluding edges
in either of the workpiece halves. Euclidean Clustering is
utilized to separate the two work piece halves into two
separate clusters. The goal of the clustering technique is to
divide unorganized point cloud model into smaller parts
based on spatial decomposition. An object cluster is defined
as follows, where dth is maximum imposed distance
(a) Input Cloud threshold: Let Oi = pi P be a distinct { ę } from Oj
cluster
= pj P if: { ę }
2
min " pi − pj " ≥ dth ... (1)
All the steps in Euclidean Clustering can be summarized as
follows:
z A Kd-tree representation is superimposed on input point
cloud P, C is the empty list of clusters and a Q is the
array of points that need to be clustered.

(b) Downsampled Cloud z Perform the following steps as we cycle through each
Figure 3. Prefiltering of cloud.
point pi ę P :
• Add pi to the array of points Q and for every point
2) Random Sample Consensus (RANSAC) Workpiece
pi ę Q do:
Segmentation: In this step the welding workpiece was
segmented from the entire workspace of the manipulator • · for the set P i of point
Search k neighbors of p i
so that the noisy points, other objects and obstacles can in a sphere with radius r < dth;
k k
be rejected from the input point cloud (Fig. 4). The · For every neighboring point p i ę P i ,
RANSAC model iterates through different Sample which is unprocessed, add it to Q;
Consensus Models namely spherical, planar and z Once every point in Q is classified, add Q to the list of
cylindrical, tries to estimate model of work piece and clusters C, and reset Q
then segments out the workpiece from the cloud [15].
RANSAC algorithm for segmenting a geometrical z Euclidean clustering ends once all points p ę P are
i
object in point clouds can be summarized with the clustered into the clusters list C
following steps: 4) Identifying the Interface Edge: After Euclidean
Clustering, we accessed two different workpiece halves and
z A random subset of points from the entire point
saved them in two different point clouds. Further, only one
cloud are considered as the hypothetical inliers for cluster of the workpiece which is half 2 (right) was considered.
the geometrical model. In this right half, we must reject all the other edges not related
z All the coefficients of the geometric model are to the edge of interface which forms the weld seam along with

729

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on June 02,2021 at 08:31:27 UTC from IEEE Xplore. Restrictions apply.
 its other half. For achieving this, it is required to find how isolate the edges related to the workpiece half 2 through the
the left cluster (half 1) is oriented and translated with PCL extract indices class directly. But since the workpiece
respect to this particular half 2. Assuming the current half cluster from Branch 1 is an unorganized point cloud and
workpiece half 2 has four edges: top, bottom, left and right. the high curvature edge cloud is an organized one, hence it
Any one of these four edges could be the interface edge was required to use alternative methods to find the edges. The
along which the weld seam is present. To identify the pose KdtreeFLANN PCL class was used for performing fast
of the second cluster, the PCL Centroid Class was used to approximate nearest neighbor searches in the high curvature
independently find centroids of both the workpiece halves. edges cloud around each point in the workpiece half cloud.
Furthermore, both the spatial centroid points were joined Kd-tree is a data structure in computer science and is a binary
with a 3D line (Fig. 5). Based on the relative location of search tree where constraints are imposed on it to perform
both the centroids and by comparing x and y coordinates it nearest neighbor searches. All the points in the high curvature
was concluded which of the edges is the edge of interface. cloud which are declared inliers to our search radius of 0.1
Rest of the edges are not required and can be rejected. In were exported to another cloud called workpiece half edges.
this case the left edge of the workpiece half 2 is the edge 3) Edge of Interface Extraction based on Centroid Shift:
of interface and the top, bottom and right edges can be The workpiece half edges cloud was available and also it was
ignored. known from Branch 1 that the left edge of the workpiece half
2 is the interface edge. Since in point clouds there are no
B. Branch 2 individual edges but just a grid of points, we were required to
1) Scene Edge Segmentation : The most reliable and extract all the points corresponding to this left edge from the
real time algorithms for edge segmentation in PCL require workpiece half 2 edges cloud. The left edge points from the
or- ganized point clouds. So all the Branch 2 operations workpiece half edges cloud was computed using the Centroid
were performed on the default input point cloud of the Shift Approach. The KdTreeFLANN nearest neighbor search
workspace scene. Reliable 3D edges are computed by PCL class was employed again to find the neighbors of each
exploiting depth discontinuities, photometric RGB point in the workpiece half edges cloud. These nearest
intensity changes and by backprojecting 2D edges to 3D neighbors were exported to temporary cloud temp. Since cloud
point clouds instead of complex 3D processing [9]. Out of temp is a cloud consisting of neighbors of an edge point, it is
all the types of edges, high curvature edges were chosen to asymmetrically distributed. If the centroid of this temp cloud
give the best results. High curvature edges occur where is calculated, it shifts away from the edge towards the center
surface normals change abruptly. Apply a variant of of workpiece half 2. This shown in Fig. 7(a) and Fig. 7(b).
Canny Edge Detection, the regions with higher degree of This is the fact that has been exploited to classify the edge
normal variations are detected by applying Sobel operator point as top, bottom, left or right edge point. Since the edge of
to the x and y components of the normals which are interest lies to the left side of the workpiece half 2, it can be
detected from the point clouds. At the end of this step we easily concluded that the centroid of the neighbors cloud of
have a point cloud of high curvature edges of the entire every left edge point will shift towards the right. We iterated
scene (Fig. 6). through every point in the workpiece half edges cloud. Further,
the difference between the co-ordinates of the edge point and
the centroid point of the neighborhood cloud temp was
computed. For a top edge, only the y coordinate of the
centroid varied while the x and z co-ordinates remained almost
constant. Similarly, for a left edge, the difference between the
y and z co-ordinates was zero while the x co- ordinate had
Figure 5. Line joining the two workpiece centroids. shifted significantly by some distance. Thus, the points
corresponding to the edge of interface were extracted and
pushed to the final segmented weld seam cloud. This left edge
of the workpiece half 2, which is in principle an occluding
edge is the interface edge between the two halves of the
welding workpiece that needs to be welded together. This
interface edge also accurately overlaps our desired weld seam
in which we are interested in. Thus, the goal of extracting the
weld seam in real time in a computationally efficient way was
achieved (Fig. 7(c)).
Figure 6. Edges in the input cloud.

2) Workpiece Edge Segmentation: As it was required

in finding the edges for the welding workpiece, workpiece
edges from the entire scene edges were isolated. As
discussed earlier, workpiece half 2 from Step 3 in Branch 1
was considered as the target cloud. All the rest of our
processing was done only on this half. It is possible to (a) Shifted Centroids Along Edge

730

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on June 02,2021 at 08:31:27 UTC from IEEE Xplore. Restrictions apply.
 different input-output devices with minimal lines of code. A
Microsoft Kinect Sensor was placed in workspace and data
from it was accessed using the iai kinect2 ROS package [19].
The entire point cloud processing and proposed pipeline was
implemented on a Linux computer with DDR4 SDRAM
2133Mhz 16GB, 512GB SSD, 6th Gen Intel Core i7 and in
C++ language. However, it was also required to have a
windows operating system for the Microsoft ActiveX library,
(b) Assymetrical Neighborhood hence one Linux computer acquired data from the sensor and
processed it while the windows computer controlled the
robotic arm. ROS cannot be installed on Windows computers.
Hence in the present work, ROS in MATLAB was employed
to allow MATLAB/ Simulink to communicate with other ROS
computers on the same network. Further, both the Linux
computer and the Windows computer were connected using
(c) Segmented Weld Seam ethernet cables and connection was enabled so that data can
Figure 7. Segmentation of weld seam from all the edges. be seamlessly transferred between each other. Each computer
was made available and visible in the ROS network for
IV. HARDWARE EXPERIMENTAL SETUP bilateral connectivity. The Linux computer published data
The 5-DOF TAL Brabo! manipulator was employed to while the windows computer subscribed this data and
trace the weld seam extracted by our proposed pipeline. performed further actions. After successful extraction of the
This manipulator does not allow to communicate with the edge, x, y, z co-ordinates of the weld seam were sent from the
servo motors directly rather trio motion controller is used Linux computer to the windows computer running MATLAB
to control the motors of the manipulator by preset factory through ROS multiple machines setup. This data was then fed
conditions. Furthermore, Trio motion controller provides to the neural network and real world co-ordinates were
only ActiveX support to interface the robot to the obtained. The output co-ordinates of the network were sent to
controller. The ActiveX control standard is developed by the controller and the robotic arm traced the weld seam in real
Microsoft and it describes modular, reusable software time.
components that can be used universally by any
V. ROBOTIC MANIPULATOR MOTION PLANNING
environment that supports the standard and in any
programming language. MATLAB was identified as the As mentioned earlier, the TAL Brabo! robotic arm does
best way to interact with the controller through this not allow implementation of inverse kinematics to control its
ActiveX COM object. An ActiveX window was created motion. Therefore, Artificial Neural Networks (ANN) were
and connection was established using the function employed to translate the point cloud x, y, z co-ordinates to the
‘actxcontrol’ supported by MATLAB so that the robotic real world X, Y, Z co-ordinates, which can be sent to the trio
arm can be controlled using different public properties controller [5]. An artificial neuron is a computational model
and methods of the PC Motion COM object. The entire which consists of nonlinear statistical data modeling tools
hardware setup is shown in Fig. 8. where the complex relationships between inputs and outputs
are modeled or patterns are found. The fundamental processing
element of a neural network is a neuron. The weights of these
neurons are adjusted on the basis of a given data. In first
feedforward pass of the neural network training, random
weights and biases are assigned to the neurons. Later, at every
epoch of the backpropagation pass, these weights are updated
so that the error metric is minimized.
A. Data Collection and Architecture
Supervised learning for ANN’s was employed where input
Figure 8. Experimental setup. and output vectors were rendered to the network. A hexagonal
plane was segmented by the RANSAC algorithm. Further,
To create a real time welding process, it is necessary to centroid of the segmented plane was computed and pushed
communicate continuously with the robotic manipulator. to the point cloud. The x, y and z co-ordinates of the centroid
The information about the control actions and movement were recorded from the point cloud of the plane. These
commands should be communicated to the controller of centroid co-ordinates served as input to the Neural Network.
the manipulator with high speed and accuracy. Hence, the The robotic arm was then manually moved to the center of the
widely used Robot Operating System (ROS) framework plane and real world coordinates were recorded as target data.
was adopted for data acquisition and data communication Furthermore, the plane was moved and held at different
[18]. ROS is a meta-operating system which runs on top of locations, heights and orientations and 50 readings were
a Linux computer and helps in the intercommunication of recorded so that the workspace of the robotic arm was covered

731

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on June 02,2021 at 08:31:27 UTC from IEEE Xplore. Restrictions apply.
 majorly. The position of the Kinect sensor was fixed to where TP stands for True Positives which are correctly
avoid errors. The neural network consisted of 3 neurons in identified points, FP stands for False Positives for wrongly
the input layer, 9 neurons in the hidden layer and 3 detected points, FN stands for False Negatives, representing
neurons in the output layer. The architecture is shown in false rejections, i.e. points that belong to the weld seam in the
figure below. After processing the input point cloud, the ground truth but are not extracted by our PCL pipeline [21]. In
weld seam was extracted by the proposed pipeline. The x, the present work we have obtained the optimal performance of
y, z co-ordinates of this segmented weld line were the proposed methodology with Cluster Tolerance of 0.8 cm,
published by the Linux computer. The windows computer 0.1 cm radius of sphere for nearest neighbor search while
received these co-ordinates in MATLAB through ROS calculating shift in centroid and 0.01 cm as threshold for
Subscriber in real time. These end- coordinates of the weld centroid shift.
seam were input values to the Neural Network. The pre-
trained Neural Network model then mapped the point TABLE I. PRECISION AND RECALL METRICS
cloud co-ordinates to the X, Y, Z co-ordinates of the end
effector. The tracing of the weld seam was experimentally
verified by sending the output values (X, Y, Z) from the
network to the trio motion controller of the robotic arm.
The robotic arm moved to these co-ordinates and traced 
the weld seam.
B. Performance of the Motion Planning Algorithm
Levenberg Marquardt (LM), and Scaled Conjugate
Gradient (SCG) algorithms were used to train and compare the
performance of the neural networks keeping all the parameters
constant for comparison. The data was divided into following
Figure 9. Neural network architecture. ratio: Training set: 65LM produced least mean squared error
and fit more data points as compared to the SCG algorithm
VI. EXPERIMENTAL RESULTS AND ANALYSIS (Figure 11). This algorithm also performed well on unseen
A. Evaluation Metrics of the Proposed Weld Seam data. The error on unseen data was very less, hence Levenberg
Extraction Marquardt (LM) with a learning rate of 0.01 was chosen for
the neural network training. To verify the performance of the
The proposed weld seam extraction pipeline was algorithm in real time, we conducted several experiments. The
evaluated on three different welding scenarios. In all three path traced by the robotic arm was satisfactory and it can be
scenarios, the welding workpiece was always cylindrical concluded that the motion of the arm can be controlled using
in shape but the welding seams were cut radially straight, Neural Networks.
radially inclined and radially curved. The pipeline is
independent of the shape of the workpiece and will work
flawlessly on any geometrical shape spherical, cylindrical,
conical or planar. The obtained results are shown in the
Fig. 10 below.

Figure 10. (a) Inclined Cut (b) Curved Cut (c) Cross-section. 
Figure 11. Performance of different algorithms.
We have chosen precision and recall estimates as the
evaluation metric in analyzing the performance of our
pipeline in all the three scenarios. Ground truth of the weld
seam has been calculated from the 3D point cloud by
manually selecting the points along the desired trajectory
in Mesh Lab [20].

 (a)

732

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on June 02,2021 at 08:31:27 UTC from IEEE Xplore. Restrictions apply.
 [4] L. A. Fernandes, A. Saraiya, V. Kalaichelvi and R. Karthikeyan, ”Per-
formance Evaluation of Digital Image Processing and Artificial Neural
Networks for Weld Line Detection of Robotic Manipulator,” 2017 9th
IEEE-GCC Conference and Exhibition (GCCCE), Manama, 2017, pp. 1-
6.
[5] S. H. Rao, V. Kalaichelvi and R. Karthikeyan, ”Real-Time Tracing Of A
Weld Line Using Artificial Neural Networks,” 2018 IEEE/ACIS 17th
International Conference on Computer and Information Science (ICIS),
Singapore, 2018, pp. 275-280.
(b)
[6] R. B. Rusu and S. Cousins, ”3D is here: Point Cloud Library (PCL),”
Figure 12. Real time tracing using 5 DOF robotic arm. 2011 IEEE International Conference on Robotics and Automation,
Shanghai, 2011, pp. 1-4.
VII. CONCLUSION AND FUTURE WORK [7] S. M. Ahmed, Y. Z. Tan, G. H. Lee, C. M. Chew and C. K.
In this paper, a novel approach has been demonstrated Pang, ”Object detection and motion planning for automated welding of
tubular joints,” 2016 IEEE/RSJ International Conference on Intelligent
to extract the 3D weld seam from noisy background. Robots and Systems (IROS), Daejeon, 2016, pp. 2610-2615.
Experiments were performed with different weld seam [8] X. Lin, J. R. Casas and M. Pardas, ”3D point cloud segmentation using
curves. The precision obtained was very close to ideal a fully connected conditional random field,” 2017 25th European Signal
readings. It can be concluded that the proposed pipeline Processing Conference (EUSIPCO), Kos, 2017, pp. 66-70.
works for all types of weld cuts and also allows [9] C. Choi, A. J. B. Trevor and H. I. Christensen, ”RGB-D edge detection
placement of the workpiece in any orientation and and edge-based registration,” 2013 IEEE/RSJ International Conference
alignment so as to utilize the entire 3D workspace of the on Intelligent Robots and Systems, Tokyo, 2013, pp. 1568-1575.
robotic arm. Further, using MATLAB and Neural [10] Ni, Huan & Lin, Xiangguo & Ning, X & Zhang, Jixian. (2016). Edge
Detection and Feature Line Tracing in 3D-Point Clouds by Analyz-
Networks, the control of robotic arm to trace the weld ing Geometric Properties of Neighborhoods. Remote Sensing. 8. 710.
seam is presented. The performance of the neural network 10.3390/rs8090710.
was analyzed by tuning various parameters. Levenberg [11] C. Weber, S. Hahmann and H. Hagen, ”Sharp feature detection in
Marquardt algorithm was chosen as it performed very well point clouds,” 2010 Shape Modeling International Conference, Aix-en-
on unseen data. The robot traced the weld seam in real Provence, 2010, pp. 175-186.
time for all three weld seam profiles with satisfactory [12] T. Hackel, J. D. Wegner and K. Schindler, ”Contour Detection in
results. However, one limitations of our algorithm is the Unstructured 3D Point Clouds,” 2016 IEEE Conference on Computer
requirement of organized point cloud. In present work, two Vision and Pattern Recognition (CVPR), Las Vegas, NV, 2016, pp.
1610- 1618.
separate branches are considered for processing and are
[13] R. Bormann, J. Hampp, M. H a¨gele and M. Vincze, ”Fast and accurate
later combined. This work can be extended such that the normal estimation by efficient 3d edge detection,” 2015 IEEE/RSJ
entire processing of weld seam is carried out on International Conference on Intelligent Robots and Systems (IROS),
unorganized point cloud. Secondly, the depth sensor is Hamburg, 2015, pp. 3930-3937.
fixed at a static position in the workspace of the [14] Ahmed, Syeda Mariam, Yan Zhi Tan, Chee-Meng Chew, Abdullah
robotic arm. The dynamic welding process can be Al Mamun and Fook Seng Wong. “Edge and Corner Detection for
implemented by placing the Kinect sensor on the robotic Unorganized 3D Point Clouds with Application to Robotic Welding.”
CoRR abs/1809.10468 (2018): n. pag.
arm. In this paper forward kinematics is employed. An
[15] Schnabel, Ruwen & Wahl, Roland & Klein, Reinhard. (2007). Efficient
attempt can be made to circumvent the given controller of RANSAC for point-cloud shape detection. Comput. Graph. Forum. 26.
the robotic arm so that inverse kinematics can be adopted. 214-226. 10.1111/j.1467-8659.2007.01016.x.
[16] A. Fischler, Martin & C. Bolles, Robert. (1981). Random Sample
ACKNOWLEDGMENT Consensus: A Paradigm for Model Fitting with Applications To Image
The authors are extremely grateful for the constant Analysis and Automated Cartography. Communications of the ACM. 24.
381-395. 10.1145/358669.358692.
support, encouragement and motivation given to us by the
authorities, colleagues and friends at BITS Pilani, Dubai [17] C. Choi, A. J. B. Trevor and H. I. Christensen, ”RGB-D edge detection
and edge-based registration,” 2013 IEEE/RSJ International Conference
Campus. We are also thankful for all the hardware and on Intelligent Robots and Systems, Tokyo, 2013, pp. 1568-1575.
software facilities provided to us for carrying out this [18] Quigley, M & P. Gerkey, B & Conley, K & Faust, J & Foote, T & Leibs,
research work. J & Berger, E & Wheeler, R & Y. Ng, A. (2009). ROS: An open-source
Robot Operating System. ICRA Workshop on Open Source Software. 3.
REFERENCES 1-6.
[1] Doukas, Christos & Pandremenos, Ioannis & Stavropoulos, [19] T. Wiedemeyer, “IAI Kinect2,” https://github.com/code-iai/iai kinect2,
Panagiotis. (2011). Machining with Robots: A Critical Review. Institute for Artificial Intelligence, University Bremen, 2014 – 2015,
accessed June 12, 2015.
[2] Liu Suyi, Liu Lingteng, Zhang Hua, Bai Jianjun and Wang
Guorong, ”Study of robot seam tracking system with laser vision,” [20] Cignoni, Paolo & Callieri, Marco & Corsini, Massimiliano & Dellepi-
2009 Interna- tional Conference on Mechatronics and Automation, ane, Matteo & Ganovelli, Fabio & Ranzuglia, Guido. (2008). Mesh- Lab:
Changchun, 2009, pp. 1296-1301. an Open-Source Mesh Processing Tool.. Computing. 1. 129-136.
10.2312/LocalChapterEvents/ItalChap/ItalianChapConf2008/129-136.
[3] Ma, Hongbo & Wei, Shanchun & Sheng, Zhongxi & Lin, Tao &
Chen, Shanben. (2010). Robot welding seam tracking method based [21] Raghavan, Vijay & S. Jung, Gwang & Bollmann, Peter. (1989). A
on passive vision for thin plate closed-gap butt welding. The Critical Investigation of Recall and Precision as Measures of Retrieval
International Journal of Advanced Manufacturing Technology. 48. System Performance.. ACM Transactions on Information Systems. 7.
945-953. 10.1007/s00170- 009-2349-4. 205-229. 10.1145/65943.65945.

733

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on June 02,2021 at 08:31:27 UTC from IEEE Xplore. Restrictions apply.