0% found this document useful (0 votes)
17 views5 pages

Paper 8 Robot

paper8
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
17 views5 pages

Paper 8 Robot

paper8
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

International Journal of Engineering and Advanced Technology (IJEAT)

ISSN: 2249 – 8958, Volume-9 Issue-1, October 2019

Data -Enhanced Convolution Neural


Networks for Wall Following Robot Navigation
Sandip Kumar Singh


Abstract: Machine learning has been used for solving the H., & Juang, C. F. propose evolutionary wall-following
Robot Navigation Task through the wall-following control. The control of a mobile robot using an interval type-2 fuzzy
wall-following control involves the movement of the robot in some controller (IT2FC) with species- differential -evolution-
directed direction maintaining a constant distance from a given
activated continuous ant colony optimization [4]. Das, A. K
wall. The path of the movement of robot is measured by ultrasonic
sensors. Many machine learning methods have been used for this
describes the method of real-time estimation on a car-like
problem, as classifiers, but Convolution Neural Networks (CNN) robot using a single omnidirectional camera as a sensor [7].
outperforms them all with almost 98% of accuracy. This study The concepts of car maneuvers, fuzzy logic control (FLC),
compared the performance of five classifiers SVC, MLR, ANN, and sensor-based behaviors are merged to implement the
CNN-1D, and CNN-2D, which play the part of controller in the human-like driving skills by an autonomous car-like mobile
navigation work. We have used the ultrasonic sensor data to robot (CLMR) by Li, T. H [6]. Millan et al. used machine
understand the hidden pattern in the navigation work and learning along with asynchronous electro cephalogram in
classified the actions by robot in terms of different motions
advanced robotics [9] for control of a mobile robot. Ge, S. S
performed by robot in response to it.
The classification reports of CNN-2D and CNN-1D with describes the problem of goals unreachable with obstacles
Artificial Neural Networks (ANN) have also been presented in this nearby when using potential field methods for mobile robot
paper. The smart Data-Enhancement used in proposed method path planning [10]. Negishi, Y [3] describes the navigation of
significantly improves the classification performance of all a mobile robot in unknown tactic environments using an
classifiers, especially CNN. omnidirectional stereo and a laser range finder. Brooks et al.
used a new design methodology for control systems, known as
Keywords: Convolution Neural Networks (CNN), subsumption architecture. It decomposed the problems in
Wall-following robot navigation, Multinomial logistic Regression
task- achieving behaviors [5]. Classifying data is a prime task
(MLR), Support Vector Classifier (SVC)
of machine learning. Support vector machines (SVM) are
used as binary classifiers. They are used both as supervised
I. INTRODUCTION
learning methods and unsupervised learning methods. When
A robot is an automatically working machine and can do some the data is labelled, classification by SVM is called SVC that
task that human can do. A mobile robot is specially designed is support vector classifier. In case of unsupervised learning
to operate in environments such as automated assembly halls, where data is unlabelled, it is clustered in two distinct groups
factories, or warehouses [2]. The wall-following robot with the help of SVM. The Multinomial Logistic Regression
navigation is the movement of robot keeping a fixed distance (MLR) is a further extension of Logistic regression, which is a
from wall. The path of the movement of the robot is measured binary classifier. MLR classifies multiclass data both linearly
by ultrasonic sensors. [11]. The ultrasonic sensors are used in and nonlinearly. Primary Logistic regression classifier is
many numbers to generate data for wall and path following useful in binary situations like on /off, fail /pass, live/ dead,
robots. This data is used as a feature for pattern learning and etc., where classes are discretely separable. In such situations,
further utilized to control the robot motion. The immediate the classes can be labelled between "0" or "1". Multinomial
response of the robot after learning the sensor data is called Logistic Regression is used for the dependent variables which
real time obstacle avoidance. The sensor data is converted to are categorically equivalent. This categorical equivalence of
usable data with the help of complex algorithms. Juang et al. dependent variables is also called nomiality. This multiclass
[1] proposed fuzzy controller (FC) based reinforcement ant Logistic Regression is also called Multinomial Logistic
optimized design method, and applied it to Regression. We propose, a CNN based classification solution
wheeled-mobile-robot wall-following control. Ando, Y of wall-following control of robot navigation with
presents a method for an autonomous mobile robot with a outstanding performance of 98% classification accuracy.
sonar-ring to follow walls [8]. The sonar-ring consists of
multiple ultrasonic range sensors. The proposed
wall-following algorithm makes a robot able to follow a wall
in various shapes. Hsu, C.

Revised Manuscript Received on October 15, 2019


Sandip Kumar Singh Department of Mechanical Engineering, V B S
Purvanchal University Jaunpur (U.P.), India
[email protected]

Published By:
Retrieval Number: A1361109119/2019©BEIESP Blue Eyes Intelligence Engineering
DOI: 10.35940/ijeat.A1361.109119 4426 & Sciences Publication
Data -Enhanced Convolution Neural Networks for Wall Following Robot Navigation
II. THE BASIC CONCEPT OF CNN AND PROPOSED CNN consists of some convolutional and subsampling layers
APPROACH which are followed by fully connected layers. There are k
The convolution neural networks (CNN) are different from number of filters in the convolution layer. The size of these
artificial neural networks (ANN). The ANN consists of one filters is m × n × q, where m is the dimension of image and n is
input layer and other output layer with one hidden layer in smaller than m, k is number filters, r is the number of channels
between. On other hand, in CNN there are convolution layers and q can be equal to r, or less than it. Each map is then
followed by sub- sampling layers in between the input and sub-sampled typically max-pooling over p x p regions with p
output layers [12-13]. The input to the convolution layer is ranges between 2 to 5 for smaller and larger inputs
given in form of an image of m ×m ×r size, where r represents respectively. Figure 3 represents the architecture of CNN
the number of channels which is equal to 3 for RGB image. A with number of convolution and sub- sampling layers.

Fig. 1. Flow chart of Proposed Method

Fig. 2. The convolutional layer with pooling

Published By:
Retrieval Number: A1361109119/2019©BEIESP Blue Eyes Intelligence Engineering
DOI: 10.35940/ijeat.A1361.109119 4427 & Sciences Publication
International Journal of Engineering and Advanced Technology (IJEAT)
ISSN: 2249 – 8958, Volume-9 Issue-1, October 2019

Let be the error related term for the st layer is the input to the -th layer, and is the input image. The
in the network with a cost function , operation is the “valid” convolution
where ( are the parameters and are the data between th input in the th layer and the error w.r.t.
the layer is densely connected to the the th filter.
-st layer, then the error for the -th layer is computed In proposed CNN, we have used 5 layers which 600 hidden
as units. There are four response classes, accordingly output
layers consists of 4 units. Hyperbolic tangent function has
( ) been used activation function. Decision has used zero bias 0.
Where ( is derivative of activation function 9 momentum, and 0.0006 learning rate.
and the gradients are
III. DATA DESCRIPTION
If the -th layer is a convolutional and subsampling layer, then The data was recorded @ 9 samples per second and
the error is propagated through as generated a database with 5466 examples. The location of
different items at the test place can be seen as in figure 5. The
= unsample ) ) SCITOS G5 robot has been used for this purpose. The
collection of data is done while the robot moves with the help
Where k indexes the filter number, and ) is the of walls in a clockwise direction four times. The robot uses 24
derivative of the activation function.
ultrasound sensors to navigate, which are arranged circularly
Lastly, to calculate the gradient w.r.t to the filter
around its "waist." The numbering of the ultrasound sensors
maps, the same way we flip the filters in convolution starts at the front of the robot and increases in the clockwise
layer direction. The data contains the raw values of the
measurements of all 24 ultrasound sensors and the
corresponding class labels-- Move-forward,
Slight-Right-Turn, Sharp-Right-Turn, Slight-Left-Turn.
)
Sampling is done at a rate of 9 samples per second. This data
Where, is gradient of parameter W w.r.t. to has been taken from UCI data repository [12].

filter, is gradient of parameter b w.r.t. to filter,

Fig. 3. 2D CNN configuration

Fig.4. Sketch of the robot’s navigation environment

Published By:
Retrieval Number: A1361109119/2019©BEIESP Blue Eyes Intelligence Engineering
DOI: 10.35940/ijeat.A1361.109119 4428 & Sciences Publication
Data -Enhanced Convolution Neural Networks for Wall Following Robot Navigation
Slight-Right-Turn 1 34 497 4 0
IV. RESULT AND ANALYSIS Sharp-Right-Turn 2 14 5 1690 0
Slight-Left-Turn 3 0 3 0 252
In Machine Learning prediction can be done in two ways, one
is predicting positive results and the other is predicting TABLE III. Classification Report of CNN-2D
negative results. The outcome of such predictions may be true
positive and false positive, true negative and false negative CNN-2D
results. Hence, total number of outcomes of any kind of Class/label precision recall f1-score support
predictions is of four types, which is true positive, true Move-forward 0 0.97 0.96 0.97 1725
negative, false positive & false negative. Slight-Right-Turn 1
Precision is optimistically correct predictions (in fractions) 0.93 0.93 0.93 535
with respect to total optimistic predictions. Sharp-Right-Turn2
Recall is optimistic predictions with respect to total number 0.97 0.99 0.98 1709
of correct predictions. It is also in fractions. Slight-Left-Turn 3 1 0.99 0.99 255
The occurrence of high-precision and high recall indicates Avg/ total 0.97 0.97 0.97 4224
that the predictor is capable of predicting correct and positive Confusion Matrix
results simultaneously, which is sign of most effective Class 0 1 2 3
Move-forward 0 1651 28 44 2
predictor.
Slight-Right-Turn 1
The combined effect of precision and recall is represented in
39 491 5 0
terms of f1-score. Mathematically f1- score is the harmonic
Sharp-Right-Turn 2
mean of precision and recall.
11 4 1694 0
The precision-recall curve: The precision-recall curve depicts Slight-Left-Turn 3 2 0 0 253
the relation between Precision and recall for different values.
The high president implies that there is a low false-positive
rate. The high recall means a low false-negative rate. When
Precision and recall both are large, the area under curve is also
very large, which is the most desired situation. Figs 5,6 & 7
show that the classifier is predicting correctly, and all
predictions are positive.
The Precision-recall curves are generally useful in binary
classification to understand the output of a machine learning
classifier. The precision -recall technique is extended to a
multiclass situation in our case.it is achieved by binarizing the
output of result.
TABLE I. Classification Report of ANN
ANN
Class/label precision recall f1-score support
Move-forward 0 0.95 0.96 0.95 1725
Slight-Right-Turn
1 0.95 0.88 0.91 535
Sharp-Right-Turn
2 0.97 0.98 0.97 1709
Slight-Left-Turn 3 1 0.96 0.98 255
Avg/Tot 0.96 0.96 0.96 4224
Confusion Matrix
Class 0 1 2 3
Move-forward 0 1655 22 48 0
Slight-Right-Turn
1 57 472 6 0
Sharp-Right-Turn
2 29 3 1677 0
Slight-Left-Turn 3
8 0 3 244

TABLE.II Classification Report Of CNN-1D


CNN-1D
suppo Fig.5. Precision-Recall curves for SVC and MLR
Class/label precision recall f1-score rt
Move-forward 0 0.97 0.96 0.96 1725
Slight-Right-Turn 1 0.94 0.92 0.93 535
Sharp-Right-Turn 2 0.97 0.99 0.98 1709
Slight-Left-Turn 3 0.99 0.99 0.99 255
Avg/ total 0.97 0.97 0.97 4224
Confusion Matrix
Class 0 1 2 3
Move-forward 0 1655 27 43 0

Published By:
Retrieval Number: A1361109119/2019©BEIESP Blue Eyes Intelligence Engineering
DOI: 10.35940/ijeat.A1361.109119 4429 & Sciences Publication
International Journal of Engineering and Advanced Technology (IJEAT)
ISSN: 2249 – 8958, Volume-9 Issue-1, October 2019

Networks(CNN) are very effective tools for wall-following


control of robot navigation.The Confusion table of all
classifiers show maximum error in class1 i.e.slight right turn
movement,out of four motions of robot.Class3 i.e.slight left
turn, is very well-identified by all classifiers.We also notice
that proper data enhancement improves the classification
efficiency of all classifiers particularly in those cases where
data size is not sufficiently large.

REFERENCES:
1. Juang, C. F., & Hsu, C. H. (2009). Reinforcement ant optimized fuzzy
controller for mobile-robot wall-following control. IEEE Transactions
on Industrial Electronics, 56(10), 3931-3940.
2. Roslinamaria, A. R. (2005). Wall Following Mobile Robot.
3. Negishi, Y., Miura, J., & Shirai, Y. (2004, September). Mobile robot
navigation in unknown environments using omnidirectional stereo and
laser range finder. In Intelligent Robots and Systems, 2004. (IROS
2004). Proceedings. 2004 IEEE/RSJ International Conference
on (Vol. 3, pp. 2737-2742). IEEE.
4. Hsu, C. H., & Juang, C. F. (2013). Evolutionary robot wall-following
control using a type-2 fuzzy controller with species-DE-activated
continuous ACO. IEEE Transactions on Fuzzy Systems, 21(1),
100-112.
5. Brooks, R. A., & Connell, J. H. (1987). Asynchronous distributed
control system for a mobile robot. Mobile Robots I, 727, 77-85.
6. Li, T. H., Chang, S. J., & Chen, Y. X. (2003). Implementation of
human-like driving skills by autonomous fuzzy behavior control on an
FPGA-based car-like mobile robot. IEEE Transactions on Industrial
Electronics, 50(5), 867-880.
Fig.6. Precision-Recall curves for ANN and CNN-1D
7. Das, A. K., Fierro, R., Kumar, V., Southall, B., Spletzer, J., & Taylor,
C. J. (2001). Real-time vision-based control of a nonholonomic mobile
robot. In Robotics and Automation, 2001. Proceedings 2001 ICRA.
IEEE International Conference on (Vol. 2, pp. 1714-1719). IEEE.
8. Ando, Y., & Yuta, S. I. (1995, May). Following a wall by an
autonomous mobile robot with a sonar-ring. In Robotics and
Automation, 1995. Proceedings., 1995 IEEE International
Conference on (Vol. 3, pp. 2599-2606). IEEE.
9. Millan, J. R., Renkens, F., Mourino, J., & Gerstner, W. (2004).
Noninvasive brain-actuated control of a mobile robot by human
EEG. IEEE Transactions on Biomedical Engineering, 51(6),
1026-1033.
10. Ge, S. S., & Cui, Y. J. (2000). New potential functions for mobile robot
path planning. IEEE Transactions on robotics and automation, 16(5),
615-620.
11. Van Turennout, P., Honderd, G., & Van Schelven, L. J. (1992, May).
Wall-following control of a mobile robot. In Robotics and Automation,
1992. Proceedings., 1992 IEEE International Conference on (pp.
280-285). IEEE.
12. Freire, A. L., Barreto, G. A., Veloso, M., & Varela, A. T. (2009,
October). Short-term memory mechanisms in neural network learning
of robot navigation tasks: A case study. In Robotics Symposium
(LARS), 2009 6th Latin American (pp. 1-6). IEEE.
Fig.7. Precision-Recall curves for CNN-2D 13. Singh, S. K., Kumar, S., & Dwivedi, J. P. (2017). Compound fault
prediction of rolling bearing using multimedia data. Multimedia Tools
V. CONCLUSION and Applications, 1-18.
Figures(5, 6, and 7) along with Tables (I,II, and III) represent
the performance of 5 classifiers for wall following problem of AUTHOR’S PROFILE
robot navigation in the form of. Precision-recall curve and
Confusion matrix. The Precision -recall curve and Confusion Sandip Kumar Singh is presently employed as
matrix are well-established measures of the performance of an Associate Professor in the Department of
Mechanical Engineering at V B S Purvanchal
machine learning classifiers. Figure 6 shows that the Support Universit,y Jaunpur (U.P.), India. He completed
Vector Classifiers( SVC) and Multinomial Logistic his B.Tech degree from Kamla NehruInstitute
Regression(MLR) have 92 percent and 97 percent average of Technology( KNIT) Sultanpur(U.P.), and did
area under curve (AUC), respectively, for Precision -recall M.Tech. from National Institute of Technology
(N I T) Kurukshetra. He has done Ph.D. from
curve. Figure 7 shows the performance of artificial neural
Indian Institute of Technology (IIT BHU) Varanasi. His area of interest is
network (ANN), which is slightly less effective than Machine Learning and Structural Health Monitoring.
MLR.The two deep learning models CNN- 1D and .
CNN-2D perform excellently with the maximum area under
curve for Precision- recall curve as 97 percent and 98
percent, respectively. It exhibits that Convolution Neural

Published By:
Retrieval Number: A1361109119/2019©BEIESP Blue Eyes Intelligence Engineering
DOI: 10.35940/ijeat.A1361.109119 4430 & Sciences Publication

You might also like