IRJET- Blind Navigation System using Artificial Intelligence

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 05 Issue: 03 | Mar-2018 www.irjet.net p-ISSN: 2395-0072
© 2018, IRJET | Impact Factor value: 6.171 | ISO 9001:2008 Certified Journal | Page 601
Blind Navigation System Using Artificial Intelligence
Ashwani Kumar1, Ankush Chourasia2
1,2 Dept. of Electronics and Communication Engineering, IMS Engineering College, INDIA
---------------------------------------------------------------------***--------------------------------------------------------------------
Abstract - In order to provide the blind people hearable
environment, this project focuses on the field of assistive
devices for visual impairment people. It converts the visual
data by image and video processing into an alternate
rendering modality that will be appropriate for a blind user.
The alternate modalities can be auditory, haptic, or a
combination of both. Therefore, the use of artificial
intelligence for modality conversion, from the visual modality
to another.
Key Words: Hearable Environment, Visual Data, Image
Processing, Video Processing, Artificial Intelligence
1. INTRODUCTION
Imagine how is the life of a blind person, their life is full of
risk, they can't even walk alone through a busy street or
through a park. they always need some assistance from
others. They are also curious about the beauty of the world,
there is excitement to explore the world, and to be aware of
what is happening in front of them. Even though they can
find their own things without anyone's need. The predicted
cases will rise from 36 million to 115 million by 2050 for
blind peoples if treatment is not improved by betterfunding.
A growing ageing population is behind the rising numbers.
Some of the highest rates of blindnessandvisionimpairment
are in South Asia and sub-Saharan Africa. The percentage of
the world's population with visual impairments is actually
falling, according to the study. But because the global
population is growing and more people are living well into
old age, researchers predict the number of people with sight
problems will soar in the coming decades. Analysis of data
from 188 countries suggests there aremorethan200million
people with moderate to severe vision impairment. That
figure is expected to rise to more than 550 million by 2050.
This project contains three main parts, a raspberry pi 3
(powered by android things), camera and artificial
intelligence. When the person presses the button on device,
the camera module starts to take a pictures and analyze the
image using Tensorflow (open source library for numerical
computation ) and detect what is that picture is about, and
then using a speaker or headphone, the device will voice
assist the person about that picture 1).
1.1 Raspberry Pi
The Raspberry Pi was developed in the United Kingdom by
Raspberry Pi Foundation. It is a series of small single board
computers, basically developed to promote basic computer
science in schools and other developing countries.
The Raspberry Pi can be used for various purpose.
Depending on the user requirements it can be customized.
We are using Raspberry Pi for image and video processing.
The Raspberry Pi 3 uses a Broadcom BCM2837 SoC with a
1.2 GHz 64-bit quad-core ARM Cortex-A53 processor, with
512 KB shared L2 cache.
Raspberry Pi 3 Model B has1 GB of RAM. The Raspberry Pi 3
(wireless) is equipped with 2.4 GHz WiFi 802.11n (150
Mbit/s) and Bluetooth 4.1 (24 Mbit/s) based on Broadcom
BCM43438 FullMAC chip with no officialsupportforMonitor
mode but implemented throughunofficialfirmwarepatching
and the Pi 3 also has a 10/100 Ethernet port.
The Raspberry Pi may be operated with any generic USB
computer keyboard and mouse. ItmayalsobeusedwithUSB
storage, USB to MIDI converters, and virtually any other
device/component with USB capabilities. Other peripherals
can be attached to the various pins and connectors on the
surface of the Raspberry Pi 0.
The Raspberry Pi is connected to the system via a Rj45cable.
Raspberry pi consists of different slots for performing
different functions. Raspberry pi Foundation provided a
Raspbian operating system for Raspberry Pi. Python and
Scratch are the main programming language used, and also
support many other languages.
1.2 Raspberry Pi Camera Module
In April 2016, the original Camera Module was replaced by
Raspberry Pi Camera Module. The Camera Module consists
of Sony IMX219 8-megapixel sensor (compared to the 5-
megapixel Omni-Vision OV5647 sensor of the original
camera). Camera module can take video and still
photographs. Libraries bundled in the camera canbeusedto
create effects. It supports 1080p30, 720p60, and VGA90

video modes, as well as still capture. It attaches via a 15cm
ribbon cable to the CSI port on the Raspberry Pi.
The camera works with all models of Raspberry Pi 1, 2, and
3. It can be accessed through the MMAL and V4L APIs, and
there are numerous third-party libraries built for it,
including the Pi Camera Python library. The camera module
is very popular in home security applications, and in wildlife
camera traps.
1.3 Artificial Intelligence
The field of creation of intelligent machines that work like
humans and respond quickly, in computer science is known
as Artificial intelligence. The core part of AI research is
Knowledge engineering. Machines can react and act like
humans only when they have abundant information related
to the world. To implement knowledgeengineering,Artificial
intelligence should have access to objects, categories,
properties, and relations. To initiate common sense,
reasoning and problem-solving power in machines, it is a
difficult and tedious task. Machine learning is anotheroneof
the core parts of AI. Learning without any kind of
supervision requires an ability to identify patterns in
streams of inputs, whereas learning with adequate
supervision involves classification and numerical
regressions. Classification determinesthe categoryanobject
belongs to and regression deals with obtaining a set of
numerical input or output examples, thereby discovering
functions enabling the generation of suitable outputs from
respective inputs. Mathematical analysisofmachinelearning
algorithms and their performance is a well-definedbranchof
theoretical computer science often referred to as
computational learning theory.
Machine perception deals with the capability to use sensory
inputs to deduce the different aspects of the world, while
computer vision is the power to analyze visual inputs with a
few sub-problems such as facial, object and gesture
recognition.
Artificial neural networks (ANNs)
or connectionist systems are computingsystemsinspiredby
the biological neural networks. An ANN is based on a
collection of connected units or nodes called artificial
neurons. Each connection (analogousto a synapse)between
artificial neurons can transmit a signal from one to another.
The artificial neuron that receives the signal can process it
and then signal artificial neurons connected to it0.
In common ANN implementations, the signal at aconnection
between artificial neurons is a real number, and the output
of each artificial neuron is calculated by a non-linear
function of the sum of its inputs. Artificial neurons and
connections typically have a weight that adjusts as learning
proceeds. The weight increases or decreases the strength of
the signal at a connection. Artificial neurons have a
threshold. Only if aggregate signal crosses that threshold,
then the signal is sent. Artificial neurons are generally
organized in layers. Different layers have different functions
and perform different kinds of transformations on their
inputs. Signals travel from the first (input) to the last
(output) layer, possibly after traversing the layer’s multiple
times.
2. The model used for Image Classification:
Convolutional Neural Network (CNN)
A convolutional neural network is a class of deep, feed-
forward artificial neural networks that have successfully
been applied to analyze the visual image.
CNNs use a multilayer perceptron’s to obtain
minimal preprocessing. They are also known as space
invariant artificial neural networks (SIANN), due to their
shared-weight architecture and translation invariance
characteristics. The deep convolutional neural network can
achieve reasonable performance on hard visual recognition
tasks, matching or exceeding human performance in some
domains. This network that we build is a very smallnetwork
that can run on a CPU and on GPU as well.0.
CNN is composed of convolutional modules of the stack that
performs feature extraction. Each module has a
convolutional layer followed by a pooling layer. The last

convolutional module is followed by single or denser layers
that perform classification. The final dense layer in CNN
contains a single node for each target classinthemodel,with
a softmax activation function to generate avaluebetween0–
1 for each node (the sum of all these softmax values is equal
to 1). We can interpret the softmax values for a given image
as relative measurements of how likely it is that the image
falls into each target class.
Convolutional layers, which apply a specified number of
convolution filters to the image. For each sub-region,
mathematical operations are performed by layer to produce
a single value in the output feature map. Convolutional
layers then typically apply a ReLU activation function to the
output to introduce nonlinearities into the model.
The rectifier is an activation function defined as thepositive
part of its argument:
f(z)= z+ = max(0,z), where z is the input to a neuron.
A smooth approximation to the rectifier is the analytic
function f(z) = log (1+ exp z), which is called
the softplus function. The derivative of softplus is
f’(z)= exp (z) / (1+exp z) = 1/(1+exp(-z))
i.e. the logistic function.
Convolutional Layer: Applies 32 5x5 filters (extracting 5x5-
pixel subregions), with ReLU activation function.
We are using CIFAR-10 classification to classify RGB 32x32
pixel images. The reason CIFAR-10 was selectedwasthatitis
complex enough to exercise much of TensorFlow'sability to
scale to large models. At the same time, the model is small
enough to train fast, which is ideal for trying out new ideas
and experimenting with new techniques.
Model Architecture
The model CIFAR-10 is a multi-layer architecture consisting
of alternating convolutions and nonlinearities. These layers
are followed by fully connected layersleading into asoftmax
classifier 0. This model achieves a peak performance of
about 86% accuracy within a few hoursof training time on a
GPU. It consists of 1,068,298 learnable parameters and
requires about 19.5M multiply-add operations to compute
inference on a single image. By using CIFAR-10 Model the
images are processed as follows:
1. They are cropped to 32 x 32 pixels, centrallyforevaluation
or randomly for training.
2. They are approximately whitened to make the model
insensitive to dynamic range.
This is a good practice to verify that inputs are built
correctly.
Reading imagesfrom disk and distorting themcanuseanon-
trivial amount of processing time. To prevent these
operations from slowing down training, we run them inside
16 separate threads which continuously fill a Tensor
Flow queue.
Here is a diagram of this model:
Poolinglayers, which down sample the imagedata extracted
by the convolutional layers to reduce the dimensionality of
the feature map in order to decrease processing time. We
used max pooling algorithm, which extracts sub-regions of
the feature map (e.g., 2x2-pixel tiles), keeps their maximum
value, and discards all other values 0.
The most common form of pooling is Max pooling where we
take a filter of size F*F and apply the maximum operation
over the F*F sized part of the image.
If you take the average in place of taking maximum, it will be
called average pooling, but it’snot very popular.Ifyourinput
is of size w1*h1*d1 and the size of the filter is f*f with stride
S. Then the output sizes w2*h2*d2 will be:

W2= (w1-f)/S +1
h2=(h1-f)/s+1
d2=d1
Most common pooling is done with the filter of size 2*2 with
a stride of 2. As you can calculate using the above formula, it
essentially reduces the size of input by half 0.
Dense (fully connected) layers, which perform classification
on the features extracted by the convolutional layers and
down sampled by the pooling layers. In a dense layer, every
node in the layer is connected to every node in thepreceding
layer.
Dense Layer 1: 1,024 neurons, with dropout regularization
rate of 0.4 (probability of 0.4 that any given element will be
dropped during training)
Dense Layer (Logits Layer) 2: 10 neurons, one for each digit
target class (0–9).
Logits Layer, the final layer of our neural network is the
logits layer, which will return the raw values for our
predictions. The logit model is a regression model where
the dependent variable (DV) is categorical.wheretheoutput
can take only two values, "0" and "1", which represent
outcomes such as pass/fail or win/loss. Cases, where the
dependent variable has more than two outcome categories,
may be analyzed in multinomial logistic regression, or, ifthe
multiple categories are ordered, in ordinal logistic
regression 0.
Our final output tensor of the CNN, logits, has shape [batch
size, 10]
Generate Predictions, the logits layer of our model returns
our predictions as raw values in a [batch size, 10]-
dimensional tensors. Let'sconvert these raw valuesintotwo
different formats that our model function can return:
The predicted class for each example: a digit from 0–9.
The probabilities for each possible target class for each
example: the probability that the example is a 0, is a 1, is a 2,
etc.
Calculate Loss, for both training and evaluation, we need to
define a loss function that measureshow closelythemodel's
predictions match the target classes. To calculate cross
entropy, we use One hot encoding technique. One-hot is a
group of bits among which the legal combinations of values
are only those with a single high (1) bit and all the others
low (0). This cost that will be minimized to reach the
optimum value of weights.
Configure the Training Op, we configure our model to
optimize this loss value during training. We'll use a learning
rate of 0.001 and stochastic gradient descent as the
optimization algorithm 0.
This is a small network and is not state-of-the-art to buildan
image classifier but it’s very good for learning especially
when you are just getting started. For our training set,weget
more than 90% accuracy on the validation set. As we save
the model during training, we shall use this to run on our
own images.
3. CONCLUSIONS
The goal of this research is to provide the better image
processing using artificial intelligence. By using CNN image
classifier, we predict the correct answer with more than
90% accuracy rate. By doing so we achieved the state-of-art
result on the CIFAR-10 dataset. We also use the trained
model with real time image and obtained the correct label.
We integrated tensorflow in Android studio withourtrained
model. And it is deployed in raspberry pi with the help of
Android Things Operating System.
REFERENCES
1) Dean, Jeff; Monga, Rajat; et al. (9 November 2015).
"TensorFlow: Large-scale machine learning on
heterogeneous systems" (PDF). TensorFlow.org.
Google Research. Retrieved 10 November 2015.
2) Getting Started with Raspberry Pi; Matt Richardson
and Shawn Wallace; 176 pages; 2013; ISBN 978-
1449344214.

3) Raspberry Pi User Guide; Eben Upton and Gareth
Halfacree; 312 pages; 2014; ISBN 9781118921661.
4) Xilinx. "HDL Synthesis for FPGAs Design Guide".
section 3.13: "Encoding State Machines". Appendix
A: "Accelerate FPGA Macros with One-Hot
Approach". 1995.
5) Hinton, G.; Deng, L.; Yu, D.; Dahl, G.; Mohamed, A.;
Jaitly, N.; Senior, A.; Vanhoucke, V.; Nguyen, P.;
Sainath, T.; Kingsbury, B. (2012). "Deep Neural
Networks for Acoustic Modeling in Speech
Recognition --- The shared views of four research
groups". IEEE Signal ProcessingMagazine.29(6):82-
97. doi:10.1109/msp.2012.2205597
6) R. J. Williams and D. Zipser. Gradient-basedlearning
algorithms for recurrent networks and their
computational complexity. In Back-propagation:
Theory, Architectures, and Applications. Hillsdale,
NJ: Erlbaum, 1994
7) Hilbe, Joseph M. (2009), LogisticRegressionModels,
CRC Press, p. 3, ISBN 9781420075779.
8) Zoph, Barret; Le, Quoc V. (2016-11-04). "Neural
Architecture Search with Reinforcement
Learning". arXiv:1611.01578 
9) Hope, Tom; Resheff, Yehezkel S.; Lieder, Itay (2017-
08-09). Learning TensorFlow: A Guide to Building
Deep Learning Systems. "O'Reilly Media, Inc.".
pp. 64–. ISBN 9781491978481.
10) Sutton, R. S., and Barto A. G. Reinforcement
Learning: An Introduction. The MIT Press,
Cambridge, MA, 1998
11) Zeiler, Matthew D.; Fergus, Rob (2013-01-15).
"Stochastic Pooling for Regularization of Deep
Convolutional Neural Networks". arXiv:1301.3557
12) Lawrence, Steve; C. Lee Giles; Ah Chung Tsoi;
Andrew D. Back (1997). "Face Recognition: A
Convolutional Neural Network Approach". Neural
Networks, IEEE Transactions on. 8 (1): 98–
113. CiteSeerX 10.1.1.92.5813 
13) Krizhevsky, Alex. "ImageNet Classification with
Deep Convolutional NeuralNetworks".Retrieved 17
November 2013.
14) Yosinski, Jason; Clune, Jeff; Nguyen, Anh; Fuchs,
Thomas; Lipson, Hod(2015-06-22)."Understanding
Neural Networks Through Deep
Visualization". arXiv:1506.06579 
15) Graupe, Daniel (2013). Principles of Artifircial
Neural Networks. WorldScientific.pp. 1–. ISBN 978-
981-4522-74-8.
16) Dominik Scherer, Andreas C. Müller, and Sven
Behnke: "Evaluation of Pooling Operatiorns in
Convolutional Architectures for Object
Recognition," In 20th International Conference
Artificial Neural Networks (ICANN), pp. 92-101,
2010. https://doi.org/10.1007/978-3-642-15825-
4_10
17) Graupe, Daniel (7 July 2016). DeepLearrningNeural
Networks: Design and Case Studies.WorldScientific
Publishing Co Inc. pp. 57–110. ISBN 978-9r81-314-
647-1
18) Clevert, Djork-Arné; Unterthiner, Thomas;
Hochreiter, Sepp (2015). "Fast and Accurate Deep
Network Learning by Exponential Linear Units
(ELUs)". arXiv:1511.07289 
19) Yoshua Bengio (2009). LearningDeepArchitectures
for AI. Now Publishers Inc. pp. 1–3. ISBN 978-1-
60198-294-0.
20) Christopher Bishop (1995). Neural Networks for
Pattern Recognition, Oxford University
Press. ISBN 0-19-853864-2
BIOGRAPHIES
Ashwani Kumar, born in 1996,
India. He is currently pursuing
B.tech in ECE from IMS
Engineering College. His interestis
in Data Science and Machine
Learning.
Ankush Chourasia, born in 1996,
India. He is currently pursuing
B.tech in ECE from IMS
Engineering College. His interestis
in developing AndroidApplication.

IRJET- Blind Navigation System using Artificial Intelligence

More Related Content

What's hot (20)

Similar to IRJET- Blind Navigation System using Artificial Intelligence (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Blind Navigation System using Artificial Intelligence