Deep Learning Approach Model for Vehicle Classification using Artificial Neural Network

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395 -0056
Volume: 04 Issue: 06 | June -2017 www.irjet.net p-ISSN: 2395-0072
© 2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 1418
DEEP LEARNING APPROACH MODEL FOR VEHICLE CLASSIFICATION
USING ARTIFICIAL NEURAL NETWORK
*1Ms. Vijayasanthi D., *2Mrs. Geetha S.,
*1M.phil Research Scholar, Department of computer Science Muthurangam Government Arts College
(Autonomous), Vellore, TamilNadu, India.
*2Assistant Professor, Department of Computer Science, Muthurangam Government Arts College (Autonomous),
Vellore.
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - The Vehicle detection is used to identify the
vehicles in any video or image file. The process of detection
of vehicles includes the object detection by considering the
vehicles as the primary object. By taking the images form
aerial or horizontal view and from road or parking using
surveillance cameras, the detection process can be initiated.
A vehicle recognition method based on character-specific
extremely regions (ERs) and hybrid discriminative restricted
Boltzmann machines (HDRBMs) is performed by top-hat
transformation, vertical edge detection, morphological
operations, and various validations. It proposed a novel
algorithm to identify the density of vehicles by using the
vehicle detection and classification algorithm by
implementing the hybrid deep neural network over the huge
dataset of video and images that are obtained from the
satellite images. For feature extraction Non-negative Matrix
Factorization (NMF) and SVM compression is used. Where,
compression is used to increase the response time for
detection and classification. The proposed model has been
designed for the vehicle position identification as well as the
vehicle type classification using the deep neural network.
The proposed model has been tested with a standard
dataset image for the result evaluation.
Key Words: Negative Matrix Factorization, hybrid
discriminative restricted Boltzmann machines, extremely
region, and classification method
I. INTRODUCTION
Vehicle detection and classification plays crucial
role in traffic monitoring and management. The
application of vehicle detection and classification is very
vast. Vehicle detection is used on roads, highways, parking
or any other place to detect or track the number of
vehicles present on the spot. This will help the
surveillance to judge the traffic vehicles, average speed
and category of vehicle. There are number of object
detection techniques are available to detect and classify
them. Object Detection is a method that finds instances of
world objects like pedestrians, faces, vehicles and
buildings in pictures or in videos. It uses extracted
features and therefore the learning algorithms for
recognizing the instances of object class. Applications that
uses object detection method are image retrieval, security,
surveillance, automatic vehicle parking systems. Object
detection uses numerous models: Feature primarily based
on object detection, SVM classification, and Image
segmentation. There are many classification algorithms
that are being utilized for the main aim of the vehicular
detection and classification. Primarily we are using the
probabilistic, non-probabilistic or square distance based
object detection and classification mechanisms. The
classification technique like Support Vector Machine
(SVM), co-forest , k-nearest neighbor, neural network,
random forest etc are being utilized for the vehicular
detection and classification. Using neural network one
may also learn and reconcile advanced non-linear
patterns. Neural network possesses artificially intelligent
bio-inspired mechanism that may be helpful for feature
extraction. Neural network is a feedback network
wherever the feedback is forwarded to neural network.
The high pace development of technologies
Predominantly image or video processing techniques
have enabled a number of application scenarios.
Visual traffic surveillance (VTS) based intelligent
transport system (ITS) is one of the most sought and
attractive application and research domains, which has
attracted academia-industries to enable better
efficiency. The significant application prospects of ITS
systems have motivated researchers to achieve a
certain effective solution. An efficient vehicle detection
and localization scheme can enable ITS to make
efficient surveillance, monitoring and control by
incorporating semantic results, like “X-Vehicle crossed
Y location in Z direction and overtaking A Vehicle
with Speed B”. Considering these needs, in previous
works [1,2], vehicle detection, tracking, and speed
estimation model were developed. However, the
further optimization could enable more efficient ITS
solution. Developing a novel and robust system to detect

and classify the vehicle simultaneously can be of
paramount significance.
Applications of vehicle detection:
 Vehicular density evaluation in the urban areas
for the traffic shaping
 Automatic parking system to decide which vehicle
type if allowed
 For automatic counting and information gathering
about the vehicles coming in or going out of the
parking lots
 Aerial surveillance of the vehicular objects in the
urban areas
 Vehicle tracking on the roads across the countries
II. LITERATURE SURVEY
Yi-Ling Chen et. al. [1] Proposed an intelligent and novel
Video Surveillance System for self driven vehicle detection
technique and Tracking in Clouds and the installation of
surveillance of video surveillance cameras is done to keep
the vehicle dataset containing the vehicles. For detecting
any suspicious threat, human interaction is needed. There
are lots of other potential security problems that are
detected using the help of automated methods. The
methods used to detect and classify the vehicles when
uncontrolled environment is there. Proposed models
performance can be evaluated by improvement in
accuracy.
Thomas Moranduzzo. [2] Has proposed the UAV
(unmanned aerial vehicle) detection technique for images
with a catalog-based approach. Existing systems work
with monitoring operation that some areas are classified
to make the detection of vehicle faster and robust.
Concurrently, to extract features of HOG, filtering
operations are used in vertical and horizontal. Then the
orientation value of possible 36 directions which is
actually the vehicle points that is computed by searching
the highest value of similarity measure and in the end
avoids duplicity, as unmanned aerial vehicle images data
has very high pixel resolution so there may be a possibility
that a car may be identified more than once. So, in the end
of HOG extraction same car are merged. The accuracy
performance of proposed system is higher number of
possible 36 directions of movement.
Sayanan Sivaraman.[3] has proposed an Integrated lane
method for vehicle tracking, detection and localization.
Proposed system developed the Synergistic approach to
fuse the vehicle tracking and lane for the assistance of
driver. The result of proposed model is obtained by the
improved performance of vehicle and lane tracking.
Detection of vehicle has achieved an adequate accuracy
level.
Sebastian Tuermer [4] has proposed Airbone vehicle
detection in very dense urban location by using the HOG
features with Disparity Maps. The main objective of
proposed model is to analyze and describe the chain of
integrated real time processing. The input dataset consist
the two subsequent images, a global DEM, exterior
orientation data and a database. Similar or overlapped
areas are extracted by region growing algorithm. After
then the remaining parts classification of input data is
conducted that is based on features of HOG. This will
produce the faster and accurate results.
Sayanan Sivaraman [5] has proposed a model for looking
at vehicles on the road. Authors discusses the detection of
vehicle based on vision, behavior analysis and tracking.
They define the algorithm for on road vision based
detection of vehicle and also the classification algorithm.
They classify the branch of vehicles which further refers to
spatiotemporal measurements and trajectories tracking.
The proposed model achieved improvement with high
accuracy that is effective with and trajectory tracking and
spatiotemporal measurements.
Thomas Moranduzzo [6] has proposed an algorithm for
Automatic Car Counting method for UAV images. Proposed
system includes multiple steps i.e. the first step is used for
the asphalted zones screening. So that the particular area
where vehicle is detected is restricted and may reduce the
false alarms. By using this method feature extraction is
done more accurately and effectively. In the end, the key
points extracted from vehicles belongs to same vehicle is
fused together to achieve "one key point -one car". The
accuracy result of positioning for vehicles counting and
the cars within 2cm can be obtained using real UAV scene.
Chen, Bo-Hao, and Shih-Chia Huang [7] have proposed
neural networks primarily based on extraction of moving
vehicles for surveillance to intelligent traffic. Proposed
model uses the moving vehicles that can be detected in
any resolution range
III. PREVIOUS IMPLEMENTATIONS
Vehicle detection is defined as detecting the
vehicles on the basis of parameters such as color, shape
and size. Vehicles are detected usually by extracting the
vehicle queues from the satellite images. The vehicles can
be detected with the help of neural network i.e.
convolutional neural network. The complete system is
trained in order to classify, locate and detect the objects in
images. Hence this can improve the accuracy of
classification, detection and localization. The network can
be applied at multiple locations in the image using the
sliding window technique. Then the system is trained to

produce prediction of the size and location of bounding
box. A technique is defined to perform object localization
with convolution network based segmentation.
The above table has been recorded with the
elapsed time for the vehicle recognition and vehicle
detection transactions performed in the proposed model.
The average classification time has been found around 234
seconds in the all 13 transactions to recognize the 13
vehicular objects in the simulation. Also the detection time
has been recorded from the simulation, which has been
recorded around 6 seconds on an average for the all 13
transactions. The proposed model have correctly
identified the all of the vehicular objects in the given test
image for the experiments. The experimental results have
shown the effectiveness of the proposed model in the case
of vehicle detection and classification. The proposed
model has been proved to be efficient and robust object
classification system. In the future, the proposed model
can be applied on some of the standard vehicular dataset.
The proposed model can be also tested with the video data
for the vehicular detection and classification purposes.
RESEARCH GAPS
• The existing model evaluates the overall density
of the vehicles but does not classify them in order
to evaluate the traffic more effectively over the
urban roads.
• The existing model for the vehicle detection lacks
in analyzing the density over the roads after
classifying and identifying the specific type of the
vehicles in order to prepare the traffic shaping
and planning to reduce the congestion across the
busy roads in the urban areas.
• The busy tunnels, where the congestion occurs
almost every day, the traffic shaping method can
be applied to allow the computed number of
vehicles per day in order to reduce the traffic.
• The block-wise processing for the estimation of
the vehicular class with neural network makes the
whole process slower and tedious due to the
inclusion of the slider window function. The
execution time can be reduced by using the
reduced feature component with fast
classification.
• The existing models are capable of vehicular
detection only and do not produce any of the time
series based vehicular traffic density and analysis.
The exiting model does not perform any vehicle
classification based on the size like whether the
vehicle is heavy or light. The system does not
create the vehicular analytical framework for the
vehicular detection, classification and time based
analytical study.
IV. SYSTEM IMPLEMETNATION
The proposed solution aimed the vehicle
detection and classification are the models utilized
primarily for the vehicular traffic surveillance, data
collection and relevant applications. The vehicular
detection and classification models require the
hierarchical approach for the template matching based
object detection with probabilistic or non-probabilistic
classification algorithms. There are several challenges
which occur for the implementation of the state of the art
system for the vehicular classification and modeling. In
this paper, we have proposed the new age model for the
vehicular detection and classification with high density
vehicular database. The proposed model is being
developed over the low frame rate cameras which two or
three frames per second. The proposed model will
evaluate the vehicular type and classify them properly in
order to evaluate the traffic density categorized in the
vehicular type. The study obtained from the proposed
model would be utilized for the traffic management policy
making by analyzing the rush hours and the reasons
behind the congestion during the rush hour. The optimal
steps could be taken during the rush hours, such as the
heavy weight carriers can be stopped from entering the
congested highways to maximize the average traffic
movement speed. The proposed model is expected to
improve the performance of the vehicular classification
over the performance measures of precision, accuracy and
recall. In the future, the proposed model would be realized
to achieve the goal of vehicular classification and the
detection in the captured frames from the traffic
surveillance cameras.
 Improved camera calibration method by detection
of two vanishing points – camera calibration error
reduced by 50 %.
 Novel method for scene scale inference
significantly outperforming automatic traffic
camera calibration methods (error reduced by 86
%) and also manual calibration method (error
reduced by 19 %) in automatic speed
measurement from a monocular camera.
 The results show that the automatic (zero human
input) method can perform better than laborious
manual calibration, which is generally considered
accurate and treated as the ground truth. This
finding can be important also in other fields than
only in traffic surveillance.

Fig: System Architecture
The proposed method is illustrated in Figure 1,
and is based on the four main steps. First, the image is
over-segmented into a set of regions by means of the
Mean-shift algorithm. These regions are taken as the likely
locations to inspect for cars since one region—or a small
group of them—may represent a car. The second step is
devoted to the feature extraction process, where a window
around the candidate region is given as input to a pre-
trained CNN for feature extraction. Third, a linear SVM
classifier is trained to classify regions into either a “car” or
“no-car” class. The result is a binary map representing a
segmentation of the UAV image into “car” and “no car”
classes. Finally, the binary map is fine-tuned by
morphological operations and the further inspection of
isolated cars.
-------------- (1)
We considered the Mean-shift algorithm in our study, as it
is a robust feature-space analysis approach that can be
applied to discontinuity preservation, smoothing,
clustering, visual tracking, mode seeking, and image
segmentation problems. The theoretical framework of the
Mean-shift is based on the Parzen window kernel density
estimation technique, where for a given set of data
samples {Xi}i= n d-dimensional space, the kernel density
estimator at sample X is given by, Where ck,d is a
normalization constant, h is the bandwidth, and k(.) is the
profile of the kernel
The Mean-shift procedure is an efficient way of locating
these zeros without estimating the density, since images
are represented as a spatial range joint feature space. The
spatial domain denotes the locations for different pixels,
whereas the
range domain represents the spectral signals for different
spectral channels. Thus, a multivariate kernel can be
defined for joint density estimation:
----------- (2)
Where ρ is an normalization parameter and h =
[hs, hr] at is produced by Mean-shift filtering. The use of
the Mean-shift segmentation algorithm requires the
selection of the bandwidth parameter h, which (by
controlling the size of the kernel) determines the
resolution of the mode detection. It can be noted that the
Mean-shift algorithm cannot segment very large
resolution images; however, we can divide the large image
into smaller parts, apply the Mean-shift algorithm to each
part separately, and combine the result.
Fig : 1(a) original image
Fig 1(b) Mean Shift Algorithm
Fig: 1(C) Regions Filtering

Nevertheless, one advantage of the Mean-shift
algorithm is that it provides an opportunity for the early
elimination of large areas of the image based on the size of
large regions. For example, as shown in Figure 1b, most of
the asphalt regions (like roads and parking lots) are
segmented in large regions, which can be easily removed
from the search space automatically by including only
regions that have a width or height close to the average
car size in the image. By applying this simple technique,
only the regions shown in Figure 1c needed to be included
in the search space.
Mean Shift Algorithm
A kernel is a function that satisfies the following
requirements:
1.
2.
Some examples of kernels include :
1. Rectangular
2. Gaussian
3. Epanechnikov
Kernel Density Estimation
Kernel density estimation is a non parametric way to
estimate the density function of a random variable. This is
usually called as the Parzen window technique. Given a
kernel K, bandwidth parameter h, Kernel density
estimator for a given set of d-dimensional points is
Feature Extraction
Deep CNNs are composed of several layers of
processing—each containing linear as well as nonlinear
operators—which are jointly learnt in an end-to-end way
to solve specific tasks .Specifically, deep CNNs are
commonly made up of convolution, normalization, pooling,
and fully Connected layers. The convolution layer is the
main building block of the CNN, and its parameters consist
of a set of learnable filters. Each filter is spatially small
(along width and height), but extends through the full
depth of the input image. The feature maps produced via
convolving these filters across the input image are then
fed into a non-linear gating function such as the rectified
linear unit (ReLU). Next, the output of this activation
function can be further subjected to normalization (i.e.,
local response normalization) to help in generalized.
Region Classification with a Linear SVM
During this step, we went through all regions in
the image and checked if they represented a car. To do
this, we extracted a window surrounding the concerned
region and passed it to a pretrained CNN for feature
extraction. Next, the feature descriptor was classified as
either a “car” or “no-car” using an SVM classifier. This last
step was trained on a collection of image samples for both
classes. The set of positive samples was manually
annotated in the training images, while the set of negative
samples was randomly selected from the remaining areas
of the training images. The window surrounding the
concerned region could be defined in two ways:
(1) as the bounding box of the region,
(2) as a window centered at the centroid of the region
with a given size. By inspecting the regions could clearly
see that for many small regions that represented parts of
the car (like the roof or the front windshield), taking the
bounding box may not have contained sufficient car
features for high quality detection. The second option
should yield better results. Furthermore, cars in images
can have any direction; thus, if rectangular windows are us
Fine-Tuning the Detection Result
The result of Step 3 was a binary map showing all
regions that were classified as car regions. In this last fine-
tuning step, we cleaned up the final map by applying three
extra operations: (1) applying a morphological dilation
operation on the detection map to smooth out region
boundaries and merge close-by regions; (2) filling holes
that may have appeared in detected regions; and (3)
inspecting small isolated regions to improve the detection
of isolated cars and potentially reduce false alarms. the
results show that the method achieved a high true positive
(TP) rate; however, there was also a relatively high false
positive (FP) rate. This was due to some small isolated
regions of the image containing car-like signatures. First, it
was noted that cars were usually parked in groups close to
each other, and most regions detected as cars were next to
each other. Hence,
such regions merged into larger regions in the final mask,
and only a few isolated regions remained spread across
the image. Some examples of this small isolated region
where it is clear that some of them indicate a real parked
car in an isolated area, and that there were many false
positives. To remove some of these false positives

Fig: Detection results in test image
Fig : Samples of small isolated regions where some are
isolated parked cars (true positive) while others are
false positives.
CONCLUSION
Developed an efficient method for the detection and
counting of cars in UAV images, which has the following
novel features: (1) reducing the search space significantly
compared to the sliding-window approach by using the
Mean-shift algorithm; and (2) the use of deep learning
approaches to extract highly descriptive features without
the need for huge amounts of training data through the use
of pre-trained deep CNN combined with a linear SVM
classifier. The experimental results on five UAV images show
that our method outperformed state-of-the-art methods,
both in terms of accuracy and computational time. However,
despite the high capability of the method to detect car
presence and numbers, it still has a high FP rate
REFERENCES:
[1]. Sermanet, Pierre, David Eigen, Xiang Zhang, Michaël
Mathieu, Rob Fergus, and Yann LeCun. "Overfeat:
Integrated recognition, localization and detection using
convolutional networks." arXiv preprint arXiv: 1312.6229
(2013).
[2] Pawlus, Witold, Hamid Reza Karimi, and Kjell G.
Robbersmyr. "Data-based modeling of vehicle collisions by
nonlinear autoregressive model and feed forward neural
network." Information Sciences 235 (2013): 65-79.
[3] Sivaraman, Sayanan, and Mohan Manubhai Trivedi.
"Looking at vehicles on the road: A survey of vision-based
vehicle detection, tracking, and behavior analysis."
Intelligent Transportation Systems, IEEE Transactions on
14, no. 4 (2013): 1773-1795.
[4] Wang, Jinjun, Jianchao Yang, Kai Yu, Fengjun Lv,
Thomas Huang, and Yihong Gong. "Locality-constrained
linear coding for image classification." In Computer Vision
and Pattern Recognition (CVPR).
[5] Zhou, Xi, Kai Yu, Tong Zhang, and Thomas S. Huang.
"Image classification using super-vector coding of local
image descriptors." In Computer Vision–ECCV 2010, pp.
141-154. Springer Berlin Heidelberg, 2010.
[6] Chen, Xueyun, Shiming Xiang, Cheng-Lin Liu, and Chun-
Hong Pan. "Vehicle detection in satellite images by hybrid
deep convolutional neural networks." Geoscience and
Remote Sensing Letters, IEEE 11, no. 10 (2014): 1797-
1801.
[7] Abdel-Hamid, Ossama, Li Deng, and Dong Yu.
"Exploring convolutional neural network structures and
optimization techniques for speech recognition." In
INTERSPEECH, pp. 3366-3370. 2013.
[8] Abdel-Hamid, Ossama, Abdel-rahman Mohamed, Hui
Jiang, and Gerald Penn. "Applying convolutional neural
networks concepts to hybrid NN-HMM model for speech
recognition." In Acoustics, Speech and Signal Processing
(ICASSP), 2012 IEEE International Conference on, pp.
4277-4280. IEEE, 2012.
[9] Sainath, Tara N., Abdel-rahman Mohamed, Brian
Kingsbury, and Bhuvana Ramabhadran. "Deep
convolutional neural networks for LVCSR." In Acoustics,
Speech and Signal Processing (ICASSP), 2013 IEEE
International Conference on, pp. 8614-8618. IEEE, 2013.
[10] Abdel-Hamid, Ossama, Abdel-rahman Mohamed, Hui
Jiang, Li Deng, Gerald Penn, and Dong Yu. "Convolutional
neural networks for speech recognition." Audio, Speech,
and Language Processing, IEEE/ACM Transactions on 22,
no. 10 (2014): 1533-1545.
[11] Kim, Jingu, and Haesun Park. "Sparse nonnegative
matrix factorization for clustering." (2008). [12] Hoyer,
Patrik O. "Non-negative matrix factorization with
sparseness constraints." The Journal of Machine Learning
Research 5 (2004): 1457-1469. [13] O'grady, Paul D., and
Barak A. Pearlmutter. "Convolutive non-negative matrix
factorisation with a sparseness constraint." In Machine
Learning for Signal Processing, 2006. Proceedings of the
2006 16th IEEE Signal Processing Society Workshop on,
pp. 427-432. IEEE, 2006. [14] Lin, Chuan-bi. "Projected

gradient methods for nonnegative matrix factorization."
Neural computation 19, no. 10 (2007): 2756-2779.
[15] http://www.standard.co.uk/news/london/sky-at-
night-amazing-photographs-from-londons-aerial-
photographer-9017596.html?action=gallery&ino=6.
[16] O. Masato, S. Nishina, M. Kawato, “The neural
computation of the aperture problem: an iterative
process”, Lippincott Williams and Wilkins, Vol. 14, pp.
1767-1771, 2003. [14] B. Zhang, “Reliable classification of
vehicle types based on cascade classifier ensembles”, IEEE
Transactions on Intelligent Transportation Systems, Vol.
14, pp. 322-332, 2013.
[17] N.C. Mithun, N.U. Rashid, S.M.M. Rahman, “Detection
and classification of vehicles from video using multiple
time-spatial images”, IEEE Transactions on Intelligent
Transportation Systems, Vol. 13, pp. 1215-1225, 2012.
BIOGRAPHIES
1. Ms. Vijayasanthi D., M.phil Research Scholar,
Department of computer Science Muthurangam
Government Arts College (Autonomous), Vellore,
TamilNadu, India.
2. Mrs. Geetha S., Assistant Professor Department of
Computer Science Muthurangam Government
Arts College (Autonomous), Vellore, TamilNadu,
India.

Deep Learning Approach Model for Vehicle Classification using Artificial Neural Network

More Related Content

What's hot (20)

Similar to Deep Learning Approach Model for Vehicle Classification using Artificial Neural Network (20)

More from IRJET Journal (20)

Recently uploaded (20)

Deep Learning Approach Model for Vehicle Classification using Artificial Neural Network