G Lavaraju* et al.

(IJITR) INTERNATIONAL JOURNAL OF INNOVATIVE TECHNOLOGY AND RESEARCH

Volume No.11, Issue No.2, February – March 2023, 10327 - 10329.

Planning to Implement Deep Learning

Techniques for Object Tracking
G.LAVARAJU PURNA SAI KUMAR REDDY PONNAPATI
Assistant Professor, Department of IT, Malla UG Student, Department of IT, Malla Reddy
Reddy College of Engineering and Technology, College of Engineering and Technology,
Hyderabad, T.S, India Hyderabad, T.S, India
BOYILLA SIVA NANDINI RAKESH NELLI
UG Student, Department of IT, Malla Reddy UG Student, Department of IT, Malla Reddy
College of Engineering and Technology, College of Engineering and Technology,
Hyderabad, T.S, India Hyderabad, T.S, India
Abstract: Detecting objects is a fundamental challenge in computer vision. As the same item might appear
significantly differently depending on factors such as orientation, illumination, backdrop, and occlusion,
detection can be challenging. Because of recent developments in deep learning and neural networks, we
no longer need to come up with new heuristics on the fly in order to solve this kind of issue. Using a
convolutional neural network (CNN) method, the "Object Detection" project can quickly and accurately
identify objects in images. The identification of objects in arbitrary positions and orientations is made
possible by a number of techniques, including "You Only Look Once" and other convolutional neural
networks. We begin by using a convolutional neural network that has already been trained.
Keywords: Object Detection; CNN; Multi-Scale Object Detection Intraclass Variance; Security;
I. INTRODUCTION: variety of criteria [2]. Deep convolutional neural
networks (DCNNs) have increased in significance
The primary emphasis of the majority of
for object identification as a result of the recent
organizations’ programmers was on developing the
surge in popularity of deep learning methods.
software and hardware interfaces for their own
When it comes to object recognition, deep
image processing systems. Although this has
learning-based systems have an advantage over
helped, it has not yet led to breakthroughs in
more conventional approaches since they can
tackling common problems like facial recognition,
automatically learn both low- and high-level visual
vehicle number plate recognition, road sign
attributes. Deep learning-learned picture
recognition, remote and medical picture analysis,
characteristics outperform manually created ones in
etc. Of the several deep learning network designs,
terms of realism. As a result, deep convolutional
one is commonly employed in these situations.
neural network-based object identification
When it comes to identifying images, many people
algorithms are the primary emphasis of this review
turn to convolutional neural networks (CNN). The
article, although more conventional object
model is provided with a variety of inputs,
detection techniques will be briefly covered [3].
including location, illumination, and orientation, all
This work provides a comprehensive overview and
of which are crucial for achieving a more accurate
analysis of recent advances in deep learning
forecast [1]. This yields the resulting image's
attempts to learn object detection systems, covering
corresponding "region of interest'. The detection
such topics as foundational networks, loss
accuracy across the bounding boxes is shown next
functions, ongoing training, classical image
to each discovered object's label. The
recognition structures, difficult issues, datasets,
Convolutional Neural Network (CNN) mobile
evaluation metrics, applications, and future
network is now being used in this research, while
research directions.
the single-shot multibox (SSD) technique is being
used to identify objects. Object detection is the II. PROBLEM STATEMENT:
process of identifying and pinpointing certain items
Even though Fast RCNN and Faster-RCNN
in a still or moving picture. Its vast usefulness has
employed additional methods to speed up
led to a rise in popularity in recent years. Recent
inference, they still couldn't compete with a real-
advances in deep learning-based object detectors
time system because of the time it took to compute
are reviewed in this article. Some of the most
regions. On the other hand, a YOLO system was
common foundational designs for recognition tasks
able to get around the bottleneck by combining the
are introduced, together with a brief summary of
region proposal and categorization processes into a
benchmark datasets and assessment measures.
single regression problem [4]. This allowed the
Modern, lightweight categorization methods for
system to normalize production from image pixels
edge devices are also covered. Finally, we evaluate
to bounding two box position information and class
the effectiveness of different designs using a
probabilities during the evaluation of each full

2320 –5547 @ 2013-2023 http://www.ijitr.com All rights Reserved. Page | 10327

 G Lavaraju* et al.
(IJITR) INTERNATIONAL JOURNAL OF INNOVATIVE TECHNOLOGY AND RESEARCH
Volume No.11, Issue No.2, February – March 2023, 10327 - 10329.
image. Due to the fact that the whole diagnosis first layer stands out as a complete convolution.
pipeline is a single network, it is possible to Images are resized and cropped before being fed
directly adjust it for improved detection into the model, among other basic actions. In data
performance end-to-end. Yolo is the first platform preparation, the raw data is transformed into a more
to attain a real-time identification benchmark of 45 usable, standard, and interpretable format. Numpy
frames per second (on GPU) and a MAP (mean and OpenCV are used in the suggested technique to
average precision) of 63.4 percent on VOC 2007; process video and picture data. Knowledge transfer
nevertheless, it still has a disadvantage when it and insight finding via encodings are at the heart of
comes to identifying tiny objects. data visualization, which in turn allows for the
transformation of abstract data into meaningful
III. PROPOSED METHODOLOGIES:
representations. The common objects in context
We propose this approach with a single-action database were used to pre-train the SSD MobileNet
multibox sensor structure to prevent the model. There are 90 distinct classes in this model.
aforementioned problems from ever occurring. We
developed a mobile-network model using Tensor
Flow and OpenCV for accurate and reliable object
detection. Using faster-anchor RCNN's box
proposal method in conjunction with multi-scale
features in the detection layer made it easier to spot
tiny items. Accuracy and precision are greatly
required because of the nature of the task this
software is designed for [5]. As was previously
indicated, the suggested model can recognize
around 90 distinct items. The model will be
specially trained with the other items as part of
upcoming improvements to improve its detection
capacity. The range of things that may be detected
by the MobileNet will be expanded via the use of
transfer learning, in which the employed network
will be taught new objects.
IV. ENHANCED SYSTEM:
The focus is mostly on the user interface. When .
creating a user interface, developers often turn to
the tKinter library. There's some descriptive
language there, with buttons to choose between live
object identification and object detection through
video, among others [6]. It also supports the import
of mp4 files for object detection in videos.
Detection in this project may be performed in real
time or after the user has uploaded an mp4 file for
analysis. In order to identify items, the project
employs a Single Shot Multi Box Detector
MobileNet that has already been pre-trained. Faster
execution and more accurate object identification Output Screens: Live Image Detections
are achieved by extracting and using a frozen V. CONCLUSIONS:
inference graph that contains the pre-trained
evaluations of the objects. The project relies on In this project, we use deep learning to make a real-
sessions, which enable users to run graphs time system that can automatically tell if the
(specifically, frozen inference graphs). Objects are network's inference graph, which is made up of the
identified in each frame of the live stream or the computed weights of objects in different categories,
video file, and once identified; a rectangle box is has been frozen. Labels have been supplied for
created around the item and labelled as determined GUT that is both concise and descriptive, making it
by the labelled map. Mobile Net, an efficient CNN simple to understand how to use it. This object
architecture, was developed specifically for use in identification software can both find objects in real
mobile and integrated vision applications as an time and in videos that have already been taken.
object detector. To create lightweight deep neural Each detected item is given a name, and the
networks, this design relies on depth-wise precision with which it was identified within the
separable convolutions. Mobile Net's foundational defined bounds is also supplied. The range of
layers are made up of depth-separable filters. The things that may be detected by the mobile network

2320 –5547 @ 2013-2023 http://www.ijitr.com All rights Reserved. Page | 10328

 G Lavaraju* et al.
(IJITR) INTERNATIONAL JOURNAL OF INNOVATIVE TECHNOLOGY AND RESEARCH
Volume No.11, Issue No.2, February – March 2023, 10327 - 10329.
can be expanded via the application of transfer
learning, in which the employed network is trained
with new objects.
REFERENCES:
[1] Agarwal, S., Awan, A., and Roth, D. (2004).
Learning to detect objects in images via a
sparse. part-based representation IEEE
Trans. Pattern Anal Mach Intell 26,1475-
1490 doi: 10.1109/TPAMI 2004 10

[2] Alexe, B., Deselaers, T., and Ferrari, V.

(2010) What is an object in Computer
Vision and Pattern Recognition (CVPR),
2010 IEEE Conference on (San Francisco,
CA IEEE), 73-80 doi: 10.1109/CVPR 2010
5540226.

[3] Wu, R.B. Research on Application of

Intelligent Video Surveillance and Face
Recognition Technology in Prison Security.
China Security Technology and Application.
2019,6: 16-19.

[4] Tian, J.X., Liu, G.C., Gu, S.S., Ju, Z.J., Liu,
J.G., Gu, D.D. Research and Challenge of
Deep Learning Methods for Medical Image
Analysis. Acta Automatica Sinica,2018, 44:
401-424.

[5] Jiang, S.Z., Bai, X. Research status and

development trend of industrial robot target
recognition and intelligent detection
technology. Guangxi Journal of Light
Industry, 2020, 36: 65-66.

[6] P. Viola and M. J. Jones, “Robust real-time

face detection,” International Journal of
Computer Vision, vol. 57, no. 2, pp. 137–
154, 2004.

2320 –5547 @ 2013-2023 http://www.ijitr.com All rights Reserved. Page | 10329