Special Issue - 2021 International Journal of Engineering Research & Technology (IJERT)

ISSN: 2278-0181
NCREIS - 2021 Conference Proceedings

Multiple Object Tracking using Deep Learning

with YOLO V5

Abhinu C G Aswin P
UG Student, Department of Computer Science UG Student, Department of Computer Science
and Engineering and Engineering
College of Engineering Kidangoor College of Engineering Kidangoor
Kottayam-686583, India, Kottayam-686583, India,

Kiran Krishnan Bonymol Baby

UG Student, Department of Computer Science UG Student, Department of Computer Science
and Engineering and Engineering
College of Engineering Kidangoor College of Engineering Kidangoor
Kottayam-686583, India, Kottayam-686583, India,

Dr. K S Angel Viji

Associate Professor, Department of Computer Science and Engineering
College of Engineering Kidangoor
Kottayam-686583, India,

Abstract - The MOT (Multiple Object Tracking) is an trained and is used as a model for the system which can detect
important tool in the modern world. It has various uses like object objects in different frames comparing to the objects that was
detection, counting objects, security tools ,etc. The Object
tracking is a prominent technology in image processing which has provided in the model by mapping the same pattern of model
a large future scope. The MOT has made significant growth in a in the frame [2].
few years due to deep learning, computer vision, machine YOLO(“You Only Look Once”), OPENCV,
learning, etc. This paper aims to provide a software solution that PYTORCH,COCO dataset, TKINTER with MYSQL(MySQL
keeps track of the objects so that it can handle object list and
count. By using YOLO “You Only Look Once” Technology with is optional),GPU are the methodology used to detect, count
the help of Pytorch, the system aims in object detection, tracking and track the objects in MOT.The proposed system uses the
and counting. Also unlike the general yolo object detection tool Latest YoloV5 which is used to detect the objects.YoloV5 uses
which detects all objects at the same time ,this MOT system also
pytorch classifier for training as well as detection. Yolo begins
detects only objects which are needed to be detected by the user
and thus helps in improving the performance of the system. its journey with darknet technology ,which was later
developed to yolov2 ,then yolo v3 and later to yolo v4 [9].And
Keywords - Multiple Object Tracking (MOT); YoloV5; now for easy building of object detection yolo v5 was
Deep Learning; Dataset/Model
introduced leading to better performance of object detection
I. INTRODUCTION [6].Yolo V5 is constructed with Pytorch Classifier in deep
Tracking is one of the necessary technologies needed for the learning and after object detection the opencv module is used
upcoming world. Tracking can broadly be divided into for inputting real-time or file format video input to the
multiple Object Tracking (MOT) and single object tracking. algorithm and also tracks, and counts the objects detected in
Multiple Object Tracking (MOT) plays an important role in the output obtained making the system an efficient MOT
solving many basic problems in computer vision [1].Tracking system. And also the Tkinter makes the MOT easy for user
multiple objects in videos requires detection of objects in interaction making it user to choose particular model according
individual frames and combining those across multiple frames. to the user requirements and only the particular model object
Many Computer Vision techniques have been used to build is detected.
MOT systems, and day to day the technology is growing The proposed system can be used for various object crowded
rapidly providing an area of opportunities called image environments for detection of particular classified objects
processing is done by providing a labeled dataset which is according to the environment. The MOT makes users take

Volume 9, Issue 13 Published by, www.ijert.org 47

Special Issue - 2021 International Journal of Engineering Research & Technology (IJERT)
ISSN: 2278-0181
NCREIS - 2021 Conference Proceedings

count of the same type of objects and also used to detect the Opencv does few applications like video image stitching,
particular object from the bulk crowd helping users to save navigation, Medical analysis, etc.
time in searching for a particular object.

II. RELATED WORK

A. Object Detection
Object detection consists of three stages: classification,
detection and segmentation [3].Classification is the to classify
the image with a unique id in order to identify it while object
detection. Detection is the next step where using the trained
model the object is detected in frames. This gives a good
perception for the system giving the idea about the image. This
step detects the objects using the model and provides the
location of the object in that frame. The next stage is the
segmentation where the detected object is described and
segmented for better understanding(commonly used
coordinate representation of rectangular detection box as
Fig. 1. Object Detection
shown in Fig. 1.).
III. PROPOSED SYSTEM
B. YOLO V5
The original author of yolo algorithm is Joseph RedMon. The system begins with login page and then after optioning,
When he started the yolo algorithm construction and when it taking video or providing recorded video as input to the system
didn't have significant progress in another author Alexey of object detection. Then the inputted data is processed into
Bochkovskiy published a paper on yolo and then after that a frames for detection. During detection the The Yolo V5 uses
series of yolo arrived which led to yolov2,yolov3 and then upto the Model or dataset to detect the object in the input data. By
yolov4. In the heat of yolov4 the Ultralytics LLC team on may using the model after detecting objects the detected objects are
30,2020 issued YOLOV5 [6].The Yolov5 took a move from classified and represented by using labels and bounding boxes
darknet to pytorch achieving 140 FPS in Tesla P100 where as around the detected objects and then it is processed into output
in yolov4 only 50 FPS.Yolo V5 has the same advantages and video format (As shown in Fig. 2.).
has almost similar architecture as yolo v4.Yet Yolov5 makes A. Input
it convenient to train and detect objects compared to yolov4. Using OpenCV we can access camera module and also we can
C. COCO Dataset add video files in different formats. Using OpenCV the the
COCO dataset is a scope object recognition, segmentation, and real-time video frame is collected from camera lenses. Before
captioning dataset. It begins with object segmentation where collecting the real-time video frame we assigned a Tkinter
image division takes place in order to discover image window prompt which signifies or collects from the user about
boundaries and objects in it. which dataset or model should be used for detection. Once the
It is used for labelling using bounding boxes in image. After input option is received the user option is sent to neural
that in recognition in context, a basic correlation architecture network module and the camera is enabled using OpenCV and
is represented between the image and the objects in it. After starts collecting the video frames from the camera lens.
this we collect or gather the same type of color or grey levels.
This is super pixel stuff segmentation where they help in
featuring important areas and also they can reduce the input
element for calculations.
D. Open CV
Opencv is an open source library for computer vision modules
like image processing, camera access,etc [4].Now Gpu is also
included in the Opencv module which is an essential element
for pytorch [5].Opencv technology developed in a fast rate and
supports many algorithms to make the algorithm
efficient,especially in image processing field. Opencv in
python supports libraries like numpy, matplot,utilis,etc. Fig. 2. General working system

Volume 9, Issue 13 Published by, www.ijert.org 48

Special Issue - 2021 International Journal of Engineering Research & Technology (IJERT)
ISSN: 2278-0181
NCREIS - 2021 Conference Proceedings

B. Neural Network(YOLO V5) D. Weights / Model

Here using the input received from the camera or video the Here the labelled dataset obtained from the framework must be
input data is classified into frames and each frame is sent to configured with “.yaml” (YAML Ain't Markup Language)
yolo detection algorithm with the model which user selected. extension format file which can be used to append the text
The model can be a predefined model that is, COCO dataset label to the algorithm. Once the YAML file is configured it is
model or we can create custom models for detection. Once the set up in algorithm and using pytorch the given dataset gets
detection is done, it is bounded with boxes and labels where trained using GPU according to the epochs given in algorithm
the object is found and is send to output section where the for training and with test dataset the testing of trained model
detected frames are collected and are then compressed into after completion also takes place, predicting the objects in test
output format. Before merging, the detected frames are used image. Once the objects are predicted the model is compressed
for tracking, counting and sorting using OpenCV and also for into the yolo model format which is configured using the pre-
better results DeepSORT is also used for sorting and tracking trained model that is using COCO dataset model. Once this is
of objects [8]. done the model file with the corresponding test result potted in
graphs and texts are written in output folder where the
C. Dataset
evaluation and testing of the model can be done by using it in
This field is used for creating custom dataset from raw images
a detection algorithm.
in order for creating a custom model which can be used for
The activity structure or system architecture of the proposed
detection. For this the first thing is used to collect the raw
MOT model is shown in Fig. 3.
images from various sources and create a dataset. Then from
the dataset images the objects must be annotated and labelled E. Integration
from the images. Here the above five modules are intersected to make a single
For this Python frameworks like “Labeling” is used for system where at front a login page using MySQL can also be
annotating and labelling of the objects. Once the annotating set up but it is considered as optional as opencv mostly works
and labeling is done then the dataset is split into train and test under a secure arena . After integrating the modules it is
images in percentage of 70% for train and 30% for test as it is compressed into an executable file which is in “.exe” file
the general ideal percentage used for training. Once this is format output, which can be used as a MOT application.
done it can send to yolo training algorithm where the dataset
can be trained and model can be created using COCO dataset
model.

Fig. 3. System Architecture of the proposed system

Volume 9, Issue 13 Published by, www.ijert.org 49

Special Issue - 2021 International Journal of Engineering Research & Technology (IJERT)
ISSN: 2278-0181
NCREIS - 2021 Conference Proceedings

IV. RESULTS AND PERFORMANCE EVALUATION

The performance of YOLO is measured using mainly three

terms: mAP, Precision, Recall [7].
mAP is a measure that combines recall and precision for
detecting the accuracy of the object, Where it is calculated with
the average precision value for recall value over 0 to 1 with
IOU that is intersection over union from 0.5 to 0.95.
AP = 𝛴𝑟=1
𝑅 𝑃
𝑟 ()
𝑅
mAP = 1 𝛴 𝐴𝑃 (2)
𝑁 𝑘
Precision is used to measure how accurately the objects are
predicted. Precision can describe how good the model is at
predicting the positive class.
Precision= 𝑇𝑃 (3)
( 𝑇𝑃 + 𝐹𝑃 )
Recall will define how good the objects or classes are detected.
Which means, recall or sensitivity calculates the proportion of
actual positives that are correctly found.
Recall = 𝑇𝑃 (4)
( 𝑇𝑃 + 𝐹𝑁 )
IOU stands for Intersection Over Union.
Fig. 4. Precision, Recall, [email protected], [email protected]:0.95, of 200 images of class
IoU is used to calculate between two boundaries of an image
key plotted in graph
to check whether they have overlapped or not,if so, to know at
what rate it is we use IoU. we will be predefining an IoU
threshold (say 0.5 in our case) in order to find whether the
detection was true positive or false positive.
IoU = 𝐴𝑟𝑒𝑎 𝑜𝑓 𝑂𝑣𝑒𝑟𝑙𝑎𝑝 (5)
𝐴𝑟𝑒𝑎 𝑜𝑓 𝑈𝑛𝑖𝑜𝑛
The Yolo Accuracy is found using mAP. In COCO mAP, a
101-point interpolated AP definition is used in the calculation.
AP corresponds to the average AP for IoU from 0.5 to 0.95
with a step size of 0.05.
Consider the custom dataset model of a single class, say key,
with a dataset of 200 images which was trained with the COCO
dataset with the “yolov5m6.pt” model which has 51-52
GFLOPS.It was trained under 200 epochs using Pytorch
Classifier using Tesla T4 GPU.The results of the main three
terms discussed above were plotted in a graph as shown below
in Fig. 4.
Fig. 5. PR curve of Key dataset model
The results clearly define that at a range of 0-100 images the
prediction was growing and mAP was increasing and while V. CONCLUSION
reaching the last images the mAP was almost near to one say In this work, we proposed a Multiple Object Tracking system
for the sample dataset key the Accuracy or mAP is using yolov5 that can detect objects which were trained and
approximately 95.39% while training as the Confidence also they can track and take count of objects in each frame.
threshold value given was 0.001.Thus we were able to train a This MOT system has various real-time applications like
good custom yolo model which can be used for detection. The detecting particular objects from object crowded
model quality is represented using PR curve shown below in environments, tracking a particular type of object or detecting
Fig. 5.While using this model for detection, using this data the a set of object classes or counting a particular object. The MOT
tracked object in each frame was occurring. The detected is efficient enough to detect objects even on CPU GPU. This
object from the crowd had mAP values from 0.2-0.9 according MOT requires a local system with GPU which may not be
to the clarity of the image in real time. affordable to all systems but with the help of free cloud sources
like Google Colab we can use GPU and also make a working

Volume 9, Issue 13 Published by, www.ijert.org 50

Special Issue - 2021 International Journal of Engineering Research & Technology (IJERT)
ISSN: 2278-0181
NCREIS - 2021 Conference Proceedings

MOT system which can custom train objects with raw-images.

And can be used in our local systems. Therefore using the
yolov5 algorithms which uses pytorch. The MOT system is
capable of tracking various objects which were trained. As a
sample the system was trained with a class of object key with
200 images as dataset and after 200 epochs the dataset formed
a yolo model with 95.39% mAP which was used for detection
.While the custom model was used the object detection and
tracking had an accuracy prediction in range of 20-90%
according to the clarity of image and appearance of object in
image. As this was done in CPU GPU the prediction was better
than expected but was not effective as like GPU as GPU
system provided a better and accurate prediction of the object
this is as the MOT requires to process a model which has
50+GFLOP data. Thus excluding that limitation. The MOT
using Deep Learning and yolov5 is executable according to
user needs.

ACKNOWLEDGMENT
We would like to convey our special thanks with gratitude to
our Tutors and Head of Department as well as our principal
who gave us a wonderful chance to do this project, which
helped us in involving a lot of Research and we are really
thankful to them.

REFERENCES
[1] Li Tan, Xu Dong, Yuxi Ma, Chongchong Yu ,“ A Multiple Object
Tracking Algorithm Based on YOLO Detection” ,IEEE, 2018
[2] ShiJie Sun, Naveed Akhtar, HuanSheng Song, Ajmal Mian, Mubarak
Shah ,” Deep Affinity Network for Multiple Object Tracking ”,
JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER
2017.
[3] HASITH KARUNASEKERA , HAN WANG , (Senior Member, IEEE),
AND HANDUO ZHANG “ Multiple Object Tracking With Attention to
Appearance, Structure, Motion and Size”,IEEE, 104423-104432
,VOLUME 7, 2019
[4] Shriharsha S. Veni,Ananda S. Hiremath,Mahalakshmi Patil, Mayuri
Shinde, Aishwarya Teli “ Video-Based Detection, Counting and
Classification of Vehicles using OpenCV ” ICICNIS 2020.
[5] Samira Karimi Mansoub,Rahem Abri,Anıl Hakan Yarıcı,” Concurrent
Real-Time Object Detection on Multiple Live Streams Using
Optimization CPU and GPU Resources in YOLOv3 ” , IARIA, 2019,
ISBN: 978-1-61208-716-0.
[6] Bin Yan , Pan Fan , Xiaoyan Lei , Zhijie Liu and Fuzeng Yang,” A
Real-Time Apple Targets Detection Method for Picking Robot Based on
Improved YOLOv5 ”, Remote Sens. 2021,1619.
[7] Fadhlan Hafizhelmi Kamaru Zaman,Syahrul Afzal Che Abdullah,
Noorfadzli Abdul Razak, Juliana Johari, Idnin Pasya, Khairil Anwar Abu
Kassim,“ Visual-Based Motorcycle Detection using You Only Look
Once(YOLO) Deep Network” , ICATAS-MJJIC 2020,
doi:10.1088/1757- 899X/1051/1/012004.
[8] Hasith Karunasekera, Handuo Zhang and Han Wang, ” Real Time
Multiple Object Tracking using Deep Features and Localization
Information ”, ICCA,IEEE,July 16-19, 2019.
[9] KangUn Jo, JungHyuk Im, Jingu Kim, Dae-Shik Kim, ” A Real-time
Multi-class Multi-object Tracker using YOLOv2 ”, IEEE
ICSIPA,Malaysia,
September 12-14, 2017.

Volume 9, Issue 13 Published by, www.ijert.org 51