Real-time Object Tracking and Counting with

YOLOv8 and ByteTrack

Shriraksha Channalli Sagar Mulagund
Electronics and Communications Eng Electronics and Communications Eng
KLE Technological University Hubli KLE Technological University Hubli
01fe20bei019@kletech.ac.in 01fe20bei042@kletech.ac.in
Sahana Ankad Ritesh Vernekar
Electronics and Communications Eng Electronics and Communications Eng
KLE Technological University Hubli KLE Technological University Hubli
01fe20bei040@kletech.ac.in 01fe21bei403@kletech.ac.in

Abstract—Object detection has several applications, including Overall, the automobile industry benefits from the use of
surveillance, autonomous driving, and robots. The key tasks vehicle identification and counting technology, which has
are object detection, tracking, and counting. In terms of object applications in automation, efficiency and safety.
detection and tracking, recent deep learning-based techniques
like YOLOv8 and ByteTrack have produced outstanding results. In object recognition, there are several applications for the
Nevertheless, counting objects is still a difficult task. With the critical functions of object detection, tracking, and counting.
help of supervision, we present a novel method in this work for While object tracking involves following the identified objects
tracking and counting objects using YOLOv8 and ByteTrack. The across multiple frames in a video, object detection involves
feature maps produced by YOLOv8 and ByteTrack are used in determining the presence and location of objects in an image
our suggested method to track the detected items over several
frames and precisely count them. We also include a supervision or video. On the other hand, object counting entails calculating
mechanism that raises the counting accuracy even higher. Our the quantity of items in a specific scene.
experimental findings show that, in terms of accuracy and speed, Various object detection and tracking benchmarks have
our suggested strategy performs better than current state-of- shown impressive results for deep learning-based methods.
theart methods. Index Terms—object detection, object tracking, Nevertheless, counting objects accurately is necessary for
object counting, deep learning, YOLOv8, ByteTrack, supervi
many applications, including surveillance and autonomous
I. INTRODUCTION vehicles, and object counting remains a difficult task. With
the help of supervision, we present a novel method in this
Vehicle detection and counting technology has a number of work for tracking and counting objects using YOLOv8 and
significant uses in the automotive industry. Advanced driver ByteTrack.
assistance systems (ADAS), that assist drivers in preventing
crashes with other vehicles on the road. For example, a vehicle II. RELATED WORK
with a forward-facing camera and vehicle identification soft- Numerous methods for object detection, tracking, and
ware will recognise surrounding vehicles, gauge their distance counting have been put forth. Popular deep learning-based
from it, and alert the driver if they are approaching too closely. object detection framework YOLO (You Only Look Once)
This may enhance traffic safety and assist in minimising acci- achieves real-time performance and high accuracy. Further
dents. Autonomous vehicles can also use vehicle recognition increasing the detection accuracy is YOLOv8, the most re-
and counting technology in addition to ADAS to help them cent version of YOLO. However, YOLO lacks the ability to
navigate through traffic and prevent crashes. Autonomous count and track objects. Popular object tracking technique
vehicles can decide how to navigate through traffic by utilising SORT (Simple Online Real-time Tracking) combines online
cameras and sensors to determine the presence and position monitoring with object detection. A SORT extension called
of other vehicles on the road. Numerous real-world uses exist Deep SORT (Deep Simple Online Real-time Tracking) makes
for vehicle detection and counting, such as traffic control, use of deep appearance traits to increase tracking precision.
surveillance, and security. It can also be utilised in smart cities Despite the fact that SORT and Deep SORT support object
to monitor traffic, ease congestion, and boost the effectiveness tracking, they do not support object counting. According to
of transportation. recent studies, adding supervision can increase the accuracy
The technology for detecting and counting vehicles is also of object counting. To calculate the number of items in an
helpful for fleet management. Managers can optimise routes, image,an approach that combines an area proposal network
boost productivity, and cut costs by keeping an eye on the and a counting network. On a number of benchmarks for
number and position of vehicles in a fleet. object counting, this method delivered state-of-the-art results.
A. ByteTrack: Multi-Object Tracking by Associating Every
Detection Box
They introduce the BYTE data association method for
multi-object tracking, which is straightforward yet effective.
Existing trackers can benefit consistently from the addition
of BYTE. We also suggest the powerful tracker ByteTrack,
which ranks first among all the trackers on the leaderboard
with 80.3 MOTA, 77.3 IDF1, and 63.1 HOTA on the MOT17
test set with 30 FPS. Because of its precise detection capa-
bilities and the assistance of associating low score detection
boxes, ByteTrack is extremely resistant to occlusion. It also
clarifies the optimal way to use detection results to improve
multiobject tracking. ByteTrack’s excellent precision, quick
speed, and simplicity can make it a popular choice in real
world applications. [1] Comparisons of several trackers are
shown in Fig 1.

Fig. 2. Fig. 2. The suggested YOLOv4 is compared to various cutting-

edge object detectors. YOLOv4 performs equally well as EfficientDet while
operating twice as quickly. increases the AP and FPS of YOLOv3 by 10
percent and 12 percent respectively.

C. Comprehensive Review Of Yolo: From Yolov1 To Yolov8

And Beyond

This paper aims to provide a thorough analysis of the

evolution of the YOLO framework, from the first YOLOv1
to the most recent YOLOv8, highlighting the most significant
innovations, differences, and advancements in each iteration.
The first part of the article examines the fundamental ideas
and design of the original YOLO model, which laid the
Fig. 1. Fig. 1. Comparisons of several trackers using MOTA-IDF1-FPS on groundwork for later developments in the YOLO family. Then,
the MOT17 test set. FPS (running speed) is the horizontal axis, MOTA is the from YOLOv2 to YOLOv8, we delve into the improvements
vertical axis, and IDF1 is the radius of the circle. Our ByteTrack outperforms and refinements made in each version. These enhancements
all prior trackers, achieving 80.3 MOTA and 77.3 IDF1 on the MOT17 test
set with 30 FPS running speed. cover a range of topics, including input resolution scaling, loss
function adjustments, anchor box modifications, and network
B. YOLOv4: Optimal Speed and Accuracy of Object Detection design. We seek to provide a comprehensive knowledge of
the evolution of the YOLO framework and its effects on
They provide a cutting-edge detector that outperforms all
object detection by looking at these advances.
competitive detectors in terms of speed (FPS) and accuracy
(MS COCO AP50...95 and AP50). The detector presented
can be trained and utilised on a standard GPU with 8–16 The paper illustrates the trade-offs between speed and
GB of VRAM, enabling a wide range of applications. The accuracy that have emerged over the development of the
original theory behind single-stage anchor-based detectors has framework and discusses the unique improvements of each
been validated. Numerous features have been verified, and YOLO version. This emphasises how crucial it is to take the
some have been chosen for use in order to increase the environment and requirements of particular applications into
accuracy of the classifier and the detector. Future research account when choosing the best YOLO model. We conclude
and development can make advantage of these aspects as best by speculating on the potential directions for more research
practises. [2] Comparisons of several models with YOLOv4 and development that will influence the development of real-
is shown in Fig 2. time object detection systems. [3] Comparisons of YOLO
models is shown in Fig 3.
 occlusions and tracks things that momentarily vanish from
view by comparing the appearance of objects across frames
using a Siamese network design.Additionally, it uses a global
trajectory optimisation algorithm to boost object tracking
precision over time. An open-source application called Super-
vision aids machine learning practitioners in visualising and
evaluating model performance. It operates by establishing a
connection to a cloud storage platform, like Amazon S3, and
then automatically obtaining model predictions and ground
truth labels. The user may then study model performance
indicators like precision, recall, and accuracy and rapidly spot
areas where the model is off. Supervision then displays this
information in an easy-to-use online interface. Additionally,
the tool has functions for searching and filtering through
sizable datasets as well as for annotating data to enhance
model performance. Comparisons between different trained
YOLOv8 is shown in Fig 4 and Fig 5.

III. PROPOSED METHOD

The three main steps of our proposed method are object
detection, object tracking, and object counting. To begin with,
we use YOLOv8 and ByteTrack to identify the objects in the
video’s opening frame. The bounding boxes and confidence
ratings of the detected objects are obtained. Second, we track
the detected objects over succeeding frames by using the
feature maps produced by YOLOv8 and ByteTrack. Based
on the objects’ mobility and appearance, we monitor them
using a correlation filter-based method. In order to determine
how many objects are present in each frame, we examine the
monitored objects’ trajectories. We also include a supervision
mechanism that raises the counting accuracy even higher.The
counting process is monitored by the supervisory mechanism,
which estimates the number of objects in each frame using an
external module.
The most recent version of YOLO is called YOLOv8 by
Ultralytics. YOLOv8 expands on the success of earlier editions
as a cutting-edge, state-of-the-art (SOTA) model by adding
new features and enhancements for improved performance,
flexibility, and efficiency. The whole spectrum of visual AI
tasks, including as detection, segmentation, posture estima-
tion, tracking, and classification, are supported by YOLOv8.
Because of its adaptability, YOLOv8 can be used in a variety
of contexts and applications.
Researchers at the Chinese artificial intelligence company
Megvii Technologies created the object tracking system called
ByteTrack. It tracks objects in video footage using deep
learning and has produced cutting-edge outcomes on a number
of benchmark datasets. A deep neural network is used by the
ByteTrack technique to first identify objects in the first frame
of a video. Then, it gives each object a special ID, and it
keeps track of it throughout the rest of the video by using a
motion model to predict where it will be in each succeeding
frame and updating its ID as necessary. The technique handles
A. Algorithm
Step 1:Load the video file and initialize a frame generator.
Step 2:Load the YOLOv8 model and fuse its layers.
Step 3:Initialize a BYTETracker instance with predefined
parameters.
Step 4:Initialize a Roboflow Supervision instance with start
and end
Step 5:Loop over the video frames and perform object detec-
tion on each frame using the YOLOv8 model.
Step 6:Filter out detections for classes that are not of interest
(e.g., pedestrians) and convert them to a format that can be
consumed by the BYTETracker algorithm.
Step 7:Pass the filtered detections to the BYTETracker algo-
rithm, which performs object tracking by associating detec-
tions with existing tracks.
Step 8:Annotate the video frame with the tracked object
bounding boxes and their labels.
Step 9:Increment the LineCounter with the tracked object
bounding boxes and annotate the video frame with the count.
Step 10:Write the annotated video frame to an output video
file.
B. Block Diagram

IV. RESULTS
We evaluate our proposed method’s performance on
YOLOv8: The YOlOv8 model trains and tests the outcomes
using the COCO dataset. To track and count the incoming
and exiting vehicles, we used Bytetracker and Supervision
Fig. 3. The above Fig 6 shows block diagram of Real-time object tracking and
counting with YOLOv8 and BYTETrack.The essential components of a object along with the pretrained YOLOv8 from Ultralytics. For
tracking and counting system are object detection using YOLOv8,BYTETrack more accurate and quick tracking and detection, we used
for vehicle track,Supervision for counting ,Image annotate,Line counter. GPU acceleration. We used YOLOv8 since it works well and
performs better than its predecessors. Real-time performance
C. Dataset is also achieved by our proposed method.
In this paper, we used the YOLO model, which trains and A. Input of Video file
assesses object identification and tracking models using the
Common Objects in Context (COCO) dataset. The COCO
dataset, which consists of more than 330,000 images and
2.5 million instances of objects categorised with 80 different
categories, is a commonly used benchmark in computer vision
research. It is suitable for a variety of applications, including
object identification, segmentation, and tracking since it offers
a wide range of object sizes, occlusions, and backdrops. We
trained and assessed our models using the COCO dataset, and
we presented the findings using the COCO evaluation criteria.
Some examples of the dataset with vehicle detection are shown
in 7 and 8.
 V. CONCLUSION
In this paper, we developed a novel method for counting and
tracking objects with the help of supervision, using YOLOv8
and ByteTrack. In terms of object counting precision and pro-
cessing speed, our suggested strategy delivered cutting-edge
results. We also showed how effectively adding a supervisory
system might significantly raise counting accuracy. Numerous
potential uses for our method include robotics, autonomous
vehicles, and surveillance.
VI. FUTURE WORK
Future research will examine the viability of incorporating
our suggested approach into actual systems, like security
cameras and self-driving cars. We also intend to examine the
efficacy of combining our suggested approach with other deep
learning-based object recognition and tracking techniques,
such as Efficient Det and SiamFC. In order to increase
counting accuracy, we also intend to look into the supervision
mechanism’s ideal design further.
VII. REFERENCES
1) Zhang Y, Sun P, Jiang Y, Yu D, Weng F, Yuan Z, Luo
P, Liu W, and Wang X (2021). ByteTrack: Multi-Object
Tracking by Associating Every Detection Box.
2) Bochkovskiy A, Wang C.Y. and Liao H.Y.M.(2020).
YOLOv4: Optimal Speed and Accuracy of Object De-
tection.
3) Terven, J., and Cordova-Esparza, D. (2021). A Compre-
hensive Review of YOLO: From YOLOv1 to YOLOv8
and Beyond.
4) Ultralytics,Glenn Jocher, Ayush Chaurasia, Alex Stoken,
Jirka Borovec, NanoCode012, Yonghye Kwon, Kalen
Michael, TaoXie, Jiacong Fang, imyhxy, Lorna, Zeng
Yifu), Colin Wong, Abhiram V, Diego Montes, Zhiqiang
Wang, Cristi Fati, Jebastin Nadar, Laughing, . . . Mri-
nal Jain. (2022). ultralytics/yolov5: v7.0 - YOLOv5
SOTA Realtime Instance Segmentation (v7.0). Zenodo.
https://doi.org/10.5281/zenodo.7347926