A Report on Major Project

Deep Learning for Player Motion Tracking & Action Prediction in

Badminton
Submitted in partial fulfillment of the requirements for the degree of

BACHELOR OF TECHNOLOGY IN
INFORMATION TECHNOLOGY

2024 – 2025
By

SUSHRUT ANIL PATIL(12111187)

AMAN MANOJ MANAKSHE(12110571)
ADWAIT KIRAN MALI (12111521)
TEJAS SANJAY GAMBHIR(12111069)

Under the Guidance of

PROF. KAVITA SULTANPURE

DEPARTMENT OF INFORMATION TECHNOLOGY

Bansilal Ramnath Agarwal Charitable Trust’s

VISHWAKARMA INSTITUTE OF TECHNOLOGY

(An Autonomous Institute affiliated to Savitribai Phule Pune University) Pune – 411037
 BANSILAL RAMNATH AGARWAL CHARITABLE TRUST’S
VISHWAKARMA INSTITUTE OF TECHNOLOGY
(An Autonomous Institute affiliated to Savitribai Phule Pune University)

PUNE– 411037

C E R T I F I C A T E​
This is to certify that the Major Project titled

“Deep Learning for Player Motion Tracking & Action Prediction

in Badminton”
submitted by
Sushrut Anil Patil
Aman Manoj Manakshe
Adwait Kiran Mail
Tejas Sanjay Gambhir

is a bonafide work carried out by them under the supervision of me, Prof. Kavita
Sultanpure and it is approved by me for submission. Certified further that, to the
best of my knowledge, the report represents work carried out by the students as
the Major project, prescribed by the Savitribai Phule Pune University in the
academic year 2024-25.

Guide Head
(Prof. Kavita Sultanpure) (Prof. Dr. Premanand Ghadekar)
Department of Information Technology Department of Information Technology

Signature of External Examiner Date:

 ACKNOWLEDGEMENT

We, the members of Group No. 4, would like to sincerely thank everyone who
helped us finish this project without any problems. First and foremost, we would like to
express our sincere gratitude to PROF. KAVITA SULTANPURE, our mentor and leader,
for all of her help, advice, and encouragement during this project. Their knowledgeable
advice and helpful criticism have greatly influenced how our work has turned out.

We are also deeply grateful to the Information technology Department of the

Vishwakarma Institute of Technology, Pune for providing the resources and infrastructure
necessary to carry out this project. Access to facilities, data, and tools was essential to the
project's successful completion. Finally, we would like to thank our friends and family
for their patience and constant support as we completed this project. Their confidence in
us kept us going when things became tough.

This project was made possible by the collective efforts of everyone involved, and
we are truly thankful for the opportunity to learn, grow, and achieve this milestone
together.

Sushrut Anil Patil (Btech IT-A1-08)

Aman Manoj Manakshe (Btech IT-A-08)
Adwait Kiran Mail (Btech IT-A1-09)
Tejas Sanjay Gambhir (Btech IT-A-04)
CONTENTS​ Page No.

1.​ SYNOPSIS
1.1.​Project Title​ 1
1.2.​Project Domain​ 1
1.3.​Internal Guide​ 1
1.4.​Technical Keywords​ 1
1.5.​Problem statement​ 1
1.6.​Abstract​ 1
1.7.​Objectives​ 2
2.​ INTRODUCTION
2.1.​ Problem definition and scope​ 5
2.2.​ Motivation​ 6
3.​ LITERATURE REVIEW
3.1.​ Literature review​ 8
3.2.​ Comparison Table​ 11
4.​ PROJECT PLAN
4.1.​ Time and Scheduling​ ​ ​ ​ ​ ​ ​ 16
4.2.​ Work Breakdown Structure​ ​ ​ ​ ​ ​ 17
5.​ SOFTWARE REQUIREMENT SPECIFICATION
5.1.​ Functional​ ​ ​ ​ ​ ​ ​ ​ 18
5.2.​ Non-Functional​ ​ ​ ​ ​ ​ ​ ​ 21​ ​ ​
​ ​ ​ ​ ​
6.​ PROPOSED SYSTEM
6.1.​State Flow Diagram​ ​ ​ ​ ​ ​ ​ 24
6.2.​System Architecture Diagram​ ​ ​ ​ ​ ​ 25
6.3.​Sequence Diagram​​ ​ ​ ​ ​ ​ ​ 26​
6.4.​Use Case Diagram​ ​ ​ ​ ​ ​ 27
6.5.​ Feasibility Study​ ​ ​ ​ ​ ​ ​ ​ 29
6.2.1.​ Technical Feasibility​
6.2.2.​ Operational Feasibility
6.2.3.​ Economic Feasibility​
6.6.​Estimation​ ​ ​ ​ ​ ​ ​ ​ ​ 30
7.​ RESULTS AND DISCUSSION
7.1​ Results​ 33
7.2​ Training & Processing​ 35
8.​ DEPLOYMENT AND MAINTENANCE
 8.1​ Deployment Strategy​ 38
8.2​ Maintenance Strategy​ 38
8.3​ Scalability Considerations:​ 39
9.​ CONCLUSION AND FUTURE SCOPE​ 41

REFERENCES​ 43
LIST OF FIGURES

1.​ State Flow Diagram

2.​ System Architecture
3.​ Sequence Diagram
4.​ Use case diagram
5.​ List of stocks
6.​ Dataset Creation
7.​ Player Detection and Tracking
8.​ Shot classification Result

LIST OF TABLES
1.​ Comparison table of key research papers
2.​ Performance Metrics
3.​ Video analysis Performance
CHAPTER 1
SYNOPSIS
1.1​ Project Title
Deep Learning for Player Motion Tracking & Action Prediction in Badminton

1.2​ Project Domain

Deep Learning & Sports Analytics

1.3​ Internal Guide

Prof. Kavita Sultanpure

1.4​ Technical Keywords

Sports Analytics, Deep learning, Shot Classification , Yolo, Court Detection

1.5​ Problem Statement

Manually analyzing badminton matches is frequently laborious, slow, and prone to uneven
results and omitted details. In order to keep track of player positions, movements, and
shot types, coaches and analysts typically have to pause and rewind tape continuously.
This reduces the amount of valuable information that can be extracted from a match and
takes a lot of time.

Currently, no user-friendly technology exists that provides real-time, full analysis of

badminton matches. The majority of systems on the market are either too expensive, too
specialized, or too complicated, and they frequently need specialist knowledge to function.
As a result, many training facilities, trainers, and athletes do not have access to efficient
performance evaluation methods.

The goal of this project is to close that gap by providing a clever, user-friendly system that
can recognize player movements and shot types from match footage automatically. It
improves accuracy, decreases manual labor, and makes video analysis available to all
players, regardless of skill level.

1.6​ Abstract
Monitoring player performance and movement is becoming more and more important in
sports. Athletes and coaches depend on prompt and precise feedback to improve
performance and hone their tactics. However, this degree of analysis in badminton is still
​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 1
 mostly done by hand, using event logging and match observation, which is laborious and
prone to human mistake.

While some sports have adopted advanced tools for performance analysis, these
technologies are often either unsuitable for badminton or too expensive and complex for
widespread use. As a result, many coaches and players—especially at the training
level—lack access to tools that can quickly and effectively evaluate matches. This creates
a clear need for an accessible and affordable way to analyze the flow and dynamics of a
badminton game.

This project aims to bridge that gap. It offers a simple, user-friendly solution that allows
anyone—even those without technical expertise—to analyze badminton matches using just
video footage. The system helps users identify player positions, movements, and shot
types, ultimately supporting better performance and smarter game strategies.

1.7​ Objectives
To build a system to automate badminton match analysis using advanced deep learning
models.

A. Methodological Framework

The system employs advanced deep learning models and high-level computer vision to
analyze badminton matches automatically.
●​ YOLOv8n is used to detect players in real time.
●​ Court lines and zones of interest are detected using ResNet50.
●​ Shots are categorized into specific groups by r3d_18.To ensure precision and ease of use,
these models process video frames from a static camera view.

B. Data Handling and Tool Development

To extract flawless rally segments and label frames with the appropriate shot
classifications, custom tools are developed using Python and OpenCV.
●​ The information includes 10 types of shots (5 for each player).
●​ Annotation ensures high-quality training data, which improves model performance.
●​ This configuration eliminates redundant frames and addresses only relevant match
​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 2
 material.
C. User Interface and Accessibility
An intuitive and user-friendly Graphical User Interface (GUI) allows them to:
●​ View Post match videos.
●​ Examine outputs that consist of player tracking, movement analysis, and shot detection.
●​ The method enables access to valuable insight for non-technical users like players and
coaches without the requirement of technical skills.

D. Methodological Framework

The system reduces human labor and saves time in match analysis. Reliable information is
given to analysts and trainers for training strategy design. Individualized visual feedback
regarding their performance and location is given to the participants. In conclusion, it
improves badminton analysis's effectiveness, intelligence, and accessibility.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 3
 CHAPTER 2
INTRODUCTION

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 4
2.1​ Problem definition

Analyzing badminton matches manually is often a time-consuming and inconsistent

process. Coaches, analysts, and players frequently rely on traditional methods such as
pausing and rewinding videos to track player movements, positions, and shot types. This
manual effort is not only labor-intensive but also susceptible to human error and
oversight, limiting the depth and accuracy of insights gained from match reviews.​
​
Despite technological advancements in sports analytics, most existing systems for match
analysis are tailored to more mainstream sports and are often prohibitively expensive,
overly complex, or require specialized knowledge to operate. This renders them
inaccessible to many badminton players, coaches, and training academies—particularly at
the amateur or grassroots level.

A major issue is the lack of an intuitive, real-time tool that provides comprehensive
analysis specifically for badminton. Without such a solution, many players miss out on
valuable performance insights that could help refine technique, improve strategy, and
ultimately elevate their game. Additionally, the absence of automation and intelligent
tracking prevents analysts from generating consistent and actionable data at scale.

To address this gap, there is a clear need for an affordable, easy-to-use system that can
automatically analyze badminton match footage. Such a tool should accurately detect
player movements and classify shot types without requiring expert knowledge, making
advanced game analysis accessible to all levels of the sport.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 5
2.2​ Motivation

The demand for data-driven decision-making in sports has grown significantly, as athletes
and coaches strive to gain a competitive edge through performance optimization. While
advanced analytics are widely used in sports like football and tennis, badminton has
lagged behind in adopting such technologies due to limited accessibility and high costs.

With the global rise in interest in badminton—especially in countries like India, China,
and Indonesia—players at all levels are seeking ways to improve their performance
through deeper insights. However, most lack access to specialized analysts or expensive
systems that can provide detailed match breakdowns. This project aims to democratize
match analysis by offering a tool that brings professional-level insights to any user with a
video camera.

Manual video review is tedious and time-consuming, often requiring hours of playback
and subjective note-taking. This inefficiency can slow down training feedback loops and
lead to inconsistent evaluations. Automating this process helps players and coaches save
time while gaining accurate and actionable feedback, improving performance with every
match.

Moreover, as video content becomes more prevalent in training and scouting, the ability
to extract structured data from match footage will become a key differentiator. This
project seeks to empower not only athletes and coaches but also academies and
broadcasters with technology that simplifies, accelerates, and enhances match analysis.

By making badminton analysis smarter, faster, and more accessible, this project supports
the broader movement toward technology-enabled sports performance and opens new
opportunities for innovation in the game.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 6
 CHAPTER 3
LITERATURE REVIEW

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 7
3.1​ Literature review
[1] This paper investigates the dynamics of ice hockey strategies by integrating YOLOv8
for player detection and Gephi for network analysis in sports education. The study
leverages YOLOv8’s real-time object detection capabilities to track player movements
and interactions, while Gephi visualizes and analyzes the strategic patterns of play. This
combined approach offers a novel framework for understanding team dynamics,
facilitating more informed decision-making in coaching and training. By bridging
advanced computer vision and network analysis, the paper provides valuable insights into
the strategic complexities of ice hockey, with applications in sports education and
performance analysis.
[2] This paper examines the use of deep learning in sports skill learning through a
focused case study and performance evaluation. It explores how advanced deep learning
models can analyze and quantify skill execution in athletes, providing insights into
technique improvement and performance optimization. The study highlights the
integration of motion capture data and video analysis to train models capable of assessing
fine-grained actions in sports. The performance evaluation underscores the effectiveness
of these methods in real-world applications, offering a roadmap for enhancing training
methodologies and advancing skill learning in sports.
[3] This paper provides a comprehensive review of the applications of deep learning in
sports video analysis, highlighting its transformative impact on the field. It discusses
various deep learning techniques used for tasks such as player detection, action
recognition, event detection, and strategy analysis. The paper emphasizes the role of
advanced architectures like convolutional neural networks (CNNs) and recurrent neural
networks (RNNs) in extracting meaningful insights from complex sports scenarios. By
summarizing recent advancements and identifying current challenges, the study serves as
a valuable resource for researchers aiming to leverage deep learning in sports video
analytics.
[4] This paper introduces a deep learning model with an adaptive learning rate designed
for fault diagnosis. Fault diagnosis systems are critical for identifying and addressing
operational issues in complex systems. The proposed model adapts its learning rate
dynamically during training, enabling improved convergence and enhanced accuracy in

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 8
fault detection. By addressing challenges like overfitting and slow learning, the adaptive
learning rate mechanism ensures robust and efficient diagnosis performance across
diverse fault scenarios, contributing to the reliability and efficiency of modern industrial
systems.
[5] This paper explores the application of deep learning algorithms to automatically
detect unexpected accidents under challenging CCTV monitoring conditions in tunnels.
Accidents in tunnels often occur in low-visibility environments and are difficult to detect
due to poor lighting, camera quality, and environmental factors like smoke or dust. The
study introduces a robust framework that leverages advanced deep learning models to
enhance the accuracy and speed of accident detection. This approach aims to improve
monitoring efficiency and provide timely alerts, thereby enhancing safety and minimizing
response times in tunnel environments.
[6] This paper presents a method for accurate ball detection in field hockey videos using
the YOLOv8 algorithm. Ball detection in field hockey is challenging due to the ball's
small size, rapid movement, and frequent occlusions. The study demonstrates how
YOLOv8’s advanced object detection capabilities effectively address these challenges by
delivering high precision and real-time performance. By optimizing the model for the
unique requirements of field hockey, the research highlights its potential for enhancing
video analysis, aiding in performance evaluation, and supporting automated officiating
systems in the sport.
[9] This paper explores the application of computer vision to automate the notation
process for tactical analysis in badminton. By leveraging advanced image processing and
machine learning techniques, the study proposes an automated system capable of tracking
players' movements, shot types, and game strategies in real-time. The system facilitates
the extraction of detailed match statistics, offering insights into players' tactics and
performance. The paper highlights how computer vision technologies can significantly
reduce the manual effort involved in match analysis, providing coaches and analysts with
more efficient and accurate tools for evaluating player strategies and improving training
methodologies.
[10] This paper focuses on badminton video analysis using player tracking and pose
trajectory estimation. The study employs advanced computer vision techniques to track
players’ movements and estimate their pose trajectories during matches. By capturing key

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 9
body positions and movement patterns, the approach enables a deeper understanding of
player behavior and shot execution. The paper highlights how accurate tracking and pose
estimation can be used to analyze tactical decisions, optimize player training, and assist
in performance evaluation. This method offers a valuable tool for automating the analysis
of badminton games and enhancing coaching strategies.
[11] This paper investigates the application of motion recognition technology for
analyzing badminton players' actions in sports video images. Using advanced computer
vision and machine learning algorithms, the study focuses on recognizing and classifying
various player movements, such as jumps, smashes, and net shots. By processing video
images, the proposed system automatically detects and categorizes different actions,
enabling detailed performance analysis. The paper demonstrates the potential of motion
recognition to assist coaches and analysts in evaluating player techniques and improving
training methods, ultimately contributing to enhanced performance in badminton.
[12] This paper presents a badminton path tracking algorithm that combines computer
vision with ball speed analysis to track the trajectory of the shuttlecock during a match.
The study uses high-precision tracking methods to capture the shuttlecock's path, while
simultaneously analyzing its speed and movement patterns. By integrating these
techniques, the proposed algorithm offers enhanced accuracy in determining the
shuttlecock's flight trajectory and predicting its potential landing spot. This method aims
to improve game analysis, provide valuable insights for players and coaches, and
optimize training strategies by offering a detailed understanding of shuttlecock dynamics
in badminton.
[13] This paper explores the application of video analysis technology in badminton sports
training, emphasizing its role in enhancing player performance through detailed match
analysis. By utilizing advanced computer vision and machine learning algorithms, the
study discusses how video analysis can capture player movements, shot execution, and
tactical strategies. The paper highlights the benefits of using video data to break down
complex game situations, providing coaches and players with actionable insights for
technique improvement and strategy optimization. The research underscores the potential
of video analysis to revolutionize badminton training, making it more data-driven and
efficient.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 10
 3.2​ Comparison Table

Title Methodology Limitations Findings

Despite its effectiveness, The ODTS achieved

An application of a The researchers the system faced strong detection accuracy,
deep learning developed an Object challenges due to poor with Average Precision
algorithm for Detection and Tracking image quality in tunnel (AP) scores of 0.8479 for
automatic detection System (ODTS) that environments. Some cars, 0.7161 for people,
of unexpected uses a deep learning false positives occurred, and 0.9085 for fire. It
model, Faster R-CNN, to such as sunlight or successfully detected all
detect vehicles, people, vehicle lights being test incident types quickly,
and fires from CCTV misclassified as fires and demonstrating the
footage. A conventional long black objects as potential for real-time
object tracking pedestrians. Moreover, tunnel monitoring.
algorithm was integrated the study’s validation Additionally, the system is
to assign unique IDs to was limited to only a few scalable, allowing
objects and monitor their video cases, which may detection accuracy to
movement across not cover the full range improve as more training
frames. The system was of tunnel incident data becomes available
trained on a dataset of scenarios without modifying the
tunnel events and tested core algorithm.
on real accident footage
to detect incidents such
as wrong-way driving,
pedestrian entry, fire,
and stopped vehicles. It
could identify these
incidents in under 10
seconds..

A DeepLearning The authors propose the While the AdaFD The proposed AdaFD
Model with Adaptive AdaFD framework, method shows promising method demonstrates
Learning Rate for which utilizes a results, the paper does higher accuracy and better
Fault Diagnosis convolutional neural not extensively discuss adaptability compared to
network (CNN) as its its limitations. However, other common methods. It
core model. The training as with many deep also exhibits advantages
phase incorporates an learning approaches, in terms of training
adaptive learning rate potential limitations may efficiency, making it
strategy that include the need for large suitable for real-time fault
differentiates between amounts of labeled data diagnosis applications.
weight and bias for training and the
parameters, aiming to computational resources
accelerate the training required for model
process. deployment.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 11
Deep learning in This review paper Despite the The review finds that deep
sport video analysis: systematically compares advancements, the paper learning models,
a review traditional handcrafted notes that deep learning particularly those
approaches with deep approaches are combining CNNs and
learning techniques in data-intensive, requiring LSTMs, outperform
sport video analysis, large, annotated datasets traditional methods in
focusing on human for effective training. accurately recognizing
activity recognition. It The high computational complex actions in sports
examines various deep demands necessitate videos. These models
learning architectures, powerful hardware, have shown success in
including 2D-CNNs, which may not be applications like event
3D-CNNs, RNNs, and accessible to all detection in soccer, action
LSTMs, highlighting researchers. Additionally, recognition in tennis, and
their capabilities in most existing studies player tracking in hockey.
capturing spatial and focus on a limited range The integration of spatial
temporal features. The of sports, such as soccer and temporal feature
authors analyze recent and basketball, indicating extraction has enhanced
studies across multiple a need for broader the precision and
sports to assess the application across efficiency of sport video
effectiveness of these diverse sporting analysis .
models in performance disciplines.
evaluation and tactical
analysis.

Despite its promising The AHO-XGB model

Deep learning in The study introduced an results, the study faced demonstrated superior
sports skill learning: Artificial Hummingbird certain limitations. The performance compared to
a case study and Optimized XGBoost dataset was confined to traditional methods like
performance (AHO-XGB) model to participants from three Support Vector Machines
evaluation enhance sports skill universities, which may (SVM) and Naive Bayes
learning. This model affect the generalizability classifiers. Specifically, it
combines the XGBoost of the findings to broader achieved an accuracy of
algorithm with an populations. 92.6%, precision of
optimization technique Additionally, the model's 90.5%, and recall of
inspired by reliance on high-quality 94.3%. These results
hummingbird foraging sensor and video data indicate the model's
behavior, aiming to may pose challenges in effectiveness in accurately
improve predictive environments with predicting and assessing
accuracy in assessing limited technological sports skills, thereby
athletic performance. resources. offering valuable insights
Data were collected for coaches and athletes to
from 250 students and tailor training programs
20 faculty members effectively.
across three universities,
focusing on various
physical education
metrics.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 12
 The study integrates The study's scope was The analysis revealed
Exploring the
computer vision and limited to a specific distinct passing patterns
Dynamics of Ice social network analysis dataset, which may not and community structures
Hockey Strategies to examine ice hockey capture the full within teams. Players with
Using YOLOv8 and strategies. Utilizing the variability of real-world high weighted degree
Gephi in Sports McGill Hockey Player games. The accuracy of centrality were identified
Education Tracking Dataset, the pass detection based on as key facilitators in
researchers employed proximity may be offensive or defensive
YOLOv8 for player affected by occlusions or plays. Community
detection and ByteTrack rapid movements. detection highlighted
for multi-object tracking. Additionally, the subgroups of players who
Passing interactions approach requires frequently interacted,
were identified based on high-quality video data providing insights into
player proximity to the and computational team dynamics and
puck, constructing a resources, which might strategy execution.
directed network where limit its applicability in
nodes represent players some educational
and edges denote passes. settings.

Accurate ball The researchers While the system showed The proposed
detection in field developed a ball promising results, certain YOLOv8-based approach
hockey videos using detection system limitations were demonstrated high
YOLOV8 algorithm utilizing the YOLOv8 identified. The model's accuracy and efficiency in
object detection performance may be detecting the hockey ball
algorithm. They affected by challenging within video frames. The
compiled a specialized conditions such as model achieved
dataset by combining occlusions, varying significant performance
two annotated field lighting, and complex metrics, indicating its
hockey ball datasets backgrounds commonly effectiveness in real-time
from Roboflow present in field hockey ball detection. By
Universe, totaling 4,700 videos. automating the ball
images. The dataset was detection process.
divided into training
(87%), validation (9%),
and testing (4%) subsets.

The authors developed a The study acknowledges The implementation of the

Application of
computer vision-based several limitations in the dynamic window
Computer Vision to system to automate the proposed system. The approach significantly
Automate Notation notational process in accuracy of player improved the accuracy of
for Tactical Analysis badminton tactical tracking and player position
of Badminton analysis. The system segmentation may be assignments to court
employs motion tracking affected by factors such segments. The system
to identify player as occlusions, varying demonstrated the potential
positions throughout a lighting conditions, and to automate the tactical
match. camera angles. analysis process,
Additionally, the providing a more
system's performance is objective and efficient
contingent on the quality method compared to
and consistency of the traditional manual
video footage. notation. ​

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 13
 The authors developed The system's The proposed system
Video Analysis
an intelligent video performance may be effectively tracks player
Technology and Its analysis system for influenced by factors movements and
Application in badminton training, such as lighting recognizes actions during
Badminton Sports integrating computer conditions, camera training sessions. By
Training vision, pattern angles, and occlusions, automating the analysis
recognition, and image which can affect the process, it offers coaches
processing technologies. accuracy of motion detailed insights into
The system utilizes tracking and recognition. player performance and
particle filter prediction Additionally, the technique. This
for adaptive separation complexity of real-world automation facilitates
of moving targets and training environments more efficient training
tracks athlete may pose challenges in sessions and allows for
trajectories. maintaining consistent targeted improvements in
analysis quality player skills.

The study employs an The accuracy of tracking The proposed method

Badminton Video
adaptive Kalman filter to and pose estimation may effectively tracks players
Based on Player track players and the be compromised in and the shuttlecock, even
Tracking and Pose shuttlecock in broadcast scenarios with severe under challenging
Trajectory badminton videos. occlusions or low-quality conditions such as
Estimation. Trajectory confidence video inputs. Real-time occlusions and rapid
estimation and a particle processing requires movements. Pose
filter are utilized to significant computational estimation provides
enhance tracking resources, which might detailed insights into
accuracy, addressing limit its applicability in player actions, facilitating
challenges like occlusion resource-constrained a comprehensive analysis
and misdetection. Pose settings. Further of game strategies and
estimation is achieved refinement is needed to individual performance.
using a convolutional handle diverse The system's robustness
neural network (CNN), environmental conditions makes it suitable for
extracting key joint and to improve the real-time coaching and
positions to analyze system's scalability. performance evaluation.
player movements and
shot techniques

The authors developed a The study acknowledges The proposed template

Motion recognition
motion recognition that while the system matching robot model
technology of system for badminton performs well under demonstrated the ability to
badminton players in players using a template controlled conditions, its accurately and robustly
sports video images matching approach accuracy may be affected extract human objects in
within a geometric by factors such as video images. The system
algebraic framework. complex backgrounds, achieved a high accuracy
They employed instance varying lighting rate of 90.9%, even when
templates and a new conditions, and rapid subjects were lying on the
CFRdF method to movements. These ground, indicating its
calculate the similarity variables can introduce effectiveness in various
between horizontal challenges in template postures and scenarios.
plates and detected matching and human
image values. body extraction.

Table 1: Comparison table of key research papers

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 14
 CHAPTER 4
PROJECT PLAN

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 15
4.1​ Time and Scheduling

1. February 2025:
o Problem Identification: Understand the project scope and objectives.Identify
key problems to solve.
o Requirement Gathering: Gather technical and functional requirements.
Stakeholder interviews or research..
o Technology Stack Finalization: Choose frontend, backend, and database
technologies.

2.​ March 2025:

o Frontend Development: Design and implement UI using selected
frameworks.Build reusable components and layouts.
o Backend Development: Set up server, API routes, and database schemas
.Implement authentication and role-based access.
o Mid-month Sync: Conduct partial integration test to ensure frontend-backend
compatibility.

3. ​ April 2025:
o Frontend & Backend Integration: Link frontend actions to backend logic
and APIs.
o Testing & Debugging: Perform unit testing, integration testing, and security
checks.Fix bugs, optimize performance, and improve UI responsiveness.
o Deployment: Host the application on cloud/server. Final environment setup
and testing.

4.​ May 2025:

o Documentation: Write technical documentation and final report (e.g., system
workflow, code structure).Prepare presentation materials.
o Final Review & Submission:Submit project and present to evaluators.

4.2​ Work Breakdown Structure

1.​ Project Initiation

● ​ Project Planning:
○ ​ Define project scope, goals, and objectives.
○ ​ Develop project timeline and resource allocation.
○ ​ Identify potential risks and mitigation strategies.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 16
 ● ​ Team Formation:
○ ​ Assemble project team members with relevant skills and expertise.
○ ​ Define roles and responsibilities.
● ​ Requirement Gathering:
○ ​ Conduct stakeholder interviews and workshops.
○ ​ Analyze existing systems and identify functional and non-functional requirements.

2. System Design and Development

● ​ System Architecture:
○ ​ Design the overall system architecture, including frontend, backend, and database
components.
○ ​ Select appropriate technologies and frameworks.
● ​ Frontend Development:
○ ​ Design and implement user interface (UI) for data visualization and user interaction.
○ ​ Integrate frontend with backend services.
● ​ Backend Development:
○ ​ Develop backend services for data fetching, processing, and API integration.
○ ​ Implement data storage and retrieval mechanisms.
● ​ Machine Learning Model Development:
○ ​ Train and fine-tune machine learning models for greenwashing detection and stock
recommendation.
○ ​ Integrate models into the backend system.

3. System Testing and Deployment

●​ Unit Testing: Test individual components and modules.
●​ Integration Testing: Test the integration of different components.
●​ System Testing: Test the entire system to ensure it meets functional and non-functional
​ ​ requirements.
●​ User Acceptance Testing (UAT): Involve end-users to test the system and provide ​
​ ​ feedback.
●​ Deployment :Deploy the system to the production environment. Configure server and
​ ​ database settings.

4. Project Closure
● ​ Project Evaluation:
○ ​ Assess project performance against defined goals and objectives.
○ ​ Identify lessons learned and best practices.
● ​ Documentation:
○ ​ Update and finalize project documentation.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 17
 CHAPTER 5
SOFTWARE
REQUIREMENT
SPECIFICATION

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 18
5.1​ Functional
●​ Video Input and Processing:

○​ Upload badminton match videos via the GUI.

○​ Extract frames from the video using OpenCV for analysis.

●​ Player Detection:
○​ Use YOLOv8n model to:
■​ Detect players in video frames.
■​ Track player movements and positions throughout the match.

●​ Court Keypoint Detection:​

○​ Use ResNet50 model to:
■​ Detect court boundaries and key points (center, backlines).
■​ Assist in player positioning and movement tracking.

●​ Shot Classification:​
○​ Use r3d_18 model to:
■​ Classify types of shots (e.g., backhand drive, forehand netshot).
■​ Annotate classified shots on the video.

●​ Player Tracking:
○​ Continuously track individual players’ actions during rallies.

●​ Annotated Output Generation:

○​ Generate annotated videos showing:
○​ Player movements.
○​ Court keypoints.
○​ Classified shots and positions.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 19
●​ User Interaction via GUI (PyQt5):
○​ Upload videos.
○​ View real-time or post-processed analysis.
○​ Click on video segments to view detailed insights.
○​ Export analyzed results and videos.

●​ System Modes:
○​ User Interaction Mode: Users upload videos, receive real-time feedback.
○​ Processing Mode: System processes videos and outputs annotated clips.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 20
5.2​ Non Functional
●​ Performance Requirements:
○​ Frame processing time ≤ 0.5 seconds per frame.
○​ Accuracy for models ≥ 85% for both detection and classification.

●​ Reliability:
○​ System uptime ≥ 98%.
○​ Consistent detection accuracy across different lighting and video qualities.

●​ Availability:
○​ System should be available 24/7.
○​ Scheduled maintenance allowed, but with user notification.

●​ Security:
○​ Video data encryption during storage and transmission.
○​ User authentication required for data access.

●​ Portability:
○​ Compatible with all platforms that support Python and required libraries.
○​ Deployable in cloud environments for scalability.

●​ Hardware Constraints:
○​ Minimum of 16 GB RAM and NVIDIA GPU with CUDA for model inference.
○​ At least 50 GB of storage for video and result data.

●​ Memory and Storage:

○​ Sufficient memory required for frame-wise processing and storing outputs.
○​ Secure database storage for processed data.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 21
●​ User Characteristics:
○​ Designed for:
■​ Coaches and analysts.
■​ Intermediate to advanced players.
■​ Non-technical users (via GUI)

●​ Assumptions and Dependencies:

○​ Pre-trained models (YOLOv8n, ResNet50, r3d_18) are available.
○​ Users have appropriate hardware and Python environments.
○​ Labeled video datasets available for training and fine-tuning.

●​ Constraints:
○​ Dependent on input video quality for accurate performance.
○​ Real-time capabilities depend on hardware efficiency.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 22
 CHAPTER 6
PROPOSED SYSTEM

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 23
 6.1​ State Flow Diagram

​ ​ ​ ​ Figure 1: State flow diagram

The state flow diagram illustrates the sequential processing stages of the sports video analysis
system. Beginning in an IDLE state, the system branches into two parallel processing paths. The
right path handles input preprocessing, followed by keypoint identification, court detection, and
player detection stages. Once these detection processes are complete, the left path is triggered,
where shot detection occurs based on the identified elements. Results are then visualized and output
to the user, who can provide feedback on the analysis. This feedback loop allows for system
refinement before either returning to the IDLE state for new video processing or proceeding to the
Exit state when analysis is complete. This state-driven approach ensures systematic progression
through critical analysis stages while maintaining clear transitions between functional states

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 24
 6.2​ System Architecture Diagram

​ ​ ​ ​ Figure 2: System architecture

The proposed system architecture follows a layered approach for comprehensive sports video
analysis. Beginning with the Input Layer that handles video input and preprocessing, the system
progresses to a Detection & Classification Layer where court keypoints, players, and shot types are
identified. The Processing Layer then performs player tracking and movement analysis, feeding into
a User Interface Layer that presents shot analytics, player movement patterns, and position tracking.
This multi-layered pipeline culminates in a processed video output that integrates all analyzed data,
providing coaches and analysts with valuable insights for performance evaluation and tactical
decision-making.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 25
 6.3​ Sequence Diagram

Figure 3: Sequence diagram

The sequence diagram captures the process flow of the badminton analysis system, outlining the
interaction between key components to achieve the project’s goals. It begins with the system
receiving a fixed-camera video input, ensuring compatibility with the setup constraints. The
YOLOv8 model is activated to detect players and track their movements across the court, followed
by the ResNet50 model, which identifies court keypoints for spatial calibration. Subsequently, the
r3d_18 model processes the tracked movements to classify player actions such as backhand drives,
ensuring accuracy and alignment with user satisfaction goals.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 26
The system prioritizes enhanced model performance through fine-tuning for badminton- specific
contexts while maintaining cost efficiency and adherence to ethical AI standards. This workflow
ensures a seamless experience and precise action detection, aligning with the project's functional,
business, technological, and quality objectives.
​ ​ ​
6.4​ Use Case Diagram

​ ​ ​ ​ Figure 4: Use case diagram

1.​ Upload Gameplay Video

●​ Actor: User
●​ Description: User uploads a badminton gameplay video for analysis. The video must
have a fixed camera angle covering the entire court.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 27
2.​ Extract Video Frames
●​ Actor: AI System
●​ Description: The system extracts frames from the uploaded video using OpenCV to enable
further processing.
3.​ Detect Players
●​ Actor: AI System (YOLOv8n)
●​ Description: YOLOv8n model detects players in each frame and generates bounding boxes
around them.
4.​ Identify Court Boundaries
●​ Actor: AI System (ResNet50)
●​ Description: ResNet50 model identifies court keypoints and maps out court boundaries.
5.​ Classify Player Actions
●​ Actor: AI System (r3d_18)
●​ Description: The r3d_18 model analyzes frame sequences to classify player actions, such as
backhand drives or forehand lifts.
6.​ Overlay Results on Video
●​ Actor: AI System
●​ Description: The system overlays detected players, court mapping, and classified actions onto
the video frames for visualization.
7.​ View and Export Processed Video
●​ Actor: User
●​ Description: User views the processed video with overlays and can export it for further analysis.
8.​ Update Classification with User Input
●​ Actor: AI System
●​ Description: System incorporates user-provided annotations to reprocess frames and update
classifications for better accuracy.
9.​ Handle Low-Quality Input Video
●​ Actor: AI System
●​ Description: If the input video quality is poor and leads to detection errors, the system notifies
the user and suggests corrective actions or alternative input.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 28
 6.5​ Feasibility
6.5.1​Technical Feasibility:
●​ Hardware and Operating Environment : Development and testing were conducted on an Apple M2
Silicon Chip with 16 GB Unified Memory running macOS Monterey.This setup provided the necessary
computational power for both development and model training tasks.
●​ Software and Frameworks:Utilized robust frameworks such as PyTorch for deep learning and OpenCV
for computer vision tasks. These tools are well-supported and widely used in the industry, ensuring
reliability and community support.
●​ Model Implementations:Implemented ResNet50 for court keypoint detection, achieving 92.5%
accuracy. YOLOv8n was employed for player detection with a mean Average Precision (mAP) of
85.3%. R3D-18 was used for action classification, attaining an accuracy of 88.0%.
●​ System Integration:The integration of these models into a cohesive system demonstrates the project's
technical soundness and the team's proficiency in handling complex integrations.
6.5.2​Operational Feasibility:
●​ User Interface and Experience: Developed a user-friendly interface using React and TailwindCSS,
facilitating ease of use for end-users. The frontend effectively displays real-time data, including ESG
scores and stock history, enhancing user engagement.
●​ System Performance:The system processes videos efficiently, with average processing times of 30
seconds per video on CPU and 15 seconds on GPU for 5-minute 1080p videos. This performance
indicates the system's capability to handle real-time or near-real-time analysis, crucial for operational
deployment.
●​ Scalability and Maintenance:The modular architecture allows for easy scalability and maintenance.
Future enhancements can be integrated with minimal disruption to existing functionalities.`

6.5.3​ Economic Feasibility:

●​ Development Costs:Utilized open-source frameworks and tools, significantly reducing software
licensing costs. Development was conducted on existing hardware, minimizing additional capital
expenditure.
●​ Operational Costs:The system's efficient processing reduces computational resource requirements,
leading to lower operational costs. Maintenance and updates can be managed with a small technical
team, further containing expenses.
●​ Return on Investment (ROI):The system's capabilities in greenwashing detection and ESG scoring can
attract users interested in sustainable investing, opening avenues for monetization. Potential
partnerships with financial institutions or ESG data providers can generate additional revenue streams.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 29
6.6​ Estimation
6.6.1​Resources Required

●​ Hardware: The project requires high‑performance computing systems, including servers

with powerful GPUs and robust workstations, to support deep learning model training,
rapid inference for real‑time video processing, and iterative development. It also relies
on high‑resolution video capture devices for recording matches and ample local or
cloud‑based storage for archiving extensive video datasets and detailed annotations. A
reliable, high‑bandwidth, and secure network infrastructure is essential for transferring
large files, synchronizing data among team members, and protecting sensitive research
and proprietary model information.

●​ Software: The software ecosystem is built primarily on Python with deep learning libraries such
as PyTorch (or TensorFlow) integrated with OpenCV for video processing and complemented
by data handling libraries like NumPy and Pandas. Development is streamlined through tools
like PyCharm or Jupyter Notebook, while version control is managed with Git. Comprehensive
data annotation and visualization are supported by specialized annotation tools and libraries like
Matplotlib and Seaborn. Finally, a user-friendly interface is developed using GUI frameworks
like PyQt5, which enables coaches, analysts, and researchers to easily upload videos, visualize
results in real‑time, and export processed outputs.

●​ Other Resources: This project also depends on several key external resources beyond physical
hardware and software. Access to pretrained models and research data—such as weights for
YOLOv8n, ResNet50, and r3d_18—provides a crucial starting point for fine‑tuning on
badminton match data, while acquiring high‑quality, high‑resolution match videos and reliable
annotation tools ensures the dataset is both diverse and accurately labeled. Additionally, scalable
cloud platforms (like AWS, GCP, or Azure) are essential for managing heavy computation loads
and large volumes of data, and comprehensive supporting documentation and APIs further
enhance the system's integration capabilities and overall robustness.

●​ Cost Considerations: The overall cost structure encompasses several financial factors: initial
investments in hardware and software, including powerful GPUs, servers, workstations, and any
necessary licenses for proprietary tools, represent significant upfront expenses. Ongoing costs
include cloud computing expenses for scalable training, storage, and on‑demand inference, as
well as fees associated with acquiring premium video datasets or specialized annotation
services. Personnel costs, covering salaries for data scientists, computer vision engineers, and
supporting staff, are another key component, while periodic expenses for system maintenance,

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 30
software updates, and hardware upgrades must also be factored into the project's bud

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 31
 CHAPTER 7 ​
​RESULTS AND
DISCUSSION

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 32
 7.1​ Results​

●​ System Configuration:
○​ The experiments were conducted on a system equipped with the following
specifications:
• Processor: Apple M2 Silicon Chip
​ ​ ​ • RAM: 16 GB Unified Memory
• Operating System: macOS Monterey
• Frameworks Used: PyTorch, OpenCV

●​ Model Evaluations:

○​ Court Keypoint Detector (ResNet50):

■​ Model Overview:
●​ Architecture: ResNet50 (50-layer Residual Network)
●​ Purpose: Identify and locate critical court features such as sidelines,
baselines, service boxes, and net positions.
●​ Input: Static frames extracted from sports videos.
●​ Output: Coordinates of predefined keypoints on the court.
■​ Performance Metrics:

Metric Value Explanation

Accuracy 92.5% Out of all predicted

keypoints, 92.5% were
correctly positioned
relative to ground truth.

Precision 90.0% Of all detected keypoints,

90% were correct (low
false positives).

Recall 91.5% Of all actual keypoints,

91.5% were correctly
detected (low false
negatives).

F1 Score 90.7% Balanced metric combining

precision and recall (ideal
for assessing performance
on keypoint tasks).

​ ​ ​ ​ Table 2: Performance Metrics

■​ Strengths:
●​ High accuracy and recall make it reliable for consistent court tracking.
●​ Effective even under varying lighting and camera angles.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 33
 ■​ Limitations:
●​ Performance may degrade under occlusion (e.g., when players block
court lines).
●​ Accuracy depends on frame quality and consistent resolution.

○​ Player Detector (YOLOv8n):

■​ mAP (mean Average Precision): 85.3%
■​ Inference Time:
●​ CPU: 45 ms per frame
●​ GPU: 12 ms per frame
○​ Action Classifier (R3D-18):
■​ Accuracy: 88.0%
■​ Training Time:
​ ​ ​ CPU: 48 hours for 100 epochs
​ ​ ​ GPU: 8 hours for 100 epochs

●​ Video Analysis Performance:

The performance of the video analysis system was benchmarked using a standard video file
(duration of 5 minutes, resolution of 1920x1080). Results are as follows:

​ ​ ​ ​ ​ Table 3: Video analysis performance​

●​ Statistical Data:
To assess the overall performance of the models, multiple videos were analyzed under varying
conditions. The summarized results are:

Total Videos Analyzed: 50

• Average Processing Time per Video:
o CPU: 30 seconds
o GPU: 15 seconds ​

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 34
 7.2​ Training & Process:

●​ Dataset Creation: A comprehensive dataset was created to train the models

effectively. Videos of badminton matches were processed using OpenCV to
extract rallies, eliminating frames that focused on non-action areas, such as the
audience or referees. An annotation tool was developed with key binding
features to label shots into predefined classes, including backhand drive,
forehand clear, and others. This dataset ensured the inclusion of diverse
scenarios and player actions, forming the foundation for training the deep
learning models.

​
​ ​ ​ ​ Figure 6: Dataset Creation

●​ Player Detection:The YOLOv8n model was employed for player detection due
to its speed and accuracy. This model was trained to identify and localize
players on the court, providing bounding boxes that served as input for
subsequent processes. The model's performance was fine tuned using the
annotated dataset to achieve high precision, even in complex scenarios like
occlusions and overlapping players.

​ ​

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 35
​ ​ ​ ​ Figure 7: Player Detection and Tracking

​ ​ ​ ​ Figure 8: Shot Classification Result

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 36
 CHAPTER 8
DEPLOYMENT
AND
MAINTENANCE

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 37
 To ensure scalability, reliability, and maintainability, the following deployment and
maintenance plan has been developed for the backend of the sustainable investing platform.

1. Deployment Strategy:

● Containerization (Docker): The backend application will be containerized using

Docker to ensure consistent execution across different environments (development, staging,
production). This simplifies deployment and improves portability.

● Orchestration (Kubernetes or similar): A container orchestration system like

Kubernetes will be used to manage the Docker containers in a production environment.
Kubernetes provides automated deployment, scaling, and health monitoring, ensuring high
availability and resilience.

● Cloud Platform (AWS, Google Cloud, or Azure): The application will be deployed
to a cloud provider (e.g., AWS, Google Cloud, Azure) leveraging their managed Kubernetes
services (EKS, GKE, AKS). This provides scalability, reliability, and cost-effectiveness. The
choice of provider will depend on factors like cost, existing infrastructure, and specific
service needs.

● CI/CD Pipeline: A continuous integration and continuous deployment (CI/CD)

pipeline will be implemented using tools like GitLab CI, Jenkins, or GitHub Actions. This
automates the build, testing, and deployment process, reducing manual effort and
minimizing errors. The pipeline will include automated tests (unit, integration, and
potentially end-to-end tests) to ensure code quality and prevent regressions.

● Infrastructure as Code (IaC): Infrastructure will be managed using Infrastructure as

Code (IaC) tools like Terraform or CloudFormation. This allows for automated provisioning
and management of cloud resources, improving consistency and reproducibility.

2. Maintenance Plan:

● Monitoring and Logging: Comprehensive monitoring and logging will be

implemented using tools like Prometheus, Grafana, and Elasticsearch. This will enable
proactive identification and resolution of issues, ensuring application stability and
performance. Key metrics to monitor include API response times, database query
performance, and resource utilization.

● Regular Updates and Patches: The backend application, libraries, and underlying
infrastructure will receive regular updates and security patches to address vulnerabilities and
improve performance. A patching schedule will be established and strictly adhered to.

● Backup and Disaster Recovery: Regular backups of the database and application
code will be performed and stored in a geographically separate location. A disaster recovery
plan will be established to ensure business continuity in case of unforeseen events.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 38
 ● Performance Optimization: Regular performance testing and optimization will be
conducted to identify and address bottlenecks. This may involve database tuning, caching
optimization, and code refactoring.

● Security Audits: Regular security audits will be conducted to identify and address
potential security vulnerabilities. This includes vulnerability scanning, penetration testing,
and security code reviews.

● Documentation: Thorough documentation of the backend application, deployment

process, and maintenance procedures will be maintained to facilitate troubleshooting and
future development.

3. Scalability Considerations:

The architecture should be designed for scalability from the outset. This includes:

● Horizontal Scaling: The ability to easily add more instances of the application to
handle increased load. Kubernetes will be instrumental in this.

● Database Scalability: MongoDB's inherent scalability features will be leveraged,

potentially utilizing sharding for large datasets.

● Caching Strategy: A robust caching strategy (Redis) will minimize database load
and improve response times.

This comprehensive deployment and maintenance plan ensures the long-term stability,
reliability, and scalability of the backend application, supporting the platform's growth and
continued success. Regular review and updates to this plan are crucial to adapt to evolving needs
and technological advancements.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 39
 CHAPTER 9
CONCLUSION AND FUTURE SCOPE

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 40
Conclusion:

The Badminton Action Analysis System marks a significant achievement in automating the
recognition and classification of player actions and shots in badminton, providing valuable insights
for coaches and players. The system's design leverages powerful deep learning models like
YOLOv8n, ResNet50, and r3d_18 to detect players, track court keypoints, and classify shots with
remarkable accuracy. By combining state of-the-art models and techniques such as hybrid deep
learning architectures and frame level annotations, the system is capable of analyzing match
footage with minimal manual intervention. A key success of this project lies in its ability to
integrate different models for player detection, court keypoint extraction, and shot classification.
This multi-component approach ensures that the system can handle the complexities of badminton
gameplay, such as rapid movements and varied shot types. The use of fixed camera angles also
simplifies the task by reducing the challenges typically associated with tracking players in dynamic
camera settings.The system’s modular design allows for scalability and future improvements.
Future developments could enhance its capabilities by incorporating multi-camera support,
real-time analysis, and advanced shot prediction models, providing coaches and analysts with even
more actionable insights. Additionally, integrating wearable sensors and cloud-based solutions
could extend the system's utility beyond match analysis, enabling comprehensive performance
tracking over time. Overall, the Badminton Action Analysis System represents a robust and
efficient tool for analyzing badminton performance. Its combination of cutting-edge machine
learning models and its user-friendly interface make it an invaluable asset for coaches and athletes
seeking to gain a deeper understanding of player actions and strategies. With continuous vements
and future enhancements, this system holds great promise for further revolutionizing the way
sports analytics is approached in badminton.

Future Scope:

1.​ Multi-Camera Support:

○​ Description:The current system operates with a fixed camera angle, but integrating
support for multi-camera setups would increase flexibility. This would allow for
more comprehensive tracking of player movements and a deeper analysis of shot
execution from multiple perspectives.
○​ Future Enhancement:The addition of multiple camera feeds could lead to 3D
player tracking and a more detailed understanding of player positions and actions.
Techniques like multi-view geometry and camera calibration would be incorporated
to reconcile multiple perspectives.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 41
2.​ Real-Time Action Recognition:
○​ Description: The system presently processes video footage in batches or segments.
Real time recognition would enable immediate feedback for coaches or analysts,
enhancing decision-making during live matches or training.
○​ Future Enhancement: Developing real-time processing capabilities would require
optimization of the current models and implementation of low-latency inference
techniques. Streaming video inputs could be processed in real time to detect and
classify actions as they happen.
3.​ Advanced Shot Prediction and Strategy Analysis:
○​ Description: Expanding the system’s capabilities to not only detect shots but
predict them based on the player’s body language, position, and previous shots
would provide strategic insights.
○​ Future Enhancement: Integrating machine learning techniques like predictive
modeling, reinforcement learning, and historical match analysis could allow the
system to anticipate the next shot and even suggest strategies to improve gameplay.
This would make the system highly beneficial for coaching, where predicting future
actions is crucial.
4.​ Enhanced Customization and Personalization:
○​ Description: While the system currently categorizes shots and player actions,
allowing users to define and customize action categories and shot types would make
the system more adaptable for specific needs.
○​ Future Enhancement: The inclusion of a user-friendly interface where coaches or
analysts can add their own shot classifications or adapt the model for specific player
types could enhance the system’s usability. Customizable templates for shot
classifications or predefined action sets could also improve user experience.

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 42
 REFERENCE

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 43
[1]​.Kyu Beom Lee, Hyu Soung Shin ." An application of a deep learning algorithm for
automatic detection of unexpected
[2]​accidents under bad CCTV monitoring conditions in tunnels ".IEEE,2019.
[3]​Xiaodong Zhai1, Fei Qiao1."ADeepLearning Modelwith Adaptive Learning Rate for Fault
Diagnosis".IEEE,2021
[4]​[3] Keerthana Rangasamy,Muhammad Amir As'ari,Nur Azmina Rahmad,Nurul Fathiah
Ghazali."Deep learning in sport video analysis: a review".ResearchGate,2020
[5]​ Diandong Lian1 . "Deep learning in sports skill learning: a case study and performance
evaluation ".ResearchGate,2024.
[6]​Boyang Zhang, Tommy Löwendahl."Exploring the Dynamics of Ice Hockey Strategies Using
YOLOv8 and Gephi in Sports Education ".ResearchGate,2019
[7]​Suhas H. Patel ,Dr. Dipesh Kamdar ."Accurate ball detection in field hockey videos using
YOLOV8 algorithm".IJARIIT,2019
[8]​
[9]​ Kokum Weeratunga1,3, Khoo Boon How2, Anuja Dharmaratne1, Chris Messom1."
Application of Computer Vision to Automate Notation for Tactical Analysis of
Badminton".ICARCV 2014
[10]​ LI Yunwei1, Jiang Shiwei2."Video Analysis Technology and Its Application in
Badminton Sports Training ".ICAACE 2019
[11]​ Cai Cuiping."BAdminton Video Based on Player Tracking and Pose Trajectory
Estimation".IEEE,2021.
[12]​ Lei Fanga,∗ , Mingsi Sunb." Motion recognition technology of badminton players in
sports video images".Science Direct, 2018
[13]​ Shumin Zhang."Badminton Path Tracking Algorithm Based on Computer Vision and Ball
Speed Analysis". Hindawi,2019

​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 44