1.

​ INTRODUCTION

Physiotherapy plays an essential role in the rehabilitation process by helping individuals recover
from injuries, manage chronic conditions, and restore mobility through guided exercises.
Traditional physiotherapy methods require in-person supervision by trained healthcare
professionals to ensure correct exercise execution and monitor progress. However, the increasing
demand for rehabilitation services, combined with limited access to specialized facilities and
professionals, presents significant challenges. Factors such as time constraints, geographical
limitations, and the high cost of repeated clinical visits often hinder patients' ability to adhere to
rehabilitation programs effectively. These challenges emphasize the need for smart, automated
systems that can provide remote physiotherapy assistance, ensuring that patients perform exercises
correctly and consistently without continuous professional supervision.

Recent advancements in artificial intelligence (AI) and computer vision have opened new avenues
for developing automated physiotherapy tracking systems. Computer vision techniques,
particularly pose estimation, have become powerful tools in human movement analysis by
detecting and tracking skeletal key points from video streams. By integrating pose estimation with
AI-driven models, it is possible to classify exercises, count repetitions, and provide real-time
feedback on form accuracy. Such systems can enhance the accessibility and effectiveness of
physiotherapy by allowing patients to conduct rehabilitation exercises at home using standard
webcam setups. Despite these advancements, several challenges remain unresolved in AI-based
physiotherapy solutions. Existing models often rely on frame-by-frame analysis, disregarding the
sequential nature of exercise movements, which leads to inaccurate classifications. Moreover,
many systems depend on raw (x, y, z) joint coordinates, making them highly sensitive to variations
in camera angle, lighting, and user positioning. Addressing these limitations is crucial for
developing robust and reliable AI-assisted rehabilitation systems.

To overcome these challenges, this research presents "Intelligent Healing Spaces," an AI-powered
web-based physiotherapy tracking system that utilizes real-time pose estimation and deep learning
techniques for accurate movement analysis. The system leverages the Mediapipe framework for
detecting and tracking skeletal key points and integrates a hybrid architecture combining

1
Pimpri Chinchwad College of Engineering,PCCOE,Pune
Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM)
networks. The CNN component is responsible for extracting spatial features from pose data, while
the BiLSTM component captures the temporal dependencies inherent in sequential exercise
movements. By analyzing motion sequences in both forward and backward directions, the
BiLSTM network significantly enhances the accuracy of exercise classification and error
detection.

A key innovation in this research is the shift from using raw joint coordinates to a pose-based
feature engineering approach, focusing on joint angles and relative distances between body parts.
This method ensures camera perspective invariance, reducing the model's sensitivity to variations
in user positioning and environmental conditions. The system is trained on a diverse dataset that
includes both real-world and synthetic exercise videos, such as the Physiotherapy-Video dataset
from Kaggle and the InfiniteRep dataset. The incorporation of synthetic data enhances the model's
generalization capabilities, enabling it to perform reliably across different environments, lighting
conditions, and execution styles.

The proposed system is deployed as a web-based platform, allowing users to perform rehabilitation
exercises remotely with a standard webcam. The platform provides real-time feedback on exercise
form, counts repetitions, and offers corrective suggestions to improve execution accuracy. This
functionality is particularly beneficial for patients with mobility limitations or those residing in
remote areas, as it eliminates the need for frequent visits to rehabilitation centers. The system's
ability to operate in real-time with high accuracy not only enhances patient engagement but also
improves adherence to prescribed rehabilitation routines, which is critical for achieving effective
recovery outcomes.

Experimental results demonstrate the system's effectiveness in classifying multiple physiotherapy

exercises with high precision and accuracy. The model's ability to provide real-time feedback on
form and execution speed ensures that patients can correct their movements promptly, minimizing
the risk of further injury. Furthermore, the use of a hybrid CNN-BiLSTM architecture allows the
system to capture both spatial and temporal patterns in exercise movements, addressing the
limitations of existing frame-based classifiers that fail to account for sequential dependencies.

2
Pimpri Chinchwad College of Engineering,PCCOE,Pune
In addition to presenting the architecture and implementation details, this research also explores
the limitations of current AI-based physiotherapy systems, such as sensitivity to environmental
factors and limited generalization to diverse user profiles. By addressing these challenges through
innovative feature engineering and robust model architecture, Intelligent Healing Spaces represents
a significant step forward in the development of AI-assisted rehabilitation systems. The study not
only contributes to the expanding field of AI-driven healthcare but also highlights the potential of
integrating deep learning and computer vision techniques to transform traditional physiotherapy
practices.

In conclusion, the proposed AI-powered physiotherapy tracking system offers a promising solution
to the limitations of conventional rehabilitation methods by providing real-time feedback, high
accuracy, and accessibility. By bridging the gap between traditional physiotherapy and AI-powered
rehabilitation, Intelligent Healing Spaces has the potential to revolutionize the way physiotherapy
is delivered, making it more effective, accessible, and personalized. Future work will focus on
expanding the range of supported exercises, enhancing the precision of feedback mechanisms, and
exploring the integration of wearable sensors to provide a more comprehensive rehabilitation
experience.

​
1.1 OVERVIEW

The increasing demand for accessible and effective physiotherapy solutions has led to the
integration of artificial intelligence (AI) and computer vision technologies in rehabilitation.
Traditional physiotherapy often requires in-person supervision, which poses challenges for
individuals with mobility issues, time constraints, or limited access to healthcare facilities. To
address these limitations, AI-driven physiotherapy tracking systems have emerged, utilizing pose
estimation and deep learning techniques to monitor patient movements in real time.

This report presents Intelligent Healing Spaces, an AI-powered physiotherapy tracking system
designed to provide real-time feedback and movement analysis. The system leverages
Mediapipe-based pose estimation to extract key skeletal points from video input and utilizes

3
Pimpri Chinchwad College of Engineering,PCCOE,Pune
Bidirectional Long Short-Term Memory (BiLSTM) networks for sequence-based movement
classification. Unlike conventional models that analyze individual frames in isolation, BiLSTM
networks consider both past and future frames to improve classification accuracy, making them
well-suited for complex physiotherapy exercises.

A key innovation of this system is the use of joint angle features instead of raw coordinate-based
pose data. This approach enhances robustness to variations in camera angles and user positioning,
ensuring consistent and accurate movement tracking across different environments. The system is
implemented as a web-based platform, enabling users to perform rehabilitation exercises at home
while receiving real-time feedback on exercise accuracy and performance.

The primary objectives of this research are:

●​ To develop an AI-powered physiotherapy tracking system that accurately classifies

rehabilitation exercises.

●​ To integrate pose estimation and deep learning techniques for real-time movement analysis.

●​ To improve robustness by utilizing joint angle-based feature extraction.

●​ To enhance user accessibility through a web-based implementation for home-based and

clinical use.

Through extensive experimentation and evaluation, the system demonstrates high classification
accuracy, effective repetition counting, and reliable feedback on exercise form. The findings
contribute to the advancement of AI-assisted healthcare, offering a scalable and intelligent solution
for modern physiotherapy.

4
Pimpri Chinchwad College of Engineering,PCCOE,Pune
1.2 MOTIVATION

Physiotherapy plays a crucial role in the rehabilitation of patients recovering from injuries,
surgeries, or neurological conditions. However, traditional physiotherapy methods often require
in-person supervision, which can be challenging due to factors such as limited access to
physiotherapists, high treatment costs, and patient mobility constraints. Additionally, the
effectiveness of rehabilitation largely depends on the accuracy and consistency of exercises, which
may not always be monitored correctly in home-based therapy.

With the rise of Artificial Intelligence (AI) and Computer Vision, there is an opportunity to
develop an intelligent system that can automate movement tracking, provide real-time feedback,
and assist patients in performing physiotherapy exercises correctly. This project is motivated by
the following key factors:

Limited Access to Physiotherapists: Many patients, especially in remote or underprivileged areas,

do not have regular access to professional physiotherapy services. An AI-powered system can
bridge this gap by enabling remote monitoring. Patients performing exercises at home often lack
proper guidance, leading to incorrect movements that can delay recovery. A smart tracking system
can provide instant feedback to ensure proper form.

Technological Advancements in AI & AR/VR: The rapid growth of pose estimation techniques
and deep learning models makes it possible to develop an efficient and accurate physiotherapy
tracking system. Gamification elements and real-time performance tracking can increase patient
motivation and adherence to therapy routines.

Scalability and Cost-Effectiveness: A digital physiotherapy solution can significantly reduce the
dependency on frequent in-person consultations, making rehabilitation more affordable and
accessible to a larger population. By integrating pose estimation, deep learning, and real-time
tracking, this project aims to create an intelligent, scalable, and engaging physiotherapy system
that enhances rehabilitation outcomes while making physiotherapy more accessible to all.

5
Pimpri Chinchwad College of Engineering,PCCOE,Pune
1.3 PROBLEM STATEMENT AND OBJECTIVE

Problem Statement

Physiotherapy plays a crucial role in the recovery and rehabilitation of patients suffering from
injuries, post-surgical conditions, or musculoskeletal disorders. However, traditional physiotherapy
methods often require in-person consultations, which may not always be feasible due to limited
access to physiotherapists, time constraints, and patient compliance issues. Additionally,
home-based physiotherapy lacks real-time feedback, making it difficult for patients to ensure they
are performing exercises correctly.

There is a need for a web-based physiotherapy application that enables real-time motion tracking,
personalized exercise plans, and automated progress monitoring to improve rehabilitation
outcomes and accessibility.

Objective

The objective of this project is to develop a web-based physiotherapy application that leverages
computer vision and motion tracking to provide an interactive and effective rehabilitation
experience. The key goals include:

●​ To implement real-time motion tracking using pose estimation for accurate movement
analysis and feedback.
●​ To develop a device-compatible web interface that allows patients to perform guided
exercises from home.
●​ To enable automated progress tracking and generate data insights for physiotherapist
evaluation.
●​ To provide personalized and adaptive exercise plans tailored to individual patient needs.

This system bridges the gap between in-person therapy and remote rehabilitation by offering a
cost-effective, accessible, and intelligent physiotherapy solution.

6
Pimpri Chinchwad College of Engineering,PCCOE,Pune
1.4 SCOPE OF WORK

The development of a web-based physiotherapy application is centered on enhancing accessibility,

accuracy, and efficiency in rehabilitation by incorporating real-time motion tracking. The project
encompasses several key components. The system development phase involves designing a
responsive web interface compatible with desktops, tablets, and mobile devices, as well as
implementing secure backend systems for user authentication and storing patient data such as
progress, exercise history, and feedback. The frontend includes an interactive dashboard that
enables patients to access their exercise plans and receive real-time feedback on their performance.
In the motion tracking and pose estimation component, models like OpenPose, MediaPipe, or
TensorFlow PoseNet are integrated to track body movements and posture in real time, with built-in
error detection mechanisms that notify users of incorrect motions. The physiotherapy exercise
module offers a categorized library of exercises based on body region and injury type,
customizable plans tailored to individual needs, and adaptive difficulty levels that adjust with
patient progress. Progress monitoring and analytics are achieved through tracking key movement
metrics over time, generating insights and reports for physiotherapists, and providing visual
summaries to motivate patients. In terms of security and user management, the system ensures data
privacy through secure authentication, encrypted storage of health records, and role-based access
for both patients and physiotherapists. Finally, the testing and deployment phase includes
functional and usability testing to ensure a smooth user experience, performance optimization for
real-time processing, and deployment on a scalable cloud platform to guarantee accessibility and
reliability. Overall, the project aims to deliver a comprehensive, interactive, and user-friendly
physiotherapy solution that streamlines rehabilitation through intelligent tracking, personalized
care, and insightful progress analysis.

7
Pimpri Chinchwad College of Engineering,PCCOE,Pune
1.5 METHODOLOGIES OF PROBLEM SOLVING

To develop a robust web-based physiotherapy application, a structured methodology is followed to

ensure accurate motion tracking, personalized rehabilitation, and effective patient engagement. The
problem-solving approach consists of several key phases, beginning with problem analysis and
requirement gathering. This phase involves conducting thorough research on common
physiotherapy exercises and understanding patient rehabilitation needs. It also identifies the
limitations of traditional physiotherapy methods and existing digital solutions. Core functionalities
such as real-time motion tracking, progress monitoring, and personalized therapy plans are defined
during this phase to guide the system’s design and development.

The next phase focuses on technology selection. For frontend development, technologies such as
React.js or Angular are chosen to create a responsive and interactive user interface. The backend is
built using Node.js with Express.js to manage user authentication and handle database operations.
For database management, MongoDB or PostgreSQL is employed to store patient information,
exercise logs, and performance metrics. Motion tracking is enabled by integrating pose estimation
models such as MediaPipe, OpenPose, or TensorFlow PoseNet. The entire application is deployed
on scalable platforms like AWS, Firebase, or Heroku to ensure broad accessibility and system
scalability.

System design and development follow, covering several components. The user interface (UI) and
user experience (UX) design are crafted to provide an intuitive dashboard for both patients and
physiotherapists. An interactive exercise module with visual guidance is implemented, ensuring
compatibility across various devices. Motion tracking and pose estimation use webcam-based
analysis to monitor patient movements, detect body joints, validate posture, and provide real-time
feedback with error correction. The personalized therapy module stores predefined exercises with
detailed instructions and allows physiotherapists to customize plans based on individual patient
needs. It also adjusts difficulty levels based on progress tracking. Additionally, a progress
monitoring and analytics module records performance data, generates visual reports, and provides
feedback mechanisms to encourage patient adherence and improvement.

8
Pimpri Chinchwad College of Engineering,PCCOE,Pune
In the testing and validation phase, rigorous unit testing, integration testing, and usability testing
are conducted. Real-world testing with physiotherapists and patients helps validate the system’s
accuracy and usability. System performance is optimized to deliver real-time feedback with
minimal latency, ensuring a smooth user experience.

The final phase involves deployment and maintenance. The application is deployed on a scalable
cloud platform, followed by the implementation of regular updates and security patches.
Continuous user feedback is collected to enhance features and functionalities over time.

By following this structured methodology, the web-based physiotherapy application aims to

deliver a user-friendly, efficient, and data-driven rehabilitation solution. It ensures real-time
movement tracking, personalized therapy plans, and improved patient engagement, effectively
bridging the gap between traditional in-person therapy and remote digital solutions.

9
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 ​2. LITERATURE SURVEY

​
2.1 REVIEW OF RECENT LITERATURE

The study conducted by (Ullah et al., 2023) examines the utilization of metaverse technology in
the healthcare sector, emphasizing its potential advantages such as better patient involvement
and improved medical instruction. This study fills the void in existing research by examining the
difficulties and remedies associated with the integration of metaverse technologies in healthcare
settings. The study is expected to provide valuable insights on how to address and overcome the
technological limits, legislative obstacles, and privacy concerns that are involved with the
deployment of the metaverse. In summary, it offers significant suggestions for successfully
incorporating metaverse technology into healthcare procedures, providing direction for future
research and implementation endeavors in this new domain.
In his paper, (Bhugaonkar et al., 2022) explores the incorporation of augmented reality (AR),
and virtual reality (VR) technology in the field of healthcare. It highlights improved patient care
through increased involvement and remote accessibility, as well as more efficient medical
training through realistic simulations. The challenges encompass technology constraints,
legislative considerations, and the guarantee of fair and equal access. In summary, the study
offers a thorough analysis of the possible advantages, uses, and difficulties linked to
incorporating these technologies into healthcare provision.
The study conducted by (Chandrasekera & Yoon, 2018) explores the influence of Augmented
Reality (AR) and Virtual Reality (VR) on learning styles in the field of creative design. This
underscores a notable disparity in comprehending their precise impact on the acquisition of
creative design skills and underscores the necessity for thorough examinations of their actual
integration in educational settings. The study intends to fill these gaps in order to enhance our
understanding of how AR and VR might improve learning outcomes in creative design
disciplines.
The research conducted by (Fiske et al., 2019) examines the impact of artificial intelligence (AI)
on improving the metaverse, with a particular focus on its advantages in user engagement,

10
Pimpri Chinchwad College of Engineering,PCCOE,Pune
content generation, and adaptable encounters. The text addresses ethical issues related to data
privacy and prejudice, emphasizes the difficulties in achieving interoperability, and explores the
necessity of developing scalable AI systems. The paper highlights the significance of user
acceptance and explores potential areas for future research, thereby enhancing our
comprehension of the integration of AI into the metaverse.
The study conducted by (Sriati et al., 2023) evaluates the influence of AR/VR interventions on
mental health, emphasising their beneficial outcomes in terms of knowledge acquisition, attitude
enhancement, empathy cultivation, and reduction of stigma towards individuals with mental
diseases. The analysis reveals several gaps, such as the necessity for long-term studies on the
impacts, considerations for diversity, issues over scientific region, and difficulty in integrating
into current mental health services and education. These findings enrich our understanding of the
potential advantages and difficulties of using AR/VR therapy in mental health. They also
identify areas for future research to improve their effectiveness.
The study conducted by (Dutmer et al., 2019) explores the ethical consequences of integrating
embodied AI into mental healthcare. It highlights the possible advantages, such as the
introduction of novel treatment methods and the ability to cater to marginalized populations. The
text focuses on ethical considerations such as the prevention of harm and the transparency of
algorithms. It also advocates for additional research on the incorporation and acceptability of
embodied AI in therapeutic environments. This study highlights the importance of conducting
thorough research on various mental health conditions and continuously evaluating the
long-term impacts. This contributes to the development of ethical advancements in the treatment
of mental health.
Research on augmented reality (A.R.) has experienced a significant increase, leading to
comprehensive scientific metric studies aimed at monitoring its development and practical uses
(Borgohain et al., 2022). Advancements in methodology have made it possible to conduct
thorough bibliographic analysis, which have shown trends such as the dominance of conference
proceedings and articles as the main types of publications. The presence of recurring keywords
such as A.R. and virtual reality highlights the main themes in academic pursuits. In terms of
geography, the United States stands out as a key center for A.R. research, with Germany.

11
Pimpri Chinchwad College of Engineering,PCCOE,Pune
2.2 GAP FINDINGS / COMMON FINDINGS FROM LITERATURE

Common Findings From Literature

Emerging technologies such as the metaverse, augmented reality (AR), and virtual reality (VR) are
increasingly being adopted in the healthcare sector, as highlighted by studies like those of Ullah et
al. (2023) and Bhugaonkar et al. (2022). These technologies offer numerous benefits, including
enhanced patient engagement, improved remote accessibility, and more effective medical training.
However, their implementation is not without challenges. Technological limitations, legal
complexities, and data privacy concerns are consistently cited as obstacles (Bhugaonkar et al.,
2022; Fiske et al., 2019; Dutmer et al., 2019). Additionally, there is a growing emphasis on the
need for ethical guidelines, transparency in AI algorithms, and fairness in healthcare access.

Artificial Intelligence (AI) plays a crucial role in advancing these digital healthcare solutions. It
enhances user interaction, supports personalized content creation, and enables adaptive
experiences within metaverse environments (Fiske et al., 2019). Specifically, AI-driven
physiotherapy applications have demonstrated effectiveness in improving posture correction and
rehabilitation outcomes (Rahman et al., 2021). In mental health care, AR/VR interventions have
shown promising results in increasing knowledge, fostering empathy, and reducing stigma (Sriati
et al., 2023). Nonetheless, there is a call for more long-term research, inclusion of diverse
populations, and better integration into current mental health systems.

Furthermore, there is a growing global interest in AR research, with the United States and
Germany leading in contributions, according to Borgohain et al. (2022). The field is largely driven
by academic pursuits, as evidenced by the dominance of conference proceedings and journal
publications. This trend underscores the rapid development and sustained scholarly attention
toward AR/VR technologies in healthcare.

Gap Findings in Literature:

12
Pimpri Chinchwad College of Engineering,PCCOE,Pune
Current research in digital physiotherapy is largely concentrated on AR/VR-based rehabilitation,
with limited emphasis on web-based physiotherapy applications that utilize AI and motion tracking
technologies (Bhugaonkar et al., 2022; Ullah et al., 2023). While the short-term benefits of AR/VR
interventions in healthcare have been acknowledged, there is a notable lack of long-term studies
evaluating their effectiveness, patient adherence, and clinical outcomes (Sriati et al., 2023).
Furthermore, the potential of AI in enabling automated motion tracking and providing real-time
feedback for physiotherapy remains underexplored, as most existing studies continue to focus on
immersive AR/VR environments rather than web-based platforms (Rahman et al., 2021).

Another pressing issue is the absence of standardized guidelines for ensuring data privacy and
security in web-based healthcare solutions. Although privacy concerns have been raised in the
context of metaverse and AR/VR applications (Fiske et al., 2019; Dutmer et al., 2019), concrete
strategies for safeguarding sensitive data in AI-driven online physiotherapy tools are yet to be
developed. Additionally, challenges related to AI ethics, including algorithmic bias, transparency,
and user acceptance, have been highlighted (Dutmer et al., 2019). However, further investigation is
required to specifically address these concerns in the context of AI-assisted physiotherapy,
ensuring equitable and trustworthy patient care.

13
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 3. SOFTWARE REQUIREMENTS SPECIFICATIONS

3.1 FUNCTIONAL REQUIREMENTS

The web-based physiotherapy application is equipped with multiple intelligent system features to
ensure accurate, personalized, and engaging rehabilitation experiences. One of the core
functionalities is real-time pose estimation, which uses frameworks like MediaPipe or OpenPose to
identify skeletal key points and track body posture. It continuously monitors joint positions, limb
angles, and movement trajectories to distinguish between correct and incorrect postures using only
a webcam or smartphone camera. The system is optimized for low-latency execution to provide
seamless feedback during exercises.

Another significant feature is the AI-based exercise classification system, which leverages
BiLSTM or LSTM models trained on custom datasets of rehabilitation exercises. This enables the
system to recognize a variety of movements such as squats, lunges, bicep curls, and planks, and to
assess their correctness by comparing them with predefined motion patterns. Closely linked is the
repetition counting feature, where the AI model tracks joint movements to count accurate
repetitions, avoiding false positives by using predefined movement thresholds. It also provides
immediate visual or audio cues and allows sensitivity adjustments to match user variability.

To enhance user engagement, the system includes real-time performance feedback that evaluates
posture, movement speed, and range of motion. It uses a color-coded scoring system to indicate
accuracy, provides personalized tips, and concludes each session with a performance summary.
The application is accessed through a web-based interface designed for cross-device compatibility,
removing the need for software installation and including an intuitive dashboard for both users and
therapists.

Further supporting personalized care, customizable exercise programs allow users and therapists to
tailor workouts, set repetition goals, and modify exercise difficulty based on the patient’s recovery.
Integrated reminder notifications help users maintain consistency. The data logging and progress

14
Pimpri Chinchwad College of Engineering,PCCOE,Pune
tracking module records session history, repetition counts, and accuracy metrics, visualized
through progress charts and available for export to share with physiotherapists.

With multi-user support, the system accommodates multiple patient profiles, allowing therapists to
remotely manage and track individual progress. Cloud storage ensures secure and personalized
data access. Enhancing usability, voice and visual assistance is offered through real-time audio
instructions, on-screen markers, and step-by-step tutorials to guide users through exercises. To
increase motivation, gamification features such as rewards, achievement badges, leaderboards, and
streak tracking are embedded to foster user engagement and consistency.

A standout feature is the AI-powered error detection and correction module, which identifies
incorrect joint positions and deviations from ideal movements. It offers on-screen corrections and
adapts feedback based on the user's body type and motion characteristics. Lastly, the application
provides cross-browser and multi-platform support, functioning smoothly on Chrome, Firefox,
Edge, and Safari, and across operating systems like Windows, macOS, iOS, and Android. It is also
optimized for low-end hardware to ensure broad accessibility, making it a comprehensive,
intelligent, and user-centric solution for remote physiotherapy.

15
Pimpri Chinchwad College of Engineering,PCCOE,Pune
3.2 EXTERNAL INTERFACE REQUIREMENT

This section outlines the external interfaces required for the web-based physiotherapy application,
including user interface, hardware, software, and communication interfaces.

The user interface of the physiotherapy system is crafted to offer an intuitive and seamless
experience for both patients and physiotherapists. Patients can access a dedicated dashboard that
displays assigned exercises, tracks progress over time, and provides feedback based on
performance. Physiotherapists are provided with a separate dashboard that allows them to manage
patient records, assign or modify exercise routines, and monitor each patient's recovery remotely.
The exercise guidance screen delivers real-time feedback through motion tracking while offering
visual instructions to help users maintain correct posture and form. The overall design ensures
simple navigation, compatibility with a range of devices including desktops, tablets, and
smartphones, and can incorporate multilingual support to improve accessibility for users from
diverse linguistic backgrounds.

On the hardware front, the system interacts efficiently with devices necessary for real-time motion
tracking and user engagement. It utilizes a standard webcam or integrated camera to capture user
movements for posture analysis. An optional microphone can be used to receive voice commands
or provide audio feedback, enhancing the interactive experience. The application is responsive to
various screen sizes, allowing it to function well on PCs, laptops, tablets, and mobile phones.
Additionally, it supports cloud-based or local storage systems to maintain patient progress reports
and historical data securely.

The software interface integrates a range of components to ensure smooth operation. It is

compatible with operating systems such as Windows, macOS, and Linux, and works across widely
used web browsers including Google Chrome, Mozilla Firefox, Microsoft Edge, and Safari. The
backend system is developed using frameworks like Django or Flask for Python, or alternatively
Node.js, to manage application logic and server-side operations. The frontend is built using
modern frameworks such as React.js, Vue.js, or Angular to deliver a responsive and engaging user
experience. Patient records, exercise data, and progress metrics are stored using relational

16
Pimpri Chinchwad College of Engineering,PCCOE,Pune
databases like MongoDB, while motion tracking is facilitated by AI frameworks such as OpenCV,
TensorFlow, or MediaPipe.

Effective communication between patients, physiotherapists, and supporting systems is enabled

through robust communication interfaces. The application requires internet connectivity for
real-time interactions, remote monitoring, and cloud-based data storage. It includes API integration
to support authentication through services like Google OAuth, cloud storage using platforms such
as AWS S3, and AI model hosting. The system can send email or SMS notifications to remind
users of scheduled exercises or update them on progress. For real-time physiotherapy sessions,
WebSocket technology is employed to ensure smooth and instant feedback and interaction between
users and therapists.

3.3 NON-FUNCTIONAL REQUIREMENTS

This section outlines the non-functional requirements that ensure the reliability, security, and
performance of the web-based physiotherapy application.

3.3.1 Performance Requirements

The system is expected to be highly responsive, scalable, and capable of managing multiple
concurrent users without experiencing noticeable delays. Page load times should not exceed three
seconds under normal load conditions. For real-time motion tracking, latency should remain at or
below 150 milliseconds to ensure seamless user interaction. Initially, the system must support at
least 100 concurrent users and be designed to scale effortlessly to accommodate increased demand.
The AI-based motion tracking and posture correction module should maintain a minimum
processing speed of 30 frames per second (FPS) to provide real-time feedback. To ensure
consistent availability, the system must achieve an uptime of 99.9%, minimizing any disruption for
both patients and physiotherapists. Furthermore, the solution should be deployable on cloud
infrastructure platforms such as AWS, Microsoft Azure, or Google Cloud to enable dynamic
scaling and high availability.

17
Pimpri Chinchwad College of Engineering,PCCOE,Pune
3.3.2 Safety/Security Requirements

To safeguard user data and maintain privacy, the system must incorporate robust security measures
and adhere to applicable healthcare regulations. User authentication should be implemented using
secure standards such as OAuth 2.0 or JSON Web Tokens (JWT), and role-based access control
(RBAC) must be enforced to differentiate permissions between patients and physiotherapists. All
sensitive patient data stored in the system should be encrypted using AES-256 encryption, while
data transmitted over the network should be protected using SSL/TLS 1.3 encryption protocols.
The system must comply with healthcare data regulations such as HIPAA for users in the United
States, GDPR for European users, or other region-specific laws to ensure the confidentiality of
patient information. For additional security, users should be automatically logged out after ten
minutes of inactivity. The system must also maintain detailed audit logs capturing all user
activities, including login attempts, data access, and modifications. Secure API access is essential
and should require authentication tokens, along with implementation of rate-limiting policies to
prevent misuse. To prepare for potential failures, the system should support automated daily
backups and include a disaster recovery plan for data restoration.

3.4 SYSTEM REQUIREMENTS

This section defines the necessary database, software, and hardware requirements for the AI
Physiotherapy Tracker—Intelligent Healing Spaces.

3.4.1 Database Requirements

The system requires a robust relational database to efficiently store and manage critical data such
as user profiles, physiotherapy session records, exercise logs, and AI model predictions.
PostgreSQL is the preferred choice due to its scalability, strong security features, and ACID
compliance, making it ideal for handling structured healthcare data. Optionally, MongoDB may be
employed for specific components that benefit from a flexible schema, such as real-time tracking
data and unstructured logs.

18
Pimpri Chinchwad College of Engineering,PCCOE,Pune
The database must support various types of information. This includes user data such as login
credentials, user profiles, medical history, and progress tracking details. It should also store
physiotherapy-related information like predefined exercise routines and AI-generated motion
tracking data. Session logs need to capture exercise completion records and any feedback provided
by the AI models. Additionally, the database should maintain audit logs detailing system usage,
user login activities, and API request history for security and compliance purposes.

To ensure data integrity and availability, the system should implement automated daily incremental
backups along with full backups on a weekly basis. A disaster recovery plan must also be in place,
including the use of redundant storage systems to prevent data loss and ensure continuity of service
in case of system failures.

3.4.2 Software Requirements (Platform Choice)

The web-based AI physiotherapy tracker demands a robust and modern technology stack capable
of supporting real-time pose estimation, AI-powered processing, and an interactive user interface.
On the frontend, the system should utilize React.js for building a dynamic user interface, or Next.js
for enhanced performance through server-side rendering. For UI components, Tailwind CSS or
Material UI will be used to ensure a responsive and accessible design across devices. Integration of
webcam access is facilitated through WebRTC or the Media API.

On the backend, the development will be carried out in Python, leveraging frameworks like
FastAPI or Flask for building RESTful APIs. AI model processing, particularly for human pose
estimation, will be handled using TensorFlow or PyTorch, while database interactions will use
SQLAlchemy (for SQL databases) or Mongoose (for MongoDB).

The AI and computer vision functionalities rely on powerful libraries including MediaPipe, which
is Google's framework for real-time human pose tracking, OpenCV for video and image
processing, and NumPy and Pandas for data manipulation and analysis.

Deployment and hosting will be performed on cloud platforms such as AWS, Google Cloud, or
Microsoft Azure. To ensure scalability and resilience, the system will be containerized using

19
Pimpri Chinchwad College of Engineering,PCCOE,Pune
Docker, with orchestration handled by Kubernetes. Additionally, GitHub Actions will be used for
setting up a continuous integration and deployment (CI/CD) pipeline.

3.4.3 Hardware Requirements

The system’s hardware requirements are defined to ensure smooth server-side performance and
compatibility with user devices. On the server side, a powerful processor such as Intel Xeon or
AMD EPYC with at least 16 cores is essential. The memory requirement is a minimum of 32GB
RAM, although 64GB is recommended to support high user concurrency and AI model inference.
Storage should be SSD-based, with a minimum of 1TB capacity, expandable to accommodate
AI-generated data. A dedicated GPU, such as the NVIDIA Tesla T4 or A100, is crucial for
real-time AI inference and motion tracking. Additionally, a high-speed, low-latency fiber internet
connection is necessary to support smooth data flow.

On the client side, end-user devices should be equipped with at least an Intel i5 processor (or
equivalent) and 8GB RAM to ensure performance efficiency. A 720p or higher HD webcam is
required for motion tracking functionality, and a minimum internet speed of 5 Mbps is
recommended for seamless real-time video streaming and interactions.

3.5 SDLC Model to Be Applied

For the development of the AI Physiotherapy Tracker – Intelligent Healing Spaces, the Agile
Development Model is identified as the most suitable Software Development Life Cycle (SDLC)
approach.

The Agile model offers several key benefits, starting with incremental development and
continuous feedback, allowing the system to be built in functional components. Physiotherapy
exercises, AI models, and user feedback mechanisms can be validated through iterative testing,
ensuring the system evolves based on actual usage and performance. Its flexibility and adaptability
are particularly beneficial for AI-based systems, which require frequent updates and enhancements

20
Pimpri Chinchwad College of Engineering,PCCOE,Pune
derived from real-world data. Agile supports these iterative improvements in both AI algorithms
and UI/UX design without compromising overall project timelines.

Furthermore, Agile facilitates user-centered development, which is critical in physiotherapy

applications where interaction with patients and physiotherapists directly influences design
choices. Regular usability tests and feedback loops ensure the system meets the real needs of its
users. The model also supports the integration of AI and computer vision techniques, like RNN or
LSTM models, which require continuous refinement. Agile enables model updates to be deployed
incrementally, without disrupting the system’s functionality.

The model promotes efficient bug fixing and quality assurance, with frequent testing cycles that
help catch issues early — such as pose estimation inaccuracies or system latency — thereby
maintaining high movement tracking accuracy and reliability.

For implementation, the Agile Scrum framework will be followed. Sprints will span 2–3 weeks,
and key roles include the Product Owner, who defines physiotherapy features and requirements;
the Scrum Master, responsible for guiding Agile practices; and the Development Team, which
handles AI model building, frontend and backend integration, and computer vision system
development.

21
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 4. PROJECT PLAN

4.1 PROJECT COST ESTIMATION

4.1.1 COMPUTATIONAL COST

●​ Cloud GPU Services (e.g., Google Colab Pro, AWS, or Azure ML): ₹3,000 -
₹5,000/month
●​ Local Machine (Laptop/PC with GPU for Model Training): ₹80,000 - ₹1,20,000
(one-time)
●​ Web Hosting & Server Costs: ₹2,000 - ₹5,000/month
●​ Internet & Maintenance: ₹1,000 - ₹2,500/month

Table 4.1 : Computational Costs

Item Description Cost

Cloud GPU Services (e.g., Google Colab Pro, ₹3,000 - ₹5,000/month

AWS, or Azure ML)

Local Machine (Laptop/PC with GPU for ₹80,000 - ₹1,20,000

Model Training) (one-time)

Web Hosting & Server Costs ₹2,000 - ₹5,000/month

Internet & Maintenance ₹1,000 - ₹2,500/month

Table 4.2 : Component Specifications

Component Specification
Processing Power GPU (NVIDIA RTX 3060) / CPU (Intel i7)
Memory Usage 16GB - 32GB RAM
Storage Requirements 512GB SSD + 1TB HDD
Network Latency & Bandwidth High-speed broadband (100 Mbps - 500 Mbps)

22
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 4.1.2 SOFTWARE PERFORMANCE COST

●​ AI Model Training & Optimization Tools (TensorFlow, PyTorch, Mediapipe –

Free/Open-source): ₹0
●​ Database Management (Firebase/MySQL Server Hosting): ₹500 - ₹2,000/month
●​ Software Licensing (if using premium APIs or libraries): ₹5,000 - ₹10,000
(one-time or annual)

Table 4.3: Software Performance Cost

Item Description Cost

AI Model Training & (TensorFlow, PyTorch, Mediapipe ₹0

Optimization Tools – Free/Open-source)

Database Management (Firebase/MySQL Server ₹500 - ₹2,000/month

Hosting)

Software Licensing (if using premium APIs or ₹5,000 - ₹10,000

libraries) (one-time or annual)

Estimated Total Cost: ₹90,000 - ₹1,50,000 (initial setup) + ₹6,500 - ₹12,500/month (operational)

4.2 SUSTAINABILITY ASSESSMENT

4.2.1 ENVIRONMENTAL SUSTAINABILITY

●​ The system runs on cloud-based AI, reducing the need for energy-intensive personal
computing.
●​ Uses efficient ML models (RNN, LSTM, Mediapipe) to reduce computational
power and CO₂ footprint.

23
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 Table 4.4: Computational Complexity Analysis

Aspect Details Impact

Processing Power The AI model requires real-time pose High computational load on
Requirement estimation and movement analysis. CPU/GPU.
Algorithmic RNN and LSTM architectures have O(n²) Performance may degrade with
Complexity complexity. longer input sequences.
Real-time Mediapipe-based skeletal tracking is May introduce latency on
Execution optimized for fast processing. low-end hardware.
Increased input sequences lead to higher Limits usability on
Scalability memory and processing demands. resource-constrained devices.

4.2.2 ECONOMIC SUSTAINABILITY

●​ Cost-effective web-based solution, reducing the need for expensive physiotherapy

sessions.
●​ Scalable AI reduces costs over time with minimal hardware upgrades.
●​ Open-source tools lower development and maintenance costs.

Table 4.5: Economic Sustainability Analysis

Factor Description Impact

Cost-Effective Web-based platform minimizes the Reduces long-term expenses for
Solution need for in-person physiotherapy. patients.
AI-driven model requires minimal Lowers operational costs and
Scalability hardware upgrades over time. enhances system longevity.
Open-Source Use of Mediapipe, TensorFlow, and Cuts down on development and
Tools other free libraries. maintenance expenses.

24
Pimpri Chinchwad College of Engineering,PCCOE,Pune
4.2.3 SOCIAL SUSTAINABILITY

●​ Enhances accessibility to physiotherapy for patients in remote areas.

●​ Improves adherence to rehab programs through real-time feedback.
●​ Reduces dependence on in-person physiotherapy, allowing greater flexibility and
convenience.

Table 4.6: Social Sustainability Analysis

Factor Description Impact

Enables physiotherapy access for Improves healthcare reach and
Accessibility patients in remote areas. inclusivity.
Adherence to Rehab Provides real-time feedback to Enhances recovery outcomes and
Programs encourage participation. patient engagement.
Flexibility & Reduces reliance on physical visits to Saves time and allows for
Convenience physiotherapists. personalized rehabilitation.

4.3 COMPLEXITY ASSESSMENT

4.3.1 COMPUTATIONAL COMPLEXITY

●​ The AI model relies on real-time pose estimation and movement analysis, requiring
substantial processing power.
●​ RNN and LSTM architectures used for motion tracking have O(n²) complexity,
impacting scalability with increased input sequences.
●​ Mediapipe-based skeletal tracking is optimized for real-time execution but may
introduce latency on low-end hardware.

4.3.2 ALGORITHMIC COMPLEXITY

●​ Pose estimation algorithms (Mediapipe) run in O(n) time complexity, making them
efficient for real-time tracking.

25
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 ●​ Deep learning models (RNN/LSTM) require O(n²) to O(n³) time for training but
operate efficiently at inference time.
●​ Repetition counting and movement validation involve multiple conditional checks,
keeping processing within O(n) complexity.

Table 4.7: Algorithmic Complexity Analysis

Algorithm Time Complexity Efficiency

Fast and suitable for real-time
Pose Estimation (Mediapipe) O(n) applications.
Training is computationally
expensive, but inference is
Deep Learning Models (RNN/LSTM) O(n²) to O(n³) optimized.

Repetition Counting & Movement Efficient due to conditional

Validation O(n) checks.

4.3.3 IMPLEMENTATION COMPLEXITY

●​ Integrating real-time video processing with AI models is challenging due to

synchronization issues.
●​ Ensuring cross-browser compatibility for a web-based platform requires extensive
testing.
●​ Handling diverse physiotherapy exercises demands a custom dataset with
high-quality annotations, increasing development effort.

4.3.4 RESOURCE COMPLEXITY

●​ Requires GPU-enabled cloud computing for training large AI models, increasing

infrastructure dependency.
●​ Webcam-based tracking relies on stable internet connections, which may limit
accessibility in rural areas.

26
Pimpri Chinchwad College of Engineering,PCCOE,Pune
4.4 RISK MANAGEMENT

4.4.1 RISK IDENTIFICATION

Table 4.8 : Risk Identification

Risk Type Mitigation Strategy Monitoring & Management

Optimize AI models, use high-quality Regular testing, user feedback
Technical Risks datasets. integration.
Secure data collection and Implement GDPR-compliant policies,
Data Risks anonymization. periodic audits.
End-to-end encryption, strict access
Security Risks control. Security audits, compliance checks.
User Adoption Conduct training, improve UI for Regular user engagement surveys,
Risks ease of use. feature updates.
Optimize cloud costs, seek
Financial Risks grants/funding. Budget reviews, cost-benefit analysis.

4.4.2 RISK ANALYSIS

Table 4.9: Risk Impact Level Analysis

Impact
Risk Category Level Potential Consequences
Model Misclassification High Misleads patients, affects recovery progress.
Privacy Breaches High Data leaks may result in legal consequences.
Computational Latency Medium Affects real-time performance but can be optimized.
Slows down implementation but can be addressed with
Resistance to Adoption Medium training.
Minor Tracking
Inaccuracies Low Does not significantly affect feedback.

27
Pimpri Chinchwad College of Engineering,PCCOE,Pune
4.5 PROJECT SCHEDULE

4.5.1 PROJECT TASK SET

The development of the AI Physiotherapy Tracker is structured into four key phases, each
spanning a specific time frame and focusing on distinct project goals.

Phase 1: Research and Planning (Week 1–2)​

This phase lays the foundation for the entire project.

●​ Week 1: The team will define detailed project requirements and objectives, followed by
research into advanced pose estimation techniques and suitable AI models.​

●​ Week 2: Efforts will focus on collecting and preparing datasets necessary for training the AI
models.​

Phase 2: Implementation (Week 3–6)​

The core development tasks are carried out during this period.

●​ Week 3: Implementation of pose estimation methods using libraries like OpenPose or

MediaPipe.​

●​ Week 4: Development of AI algorithms for posture analysis and motion recognition.​

●​ Week 5: Designing and building a user-friendly interface to ensure intuitive user interaction.​

●​ Week 6: Integration of a real-time feedback system that responds dynamically to user

movement.​

Phase 3: Testing and Refinement (Week 7–8)​

This phase emphasizes system validation and optimization.

●​ Week 7: Testing will be conducted to evaluate the accuracy of pose detection and to
fine-tune the system's performance.​

●​ Week 8: Bug fixing and system stabilization activities will be undertaken to ensure
reliability and smooth functionality.​

28
Pimpri Chinchwad College of Engineering,PCCOE,Pune
Phase 4: Documentation and Final Delivery (Week 9–10)​
The final phase focuses on deployment, evaluation, and reporting.

●​ Week 9: The application will be deployed to a suitable platform and user feedback will be
collected for future enhancements.​

●​ Week 10: Preparation of the final project report and presentation to summarize the entire
development process and outcomes.​

4.5.2 TIMELINE CHART

Figure 4.1: Project Timeline Gantt Chart

29
Pimpri Chinchwad College of Engineering,PCCOE,Pune
4.6 TEAM ORGANIZATION ( STRUCTURE)

Table 4.10: Team Organisation Structure

Role Responsibilities Team Member

Oversees project progress, ensures deadlines are met, and Prerna
Project Manager manages risks. Pokharkar
AI/ML Develops and trains the AI model for pose estimation and
Developer movement analysis. Rishabh Patil
Full-Stack Implements the user interface, backend, and integrates
Developer real-time feedback. Aryan Pitre
Dr. Deepali
Project Guide Provides technical guidance and project supervision. Naik

30
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 5. SYSTEM DESIGN

5.1 PROPOSED SYSTEM ARCHITECTURE/BLOCK DIAGRAM

Figure 5.1: System Architecture

5.2 DATASET/ DATABASE DESIGN

5.2.1 DATASET COMPOSITION:

●​ The dataset consists of labeled pose keypoints extracted from exercise videos.
●​ Each exercise (e.g., squats, push-ups) is mapped to a set of keypoints representing
body positions.
●​ Data is stored as CSV files containing keypoint coordinates and class labels.

31
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 5.2.2 DATA COLLECTION PROCESS:

●​ Positive Samples: Videos of correct exercises.

●​ Negative Samples: Random movements or incorrect exercises.
●​ Data augmentation techniques (flipping, rotation) are applied to improve model
generalization.

5.2.3 DATASET EXAMPLE (CSV FORMAT):

Figure 5.2: Sample of coordinates from dataset

Figure 5.3: Images from the Datasets used for training and validation of the exercise classifier.
(images are cropped due to screenshots, hence they appear to have different dimensions)

32
Pimpri Chinchwad College of Engineering,PCCOE,Pune
5.3 MATHEMATICAL MODEL

The mathematical model behind the AI Physiotherapy Tracker involves pose estimation,
movement analysis, and sequential data processing using deep learning.

Problem Formulation

The system classifies physiotherapy exercises and counts repetitions using keypoint detection and
temporal sequence modeling.

●​ Input: Video frames X={x1​,x2​,...,xn​} from a webcam.

●​ Output: Exercise classification Y and repetition count R.

Pose Estimation (Keypoint Extraction)

Using Mediapipe, each frame extracts J skeletal keypoints:

Pt​={(x1​,y1​),(x2​,y2​),...,(xJ​,yJ​)}

where (xj​,yj​) are the coordinates of the jth joint at time t.

Feature Representation

The keypoints are converted into meaningful features:

●​ Joint Angles: θj​=arctan(xj​yj)​​

●​ Normalized Distances: Dij​= {​(xi​−xj​)^2+(yi​−yj)^2} / dmax

Temporal Modeling with BiLSTM

●​ The extracted feature sequence F={F1​,F2​,...,Fn​} is fed into a BiLSTM network.

33
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 ●​ Forward and backward passes encode dependencies: htfwd​=σ(Wf​Ft​+Uf​ht−1fwd​+bf​)
htbwd​=σ(Wb​Ft​+Ub​ht+1bwd​+bb​) where htfwd​and htbwd​represent the hidden states in
forward and backward passes.

Classification & Repetition Counting

●​ The final hidden state is passed to a softmax classifier: Y=softmax(Wo​ht​+bo​)

●​ Repetitions are counted based on peak detection in movement trajectories.

5.4 UML DIAGRAMS

i) Activity diagram

Figure 5.4 : Activity diagram

34
Pimpri Chinchwad College of Engineering,PCCOE,Pune
ii) User diagram

Figure 5.5 : User diagram

iii) Sequence diagram

Figure 5.6 : Sequence diagram

35
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 6. PROJECT IMPLEMENTATION

6.1 OVERVIEW OF PROJECT MODULES

The AI Physiotherapy Tracker consists of the following key modules:

1.​ User Management Module – Handles patient registration, authentication, and profile
management.
2.​ Pose Estimation Module – Uses AI-based techniques to track body movements in real time.
3.​ Exercise Analysis Module – Compares user movement with ideal physiotherapy exercises
and provides feedback.
4.​ Feedback and Reporting Module – Generates real-time feedback on posture, movement
accuracy, and improvement suggestions.
5.​ Web-Based User Interface – Provides an interactive platform for patients and
physiotherapists to track progress.

6.2 TOOLS AND TECHNOLOGIES USED

Table 6.1: Tools and Technologies Used

Category Technology/Tool

Programming Languages Python, JavaScript

Frameworks/Libraries TensorFlow, PyTorch, OpenCV, MediaPipe
Frontend React.js, HTML, CSS
Backend Flask, FastAPI
Database PostgreSQL, Firebase
Cloud/Hosting AWS, Google Cloud, Firebase
Version Control Git, GitHub

Development Tools Jupyter Notebook, Visual Studio Code

36
Pimpri Chinchwad College of Engineering,PCCOE,Pune
6.3 ALGORITHM DETAILS

6.3.1 ALGORITHM 1 - Pose Estimation using MediaPipe

Objective: Extract and track human skeletal key points in real time from a webcam video stream to
monitor physiotherapy exercises.

Tools Used:

●​ MediaPipe Pose for keypoint extraction

●​ OpenCV for image processing
●​ NumPy/Pandas for data processing

Steps:

1.​ Capture the video feed

○​ Use OpenCV to access the webcam and continuously fetch frames.
2.​ Preprocess the frame
○​ Convert the frame to RGB format (as required by MediaPipe).
3.​ Apply MediaPipe Pose Detection
○​ Use the MediaPipe Pose model to extract 33 skeletal key points, including joints
(shoulders, elbows, knees, etc.).
4.​ Normalize key points
○​ Normalize the key point coordinates with respect to the frame size to make the
model robust to different camera positions.
5.​ Compare detected key points with predefined physiotherapy exercises
○​ Store key points of each exercise and compare the detected pose with a ground truth
dataset.
○​ Use Euclidean distance or cosine similarity to measure deviation from the ideal
pose.
6.​ Count Repetitions
○​ Implement an angle-based movement tracking technique.

37
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 ○​ Example:
■​ Track the elbow angle for an arm curl.
■​ If the angle decreases below a threshold and then increases back, count it as
one repetition.
7.​ Store results and generate real-time feedback
○​ Store the session details in a database.
○​ Display corrective feedback on posture adjustments.

Figure 6.1: Mediapipe Flowchart

38
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 Figure 6.2: Mediapipe working illustration

6.3.2 ALGORITHM 2 – Motion Analysis and Feedback Generation using LSTM/BiLSTM

Steps for Motion Analysis and Feedback Generation using BiLSTM

1.​ Data Collection

○​ Convert the sequence of key points from MediaPipe into a structured dataset.
○​ Example: A single frame will have 33 key points, each with (x, y, z) coordinates.
○​ Collect multiple frames to create a time-series dataset.
2.​ Preprocessing
○​ Normalize key point positions.
○​ Convert sequential joint positions into a vector representation.
○​ Example:Xt=[x1,y1,z1,x2,y2,z2,...,x33,y33,z33]X_t = [x_1, y_1, z_1, x_2, y_2,
z_2, ..., x_{33}, y_{33}, z_{33}]Xt​=[x1​,y1​,z1​,x2​,y2​,z2​,...,x33​,y33​,z33​](where
XtX_tXt​represents a single frame at time t)
3.​ BiLSTM Model Architecture
○​ Input Layer: Sequences of key points over time (e.g., 30 frames).
○​ BiLSTM Layers: Capture motion patterns forward and backward.
○​ Fully Connected Layers: Classify the movement into correct or incorrect posture.

39
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 ○​ Softmax Layer: Outputs a probability score for different movement classes.
4.​ Training on Custom Dataset
○​ Use a labeled dataset of correct and incorrect physiotherapy exercises.
○​ Loss function: Categorical Cross-Entropy
○​ Optimizer: Adam
5.​ Feedback Generation
○​ If movement is classified as incorrect, generate real-time feedback:
■​ “Adjust left knee position”
■​ “Slow down your movement”
○​ Display feedback via text, sound alerts, or visual indicators.

Figure 6.3: LSTM classifier flowchart

40
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 7. SOFTWARE TESTING

7.1 TYPE OF TESTING

Software testing is a critical phase in the development lifecycle that ensures a software application
functions correctly, securely, and efficiently. It involves systematically identifying defects,
verifying that the software meets user and business requirements, and optimizing performance
across different environments. Testing helps detect vulnerabilities, preventing security breaches
and ensuring data protection. By identifying bugs early, it reduces development costs and
minimizes post-deployment maintenance efforts. Additionally, it enhances user satisfaction by
delivering a smooth, error-free experience while ensuring compliance with industry standards.
Ultimately, software testing improves software reliability, making it more robust and trustworthy
for end users.

Table 7.1: Types of Software Testing

Testing
Type Purpose Applied to
Verify correctness of individual functions (pose MediaPipe pose detection,
Unit Testing estimation, movement classification). BiLSTM motion analysis
Integration Ensure all modules (pose detection, feedback
Testing system, UI) work together. Full system integration
Functional Check if the system meets physiotherapy tracking Repetition counting, posture
Testing requirements. correction
Performance AI inference time, real-time
Testing Measure system efficiency, response time, and lag. feedback delay
Usability Web interface, real-time
Testing Evaluate ease of use and user experience. feedback display
Security User authentication, data
Testing Ensure video data privacy and system integrity. handling

41
Pimpri Chinchwad College of Engineering,PCCOE,Pune
7.2 TEST CASES & TEST RESULTS

Table 7.2: Test cases and Results of the Project

Test Status
Case ID Test Scenario Expected Output Actual Output (Pass/Fail)
Detect pose from Key points detected ✅ Key points
TC-01 webcam input accurately detected correctly Pass
Count repetitions of Correct rep count ✅ Accurate rep
TC-02 an exercise displayed count Pass
Provide real-time Feedback given for ✅ Feedback
TC-03 posture feedback incorrect posture displayed Pass
Handle varying System adjusts to ✅ Robust across
TC-04 camera angles different views angles Pass
Fail

❌
Ensure minimal AI model responds (Optimization
TC-05 processing delay within 0.5 sec Delay of 0.8 sec Needed)
Handle incorrect user Incorrect posture ✅ Model detects
TC-06 movements detected errors Pass
Data stored in the ✅ Stored
TC-07 Save session history database successfully Pass

42
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 8. RESULTS

A thorough examination of our experimental results for the BiLSTM model created with both
coordinate and angular features is provided in this chapter. We compare this suggested model's
performance with three other methods: a standard LSTM model trained on the same feature sets, a
BiLSTM model using only coordinate data, and a BiLSTM variation using just angles and
normalized distances. Furthermore, we compare our implementations to the approaches.

The Final Model is tested on the test dataset of exercise videos, which is contained in the folder
called “exercise_videos” and the model is repeatedly tested for 50 epochs.

Figure 8.1: Process of Testing for 50 epochs (A snapshot is provided the model is trained in
Tensorflow so the model could recognize the type of exercise)

8.1 OUTCOMES

The AI Physiotherapy Tracker (Intelligent Healing Spaces) has been successfully developed and
tested for accurate exercise tracking, movement correction, and real-time feedback. The results
demonstrate its effectiveness in physiotherapy rehabilitation and guided exercise monitoring.

Key Outcomes of the Project -

1. Real-Time Pose Estimation & Joint Tracking

43
Pimpri Chinchwad College of Engineering,PCCOE,Pune
The system uses MediaPipe to detect key skeletal points of the human body. It extracts features
such as joint angles, limb positions, and movement trajectories to analyze posture. The pose
keypoints (e.g., shoulders, elbows, knees) are mapped in real-time, ensuring smooth tracking and
accuracy.

2. AI-Powered Posture Classification

The AI model classifies physiotherapy exercises using BiLSTM and LSTM architectures. The
system currently supports exercises such as bicep curls, squats, and planks, with the potential to
expand to other physiotherapy exercises. It identifies pose errors and provides real-time feedback
to correct form and execution.

3. Repetition Counting and Movement Analysis

The system tracks and counts exercise repetitions automatically. It detects movement variations,
such as incorrect squat depth or improper arm positioning, and provides real-time feedback if the
movement is too fast, too slow, or incorrect.

4. High Accuracy with Mixed Feature Learning

The BiLSTM model demonstrates higher accuracy compared to the standard LSTM model. By
incorporating joint angles and normalized distances, the classification is more robust and resistant
to variations in user movement. The model, tested on a custom physiotherapy dataset, achieves an
accuracy of up to 98.5%.

5. Web-Based Deployment for Accessibility

The system is implemented as a web-based application, allowing users to perform exercises at

home while enabling physiotherapists to monitor sessions remotely. The system efficiently
processes live webcam video and applies pose estimation in real-time for seamless user
experience.

44
Pimpri Chinchwad College of Engineering,PCCOE,Pune
6. Generalization Across Synthetic & Real-World Data

The model was trained on a dataset containing both real and synthetic exercise data and
generalizes well across different environments, such as home, clinic, or gym settings. However,
minor performance drops are observed in extreme lighting conditions, which is an area for future
improvement.

7. Robust Performance Despite Pose Variations

The model effectively handles pose variations, including different camera angles (front, side),
variations in user height and body type, and changes in exercise speed. Some challenges remain,
such as detecting planks due to occluded key points when the body is aligned horizontally.

8.2 RESULT ANALYSIS AND VALIDATIONS

A.​Hyperparameters

Table 8.1: Hyperparameter analysis LSTM vs BiLSTM(LSTM uses a single directional flow
to process sequences, while BiLSTM processes data in both forward and backward directions,
capturing more context but requiring more computational resources.)

Hyperparameter LSTM BiLSTM

Batch Size 36 52
Dropout Rate 0.375 0.225
Epochs 55 75
Learning Rate 0.00012 0.00035
Units 120 76

45
Pimpri Chinchwad College of Engineering,PCCOE,Pune
B. Evaluation of Test Sets on The Dataset Used For Training
Table 8.2: Evaluation of Test Sets on The Dataset Used For Training (Classification Report is
generated on the model trained to classify the exercises)

Class Precision Recall F1-Score Support

Bicep Curl 0.5 0.33 0.4 3
Plank 0 0 0 1
Squat 0.5 0.75 0.6 4
Accuracy 0.9 8
Macro Avg 0.83 0.86 0.83 8
Weighted Avg 0.94 0.85 0.85 8

C. Classification Report for LSTM model

Figure 8.2: Confusion Matrix for LSTM model (A heatmap is generated to accurately represent
the exercise which should be classified by model)

46
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 D. Accuracy and Loss for the Model

Figure 8.3: Accuracy and Loss for the Model (A line graph represent accuracy increases by the
epochs and loss decreases)

Table 8.3: Numeric Accuracy of Test set and Final Test( This shows accuracy performance of
BiLSTM and LSTM models using mixed features. BiLSTM achieved higher accuracy across all
datasets, with 0.985 on the test set and 0.872 on the final test, compared to LSTM 0.9805 and
0.8555 respectively, indicating better generalization and performance.)

My Test Video
Model Test Set Accuracy Accuracy Final Test Accuracy
BiLSTM with Mixed
Features 0.985 0.945 0.872
LSTM with Mixed
Features 0.9805 0.935 0.8555

47
Pimpri Chinchwad College of Engineering,PCCOE,Pune
8.3 SCREENSHOTS

Figure 8.4(a): Welcome Page (This is the landing page of our site first page the user encounters
this guides the user to a LOGIN, ABOUT US, BLOG and CONTACT button user can explore)

Figure 8.4(b): Welcome Page(Scrolling through the welcome page the user can read specific
information about our website and its features )

48
Pimpri Chinchwad College of Engineering,PCCOE,Pune
Figure 8.5 : Login page(User is directed to this page after pressing the login button it has 2
textboxes for Username and Password if existing user then Username and Password to be
entered if not then, Sign Up)

49
Pimpri Chinchwad College of Engineering,PCCOE,Pune
Figure 8.6 : Register page(This page is for new User After entering all new detailed they are
saved in the MongoDB database and User is prompted to Login by the same details)

50
Pimpri Chinchwad College of Engineering,PCCOE,Pune
Figure 8.7 : Exercise Menu Page(This page is reached after Login User can Choose between
different exercises)

Figure 8.9 : User Activity History(This page shows the activity history performed by the User
this is also saved in the MongoDB database)

51
Pimpri Chinchwad College of Engineering,PCCOE,Pune
Figure 8.10(a) : Demonstration of exercise (When selecting a specific exercise a timer is started
for 3 minutes various joint points are captured and a angle is measured between joints)

Figure 8.10(b) : Demonstration of exercise(Similarly when exercise is performed angle change

and after a specific angle the system captures it as one repetition)

52
Pimpri Chinchwad College of Engineering,PCCOE,Pune
Figure 8.11 : Demonstration of exercise with assistance(A person can also have a helper the
assist their limbs if one is unable to perform)

53
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 9. CONCLUSIONS

9.1 CONCLUSIONS

The AI Physiotherapy Tracker (Intelligent Healing Spaces) represents a significant advancement in

AI-driven physiotherapy and rehabilitation. By integrating real-time pose estimation, movement
analysis, and AI-based posture recognition, the system provides a highly effective and interactive
rehabilitation tool for patients recovering from injuries or undergoing physiotherapy sessions.

The system utilizes BiLSTM and LSTM architectures for accurate movement classification and
repetition counting. The pose estimation framework, powered by Mediapipe, extracts skeletal key
points in real time, allowing for precise tracking of joint movements. By leveraging both
coordinate-based and angular features, the model achieves a higher degree of accuracy and
generalization across different body types and execution styles.

Our experimental analysis reveals that the BiLSTM model with mixed features outperforms
traditional LSTM architectures, demonstrating improved accuracy and robustness in tracking
movement patterns. The system is capable of detecting multiple physiotherapy exercises,
evaluating execution correctness, and providing feedback based on identified posture deviations.
Furthermore, the integration into a web-based platform ensures accessibility and ease of use,
allowing users to perform exercises remotely under virtual guidance.

Despite the promising results, the project faces certain challenges. One of the primary limitations
is generalization across diverse environments, such as different backgrounds, lighting conditions,
and exercise settings (e.g., gyms, outdoor spaces). Additionally, plank exercises pose difficulties in
keypoint tracking due to occlusion, leading to reduced accuracy in posture evaluation. Addressing
these challenges through expanded datasets, refined feature engineering, and enhanced deep
learning models would further improve system reliability.

Overall, the AI Physiotherapy Tracker serves as an intelligent and automated physiotherapy

assistant, improving patient adherence to rehabilitation routines, reducing injury risks, and
enhancing overall recovery outcomes. The project’s findings pave the way for more advanced

54
Pimpri Chinchwad College of Engineering,PCCOE,Pune
AI-based rehabilitation tools, bridging the gap between remote physiotherapy and expert-guided
training.

9.2 FUTURE WORK

While the current implementation provides a solid foundation, there is significant scope for
improvement and expansion. Future enhancements can focus on several key areas. Expanding the
dataset to include a wider variety of physiotherapy exercises, different body types, age groups, and
environments (such as indoor, outdoor, and clinical settings) will improve the model’s robustness
and generalization. The real-time feedback mechanism can be enhanced by incorporating
voice-based guidance, haptic feedback (like vibration alerts), and an AI-powered virtual
physiotherapist capable of offering personalized recommendations based on user performance
trends.

Integration with wearable technology, such as accelerometers and gyroscopes, can enhance motion
tracking accuracy. Combining IMU data with computer vision will enable a hybrid model for
superior physiotherapy assessment. To improve accessibility, the AI Physiotherapy Tracker can be
extended into a mobile application, using edge computing for real-time pose estimation directly on
devices while ensuring compatibility across Android, iOS, and tablets.

A dedicated dashboard for physiotherapists can facilitate remote monitoring by providing detailed
analytics, progress tracking, and personalized insights. Live feedback sessions will enable
physiotherapists to supervise users virtually. Enhancements to AI models—especially through
fine-tuned BiLSTM, Transformer-based models, reinforcement learning, and self-learning
capabilities—will improve motion classification and feedback precision.

Additionally, support for multi-user and group exercises can make the system suitable for
team-based physiotherapy or fitness sessions. To ensure privacy and security, stronger data
encryption, on-device processing of sensitive motion data, and strict compliance with healthcare
regulations.

55
Pimpri Chinchwad College of Engineering,PCCOE,Pune
9.3 APPLICATIONS

The AI Physiotherapy Tracker offers a wide range of real-world applications that extend beyond
traditional rehabilitation and into the realms of fitness, sports, and broader healthcare. In
physiotherapy and rehabilitation centers, it supports professionals by enabling remote patient
monitoring, ensuring correct execution of prescribed exercises, and reducing the reliance on
in-person sessions—making rehabilitation more cost-effective and accessible. For home-based
physiotherapy, the system empowers users to perform guided exercises from the comfort of their
homes, offering real-time corrections and personalized feedback. This is particularly beneficial for
elderly or mobility-impaired individuals who require assistance in maintaining flexibility and
movement strength.

In the field of sports training and injury prevention, the tracker aids athletes and trainers in
analyzing biomechanics and movement patterns, optimizing training routines, and supporting
injury rehabilitation. Post-surgery recovery programs also benefit, as the system provides
structured exercises that promote safe and effective mobility training, while simultaneously
tracking recovery progress and alerting healthcare professionals to any concerns.

The tracker proves useful in fitness and personal training by helping users and trainers monitor
form, posture, and accuracy of movements. Its integration with smart mirrors and fitness apps
enables interactive and guided workouts that prevent injuries and enhance performance. For
elderly mobility and fall prevention, the system delivers guided exercises aimed at maintaining
balance and flexibility, reducing fall risk, and can even be linked to smart home systems to monitor
movement patterns.

Furthermore, the tracker enhances virtual healthcare and remote patient monitoring by allowing
doctors and physiotherapists to oversee rehabilitation progress without physical appointments. This
is especially valuable for patients in remote or underserved areas with limited access to healthcare
services. Overall, the AI Physiotherapy Tracker is a scalable and versatile solution that is set to
transform patient care, improve rehabilitation outcomes, and expand access to physiotherapy
across multiple domains.

56
Pimpri Chinchwad College of Engineering,PCCOE,Pune
APPENDIX 1

1.​ Paper 1 details:​

A bibliometric analysis Paper titled - “Exploring Emerging Frontiers: A Bibliometric
Analysis of Trends in Virtual Physiotherapy within the Metaverse Environment”,
Published by IGI Book Chapter.

2.​ Paper 2 Details: Research Paper based on the Project under title - “Intelligent
Healing Spaces: AI-Powered Pose Tracking for Physiotherapy Rehabilitation”,
Submitted to International Conference on Information, Implementation, and
Innovation in Technology 2025

57
Pimpri Chinchwad College of Engineering,PCCOE,Pune
3.​ Participated in M-Exhibit 2K25 in PES Modern College of Engineering,
Pune. Demonstrated the project in the Exhibition.

58
Pimpri Chinchwad College of Engineering,PCCOE,Pune
4.​ The Project is Submitted for Copyright by Government of India.

59
Pimpri Chinchwad College of Engineering,PCCOE,Pune
APPENDIX 2​
​
Plagarism Report​

60
Pimpri Chinchwad College of Engineering,PCCOE,Pune
REFERENCES

1.​ Yin, X., Liu, Z., Liu, D. et al. A Novel CNN-based Bi-LSTM parallel model with attention
mechanism for human activity recognition with noisy data. Sci Rep 12, 7878 (2022).
https://doi.org/10.1038/s41598-022-11880-8
2.​ Li, Y.; Wang, L. Human Activity Recognition Based on Residual Network and BiLSTM.
Sensors 2022, 22, 635. https://doi.org/10.3390/s22020635
3.​ Bazarevsky, Valentin, Ivan Grishchenko, Karthik Raveendran, Tyler Lixuan Zhu, Fan
Zhang and Matthias Grundmann. “BlazePose: On-device Real-time Body Pose tracking.”
ArXiv abs/2006.10204 (2020): n. pag.
4.​ S. Chadha, I. Raj and D. Saisanthiya, "Human Activity Recognition For Analysing Fitness
Dataset Using A Fitness Tracker.," 2023 International Conference on Computer
Communication and Informatics (ICCCI), Coimbatore, India, 2023, pp. 1-5, doi:
10.1109/ICCCI56745.2023.10128242.
5.​ Davids, J., Lidströmer, N., & Ashrafian, H. (2022). Artificial intelligence for physiotherapy
and rehabilitation. In Artificial Intelligence in Medicine (pp. 1789–1807).
https://doi.org/10.1007/978-3-030-64573-1_339
6.​ Aggarwal, R., & Ganvir, S. S. (2021). Artificial intelligence in physiotherapy.
Physiotherapy - the Journal of Indian Association of Physiotherapists, 15(2), 55–57.
https://doi.org/10.4103/pjiap.pjiap_3_22
7.​ G. Taware, R. Kharat, P. Dhende, P. Jondhalekar and R. Agrawal, "AI-Based Workout
Assistant and Fitness Guide," 2022 6th International Conference On Computing,
Communication, Control And Automation (ICCUBEA, Pune, India, 2022
8.​ K. L. Joshitha, P. Madhanraj, B. Rithick Roshan, G. Prakash and V. S. Monish Ram,
"AI-FIT COACH - Revolutionizing Personal Fitness With Pose Detection, Correction and
Smart Guidance," 2024 International Conference on Communication, Computing and
Internet of Things (IC3IoT), Chennai, India, 2024
9.​ A. Sinha, B. Kumar, R. Kundu, A. Sharma, M. Sharma and A. Raza, "A Comprehensive
Study of a Conversational AI-Driven Healthcare Bot with Pose Estimation for Precision

61
Pimpri Chinchwad College of Engineering,PCCOE,Pune
 Fitness Monitoring and Personalized Guidance," 2024 IEEE International Conference on
Big Data & Machine Learning (ICBDML), Bhopal, India, 202
10.​ Cao, Z., Hidalgo, G., Simon, T., Wei, S.-E., & Sheikh, Y. (2018). OpenPose: Realtime
multi-person 2D pose estimation using part affinity fields. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 43, 172–186.
11.​Abdillah, H. (n.d.). Workout fitness video. [Dataset]. Kaggle. https://
www.kaggle.com/datasets/hasyimabdillah/workoutfit ness-video.
12.​ Moran, A., Gebka, B., Goldshteyn, J., Beyer, A., Johnson, N., & Neuwirth, A. (2022).
Muscle vision: Real-time keypoint based pose classification of physical exercises. arXiv
preprint arXiv:2203.12111

62
Pimpri Chinchwad College of Engineering,PCCOE,Pune