🧠 AI Exercise Support System

A real-time AI-powered exercise posture correction system using Human Pose Estimation and Deep Learning. This project was developed as a Graduation Thesis in Data Science & Artificial Intelligence.

---

🚀 Project Overview

The system analyzes body movements through webcam input and provides real-time feedback to improve exercise posture and reduce injury risk. Main features:

• Age & Gender prediction using CNN (MobileNetV3)
• Real-time pose landmark detection (33 keypoints)
• Joint-angle based posture evaluation
• Repetition counting
• Temporal smoothing using Exponential Moving Average (EMA)
• REST API deployment with Flask

---

🏗 System Architecture

User Input (Webcam)
        ↓
MediaPipe Pose Landmark Detection
        ↓
Feature Engineering (Joint Angles)
        ↓
Posture Classification & Rep Counting
        ↓
Real-time Feedback (Frontend)

---

🛠 Tech Stack

• Python
• TensorFlow / Keras
• OpenCV
• MediaPipe
• Flask (REST API)
• NumPy / Pandas
• Git

---

🧪 Model Details

1️⃣ Age & Gender Classification

• Architecture: MobileNetV3-based CNN
• Training epochs: 200
• Metrics: Accuracy, MAE

2️⃣ Pose-Based Exercise Analysis

• 33 keypoints extracted via MediaPipe
• Angle-based feature engineering
• EMA smoothing for temporal stability
• Rule-based posture evaluation

---

📦 Installation

git clone https://github.com/TanHuy2k2/Pose_DATN.git
cd Pose_DATN
pip install -r requirements.txt

Run application:

python app.py

📈 Future Improvements

• Replace rule-based posture logic with trained pose classification model
• Deploy using Docker
• Add real-time performance optimization (GPU inference)
• Improve dataset diversity

👨‍💻 Author

Nguyễn Tấn Huy
AI Developer Intern
GitHub: https://github.com/TanHuy2k2
Project Report: https://drive.google.com/drive/folders/1wNNTclKOS913UO67UIV23rJhxyFKXk8M?usp=sharing