PosePilot 🧘

If you find PosePilot helpful, give us a ⭐!

An Edge-AI Solution for Posture Correction in Physical Exercises

Overview 📖

PosePilot is an edge-AI solution for posture correction in physical exercises, with a focus on Yoga. It integrates pose recognition and personalized corrective feedback, leveraging BiLSTM and Multihead Attention for robust, lightweight, and accurate posture analysis. The system is designed for deployment on edge devices and can be extended to various at-home and outdoor exercises.

Key features:

• Automatic human posture recognition
• Personalized, instant corrective feedback at every stage
• Lightweight and robust model for edge deployment

Dataset 🗂️

Our in-house dataset was created with videos from four angles, featuring 14 participants (ages 17-25) performing six poses. Frame-by-frame keypoint extraction using Mediapipe identified 33 keypoints for pose analysis. The dataset contains 336 videos, filmed indoors with controlled lighting.

For detailed information about the dataset structure, landmark mapping, and data format, see the Dataset Documentation.

Quick Start Guide 🚀

1️⃣ Set Up Your Development Environment

First, clone the repository and set up your virtual environment:

git clone https://github.com/gadhvirushiraj/PosePilot.git
cd PosePilot
conda create --name posepilot-dev python=3.10
conda activate posepilot-dev
pip install -r requirements.txt

2️⃣ Train Your Own Models

Train pose classification and correction models on your data:

Pose Classification Training

python train_classify.py

Pose Correction Training

python train_correction.py

Use --use_cached flag to skip data preparation and use previously processed datasets for faster training iterations.

3️⃣ Run PosePilot

Option A: Gradio Web Interface (Recommended)

Launch the interactive web interface for pose classification:

python gradio_app.py

The web interface will be available at http://localhost:7860 with the following workflow:

Upload Video → Extract Keypoints → Classify Pose → View Rendered Landmarks

Supported poses: Tree, Chair, Warrior, Downward Dog, Cobra, Goddess

Option B: Direct Prediction

For programmatic pose classification and correction:

• Key Files: classify_predict.py for pose classification, correction_predict.py for pose correction
• First Step: Convert video to keypoints using give_landmarks() function in utils.py
• Then: Use the extracted keypoints for classification or correction tasks

Gradio GUI

Citation 🏷️

If you use PosePilot in your research, please cite:

@InProceedings{10.1007/978-3-031-99568-2_17,
	author    = {Gadhvi, Rushiraj and Desai, Priyansh and Siddharth},
	title     = {PosePilot: An Edge-AI Solution for Posture Correction in Physical Exercises},
	booktitle = {Pattern Recognition and Image Analysis},
	year      = {2026},
	publisher = {Springer Nature Switzerland},
	pages     = {208--219},
	isbn      = {978-3-031-99568-2}
}