Pose Estimation using MoveNet (ONNX)

This project demonstrates human pose estimation using a MoveNet ONNX model. It supports image and video inference and provides structured pose outputs (keypoints + confidence) along with visualization.

🚀 Features

• MoveNet ONNX inference (CPU)
• Image and video pose estimation
• Keypoints parsing (17 COCO joints)
• Skeleton visualization
• Clean class-based architecture
• Compatible with swatahVision style parsers

---

📥 Model Download

Pretrained models for swatahVision are available in the Model Zoo.

🔗 https://visionai4bharat.github.io/swatahVision/model_zoo/

---

📁 Project Structure

pose_estimation/
│
├── Movenet.onnx                # Pose estimation model
├── movenet_pose.py             # Inference class (video / image)
├── pose.py                     # Pose parser (from_movenet)
├── assets/
│   └── sample.jpg              # Example input image
└── README.md

---

📦 Requirements

Install dependencies:

pip install onnxruntime opencv-python numpy

(Optional if using swatahVision)

pip install swatahVision

---

⬇️ Download the MoveNet ONNX Model

Option 1 — Download pre-converted ONNX (recommended)

1. Download MoveNet ONNX model from a model hub (example sources):

2. ONNX model zoo mirrors

3. TensorFlow → ONNX community conversions

4. Your internal model storage

5. Rename the file to:

Movenet.onnx

1. Place it inside:

pose_estimation/

---

Option 2 — Convert MoveNet to ONNX yourself

If you have TensorFlow MoveNet:

Install:

pip install tf2onnx tensorflow

Convert:

python -m tf2onnx.convert \
--saved-model movenet_saved_model \
--output Movenet.onnx \
--opset 13

Move the generated file into the project folder.

---

▶️ Run Pose Estimation (Image)

Example:

import swatahVision as sv
from pose import Pose

model = sv.Model(
    model="pose_estimation/Movenet.onnx",
    engine=sv.Engine.ONNX,
    hardware=sv.Hardware.CPU
)

image = sv.Image.load_from_file("pose_estimation/assets/sample.jpg")

outs = model(image)

pose = Pose.from_movenet(outs)

print(pose.keypoints)
print(pose.confidence)

---

▶️ Run Pose Estimation (Video)

python movenet_pose.py

Press q to exit.

---

🧠 Output Format

MoveNet returns 17 keypoints:

[x, y] coordinates
confidence score

Keypoints follow COCO order:

• nose
• eyes
• ears
• shoulders
• elbows
• wrists
• hips
• knees
• ankles

---

🛠️ How it works

Pipeline:

1. Load ONNX model
2. Preprocess frame (resize → normalize)
3. Run ONNX inference
4. Parse keypoints via Pose.from_movenet
5. Draw skeleton

---

✅ Supported Tasks

• Pose estimation
• Real-time video pose
• Visualization
• Framework integration (swatahVision)

---

📌 Notes

• Model expects 192×192 input
• Works on CPU
• Confidence threshold can be adjusted
• Compatible with other pose ONNX models with similar output

---

🤝 Contributing

1. Fork the repository
2. Create a branch
3. Make changes
4. Open Pull Request

---