AI & data
Hand gesture recognition
A complete vision pipeline, from data to webcam: a CNN recognizes three hand gestures in real time.
- Trained on synthetic data
- Individual project
- gestures: fist, open palm, V sign
- 3
- generated images
- 1,200
- metrics, plus a confusion matrix
- 5
- frames of temporal smoothing
- 5
01Starting point
Context & mission
The problem
Controlling a machine with a gesture takes two very different things: a model that classifies well, and a real-time app that stays stable frame after frame.
The mission
Build both: the model development workflow, then live inference from the webcam.
My role
Designed and built on my own.
02Execution
What I did
Synthetic data
1,200 programmatically generated 28×28 grayscale images (circle, rectangle, V shape), with variations in position, size and noise.
Preprocessing
Grayscale, CLAHE, Gaussian blur and normalisation with OpenCV.
Training and evaluation
A CNN with augmentation and callbacks, evaluated on accuracy, precision, recall, F1, AUC and a confusion matrix.
Real-time inference
HSV skin segmentation to find the hand, region extraction, prediction, then smoothing over the last five frames to avoid class flicker.
03Design
Architecture
Two connected workflows: numbered notebooks to develop the model, and a modular Python app (detection, preprocessing, prediction, display) launched from app.py.
- Webcam
- HSV detectionSkin · region
- PreprocessingCLAHE · blur
- CNN3 classes
- SmoothingLast 5 frames
04Impact
Results
- A complete chain, from data generation to live prediction.
- Predictions stabilized frame after frame.
05Wrap-up
Takeaways
What this project provesThat I can assemble a full vision pipeline, and say clearly where its evidence stops.
06Technology
Full stack
- Python
- TensorFlow
- Keras
- OpenCV
- CNN
- NumPy
- Jupyter