GestureFlow logotype

GestureFlow

Educational

Completed

An application that recognizes five hand gestures in real time from a regular webcam, with 99.5% accuracy on people the model never saw. Team project for the computer vision module at IMT Mines Alès.

View Project →

Project Genesis

GestureFlow is a project from the Computer Vision specialization module at IMT Mines Alès. Two topics were offered, gesture recognition and emotion recognition: with Jules, Martin and Matthias, we chose hand gestures. The tools were set, MediaPipe to locate the hand and scikit-learn for the learning, and the project ended with a final defense.

We built the whole chain ourselves: recording our own poses (735 in total, across the four of us), normalizing them, training a model, then recognizing gestures live from the webcam. Normalization was key. By bringing every hand to the same scale and orientation, a small hand far away and a large hand up close become comparable, and the model only has to learn the shape of the gesture.

To measure the result, the model is trained on three people and tested on the fourth, in turn. That is what this project taught me most: a model tested on the people it was trained on gives a flattering picture, and only an evaluation on strangers shows whether it really generalizes. Under these conditions, it recognizes 99.5% of poses.

Team

Technical Choices

Detection
MediaPipe Hands, 21 hand landmarks in 3D
Normalization
Landmarks rescaled and rotated into a reference position
Features
78 values per sample, including 15 angle and distance measures
Model
Random Forest with 200 trees (scikit-learn)
Evaluation
Trained on three people, tested on the fourth, for each of them

Project Roadmap

  1. Data collection

    735 hand poses recorded from four people, for five gestures.

    Completed
  2. Training and evaluation

    4 poses misread out of 735, on people left out of training.

    Completed
  3. Live recognition

    Real-time gesture prediction from the webcam.

    Completed