Research · 2025

IMU Gesture Classifier

Fifteen wrist gestures, live, from six channels of a watch, split into when and which.

Public

AIIMUPython

What it does

Recognizes fifteen hand gestures from a smartwatch in real time, from six channels, linear acceleration and angular rate, at 50 Hz over UDP.
Two stages: a small binary entry detector runs on every window and decides when a gesture has started; only then does a fifteen-class model run on the 2.5 s that follow.
Six architectures in the grid, MLP, LSTM, GRU, TCN, 1D CNN and CNN-LSTM. Both training scripts write the full grid, not the winner alone.
Deployed pair: LSTM entry detection and TCN classification, the models that drive IVO's gesture commands.

A detection opens a 2.5 s buffer; the classifier runs over several windows inside it and the highest-confidence result wins, so the gesture need not start exactly where the detector fired.
After a recognized gesture the detector is suppressed for a configurable interval, two seconds by default, so one movement cannot produce a burst of detections.
Fixed protocol: 70/15/15 split, seed 42, early stopping after 15 epochs without improvement, accuracy and macro F1 with a confusion matrix. No benchmark numbers are committed, so none are claimed.

Recordings, pretrained checkpoints and training reports are on Google Drive, linked from the README.
Eight tests run without recordings or a checkpoint.