| | --- |
| | tags: |
| | - onnx |
| | - gesture-recognition |
| | - time-series-classification |
| | - android |
| | - on-device |
| | - scikit-learn |
| | datasets: |
| | - ravenwing/cheedeh-IMU-data |
| | library_name: onnxruntime |
| | task_categories: |
| | - time-series-classification |
| | metrics: |
| | - accuracy |
| | - f1 |
| | --- |
| | |
| | # cheedeh-gesture-classifier |
| |
|
| | ONNX model for classifying phone air-gestures from accelerometer data. Designed for on-device inference on Android. Trained with scikit-learn (StandardScaler + SVM rbf), exported to ONNX. |
| |
|
| | **Classes:** `z`, `m`, `s`, `o`, `none` |
| |
|
| | ## Files |
| |
|
| | | File | Description | |
| | |------|-------------| |
| | | `gesture_classifier.onnx` | Inference model (StandardScaler + SVM, ONNX opset 15) | |
| | | `label_map.json` | Maps output class index (0–4) to gesture name | |
| |
|
| | ## Model Details |
| |
|
| | | Property | Value | |
| | |----------|-------| |
| | | Architecture | StandardScaler + SVM (rbf, C=10, gamma=scale) | |
| | | Input | 52 hand-crafted features from 3-axis accelerometer | |
| | | Output | Class index (int64) + probabilities (float32\[5\]) | |
| | | Test accuracy | 0.759 | |
| | | Macro F1 | 0.793 | |
| | | Training samples | ~372 | |
| | | Test samples | ~54 | |
| |
|
| | ## Usage |
| |
|
| | Input tensor: `float32[1, 52]` — 52 features extracted from a 100-point resampled accelerometer gesture. |
| | Output tensors: `int64[1]` (class index), `float32[1, 5]` (class probabilities). |
| |
|
| | For data collection and inference implementation see [cheedeh-collect](https://github.com/raven-wing/cheedeh-collect). |
| |
|
| | ## Input Sensor Requirements |
| |
|
| | - **Sensor type:** `TYPE_LINEAR_ACCELERATION` (gravity-compensated, m/s²) |
| | - **Sample rate:** ~50 Hz (interpolated to exactly 100 points before feature extraction) |
| | - **Gesture duration:** typically 0.5–3 seconds |
| |
|
| | The `none` class represents background / non-gesture motion. |
| |
|
| | ## Training |
| |
|
| | Trained on the [cheedeh-IMU-data](https://huggingface.co/datasets/ravenwing/cheedeh-IMU-data) dataset collected with the [cheedeh-collect](https://github.com/raven-wing/cheedeh-collect) Android app. Training pipeline at [cheedeh-learn](https://github.com/raven-wing/cheedeh-learn). |
| |
|
| | Class weights were balanced during training to handle imbalanced class distribution. |
| |
|
