Wearable IMU Activity Segmentation Pipeline
Public inference assets for the multi-scale wearable IMU activity-segmentation pipeline. The repository contains the selected 3 s, 5 s, and 8 s PyTorch checkpoints used by the Python pipeline and their ONNX exports used by the Android app.
Model
The classifier combines a compact one-dimensional CNN with a bidirectional LSTM. Three temporal views produce six-class posteriors. Local-Boundary Scale Arbitration (LBSA) fuses the views, and the Temporal Record Layer (TRL) converts the posterior sequence into activity records.
Input is a 100 Hz wrist-IMU stream with six channels in this order:
ACC_X, ACC_Y, ACC_Z, GYRO_X, GYRO_Y, GYRO_Z.
| Index | Class |
|---|---|
| 0 | Background |
| 1 | Badminton |
| 2 | Jump rope |
| 3 | Fly |
| 4 | Running |
| 5 | Table tennis |
Files
saved_models/: selected PyTorch checkpoints, normalization parameters, and the ensemble configuration.android_realtime_app/app/src/main/assets/: selected ONNX exports and the Android normalization parameters.model-assets.json: file sizes and SHA-256 checksums.
The 3 s fallback files duplicate the selected 3 s model for backward compatibility. The manifest makes that relationship explicit and provides an integrity check for every published asset.
Download
Download the Python assets into a source checkout:
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="config-h/Wearable-IMU-Activity-Segmentation-Pipeline",
allow_patterns=["saved_models/*", "model-assets.json"],
local_dir=".",
)
The GitHub project also provides a checksum-verifying download command and downloads missing assets automatically before inference.
Intended use
- Research and reproducible evaluation of the associated activity-segmentation pipeline.
- The public Gradio demo.
- On-device Android inference with the supplied ONNX exports.
These weights are not validated for clinical, safety-critical, or unrestricted cross-device use. Performance can change with sensor placement, sampling stability, calibration, device hardware, and population shift.
Project links
License
The model assets are released under the Apache License 2.0.