Apppublic

config-h/Wearable-IMU-Activity-Segmentation-Pipeline

sourceHugging Faceapache-2.0updated 16d agoView on Hugging Face

0likes

App README

English · <a href="README_zh.md">中文</a>

<h1 align="center">An End-to-End Wearable IMU System for Segment-Level Activity Recognition via Multi-Scale Arbitration and a Temporal Record Layer</h1>

A multi-scale wrist-IMU pipeline that converts continuous 100 Hz signals into timestamped activity records.

<a href="https://rudykon.github.io/Wearable-IMU-Activity-Segmentation-Pipeline/">Website</a> · <a href="https://rudykon.github.io/Wearable-IMU-Activity-Segmentation-Pipeline/demo/">Browser Demo</a> · <a href="https://huggingface.co/config-h/Wearable-IMU-Activity-Segmentation-Pipeline">Models</a> · <a href="https://github.com/rudykon/Wearable-IMU-Activity-Segmentation-Pipeline/releases/tag/v0.1.0-research-preview">Research release</a> · <a href="README_zh.md">中文</a>

Participant recordings are not distributed on GitHub. Public model weights are hosted on Hugging Face and downloaded with integrity verification.

Overview

Window classifiers provide local activity evidence, but a long recording needs complete records: activity class, start time, and end time. This project builds those records with three components:

1.Multi-scale posterior models analyze 3-, 5-, and 8-second views of the same six-channel wrist IMU stream.
2.Local-Boundary Scale Arbitration (LBSA) emphasizes short-window evidence near transitions and longer context in stable regions.
3.Temporal Record Layer (TRL) converts the fused timeline into deterministic segment records for record-level evaluation.

<a href="docs/assets/fig02overallframework.png"> <img src="docs/assets/fig02overallframework.png" alt="Overall framework from wrist IMU input to activity records" width="92%"> </a> Three scale-specific models produce aligned posteriors; LBSA fuses them, and TRL constructs activity records.

Results

Evidence	Value
Sensor data	259.6 h
Independent external test	37 recordings / 114 segments
Segment performance	0.89 mean-user F1 / 0.90 micro-F1

External-test records are matched one-to-one with same-class references at IoU > 0.5. See the results page for the fixed-variant comparison, per-activity outcomes, representative cases, and limitations.

Quick Start

bash

git clone https://github.com/rudykon/Wearable-IMU-Activity-Segmentation-Pipeline.git
cd Wearable-IMU-Activity-Segmentation-Pipeline
conda env create -f environment.yml
conda activate imu-activity-pipeline
python -m pip install -e .
python tests/smoke_test.py

The CPU-safe smoke test verifies the public package and file interfaces. It requires neither participant data nor trained checkpoints.

Data and Models

Resource	Access
Participant recordings	Not stored on GitHub; follow the dataset access instructions
PyTorch and ONNX weights	Public Hugging Face model repository
Asset hashes and licenses	Asset documentation

Missing public weights are downloaded into the expected local paths and checked against model-assets.json before use.

Documentation

Page	Purpose
Method	Task, dataset, posterior models, LBSA, and TRL
Results	Independent external-test evidence and failure cases
Supplementary analyses	Development diagnostics, portability, and Android evidence
Reproduce	Installation, data, models, training, inference, and evaluation
Browser Demo	Run the real ONNX pipeline with visitor-local WebGPU/WASM compute

Citation and License

Until an archival paper citation is available, cite the v0.1.0 research preview as described on the citation page. Repository-authored source and public model assets are licensed under Apache-2.0; datasets and third-party dependencies retain their own terms.