Skip to content

Wearable IMU activity records

An End-to-End Wearable IMU System for Segment-Level Activity Recognition via Multi-Scale Arbitration and a Temporal Record Layer

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

Android Demo Browser Demo Method GitHub
259.6 h sensor data 37 external-test recordings 0.90 micro-F1
Live IMU stream 100 Hz · 6 channels
3 s model 5 s model 8 s model LBSA TRL
Badminton
Jump rope
Running
Continuous signal Stable records

Problem and output

Window-level predictions do not directly provide reliable activity records. A long-session system must recover the activity class, event count, duration, and boundaries from continuous sensor streams.

6-channel wrist IMU stream {activity, start, end}

The evaluated task covers five sports recorded from a 100 Hz wrist accelerometer and gyroscope. Background motion supports decoding but is not emitted as a workout record.

Method

Overall framework from wrist IMU input through three scale-specific models, LBSA, and TRL to activity records
IMU stream → 3/5/8-second posteriors → LBSA → TRL → activity records.
01

Multi-scale models

Scale-specific CNN–BiLSTMs analyze 3-, 5-, and 8-second views.

02

Boundary-aware fusion

LBSA favors short evidence near transitions and longer context in stable regions.

03

Record construction

TRL smooths, decodes, merges, refines, and filters the fused timeline.

Read the full method

Evidence

The operating point was frozen before evaluation on the independent 37-recording external test.

259.6 hsensor data
37 / 114test recordings / segments
0.90micro-F1
0.89 mean-user F1 · LBSA + TRL

Evaluated scope

Five activities under the studied device, wrist placement, and long-session protocol.

New devices, placements, users, activities, and deployment conditions require new validation.

Inspect results and failure cases

See the pipeline run

Actual Hugging Face demo output showing paper-style wrist IMU channel notation and timestamped activity records Run locally in this browser →

Browser demo

Signals, probabilities, timeline, and records in one view

Run the public 3-, 5-, and 8-second ONNX models with the bundled synthetic recording or a compatible wrist-IMU file. Samples stay on the visitor's device.

  • Six-channel signal plots
  • Decoded activity timeline
  • Record table and CSV export
Run browser Demo Demo source Compare the HF server Demo

Reproduce and inspect

Source

GitHub

Browse the package, experiments, Android app, issues, and release history.

 Weights

Models

Download verified PyTorch and ONNX assets from Hugging Face.

 Code

Quickstart

Verify the public package and run authorized-data inference.

 On-device demo

Android APK

Download the app and a synthetic sample for offline on-device inference.

Participant recordings are not distributed on GitHub. The reproduction guide separates public verification from workflows that require authorized data.