WRIST IMU → WHOLE-BODY GAIT

Reconstruct whole-body gaitfrom the swing of one arm.

A wrist sensor only ever observes the arm. But that arm swing is driven by the pelvis, the trunk and the legs. gait synthesises the wrist signal a whole-body walking model would have produced, then minimises the difference against the recording — recovering the motion the sensor never saw.

Open the viewerView on GitHub
6ch
INPUT / ACCEL+GYRO
51.2Hz
SAMPLE RATE
46dof
MODEL PARAMETERS
45
OPEN DATASETS INDEXED
FIG. 01  /  Estimated whole-body motion, played back from real data
01
Why the wrist

Gait is the cheapest biomarker we have, and the wrist is the only place people keep a sensor.

Gait speed has been reported to track survival in older adults in a large pooled analysis, and changes in walking have been studied as a window onto cognitive decline and frailty.

The obstacle was always measurement. Conventional gait analysis assumes a motion-capture lab, or sensors taped to the trunk and feet. Both capture a few minutes on a day someone decided to measure.

Exactly one sensor gets worn all day, for weeks, inside ordinary life. It sits on the wrist. If whole-body gait can be recovered from there, gait analysis leaves the lab.

6
axes
three of acceleration, three of angular rate — nothing else
1
mounting site
the same position as any existing smartwatch
0
extra sensors
no trunk or foot-mounted units required

On gait speed and survival: Studenski S, et al. Gait Speed and Survival in Older Adults. JAMA. 2011;305(1):50–58. — prior work by others, not a result of this project.

02
Method

Solving the inverse problem by iterating the forward one.

Rather than regressing joint angles directly from the wrist signal, an anatomically plausible whole-body model is run forward and its parameters are moved until its output matches the recording. Every solution is therefore a way a body could actually have walked.

  1. 01

    Observe

    Six-axis time series — acceleration and angular rate — come off the wrist device. Gravity is separated, the signal is band-passed, and walking segments are detected.

  2. 02

    Forward model

    A whole-body walking pattern parameterised by cadence, stride length, arm swing and trunk lean drives a skeleton through forward kinematics.

  3. 03

    Virtual sensor

    A virtual IMU is mounted at the model's wrist. Its rigid-body motion yields the signal this model would have produced, mount tilt included.

  4. 04

    Minimise

    The difference between synthesised and measured signal becomes the objective. Parameters are optimised inside biomechanical range-of-motion limits; where it converges is the estimated gait.

02 → 03 → 04 → 02iterate until convergence
03
Parameters

The output is not a number. It is a moving body.

The optimiser returns a complete parameter set for the walking pattern. From it the full joint trajectory is reconstructed, then played back and inspected in the viewer.

Cadence
cadenceHz
Stride length
stride_lengthm
Arm swing amplitude
arm_swing_amprad
Trunk lean
trunk_leanrad
Vertical sway
vertical_ampm
Lateral sway
lateral_ampm
Hip flexion
hip_swing_amprad
Knee flexion
knee_swing_amprad

Pelvis and torso rotation, elbow flexion, forearm pronation and the heel-strike impulse response are modelled alongside these.

04
Validation

How much of a known parameter set comes back.

A round-trip test synthesises a signal from known gait parameters, runs estimation on it and compares against ground truth. These are the tolerances that test actually enforces.

quantitytolerance
Cadence< 0.05 Hz
Arm swing amplitude< 3.0°
Trunk lean< 4.0°
Vertical sway< 0.005 m

This is round-trip accuracy against synthetic signal. It is not a claim of clinical validity on human recordings; evaluation on real subject data is ongoing.

05
What comes next

Building shared ground for wrist-based gait analysis.

A method alone does not move a field. It needs shared data, a shared yardstick, and an implementation anyone can rerun. gait aims to hold all three in one place.

LandscapeAvailable

Who measures gait, and with what

The recorders, reference systems and open algorithms actually used in gait research, and the labs to watch. Every entry is verified against a primary source; what remains unverified is listed explicitly as a gap.

Open
DatasetsAvailable

An index of public datasets

Public datasets containing wrist-worn IMU during walking, indexed by mounting site, sampling rate, ground-truth type and access route. Every field is checked against a primary source and carries the date it was checked.

Open
BenchmarkIn preparation

Reproducible evaluation

An evaluation suite that compares methods on identical splits and metrics, scoring step count, cadence and walking speed per event.

PapersIn preparation

Papers with runnable code

Publications from this platform, each paired with the code that regenerates its figures.

Open sourceAvailable

Library and API

The Python library, the CLI, the FastAPI server and this viewer are all public.

06
Get involved

Looking for people who have something they need to measure.

This platform grows on research funding and real recordings. There are three ways in.

Research collaboration

Open to pharmaceutical, insurance and healthcare groups who want gait as an outcome measure, and to manufacturers with a wrist device. Endpoint-measurement pilots included.

Start a conversation

Contribute data

Labelled wrist-IMU walking data is the scarcest resource in this area. If you hold a dataset that could be shared, please get in touch.

Talk about data

Sponsor

Funds compute and the upkeep of public data. Even small amounts function as a signal that this field needs shared infrastructure.

Support the work