All researchProject overview

Wearable biomechanics · 2025

Reconstructing 3D kinetics from smart-insole loading

Estimating 3D forces, moments, and centre of pressure from the vertical loads measured by smart insoles.

Qatar University · Pressure-only baseline for the AI-powered smart-insole project

Two panels of gait-cycle curves, left foot above right. In (a), vertical force, the Ke2KeNet estimate follows the treadmill reference more closely than the summed insole signal through loading, midstance and unloading. In (b), mediolateral moment, the estimate captures the main peaks but departs from the reference during swing. Gait events such as heel strike, foot flat and toe off are labelled.
Figure 5 from Mahmud et al. (2026). Phase-aligned treadmill reference, summed insole loading and Ke2KeNet estimates for one participant walking at 0.9–5.4 km/h: (a) vertical force and (b) mediolateral moment. Shading shows one standard deviation.View full-size figureSource article© 2025 IEEE · author reuse

Approach

How much of the force and moment profile measured by an instrumented treadmill can be recovered from insole signals?

Ke2KeNet maps the 16 vertical-loading channels of each Moticon insole to eight treadmill-referenced waveforms per foot: three forces, three moments, and the anterior–posterior and mediolateral centre of pressure. Separate one-dimensional models were trained for each output and compared with instrumented-treadmill recordings across walking and running speeds. A sensor-relevance analysis then tested whether all 16 pressure channels were needed.

  1. Smart-insole pressure signals
  2. Estimated forces, moments and centre of pressure
  3. Comparison with treadmill measurements

Reported results

For the average gait cycles, Pearson correlation exceeded 0.95 for every output except the right-foot vertical moment, and the estimated vertical force was closer to the treadmill than the summed insole channels. Reducing the layout from 16 to eight pressure locations per foot gave broadly comparable accuracy.

Average-cycle correlation, all outputs but one
> 0.95
Pressure sensors per foot after layout analysis
16 → 8
Healthy participants, five subject-independent folds
9
Insole maps showing which of the 16 Moticon sensor regions are relevant to each kinetic output, a consolidated layout with eight retained sensors per foot highlighted in green, and two variance-inflation matrices before and after channel reduction.
Figure 4 from Mahmud et al. (2026). The 16-sensor Moticon layout, the sensing regions relevant to each kinetic output, the consolidated eight-sensor layout, and variance-inflation matrices before and after channel reduction.View full-size figureSource article© 2025 IEEE · author reuse

My contribution

I led preprocessing, model development, subject- and speed-wise evaluation, analysis, and manuscript preparation.

Study scope

The results come from nine healthy participants under controlled treadmill conditions. Performance varied across participants and speeds, particularly for some moments.

Ongoing work

This study is the foundation of the AI-powered smart-insole project, which now combines plantar loading with wearable motion, tests transfer across devices and develops the insole hardware. The project overview sets out each stage.

Run the insole workflow

Use Python 3.12 and the tagged code release v0.1.0. The demo uses a small U-Net and synthetic signals to exercise the insole workflow on CPU without study data or pretrained weights.

  1. Get the release

    git clone --branch v0.1.0 --single-branch https://github.com/Sakib1263/Unimodal-3D-Kinetics-Estimation-from-Insole-or-IMU.git
    cd Unimodal-3D-Kinetics-Estimation-from-Insole-or-IMU
  2. Install and run the demo

    Create an environment with python -m venv .venv. Activate it with source .venv/bin/activate on Linux or macOS, or .\.venv\Scripts\Activate.ps1 in PowerShell. Then run the following commands.

    python -m pip install -r requirements.txt
    python -m kinetics demo --config configs/insole.json
  3. Check the output

    The command prints Saved run: …/runs/demo_Fy_LL_UNet_…. Open that directory and check that status.json reports "status": "complete". The run also writes summary.json and fold-level predictions. Those results describe synthetic inputs.

Verified on CPU. A fresh checkout of v0.1.0 completed the synthetic fold and produced finite reference and prediction arrays. The demo verification record lists the tested versions, completion status, and expected array shape.

Work with the study data

Huawei Wang’s processed treadmill and insole recordings are publicly downloadable. Use the data preparation guide to create the expected input files. To reproduce the article evaluation, match the preprocessing, target scaling and participant folds described in the study.

python -m kinetics inspect --config configs/insole.json

The article’s training-ready exports and pretrained checkpoints are separate from the release. The verification record covers synthetic inputs only.