DIGITAL HEALTH · MVP DEVELOPMENT · COMPUTER VISION

Proving camera-guided rehabilitation on everyday devices.

We worked with Allimb to build an installable web MVP that could teach a rehabilitation exercise, observe movement through the device camera, count repetitions and provide corrective visual and spoken feedback.

Person completing a rehabilitation exercise with body pose landmarks overlaid — ALLIMB · Product MVP

Client: Allimb
Sector: Digital rehabilitation
Engagement: Product MVP
Delivered: 2020–2021

HARDWARE

Consumer devices

Phone, tablet or laptop camera

VISION

On-device

Pose estimation and classification in the browser

EXPERIENCE

Complete session

Instruction, correction and score

What the MVP established

A working MVP that could guide a shoulder exercise, recognise relevant positions, count repetitions, flag common mistakes and show a session result.

The core experience ran on standard consumer devices without dedicated motion-capture hardware.
The engagement established both the feasibility of camera-guided feedback and the practical constraints that a larger health platform would need to address.

THE PROBLEM AND HYPOTHESIS

A movement model was necessary, but it was not enough to create a rehabilitation session.

A patient using the product at home would not have a specialist configuring the camera, checking whether their body was visible or explaining every correction. The software had to teach the movement, guide device placement, determine when the patient was ready and convert frame-level predictions into a coherent set of repetitions.

Consumer browsers added practical constraints. Camera permissions, orientation sensors, low-power behaviour, media playback and large tutorial files behaved differently across devices. A model could classify one pose correctly while the surrounding session still failed.

Camera-guided rehabilitation session showing pose landmarks, recognised movement states and corrective timing feedback

SESSION AND VALIDATION FLOW

The technical assumption was tested inside a complete session.

Patient actions and system stages were evaluated together on ordinary consumer devices rather than in an isolated model demonstration.

01 · Human decision
Device check
02 · Human decision
Tutorial
03 · Human decision
Positioning
04 · Test stage
Pose estimation and classification
05 · Test stage
Repetition and correction
06 · Human decision
Session score

ENGINEERING DECISIONS

Making the MVP work on ordinary devices.

Process camera frames locally

Pose estimation and exercise classification run in the browser. This reduced dependence on continuous video upload and provided a useful starting point for a privacy-conscious health experience.

Privacy and responsiveness

Design around device failure modes

The application explicitly handles orientation, missing sensors, camera errors, iOS playback constraints and low-power behaviour. These edge cases were treated as part of the MVP rather than deferred until after the model demonstration.

MVP judgement

VALIDATION LIMITS

What this MVP did not establish

The work did not establish clinical efficacy, replace clinical judgement or validate a broad exercise library. It proved a bounded camera-guided experience and exposed the device, media and interaction risks for a larger product.

RESPONSIBILITY