Case study · computer vision
Movement assessment without diagnostic overreach.
CortexMD is presented as a pose and posture-analysis prototype. It does not claim to diagnose a condition or replace clinical evaluation.
Problem
Turn video landmarks into interpretable movement feedback.
The project explores whether a real-time pose pipeline can identify exercise-specific posture deviations using landmark geometry and explicit movement rules.
Architecture
Landmarks first, then geometry, then feedback.
Implementation
Exercise-specific analysis on 33 three-dimensional landmarks.
Each frame produces 33 3D landmarks. Joint angles and spatial relationships are derived from those points, then evaluated against movement-specific criteria for more than 15 exercises.
Reported results
Fast enough for interactive feedback.
Evaluation boundary
The accuracy number needs a reproducible protocol.
The approximately 92% figure is a project measurement for posture-error classification, not diagnostic accuracy. A public release should document the test set, labeling process, class balance, subject diversity, camera placement, threshold selection, and whether examples overlap across development and evaluation.
Demonstration
Annotated movement demo in preparation.
A public demo will show landmarks, measured angles, feedback timing, and representative correct/incorrect movement examples.
Limitations
Pose estimation is sensitive to context.
Camera angle, occlusion, lighting, clothing, body diversity, and movement speed can all affect landmark quality. The project is therefore described as a movement-analysis prototype, not a clinical tool.
Lesson
Precise wording is part of responsible engineering.
A strong computer-vision demo does not need an unsupported medical claim. Clear measurement boundaries make the engineering more credible.