ai · United States · 2025

VueMotion

Athletic performance analysis from ordinary phone video

Duration9 months
Team3 engineers, 1 ML specialist
StackPython · PyTorch · MediaPipe · React · FastAPI
Livevuemotion.com

The problem

Motion capture that gives coaches useful biomechanical feedback normally requires a lab, markers and a specialist to interpret the output.

Coaches in the field have a phone and about ninety seconds of attention between drills.

How we approached it

We ran pose estimation on-device to avoid upload latency and to keep athlete video off our servers entirely — a privacy decision that also removed a whole class of compliance work.

The hard problem was not detection but interpretation: raw joint coordinates mean nothing to a coach. We built a rules layer that converts angle sequences into plain-language observations tied to specific frames.

Architecture

Capture

Standard phone camera, 60 fps, no markers or calibration rig.

Pose

On-device landmark extraction; only derived keypoints leave the phone.

Smoothing

Temporal filtering to remove jitter without flattening real movement.

Kinematics

Joint angles, velocities and asymmetry indices computed per frame.

Interpretation

Rules layer maps sequences to coach-readable observations.

Report

Frame-linked feedback, shareable with the athlete.

Outcome

  • Analysis turnaround went from days in a lab to under a minute on the sideline.
  • No athlete video is transmitted or stored, which removed the client's largest compliance obstacle.

60 fpson-device motion capture.

Tell us the hard part. That is the bit we want.

A 20-minute call, no deck. Bring the problem you have not been able to hand to anyone else.