Computer Vision · Sports Analytics

Measuring swimmer velocity from a single poolside video

A desktop tool that turns poolside video into real-world speed data — without unreliable wearables, lengthy setup, or days of manual post-processing.

Philip Paskov in collaboration with Marek Polach Student Competition Showcase
Hip Light Tracking application showing video playback with velocity graph overlay
The analysis workspace — synchronized video playback and live velocity graphs.

The problem

Measuring how fast a swimmer moves through the water usually means strapping on cables or other body-mounted sensors. In practice, that workflow breaks down quickly.

  • Unreliable in the pool. Cables snag, adhesives fail, and sensors drift or drop out mid-set — especially during turns, underwater phases, and high-intensity training.
  • Slow to set up. Fitting hardware, syncing devices, and checking signal quality can take longer than the swim itself, which makes routine use impractical for coaches.
  • Slow to analyze. Raw sensor data often needs export, alignment, and manual cleanup before you get a usable velocity curve — turning one session into hours of work.

Video looks like the simpler path, but poolside footage usually comes from a moving camera. Track pixels naively and you measure the camera pan as much as the swimmer — so neither traditional sensors nor raw video alone give a trustworthy answer.

The solution

Hip Light Tracking detects a small light on the swimmer's hip and reference markers along the lane rope. The app includes parallax calibration (lane width + camera distance) and water surface marking to convert pixels → real-world meters correctly.

Parallax calibration

Two marker rows at different depths give separate px→m scales, so perspective distortion is handled properly.

Water surface marking

Two clicks on the water line anchor depth and help relate hip position to the pool geometry.

Camera motion correction

Reference markers move with the camera; subtracting their velocity isolates the swimmer's true speed.

Parallax calibration interface with front and back marker rows marked on pool video
Calibration screenshot Drop image at assets/calibration.png
Interactive calibration turns pixel motion into meters — even with a moving camera.

See it in action

From raw pool footage to corrected velocity curves and annotated output video — the full pipeline runs in a single desktop session.

End-to-end demo — detection, calibration, velocity correction, and export.

Session outputs

Every session exports a corrected_velocity.csv plus a rendered video with overlaid metrics — ready for coaching review or further analysis.

Data

corrected_velocity.csv

Frame-by-frame hip position, velocity, and acceleration in real-world units (m/s).

frame,x,y,vx_m,vy_m,ax_m
142,812.4,391.2,1.84,-0.03,0.12
143,818.1,390.8,1.91,-0.02,0.09
144,824.0,390.5,1.88,-0.01,-0.04
Video

Annotated export

Rendered MP4 with detection overlays, marker labels, and live speed graph burned in.

  • output_with_markers.mp4
  • output_trajectory_graph.mp4
  • output_velocity.mp4
Exported velocity graph and annotated output video
Example session export — velocity data and annotated video side by side.

Built with

A focused stack chosen for fast iteration on detection quality and a responsive analysis UI.

  • RF-DETR Custom-trained detection for hip light, rope markers, and floor references
  • PyQt6 Desktop UI for video playback, calibration workflows, and batch processing
  • pyqtgraph Live velocity, speed, and acceleration graphs synced to video frames
  • OpenCV Frame I/O, overlay rendering, and export pipeline
Application architecture and technology stack overview
Pipeline working together in one video frame.