Point the camera
at the court

Pose tracking draws a stick figure over you, and an experimental detector follows the ball and traces its path. Your video never leaves this device.

Session time: --
Player tracking: --
Ball tracking: --
Movement: --
Player confidence: --
Ball confidence: --
Court position: --
Recording: --

AI Coach

Waiting

Start the camera and I'll watch your position and movement.

Coaching only uses what the camera can actually measure.

Settings

Camera & view Camera Mirror previewAuto mirrors the front camera

Overlay Show skeleton Show ball Ball circle Ball label Ball trail Show confidence Trail length40

Ball detection Detection sensitivity60 Ball colour

DetailHigher finds smaller balls, costs more CPU Debug viewShows the colour mask and candidates

Player & court Pose modelFull is steadier, uses more CPU Court area Coaching

Recording RecordOverlay is burned into the video MicrophoneAsks for permission when recording

Technical notes

What is real, what is limited

Pose: MediaPipe Pose Landmarker (33 landmarks) runs in your browser (WebAssembly, GPU if available). One person is tracked, the most prominent one. Landmarks are smoothed with a One-Euro filter.
Ball: classical computer vision, not a trained pickleball model. It segments ball-coloured pixels, checks size and roundness, requires motion, then follows candidates with a gated alpha-beta tracker. It is experimental: a fast, small, blurry ball can be missed, and a ball-coloured moving object can fool it. Use Calibrate ball colour and the Debug view.
Never faked: if the ball is not found it says Searching. Predicted positions are drawn dashed, labelled predicted, and are not added to the trail.
Court: not auto-detected. You click four corners; a homography then maps the player's feet to feet on the court (an estimate that assumes flat ground).
Coaching only uses measured things: position, movement, tracking quality, and approximate knee bend and stance width from the 3D pose. No paddle angles, no shot types.
Recording: the overlay is composited on a canvas and recorded with MediaRecorder, so keep this tab visible while recording (browsers throttle hidden tabs). “Camera only” records the raw stream and keeps working in the background.
Camera access needs HTTPS or localhost. Opened straight from a file, Chrome and Edge normally allow it (they ask every time); Safari and phones need a served page. Locally: py -m http.server 5180 then open http://localhost:5180.

Data flow

Architecture hooks, not built yet

Shot detection (serve, dink, volley, drive, lob, overhead, third-shot drop) → CoachEngine.registerAnalyzer()
Paddle tracking, swing and contact-point estimation, ball speed in real units (needs court scale), rally length, heatmaps, session history, AI summaries.

Point the camera
at the court

Something went wrong

Replay

Session summary

Coaching events

Technical notes

What is real, what is limited

Data flow

Architecture hooks, not built yet

Point the cameraat the court

Something went wrong

Replay

Technical notes

What is real, what is limited

Data flow

Architecture hooks, not built yet

Point the camera
at the court