Computer vision turns ordinary court cameras into a coach and statistician: real-time ball tracking, shot analysis, speed in km/h, automatic scoring and highlights.
The TrackNetV3 neural network finds the ball even at high speed and with motion blur; a Kalman filter builds the trajectory and detects shots and bounces.
Shot classification
8 shot types based on player pose and ball trajectory: forehand, backhand, smash, volley, lob, serve, bandeja, vibora — with confidence scores.
Speed in km/h
After court calibration (4 clicks on the corners), the system converts pixels to meters and calculates the true speed of every shot.
Heat maps
Where the ball travels most and how players move — a visual map of court zones for tactical analysis.
Auto-highlights
The system automatically cuts the best rallies and smashes into short clips — ready to share right after the match.
Automatic match scoring
Full 15-30-40 scoring, games, sets and tie-break — no referee needed, live score on screen.
Replay and stats
Rally timeline, JSON/CSV export, replay of any moment — for coaches and progress tracking.
Shot audio fusion
A microphone picks up the sound of the ball strike (FFT 1–8 kHz) and pinpoints the moment of contact — more accurately than video alone.
Court calibration
Homography from the 4 court corners: serve zones, “out” calls, distances and speeds — mapping video to real court geometry.
How it works
Standard court cameras + a GPU server. No racket sensors, no chips in the ball.
1
Cameras capture the court
1–4 cameras around the perimeter. Frames are streamed to the server via WebSocket.
2
GPU recognition
TrackNetV3 finds the ball; RTMPose+YOLOX track the poses of up to 3 players.
3
Real-time analytics
Shot type, speed, zones, score — all computed on the fly and overlaid on the video.
4
Results for players
Scoreboard, highlights and match stats — right after the game.
Modes and drills
From single-shot practice to a full match with automatic scoring.
🎾
Match (full mode)
Automatic 15-30-40 scoring, HUD, heat map, replay, highlights and audio.
🧱
Wall rally
A series of wall ↔ racket bounces — control and rhythm training.
➡️
Forehand
Forehand count with technique and speed assessment.
⬅️
Backhand
Backhand count with technique analysis.
🚀
Serve
Serves by court zone (after calibration) + speed.
✈️
Volley
Shots taken out of the air before the bounce — net play practice.
Technology
An open computer vision stack under Apache-2.0 and MIT licenses — no vendor lock-in.
TrackNetV3
Dedicated ball detector over 8 frames + gap inpainting (MIT).
RTMPose + YOLOX
Poses of up to 3 players, 17 keypoints, on GPU (Apache-2.0).
Kalman + homography
Trajectory smoothing and pixel-to-court-meter conversion.
Edge / GPU
Discrete server GPU (RTX). Everything is processed at the court, no cloud.
Get in touch with us
We'll show the system in action on your court and spec the equipment for your club.