Tennis Performance AI

Overview

An applied computer-vision platform for racquet sports — tennis first, with the same architecture portable to badminton, pickleball, and squash. Point a single camera at the court (240 FPS for serve-speed accuracy, but ordinary 60 FPS works for most metrics) and the system extracts ball trajectory, player positions, shot type (forehand, backhand, slice, drop), placement zone, depth, and outcome (in, out, winner, error). It builds a structured event log per rally, computes match-level statistics, and generates actionable coaching insights by mining for patterns ('on wide AD-court serves above 115 mph, the returner's backhand drifts long on the next shot'). Output is a coach-ready dashboard plus a per-shot data file you can analyze.

The Challenge

Sports vision is one of the harder applied CV problems. The ball is small, fast, and frequently occluded. Camera angles vary. Court color and lighting shift. Player tracking has to disambiguate two athletes in similar clothing. And speed measurement from a single camera requires careful calibration. None of these are showstoppers individually — together they make off-the-shelf vision models useless. We engineer a stack that's accurate on real-world phone-recorded matches, not just stadium broadcast feeds.

Our Approach

We pair a custom-trained ball-detection model with a player-tracking model and a shot-classification head. Court keypoint detection lets us geometrically calibrate the scene from any reasonable angle, which means accurate placement and depth without manual setup. Ball speed is computed using inter-frame physics rather than raw pixel velocity, so a 60 FPS phone camera gets within a few mph of radar gun readings. Rally-level analytics are generated by an LLM that mines the structured event log for actionable patterns. The coach dashboard runs in the browser; data is downloadable as CSV or JSON for advanced analytics.

Key Features

• Single-camera setup — no instrumented court required
• Frame-by-frame ball and player tracking
• Shot classification (forehand, backhand, slice, drop, lob)
• Placement, depth, and zone analytics per shot
• Ball-speed estimation from standard frame rates
• Rally-level pattern mining and coaching insights
• Coach dashboard plus raw data export (CSV/JSON)
• Architecture extends to other racquet sports

Results

Try It Yourself