AI Golf Launch Monitor

Project Story

I built this because I was frustrated by how little feedback I had after a shot. I could see the ball slice or hook, but I wanted to know what changed in the first few frames after impact. Since commercial launch monitors are expensive and not always available, I tried to see how far I could get with a normal video, computer vision, and a physics model.

Why I Built It

The hard part is that a golf ball is tiny, fast, and often blurred. Classical OpenCV could find it in some clips, but shadows, grass texture, and motion blur caused false detections. On top of that, a simple projectile model is only an approximation because real ball flight depends on drag, lift, spin, and wind.

What I Built

I built a software-only pipeline: load the swing video, detect the ball, calibrate pixels into metres, fit the first part of the trajectory, and turn that into launch speed, angle, and carry. When classical tracking became unreliable, I labelled my own golf-ball dataset and trained a YOLOv8 model to get cleaner detections.

What I Did

Created and labelled a custom golf-ball detection dataset
Trained and fine-tuned YOLOv8 for golf swing video
Wrote the calibration, trajectory fitting, and carry-estimation code
Compared the outputs with TrackMan-style reference data and kept refining the assumptions
Built the analysis pipeline and demo presentation for AWS KIRO Hackathon

How I Built It

Build Flow

Video input -> frame extraction -> YOLO detection (ball/club) -> impact frame identification -> trajectory tracking -> physics-based parameter estimation.

How It Works

The system finds the ball frame by frame, focuses on the first part of flight, and converts image movement into real-world speed using a scale factor from the scene. I use least-squares fitting over the early trajectory instead of trusting one noisy frame. From there, the physics model estimates carry, and I used TrackMan comparisons to see where simple projectile assumptions started to break down.

Software

Python
YOLOv8
OpenCV
Ultralytics

What Was Hard

The ball is only a few pixels wide in many frames and disappears quickly after impact
Pixel-to-metre calibration has to be right or every downstream number is wrong
Motion blur and shadows made early OpenCV tracking fragile
The physics model had to be honest about what it did not know, especially spin and drag

Detection pipeline diagram — Build Reference

Pipeline from video input to detection, tracking, launch-parameter estimation, projectile modelling, and output metrics.

Click to inspect

Demos and Build Notes

Product Demos

Final end-to-end launch monitor demo: video ingestion, slow-motion factor handling, AI ball detection, trajectory tracking, calibrated metres-per-pixel scaling, least-squares launch estimation, and projectile-flight output.

First successful stable ball tracking using a tuned OpenCV pipeline with contour filtering and a launch-zone region of interest.

Classical OpenCV tracking versus AI-assisted tracking. The trained YOLOv8-Nano detector reduced false positives and provided cleaner ball centroids for the physics model.

Code walkthrough of the video-processing, detection, calibration, regression, and projectile-motion pipeline.

Photos and Notes

Golf launch monitor ideation notes — Initial engineering brief: commercial radar systems are expensive, so the project targets ordinary video, ball tracking, pixel-to-metre calibration, and physics-based launch estimates.

Golf launch monitor flow diagram — Software flow diagram from raw video input through OpenCV tracking, velocity and angle calculation, projectile physics, and final numerical output.

Golf ball tracking progression screenshots — Tracking progression from setup localisation to noisy motion contours and then filtered trajectory attempts, motivating the final region-of-interest approach.

Pixel to metre calibration calculation — Calibration step deriving a pixel-to-metre scaling factor from known objects in the frame, allowing image-space motion to become real-world velocity.

Projectile-motion equations and code — Projectile-motion derivation and code implementation for velocity components, flight time, apex height, and carry distance.

Roboflow golf ball labelling dataset — Custom dataset preparation in Roboflow, labelling golf-ball frames for YOLOv8-Nano training to replace the less reliable contour-based tracker.

AWS Kiro Hackathon win recognition — AWS Kiro Hackathon recognition for the launch monitor project and its software-only approach to accessible golf analytics.

Technical Evidence

Velocity and launch-angle calculation code — Velocity and launch-angle calculation using least-squares regression over early trajectory points, scaled by metres per pixel, to reduce frame-to-frame jitter.

Measured golf launch monitor data compared with Trackman reference data — Performance validation against Trackman reference values for professional swings. Launch angle, initial velocity, and carry were typically within +-5%, while apex and flight time showed limitations from unmodelled spin and drag.

Before AI golf tracking results — Before AI integration: classical tracking was more vulnerable to shadows, background texture, and motion blur.

After AI golf tracking results — After AI integration: YOLO-based ball detection produced cleaner positions for downstream velocity and physics calculations.

Downloads

Trackman comparison dataset (XLSX)

Results & Impact

Winner - AWS KIRO Hackathon 2025
Functional prototype estimating key launch parameters from video
Custom dataset and trained detection model for golf-specific scenarios
In controlled tests, launch angle, velocity, and carry were typically within about 5% of reference values

What I Learned

Domain-specific training data dramatically improves detection accuracy
Physics modelling bridges the gap when hardware sensors are unavailable, but assumptions about spin and drag must be made explicit
Golf video analysis requires careful frame rate and camera positioning

Next Steps

Improve calibration with reference markers or dual-camera setup
Expand to spin estimation and club path analysis
Integrate with Golf Dual View for multi-angle capture

Back to all projects