AI Snooker
AI SCORING SYSTEM — Computer vision that referees snooker in real time
An edge-AI system that watches snooker tables through ceiling cameras, detects balls and cues with a custom-trained model on Jetson hardware, and drives a live scoreboard, food ordering, and cashier system for the venue.
DETECTION
SCOREBOARD
HARDWARE
CASHIER
TRAINING DATAProject Overview
// THE CHALLENGE
Players wager a small amount for table rental. The venue needed a scoreboard tablet and a cashier tablet; both stream money and balances from the server. When someone scores, the server must calculate how much each person pays based on the game rules they set.
// THE AI SOLUTION
A ceiling camera feeds a Jetson Nano running MobileNet-SSD for real-time ball detection (30 FPS) at the venue. When a customer rents a table, the scoreboard is enabled and the table lights turn on. When the AI detects a score, the Jetson sends it to the scoreboard and the cloud calculates what each person pays.
// THE IMPACT
A scoreboard tablet and a cashier tablet run the Flutter app, both streaming money and balances from the server. When the AI detects a score, the server applies the game rules and calculates how much each player pays for table rental — fully automated.
By The Numbers
System Architecture
Edge AI hardware mounted above the table, talking to venue tablets and the cloud.
EDGE DEPLOYMENT
Ceiling-mounted hardware setup for optimal bird's-eye view (~1.7m).
NVIDIA JETSON NANO
Edge AI compute module powering real-time object detection at the venue.
Data Flow
WHEN A TABLE IS RENTED
- 1Customer rents table (rent per minute)
- 2Scoreboard is enabled
- 3I/O switch turns the table lights on
- 4Jetson is always on, watching
BALANCE — CLOUD IS SOURCE OF TRUTH
- 1Cloud holds amount and balance
- 2Scoreboard streams balance from cloud
- 3Cashier streams balance from cloud
WHEN SOMEONE SCORES
- 1Jetson sends score to scoreboard
- 2Cloud calculates new balances by game rules
- 3Scoreboard shows updated balance
AI & Object Detection
REAL-TIME DETECTION (OVERHEAD)
Computer vision model tracking balls, cues, and pockets from a bird's-eye perspective.
PRECISION & COLOUR
Fine-tuned MobileNet-SSD model handling small objects and critical colour differentiation.
Detection, Live
Live view of the AI detecting balls, cues, and pockets in real time as the game is played.
REAL-TIME DETECTION — RAW MODEL OUTPUT
Detection Capabilities
OBJECTS
Detects all ball colours (Red, Yellow, Green, Brown, Blue, Pink, Black, White), cues, cushions, pockets, and D-zone.
PERFORMANCE
Optimized for Edge AI on Jetson hardware, achieving ~30 FPS for smooth, real-time tracking capabilities.
TRAINING
Trained on 4,000+ custom annotated images using cloud H100 GPUs. Custom fixes for bounding boxes and resolution.
Annotation Samples
No pre-existing dataset existed for this problem — every training image was collected and annotated from scratch.
RED BALLS
Annotation samples for red ball detection.
YELLOW BALLS
Annotation samples for yellow ball detection.
POCKETS
Annotation samples for pockets (bags).
CUES
Annotation samples for cue detection.
ANNOTATIONS CLOSE-UP
Close-up view of bounding boxes and labels.
Difficulties Faced
IMBALANCED DATASET
Some classes (e.g. red balls) had far more examples than others, which can bias the model. Careful sampling and augmentation ensured all classes were represented during training.
SSD BOX SIZES
Fixed by changing the SSD box sizes to the range that fits the objects we detect.
More Challenges Solved
SMALL OBJECTS AT DISTANCE
From the ceiling camera, balls are only around 40–60 pixels. We increased input resolution and tuned anchor sizes so the model could detect and classify them reliably.
REAL-TIME FPS ON EDGE
Running the detector at usable frame rate on the original Jetson Nano was tight. We moved to Jetson Orin Nano, which is up to 80× faster, so inference keeps up with the camera feed.
DETECTION JITTER
Raw detections jumped between frames, which looked bad and could affect scoring. We ignore sudden color changes along a ball's path so the scoreboard sees stable, consistent detections.
CAMERA DISTORTION AT EDGES
Wide-angle lens distortion made balls near the edges appear oval. Instead of lens calibration, we added distorted data to training so the model learned to handle oval shapes at the frame edges.
Real-Time Scoreboard
The AI doesn't just see balls; it understands the game.
COORDINATE STREAMING
Streams X,Y coordinates of every ball in real-time.
POT DETECTION
Tracks ball paths to detect when a ball is "lost" near a pocket.
INSTANT SCORING
When a person scores, the app sends the event to the server for calculation.
SCOREBOARD & CASHIER
Scoreboard shows what each person pays by game rules. Cashier streams balances per table, takes food orders, and sees orders placed from the scoreboards.
Venue Apps
Flutter apps running on venue tablets — all streaming live state from the cloud.
REAL-TIME STREAMING
Live X,Y coordinates of every ball streamed from the Jetson to the scoreboard and cloud.
SCOREBOARD
Streams money and balances from the server; score events trigger calculation of what each pays by game rules.
FOOD ORDERING
Food ordering is available on the same scoreboard tablet.
CASHIER
Streams balances for each table, takes food orders, and sees orders placed from the scoreboards.
More Features
PRINTING
Staff can print receipts, tickets, or reports directly from the cashier app — one workflow, one tablet.
I/O SWITCH CONTROL
Table lights and venue equipment controlled via I/O switches over local WiFi — no cloud required for on/off control.
REMOTE JETSON CONTROL
TeamViewer access to the ceiling-mounted Jetsons for updates, diagnostics, and configuration — no ladder required.
Why It Matters
NICHE AI DATASET
Built from scratch with no pre-existing dataset. Collected and annotated 4,000+ images to solve a specific, real-world problem.
EDGE PERFORMANCE
Successfully deployed a complex MobileNet-SSD model on resource-constrained Jetson hardware while maintaining real-time FPS.
ROBUSTNESS
Handled environmental challenges like lighting, exposure, and similar object shapes/colors through rigorous tuning.
Technology Stack
WANT SOMETHING
LIKE THIS?
From first sketch to app-store launch — tell us what you're building and we'll map the way there.
DASHBOARD
WORKOUTS
MEDITATION
MEALS
STATS
EXECUTIONAura
WELLNESS APP
An all-in-one health companion that integrates fitness workouts, mindfulness meditation, and nutrition planning. Personalized dashboards, activity tracking, and gamified challenges for a balanced lifestyle.