Implementation of AI bone recognition

The goal of the project is to create new game mechanics.

The dice physically fall out in the game tower.
The system must:

• record the result of the throw
• recognize denominations
• define a combination
• calculate the coefficient
• send the result to SAS
• display the total on the screen

Recognition should be instant and error-free, as the result directly affects financial calculations.

This is not an auxiliary tool — it is the computing core of the new game.

The project was not trivial due to:

• work in a real physical environment (glare, glass, light angles)
• The need to ignore bone reflections
• Synchronization of three cameras
• lack of margin for error
• real-time processing
• Direct integration with the financial logic of the game

Any mistake would mean incorrect calculation of winnings.

Every throw:

• it is recorded by three independent cameras
• processed by a neural network
• creates three entries in the database
• passes majority validation

Result Confirmation rule:

Match of at least 2 out of 3 recognitions → the result is accepted
Otherwise → error

This significantly increases the reliability of the system.

3 models have been developed, trained and tested:

• bone detection
• face value recognition
• classification of the combination

Stages:

• data collection in real conditions
• markup
• augmentation
• training
• GPU optimization
• Stress testing

The inference occurs on the GPU on the customer’s side.

Basic game (3 dice)

Combinations:

111 / 222 / 333 / 444 / 555 / 666 → ×1000 to the bid

If there are no threes:

4 or 17 → ×500
5, 6, 15, 16 → ×200
7, 8, 13, 14 → ×100
9–12 → ×0

Formula:

Game_Result = Σ (Dice_Koeff × Player_Bet)

Launch through a 12-sided cube.

Examples:

• three identical ones are separate coefficients
• Three pairs — 750
• Six different ones — 1,500

Additional matching dice increase the multiplier

The results of all the throws are summed up.

System:

• retrieves player data from SAS
• defines the scenario
• recognizes the result
• calculates the coefficient
• transmits the total back to SAS
• displays the visualization on the screen

It’s not just bone recognition.

This is a financially critical real-time AI system
that is directly involved in calculating winnings
and ensures the reliability of the new game.

Our results:

• Recognition accuracy: 99%
• Delay: instant (real-time)
• Errors in industrial operation: 0
• correctly ignoring reflections in glass
• The system operates in an industrial environment
• AI has become the core of the new game mechanics.