I would take the example of the most common roulette wheel and based on it I would make a computer. Which means it would only work against one type of roulette wheel, so I would only look for that type of roulette wheel. The reason is that the size of a roulette wheel is not always the same.
I would create a server with which my computer / cell phone would communicate. There would be a 3D roulette on the server with the wheel I want to beat, and when the camera recognizes the wheel, the ball, their speed, the exact position of the ball with respect to the diamonds and the exact position of the wheel with respect to where 0 is relative to the diamonds, this data was sent to the server, and the server would simulate / calculate where the ball would fall exactly or approximately. Everything would take place quickly to get an early “prediction”.
It would be ideal for 2 people to play in the casino so that one bets and the other stands still so that the camera has no recognition problems.
It is by no means easy to realize this, especially since I only have experience with programming, but not artificial intelligence, physics, …