CASE
Casino Roulette Recognition
Summary
However improbable this might seem, even such business as casinos tries to make their every customer comfortable and happy, even though sometimes people are in special needs. This is why our client needed:
- Recognition of ball placement on a roulette wheel.
- A feature of smart casino tables for people with impaired sight and hearing.
- Device built-in solution with real-time performance and high sensitivity to mistakes.
5 months
Project length
2 people
2 Deep Learning Engineers.
Technologies
- C++
- TensorFlow
- OpenCV
- YOLOv3-tiny.
- TensorFlow
- OpenCV
- YOLOv3-tiny.
Tech Challenge
- Real-time performance with no GPU.
- Robust to dim light solution.
- Highly optimized C++ code.
Solution
As straight-forward classification would not work, we decided to split our solution into two parts:
- Tune YOLOv3-tiny, perfect network for fast and accurate object detection, to detect a ball and the zero cell on a roulette wheel.
- Since we always know where the center of the wheel is, we could find a sector of the circle with three points (the ball, the zero and the center of the wheel) and then by calculating the angle of the sector find out on which cell the ball stopped.
