I developed "Rock Paper Scissors" game machine on Nucleo L401RE board with Panasonic AMG8833 in January 2019, just after the first relase of STM32Cube.AI (December 2018) became available for developers: the demo in on YouTube. I remember my excitement on STM32Cube.AI -- AI can run on such a tiny device!
In this project, I improve the past project.
Note: STMicroelectronics also developed a toy similar to mine, but with VL53L5CX.
Step 1: Data collection of 8x8 matrix temperature data.
[AMG8833]--I2C-->[NUCLEO-L476RG] ----UART----> [Thermography GUI] --> CSV files of 8x8 matrix data saved in "/data" folder
Step 2: Training DNN and generate a Keras model in HDF format.
Step 3: Generate code with the HDF file.
Keras model (.h5) --> CubeIDE with STM32Cube.AI --> Generated code
Step 4: Add code to the generated one to make it a "Rock Paper Scissors" game machine.
Step 5: Run the code on the device to play "Rock Paper Scissors".
=> Thermography
=> Training DNN with DCT Type-II coefficients as heatmap feature
The DNN model's recognition performance is much better than that of my old project in 2019: overfitting occured in the old model. The performance has been improved by dropping DCT coefficients of higher frequencies drastically and adding another dense layer.
=> RockPaperScissors with CubeIDE and STM32Cube.AI
The hardware part of "RockPaperScissors" on STMicroelectronics NUCLEO-L476RG board with the latest version of CubeIDE/CubeMX and STM32Cube.AI.
I have been actively using Blender since the COVID19 pandemic for prototyping something in a virtual world. This time I use Blender and GIMP to make a digital twin of my "Rock Paper Scissors" game machine.
Animations of the digital twin on YouTube:
I used Python to animate text on the LCD in the 3DCG scene. The script must be run just once after the blender file is opened.
I also used Python to generate DNN programatically.
