This uses Keras2c: https://github.com/f0uriest/keras2c
Described here: https://control.princeton.edu/wp-content/uploads/sites/418/2021/04/conlin2021keras2c.pdf
Keras2c seems to have a slight precision loss over using ONNX Runtime or Keras directly, and runs at roughly half the performance of what ONNX Runtime would in respect to model predictions per second but it's still very performant running at around 200-250 FPS where was ONNX could easily hit 400-500 FPS.
The benefit of using Keras2c is that it is a more portable solution, it requires less dependencies and setup time, you can just install clang and libx11-dev from your package manager, execute compile.sh and the resultant binary will just run.
I have also provided pre-compiled binaries.