What are the possible solutions for determining the user through the webcam?

Question

Neural networks

Good afternoon, I'm trying to implement the background blur functionality for the user when broadcasting from a webcam. I came across bodypix https://github.com/tensorflow/tfjs-models/tree/mas... but this solution produces low fps.
PS Colleagues in the workshop advised OpenCV, but how can it help with this task?
Tell me which way to look? What are the possible solutions for this functionality?
Thanks for answers.

Answer 1

Apparently so
https://docs.opencv.org/3.4/df/d6c/tutorial_js_fac...

Answer 2

If the camera and the background are relatively static, you can calculate changes relative to specially taken frames and blur only what has not changed for a while (tweak the parameters, how many and what changes to consider as such) - the fastest solution, but there will be artifacts and blurring of clothes, if not stir.
Since in this task it is stupid to work with matrices - opencl is ideal, it is quite possible that there are already ready-made filters there.
The project on your link above for me is so ideal for the formulated task, most likely incorrectly configured hardware slows down (there are no drivers for the video card) or there is no GPU at all

The ResNet architecture is recommended for computers with even more powerful GPUs.