How to optimize stream localization of objects in TensorFlow, OpenCV and Python [CLOSED]?

Question

Python

Hello again! In an incomprehensible magical way, I installed OpenCV. I found a guide ( original in English ), everything works well, but there are two problems: 1. localization is sooooo crooked (there are suspicions that this is due to the training set) 2. Detection on normal images takes a decent amount of time, 1 - 8 seconds and the process itself eats 1 GB of RAM and loads the processor by 100%! !!!
But TensorFlow and OpenCV seem to be written in c++ and very well optimized, and the snake is just a wrapper?
Several questions:

1. Why is everything so crookedly detected?
   
2. Is it possible to implement video processing on Raspbery Pi with these tools, or rather on Orange Pi with 1 GB of RAM and at least 15 - 30 FPS?
   
3. How to mono optimize this code?
   
4. Are there analogues more adapted for these purposes?
   
5. Is c++ needed here?
   
6. Are there examples of something like this?
   

I would be very grateful to those who can answer!
PS The ultimate goal is to control the robot depending on the environment, interaction with other objects (doors, fire extinguishers).

Answer 1

you have the answers in your article, see in short,
on other datasets it can be faster, but what you want is in five years

Answer 2

Look at Intel Movidius, it's a special device that can be paired with Raspberry. It can significantly increase the recognition speed
https://www.pyimagesearch.com/2018/02/12/getting-s...
Judging by the description, the piece of iron is interesting for its money, but it is only for recognition, you cannot learn the network with it .