Hey
@Hacky
You are doing a great job with drone gesture control! Very impressive. Reminds me of the Minority Report movie. Perhaps it may lead to combining gesture reading with a dancing drone swarm reacting to human movements
To answer your questions:
I am actually using MoveNet for body coordinates detection. But the bounding box is drawn using derivatives of actual points (multiples to the ears coordinates to be precise), which may cause box shaking you noticed. Probably I have to use a better formula for the box.
The fragment with a yellow dot is one of the earliest experiments using PoseNet. And the dot is at the exact point returned from the model. I switched to MoveNet later since it is newer and less resource-consuming.
I would be glad to know your suggestions on what other simpler person detection models you mentioned could we try?
Running on mobile: since this project has a server part for video feed processing, it can not be entirely run on a mobile device in its current state.
Possible option: run a server connected to
Tello on a machine simultaneously connected to
Tello and a second network (or connect your
Tello to an arbitrary AP, I can't do that with mine).
Then open a browser page on mobile connected to that network. The pose estimation is executed in the browser, so the heavy load will be shifted to the mobile device. I guess modern high-end smartphones should handle it perfectly.
Here are the official demos of MoveNet and PoseNet to estimate how they behave on mobile:
MoveNet:
PoseNet: