Investigate performance with multiple humans in camera view

[hello-lamsey]

Description

Currently, there is not a reliable way to disambiguate keypoints belonging to different humans in the view of the d435i. We need to check the implementation in stretch_deep_perception, and then add logic to the pose estimator in this repository to ensure proper disambiguation between humans.

Mixing up keypoints between humans, or using keypoints for a human that is not the subject of interest for tablet showing, can lead to unexpected robot motions.

Possible Extension

Linked with stretch_web_teleop - add a feature to click on an individual human in the camera feed to show them the tablet.