head pose orientation solution for an IP camera

[AhmadShaik]

Hi,

I am trying to develop a head pose orientation solution for an IP camera. The process I have used is as follows.

Used a Yolov8-pose model to detect the 2D skeletons
Used MotionBERT to convert the skeletons from 2D space to 3D space
Applied few conditional logics and ML model to detect the head orientation.
This apporach is giving consistent results, when the person is looking stright towards the camera, but its failing when person is looking at the other sides. One thing I have observed was, whenever some body parts are not visible the skeleton points locations are not accurate because of that the orientation detection is not accurate.

Here are few questions I have.

Is the head pose estimation pipeline good or are there anything missing
Is there any way to improve the accuracy when bodyparts are not visible. Or occlusion will be a limitation for 2D skeleton detection.