This geometric computation method is faster than raycast.

Limitations: 

-It won't delete points occluded by objects within the field of view. 

-Points behind the camera are not deleted. I added a second node group (Camera Geometric Culling Plus) that deletes these points inside a box volume if needed, when animating a camera ride for instance. 

How to use it:

The node need to be driven by the actual camera focal length and image ratio (is width divided by height, this can be calculated with drivers either). 

The scale factor is usefull to adjust the culling effect outside the camera's edges to avoid instances popping on screen edge during animations. 

Note:

The math and geometry node setup comes from Claudio Piciarelli's youtube tutorial:

https://www.youtube.com/watch?v=dxIni7l1ZVs

Special thanks to him.