How to handle collisions based on what is visible on screen

Hello everyone,

I was wondering if it was in anyway possible to handle/generate collisions based on what is shown on screen.
To illustrate my question here is a video of what I am trying to achieve:
link text