If you want accuracy you should track the scene, solve the camera and reconstruct the camera moves and 3d environment. Watch the "match track blend" series of videos by Sebastian Keoenig to understand the process https://www.youtube.com/watch?v=U7AjIL_fDtU
Then there is fspy (https://fspy.io/), a piece of software and a blender addon to estimate camera placement and lens based on the vanising points in an image. worth trying that too.
Manually placing the camera is not only tedious, but very hard to do accurately just "by-eye', as it involves matching field of view, distortion, camera rotation and placement on top of estimating the scale and distance to the 3d objects. The camera placement, rotation and field of view is only valid for a single point in space and not others. That is a highly complex task.
If you still want to painfully guesstimate it yourself use the camera background image feature.

It will display an image (or video frames) as background in the camera view only.
To match it with your scene lock the camera to view as explained in the following post:
https://blender.stackexchange.com/a/135150/97988