Hi organization-ticket.

Thank you very much for the file!

The problem with this shot is manifold.

The Object Tracker needs first a Camera Track (as you did); otherwise, it might not really know what is going on. However, since this is from one location with little motion, hence no data in space that is reliable.
In other words, the shot is done pretty much like a lock-off shot, meaning it has no parallax change in the main part of the image.
The Footage is obviously from a phone, and there is a more complex problem. It typically has a rolling shutter, which prevents clean tracking.
On top of that, most of those shots have stabilization automatically applied, which means the image is moved inside of an overscan, and that doesn't change the perspective inside the frame, meaning it is wrong.
If all of that is in this shot, I can only assume.
Besides, there was no lens grid used in the shot, nor was the focal length set. I did not find a mask on the track to exclude the train first, to get a better result.

I mentioned that not to critique your setup, but to illustrate the problem.

Please merge this example into your tracked scene.
CV4_2026_drs_25_TRws_11.c4d

The idea is simple: if the camera was tracked in its rotation correctly, while the track of the "Schwebebahn" or Monorail is pretty straight in that section, I used the part where the camera is perpendicular to the train, and placed a cube along its side. Again, this is on a nearly straight portion of the track. The idea here was to have the straight line established, and then move the cube with the train step by step, meaning every 50 frames or less.

You could define the position of the train with the Camera Calibrator for a single frame, to align the train. I used that "often" when a drone shot was used that just rotated a little bit in the air. Similar case here, except locking up instead of down.

Another idea is a little bit of a hack. A traveling mask tag (Mask Constrained is rotoscoped to get only the train. Then you track only what is inside the mask. The tracker assumes now that the train is like a building, and the camera moves as this is the only part of the whole shot that has parallax in it. Again, this is more of a hack, as the train is very small. But to get the train tracked to have an object representing it works.
My quick test is below.

Yes, I re-rendered the images to get a 0-303 frame number sequence. I have the feeling that two million as a frame number is not working with my computer as the frame sequence starts.
The file is a very rough draft, quickly done. Not for production, just as a demo.
CV4_2026_drs_25_TRws_01.c4d

My best wishes for your project