Humans don't only have vision. Humans have an ocular-vestibular system that makes it possible to perform SLAM[1]-like processing; when in motion, your balance and visual input like the parallax effect can help the brain build a mapped 3D volume and determine your position, orientation and velocity and your relation to other objects within that volume.

VSLAM, visual SLAM, which is what you have to employ if relying largely on vision-only, is so far cruder than that.

I'm sure Tesla has some sensor fusion that also employs GPS and short-range radar/ultrasonar to help.

You can get some velocity and positional data through GPS, but it will have quite a lot of latency and is very coarse (at least the commercially accessible part).

SLAM is much, much easier to do with LIDAR since that will give you a point cloud 3D model. With VSLAM, you have to infer the depth data and build it up over time, and it's a much harder problem to solve reliably. OTOH, LIDAR's Achilles heel is, however, heavy rain/sleet/snow. Waymo has shown off using deep learning to filter the bulk of the noise out, but it looks like it's still early days there.

Vision isn't great for sleet/snow either, to be fair.

[1]: simultaneous localization and mapping