[quantumduck]

Yeah, vision is indeed used only to localize, not really to recognize the track. Low-key this is what I don't like Waymo/Cruise approach to autonomous driving - they use HD maps to localize themselves and I see that as finding an easy way out instead of tackling the harder problem.

[danuker]

A humorously-named method I like is LSD-SLAM: Large-Scale Direct Simultaneous Localization and Mapping.

It is fast, and it produces the richest point clouds I've seen, from even monocular video. The "Direct" comes from directly using pixels, not features such as SIFT/SURF.

I suppose it could be used in environments with not too many reflective or transparent objects.