I've posted many times about camera placement, night vision performance and also (a bit less) about cleaning issues. There are plenty of inexpensive ways Tesla could enhance the camera placement and viewpoint geometry.
But the rest here is prefaced by the observation that, even given my strong opinions, I also respect and am willing to defer to Tesla's engineering decisions - but it would be a lot easier to do so if they would have covered these topics in a little bit more detail at one of their presentations. I I would urge them to address this at the next autonomy day if there is one coming up.
The periscope concept you mentioned - I won't say it's crazy, but I think it's probably overcomplicated, relatively expensive and doesn't contribute to the vision in a consistent full time way. You identified that it's needed at intersections, but actually the designers want to get the most out of every camera position, full-time. Accidents don't just happen at intersections, edge cases are just that, so why build a complex mechanism to tuck away the camera?
My suggestion was corner cameras integrated into headlights and maybe tail lights. I think this is highly advantageous regarding viewpoint geometry for both looking at cross traffic and, in the forward direction, looking past left-turning traffic to see the high-speed oncoming lanes that are dangerously blocked by those waiting cars. I also think the keeping-clean problem it's helped by a) known surface-treatment methods and b) widely known and relatively inexpensive washer modules if needed.
Though it's still not entirely clear what is coming with HW4, it looks like Tesla has continued to ignore these expert suggestions
. However, if I understand the early leaks about forward plus rear cameras in the repeater modules, these could do almost as well in looking to the side (but not quite as well placed to look past roadside obstacles and adjacent cars), and at least an improvement over the central windshield cameras for looking past oncoming left-turning traffic for UPLs (but still somewhat restricted and sensitive to the ego car's pose).
Regarding night vision, I don't think thermal IR sensing is quite as important as some people do (Sandy Munro is a strong believer in FLIR). I do think that near-infrared (optical) night-vision sensitivity is helpful, and could easily be improved with matching-wavelength IR LED emitters in the headlights, and even in the repeaters etc. This is like common and now cheap surveillance cameras. The beam pattern of the IR emitters doesn't have to be dipped. (I see this as a potential alternative to the smart Matrix adaptive headlights, which seem to have really stalled in terms of widespread adoption. It's not clear that the new Highland M3 even has them.)
I do agree with others that Foveon sensors are not as efficient, nor as versatile in the engineering of color channel selectivity. But I think the biggest problem is that they're proprietary and not widely adopted, so they don't meet the ultra-cheap characteristic that Tesla would be looking for.
Regarding depth sensing, I'm very hopeful regarding the forward HD radar, but we don't have much more detail about that. I don't perceive that there's much chance of side-ooking HD radar. Note that even the more economical Lidar, like the first math production units from Luminar, seem to be configured as primarily forward looking sensors, and it does seem that HD radar is probably I'm more sensible and economical alternative in this current design generation.
