What you are probably referring to is not SLAM but just localization with HD Maps. That is now at a pretty good level of reliability in many of the projects out there. It depends on the quality of the maps of course, and the quality of the localization.
However, LIDAR is still preferred. Those attempting to do it with cameras are aware of the issues with variation of illumination and choose algorithms to work with that. Many algorithms will not work in a freshly snow covered road, or after an ice storm, or in heavy rain or snow etc. but some are better than others. People consider it good enough to put robots out there, though they may shut them down in certain weather.

The answer today is “As many as can be effective together, almost regardless of cost.”

For the first few years of deployment of robocars, safety is paramount. Well, it’s always paramount, but in those early years, even more so, because it’s not about making money or slick design yet. It’s about making it work, safely, and developing public trust.

That doesn’t mean you go crazy. Sometimes extra sensors just add complexity and confusion, so you want to tune for how well you can fuse the data from the sensors.

While cost is much less important than it will be in future, it still plays some role. A $75,000 LIDAR (which is commonly used for research) is just at the upper range of what’s practical for a taxi, and above what’s practical for a private car in volumes. However, the price of these is already falling fast.

In general you need to sense a variety of things:

– Pretty much everything in front of the vehicle to a long range — over 200m for highway driving, 100m for lower speed urban driving. 100m range lidar, camera and 250m range radar are popular here, along with 250m range lidar. Camera is also required.

– You must sense the the sides (left and right) when crossing or turning onto streets with fast traffic. Typically side facing radar is popular here.

– You want shorter range sensors directly behind for backing up. Here, stereo camera and lidar can function well, and even ultrasonics.

– You want longer range detection of vehicles moving up fast in other lanes for us in lane changing. Radar is good here

– You want general awareness of the surroundings of the vehicle at short range to the sides.
Cameras, ultrasonics and wide-angle lidar are popular here, as well as shorter range phase based time of flight.

Currently the most important challenge faced by autonomous car in computer vision is how to run most of the algorithm in real time and that too in complex and cluttered environment. Making a vision model which can achieve high accuracy in no more a problem today because of deep learning but making it work in real time that too in more complex environments like in Indian roads is.
Today computer vision algorithms for autonomous vehicles are getting a direct competition from LiDAR because LiDAR can solve most of the problems which is much complex for vision algorithms and at at same time can also help vehicles with mapping and localizing.
Vision fails in SLAM thus which is why its usage in autonomous vehicle is now limited for lane marking and generation and to make low cost ADAS systems.