In order to support level 4 and 5 autonomous vehicles some vendors are looking to machine vision and deep learning to provide the needed object recognition. This article in Electronic Design explains that in addition to machine vision, radar and lidar will also be needed to provide the needed imagery. Needless to say this is a very complicated problem.
As an alternative MobilEye (recently acquired by Intel for $15.3 billion) plans to rely on crowd sourcing for a real time updating of the highway network.
If there are challenges with the more traditional approach what will the problems be like aggregating crowd sourced data and updating the high definition maps in real time? It’s the last 20% of the problem that there are still no solution for.