Driver assistance systems and certain automated driving functions need precise information about the current visibility of the driver. This information is almost impossible to determine with the sensors of an autonomous vehicle (especially LIDAR and RADAR).

Therefore, we trained a neural network using synthetically generated training data and real driving videos. The neural network enables the prediction of visibility with a maximum tolerance of five meters.

A realistic virtual environment was generated that produces and automatically labels large amounts of data. This data was supplemented by real driving videos and trained afterwards in the Microsoft Azure Cloud.