Increasing road safety with artificial intelligence

In the context of the mobility of the future, the words “autonomous vehicles” come up over and over. However, if an AI system takes the wheel instead of a human being, then that system must have adequate safety levels − in fact, it should ideally be even saf­er than a human driver. Until now, the data required to set up new driving functions and, where necessary, to train artificial intel­ligence programs has mostly been generat­ed from test drivers. However, the driving functions are becoming ever more compre­hensive. Estimates suggest that it would take 2.5 billion kilometers of test driving just to make autonomous vehicles safe for high­way use. However, realistic simulations of road traffic could be a promising solution for achieving AI safety in a more efficient way. Such simulations would have to pro­vide valid models not only for typical traffic conditions, but also for very rare accidents and the scenarios in which they occur. To achieve this, the simulations must undergo data-driven optimization. However, very lit­tle data actually exists regarding the occur­rence of accidents and critical scenarios that do not actually result in accidents − in fact, there is no valid way of identifying critical scenarios at present.

Researchers at Fraunhofer EMI are tackling both of these issues in project KIsSME. The team is developing a filter that can specifi­cally identify critical scenarios and record the related data. The factors contributing to whether the data can be considered critical are drawn from various aspects of road traf­fic. The researchers intend to combine indi­vidual metrics with differing complexity lev­els, properties and scales to form sets of metrics for specific scenarios. These sets would be represented by a single criticality value. Some of the most important metrics here include time, e.g. the time to collision, and the physical dimension, e.g. the dis­tance between road users. In project AVEAS, the Fraunhofer EMI team is develop­ing processes such as data-driven optimiza­tion of traffic flow simulations.