AI to give autonomous cars 'social skills'

(Hier finden Sie die deutsche Version des Beitrags)

Future autonomous cars will be able to interpret the behavior of other road users, especially pedestrians, in order to react better to them. The Fraunhofer Institute of Optronics, System Technologies and Image Exploitation (IOSB) in Karlsruhe has been working on this since April 2021, and has now presented a prototype of its system. The institute talks about teaching self-driving cars "social competence.

This will be made possible with a stereo camera that can "see" spatially and thus detect the exact position of people passing by, and an AI algorithm that can detect the positions of limbs and draw conclusions from them, Fraunhofer IOSB explains. The institute has already been able to demonstrate it successfully, and now the AI is being trained further and the system refined. Later, it should be able to recognize pedestrians' intentions as accurately as possible in all conceivable situations.

Fraunhofer IOSB is using a system called Advanced Occupant Monitoring, which it previously worked on. This involves recording the body posture of vehicle occupants in 3D, analyzing their movement behavior and classifying their actions. The idea is not only to detect drivers who fall asleep, but to distinguish between different activities and their associated levels of distraction. Safety systems and comfort functions could benefit from this, says Fraunhofer IOSB.

Automated cars in mixed traffic scenarios

The research project, called "INITIATIVE," is about recognizing, for example, whether pedestrians want to cross a street. A human driver usually does this intuitively and without thinking, based on location, line of sight and gestures. Fraunhofer IOSB wants to teach cars to do this using artificial intelligence. This requires a great deal of training data. The prototype now presented can estimate whether a pedestrian wants to cross the street.

In the long term, the Fraunhofer research team is aiming for more: AI-supported "adaptive communication" is to be made possible for different road users so that automated vehicles can be integrated in mixed traffic scenarios. To that end, communication interfaces will be developed for a vehicle's interaction with other road users and with its own occupants. This includes a car that uses a light indicator to tell a pedestrian who wants to cross the street that it is going to stop or pass.

Researchers at the University of Tokyo, among others, are also working on a technology that will make robot cars assessable. For a test, they equipped an electric golf cart with two spherical glowing eyes at the front and determined that the road safety of pedestrians could be increased by moving robot eyes. Jaguar Land Rover opted for light bars in a similar research project. Six years ago, the startup drive.ai presented a display panel for autonomous cars. Fraunhofer IOSB has also developed software to automatically detect dangerous situations in video images and alert police officers.

(anw)