The Technical University of Munich has developed a new vehicle warning system that can issue an alarm 7 seconds in advance

According to foreign media reports, researchers at the Technical University of Munich (TUM) have developed a new vehicle warning system. The system uses artificial intelligence to learn thousands of real-world traffic scenarios. The results show that when the system is applied to self-driving cars, it can detect potential dangerous situations that the car itself cannot detect, and issue an alert to the car 7 seconds in advance, with an accuracy of more than 85%. The BMW Group also participated in the research of this project.

(Image source: Technical University of Munich)

To ensure the safety of future self-driving cars, development work often uses complex models that allow the car to analyze the behavior of all traffic participants. But what happens if the model cannot handle complex or unforeseen situations?

Professor Eckehard Steinbach, member of the board of directors and chair of media technology at the TUM Munich School of Robotics and Machine Intelligence (MSRM), and his team have made breakthrough progress in this area. In past scenarios, autonomous driving test vehicles were tested to their limits in real road conditions, but usually the vehicle had a human driver take over, either because the car sent a signal for intervention or because the driver took the initiative to intervene for safety reasons. Thanks to artificial intelligence, the system can learn from these past scenarios.

The technology uses sensors and cameras to capture surrounding conditions and record vehicle status data, such as steering wheel angle, road conditions, weather, visibility, and speed. The AI ​​system based on recurrent neural networks (RNNs) learns to use data to identify patterns. If the system finds a pattern in a new driving scenario that the control system could not handle in the past, it will issue an advance warning to the driver about possible dangerous conditions.

“To make cars more autonomous, many existing approaches look at the vehicle’s current understanding of traffic and try to improve the models it uses,” said Steinbach. “The big advantage of our new system is that it completely ignores what the car is thinking. Instead, we develop new patterns based on data from real-life events, allowing the AI ​​to spot potentially dangerous situations that the patterns wouldn’t have recognized or hadn’t yet spotted. As a result, the system provides a safety feature that detects when and where the vehicle is acting dangerously.”

The research team and the BMW Group tested autonomous vehicles using this technology on highways and analyzed about 2,500 situations in which the driver had to intervene. The study showed that artificial intelligence can predict potentially dangerous situations up to 7 seconds in advance with an accuracy of more than 85%. The application of this technology requires the support of a large amount of data, because if the scenario has occurred before, artificial intelligence can only be limited to the system's recognition and prediction. Christopher Kuhn, one of the members of the study, said: "As more and more autonomous vehicles take to the road, the data will actually be generated by themselves. Each test of a potentially dangerous situation will bring new training examples." Central data storage helps vehicles learn from recorded data from the entire fleet.