Automotive Safety and Autonomous Driving

Besides electrification and connectivity, autonomous driving is one of the hot topics in the automotive sector. In contrast to conventional cars where the driver has to perform all aspects of the driving task, fully autonomous cars allow the driver to take their hands off the steering wheel and concentrate on other activities at all times. Although a lot of research is still required to make this possible and ensure a high level of safety, several advanced driver assistance systems (ADAS) that perform at least temporarily single aspects of the driving task are already available. These systems are either implemented based on classical model-based approaches or, more recently, by methods of machine learning.

Highly automated driving functions and ADAS have the potential to reduce the amount and the severity of human-induced accidents. In addition, they can also increase passenger comfort while simultaneously reducing emissions.

Automotive safety functions like automatic emergency braking (AEB), that intervene in dangerous situations which the driver is not able to control, form an important class of ADAS. The perception of the environment surrounding a car is crucial for detecting dangerous situations in which those functions have to react appropriately.

Here, one of our main interests lies in how unavoidable measurement errors made by sensors affect the performance of the functions using the measurements of those sensors in order to interpret the driving situation. In particular, we develop methods and algorithms for designing sensors and functions such that their performance meets the desired specifications in a robust manner despite these measurement errors.

Another aspect of our research deals with how to fuse the information of multiple sensors to provide the necessary information for the aforementioned safety functions. More specifically, we extend classical static grid-based approaches for dynamic environments by estimating velocities.

As more and more autonomous driving functions depend on machine learning, validating these functions becomes increasingly important. Therefore, another interest of our research lies in developing methods to validate machine learning algorithms while implementing state-of-the-art deep learning techniques.