Intelligent, highly automated vehicles have the advantage that, above all, the safety of all persons involved in traffic can be significantly increased. For the realization of such vehicles an exact and secured knowledge of the vehicle environment with a resolution in the centimeter range is necessary. An approximation of this can be achieved, for example, by redundant sensors, but also by fusing the vehicle's own information with information from other sources via communication channels such as Vehicle2X. However, as the complexity of such systems increases, so does the challenge of ensuring the functional reliability and operational protection of the overall system. New inherent monitoring techniques are therefore needed to ensure maximum safety while maintaining cost-effectiveness.
Aim of the research project
Therefore, a module-by-module development of an observer structure is proposed, which finds inconsistencies within the chain of effects of automated driving with the help of machine learning methods, but also classical control engineering observer approaches. Inconsistencies refer, for example, to possible manipulative third-party interventions or erroneous perception such as overlooking obstacles. The development initially focuses on the area of perception and is supplemented by subsequent elements of the chain of effects.
The observer is seen as a superordinate instance, which regards the system to be observed as a composition of several black box models - i.e. the inner mathematical models and correlations are not known or assumed to be known. Only the predefined input and output variables and the characteristics derived from them are available for observation. This approach has the effect that the system complexity can be reduced so clearly, since not each possible, in the system available and possibly superfluous information is used. It is also important that the concept is well generalizable and transferable to allow modular development. For this purpose, it is important that the observer can be extended as efficiently as possible by training units and feature inputs.
In general, it is also relevant to be able to guarantee the applicability on a control unit. A high parallelizability of the tasks as well as a mathematically less complex applicability are important for this.