Towards Runtime Detection of Novel Traffic Situations

Kurzfassung

Automated Vehicles developers need to define an Operational Design Domain (ODD) where such vehicles can operate safely. In order to extend the defined ODDs, the developers base their decision after detailed analysis of recorded data from multiple data collection drives. For the acquired data, it is important to know whether it is known traffic situation information (inside the automated vehicle's ODD) or novel information that can be used to expand the ODD. The large amount of data that is generated by a modern vehicle's sensors makes data storage and efficient analysis for expanding ODDs hardly feasible (most of the current approaches record all sensor data and then post-process the data using AI-based methods and finally perform manual checks in order to find the novel data). Hence, there is a need to classify traffic situations as novel at system runtime for an appropriately abstract notion of novelty so that the conceptually same traffic situation, e.g. on two similar days, is not considered novel only because of the different date. We propose a new methodology for detection of novel traffic situations at system runtime. The methodology is based on a traffic catalogue that consists of abstract traffic situation descriptions, which are a formalized representation of sets of concrete traffic situations. Continuous, automatic checks for satisfaction of the current traffic situation against the traffic catalogue provides verdicts about the novelty of the current traffic situation. Using an illustrative example, we show how domain experts can utilize the detected novelties to create such a traffic catalogue such that the novelties are classified as known in the future. The proposed method doesn't require any pre-training of an AI-based classifier and is human understandable, explainable and traceable.