When it comes to a failsafe operation of AD (automated driving) the robustness and reliability of its components are critical. In current AD systems traditional computer vision is combined, or even replaced by deep learning (DL) forecasters.

One major drawback of DL is the black-box nature, which results in a lack of interpretability and explainability for both passengers and developers. It is not obvious why decisions are made and how the inner workings are influenced by, for example, the operating domain.

A second characteristic arises from the probabilistic nature of DL, which aggregates knowledge about the underlying data distribution from samples. Therefore, specification of the domain during development (training) is mandatory for later operation (inference). This attribute results in huge efforts during testing and validation to ensure the necessary robustness against domain changes, consistency in data for following processing steps, and a failsafe operation in total.

The proven way to quantify the confidence in estimates of e.g. sensor systems is to determine their uncertainty. This concept is also used in current approaches to quantify the (predictive) uncertainty in the estimates of a DL forecaster. The predictive uncertainty can be further decomposed into the aleatoric component, which is induced by the randomness itself, and the epistemic uncertainty. The epistemic uncertainty is due to the design of the network and e. g. the limited amount of training data available.

In the research work, two perspectives will be taken:

1. Efficient testing/validation to identify weaknesses.

2. Improve system robustness and reliability by self-awareness.

In both cases, methods of the field of uncertainty quantification will be applied.