In order to deploy the obvious potentials of DNN-based systems, it is necessary to develop practicable solutions for industrial topics that are not in the focus of common research in the field. For most automotive applications the needed training data imply very high measurement and annotation efforts. These efforts become even more critical when the system is sensitive towards different application environments. Networks trained in a single environment take non-relevant characteristics of the specific environmental conditions in an uncontrolled way into account and therefore data must be recorded repetitively for different environments. Consequently, the available means to reduce required training efforts are limited. Notwithstanding all these efforts, the operational reliability of DNN is hardly provable and hence the regulatory safety requirements for automotive applications cannot be fulfilled.This project will investigate and develop methods for invariant-salient information separation which will improve the robustness and invariance of DNNs to changes irrelevant to the application problem. The efficiency of the resulting background invariant DNN will be tested on an IR-based camera system in the vehicle interior to classify passengers independent of the environment. In order to pave the way to safety critical applications, the aforementioned methods have to be complemented by mathematical concepts to quantify the statistical reliability of such a DNN. A measure of confidence will be defined to detect whether the input is in some sense far from what has been seen during training and can thus not be classified with sufficiently high confidence. Finally, as an immediate consequence of the previous points, the goal is to reduce the overall data acquisition effort and improve the theoretical foundation of DNN-based systems for industrial applications.