ISBN 978-3-8439-4647-6

978-3-8439-4647-6, Reihe Ingenieurwissenschaften

Martin Pöllot
Contributions to Visual Automotive Human Machine Interface Testing

215 Seiten, Dissertation Universität Erlangen-Nürnberg (2020), Softcover, A5

Zusammenfassung / Abstract

The amount of driver assistance, entertainment, and information features in automobiles is increasing rapidly over the past few years. Human machine interfaces are the means to communicate between car and passengers. With the regard of automated testing of these interfaces, a typical scenario in the development of automotive infotainment systems is revealed. Conventional approaches for testing, based on empirical thresholds, however, are not able to provide reliable detection rates and require a lot of maintenance. This is where model-based testing approaches come into play. By applying model-based approaches, the acquisition of empirical thresholds is abandoned in favor of models learned by algorithms. Relying on an underlying model, the algorithms exhibit higher capabilities in detecting errors that approaches based on empirical thresholds are not taking into account.

This thesis concentrates on automated testing in the context of a model-based end-to-end approach. After introducing the main approaches and components for evaluating and setting up an automated testing environment, a first component for interacting with the infotainment system is provided, which could bring human-like input patterns forward. The main part of this thesis focuses on the detection of irregular motion in the navigation context. By applying a model-based approach that learns the normal behavior of the system, irregularities could be found with relative ease in comparison to stale and fixed threshold-based approaches. Furthermore, in the domain of screen content verification, a deep neural network approach is dedicated to detect and classify icons in a given screen. This approach could speed up the state-of-the-art sliding window pattern matching approach and also be more robust to design changes and noise. All contributions provide extensive experiments on custom data sets and demonstrate the advantage over conventional methods.