Herr Igor Anpilogov hat seine Bachelorarbeit an der TU Clausthal erfolgreich abgeschlossen. Das Thema seiner Arbeit lautet: “ Selection and Transition among different Behavioural Patterns of End-To-End Based Driving Functions“.
Abstract:
The field of autonomous vehicles is driven by complex End-To-End systems, mainly presented as deep neural networks. The complexity and the computational power behind these systems proved to be effective at driving the field forward, but the same features hinder the understanding and proving the behavioral attributes of deep neural networks, which can be a big concern for safety and effectiveness of the system. In this paper I propose the Behavior Selector - a driving system architecture using selection and transition between smaller neural networks to achieve performance close to much bigger End- To-End systems. The architecture based on smaller, highly specialised networks makes it
easier to prove the behavioral features of the networks separately, as well as the driving system as a whole.
