Traffic Scenario Detection – A Comparison of CNN-RNN and Transformer-based Architectures in the Context of Autonomous Driving

Authors

• Nicole Grieger, Paderborn University, GermanyFollow
• Hasan Okumus, Paderborn University, GermanyFollow
• Sven Burdorf, dSPACE GmbH, GermanyFollow
• Dennis Kundisch, Paderborn University, GermanyFollow

Abstract

The precise classification of vehicle maneuvers is essential for enhancing the safety and reliability of autonomous vehicles. While deep learning methods have advanced maneuver prediction, most studies focus solely on lane changes of either the ego vehicle or surrounding vehicles, neglecting the combination of both in more diverse traffic scenarios. Moreover, these studies only implement CNN -RNN or closely related approaches. What is missing is a comparison to emerging attention -based approaches. The presen t study fills this gap by classifying more divers e maneuvers, including lane changes and cut -ins of the ego and surrounding vehicles, using a dataset from dSPACE GmbH. We compare a CNN -RNN with a transformer-based architecture to identify the best approach for the underlying task. Our results show that tr ansformers outperform CNN -RNNs in capturing spatio -temporal dependencies inherent in our data. Thus, this study advances research on transformer -based models for autonomous driving and offers a cost -effective alternative to manual scenario classification

Recommended Citation

Grieger, Nicole; Okumus, Hasan; Burdorf, Sven; and Kundisch, Dennis, "Traffic Scenario Detection – A Comparison of CNN-RNN and Transformer-based Architectures in the Context of Autonomous Driving" (2025). MCIS 2025 Proceedings. 31.
https://aisel.aisnet.org/mcis2025/31