PUB - Publikationen an der Universität Bielefeld

The quest for fully autonomous vehicles has made significant strides in recent years, with partial automation becoming increasingly common in modern vehicles. However, achieving widespread public acceptance of autonomous systems hinges on ensuring their safety and reliability. Central to the functionality of autonomous vehicles is their perception of the surrounding environment, which informs decision-making processes in real-time. Current perception systems in the automotive context require objects to be directly visible to be detected. However, humans already use visual cues such as shadows during the day or light reflections during the night to draw assumptions about the presence of objects before they are directly visible. The effect is called anticipatory vehicle detection and is crucial to safe and comfortable driving. This thesis investigates for the first time the anticipatory detection of oncoming vehicles at night in urban environments, aiming to bridge the gap between human and computer perception. Through a series of experiments and analyses, this work seeks to understand the feasibility of anticipating oncoming vehicles before they are directly visible, leveraging visual cues such as light reflections. The thesis comprises three different main contributions in this realm: - A better understanding of human anticipation capabilities in urban nighttime driving scenarios, which is addressed via a sequence of user studies and their evaluation, - a novel dataset to address the task of detecting oncoming vehicles at night in urban scenarios based on light reflections, with a particular focus on the challenge of achieving high-quality annotations in the light of limited human effort, and - a proposal of a pipeline to address anticipatory vehicle perception in cities at night and an extensive evaluation of the capabilities of deep learning foundation models for this task, with a focus on the implementation and evaluation of image and video-based deep learning architectures in this context. In more detail, the research begins with an examination of human anticipation abilities in urban nighttime driving scenarios. Results indicate that while humans can indeed anticipate oncoming vehicles before direct visibility, the potential for anticipation is slightly diminished compared to rural environments. Light reflections emerge as a key visual cue used by humans to anticipate oncoming vehicles in urban settings. Subsequently, a dataset specifically tailored for detecting light reflections caused by oncoming vehicles in urban environments is developed. Annotation strategies are explored, and the dataset is made publicly available to facilitate further research in the field. Building upon insights gained from the dataset creation, state of the art semantic segmentation approaches are evaluated for their efficacy in detecting light reflections in urban environments. While promising, the results underscore the need for further improvement, particularly in reducing false positive detections. Overall, this thesis contributes to advancing the field of autonomous driving by addressing the critical challenge of anticipatory vehicle detection at night in urban scenarios, with implications for safety, comfort, and user acceptance.