DAnoScenE: a driving anomaly scenario extraction framework for autonomous vehicles in urban streets

Abstract

Autonomous vehicles (AVs) hold great potential to improve traffic safety. However, urban streets present a dynamic environment where unforeseen and complex scenarios can arise. The establishment of a systematic framework to extract a variety of vehicle driving scenarios could empower AVs to learn from and effectively navigate various situations. This study introduces a driving anomaly scenario extraction (DAnoScenE) framework tailored for AVs operating in urban street settings. The Waymo Open Motion Dataset (WOMD) is used to showcase the framework’s capability to capture an extensive range of realistic driving anomaly scenarios. The central process involves the detection and labeling of driving anomalies. To avoid the erroneous detected and labeled driving anomalies arising from issues such as outliers and noise within vehicle track data, a two-step approach is introduced to analyze and rectify vehicle movement parameters in raw data. To comprehend these driving anomalies and their associated scenarios, manual labeling identifies causative factors of scenarios such as traffic signals and behaviors of other agents, forming three scenario groups: Signal Interaction, Agent Interaction, and Other. A multimodal model is developed to classify scenario groups, complemented by a segmentation process that further divides groups into specific scenarios based on simple conditions. The results show that recognition accuracy for driving anomaly scenario groups achieved 98.4%, and the scenario segmentation method achieved 100% accuracy by simple conditions. The proposed framework provides valuable support for the advancement of autonomous driving algorithms and comprehensive AV testing, with a specific emphasis on navigating abnormal driving environments.