Autonomous driving decisions depend on the behavior of surrounding agents, which is inferred from their observed past motion and the predicted future trajectories. Modern decision making algorithms increasingly rely on machine learning, which, however, may fail under uncertain conditions caused by anomalies, distribution shifts, or inherent forecasting errors. Consequently, early detection of abnormal agent behavior and the quantification of prediction uncertainty are essential to enable safe driving. This thesis presents methods and datasets to identify and quantify uncertainty in agent behavior, addressing both the observed motion and the predicted trajectories. The first contribution proposes a spatio-temporal graph auto-encoder (STGAE) combined with kernel density estimation (KDE) for anomaly detection in the past motion, focusing on multi-agent highway scenarios. The approach encodes agent interactions into a latent space and models normal behavior with a non-parametric distribution. Abnormal maneuvers are detected when their latent representations fall into low-density regions. A simulated highway dataset with diverse agent interactions is introduced, demonstrating that the method outperforms linear baselines, reconstruction-based models, and one-class support vector machine (OC-SVM), achieving an area under the receiver operating characteristic curve (AUROC) of 86.3 % in real time. The second contribution extends anomaly detection to complex urban environments by introducing R-U-MAAD, a benchmark for unsupervised anomaly detection in realistic urban driving settings. The dataset that is part of the benchmark integrates real-world multi-agent trajectories with simulated abnormal behaviors across diverse map topologies and provides timestamp-wise annotations covering 13 anomaly classes. A standardized training and evaluation protocol is introduced, alongside eleven baseline models. Results show that deep reconstruction-based methods perform best overall, while interaction-aware models outperform those without interaction modeling. The third contribution introduces a lightweight extension to encoder–decoder trajectory prediction models for joint out-of-distribution (OOD) detection and uncertainty quantification. The approach augments the latent space of the prediction model with a Gaussian mixture model (GMM) for OOD detection and an error regression network for supervised uncertainty estimation. Evaluated on the Shifts dataset, the method outperforms deep ensembles and Monte Carlo dropout in both tasks, while also achieving state-of-the-art trajectory prediction accuracy. Notably, the additional modules increase runtime by only 5 % over the baseline predictor, supporting real-time deployment. Together, these contributions provide building blocks for uncertainty-aware safety monitoring in autonomous driving. Their efficiency and reliance on the structured environment representations, as provided by perception modules, facilitate integration into modular driving stacks. Future work should evaluate closed-loop interactions with planning the integration into end-to-end prediction frameworks.
Julian Wiederer (Thu,) studied this question.