Camera-based Anomaly Detection with Generative World Models

Although huge improvements in the field of autonomous driving have been made in recent years,
dealing with unexpected situations remains a challenging task. Anomaly detection techniques aim
to detect those unknown cases. While generative world models have shown promising results
regarding the perception of the environment and predicting future driving conditions, they are
rarely utilized in anomaly detection for autonomous driving. This thesis presents a novel anomaly
detection method which leverages the advantages of world models and uses feature extraction,
reconstructed observations, and predictions of future observations in order to detect corner cases
in automated driving. The proposed anomaly detection model works fully unsupervised and does
not require anomalies in training data.