Anomaly Detection for Autonomous Driving

With small fleets of autonomous vehicles of SAE level 4, i.e., such without a safety driver, publicly available, the adoption of autonomous vehicles will only continue to increase. Embedded within shared mobility solutions, this technical advancement can lead to a more sustainable, safe, and comfortable future. Scaling autonomous vehicles more broadly, however, requires handling a wide variety of challenging scenarios, especially those with often rare anomalies. With rising fleet sizes, such scenarios appear with increasing frequency. As many Machine Learning systems follow a closed-world assumption based on a set of known classes, such unknowns remain challenging. This dissertation addresses anomaly detection for autonomous driving from a holistic perspective, contributing to the generation of scenarios with anomalies, the detection of anomalies, and the handling of anomalies. The first part addresses external anomalies, i.e., such that occur in the environment. Generating scenarios involves providing normal data to train models and creating scenarios with anomalies to evaluate anomaly detection methods. Based on a theoretical systematization of anomalies from the literature, scenarios from all anomaly layers can be created. As generating such external anomalies is often dangerous or infeasible, data is provided through a simulation engine. Based on these scenarios, an anomaly detection method is presented, which is trained on unlabeled sensor data alone. It leverages a world model as a representation of normality, utilizing both camera and LIDAR data. Once detected, anomalies can be integrated into the training process of Neural Networks, removing their status as anomalies. The presented approach handles previously detected anomalies where controlled traffic rule exceptions are required. To achieve this, a situation-aware reward for Reinforcement Learning is introduced. Next to challenges induced by external anomalies, the driving task can be equally impacted by internal anomalies, such as model failures. This dissertation contributes to the field of internal anomaly detection by detecting model failures without the need for labeled evaluation sets. This is achieved by analyzing the disagreements between two models trained on the same task, but with different learning paradigms. Based on real-world data, the method successfully reveals categorical model failures, most often in seemingly normal situations. Summarizing, this dissertation presents a holistic set of contributions to the field of anomaly detection for autonomous driving, addressing the generation, detection, and handling of anomalies. This is emphasized by examining both internal and external anomalies.