Anomaly Detection with Model Contradictions for Autonomous Driving

Anomaly detection is a critical aspect of safe autonomous driving systems, where detecting and understanding uncommon and unpredictable scenarios, often referred to as corner cases or anomalies, is crucial for ensuring the safety of passengers and pedestrians. In this bachelor's thesis, I quantitatively evaluate an anomaly detection method proposed by Sartoris that utilizes lidar data for detecting anomalies. The method combines a supervised and a self-supervised part to detect motion anomalies in the environment. By analyzing the discrepancies between the two parts, the detection method identifies points that deviate from the expected behavior, indicating potential anomalies.

This evaluation utilizes the CODA dataset, which provides the only anomaly dataset including lidar data. However, the corner cases are only labeled in the form of 2D bounding boxes. To address this limitation, I convert the 2D bounding boxes in the CODA dataset into 3D point-wise labels. The CODA dataset is then translated into the KITTI-odometry data format suitable for evaluating Sartoris' method. Additionally, improvements are proposed for the clustering algorithms used to create 3D point-wise labels, aiming to reduce the need for manual verification.
... mehr

Given the lack of suitable metrics for semantic segmentation, except for IoU, I propose two novel approaches that utilize the metrics AP, AR, and F1 to quantitatively evaluate Sartoris' anomaly detection approach on the CODA dataset. The results demonstrate the potential of this method in detecting anomalies compared to standard object detection techniques carried out by Li et al., despite the challenge of comparing my metrics with theirs.

In the outlook of this thesis, the potential extensions and improvements to the detection approach are discussed, such as fine-tuning the models for the original datasets and addressing challenging scenarios like fast turns or speed bumps. Furthermore, the need for appropriate metrics