Benchmarking Anomaly Detection on Camera and Lidar Data with 3D Voxel Representation

The research field of autonomous driving has seen rapid development in recent years. There are,
however, challenges that hinder the deployment of autonomous vehicles on the road. One of these
challenges is the detection of unknown or anomalous instances on the road. The field of Anomaly
Detection is crucial for the safe deployment of these systems, as detection failure could lead to the
execution of potentially dangerous behavior.
Most autonomous vehicles employ an array of different sensors for scene understanding. To
effectively utilize data extracted from multiple sensors, it is important to fuse all sensor data into
a common state representation.
This thesis explores anomaly detection in combination with sensor fusion representations by
evaluating anomaly detection methods for camera and LiDAR on voxel grids. The current stateof-
the-art of anomaly detection for camera and LiDAR is reviewed to identify current trends and
research gaps in the field. From the literature review, a camera-based method and a LiDAR-based
method were selected for evaluation on the FZI AnoVox benchmark, an anomaly detection dataset
that includes ground truth information on 3D surroundings in the form of a voxel grid. ... mehrAnomaly
score predictions were mapped into the voxel space to evaluate the detection performance. The
findings show that metric results on images and point clouds can significantly deviate when transformed
into voxels. Furthermore, the thesis explores effects and result changes when adjusting
parameters in the voxelization process.