Topological Computer Vision for LiDAR Point Cloud Processing in Automotive Industry (engl.)

Point clouds captured by LiDAR sensors from objects in the scenes become an important source of highly resolved environmental information for the automotive industry. The highly accurate depth information of the LiDAR point clouds made them an inevitable perception component for highly automated vehicles. However, the processing of these point clouds is challenged by their irregular and sparse distribution and the presence of various artifactual effects. Topological computer vision spans through the fascinating fields of graph-topological detection and pattern recognition, geometry-aware architectures, spectral graph theory, and manifold learning. This altogether allows the topological computer vision to tackle irregularity, sparsity, and artifactual effects of LiDAR point clouds in challenging perception and detection tasks. Besides, the manifold-based processing of the point clouds through geometrical and topological models not only fosters an efficient and discriminative data representation but also helps to reduce the curse of dimensionality if the manifold gets defined in the feature space of the perception/detection task.

Requirements: Familiarity with basics of topology, geometry, and neural networks is helpful for the understanding of the martials but is not necessarily needed.
Credit Points (ECTS): —