Monday, February 26, 2024

Deep Learning-driven 3D Reconstruction of Road Users

Topic and Goal of the Thesis

Large sets of driving data form the basis for the development and validation of automated driving functions. For example, 3D laser scanners (lidars) are used to collect such driving data sets, whose raw data are evaluated in further process steps. In order to enable highly accurate object recognition in the point clouds recorded by the lidar sensors, a large amount of training data is required to teach the processing algorithms.

One way of generating large amounts of annotated data in a short time is simulation. With simulation, any number of different scenarios can be realized, which is only possible with enormous effort for real data. One disadvantage of simulation, however, is that edge cases can hardly be represented due to the limited variety in the underlying data. A complementary approach to expanding the data variety is the incorporation of real-world data, especially from road users. Lidar sensors play a central role here, as they capture the environment with high precision in the form of point clouds.

The topic of this thesis is the conversion of these point clouds into 3D meshes, which also enable the realistic modeling of unconventional road users. In the thesis, a suitable deep learning architecture shall be explored, implemented and evaluated. Numerous software components and frameworks are already available for this purpose.

Working Points

• Literature research on neural networks for 3D reconstruction from point clouds
• Development and implementation of the proposed network architecture
• Training of neural networks on publicly available data sets
• Evaluation of the proposed methods

Requirements

• Good English or German language skills
• Experience with Python
• Experience with Machine Learning is an advantage (not a must)
• Reliability, commitment and enjoyment of working independently

Note: Please attach brief resume and grade summary.