[Robotics] Master’s thesis: Machine learning based radar point cloud filtering for odometry and mapping

15.01.2025, Abschlussarbeiten, Bachelor- und Masterarbeiten

New 4D imaging radar sensors are emerging as a promising sensing modality. With high-resolution 3D geometry information and velocity measurements, they approach lidar-level performance. In addition, 4D radars provide additional data, such as variances or radar cross-section estimates.
In this project, the goal is to minimize the manual tuning efforts necessary to implement radar odometry on different radar sensors while maximizing the use of all information available for each radar point. To incorporate all information, a Deep Neural Network (DNN) should be designed to compute weights and descriptors for each radar point. Minimal tuning effort should be achieved by embedding the radar filter network into an unsupervised learning pipeline, similar to [1] and [2].

Possible work packages are:
- Literature review of network architectures for radar data
- Implementation of different network architectures
- Adjustment of existing training framework to the implemented network
- Training data generation using available robots
- Evaluation on different downstream tasks

Prerequisites:
- Programming experience in Python and C++
- Experience with ROS, Pytorch/Tensorflow, and Docker preferable
- Knowledge of robotics
- Independent and self-sufficient way of working

The thesis is supervised by the chair “Perception for Intelligent Systems” (Prof. Lilienthal). We are a new, growing chair with a flat hierarchy and many national and international partners. You can find more information about our research at https://www.ce.cit.tum.de/pins/startseite/

References:
[1]: Burnett, Keenan, et al. "Radar odometry combining probabilistic estimation and unsupervised feature learning." arXiv preprint arXiv:2105.14152 (2021).
[2]: Zhu, Simin, Alexander Yarovoy, and Francesco Fioranelli. "Deepego: Deep instantaneous ego-motion estimation using automotive radar." IEEE Transactions on Radar Systems 1 (2023): 166-180

Kontakt: maximilian.hilger@tum.de