Informed, Constrained, Aligned: A Field Analysis on Degeneracy-aware Point Cloud Registration in the Wild

Turcan Tuna†, Julian Nubert†§, Patrick Pfreundschuh*, Cesar Cadena†, Shehryar Khattak‡, Marco Hutter†

†Robotics Systems Lab, ETH Zürich.

*Autonomous Systems Lab, ETH Zürich.

‡NASA Jet Propulsion Laboratory, California Institute of Technology, USA

§ Max Planck Institute for Intelligent Systems, Stuttgart, Germany.

Pre-Print Version

Pre-Print PDF

Supplementary

Video

LiDAR Degenerate Datasets

Datasets

Github

Code

This work deploys and compares multiple degeneracy mitigation methods with real-life field experiments in the loop with a LiDAR SLAM solution.

Analyzed Methods

An overview of all the investigated methods, highlighting at which stage these methods affect the point cloud registration process. The letters P and A mark passive and active degeneracy mitigation methods, respectively. The star symbol indicates that the method is investigated in the context of degenerate point cloud registration for the first time in this work.

Results

As a recommendation, if computational resources and fine-tuning are available, NL-Reg. method is a good option. If deployed on a computationally limited system, with fine-tuning availability Ineq. Con. method has great potential since this method (thanks to the formulation of the QP problem) can also accept equality constraints. On the other hand, if fine-tuning is not an option, Eq. Con. method performed reliably throughout the experiments with good accuracy. However, if the freedom of optimization is of utmost importance, the TSVD method showed the same reliability in exchange for accuracy but without adding hard-constraints to the optimization. As a sampling approach, RMS by Petracek et al. provided a speed-up to the framework throughout all the experiments and performed the best in the urban setting. Although this method did not perform well in other experiments, if the environment is feature-rich, it could be a strong candidate for many frameworks to mitigate the effect of mild degeneracy.

The results and the recommendations done in this work are to provide an objective view of the problem of degenerate point cloud registration. However, the reader should be aware of the differences between the systems used in this work, which use a voxel-based mapping backend, external pose prior, and scan-to-submap registration.

Citation

Tuna, T., Nubert, J., Pfreundschuh, P., Cadena, C., Khattak, S., & Hutter, M. (2024). Informed, Constrained, Aligned: A Field Analysis on Degeneracy-aware Point Cloud Registration in the Wild

Corresponding Author Email: tutuna@ethz.ch

Page updated

Report abuse