LocNDF: Neural Distance Field Mapping for Robot Localization

The system has been tested with Ubuntu-20.04 and Python3.9. Install the following dependencies if needed:

apt-get install libgl1-mesa-glx libx11-dev ldconfig

For the installation of the python package simply clone the repo and pip install it.

git clone [email protected]:PRBonn/LocNDF.git
cd LocNDF
pip install .

Optionally one can use pip install -e . for the editable mode if you plan to change the src code.

The following commands are expected to be executed in this root directory.

First, get the data: Download the Apollo Southbay dataset and place it in data/ (or create a symlink). The ColumbiaPark/ set is enough for the examples.

For training a single model you can configure config/config.yaml and run scripts_pose_tracking/train.py -c config/config.yaml.

Registering a scan to the trained model can be done using python3 scripts_pose_tracking/register_scan.py -c experiments/refactor/lightning_logs/version_7/checkpoints/best.ckpt while only visualizing the meshed result one can use python3 scripts_pose_tracking/mesh_it.py -c experiments/refactor/lightning_logs/version_7/checkpoints/best.ckpt .

For the training of multiple key-poses you can use the config/config_mapping.yaml file and run python3 scripts_pose_tracking/train_mapping.py -c config/config_mapping.yaml .

Tracking the car pose in the trained submaps can be done using python3 scripts_pose_tracking/pose_tracking.py experiments/PATH-TO-THE-CHECKPOINTS/best-v*.ckpt -vis.

Pretrained models can be downloaded here and should be placed under /experiments. Those models can be used as explained above.

Most importantly implement your own dataloader. An example can be seen in src/loc_ndf/datasets/datasets.py. Second, exchange the dataloader in the training script by you dataloader. Ready to train.

The data can be downloaded here. The training data consists of poses.txt and scans/*.npy. The evaluation is done on seq1 to seq5 using the provided scans as well as the odometry.txt. The extrected files are expected to be in in data/.

For training a model you can configure scripts_mcl/config.yaml and run scripts_mcl/train.py.

After training a model, one can run the MCL example, e.g. (scripts_mcl/run_mcl.py -c PATH-TO-YOUR_CKPT -i data/indoor_scan_poses_2d/seqX -cal data/indoor_scan_poses_2d/base2laser.txt -o out_poses.txt) with the trained model.

The Pretrained models can be downloaded here and should be placed under /experiments.

If you use this library for any academic work, please cite the original paper.

@article{wiesmann2023ral,
author = {L. Wiesmann and T. Guadagnino and I. Vizzo and N. Zimmerman and Y. Pan and H. Kuang and J. Behley and C. Stachniss},
title = {{LocNDF: Neural Distance Field Mapping for Robot Localization}},
journal = ral,
volume = {8},
number = {8},
pages = {4999--5006},
year = 2023,
issn = {2377-3766},
doi = {10.1109/LRA.2023.3291274},
codeurl = {https://github.com/PRBonn/LocNDF}
}