LocNDF

LocNDF: Neural Distance Field Mapping for Robot Localization

Installation

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 git@github.com: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.

Usage

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

Pose tracking in Apollo Southbay

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.

Training one Submap

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 .

Training multiple key-poses

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.

Training on your own data

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.

2D - MCL

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/.

Training

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.

Citation

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}
}

lhthq/LocNDF