Poisson Surface Reconstruction for LiDAR Odometry and Mapping

Surfels	TSDF	Our Approach

Table: Qualitative comparison between the different mapping techniques for sequence 00 of the KITTI odometry benchmark.

This repository implements the algorithms described in our paper Poisson Surface Reconstruction for LiDAR Odometry and Mapping.

This is a LiDAR Odometry and Mapping pipeline that uses the Poisson Surface Reconstruction algorithm to build the map as a triangular mesh.

We propose a novel frame-to-mesh registration algorithm where we compute the poses of the vehicle by estimating the 6 degrees of freedom of the LiDAR. To achieve this, we project each scan to the triangular mesh by computing the ray-to-triangle intersections between each point in the input scan and the map mesh. We accelerate this ray-casting technique using a python wrapper of the Intel® Embree library.

The main application of our research is intended for autonomous driving vehicles.

Table of Contents

• Running the code
  • 0) Requirements
  • 1) Datasets
  • 2) Converting from .bin to .ply
  • 3) Runing the puma pipeline
  • 4) Inspecting the results
• Where to go next
• Citation

Running the code

NOTE: All the commands assume you are working on this shared workspace, therefore, first cd apps/ before running anything.

0) Requirements

If you plan to use our docker container you only need to install docker and docker-compose.

If you don't want to use docker and install puma locally you might want to visit the Installation Instructions

1) Setting up the environment

First, you need to indicate where are all your datasets, for doing so just:

export DATASETS=<full-path-to-datasets-location>

This env variable is shared between the docker container and your host system(in a read-only fashion).

So far we've only tested our approach on the KITTI Odometry benchmark dataset and the Mai city dataset. Both datasets are using a 64-beam Velodyne like LiDAR.

Then you need to expose your user id and group id to have access to the artifacts of the pipelines:

export UID
export GID

2) Converting from .bin to .ply

All our apps use the PLY which is also binary but has much better support than just raw binary files. Therefore, you will need to convert all your data before running any of the apps available in this repo.

docker-compose run --rm apps bash -c '\
    ./data_conversion/bin2ply.py \
    --dataset $DATASETS/kitti-odometry/dataset/ \
    --out_dir ./data/kitti-odometry/ply/ \
    --sequence 07
    '

Please change the --dataset option to point to where you have the KITTI dataset.

3) Runing the puma pipeline

Go grab a coffee/mate, this will take some time...

docker-compose run --rm apps bash -c '\
    ./pipelines/slam/puma_pipeline.py  \
    --dataset ./data/kitti-odometry/ply \
    --sequence 07 \
    --n_scans 40
    '

4) Inspecting the results

The pipelines/slam/puma_pipeline.py will generate 3 files on your host sytem:

results
├── kitti-odometry_07_depth_10_cropped_p2l_raycasting.ply # <- Generated Model
├── kitti-odometry_07_depth_10_cropped_p2l_raycasting.txt # <- Estimated poses
└── kitti-odometry_07_depth_10_cropped_p2l_raycasting.yml # <- Configuration

You can open the .ply with Open3D, Meshlab, CloudCompare, or the tool you like the most.

Where to go next

If you already installed puma then it's time to look for the standalone apps. These apps are executable command line interfaces (CLI) to interact with the core puma code:

├── data_conversion
│   ├── bin2bag.py
│   ├── kitti2ply.py
│   ├── ply2bin.py
│   └── ros2ply.py
├── pipelines
│   ├── mapping
│   │   ├── build_gt_cloud.py
│   │   ├── build_gt_mesh_incremental.py
│   │   └── build_gt_mesh.py
│   ├── odometry
│   │   ├── icp_frame_2_frame.py
│   │   ├── icp_frame_2_map.py
│   │   └── icp_frame_2_mesh.py
│   └── slam
│       └── puma_pipeline.py
└── run_poisson.py

All the apps should have an usable command line interface, so if you need help you only need to pass the --help flag to the app you wish to use. For example let's see the help message of the data conversion app bin2ply.py used above:

Usage: bin2ply.py [OPTIONS]

  Utility script to convert from the binary form found in the KITTI odometry
  dataset to .ply files. The intensity value for each measurement is encoded
  in the color channel of the output PointCloud.

  If a given sequence it's specified then it assumes you have a clean copy
  of the KITTI odometry benchmark, because it uses pykitti. If you only have
  a folder with just .bin files the script will most likely fail.

  If no sequence is specified then it blindly reads all the *.bin file in
  the specified dataset directory

Options:
  -d, --dataset PATH   Location of the KITTI dataset  [default:
                       /home/ivizzo/data/kitti-odometry/dataset/]

  -o, --out_dir PATH   Where to store the results  [default:
                       /home/ivizzo/data/kitti-odometry/ply/]

  -s, --sequence TEXT  Sequence number
  --use_intensity      Encode the intensity value in the color channel
  --help               Show this message and exit.

Citation

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

@inproceedings{vizzo2021icra,
author    = {I. Vizzo and X. Chen and N. Chebrolu and J. Behley and C. Stachniss},
title     = {{Poisson Surface Reconstruction for LiDAR Odometry and Mapping}},
booktitle = {Proc.~of the IEEE Intl.~Conf.~on Robotics \& Automation (ICRA)},
codeurl   = {https://github.com/PRBonn/puma/},
year      = 2021,
}