/SC-LIO-SAM

LiDAR-inertial SLAM: Scan Context + LIO-SAM

Primary LanguageC++

SC-LIO-SAM

version 2021-06-24

What is SC-LIO-SAM?

Scan Context: A fast and robust place recognition

  • Light-weight: a single header and cpp file named "Scancontext.h" and "Scancontext.cpp"
    • Our module has KDtree and we used nanoflann. nanoflann is an also single-header-program and that file is in our directory.
  • Easy to use: A user just remembers and uses only two API functions; makeAndSaveScancontextAndKeys and detectLoopClosureID.
  • Fast: A single loop detection requires under 30 ms (for 20 x 60 size, 3 candidates)

Examples

We provide example results using MulRan dataset, which provides LiDAR and 9dof IMU data. You can see the parameter file (i.e., params_mulran.yaml) modified for MulRan dataset.

example 1: KAIST02 of MulRan dataset

example 2: Riverside03 of MulRan dataset

  • As seen in the above video, the combination of Scan Context loop detector and LIO-SAM's odometry is robust to highly dynamic and less structured environments (e.g., a wide road on a bridge with many moving objects).

How to use?

  • We provide a tutorial that runs SC-LIO-SAM on MulRan dataset, you can reproduce the above results by following these steps.
  1. You can download the dataset at the MulRan dataset website
  2. Place the directory SC-LIO-SAM under user catkin work space
    For example, 
    cd ~/catkin_ws/src
git clone https://github.com/gisbi-kim/SC-LIO-SAM.git
cd ..
catkin_make
source devel/setup.bash
roslaunch lio_sam run.launch # or roslaunch lio_sam run_mulran.launch
  3. By following this guideline, you can easily publish the MulRan dataset's LiDAR and IMU topics via ROS.

Dependency

  • All dependencies are same as the original LIO-SAM

Notes

About performance

  • We used two types of loop detetions (i.e., radius search (RS)-based as already implemented in the original LIO-SAM and Scan context (SC)-based global revisit detection). See mapOptmization.cpp for details. performSCLoopClosure is good for correcting large drifts and performRSLoopClosure is good for fine-stitching.
  • To prevent the wrong map correction, we used Cauchy (but DCS can be used) kernel for loop factor. See mapOptmization.cpp for details. We found that Cauchy is emprically enough.

Minor

  • We used C++14 to use std::make_unique in Scancontext.cpp but you can use C++11 with slightly modifying only that part.
  • We used a larger value for velocity upper bound (see failureDetection in imuPreintegration.cpp) for fast motions of a MulRan dataset's car platform.
  • The some code lines are adapted for Ouster LiDAR. Thus, if you use an other LiDAR, please refer the original author's guideline and fix some lines.
  • A LiDAR scan of MulRan dataset has no ring information, thus we simply made a hardcoding like int rowIdn = (i % 64) + 1 in imageProjection.cpp to make a ring index information that LIO-SAM requires, and it works. However, if you use an other LiDAR, you need to change this line.

Applications

  • With our save utility accompanied with this repository, we can save a set of keyframe's time, estimated pose, a corresponding point cloud, and Scan Context descriptors. The estimated poses are saved as a file named optimized_poses.txt and its format is equivalent to the famous KITTI odometry dataset's pose.txt file. For example:

  • If you use the above saved files, you can feed these data to Removert and can removing dynamic objects. No GT labels or external sensor data such as RTK-GPS is required. This tutorial guides steps from running SC-LIO-SAM to save data to Removert to remove dynamic objects in a scan. Example results are:

  • For the safe and light-weight map saver, we support off-line scan merging utils for the global map construction within user's ROI (see tools/python/makeMergedMap.py, for the details, see the tutorial video)

Cite SC-LIO-SAM

@INPROCEEDINGS { gkim-2018-iros,
  author = {Kim, Giseop and Kim, Ayoung},
  title = { Scan Context: Egocentric Spatial Descriptor for Place Recognition within {3D} Point Cloud Map },
  booktitle = { Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems },
  year = { 2018 },
  month = { Oct. },
  address = { Madrid }
}

and

@inproceedings{liosam2020shan,
  title={LIO-SAM: Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping},
  author={Shan, Tixiao and Englot, Brendan and Meyers, Drew and Wang, Wei and Ratti, Carlo and Rus Daniela},
  booktitle={IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages={5135-5142},
  year={2020},
  organization={IEEE}
}

Contact

  • Maintainer: Giseop Kim (paulgkim@kaist.ac.kr)

Contributors

  • Minwoo Jung: made the original LIO-SAM runs on the MulRan dataset.

Acknowledgement

  • SC-LIO-SAM is based on LIO-SAM (Tixiao Shan et al., IROS 2020). We thank Tixiao Shan and the LIO-SAM authors for providing a great base framework.

Update history

  • 2021.06.23
    • yaml file is reformatted to support the compatible form with the recent original LIO-SAM repository.
    • offline ROI global map construction python util is supported.

TODO

  • About performance
    • improve better RS loop closing (ICP test sometimes fails in reverse directions)
    • support reverse-closing of SC loops with Scan Context initial yaw
    • support SC augmentation
    • lagged RS loop closing after large drifts solved
  • About funtions for convenience
    • save extended data: nodes' time, 6D pose, node's point cloud, and corresponding SC descriptors
    • make a static map and self-labeled dynamic points by combining SC-LIO-SAM and removert.
  • Minor (not related to the performance)
    • fix the visualization issue: disappearing map points after SC loop closing
    • fix safe-save for map points after closing the program