MCS (Map Clustering Similarity)

MCS is a LiDAR-based place recognition framework merging both scan-to-scan and scan-to-map comparisons. It is accepted by IEEE TIV. A New Horizon: Employing Map Clustering Similarity for LiDAR-based Place Recognition

Differences: MCS vs. Current Methods

• Fresh perspective: Despite employing the prior map, we simplify it into multiple cluster centers for constructing a cluster descriptor, thus avoiding intricate scan-to-map matching.
• Lightweight design: Our LiDAR descriptor (OcSC) solely judges whether a region is occupied rather than computing geometric details like height, normals, and density.
• 2-phase search without KD tree: cluster descriptor + Spearman loss, OcSC LiDAR descriptor + Occupancy loss. We can achieve good efficiency without constructing a KD tree.

Dependencies

CMake
PCL
OpenCV
Boost
Eigen

File and Data Structure

Data structure

data
    |---KITTI
    |       |---00
    |            |---gt
    |                 |---gt_poses.txt
    |                 |---loop_nn.txt
    |            |---map
    |                 |---cnts_4_20_60_3.bin
    |                 |---cnts_4_20_60_3.txt 
    |                 |---map_ring_keys_4_20_60_3.txt
    |            |---pcd
    |                 |---0.pcd
    |                 |---1.pcd
    |                 |---2.pcd
    |                 |---...
    |       |---01
    |       |---02
    |       |---...
    |---...

Files

• loop_nn.txt: a loop flag file based on a 8m threshold. Take a line as an example: (835 0 1 2 3 4 5 6 7) means the 835th scan froms a loop closure with 0th-7th scan.
• gt_poses.txt: each row stores a 4*4 pose
• cnts_4_20_60_3: map cluster centers file
• map_ring_keys_4_20_60_3: map ring keys

How to use?

MCS pipeline comprises two parts: (1) offline map processing, (2) online place recognition.

1. Build the project

mkdir build && cd build
cmake .. 
make -j16

Note: In KITTI Odometry, we've generated necessary offline map files, allowing the bypass of offline map processing for direct online place recognition. When testing in other datasets, offline map processing is required to generate the necessary map files before place recognition. Move the point cloud to the "pcd" directory prior to execution.

2. Offline map processing

2.1 Create map point cloud and virtual points

./offline_map_process/create_virtual_pts_and_map

you will get a "map_sam_pro.pcd" (map) and a "ground_sam_dist_3.pcd" (virtual points) in map folder .

2.2 Create map cluster centers (4 Kb)

./offline_map_process/create_map_cluster_centers

or run the matlab script (we use this in our experiments)

k_means.m

you will get a "cnts_4_20_60_3.bin" (map cluster centers) in map folder. After this, all necessary map files are generated.

3. Online place recognition

We provide the evaluation code (recall top 1) for MCS and MCS-BF. MCS (two-phase search) balances the runtime and accuracy, while MCS-BF (one-phase search) may showcase better accuracy.

3.1 Run MCS

./online_place_recognition/mcs_recall_1

3.2 Run MCS-BF

./online_place_recognition/mcs_bf_recall_1

Citation

If you think this work is useful for your research, please consider citing:

@ARTICLE{10416737,
  author={Shi, Pengcheng and Xiao, Yilin and Chen, Wenqing and Li, Jiayuan and Zhang, Yongjun},
  journal={IEEE Transactions on Intelligent Vehicles}, 
  title={A New Horizon: Employing Map Clustering Similarity for LiDAR-based Place Recognition}, 
  year={2024},
  volume={},
  number={},
  pages={1-11},
  doi={10.1109/TIV.2024.3360321}}

Acknowledgement

Thanks to the source code of some great works such as linefit

Contact

For any questions or bug reports, please create an issue or contact me at [email protected].