PO-SLAM: Planetary target-Oriented Simultaneous Localization and Mapping

intro: This repo proposes a SAM-augmented target-oriented SLAM framework that enables planetary rovers to identify target, estimate the relative position, and reconstruct/represent the target from abstraction to precision. This object SLAM algorithms can work in an unstructured, weakly textured, and lunar terrain environment.

1. Structure

flowchart LR
	id1(PO-SLAM) --> id2([datset]) --> id9[[SePT_Stereo-Planetary-Tracks Dataset]]
	id1(PO-SLAM) --> id3([Examples]) --> id10[[Save .cc/.yaml/timestamp files]]
	id1(PO-SLAM) --> id4([include]) --> id11[[Head files for PO-SLAM]]
	id1(PO-SLAM) --> id5([src]) --> id12[[Source files for PO-SLAM]]
	id1(PO-SLAM) --> id6([protobuf-redis]) --> id13[[Save defined protos and subscribe source files]]
	id1(PO-SLAM) --> id7([results_line_segments]) --> id14[[Save the demi-dense model for each object]]
	id1(PO-SLAM) --> id8([Thirdparty]) --> id15[[Reliable thirdparty files, such as g2o]]

Loading

2. Quick Running

2.1 Prerequisites

• Prerequisites are the same as EAO-SLAM. If compiling problems met, please refer to semidense-lines and ORB_SLAM2.
• The code is tested in Ubuntu 20.04, opencv 3.4.4, Eigen 3.3.7, boost 1.63.0.
• Open source dataset: SePT Dataset

2.2 Building

chmod +x build.sh
./build.sh

2.3 Running demo

bash sept_examples.sh

or

./Path_to_executeFile DataAssoParam ./Path_to_vocabularyFile ./Path_to_yaml Path_to_DataFolder ./Path_to_timestampFile

2.4 Results

The mean matching accuracy figure based on SOTA matching baselines:

Object instances extraction results:

Trajectories's results (camera positioning):

Trajectories's results (relative positioning):

Rocks reconstruction results:

 

2.5 Eval

• CameraTrajectory[EuRoC]: save in evo_output
• ObjectRelativePoses[EuRoC]: save in evo_output
• Line3D++: save in results_line_segments
• Related cmds are shown in: evo_output/evo_cmd.txt

3. Examples

3.1 Input Stereo images

The PO-SLAM inputs stereo images (stereo baseline: 150mm)

 

3.2 Prepare object extraction for data association

The patch2pix is utilized for image matching (left camera), an example result is:

The keypoints work as prompt in SAM, then we got:

 

As shown in left image above, the minimum bounding boxes (i.e., [objectID, x, y, w, h, confidence]) are obtained and saved as txt files, which can be used as offline input for PO-SLAM.

PS: these boxes can generate 1FPS by running protobuf+redis scripts (subscribe_semanticStereo.cc).

3.3 Run PO-SLAM for Lunar01(SePT01)

./build.sh
bash sept_examples.sh

or

./build.sh
./Examples/Stereo/stereo_SePT EAO ./Vocabulary/ORBvoc.bin ./Examples/Stereo/SePT01.yaml ./dataset/SePT/SePT01 ./Examples/Stereo/TimeStamps/SePT01.txt

The running demo is like:

3.4 Results

• CameraTrajectory: save in evo_output
• ObjectRelativePoses: save in evo_output
• Semi-dense reconstruction for Rock1: save in results_line_segments

4. Video

• Local: the demo video is saved in 902
• demo: Youtube

5. Notes

• First time before running: remove all cmake build files in all folders

• Defualt execute file is: stereo_SePT.cc

• Defualt segmentFusion image: left camera

• Defualt protoFile and its execute file are located in: offline_bbox.proto and subscribe_semanticStereo.cc

• To record logs, please use:

  bash sept_examples.sh > log.txt
  

6. Acknowledgement

Thanks to following works: Image-Matching-Toolbox, SAM, and EAO-SLAM.

• (Patch2Pix) Q. Zhou, T. Sattler and L. Leal-Taixé, "Patch2Pix: Epipolar-Guided Pixel-Level Correspondences," CVPR 2021, Nashville, TN, USA, 2021, pp. 4667-4676. Paper.
• (SAM) Kirillov, Alexander and Mintun, Eric and Ravi, Nikhila and et al., "Segment Anything," 2023, CoRR. paper.
• (EAO-SLAM) Y. Wu, Y. Zhang, D. Zhu, Y. Feng, S. Coleman and D. Kerr, "EAO-SLAM: Monocular Semi-Dense Object SLAM Based on Ensemble Data Association," IROS 2020, Las Vegas, NV, USA, 2020, pp. 4966-4973. Paper.

7. Contact

• Author: Yaolin Tian (email: tianyaolin21@mails.ucas.ac.cn)

• Corresponding author: Xue Wan*, Email: wanxue@csu.ac.cn

• [Updates]: The paper has been accepted by TGRS (aim for lunar objects)

   @article{Yaolin2025,
     title={Lo-SLAM: Lunar Target-oriented SLAM Using Object Identification, Relative Navigation and Multi-level Mapping},
     journal={IEEE Transactions on Geoscience and Remote Sensing}, 
     author={Yaolin Tian, Xue Wan, Shengyang Zhang, Jianhong Zuo, Yadong Shao, Baichuan Liu, and Mengmeng Yang},
     year={2025},
     doi={10.1109/TGRS.2025.3547292}
   }
  

  @article{Yaolin2024,
    title={LO-SLAM: Lunar Object-centric SLAM using Point Prompted SAM for Data Association},
    author={Yaolin Tian, Xue Wan, Shengyang Zhang, Jianhong Zuo, Yadong Shao, and Mengmeng Yang},
    year={2024},
    eprinttype={techRxiv},
    doi={10.36227/techrxiv.170975343.37379344/v1}
  }