This repo presents the code for my Master's thesis. Some of the work is also presented in the following paper (accepted at ICPR 2018):
- Incremental 3D Line Segment Extraction from Semi-dense SLAM, Shida He, Xuebin Qin, Zichen Zhang, Martin Jagersand (arxiv)
Our method simplifies the per-keyframe pointcloud produced by a semi-dense SLAM system using 3D line segments. Then, we take the extracted 3D line segments and reconstruct a surface using them. In this way, surface of the scene viewed by the camera can be reconstructed while the camera is exploring. Our method produce accurate 3D line segments with few outliers, which makes the reconstructed surface more structually meaningful than surface reconstructed using feature points or random selected points. See my thesis page for more details.
This version of software is based on ORB-SLAM2. We use the semi-dense version of ORB-SLAM which is implemented based on the technique described in this paper. We release this software under GPLv3 license. See Dependencies.md for other dependencies.
If you use this software in an academic work, please cite:
@inproceedings{he18icpr,
title={Incremental 3D Line Segment Extraction from Semi-dense SLAM},
author={Shida He, Xuebin Qin, Zichen Zhang, Martin Jagersand},
booktitle={International Conference on Pattern Recognition, {ICPR} 2018},
year={2018}
}
The software is tested in 64-bit Ubuntu 14.04.
We use the new thread and chrono functionalities of C++11.
We use Pangolin for visualization and user interface. Dowload and install instructions can be found at: https://github.com/stevenlovegrove/Pangolin.
We use OpenCV to manipulate images and features. Dowload and install instructions can be found at: http://opencv.org. Required at leat 2.4.3. Tested with OpenCV 3.1.0.
Required by g2o (see below). Download and install instructions can be found at: http://eigen.tuxfamily.org. Required at least 3.1.0.
BLAS and LAPACK libraries are requiered by g2o (see below). On ubuntu:
sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
We use modified versions of the DBoW2 library to perform place recognition and g2o library to perform non-linear optimizations. Both modified libraries (which are BSD) are included in the Thirdparty folder.
We provide some examples to process the live input of a monocular, stereo or RGB-D camera using ROS. Building these examples is optional. In case you want to use ROS, a version Hydro or newer is needed.
Apart from the dependencies list above for ORB-SLAM2, additional libraries are needed.
sudo apt-get install libcgal-dev
sudo apt-get install libboost-all-dev
We use EdgeDrawing edge detector in our system. The library is available as binary files. The version for 64-bit Linux is included in the Thirdparty folder.
We compare our system against methods using EDLines and Line3D++. EDLines is avaiable as binary files and the version for 64-bit Linux is included in Thirdparty folder. We also include Line3D++ in the Thirdparty folder and it will be compiled when running the build.sh script.
Similar to ORB-SLAM2, build.sh can build the Thirdparty libraries and semi-dense ORB-SLAM2 with 3D line segment extraction and surface reconstruction. Please make sure you have installed all required dependencies (see section 1).
Execute:
chmod +x build.sh
./build.sh
Sometimes the build will fail with errors related to c++11 standard, and simply running the script again should fix the issue.
The system can be run in the same way of running ORB-SLAM2. Here are examples for running on EuRoC and TUM-RGBD dataset (monocular). It can also run using ROS. See ORB-SLAM2 for more details.
-
Download a sequence (ASL format) from http://projects.asl.ethz.ch/datasets/doku.php?id=kmavvisualinertialdatasets
-
Execute the following first command for V1 and V2 sequences, or the second command for MH sequences. Change PATH_TO_SEQUENCE_FOLDER and SEQUENCE according to the sequence you want to run.
./Examples/Monocular/mono_euroc Vocabulary/ORBvoc.bin Examples/Monocular/EuRoC.yaml PATH_TO_SEQUENCE_FOLDER/mav0/cam0/data Examples/Monocular/EuRoC_TimeStamps/SEQUENCE.txt
./Examples/Monocular/mono_euroc Vocabulary/ORBvoc.txt Examples/Monocular/EuRoC.yaml PATH_TO_SEQUENCE/cam0/data Examples/Monocular/EuRoC_TimeStamps/SEQUENCE.txt
-
Download a sequence from http://vision.in.tum.de/data/datasets/rgbd-dataset/download and uncompress it.
-
Execute the following command. Change
TUMX.yamlto TUM1.yaml,TUM2.yaml or TUM3.yaml for freiburg1, freiburg2 and freiburg3 sequences respectively. ChangePATH_TO_SEQUENCE_FOLDERto the uncompressed sequence folder.
./Examples/Monocular/mono_tum Vocabulary/ORBvoc.bin Examples/Monocular/TUMX.yaml PATH_TO_SEQUENCE_FOLDER
In default configuration, the sparse SLAM system run once and then all other processing (including semi-dense mapping, 3D line segment extraction and surface reconstruction) happens offline afterwards. Note the offline processing can take a long time which makes the windows unresponsive. Please wait until the processing is finished in order to save the results. Online processing can be enabled by uncomment the OnlineLoop macro in ProbabilityMapping.cc file. However, this is not recommended due to the fact that real-time performance is not yet guaranteed.
The results are saved in the directory results_line_segments under a subdirectory named by the starting time.
In each result directory:
- info.txt reports time usage and used parameters.
- model.obj is the reconstructed mesh of the scene.
- semi_pointcloud.obj is the raw semi-dense pointcloud from semi-dense mapping.
- line_segments.obj contains the extracted 3D line segments before clustering.
- line_segments_clustered_incr.obj contains the clustered line segments.
- line_segments_edlines.obj contains the line segments extracted using decoupled line segment fitting.
- Files with Line3D++ in their names are created by Line3D++.