rgbd-slam-practice

Overview

This project is related to RGB-D SLAM (Simultaneous Localization And Mapping). It is only a small practice for me to go through the entire pipeline of keypoint-based visual SLAM and to pick up my C++ programming. The code largely depends on this tutorial.

Prerequisite

g2o: I encountered several problems with this library and I have modified it a bit. The one works well is in 3rdparty/ directory. What I've done in my uploaded g2o package are,

$ git clone -b c++03 https://github.com/felixendres/g2o.git
  # add “add_compile_options(-std=c++11)” to CMakeLists.txt
$ cd g2o
$ mkdir build && cd build
$ cmake .. -DG2O_BUILD_EXAMPLES=OFF
$ make -j4

PCL: For Ubuntu 16.04, you can directly install it using sudo apt-get install libpcl-dev pcl-tools. Otherwise, you can build it from source.
OpenCV: Remember to check the version being used in CMakeLists.txt.

How to Run

Build the code,

$ cd ${RGBD-SLAM-PRACTICE-ROOT_DIR}/build
$ cmake ..
$ make

A set of RGB and depth images are required for RBGD-SLAM. Normally, a pair of RGB and depth images should have the same name but are placed in ./data/rgb_img and ./data/depth_img directories resepectively. The images in a directory will be read in numerical order of the filenames. There are several adjustable parameters specified in parameters.txt. Run the code,

$ cd ${RGBD-SLAM-PRACTICE-ROOT_DIR}
$ ./bin/slam ./parameter.txt
$ pcl_viewer ./results/result.pcd # other outputs are also in ./results/ directory

Results

Using data provided by the tutorial, we can construct 3D model in point cloud like,

Notes

If you are using other detector or descriptor like SIFT or SURF, you need to compile OpenCV with nonfree package.
For Ubuntu 16.04, g2o_viewer is broken (*** Error in `g2o_viewer`: realloc(): invalid pointer: 0x00007fe7d3d0c820 ***), at least on my computer.
There is a mysterious segmentation fault at the end of the slam program and it may be related to g2o package I guess.