RGBD-SLAM-tutorial

Reference：

• [Tutorial Blog:一起来做RGBD-SLAM系列]. Gao Xiang
• [slambook-en]. Gao Xiang
• slambook code
• TUM dataset

How to run the code:

• Preparation: download openCV 3.4+ and PCL, Eigen3, Ceres, g2o, and other necessary libraries
• Step 1: download the dataset from TUM website above, then put it into [$repo_folder]/data/rgb_png and [$repo_folder]/data/depth_png, along with the data/rgb.txt, and data/depth.txt for dataset index. (The code about how to extract images and depths data are written in slamBase.h, you can change the path there)
• Step 2: build and compile the code

mkdir build
cd build
cmake ..
make

• Step3: run the code in /bin folder: ./bin/visualOdometry

Code structure

/src has the main source files.

• slamBase.cpp has most self-defined functions to compute keyPoints and Descriptors, estimate Motion, join point clouds and so on.
• visualOdometry.cpp implemented the VO.
• slamEnd.cpp integrated the G2O library for optimization.
• slam.cpp improved based on slamEnd and vo, by adding keyFrames extraction and check_loop_closure. It is the complete version based on the above steps, yet still has some bugs.

Bugs for future improvement

• core dump at some frames
• some wierd DTL algorithm bugs:

DLT algorithm needs at least 6 points for pose estimation from 3D-2D point correspondences. 
(expected: 'count >= 6'), 
where'count' is 5

• too many consecutive not enough inliers, maybe the distance between the frames are too far.

solving PnP
inliers: 0
inliers not enough, abandoning this frame

Example of results

(core dump, to be continued...)