This visual SLAM implementation follows the ORB-SLAM paper (Mur-Artal et al., 2015), and implements map initialization (Map::initializeMap()), tracking (Map::localTracking()) and new point mapping (Map::localMapping()) using OpenCV algorithms. Bundleadjustement is also implemented to optimize the estimated map points and poses (Map::BundleAdjustement()) using g2o.
Aim is to migrate this program to ROS environment, and use deep learning to do 3D modeling in real-time. Here is what is currently being produced offline using https://github.com/ardaduz/deep-video-mvs.
Requirements:
- OpenCV https://opencv.org
- Eigen3 http://eigen.tuxfamily.org
- In linux: sudo apt install libeigen3-dev
- g2o https://github.com/RainerKuemmerle/g2o
- Easy3D https://github.com/LiangliangNan/Easy3D
- cmake (compiled with VERSION 3.11.0)
- C++14 compiler
Libraries (apart from Eigen3) can be installed from the links or just by building them from the /libs folder. It is recommended to make with sudo make install, as this installs them to /usr/local/include folder where the program looks for them by default.
After installing the requirements, out of source build can be achieved by the following command sequence:
- cd path-to-cmakelist
- mkdir build
- cd build
- cmake ../
- make
This creates a run file named “run_slam”.
- Add filtering for bad poses and points