xuankuzcr/global-lvba

Global-LVBA is a research tool for robots and vehicles that build maps of their surroundings using both cameras and LiDAR, a sensor that measures distance with laser pulses. It is meant to run after a system like FAST-LIVO2 has already produced an initial map and set of camera positions, and its job is to clean up and sharpen that result. When a robot moves through a space while scanning it, small errors build up over time. These errors can make the resulting 3D point cloud map look layered or smeared instead of solid, and can leave the camera's recorded positions slightly off from where objects actually are. Global-LVBA takes the whole recorded path at once, rather than looking at short overlapping windows, and adjusts everything together so the point cloud lines up more consistently and the camera positions match the scene down to the pixel level. The project is written in C++ and depends on several specialized libraries: Ceres Solver for the underlying math optimization, Eigen for linear algebra, OpenCV for image processing, and Python 3.8 as a supporting tool. Building it requires compiling a bundled component called SiftGPU and then building the whole project with catkin, the build system used by the ROS robotics framework. To use it, you need a specific folder structure containing timestamped images with recorded camera poses, timestamped point cloud files with recorded LiDAR poses, and optionally a COLMAP feature matching database, since the README notes that COLMAP produces better feature matches than the built in matching tool. A ready made example dataset is provided so you can test the pipeline before using your own recordings. The project is released under the MIT license, meaning it can be used freely, including for commercial purposes, as long as the copyright notice is kept.