This project (proposal) aims at implementing a complete system for creating realistic 3D maps from SLAM algorithms. In my proposed approach, it's highly related with offline-SLAM.
Sparse online SLAM algorithms only produce sparse pointcloud, which is not ideal for the human. Surface reconstruction would produce more human-friendly maps. However, several issues need to be solved for performing surface reconstruction on real-time SLAM result. This includes globally optimize 3D points, enforcing smoothness,etc.
Besides, it's possible to solve the problem via an offline-SLAM approach (with surface reconstruction), because it's highly related with that.
The main part of this project is related to offline-SLAM, to compensate some trade-offs made by real-time SLAM for real-time consideration. This includes:
- pose-graph optimization
- bundle adjustment
- enforcing smoothness
- surface reconstruction
- additional details about improving pointcloud accuracy
This project may combine several technicals used in state-of-art SLAM algorithm.
Realistic 3D maps
Realistic 3D maps refer to maps with surface reconstruction (with color) so that they are more human-friendly.