Year: 2022
Author: Ran Duan
- Observation: 2D-3D coresp with low reprojection error => high pose estimation error
- Idea: implicit pose error is better for outlier rejection, possible with stereo only
- Perform as well as IMU aided system
- Stereo view => prior knowledge of orientation
- Why visual pose estimation can produce large error? => bias in orientation of inliers
- It is still chalenging to develop a stereo VO solution
- Inlier/outlier polynomial system ? using the squared norm of error ?
Methodology:
- grid Feature Detection to offer a proper feature distribution
- Stereo intialization , KLT tracking and pnp for pose estimation
- Typical VO
- the pose of a landmark is evaluated by interpolating between the last estimation and the current one, and it is declared as an outlier with an SOS feasability test (?)
- Build an optimization problem based on the 2D-3D projection $$ p_k = RP_k + t \ R = (I-[c]{\times} )^{-1}(I + [c]{\times}) \ (I-[c]{\times} ) p_k \times \left ( (I + [c]{\times})P_k + (I-[c]_{\times} )t \right) = 0 $$
- SOS theory ultra hard to understand :o