Author: Leutenegger
Year: 2013
- tight fusion of IMU and reprojection in a single cost function
- marginalization to maintain a bounded sized optimization window
- historically fusion is based on filtering
- nonlinear estimation is good to reduce suboptimality due to linearization
- IMU introduces temporal constraints
- landmarks are represented in homogeneous coordinates
$p = [u^T s]^T$ - perturbation on the tangent space of state space: normal vector space for position, speed and bias and axis-angle perturbation for quaternion
- VOnly has 6DOF, visual inertial 4DOF as gravity vector fixes two
- standard reprojection error
- accelerometer bias is modeled as a biased random walk
- descriptors (BRISK) are extracted oriented along the gravity direction (thx to IMU)
- brute force matching for keypoints of the local map
- outlier rejection performed with a chi squarred test on image coordinates obtained with IMU prediction
- marginalization on landmarks that were visible on the marginalized KF but not on the current one