Author: Shen
Year: 2015
- superior size, weight, and power (SWaP) characteristics of a VINS setup
- first to demonstrate autonomous navigation with a VINS setup
- Perform non linear optimization on a bunch of KF to estimate the gravity vector, all the robot states and the depth of the features
- all details in : Initialization-free monocular visual-inertial estimation with application to autonomous MAVs
- The bias are not estimated?
- Pre integration using quaternions
- Compute the covariance using the time derivative of the deltas
- Information matrix of projection factors = FOCAL^2 * I
- Two way marginalization scheme: in hovering (vol stationnaire) the scale ambiguity may arise as keyframes are voted even with no motion to prevent from IMU divergence: marginalize the most recent state instead of the last in this case
- All the features that were first observed in a marginalized frame are also marginalized
- 30 KF and 200 features in the sliding window
- 30 pixels of parallax (after rotation compensation ?) before being included in the sliding window
- Add KF every 0.1s
- Bias experimentally around 0 => not included in optimization