https://s3-us-west-2.amazonaws.com/secure.notion-static.com/c5bcc913-6f77-44bf-bbba-04f1758f0bfc/08421746.pdf
- Arxiv 업로드 버전 (복잡하지만 증명수식 포함)
https://s3-us-west-2.amazonaws.com/secure.notion-static.com/a7976229-50df-4582-aae1-6cb8bac6fe8e/1708.03852.pdf
0. Summary
Vision:
- KLT Sparse Optical Flow + Shi-Tomasi tracker
- Keyframe selection (Parallax/Quality criterion)
IMU:
- Pre-integration to avoid re-propagation
- Bias adjustment using First-order approximation, or re-propagation if a change in estimation is above a certain threshold
VINS-Estimator initialisation (Visual-Inertial Alignment):
- Vision-only SfM (five-point algorithm, PnP method, global full bundle adjustment).
- This method does not comprise scale information.
- Gyroscope bias calibration + IMU re-propagation, based on rotation information from SfM
- Velocity, Gravity, Metric Scale initialisation, based on SfM and newly re-propagated IMU path after gyroscope bias calibration
- Gravity refinement, based on the gravity vector calculated in the previous step
- Recover position, orientation, velocity, scale relative to the world frame, based on all the previous steps.
Tightly-coupled VIO cannot work without Initialisation step!!