Author: Miao
Year: 2021
- new cmaera model for image rectification
- camera in a waterproof housing => refraction to be rectified
- New calibration method that separates lens distortion and water air refraction
- robust front-end for keypoint tracking
- sparse optimization model that integrates both photometric and reprojection errors
- Underwater dataset
Preprocessing:
- IMU pre integration on manifold (classic)
- takes refraction into account in camera calibration
Data association:
- local optical flow tracking
- corners detected first, then strong gradient point if not enought
- KF selection: number of long termed tracked features and avg moving distance of tracked features
- LKT often failed when environnement illumination changes => new tracking method based on both optical flow and illumination changes estimation
Windowed optimization:
- classical bundle adjustement for reprojection error
- direct residuals = difference between patches intensity around points
- direct and feature based residuals are unified in a common residuals
- classi marginalization