Author: Yin
Year: 2022
- VISLAM in dynamic envirronment: couple stereo scene flow with IMU for dynamic feature detection
- Tightly couples dynamic (virtual landmarks) and static features in non linear optimization (instead of discarding dynamic features)
- uncertainty estimation of the scene flow
- in soa: semantic segmentation or geometric methods to remove dynamic features (for vision only)
- dynamic feature detection is a thread running in parallel
- scene flow is a 3D vector field define for each point motion in the world frame (should be zero for static points)
- to obtain scene flow $\delta M $ between adjacent frames
$i$ and$j$ : use KLT for frame to frame matching and SGM (semi global matching) for disparity to estimate the lmk$n$ in cam frame $ { p^{c_i}_n, p^{c_j}n } $ and use IMU preintegration to get ${}^{c_i} T{c_j}$ $$ \delta M_n = p^{c_i}n - {}^{c_i} T{c_j} p^{c_j}_n $$ - histogramm analysis demonstrates that scene flow is corrupted by depth noise and other factor => thresholding doesn't work
- So compute the covariance of the scene flow taking into account: IMU pre integration, SGM disparity estimation, KLT tracking
- then uses
$\chi^2$ rejection - virtual landmark: use a constant motion model to predict its future pose
- Loop closure module using only static features
- better than ORBSLAM & co on dynamic env