Author: Forster
Year: 2017
- Semi-direct: track high intensity gradient pixels with direct methods and joint optimize with feature based methods
- introduce sparse image alignement algorithm (direct method for frame-to-frame motion)
- 3D features creation with Bayesian recursive depth estimation
- pose refinement with BA (BA used only for refinement)
- very fast: select features only on KF, direct method permits to get rid of data association, sparse reconstruction
- generalization of SVO to edgelet, wide fov and multi cam
- motion estimation thread: frame to frame algnement + image alignement with map + BA refinement
- mapping thread: at each KF, 3D point init if the uncertainty of the bayesian filter is low enough
- use scaramuzza model for fisheye
Discussion:
- tracks less features than LSD SLAM, more efficient with high frame rate cameras
- high speed mode: refinement with only the latest pose
- high accuracy mode: smoothing with iSAM2