Vision-based mobile robot localization and mapping using scale-invariant features

Stephen Se, David Lowe, Jim Little (2001)

Key points

Robot odometry data can only give a rough estimate of motion, but can be used to limit the search area for matching SIFT features
- Camera ego motion is determined using least squares minimization
Maintain a database containing SIFT landmarks (different types) observed and use it to match features in subsequent views
- Permanent landmarks: when the environment is clear, mark stable features as permanent, which means they will be kept in the database even after not being seen for many frames, in order to increase localization performance
- Also, allow multiple "view vectors" per SIFT landmark, to deal with large changes in viewpoint
Currently, only map-reuse if the start position is the same as the last stop position (work needed in this area)
Kalman filtering is used to deal with image errors