What are the optimization techniques for SLAM problems?

The need for high definition maps, mapping of unknown GPS-denied areas, exploration and autonomous mission are the main motivation for Solving the Simultaneous Localization And Mapping Problem “SLAM”

Given a map, a robot could localize itself easily by comparing its observations to the stored map. Given a robot pose relative to a global axis “localization” makes it easy to map the environment.

A map is built by the incorporation and stitching of the robot observations, but these observations are measured relative to the robot frame of reference due to the existence of the observing sensor on-board.

To refer these observations to the map axes , robot pose needed to be known so that the observation could be transformed from the robot frame to the map frame, that is why a localization is needed for mapping.

OK, that’s seems good. Given a map we could get pose, given a pose we could get map. What if we do not know neither of them, like being in an unknown place with neither a known maps nor a localization system ?.

You have to map it yourself. But the robot has no map localize it self to build the map. That’s why it is an egg-chicken problem and where do simultaneous name came from.

If there is no noise associated with sensors, the problem could be solved easily, by tracking the robot position from the origin and map the environment consequently.

Unfortunately noise does exist and will be accumulated over time leading to catastrophic results. Probabilistic approaches are introduced then to filter and minimize sensors noise.

Filtering out these noise using different methods like least square error minimization or filtering is the main task of the SLAM problem.

Extended kalman filter (EKF) SLAM, rao-blackwellized particle filter and graph SLAM represent some approaches to solve that problem.