Amortized Constant Time State Estimation in SLAM using a Mixed Kalman-Information Filter

摘要

The computational bottleneck in all informationbasedrnalgorithms for SLAM is the recovery of the state mean andrncovariance. The mean is needed to evaluate model Jacobians andrnthe covariance is needed to generate data association hypotheses.rnRecovering the state mean and covariance requires the inversionrnof a matrix of the size of the state. Current state recovery methodsrnuse sparse linear algebra tools that have quadratic cost, eitherrnin memory or in time. In this paper, we present an approach tornstate estimation that is worst case linear both in execution timernand in memory footprint at loop closure, and constant otherwise.rnThe approach relies on a state representation that combines thernKalman and the information-based state representations. Thernstrategy is valid for any SLAM system that maintains constraintsrnbetween robot poses at different time slices. This includes bothrnPose SLAM, the variant of SLAM where only the robot trajectoryrnis estimated, and hierarchical techniques in which submaps arernregistered with a network of relative geometric constraints.