Summary
RBPF divides the slam problem into two parts: localization and mapping
drawbacks: the number of particles? and the particle-depletion problem
challenges: reduces the quantity of particles
contributions in paper [?Improved Techniques for Grid Mapping With Rao-Blackwellized Particle Filters---Giorgio Grisetti, Cyrill Stachniss, and Wolfram Burgard?]
1. the improved proposal distribution considering the movement of robot and the recent observation that gives accurate manner
2. adaptive resampling technique that reducing the risk of depletion??
Algorithm
1.mapping with known poses P(m| x, z)
2.estimate the trajectories P(x| z, u) using PF
?PF———E.X. SIR
procedure：1) sampling from proposal x(i) [how to compute proposal]
? ? ? ? ? ? ? ? ? ? 2) importance weighting w(i) —— give a recursive formulation
? ? ? ? ? ? ? ? ? ? 3) resampling? [when the resampling step should be carried out]
? ? ? ? ? ? ? ? ? ? ?4) map estimation P(m(i)| x(i), z)
Part I : improved proposal distribution
used:?odometry motion model——easy to compute but suboptimal when the sensor more precise than odometry
improved: integrating the most recent sensor observation z into proposal——optimal respect to weights——gaussian approximation for each particle
how to compute the gaussian approximation
1.scan matcher to determine the meaningful area L of likelihood function
where scan matcher can use lots of algorithms, and gradient-descent search
2.sampling in the L and evaluate the points
Part II : adaptive resampling
resampling can decrease the computation but maybe remove the good particles
give a N(eff) which represent the as threshold N/2
