Author : Ila
Year: 2010
- Variant of SLAM with only the robot trajectory is estimated and where landmarks are only used to produce relative constraints between poses
- Only loop closure links and non redundant poses taken into account ie higly informative links
- Two measures: distance between two poses and mutual information gain
- Leads to a more compact map
- Bottleneck for rt: data association that is solved with a tree based method
- Markov blanket of a node = all nodes linked or that are linked to this node
- difference made between observations linking two consecutive poses and observations linking other poses
- pose is redundant if it is too close to another pose already in the traj
- uses the formalism with information vector and information matrix this gives the following state update for a measurement
$y_i^n$ linking states$i$ and$n$ : $$ \Delta \eta = H^T \Sigma_y^{-1}(y_i^n - h(\mu_i, \mu_n) + H \mu_n ) \ \Delta\Lambda = H^T \Sigma_y^{-1} H $$ - mutual information gain = amount of uncertainty removed from the state
- erf = error function linked to a gaussian (erf(z/sigma*sqrt(2)) = probability of a centered gaussian to have a value between -z and z)
- mutual information gain: $$ I = \frac{1}{2} ln \frac{|\Lambda + \Delta \Lambda|}{|\Lambda|} $$
- matrix manipulation to accelerate the computation of
$I$ - bound pose similarity with interval arithmetic (ie. operations over intervals)
- pose tree to search for loop closure
- Very technical part about trees operations: recursive search & co
Experiments:
- real mapping with stereo vision, 3D landmarks and least square landmarks for pose estimation
- most of execution time due to nearest neighbour search