Zia Khan, Tucker Balch, Frank Dellaert

1
Zia Khan, Tucker Balch, Frank Dellaert Georgia
Institute of Technology
Problem Traditional methods for data
association assume (1) a target generates at
most one measurement (2) a measurement
originates from at most one target Assumptions
are often violated in computer vision
applications.
Rao-Blackwellization If we assume the motion is
linear Guassian, we can marginalize over the
state of the targets given a sampled data
association.
Contributions Data association method that
addresses multiple, split measurements from a
single target and merged measurements from
multiple targets. Multiple hypothesis tracking
capability in a Rao-Blackwellized auxiliary
variable particle filter framework.
Sparse Recursive Least Squares Marginalization
involves solving a large least squares problem
The sparse observation matrix represents the data
association.
Forthcoming Work Naive implementation O(Tmn2).
MCMC sampler that leverages sparse least
squares updating and downdating. O(Tn2) when
all targets are interacting O(T) when no
interactions are present T mixing time of
Markov chain m number of measurements n size
of target state space Please see the laptop
demonstration.
Markov chain runs in the space of all bipartite
graphs.