Hybrid Bayesian fusion of range-based and sourceless location estimates under varying observability

The paper proposes a hybrid Bayesian approach
for multi-sensor data fusion for 3D localization. The approach
addresses the problem of fusing range-based and sourceless
localization estimates under conditions of varying observability in
the range-based sub-system. The proposed localization approach
uses a mixture of Single-Hypothesis-Tracking (e.g. Kalman filter)
and Multi-Hypothesis-Tracking (MHT) (e.g. Particle Filters)
Bayesian filtering to improve tracking accuracy under conditions
of varying observability. Under conditions of sufficient (or no)
range measurements a single hypothesis approach is used. Under
the condition of insufficient range measurements (i.e, 1 or 2
ranges), MHT is used, since it more accurately models the
distribution of real error in the estimated positions by means of
Gaussian mixtures rather that a single Gaussian. The results show
up to 10% improvement in 3D position estimation as compared
to Single-Constraint-at-a-Time (SCAAT) approach and upto 24%
improvement compared to an Extended Kalman Filter approach
for intermittent 3 second partial range occlusions when tracking
human arm movements.