Talker: Norikazu Ikoma (Kyushu Institute of Technology)
(Joint work with Hiromu Hasegawa (MegaChips Corporation))
Title: Fusion of Multi-modal Features in Particle Filter with Dual
Weight of Belief and Plausibility for Track-Before-Detect Multiple
Target Tracking
Abstract: Dual weight of belief and plausibility have been
introduced to cope with fusion problem of multi-modal features in
observation process within a framework of track-before-detect visual
tracking by particle filter for multiple target. Observation model
consists of dual function of belief and plausibility corresponding
to conjunction and disjunction of multi-modal features. Each
particle has dual weight corresponding to the two likelihood
functions, and the two weights are updated respectively. Resampling
step involves some elaborations consisting of three steps such that;
1) normalized weights of plausibility are used as the
probability to draw with replacement, 2) uniform value is set for
the weights of plausibility after the draw with replacement, and 3)
weights of belief are adjusted for each particle. The idea of
dual weight has been extended to multiple target tracking framework
with SMC-PHD filter. Performance of the proposed method has
been demonstrated for multiple people tracking over videos captured
by a fish eye camera at ceiling.