Kalman Filter Results Fred Astaire and Ginger Rogers

  • Slides: 6
Download presentation
Kalman Filter Results Fred Astaire and Ginger Rogers Shall We Dance, MGM 1941 J.

Kalman Filter Results Fred Astaire and Ginger Rogers Shall We Dance, MGM 1941 J. M. Rehg © 2003

Classical MHT z Approach y. Tree of Kalman Filters (e. g. Reid 1979) y.

Classical MHT z Approach y. Tree of Kalman Filters (e. g. Reid 1979) y. Enumerate all assignments Blip 1 z Properties y. Requires discrete features yp(z|x) is multimodal ? z Problems for figure tracking y. No discrete feature-detector for articulated figure J. M. Rehg © 2003 Blip 2 Blip 3

Monte-Carlo Methods z Approach y. Nonparametric (sample-based) pdf updated via SIR. y. Particle filter

Monte-Carlo Methods z Approach y. Nonparametric (sample-based) pdf updated via SIR. y. Particle filter for time series data ye. g. CONDENSATION algorithm (Isard & Blake 1996) z Properties y. Can approximate arbitrary density functions y. Number of samples exponential in size of state space z Problems for figure tracking: y. Prohibitive number of samples are required due to x. High dimensions (38 states) x. Weak dynamical model (constant velocity) J. M. Rehg © 2003

Likelihood from Image Measurements z Likelihood measure from squared pixel residual error: z Template

Likelihood from Image Measurements z Likelihood measure from squared pixel residual error: z Template registration locally maximizes likelihood measure. 0. 9 J. M. Rehg © 2003 0. 01

Multiple Hypothesis Tracking Cham & Rehg, CVPR 99 Maximum Likelihood Result Prediction p(xt|Zt-1) p(xt|Zt-

Multiple Hypothesis Tracking Cham & Rehg, CVPR 99 Maximum Likelihood Result Prediction p(xt|Zt-1) p(xt|Zt- p(xt-1|Zt-1) 1) x Observation p(zt|xt) Bayes Rule p(xt|Zt) |x ) p(xt| zt)p(xt|Zp(z t-1) t t p(zt|xt) x J. M. Rehg © 2003 x x xx

Further Tracking Results J. M. Rehg © 2003

Further Tracking Results J. M. Rehg © 2003