Page 265 -

244 5 Segmentation

(a)

(b)

Figure 5.6 Probability density propagation (Isard and Blake 1998) c 1998 Springer. At the beginning of each
estimation step, the probability density is updated according to the linear dynamic model (deterministic drift)
and its certainty is reduced due to process noise (stochastic diffusion). New measurements introduce additional
information that helps refine the current estimate. (a) The Kalman filter models the distributions as uni-modal,
i.e., using a mean and covariance. (b) Some applications require more general multi-modal distributions.