Observation Model and Approximate Posterior - 

The observation model is very important since the multiple hypothesis are generated in the image observation, and this is why the system should be able to propagate the whole density function. However, it is still valuable when the observation model produces Gaussian density since the computationally expensive covariance matrices do not need to be propagated as in the Kalman filtering case. Instead, only a fixed number of sampled states are propagated with their weights, which approximated the conditional density.

Considering the equipment and time constraint, the observation model is chosen to be Gaussian as follows

where the  is chosen as in [1].

 
Figure: 10 likely posterior samples from -  : kth frame
 
 
 

Tue Mar 14 02:05:32 PST 2000