Observation

The conditional density  is multi-model in severe visual clutter, and the observation module in the system should be implemented considering this. However, it is very computationally inefficient to look at whole image at every instant, so the authors of [1] proposed to look for image features along normals from the previous object contour estimate. However, it is not very clear in [1] how the authors designed their observation model in such way that the resulting density function carries truly the multiple hypothesis about the current state. This is an open research problem, and will be discussed in the section 4.

Tue Mar 14 02:05:32 PST 2000