Right now I have a particle filter that uses SIFT features to provide estimates on the location of a airborne camera. What happens is that SIFT features are pre-calculated for a given map and they are stored in a database where they can be quickly looked up.
Currently one aerial image is taken and its SIFT features are extracted. These are then matched against the features of the database and the features that have a good match are used to generate hypothesis about the location of the vehicle.
These hypothesis are then used to power the update phase of the particle filter. This is done by firstly strengthening the particles that fall in the range of one of the hypothesis, and then adding one sample single particle for each of the new hypothesis presented.
What happens is that the particles correctly accumulate around the position on the map where the centre of the taken image lies. They are, therefore, able to find the position of the image on the map.
What I am working on is a way to aggregated the information present in the particle filter into knowledge. Currently I am trying to use expectation maximization to find Gaussian functions that best represent the knowledge in the particles.
