Optimization approach on the fusion of multiple images for classification: object recognition

The project "optimization approach on the fusion of multiple images for classification" is concerned with the search for camera settings which allow an information theoretically optimal acquisition of an observed scene. The scene acquisition is modeled as a probabilistic state estimation problem. The camera settings are optimal if they lead to the lowest possible uncertainty of the state estimation. The uncertainty is measured by the entropy of the probability density function on which the state estimation is based.

Concerning the object recognition, research aims at the problem of which sequence of object views is to be chosen for object recognition, i.e. which camera movements, if actively selectable, are most favorable. Therefore, the criterion is the mutual and flexible optimization of competing demands regarding the recognition performance as well as the required effort. Important parameters within this framework are the number of required views, the consecutive maximization of the object discriminability and the reduction of potentially cost- and time-sensitive camera actions. Especially the problem of objects with an arbitrary number of ambiguities is thought to be solved with this approach, namely the view planning algorithm.

Figure 1: Exemplary depiction of the view planning problem.

The main matter of view planning is supplemented by other essential points of interest which have to be regarded within the scope of this work. One important aspect there is the fusion of sensor data concerning the spatial and temporal dimension. The probabilistic integration of information about a series of image data as well as its spatial correlation is elementary for being able to address the view planning problem at all. Another question is how to deal with objects never seen before and how to decide on a next best view of those objects in order to raise discriminability. Furthermore, a method has to be provided that is able to make use of the knowledge accumulated so far in an arbitrary situation, that is to provide a mapping function for finding the optimal solution.