Object Recognition Using Color And Geometry Indexing

This thesis addresses the problem in computer vision of recognizing a library of objects in a scene without knowing which objects are in the scene or where they are located. Inherent in this topic is the massive number of possible sets of features from both the scene and the known objects that need to be identified. For an object recognition system to function under these conditions, it must be able to quickly eliminate most of the features in the scene which are unlikely to be a part of any objects known to the system. The few features that remain will require much less time for the system to identify, thus making such a recognition system viable in its computing requirement. I approach the problem of eliminating most of the unlikely features with the use of the color in the scene and of the model objects. Any portion of the scene that contains a color that does not appears on any of the models is eliminated. A 3D object recognition theory developed by David Jacobs is then used on the remaining portions of the scene to recognize objects by using their shape features. The work in this thesis is at an exploratory stage. Nevertheless, I will show some preliminary results in which the use of color can greatly reduce the number of possibilities that a geometric system has to consider, thereby demonstrating that this approach is promising. Thesis Supervisor: W. Eric L. Grimson Title: Professor