Automatic target recognition based on collection of evidence

Abstract

The problem of automatically recognizing an object in an image scene is very difficult. This thesis develops an image-based object recognition algorithm in which information from different features is combined using Dempster-Shafer reasoning. Specific attention is paid to cases in which only partial information is available because of occlusion or sensor limitations. The structure of the recognition system developed herein is as follows. First, some image processing techniques are used to filter out noise, detect edges, and find features in the raw image, Further preprocessing is performed to isolate objects of interest. Finally, Dempster-Shafer reasoning is used to combine evidence from the edge features into a working model of the objects seen in the raw image. The preceding object recognition system was tested on simulated data, dealing with sets of geometrical shapes, Two experiments were performed, one with un-occluded objects, and one with up to four occluded objects in each raw image. Its performance was compared to the Bayesian approach and human classification. Although Dempster-Shafer reasoning did not outperform human reasoning, it did perform considerably better than the Bayesian approach.