Automatic target recognition using location uncertainty.

摘要

In this dissertation, we present a framework for using location uncertainty information in computer vision applications. This framework is applied to the military automatic target detection and recognition application. We take a model-based approach to accumulating weak but consistent target evidence. Reliable target detection and recognition is achieved by making use of the location uncertainty information not being utilized by existing algorithms. The development of the location uncertainty measure consists of three major pieces: the relative importance of boundary points as determined by the geometric relationship between the location uncertainty of the centroid and boundary points, the relationship between the signal-to-noise ratio and the location uncertainty at the boundary points, and the optimal estimation of the image gradient at the boundary points. With sound mathematical models, the study of these sub-problems yield meaningful results useful not only in this framework, but in many other general problems as well.;The results of our experiments with real and simulated image data show that the centroid location uncertainty feature computed by the proposed framework is very effective in target detection and recognition. As a powerful addition to existing automatic target recognition algorithm modules, it has been successfully combined with the traditional matched filter to give further improved target detection and recognition performance.;Performance evaluation is always an important part in any new algorithm development. For characterizing the detection and recognition performance of computer vision algorithms, a new methodology is developed to overcome some problems with existing methods. An optimal matching problem is formulated to describe the situation. It is then transformed into an unconstrained assignment problem which enjoys an efficient solution technique: the Hungarian algorithm. This results in a one-to-one correspondence between ground-truth and declared entities and yields more precise performance measures.