Automatic Target Recognition, Wavelet Transforms and Stereo Matching
Final rept. 1 Apr 1998-31 Mar 2001
NEBRASKA UNIV LINCOLN DEPT OF COMPUTERSCIENCE AND ENGINEERING
Pagination or Media Count:
Stereo Imaging is very important in the analysis of aerial imagery. The proposed project considers the problem of stereo matching using wavelet transforms. In our method, the wavelet transformation of a signal using a wavelet basis with nth vanishing moment is regarded as nth edge of the signal after some smoothing operation. We then analyze the physical meaning of each channel of the sub-images decomposed by a wavelet. Eased on the analysis, we proposed a hierarchical algorithm that combines both feature-based and pixel-based methods together, to compute the disparity of the binocular stereo image. We have tested the proposed algorithm on a large number of natural and synthetic scenes. Our method is effective on stereo image pairs with large disparities that is a case, which is very important for applications as structure from stereo. We also proposed a general wavelet-based hierarchical matching scheme which involves a dynamic detection of interesting points as feature points at different levels of sub-band images via wavelet transform, and adaptive threshold selection based on compactness measures, a guided searching strategy for the best matching from coarse-to-fine levels. Also, some other problems are investigated including the extraction of 3D lines from range images, and dynamic shape retrieval by curve matching and active contour models, and applications of fuzzy logic to project selection. The project falls under the Nebraska DEPSCOR Research Priority Area of Information and Communications Science and Technology awarded by the Ballistic Missile Development Agency BMDO and it is administered through the AFOSR.
- Numerical Mathematics