This proposal builds upon many years of research and experience in the problems of sensor fusion, and 3D modeling. Powerful modeling and data fusion systems, previously developed by the investigators at USC and at Sentinel, will be extended with new capabilities to produce a step change in the automation of modeling and feature extraction for geospatial database systems. The new algorithms are based on our recent explorations into perceptual grouping and classification with sensor cues, feature invariants, and machine learning. Specifically, we integrate the mathematics of tensor voting, Gabor wavelets, and feature-based recognition to automate the extraction of geometric models and terrain features from aerial images and Lidar data. Early results show the feasibility and advantages of our approach. The integration of these methods to attack the feature extraction and modeling problems has never been attempted, to the best of our knowledge; therefore our effort brings a new approach to the geospatial database problem, with the potential for making significant improvements in automation, robustness, and throughput.
