| Three-dimensional models of the environment have many current uses and great potential applications for intelligence operations, biological sciences, and city planning. How best to accurately and efficiently recover the three-dimensional structure of a scene from a number of two-dimensional aerial images and additional contextual information, however, represents a significant and challenging problem. Sensor data are now widely available from aerial surveys, autonomous aerial vehicles, or other sources. Traditionally, image-space stereo algorithms have been used to turn this data into usable three-dimensional information such as digital elevation models.; We contend that a better methodology should involve the use of all available images simultaneously, as well as additional information from other sources, and propose a matching scheme that adjusts a surface model in a world coordinate system (object space) via an optimization process. A cost function is applied in a local surface optimization to minimize the differences between the object-space model and the image data at each location on the surface, taking into account other constraints. We apply constraints due to prior knowledge about the scene as well as resource constraints imposed by limited computational resources. Our model is recovered in a flexible and powerful representation that has already seen wide use in modeling for Computer Aided Design and Manufacturing (CAD/CAM): non-uniform rational B-splines (NURBS). |
