|
Considerable resources are devoted to the acquisition, storage, and measurement of stereo imagery for the purpose of extracting terrain elevation data. Improvement in the accuracy of stereo parallax measurement allows a reduction in the resolution of images collected and measured to obtain a required elevation accuracy. A reduction in resolution requirements can lead to significant savings in collection, storage, and processing costs. A neural network is described which uses image cross correlation data to determine stereo parallax disparity to fractional pixel accuracy at each pixel location in the images. Correlation values are obtained between image windows in the left view with a succession of overlay window positions in the right view. High resolution parallax disparity determination requires small image windows. Correlation peak locations for the small windows are often unreliable match points due to noise and relative parallax distortion between the images in regions of rapid elevation variation. Further processing of the correlation data is needed to reduce errors. A neighborhood of correlation data is input to a neural network. The network outputs the parallax disparity for the pixel location (in the left view) centered on the neighborhood. The network was trained and tested using stereo images of a scene containing a variety of hills, terraces, flat terrain, and roads. The parallax disparities for this stereo pair were carefully measured manually at high resolution. These measurements provided accurate sub-pixel accuracy at the lower resolution used to train and test the network. A network with two `hidden layers' was trained using the backward-error-propagation method. Network results on separate test sections of the scene show significant improvement over results using alternate methods.
|
