|
Match points between stereo image pairs can be used to derive terrain elevation data from aerial photography and to determine obstacle and target ranges for camera guided vehicles. These measurements are made possible by the relative image offsets, or parallax, produced when objects at different ranges are imaged from different angles. In other image comparison applications, such as change detection, parallax may constitute a significant nuisance, producing undesired relative image distortions. Parallax removal through pixel by pixel image matching is then necessary before image comparison can be performed. In this study fractional pixel parallax determination at each pixel location is attempted using image cross correlation calculations input to a neural network. Correlation values are obtained between image windows in the left view with a succession of overlay window positions in the right view. High resolution correlation requires small image windows. Correlation peak locations for the small windows are often unreliable match points due to noise and relative parallax distortion. Further processing of the correlation data is necessary to reduce match point errors. A section of correlation data is input to a neural network. The network outputs the parallax offset value for the pixel centered on the section. The network was trained using simulated stereo imagery so that the exact parallax offset at each pixel was known. A network with two "hidden layers", and with symmetries imposed on the cell connection weights, was trained using the back-propagation method. Network results on simulated test sets show a distinct improvement over results using correlation smoothing methods. The trained network was then used on a real image pair for elevation extraction and change detection. The resulting elevation surface and change detection difference image are illustrated and evaluated.
|
