Automatic Relative Orientation of Aerial Images

An approaclr is described for automatic relative orientation of a stereopair of digital aerial images. The concept and the implementation are based on practical conditions with respect to available a priori knowledge, speed of computation, and obtainable accuracy. Feature-based image matching using point features extracted with a modified version of the Moravec operator and a coarse-to-fine strategy ure incorporated into the approach. In higher image pyramid levels, where images are small in size and of low resolution, the entire model area is searched for interest points. In lower levels, window tracking is carried out in order lo speed up the entire procedure and to stabilize the final results. In all levels, matching is based on geometric as well as radionrelric constraints. The approach was developed as one of the automation-oriented software components of a digital photogram~netric workstation. Results obtained from ten aerial image pairs with scales ranging from 1:3,000 to 1:34,000 and scanned with a pixel size of 15 pm, thus yielding some 235 Megabytes per image, are presented. In each case, more than 150 well distributed points were extracted. The obtained root-mean-square standard deviations of the image coordinutes consistently lie between 3.2 and 3.6 pm or 0.21 and 0.24 pixels. A human operator checked the resulting models on on analytical plotter. The models were found to be free of y-parallax. The elapsed computing time was approxi~nately 4 minutes per image pair on a Silicon Graphics Iris Indigo workstation with R4000 processor. This means that the procedure runs as fast as, if not faster than, a human operutor can carry out the relative orientation while yielding the same level of accuracy. Thus, it could be shown that tlre presenled method for automatic relative orientation is operational for practical applications.