期刊名称：ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
印刷版ISSN：2194-9042
电子版ISSN：2194-9050
出版年度：2006
卷号：XXXVI Part 5
出版社：Copernicus Publications
摘要：Commercial software of digital photography, realizing cylindrical or spherical panoramas, are becoming popular. They are delivered for tourist and documentary use. For instance they are suitable for quick documentation of field excavations in archaeology. In fact their principal application consists in the realization of active explorations known as QTVR (Quick Time Virtual Reality). It has already been proved that these panoramic images also have a metric use (Luhmann, 4, 2004), (Szelisky,Kang, 15,2001). The 3D final reconstruction of object is performed by bundle adjustment of multi-station panorama. Normally rotating cameras are used instead of mosaics (Schneider, Mass, 2204). The advantage of the stitching software consists in its economy compared to the rotating cameras. The analogy between surveying and photogrammetry is in the case of the spherical panoramas almost perfect. In fact the panoramic photos are produced for projection on a sphere of the photographs having as centre of common projection the centre of the sphere (Szelisky,12,1997). Then the sphere is mapped in the image plane by the so-called longitude-latitude projection. The image points can be regarded as analogical recording of the angular observations of a theodolite having its centre in the centre of the sphere. The spherical panorama can have a field of view up to 360°x360°. They could be the ideal "theodolite". Nevertheless the camera cannot be set correctly as a theodolite. It is necessary therefore to recover two angles to set up vertical the axis of the sphere as the bi-axial compensator in a theodolite. The estimate of the angular corrections is done by means of known directions or known coordinates of points (control directions and control points), obtained by traditional theodolite, or finally with geometric constraints as horizontality or verticality of straight lines. In order to evaluate the effects of a non perfect verticality of the principal axis of the spherical panorama, a computer programme has been written, following the steps: 1) creation of a set of points laying in the unit sphere regularly spaced along meridians and parallels; 2) projection of the points in the "cartographic" plane by the latitude-longitude projection; 3) rotation of the sphere alternatively about x and y axes of small rotation angles; and monitor the shifts of the projected points in the cartographic plane; 4) back estimation of the rotation angles; 5) angular correction. However the angular corrections are not still sufficient to guarantee a reasonable accuracy in the final 3D object compilation. The formation of the mosaic of photos doesn't happen without noises. The errors have nevertheless an evident systematic behaviour and they can be filtered out with interpolation polynomials whose parameters are estimated in correspondence of control points. For this reason the network of control points has to bee quite dense. When the terrain coordinates of the panorama centre are known, the correct image position for the control points is known, that we can compare with the actual position in the image, knowing the correction vector for any control point. Therefore the correction of the observed points in the panoramic image takes place in two steps: correction for rotation, with the estimated correction angles, then further correction computed by interpolation with the corrections estimated in the nearest control points giving a Gaussian weight to the control points. We present and comment some experiences of spherical panoramas produced with the software Stitcher 4 ., by Realviz. The lens distortion is already corrected by the mosaicing software itself. But the main problem still consists in the noise occurring during the formation of the mosaic. There are different causes for the noise, the moving clouds in the sky, the persons and the traffic moving in the scene, the non perfect interior orientation parameters of the camera, the camera projection point off set from the rotating axis. The discussed examples are the panoramas taken 1) in Ancona the university campus, 2) Piazza del Popolo in Ascoli town, 3) Piazza del Campo in Siena. The used camera was an amatorial 35mm equivalent digital camera of 3 mb resolution. The panoramas have resolution of 10000x5000 pixel. Any pixel corresponds to 0.04 g, which is not a very high accuracy. The results are encouraging as far as control points is concerned. For example, in Piazza del Campo, a valid test area, having dimensions ranging from 100 to 150m in plan and 100 in height (the municipality tower), we took four panoramas, and with a reflectorless theodolite we surveyed 135 control points. The RMS of the residuals are 0.027 in planimetry and 0.009 m in height over 108 control points, observed at least in three panoramas whilst for the plotted points the results are not so good, the RMS of sigma naught are 0.16 m in planimetry and 0.05 m in altimetry for 358 points over a total amount of 385, and we had to discard the remaining 27. Similar results we got for the other test fields. So far the results are only partly satisfactory. There are still improvements to be performed: improve the resolution of the panorama, improve the quality of the stitching algorithms, improve the efficiency of the interpolation procedure
