|
Adaptive Multi-Image Matching Algorithm for the Airborne Digital Sensor ADS40
Quality control
During run time, quality measures are calculated and used for elimination of blunders and false matches. The similarity measure, the 2nd best similarity score and its distance to 1st one, the size of search window, the change of similarity measure between the 3 patch sizes, the change of position between patch sizes, the angle of dominant edge direction with the epipolar line, the residuals from 3D forward intersection, the change of final point position from the starting position of the largest patch, the approximate standard deviations for x and y pixel coordinates are the most important quality criteria calculated. The quality criteria are combined according to possible occurring errors. E.g. in case of an occlusion, the cross-correlation coefficient would be small and the change of the similarity measure would be generally decreasing from the largest to the smaller patch. Thresholds for the quality criteria are derived by a statistical analysis of their values in the given pyramid level. We will investigate a fuzzy approach for threshold calculation, and relative weights when combining together multiple quality criteria to derive a single quality measure. All quality criteria are used for each image ray individually, to detect problems occurring in individual images, e.g. occlusions. Weak rays are excluded and final 3D intersection is computed using only the good rays. Regarding multiple solutions, along the epipolar lines, the multi-patch approach may not detect such problems. Thus, additionally the consistency of height in the local neighbourhood is checked to detect and eliminate spike errors by a median-type filter. Least squares matching were used after blunder elimination with the above quality criteria. Successful matches are rechecked with harder quality criteria resulting from the least squares approach, such as number of iterations and change of shape parameters of the patch. Results The matching algorithm has been used for DSM generation in the previously mentioned datasets. Out of these datasets, the Level 1 PAN images were processed and DSM’s have been extracted. Figure 3 shows a portion of the results for the Waldkirch data for the 0th pyramid level and matching without constraints using doublets. The part of the area that has been used for DSM extraction was 2000 x 1000 pixels. The number of points that have been matched in the lower level of the last doublet is about 165,000 and after the quality control 130,000. The percentage of successful matches was around 80%. The irregularly distributed match points have been used to interpolate a regular grid with 0.5 m spacing, which was hill-shaded. The match points were also used to interpolate contours with an interval of 2 m. The results were controlled visually by overlaying contours on stereo images using SOCET SET and by inspecting individual points. The major problems encountered were with repetitive fine texture (multiple solutions), e.g. in agricultural fields. We have applied the algorithm also on Level 0 images where matching can be more problematic, due to the perturbations of the aircraft and the non perfect camera stabilization that influence the produced raw images. As the results were good, the developed algorithm can be also used in aerial triangulation for extraction of tie points or determination of a coarse DSM (latter can provide better approximations for tie point area selection and tie point matching). In this case constraints are less strict than in Level 1 images since the orientation data have not been adjusted with aerial triangulation yet, and the GPS/INS data are not perfect. In addition, the quality control procedure can be stricter than the one used when matching Level 1 images, resulting in less, but however sufficient in number and distribution, matched points. 
Figure 3. Portion of test area in Waldkirch (CH). Left: Level 1 nadir channel overlaid with contours with 2 m interval. Right top: zoomed region from the same area. Right bottom: DSM image Conclusions and Future Work ADS40 is a new camera system, which is still being fine-tuned. Thus, there are not so many data available, and even less data in regions with different land cover and relief and accurate reference DSMs for testing our matching methods. This was one of the main obstacles in our previous work and we hope that the next period significant improvements will be done in this direction. Thus, it is planned to quantitatively compare various matching methods and side aspect variants. Furthermore, the quality control procedure, probably the most critical aspect of matching, will be further developed and tested. Further work will also include reduction of DSM to DTM using also NIR information for the elimination of trees. Acknowledgements This work was performed within the project AIM, a cooperation of ETH Zurich and LH Systems, with the financial support of the Commission for Technology and Innovation of the Swiss Government. References
Page 4 of 4
| Previous | |