Duration of the Thesis: 6 months
Completion: December 2014
Tutor: Dr.-Ing. Michael Cramer
Tutor & Examiner: Prof. Dr.-Ing. Dieter Fritsch
Semi-Global Matching (SGM) algorithm is developed by Hirschmüller in 2008, which can produce very dence Digital Surface Model (DSM) efficient and robust. The algorithm matches stereo image pairs pixelwise and makes the DSM at the same ground sample distance (GSD) as original images (Hirschmüller,2008). SGM algorithm boosted the evolution of Photogrammetry. SGM method has already been applied to generate aerial photogrammetry bench marks of European Spatial Data Research Organization (EuroSDR) in 2013 (Haala, 2013). In this thesis, we will build a pipeline to implement the dense image matching method on satellite imagery and evaluate its results.
This Master thesis is mainly completed by Matlab codes. Figure 1 is the workflow chart of the pipe implementation. The whole pipeline can be divided into four work packages (WP).
|
| Figure 1. Workflow |
WP 1: The orientation of original images. Bias-compensated Rational Polynomial coefficients (RPCs) bundle block adjustment is applied. Usually, satellite imagery provides RPCs instead of in- and exterior elements for orientation. RPCs are 80 coefficients which directly build a pure mathematic model between object and image space (Grodecki, et al., 2001). And RPCs will be adjusted by additional bias-compensation model to implement the orientation.
WP 2: Image rectification. According to the property of pushbroom satellite sensors and the RPC model, the project–trajectory-based epipolarity (PTE) model is chosen to generate the epipolar image pairs (Kim, 2000). This model will project image points to object space and back-project to second image space by RPC model and then find the corresponding epipolar line.
WP 3: Dense image matching. LibTsgm is used to produce the disparity image. LibTsgm is the key library of software SURE, which implements a modified SGM method. LibTsgm not only uses the disparities of high level image pyramid as the initial disparity images of next pyramid level, but also applies the disparities to limit the disparity search range for the matching of subsequent pyramid level (Fritsch, D., et al. 2013).
WP 4: Triangulation. Forward intersection is utilized for getting very dense point cloud. The point cloud is transfer to DSM by software SURE.
The pipeline is tested on two different satellite imagery. QuickBird (QB) imagery at 0.7m GSD is acquired from the test data package of software Barista. WorldView-2 (WV-2) data is provided by Deutsches Zentrum für Luft und Raumfahrt (DLR), Oberpfaffenhofen.
Figure 2 shows the orientation result of the check points. All the Root-Mean-Square (RMS) errors in image space are less than 0.4 pixel. The RMS errors of check points in object space are almost sub-pixel level except the elevation error of WorlView-2 data (ca. 1.3m).
 |
 |
| Figure 2. RMS error of check points in orientation |
Figure 3 and Figure 4 display the vertical parallax of corresponding points on both datasets. It is obviously that all the corresponding points’ vertical parallaxes are subpixel level and the RMS is ca 0.5 pixel.
 |
| Figure 3. Vertical parallax of corresponding points on QB data |
 |
| Figure 4. Vertical parallax of corresponding points on WV-2 data |
Figure 5 presents the dense point cloud of QB data and WV-2’s point cloud is exhibited in Figure 6. The visualization is conduct by software CloudCompare. The point clouds have rebuilt the terrain with clear details but also have some noise.
 |
| Figure 5. Vertical parallax of corresponding points on WV-2 data |
 |
| Figure 6. WV-2s point cloud |
Three profiles are extracted from the WV-2 data’s DSM for further evaluation. Figure 7(a) is the chosen profiles in the orthophoto. Figure 7(b) and (c) are the DSMs generated by aerial photography and satellite imagery. Figure 7(d), (e) and (f) exhibit the three profiles comparison between two DSMs. The red line is satellite data’s DSM and blue one is the reference. The DSM produced from satellite imagery do describe the shape of the buildings but still have some elevation shift.
 |
 |
 |
| (a) | (b) | (c) |
 |
 |
 |
| (d) | (e) | (f) |
| Figure 7. DSM Profiles Analysis | ||
I. RPCs bundle block adjustment with affine model is reliable, images can be oriented. In object space the accuracy can reach sub-meter level and in image space, the accuracy can reach sub-pixel level.
II. Epipolar image generation based on projection trajectory can reach sub-pixel level accuracy.
III. After LibSURE’ s dense image matching, DSM can be generated. Although there still exist elevation shifts and noise, most basic surface information can be seen on the model.
IV. The pipeline is robust and it can be used in different datasets.
Fritsch, D., Becker, S., Rothermel, M., 2013. Modeling Facade Structures Using Point Clouds from Dense Image Matching. Proceedings Intl. Conf. Advances in Civil, Structural and Mechanical Engineering, Inst. Reserach Eng. and Doctors, pp. 57-64.
Grodecki, J., Dial, G., 2001. IKONOS geometric accuracy, Proceedings of Joint International Workshop on High Resolution Mapping from Space, Hannover, Germany, pp. 77-86 (CD-ROM).
Haala, N., 2013. The Landscape of Dense Image Matching Algorithms. IN: Photogrammetric Week’13, Ed. D. Fritsch, Wichmann/VDE Verlag, Berlin/Offenbach, pp. 271-284.
Hirschmuller, H., 2008. Stereo processing by semiglobal matching and mutual information. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 30(2), pp. 328-341.
Kim, T., 2000. A study on the epipolarity of linear pushbroom images. Photogrammetric Engineering and Remote Sensing, 62(8), pp. 961-966.