Fig. 1. Working principle of five camera oblique photography device. (a) Five images captured in 3D object space; (b) plane-directional distribution of five images captured on an exposure station
Fig. 2. There is a large overlap between the images on five different exposure stations. (a) Plane-directional distribution of the five exposure stations; (b) an example of local map from ISRPS EuroSDR dataset
Fig. 3. Principle structure of the LGO method
Fig. 4. Combining all local maps into a global map by a least-square optimization
Fig. 5. Spatial distribution of the exposure stations over the observed terrain
Fig. 6. Trajectories of the nadir cameras of the exposure stations
Fig. 7. Change of MSE with the number of iterations. (a) MSE of the BA method; (b) MSE of the LGO method
Fig. 8. Residual of all estimated camera positions
Fig. 9. Distribution of the 135 Zurich images and the construction of local maps
Fig. 10. Trajectories for the Zurich data. (a) Trajectory of only nadir images; (b) Trajectory of both nadir and side images
Fig. 11. Change of MSE with the number of iterations on the Zurich dataset. (a) BA method; (b) LGO method
Fig. 12. Residual of 135 Zurich images estimated by the BA and LGO methods. (a) X direction; (b) Y direction; (c) Z direction; (d) 3D
Parameter | Value |
---|
Area/(km×km) | 60×7 | Number of track points | 10 | Number of images | 5000 | Number of local maps | 1000 | Number of 3D points | 54337 | Number of projection points | 490086 |
|
Table 1. Parameters of the large scale simulated dataset
Parameter | BA | LGO |
---|
X | 9.6563 | 9.6516 | Y | 5.0619 | 5.0623 | Z | 0.8362 | 0.8568 | 3D | 6.3131 | 6.3117 |
|
Table 2. RMSE of camera positions estimated by the BA and the LGO methodsunit: m
Condition | BA method | LGO method |
---|
Initial MSE | Final MSE | Number ofiterations | Initial MSE | Final MSE | Number ofiterations |
---|
XYZ+5 m | 1.922×103 | 0.012 | 4 | 2.352×103 | 0.004 | 3 | XYZ+50 m | 1.998×105 | 0.012 | 5 | 2.425×105 | 0.004 | 5 | XYZ+100 m | 7.930×105 | 0.012 | 6 | 1.288×106 | 0.004 | 6 | XYZ+200 m | 3.594×106 | Singular | | 5.828×106 | 0.004 | 9 | XYZ+300 m | 2.375×1013 | Singular | | 3.188×1013 | Singular | | Ang+0.1 rad | 8.903×105 | Singular | | 9.805×105 | 0.004 | 5 | Ang+0.2 rad | 3.823×106 | Singular | | 4.013×106 | 8.343 | 22 | Ang+0.25 rad | 1.019×107 | Singular | | 6.496×106 | Singular | | Ang+0.3 rad | 2.022×107 | Singular | | 9.538×106 | Singular | | Ang+0.4 rad | 1.803×1012 | Singular | | 1.677×107 | Singular | |
|
Table 3. Results of BA and LGO methods on the large scale simulated dataset with initialization noise
Parameter | Content |
---|
Camera type | Leica RCD30 Oblique Penta | Image size /(pixel×pixel) | 9000×6732 | Focal length /mm | 53 | Pixel size /mm | 0.006 | Platform height /m | 1000 | Title angles /(°) | 35 | Along-track overlap /% | 70 | Across-track overlap /% | 50 | Ground sample distance(GSD) /cm | 6--12 | Number of images | 135 | Number of 3D points | 51672 | Number of projection points | 225952 |
|
Table 4. Overall parameters of the used Zurich dataset
Indicator | BA | LGO |
---|
Initial MSE | 2.416×106 | 1.838×108 | Final MSE | 0.989 | 0.282 | Number of iterations | 8 | 4 | RMSE along X-axis /m | 0.1896 | 0.1817 | RMSE along Y-axis /m | 0.1704 | 0.1752 | RMSE along Z-axis /m | 0.1527 | 0.1535 | RMSE in 3D space /m | 0.1715 | 0.1705 |
|
Table 5. Parameters of the BA method and the proposed method for Zurich data
Condition | BA method | LGO method |
---|
Initial MSE | Final MSE | Number ofiterations | Initial MSE | Final MSE | Number ofiterations |
---|
XYZ+5 m | 4.337×103 | 0.989 | 5 | 4.383×103 | 0.380 | 3 | XYZ+50 m | 4.879×105 | 0.989 | 6 | 4.818×105 | 0.380 | 3 | XYZ+100 m | 1.352×106 | 0.989 | 8 | 1.288×106 | 0.380 | 3 | XYZ+200 m | 6.971×106 | Singular | | 5.498×106 | 0.380 | 3 | XYZ+300 m | 3.687×1012 | Singular | | 1.677×107 | 0.380 | 3 | Ang+0.1 rad | 4.119×105 | 0.989 | 6 | 4.003×105 | 0.380 | 5 | Ang+0.2 rad | 1.897×106 | 0.989 | 7 | 1.675×106 | 0.380 | 6 | Ang+0.25 rad | 3.260×106 | 0.989 | 7 | 2.593×106 | 0.380 | 7 | Ang+0.3 rad | 4.419×106 | 0.989 | 8 | 3.349×106 | 0.380 | 7 | Ang+0.4 rad | 8.364 ×108 | Singular | | 6.680×106 | 0.380 | 8 |
|
Table 6. Results of BA and LGO methods on the Zurich dataset with initialization noise