The results of a statistical evaluation are shown in Table 2.
In this experiment, all statistical values indicated the superi-
ority of
SATM
+ relative to
SGM
. However, the relatively lower
resolution (0.5 m/pixel) of the satellite images yielded point
densities that were inferior to the lidar data.
Table 2. Statistical analysis of the experiment in Central
District, Hong Kong.
RMSE (m)
STD
(m)
Density (pts/m
2
)
(4 pts/m
2
for LiDAR data)
SGM 6.809
6.015
2.1
SATM+ 5.509
5.417
2.2
Visually, the
DSM
generated by
SATM
+ may not be as ap-
pealing as the lidar
DSM
, as shown in Figure 14. Notably,
however, the
DSM
generated by
SATM
+ only used two images,
and only the objects visible in both the two images could be
reconstructed. If multiple image views are available, the pro-
posed method could be used to generate 3D surface models
with full structures.
Conclusions and Discussion
The approaches presented in this paper address 3D surface
reconstruction in urban areas, the most complicated challenge
in the field of photogrammetry. The theoretical analysis and
experimental validation presented in this paper demonstrate
the performance of the developed approaches and yield the
following conclusions:
1. The integrated image matching and segmentation ap-
proach helps to alleviate matching ambiguities in urban
images, which are caused by the existence of buildings.
Here, occlusion filtering can eliminate the mismatches
caused by occlusions and improve the matching reliability.
The segment-adaptive similarity measurement can com-
pensate for defects caused by discontinuous disparities.
2. The proposed dense matching strategies, including local
and regional propagation, are used to propagate matching
from the robust feature matches to other regions of the im-
ages. These dense matching strategies enable an evalua-
tion of all pixels for possible matches and the isolation of
regions without any matches.
3. According to the experimental analysis, the proposed
SATM
+ performs better than a popular matching method
(i.e.,
SGM
) in terms of the preservation of feature boundar-
ies and recovery of tall buildings. These advantages are
largely attributable to feature-based matching and the
integrated matching and segmentation. Our results dem-
onstrate that
SATM
+ is able to generate 3D point clouds
comparable to or superior to those by lidar.
The research and developments presented in this paper offer
an alternative option for 3D surface reconstruction in urban
areas. Our developed approaches appear to be significant
for urban 3D reconstruction and modelling applications.
It should be noted that, this paper focuses on stereo image
matching for 3D surface reconstruction. Using multiple im-
ages (three or more) has the potential to further improve the
matching performances by taking advantage of the redundant
information, which will be our future efforts. Future works
will also include improvements to image segmentation, opti-
mization of the algorithms, and reductions in processing time.
Acknowledgments
This work was supported by grants from the Research Grants
Council of Hong Kong (Project No. PolyU 5330/12E and Proj-
ect No. PolyU 152086/15E), and a grant of strategic impor-
tance from the Hong Kong Polytechnic University (Project No.
1-ZE24).
Figure 13. A Pleiades-1 image of Hong Kong, with the study area marked by a rectangle.
146
March 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
