inspection of Figure 9 indicates that the results were very
close to the reference data. The main reason for misidentifying
some rooftops is the lack of disparity information at building
edges due to occlusion (i.e., matching failure). Pixels in this
category as well as misidentified pixels within and around the
object were not included in the true positive count notwith-
standing that the rooftop was correctly detected as a whole.
Accordingly, the completeness measure decreased and quanti-
tatively pointed out differently than the visual inspection.
On the other hand, almost one-fifth of the detection result
was missed for the results generated based on the convention-
al interpolation and normalization techniques. Interpolating
the gaps of the resulting disparity map in dense urban areas
leads to smoothing the sharp building edges and missing the
narrow streets between buildings. The consequence is that
many terrain-level objects will be elevated falsely as build-
ings. Thus, many non-building objects were detected and
classified as buildings as indicated by the high value of the
completeness measure in Table 3. The effect of the interpola-
tion on the detection result is illustrated in the upper row of
Figure 12, where three high-rise and separated building roofs
were mistakenly classified as one building object.
Data normalization by subtracting the extracted terrain
disparity from the original surface disparity map amplified
the effects of surface interpolation and filtration by local-min-
ima technique. The result was an inaccurate
TDM
that led to
missing actual building objects. The effect of normalizing the
disparity map on the detection outcome is illustrated in the
lower row of Figure 12. As shown in this figure, a few gaps
on a street between two high-rise buildings were falsely filled
and elevated by the interpolation process. Thus the local-min-
ima technique extracted incorrect ground-level information
and produced a highly elevated terrain area. This area was
erroneously removed in the subtraction process and, conse-
quently, no building roofs detected in this area of the
nSDM
.
The evaluation measure that combines both the correctness
and completeness measures is the overall quality (
OQ
) of the
detection. The
RMAD
-based detection overall quality, of 75 per-
cent as in Table 3, is affected directly by the missed roofs from
the detection. However, the achieved quality is still promising
for building roof detection using only disparity information.
Moreover, there are no building façades detected in the results
due to having
VHR
stereo images of opposite backward and for-
ward viewing angles as discussed in the Methodology Section.
In contrast, the overall quality of the detection performance
achieved based on the conventional interpolation and normal-
ization techniques was only 43 percent. This low value was af-
fected directly by the high false completeness and low correct-
ness values. This low value proves the severe negative effects
of the interpolation process in dense urban areas. In contrast,
the developed
RMAD
technique provided an efficient solution
for such dense areas. The degree of the improvement in the
tested area in this study was 32 percent. Obviously this value
depends mainly on the nature of the urban area. However, it
shows an attractive degree of improvement for the proposed
disparity normalization technique based on co-registration.
Conclusions
This paper developed a registration-based technique for dis-
parity mapping of aboveground objects (
RMAD
) based on image
rectification, epipolar-pair generation, and co-registration
methods. The aim is to mitigate the impact of occlusion on
surface representation. The developed technique achieved
this goal by eliminating the need for interpolating the gaps in
surface elevations and producing a direct approximation for
the off-terrain disparity information. The key novel concept in
the developed technique is the co-registration of terrain-level
objects in both stereo
VHR
images. Consequently, the disparity
measured for ground objects in the epipolar direction becomes
very close to zero value. This directly approximates the nor-
malized disparity information which could be easily incorpo-
rated in disparity-based building detection applications.
The implementation and testing of the approach proved its
effectiveness in dense urban areas characterized by reason-
able terrain variation. The terrain relief distortions are dra-
matically reduced after rectifying the images using the area
average elevation. This allows acceptable co-registration for
the terrain-level objects (e.g., roads).
When the generated disparity map was tested in a dispar-
ity-based building detection application, the overall quality
of the achieved result was 75 percent. The correctness of the
detection was almost 100 percent, which emphasizes the reli-
ability of the disparity information in detecting buildings. The
buildings were detected by applying a close-to-zero threshold.
This value further indicates the success of the proposed co-
registration method in approximating the normalized surface
disparity. There were no building façades wrongly detected,
taking advantage of the opposite viewing angles of the
VHR
line sensors. Additionally, the
RMAD
-based detection of build-
ing rooftops requires only one stereo-pair of
VHR
satellite
images. This is advantageous as it reduces the cost of having
multiple overlapped images to a minimum.
The developed
RMAD
technique is evaluated against the
disparity-based building detection result derived using pub-
lished techniques for epipolar generation (
RFM
-epipolarity
technique) and disparity map normalization (based on the lo-
cal-minima technique). The achieved detection result by these
techniques was substantially lower than that using the
RMAD
technique. The main reason for that is the severe negative ef-
fect of interpolating over the occluded areas in off-nadir stereo
pairs. Such a situation produces misleading information that
negatively impacts the quality of the disparity maps and all
subsequent processes including the building detection.
All these attractive features and improvements prove the use-
fulness of the developed technique for disparity-based building
detection in dense urban areas of reasonable terrain level varia-
tions. Additionally, it increases the usability of the off-nadir
stereo images in urban mapping and modeling applications.
The future work of this research will investigate overcom-
ing probable limitations to the developed method. The main
one is expected to appear when the imaged area is of a hilly
terrain. Defining the projection reference plane by the average
terrain elevation for the area may not be enough. In this situa-
tion, and for scenes larger than our test image, one may adopt
a tile-based approach for implementing
RMAD
method to co-
register terrain level features. Another challenge to the
RMAD
method is its use of
NDVI
for removing trees and vegetation.
Acknowledgments
This research is funded in part by the Libyan Ministry of
Higher Education and Research (
LMHEAR
) and in part by the
Canada Chair Research (
CRC
) program. The authors would
like to thank Airbus Defence and Space plc, for providing
the research data. The authors greatly appreciate the valuable
comments of the anonymous reviewers and the editor that
helped to improve the manuscript.
References
AASHTO, 2001.
A Policy on Geometric Design of Highways and
Streets 2001
, The Association, Washington, D.C., 242 p.
Alobeid, A., K. Jacobsen, and C. Heipke, 2010. Comparison of
matching algorithms for DSM generation in urban areas from
Ikonos imagery,
Photogrammetric Engineering & Remote
Sensing
, 76(9):1041–1050.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
July 2016
545
