The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
(a)Ratio of correct matches (b) Distance from points to scanner
Figure 7: Compare the matching results with and without
geometric constraint.
The resulting position and orientation errors for our method are
shown in Table 4. In order to evaluate the registration results,
we compared the planar patch approach (Dold and Brenner,
2006) with our proposed method using the reference orientation.
In (Brenner et al., 2008), the orientation by planar patches was
able to align SP] with SP 10 which have an overlap of 16%. In
contrast, it seems that the proposed method is more influenced
by scene contents. The proposed method fails after SP 5a , which
has a considerably larger overlap of 50%. However, the
distance between SP[ and SP 5a is 20 m, which is probably
anyhow meet the distance between two scans one would prefer
to obtain a dense city scan with few occlusions.
Pair
Total
matches
Right
matches
Right
ratio
(%)
Total
matches
Right
matches
Right
ratio
(%)
01-02
301
116
38.5
151
116
76.8
01-03
130
30
23.1
45
30
73.2
01-03a
132
32
24.2
49
32
65.3
01-04
127
16
12.6
35
16
45.7
01-05
113
10
8.8
25
10
40.0
01-05a
111
10
9.0
28
10
35.7
01-06
107
2
1.9
21
2
9.5
02-03
437
172
39.4
216
172
79.6
03-03a
1714
1438
83.9
1597
1438
90.0
03-04
328
133
34.5
162
133
82.3
04-05
609
309
50.7
342
309
90.4
05-05a
1476
1220
82.7
1352
1220
90.2
05-06
235
194
82.6
209
194
92.8
On the other hand, the proposed registration method has the
advantage of being algorithmically simple and does not rely on
the presence of planar structures. In addition, the proposed
method is conceptually simpler and faster. Therefore, if it is to
be preferred depends strongly on the application. As far as
accuracy of proposed method is concerned, one can see that the
maximum deviation from the reference is less than 10cm in
translation, and 0.2° in orientation. This is better than the planar
patch method which achieved a maximum deviation of less than
20cm in translation, and 0.5° in rotation angles.
Considering the required computation time, SIFT feature based
matching took an average of 20s. In the case of computing
Gaussian and mean curvature, it is quite time consuming to
generate the full scene of curvature map. In practice, we only
compute the Gaussian and mean curvature in the neighborhood
of matched points. This is very quick and takes around 5s. Then
a standard RANSAC will take another 5s. Thus, in total,
approximately 30s were required on average to match two scans
on a 2GHz Pentium laptop.
6. CONCLUSIONS AND FUTURE WORK
We have shown a key point based automatic method using
intensity and geometry features for the marker-free registration
of terrestrial laser scans. The method uses SIFT feature based
key points extracted from the normalized reflectance image
with geometric constraint. Results and the analysis show the
proposed method’s efficiency and robustness. The method can
be used to register laser scanning data at accuracy comparable
to that of manual registration using natural tie points.
For the future work, several issues are worth investigating.
Our approach applies the SIFT method to extract feature
points. Furthermore, other feature points, such as comer
points, can be detected as geometric primitives by using
Harris or SUSAN operators. In addition to using a geometric
constraint, other primitives can probably be used for the
prioritization of the correspondences. And finally, the
proposed algorithm offers a pair-wise registration scheme. It
can be extended into a multi-scan registration.
Table 3: Comparison of matching results with or without
geometric constraint.
Pair
A coO
A^SO
>
o
AX(m)
AY(m)
AZ(m)
01-02
-0.031
-0.001
0.022
-0.001
-0.017
0.010
01-03
0.051
-0.082
-0.078
0.018
0.021
0.038
01-03a
-0.075
-0.013
-0.032
0.032
-0.048
0.051
01-04
0.037
0.079
0.021
-0.053
-0.011
-0.025
01-05
-0.028
-0.104
0.115
-0.056
0.051
0.047
01-05a
-0.108
0.126
0.022
0.091
-0.039
0.013
01-06
—
—
—
—
—
—
02-03
-0.018
0.011
-0.009
-0.031
0.019
-0.017
03-03a
0.033
-0.009
-0.026
-0.027
-0.021
0.009
03-04
-0.031
0.018
-0.013
0.017
-0.018
0.007
04-05
0.042
-0.020
-0.016
-0.033
0.032
0.012
05-05a
-0.023
0.019
-0.021
-0.025
-0.016
-0.014
05-06
0.029
-0.014
-0.011
0.016
-0.022
0.019
Table 4: Deviation of the translation and rotation parameters
from the reference values for the registration based on
proposed method
ACKNOWLEDGEMENTS
Claus Brenner has been funded by the VolkswagenStiftung,
Germany. Zhi Wang has been funded by China Scholarship
Council.
REFERENCES
Bae, K. and Lichti, D., 2004. Automated registration of
unorganised point clouds from terrestrial laser scanners.
International Archives of Photogrammetry and Remote
Sensing 35 (Part B5), pp. 222-227.
Bamea, S. and Filin, S., 2008. Keypoint based autonomous
registration of terrestrial laser point-clouds. ISPRS Journal of
Photogrammetry & Remote Sensing, Theme Issue: Terrestrial
Laser Scanning 63, pp. 19-35.
Bendels, G. H., Degener, P., Wahl, R., Kortgen, M. and Klein,
R., 2004. Image-based registration of 3d-range data using