Huijing Zhao
[Laser Range Finder (LEF)| CCD Camer
Honzontal, |
Rotation
2
- * Vertical
Senal Port 1 Rotation
Capture Board
MÀ Serial Port 2
Frogrammable Rotatme Platiônn
Figure 1: Architecture of sensor system.
transformations aligning all range images to a global coordinate system is obtained, where the accumulation of
estimation error in pair-wise registration has to be solved.
A picture of the sensor system used in the research is shown in Figure 1. We assume the sensor's platform is set
always horizontal to the ground surface, which is easy to be satisfied in urban survey. A registration with four
degree of freedoms is studied in this research, where a horizontal rotation angle and three translation parameters
from each range frame to a global coordinate system has to be estimated. A pair-wise registration method using
"Z-image" has been developed in our previous study (Zhao and Shibasaki, 1999). In this research, we assume
all the relative transformation parameters between neighboring range image pairs (two range images with a
degree of overlay) have been obtained through the pair-wise registration. This paper contributes to a multiple
registration method, where all range images are aligned to a global coordinate system, and simultaneously
registered to achieve a well-balanced model. An experiment is conducted, 42 range images measured in the
campus of the Univ. of Tokyo are registered. A sequence of examination is conducted to test the accuracy of
the method. Two sets of ground truth are used to examine the registration accuracy. They are 1) a 1/500 scale
digital map and, 2) location of viewpoints measured by GPS with an accuracy of + 20cm.
2 REGISTRATION OF MULTIPLE OVERLAPPING RANGE IMAGES
n2
1 Study review of multiple registration methods
À number of different methods have been developed to solve the error accumulation problem in multiple regis-
tration. Y.Chen and G.Medioni, 1992 partially solved the problem by registering the newly introduced range
image with the integrated model consisting of all previously registered range images. Some of the researches
simultaneously registered all range images using weight least square methods. R.Bergevin et al. 1996 equally
distributed those accumulated registration errors by minimizing the squared sum of distance from control points
to the corresponding tangent planes in other range frames. G.Blais and M.D.Levine 1995 evaluated the dis-
placements between range images by the distance from a set of control points to their corresponding ones in
other frames. However disadvantage of the methods is not only the accuracy of extracted control points might
be significantly influenced by range error and range uncertainties, but also searching for global optimum can
be very computational heavy. Shum et al. 1994 formulated the multiple registration as a problem of principal
F.
1 ; Ï
> 12.5 9 w. is 7 P, : Range image j
2 8 7
71 Via y!
t o 7 Ey l1 :Relative V uitio) y; + 49, s
23 ES 87 Ÿ transformation x ar, sd | ill 7 Vic
A o> <O E, from the SR 2 A d. *; ,
t * » 4 i coordinate system 2 /n, A Voix ; sh F,
34 So 76 ofrange image jto ÖL J, / SR M
a . range image 1 ; 7.
45 E 65 , y x
5 (a) : (b)
Figure 2: An example of range measurement from Figure 3: Evaluating the violation in shift vector and
each facade of the building. rotation matrix.
1034 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
b c c c Cj
e
Nd fd Lj] Lj M et 3
gn "r4 OY A
135 "*
~~ ch er HA ed
E: 9 "^ . À
rn