LASER RANGE DATA REGISTRATION USING SPIN IMAGES

Xavier Mateo Prous

Department of Computer Science, Universitat Aut`onoma de Barcelona, Bellaterra, Spain

Xavier Binefa Valls

Department of Information Technologies and Communications, Universitat Pompeu Fabra, Barcelona, Spain

Keywords:

Spin images, Iterative closest point, Laser range scanner, Registration.

Abstract:

Registration of laser range data becoming from different scanner positions is still a current topic in literature.

In this paper we introduce the possibility of solving it by using spin images, which create a 2D image for

every 3D coordinate vertex in the scans. Matching between spin images allows the estimation of an initial

rigid transformation between the scans, which later can be reﬁned with ICP process in order to achieve a more

accurate registration.

1 INTRODUCTION

Laser range scanning has become a quite popular sys-

tem for the capture of 3D environments. The possibil-

ity of combining both the 3D representation obtained

from the laser and a visible/infrared camera to apply a

texture to the three-dimensional model achieves a pre-

cise representation of huge scenes with a minimum

effort.

One of the most encouraging challenges in this

kind of images is the possibility of register differ-

ent scans becoming from different unknown points of

view, achieving in this way a more wide representa-

tion and the ﬁlling of laser range shadows produced

by the reﬂection in the objects.

The usual way to achieve this pairwise registra-

tion is the so-called Iterative Closest Point (ICP) (Besl

and McKay, 1992), which performs an iterative pro-

cess in order to minimize the mean square distance

between two sets of 3D points. Nowadays, this algo-

rithm and its different derivatives (Rusinkiewicz and

Levoy, 2001) are still the most usual and effective

ways to achieve our objective. The main problem for

this algorithm is the necessity of a good initialization

if we desire that the iterative process converges to a

global minimum and not to a local minimum.

The obtaining of this initialization becomes the

main problem in most of the existing literature. Some

approaches use a simple combination of a GPS and

a Inertial Measurement Unit (IMU) (Madhavan and

Messina, 2003; Hsu et al., 2003) in order to obtain

both the position and orientation of the laser scanner

in every scanning process. This method provide on

one side the simplicity thanks to the non-existence

of complicated algorithms, but on the other side be-

comes ineffective in indoor applications or places

where the GPS signal cannot be reachable.

More recent approaches make use of the informa-

tion obtained from the visible camera attached to the

laser scanner, extracting information from the more

well-known and deeply studied 2D images process-

ing. The basic idea is to obtain characteristic key-

points from the visible images, and later achieve the

matching between the keypoints using the SIFT de-

scriptor (Lowe, 2004). Once this matching has been

ﬁnished, and assuming that the visible camera and the

laser range scanner are perfectly calibrated, we can

convert this 2D-pixel-matching in 3D-coordinates-

matching, and achieving the desired pairwise registra-

tion. Main problem of this approach is that the match-

ing of SIFT descriptors is capable to cope with small

differences is the viewpoint, but not with high differ-

ences in position as could be our case of study. Some

solutions arise to this problem making use of the 3D

information from the laser scanner (Seo et al., 2005;

Smith et al., 2008), estimating the normal of the 3D

coordinate in the real world and performing an ho-

mography of the visible image as it would be seen

from the front side of the keypoint.

Our model is based also in the use of a descriptor,

541

Mateo Prous X. and Binefa Valls X. (2009).

LASER RANGE DATA REGISTRATION USING SPIN IMAGES.

In Proceedings of the Fourth International Conference on Computer Vision Theory and Applications, pages 541-545

DOI: 10.5220/0001803005410545

 SciTePress

but this time in the 3D descriptor called spin image.

The descriptor is able to compute a 2D image for ev-

ery 3D coordinate taking into account the projection

of the other 3D coordinates in its proximity. The main

beneﬁt from this approach is the computation of the

process only using the 3D information from the laser

scanner and therefore the non-necessity of a calibra-

tion of the visible camera with the range scanner.

The structure of the document is as follows: in

Section 2 we see a review of spin images, which al-

lows a good compression of our model in Section 3.

Experimental results are explained in Section 4 and

ﬁnally conclusions and future work can be found in

Section 5.

2 SPIN IMAGES

Spin images were initially developed during the PhD

thesis of Andrew Edie Johson under the supervision

of Martial Hebert (Johnson, 1997). Later they apply

the algorithm to different cases of 3D recognition and

matching (Johnson and Hebert, 1998; Johnson and

Hebert, 1999). Nowadays they are quite used in facial

recognition (Li et al., 2006) and 3D object recognition

(Assfalg et al., 2007; Matzka et al., 2007).

Basic idea of the spin images is to represent the

proximity structure for every 3D point in a surface or

object. First step for its computation is the estimation

of the surface normal for everypoint we want to create

the spin image. Combination of the 3D point with its

normal vector is called oriented point. The oriented

point deﬁnes a plane and also a cylindrical coordinate

system. Two coordinates can then be deﬁned: a ra-

dial coordinate α and an elevation coordinate β. α

deﬁnes the distance of every point in the proximity

to the line deﬁned by the oriented point, and β de-

ﬁnes the distance of every point to the tangent plane

deﬁned by the oriented point. Their graphical repre-

sentation is shown in Figure 1, where p represents the

3D point that we want to create the spin image, n its

normal vector and x represents one of its proximity

neighbors.

Figure 1: Generation of spin image at oriented point p.

β = (x− p) ·n (1)

α =

(x− p)

− β

(2)

Once all the points in the proximity have been pro-

jected to the plane we obtain a 2D image with a cluster

of dots. At this point the second step of the spin im-

age generation starts: the 2D image can be seen as an

accumulator, resulting in darker areas where the ac-

cumulation of points is higher and lighter areas where

the accumulation is lower. For this accumulation re-

sult we must previously deﬁne a bin size, deﬁned as

the geometric width of the bins in the spin image.

The ﬁnal result of the spin image should be a gray-

level image normalized between 0 (white color) and 1

(black color).

Figure 2: Dependence of bin size in the spin image.

Spin images from two different scans representing

the same object will be similar but not exactly, so in

order to compute the possible matching between two

spin images we can use a simply correlation coefﬁ-

cient. The higher the correlation coefﬁcient, the more

probable that both points represent the same vertex in

the object or surface.

3 REGISTRATION PROCESS

As mentioned, basis of our system is the generation

of spin images. Anyway, due to computational cost

is not advisable to generate a spin image for every

3D vertex in the two scans, so some kind of charac-

teristic points selection is needed. Main advantage

of our system is that it does not require the informa-

tion of the visible camera to achieve the registration.

For this reason we can simply use the range image or

the reﬂectance image of the range scanner, which are

perfectly registered with the 3D scan since they are

captured at the same time as the 3D scan. We can use

also the visible camera image, but taking into account

that probably the calibration would not be so exact.

VISAPP 2009 - International Conference on Computer Vision Theory and Applications

542

Figure 3: Visible image, reﬂectance image and range image.

For our study case we use the reﬂectance image

of the two laser range scans, where we apply a key-

point detector in order to ﬁnd a group of characteris-

tic points. We implemented it with a Harris detector

(Harris and Stephens, 1988), but other keypoint de-

tectors can be used (Lowe, 2004; Mikolajczyk and

Schmid, 2005).

Once the keypoints from both images are selected,

we can compute the spin image of every 3D coordi-

nate associated to the keypoints. As probably both

scans will have a huge extension, we limit the gen-

eration of the spin image using only 3D coordinates

that are at a distance lower than 5 meters. With this

restriction we achieve a spin image that represents the

local structure of the feature independently to the rest

of the laser scan.

Matching between the spin images is carried out

with a simple correlation factor. Every spin image

of a scan are compared with all the spin images of

the other scan, and those that overcome a correlation

threshold (0,6 in our experiments) are selected as pos-

sible correspondences. With all the possible corre-

spondences between both scans in hand, we should

still ﬁnd the higher group of correspondences that are

geometrically consistent. We apply the algorithm ex-

plained in (Johnson and Hebert, 1998), where the con-

sistency is evaluated through the Geometric Consis-

tency Distance (W

). After applying the algorithm

the group with more elements will be considered as

the most probable one, and the rigid transformation

T from scan A to scan B is calculated by minimizing

the error considering all the correspondences of the

chosen group.

min

∑

− T(b

(3)

Finally, in order to accurately register both scans,

the ICP process is applied. Since the initialization

should be good enough the process converge to a

global minimum and the ﬁnal registration will im-

prove the previous result.

4 EXPERIMENTAL RESULTS

The experiments have been carried out with two scans

captured with a laser scanner Riegl LMS-Z420i. Both

scans capture a similar portion of a scene, containing

walls, vegetation and vehicles. Dates of capture were

different, so there is no correlation between the vehi-

cles. The positions of the scans are lightly displaced,

and also with a difference in elevation. Both scans can

be seen in Figures 4 and 5. Also, for a better scene un-

derstanding, the associated visible images are shown.

Figure 4: Scan A and its associated visible image.

Figure 5: Scan B and its associated visible image.

Execution of the different steps presented in this

paper lead us to 542 possible correspondences be-

tween the spin images. After the geometric consis-

LASER RANGE DATA REGISTRATION USING SPIN IMAGES

543

tency grouping, the group with more correspondences

is selected (167 correspondences) and the rigid trans-

formation is calculated. Result can be seen in Figure

6. As expected, result can have some minor errors, so

the Iterative Closest Point algorithm is applied in or-

der to reﬁne the registration. Final result with a more

accurate registration can be seen in Figure 7.

Figure 6: Registration result after the matching grouping.

Figure 7: Final registration result after the ICP process.

5 CONCLUSIONS AND FUTURE

WORK

This paper explains the initial developments of an al-

gorithm to achieve the pairwise registration between

laser range scans taken from different unknown posi-

tions. The registration is based in the computation of

the spin images for different speciﬁc 3D coordinates

and the later matching between them using a simple

correlation factor.

Use of spin images allows us the possibility of

working directly with the 3D data and evaluate, for

every 3D coordinate, the relationship with the other

3D coordinates in the proximity. In addition, the pro-

cessing of the visible image in order to ﬁnd an initial

approximation is not mandatory and all the process-

ing can be done only with the information obtained

from the laser scanner.

Following steps in this study will be the detec-

tion, directly in the 3D surface, of speciﬁc typical

forms: planes (useful for buildings and walls), cylin-

ders (for trees, streetlight or trafﬁc lights) or any other

forms that could be representativefor different objects

present in typical scenarios. The detection of these

typical forms will allow a ﬁltering of non-static ob-

jects (e.g. cars) and thus a better registration between

the 3D points sets. Of course spin images could be re-

ally helpful for this purpose, as they can represent the

local distribution of the 3D coordinate and its neigh-

borhood.

ACKNOWLEDGEMENTS

This work was produced thanks to the support of the

Universitat Aut`onoma de Barcelona (UAB) and the

Centro de Investigaci´on y Desarrollo de la Armada

(CIDA)

REFERENCES

Assfalg, J., Bertini, M., Bimbo, A. D., and Pala, P. (2007).

Content-based retrieval of 3-d objects using spin im-

age signatures. In IEEE Transactions on Multimedia.

Besl, P. J. and McKay, N. D. (1992). A method for registra-

tion of 3-d shapes. In IEEE Transactions on Pattern

Analysis and Machine Intelligence.

Harris, C. and Stephens, M. (1988). A combined corner

and edge detector. In Proceedings of The Fourth Alvey

Vision Conference.

Hsu, S., Samarasekera, S., and Kumar, R. (2003). Auto-

matic registration and visualization of occluded tar-

gets using ladar data. In Proceedings on SPIE Laser

Radar Technology and Applications.

Johnson, A. E. (1997). Spin-images: A representation for 3-

d surface matching. In Doctoral dissertation, tech. re-

port CMU-RI-TR-97-47, Robotics Institute, Carnegie

Mellon University.

Johnson, A. E. and Hebert, M. (1998). Surface matching for

object recognition in complex 3-d scenes. In Image

and Vision Computing.

Johnson, A. E. and Hebert, M. (1999). Using spin images

for efﬁcient object recognition in cluttered 3d scenes.

In IEEE Transactions on Pattern Analysis and Ma-

chine Intelligence.

Li, Y., Smith, W. A., and Hancock, E. R. (2006). Face

recognition using patch-based spin images. In Inter-

national Conference on Pattern Recognition.

Lowe, D. G. (2004). Distinctive image features from scale-

invariant keypoints. In International Journal of Com-

puter Vision.

Madhavan, R. and Messina, E. (2003). Iterative registra-

tion of 3d ladar data for autonomous navigation. In

Proceedings of the IEEE Intelligent Vehicles Symp.

VISAPP 2009 - International Conference on Computer Vision Theory and Applications

544

Matzka, S., Petillot, Y. R., and Wallace, A. M. (2007). De-

termining efﬁcient scan-patterns for 3-d object recog-

nition using spin images. In Advances in Visual Com-

puting.

Mikolajczyk, K. and Schmid, C. (2005). A performance

evaluation of local descriptors. In IEEE Transactions

on Pattern Analysis and Machine Intelligence.

Rusinkiewicz, S. and Levoy, M. (2001). Efﬁcient variants

of the icp algorithm. In International Conference on

3-D Digital Imaging and Modeling.

Seo, J. K., Sharp, G. C., and Lee, S. W. (2005). Range data

registration using photometric features. In Computer

Vision and Pattern Recognition.

Smith, E. R., King, B. J., Stewart, C. V., and Radke,

R. J. (2008). Registration of combined range-intensity

scans: Initialization through veriﬁcation. In Computer

Vision and Image Understanding.

LASER RANGE DATA REGISTRATION USING SPIN IMAGES

545