Geo-spatial Information Science 14(3):184-189 Volume 14, Issue 3

DOI 10.1007/s11806-011-0540-x September 2011

Article ID: 1009-5020(2011)03-184-06 Document code: A

Pattern Detection in Airborne LiDAR Data Using

Laplacian of Gaussian Filter

ZHAN Qingming1,2, LIANG Yubin2, CAI Ying1, XIAO Yinghui1

1. School of Urban Design, Wuhan University, 8 South Donghu Road, Wuhan 430072, China
2. School of Remote Sensing and Information Engineering, Wuhan University, 129 Luoyu Road, Wuhan 430079, China

© Wuhan University and Springer-Verlag Berlin Heidelberg 2011

Abstract Methods for feature detection in laser scanning data have been studied for decades ever since the emergence of the
technology. However, it is still one of the unsolved problems in LiDAR data processing due to difficulty of texture and structure
information extraction in unevenly sampled points. The paper analyzes the characteristics of Laplacian of Gaussian (LoG) Filter
and its potential use for structure detection in LiDAR data. A feature detection method based on LoG filtering is presented and ex-
perimented on the unstructured points. The method filters the elevation value (namely, z coordinate value) of each point by convo-
lution using LoG kernel within its local area and derives patterns suggesting the existence of certain types of ground ob-
jects/features. The experiments are carried on a point cloud dataset acquired from a neighborhood area. The results demonstrate
patterns detected at different scales and the relationship between standard deviation that defines LoG kernel and neighborhood size,
which specifies the local area that is analyzed.
Keywords laser scanning; point cloud; feature detection; Laplacian of Gaussian filter
CLC number P237

tern detection (segmentation or classification) in LiDAR

Introduction data has been studied for decades. Though scanning data
may be acquired from different platforms (spaceborne,
In recent years, Light Detection And Ranging (Li- airborne, terrestrial, and mobile), the processing meth-
DAR) has emerged as an effective technology, which ods usually share common principles. One kind of de-
provides valuable data in various forms and scales for tection algorithm mainly exploits geometrical deriva-
mapping and monitoring land cover features. Its use has tives (curvature, normal, etc.) to group points.[1,2] An-
increased dramatically due to availability of high-den- other kind of algorithm[3] utilizes a parameterized detec-
sity LiDAR data as well as high spatial/spectral resolu- tion method, such as the Generalized Hough Trans-
tion digital images. Nowadays, LiDAR data are often form[4] to detect regular shapes including planes, circles,
derived from one or multireturns of laser pulses and the cylinders, and spheres.
digital images usually contain multispectral bands. Pat- Some types of ground objects, such as tree leaves and

► Received on March 24, 2011.

► Supported by the National Natural Science Foundation of China (No.40871211).
► ZHAN Qingming is currently a professor and associate dean, School of Urban Design, Wuhan University. He received his BSc and MSc in photo-
grammetry and remote sensing from Wuhan Technical University of Surveying and Mapping (WTUSM), Wuhan, China, in 1985 and 1995 respec-
tively. He obtained Ph.D. in Geo-information Science from Wageningen University - ITC, the Netherlands in 2003. His research experience covers
GIS, remote sensing and applications in urban planning and management. His research interests include planning support systems, 3D city modeling
and applications, object-based analysis of remote sensing images, land-cover and land-use classification methods, etc.
► Email: [email protected]
 ZHAN Qingming, et al./ Pattern Detection in Airborne LiDAR … 185

shrubs, cannot be parameterized by regular geometrical problems that resulted from theory and applica-
shapes partly due to their penetration of laser beams. tion.[13-16]
Researchers[5-8] in the robotics and computer science Marr[17] raised the computational framework for
field have applied machine-learning-based methods to computer vision and suggested Laplacian of Gaussian
laser scanning data recognition for robotic navigation. (LoG) filter for visual information processing. Due to its
These procedures often explore statistics of spatial dis- similarity to biological vision, LoG filter has been well
tribution of returned laser points and classify returned studied for feature extraction (e.g., edge detection) in
laser data using learned classifiers (e.g., Gaussian Mix- images. Witkin[18] presented an approach called scale-
ture Models and Markov Random Fields). Although space filtering to detect image locations invariant to
different frameworks are developed, local region analy- scale change. Lowe[19] introduced Scale Invariant Fea-
sis methods and k nearest neighbors (KNN) search algo- ture Transform (SIFT) approach for feature point loca-
rithms are often exploited in these procedures. tion, which has been widely used for image matching
Methods fusing datasets from different sensors and image-based modeling. Marr and Hildreth[20] had
have drawn lots of interests in recent years. Due to its mentioned that the LoG filter could be used for detect-
complementary properties to LiDAR data, high spa- ing changes in intensity corresponding to properties of
tial/spectral digital images have been combined for physical world, like changes in reflectance, surface ori-
improving land cover mapping. Although aerial pho- entation, or distance from the viewer. Based on this ob-
tography has been used as a mapping tool for a cen- servation, it should be possible to detect changes and
tury,[9] the fusion of aerial photography and LiDAR patterns of altimetry values in LiDAR data using LoG
data has only been possible in the past few years due filter. The motivation of this research is to study and ap-
to advances in sensor design and data acquisi- ply LoG filter for pattern detection in LiDAR data. In
tion/processing techniques. There have been some at- the following sections, principles of LoG filter will first
tempts to fuse LiDAR and high-resolution imagery be introduced. Then, the convolution scheme will be
for land cover mapping. Haala and Brenner[10] applied described. Analysis of experiments and conclusions will
the Iterative Self-Organizing Data Analysis Tech- be delivered separately in Sections 2 and 3.
nique (ISODATA) algorithm to combined LiDAR-
derived DSM and three-color-band aerial images. The 1 Convolution with log filter
normalized Digital Surface Model (nDSM) was used
to classify objects that had different distribution pat- An appropriate filter for intensity changes at a
terns in elevation direction. The low-resolution Li- given scale was found to be the second derivative of a
DAR data was greatly facilitated to separate trees Gaussian.[20] It has been proved that, under some
from buildings by the near-infrared band from the ae- reasonable assumptions, LoG filter is found to be the
rial imagery. Sohn and Dowman[11] presented method appropriate one that is independent of orientation.
utilized a divide-merge scheme to obtain the recog-
1.1 Laplacian of Gaussian filter
nized building outline in a combination of multispec-
tral images and airborne laser scanning data. With the The 2-D Gaussian smoothing filter is a convolution
availability of full-waveform LiDAR data and hy- operator that is often used to remove details and
perspectral images, the problems of data fusion and noises in images. It has been shown that under some
pattern classification become more complicated. Op- conditions, the only possible scale-space kernel is the
portunities with high classification accuracy should Gaussian function,[21] which is useful for feature lo-
be achieved because of its spectral and spatial fea- calization. Furthermore, Gaussian filter provides
tures. However, there are still challenges in data “weighted” smoothing and preserves edges better
processing, waveform modeling, and measurements than average filter of similar size. The 2-D Gaussian
interpretation of full-waveform LiDAR.[12] A few distribution has the form
x2 + y2
methods have been put forward to tackle these prob- 1 −
G ( x, y ) = e 2σ 2
(1)
lems, but more efforts should be taken to investigate 2πσ 2
186 Geo-spatial Information Science 14(3): 184-189

where σ is the standard deviation. The distribution of standard deviation σ has the form of Eq.(3). Fig.3
2-D Gaussian function with mean (20, 20) and illustrates the distribution of LoG function with mean
σ =5.0 is illustrated in Fig.1. (20, 20) and σ =5.0.
x2 + y2
1 ⎡ x2 + y 2 ⎤
LoG ( x, y ) = ⎢1 + ⎥e
2σ 2
(3)
πσ 4 ⎣ 2σ 2 ⎦

Fig. 1 2-D Gaussian function with σ =5.0

As a smoothing filter, Gaussian filter reduces the

range of scales over which intensity changes can
take place. Intensity change detection is under the Fig. 3 2-D LoG function with σ =5.0

assumption that wherever a change occurs, there 1.2 Convolution scheme

will be a corresponding peak in the first directional
derivative or, equivalently, a zero-crossing in the Wack and Wimmer[22] used LoG kernel to filter
second directional derivative of intensity.[20] To de- ground objects in raster image derived from LiDAR
tect the direction of intensity changes, the sec- data so as to model terrain. However, transformation
ond-order differential operator Laplacian ∇ 2 (Eq. of point cloud data from vector format to raster for-
(2)) is often used. mat unavoidably loses accuracy and introduces un-
∂ 2 f ( x , y ) ∂ 2 f ( x, y ) certainties, which will influence both the detection of
L ( x , y ) = ∇ 2 f ( x, y ) = + (2)
∂x 2 ∂y 2 above-ground objects and the extraction of ground
Therefore, the detection of intensity changes can be points. Therefore, we choose to filter the raw point
reduced to detect zero-crossings in ∇ 2G -filtered im- cloud data instead of filtering the raster image. Eq.(3)
age. Fig.2 shows that in areas where the intensity could be written compactly as
r2
does not change, the LoG response should be zero. 1 ⎡ r 2 ⎤ − 2σ 2
LoG ( r ) = ⎢1 + ⎥e (4)
When intensity increases in the proximity, the LoG πσ 4 ⎣ 2σ 2 ⎦
response will be positive and negative, vice versa.
where r = ( x0 - xKNN ) + ( y0 − yKNN ) is the dis-
2 2

tance between a point and each of its k nearest

neighbors (KNN) within a given distance δ , which
specifies the local area of the neighborhood. The fil-
ter filters z value of each laser point by convolution of
z values of its KNN in the local area defined by δ .
There are many available implementations using
kd-tree[23] for KNN search, and we choose the freely
Fig. 2 Response of 1-D LoG filter (b) to a step edge (a) available “kd-tree for matlab.” [24]
Since convolution operation is associative, it is ef-
ficient to calculate the filter kernel once and convolve 2 Results and analysis
the image with the hybrid filter by convolution of the
Gaussian smoothing filter with the Laplacian filter. The presented method was applied to a point cloud
The 2-D LoG function centered at (0, 0) and with dataset acquired from a neighborhood area including
 ZHAN Qingming, et al./ Pattern Detection in Airborne LiDAR … 187

a church that is surrounded by several houses and

trees. The dataset contains 146972 points. The re-
search area is displayed in Fig.4, and Fig.5 shows the
corresponding elevation map. Figs.6(a-c) illustrate
convolution result with LoG filter with increasing δ .
Fig.6(a) suggests that σ =1.0 with δ =1.0 m is too
sharp for edge detection because no filtered z value is
beyond 0. The more patterns emerge with increased
δ in Fig.6(b). The Z-value of filtered points ranges
from below −1500 to beyond 500. The yellow points
indicate the areas of ground objects, where intensity Fig. 5 Elevation map
changes little. Blue and red points reflect dramatic
intensity change. These as a whole suggest the exis-
tence of boundaries between regions of uniform but
different intensities. The convolution results with
σ =1.0 slightly change with increasing δ value lar-
ger than 4.0 m. This may indicate the degree of de-
tails, which can be detected under the scale limited by
σ =1.0. Though the difference in both filtered value
is almost in an order of magnitude when compared
with Fig.6(a), Fig.7(a) demonstrates that in the same
locale area, detected features generally slightly
changes with continuously increasing scales. Trees
detected in all these experiments share almost the
same patterns. The top areas of trees are often with
the smallest filtered values, and this is because the z
values of their neighboring points are often smaller
than theirs. According to the observation from that in
Fig.2, the filtered values to some extent reflect the
difference of height between trees, and the shapes of
tree canopy begin to emerge (Fig.8). Figs.6(a)-6(c)
also show that boundaries of ground object could be

Fig. 6 Convolution result with LoG filter with

Fig. 4 The testing dataset
increasing δ
188 Geo-spatial Information Science 14(3): 184-189

tect circles (Fig.10) that resemble tree canopies and the

located trees are shown in Fig.11. Six of the seven
trees appear in the test site are correctly located, how-
ever the tower of the church is recognized as trees due
to similarity in shape.

Fig. 7 Convolution with LoG ( σ =2.0, δ =1.0 m)

extracted when there’re both positive and negative

values in the LoG-filtered point cloud. We thresh-
old the filtered value of Fig.6(b) at 500 to produce a
sketch map (Fig.9) where the extracted boundaries
represent the locations of zero-crossing points. Then Fig. 8 Convolution with LoG ( σ =3.0, δ =2.0 m)
Generalized Hough Transform (GHT)[4] is used to de-

Fig. 9 Sketch map Fig. 10 Detected circles using GHT Fig. 11 Located trees

search consumes much computation. We are consid-

3 Conclusion ering random algorithms for the improvement of
performance while keeping spatial distribution of
In this paper, we presented some experimental re- points in local areas. Besides, the presented method
sults of a method based on LoG filter to detect pat- depends on choice of σ and δ . The parameter σ
terns in airborne LiDAR data. The filtered results re- defines operation scale of LoG filter and δ specifies
flect existence of certain types of ground objects and the convolution area. Though it is more complicated
the detected patterns could be further studied and ex- to select the appropriate parameters in raw LiDAR
ploited to extract individual trees and buildings. Our points processing than that in image analysis, the
method proceeds as weighted smoothing filter on selection of σ should be related to the scale of fea-
each laser scanning point. Sketch map is extracted by tures to be detected. Parameter δ should be deter-
thresholding the filtered values and the Generalized mined by the density of the point clouds. And once
Hough Transform is then carried out to extract circles positive and negative values show in the filtered
which indicates canopy of individual trees. Moreover, points, the boundaries of ground object could be ex-
sketch map may also be used to register LiDAR data tracted. More attention will be paid in our future
with images as well as provide measurement infor- work to features detected at different scales and
mation (such as area) of buildings. Although the me- possible methods for feature detection with auto-
thod is easy to implement, the k nearest neighbors matic scale selection.
 ZHAN Qingming, et al./ Pattern Detection in Airborne LiDAR … 189

[12] Wagner W, Ullrich A, Melzer T, et al. (2004) From sin-

References
gle-pulse to full-waveform airborne laser scanners: potential
[1] Maas H G,Vosselman G (1999) Two algorithms for extract- and practical challenges[J].International Archives of the
ing building models from raw laser altimetry data[J]. ISPRS Photogrammetry, Remote Sensing and Spatial Information
Journal of Photogrammetry and Remote Sensing, 54(2-3): Sciences, 35 (Part B3): 201-206
153-163 [13] Wagner W, Ullrich A, Ducic V, et al. (2006) Gaussian de-
[2] Vosselman G, Gorte B G H, Sithole G, et al. (2004) Recog- composition and calibration of a novel small-footprint
nising structure in laser scanner point clouds[J]. Interna- full-waveform digitising airborne laser scanner[J]. ISPRS
tional Archives of the Photogrammetry, Remote Sensing and Journal of Photogrammetry and Remote Sensing, 60(2):
Spatial Information Sciences, 36: 33-38 100-112
[3] Rabbani T, Dijkman S, van den Heuvel, et al. (2007) An in- [14] Chauve A, Vega C, Durrieu S, et al. (2009) Advanced-
tegrated approach for modeling and global registration of full-waveform lidar data echo detection: assessing quality of
point clouds[J]. ISPRS Journal of Photogrammetry and Re- derived terrain and tree height models in an alpine conifer-
mote Sensing, 61(6): 355-370 ous forest[J]. International Journal of Remote Sensing,
[4] Ballard D H (1981) Generalizing the Hough transform to 30(19): 5211-5228
detect arbitrary shapes[J]. Pattern Recognition, 13(2): [15] Mallet C, Bretar F (2009) Full-waveform topographic lidar:
111-122 state-of-the-art[J]. ISPRS Journal of Photogrammetry and
[5] Anguelov D, Taskarf B, Chatalbashev V, et al. (2005) Dis- Remote Sensing, 64(1): 1-16
criminative learning of Markov random fields for segmenta- [16] Alexander C, Tansey K, Kaduk J,et al. (2010) Backscatter
tion of 3D scan data[C].Proceeding of IEEE Computer So- coefficient as an attribute for the classification of full-
ciety Conference on Computer Vision and Pattern Recogni- waveform[J]. ISPRS Journal of Photogrammetry and Re-
tion (CVPR), San Diego, CA, USA mote Sensing, 65(5): 423-432
[6] Triebel R, Kersting K, Burgard W (2006) Robust 3D scan [17] Marr D (1982) Vision[M]. San Fransisco:W. H. Freeman
point classification using associative Markov networks[C]. [18] Witkin A P (1983) Scale-space filtering[C]. Proceeding of
IEEE International Conference on Robotics and Automation International Joint Conference on Artificial Intelligence,
(ICRA), Orlando, Florida, USA Karlsruhe, Germany
[7] Munoz D, Bagnell J A, Vandapel N,et al. (2009) Contextual [19] Lowe D G (2004) Distinctive image features from scale- in-
classification with functional max-margin Markov net- variant keypoints[J]. International Journal of Computer Vi-
works[C]. IEEE Computer Society Conference on Computer sion, 60(2): 91-110
Vision and Pattern Recognition (CVPR), Miami, FL, USA [20] Marr D,Hildreth E C (1980) Theory of edge detection[J].
[8] Munoz D, Vandapel N,Hebert M (2009) Onboard contextual Proceedings of the Royal Society of London. Series B, Bio-
classification of 3-D point clouds with learned high-order logical Sciences , 207(1167): 187-217
markov random fields[C].IEEE International Conference on [21] Lindeberg T (1994) Scale-space theory: A basic tool for
Robotics and Automation (ICRA), Kobe, Japan analysing structures at different scales[J]. Journal of Applied
[9] Baltsavias E P (1999) A comparison between photogram- Statistics, 21(2): 224-270
metry and laser scanning[J]. ISPRS Journal of Photogram- [22] Wack R, Wimmer A (2002) Digital terrain models from
metry and Remote Sensing, 54(2-3): 83-94 airborne laser scanner data—a grid based approach[J]. In-
[10] Haala N, Brenner C (1999) Extraction of buildings and trees ternational Archives of the Photogrammetry, Remote Sens-
in urban environments[J]. ISPRS Journal of Photogram- ing and Spatial Information Sciences, 34 (Pt. 3B): 293-296
metry and Remote Sensing, 54(2-3): 130-137 [23] Bentley J L (1975) Multidimensional binary search trees
[11] Sohn G,Dowman I (2007) Data fusion of high-resolution sa- used for associative searching[J]. Communications of the
tellite imagery and LiDAR data for automatic building ex- ACM, 18(9): 509-517
traction[J]. ISPRS Journal of Photogrammetry and Remote [24] Tagliasacchi T (2010) Kd-tree for matlab[OL]. http:// www.
Sensing, 62(1): 43-63 mathworks.com/matlabcentral/fileexchange/21512