LIDAR POINT CLOUD CLASSIFICATION USING EXPECTATION MAXIMIZATION ALGORITHM

International Journal of Computer Science & Information Technology (IJCSIT) Vol 12, No 2, April 2020
DOI: 10.5121/ijcsit.2020.12201 1
LIDAR POINT CLOUD CLASSIFICATION USING
EXPECTATION MAXIMIZATION ALGORITHM
Nguyen Thi Huu Phuong
Department of Software Engineering, Faculty of Information Technology, Hanoi
University of Mining and Geology, Hanoi, Vietnam
ABSTRACT
EM algorithm is a common algorithm in data mining techniques. With the idea of using two iterations of
E and M, the algorithm creates a model that can assign class labels to data points. In addition, EM not
only optimizes the parameters of the model but also can predict device data during the iteration.
Therefore, the paper focuses on researching and improving the EM algorithm to suit the LiDAR point
cloud classification. Based on the idea of breaking point cloud and using the scheduling parameter for
step E to help the algorithm converge faster with a shorter run time. The proposed algorithm is tested
with measurement data set in Nghe An province, Vietnam for more than 92% accuracy and has faster
runtime than the original EM algorithm.
KEYWORDS
LiDAR, EM algorithm, Scheduling parameter, LiDAR point elevation, GMM model
1. INTRODUCTION
LiDAR (Light Detection and Ranging) is a survey method that measures the distance to a target
by illuminating the target with the surrounding laser and measuring reflected pulses with a
sensor. The difference in laser travel time and the laser wavelength can then be used to create
digital representations of the target. The essence of LiDAR technology is long laser
measurement technology, GPS / INS spatial positioning, and recognition of light reflection
intensity [1]. The pulse of the laser is sent to the ground at a certain height. Laser waves are
reflected from the ground or from object surfaces such as trees, roads or buildings, with each
pulse measuring the time required to transmit and return signals, calculating the distance from
the laser source to the object. At each laser pulse generation, the GNSS satellite positioning
system will determine the spatial position of the emitting point and the inertial navigation
system will determine the directional angles in the space of the scanning beam. With these
combined measurements, the position (spatial coordinates) of the points on the ground is
calculated [2]. From the set of reflection points, we have a point cloud.
In Vietnam, the application of LiDAR technology in many fields is not uncommon for studies
on data and application of LiDAR technology, which has been included in the development
policies of many ministries such as Ministry of Construction, Ministry of Information and
Communications and Ministry of Resources and Environment. However, LiDAR research in
Vietnam only stopped at the application and use the software with equipment. Through the
implementation process, the problems that arise are unclassified classification when classified
employees still have to use more aerial photos to get classification results as at Ministry of
Resources and Environment. In addition, the copyright issues and ability to update are the
factors that make LiDAR technology not really popular in our country. So far, studies have been
published on the application of LiDAR technology, improvements in technology and data

2
processing algorithms that are constantly published by scientists. The research and application
directions of the authors are relatively broad.
LiDAR data is a relatively large data set and has a wide coverage across the scanning area, so
sorting as a job is not easy. The point filtering process is a key for almost every application with
the LiDAR point cloud. With point filtering algorithms, parameter settings and increased
convergence thresholds increase the accuracy of the point filtering process. Based on that idea,
the authors group Z.Hui, P.Cheng, Y.Yziggah, Y.Nie proposed the algorithm of non-threshold
filtering based on the expected maximization algorithm (EM - Expectation Maximization). The
target splits the point cloud into two groups of ground and non-ground. EM is used to estimate
the maximum expected for the GMM model parameter (Gaussian Mixture Model). After a
number of iterations, the points of the point cloud are labelled, and the authors found that the
proposed algorithm worked well for points belonging to the non-ground group, with only 4.48%
of errors smaller than 8 algorithms. Point filtering math is reported in ISPRS [5]. In his doctoral
dissertation, author Artur Maligo proposed a two-class classification model, the first layer was
built on the GMM model and the second consisted of repeating an intermediate layer to select
the right grade for original classification purpose. Specifically, the GMM model is identified in
the unattended classification training class and defines a set of intermediate classes. Test results
show that the proposed system works well on two data sets with an accuracy of 0.8 - 0.89 [6].
The creation of 3D models of urban areas is currently of interest to scientists. For a data set with
a high density point, containing a lot of noise, choosing an appropriate algorithm is absolutely
necessary. Authors Qing Zhu, Yuan Li, Han Hu, Bo Wu have proposed a point cloud
classification algorithm based on multi-level semantic relationships. The input of the algorithm
is a point cloud through transformations based on the uniformity of the point and the binding of
neighbouring points, which are classified into layers with an accuracy of 93.55%. However, the
proposed algorithm is still limited with received noise [7]. The authors Suresh Lodha, David P.
Helmbold, Darren M.itzpatrick used the algorithm to maximize expectations in classifying
LiDAR ground scattering data into four groups: roads, grass, tall buildings and trees. To carry
out the classification problem, the authors used five characteristics of LiDAR data: altitude,
elevation change, laser reflectance intensity and intensity image. With 94% accuracy was
obtained for the 8 square mile area. Based on the selected parameters and models, the EM
algorithm is appropriate for the classification area [8]. Meanwhile, the team of author Zhenyang
Hui showed that the EM algorithm fully meets the requirements of the automatic DTM
extraction problem. With the error of the proposed algorithm is 16.78% lower than the
traditional PTD algorithm and reduces the system error to 31.95% [9]. The author Nallig Leal
gave the observed data set and variable values for the GMM model in the first iteration, with
each label c in the initialization cluster k, group of iteration is done by calculating the
probability value of each point in each cluster of each iteration j. Then, update the parameters
for the model and repeat until the convergence [10].
In the studies mentioned above, the expectation-maximization algorithm is an approach for
performing maximum likelihood estimation in the presence of latent variables. It does this by
first estimating the values for the latent variables, then optimizing the model, then repeating
these two steps until convergence. It is an effective and general approach and is most commonly
used for density estimation with missing data, such as clustering algorithms like the Gaussian
Mixture Model. In the EM algorithm, the estimation-step would estimate a value for the process
latent variable for each data point, and the maximization step would optimize the parameters of
the probability distributions in an attempt to best capture the density of the data. The process is
repeated until a good set of latent values and a maximum likelihood is achieved that fits the
data.

3
- E-step: perform probabilistic assignments of each data point to some class based on the current
hypothesis h for the distributional class parameters;
- M-step: update the hypothesis h for the distributional class parameters based on the new data
assignments.
The EM classification algorithm is not based on distance. The algorithm calculates the
probability for each observation belonging to each class based on the chosen distribution, the
main purpose of the EM classification algorithm is to find classification solutions to maximize
the overall probability for classification data with the required number of classes. Therefore, in
the classification problem with EM, any difference in the range or scope of the variable selected
for the analysis will not affect the classification results [11].
However, the EM algorithm is very sensitive to initialization values and is prone to errors with
local minimum. In addition, the algorithm is difficult to focus, and covariance matrices
corresponding to one or more components can become an error condition. Wishing to improve
one of the EM algorithm's disadvantages, the author improved the EM algorithm by dividing the
point cloud into smaller point clouds vertically, initialization model parameter, use EM to
conduct point cloud classification (applied to each point cloud part) and assess the degree
exactly. To improve the algorithm's convergence time, use the scheduling parameter β, where β
is initialized with a very small value (approximately 0) [12].
2. PROPOSED METHOD
2.1. Basic EM Algorithm
The EM algorithm is an iterative algorithm to estimate the maximum probability when only a
subset of data is available. The algorithm was proposed by Dempster, Laird, and Rubin in 1977.
The algorithm is stated as follows:
- Suppose for the sample dataset X = (X1, X2, …, Xn), with density conditions, fX|(x|)
with  = . We have formula:
l(;X) = logfX|(X|) (1)
for log-likelihood function [13].
Steps to implement the algorithm [13]:
- Step 1: An initial prediction is made for model parameters and a delivery probability is
generated. Sometimes called "E-Step" for the "Expected" distribution
- Step 2: New data observed based on the model – called E step
- Step 3: The probability distribution from step E is refined to include new data. This
step is sometimes called "M-step".
- Steps 2 to 4 are repeated until a stable distribution is reached.
The EM algorithm has two main steps: E - step and M – step, and described as follows [14]:
Implementation:
+ E – step: is a step that estimates the distribution of labels for a given model: find
E(Zij)
 Assumption  is known and fixed
 A (B) is the xi event drawn from f1(f2)

4
 D the observed points from xi
 The expected value of zi is denoted as P (A |D)
 We have the formula for calculating the expectation D belongs to class A
(2)
With P(D) = P(D|A)P(A) + P(D|B)P(B)
= f1(xi|1)1 + f2(xi|2)2 repeat for each xi
+ M – step: Find the maximum of log-likelihood function θ
Assumption: 1 = 2 = ; 1 = 2 = 0.5 = .
(3)
E[log ] = E (4)
Using the result found E (Zij) in E - step we have the results:
(5)
The pseudocode of the algorithm was built by the author based on the idea of document [15],
[17] as follows:
Input: DL labelled dataset, DU unlabelled dataset, parameter model f()
Output: A  DU classified and f()
Initial: distribution rule 0, parameter 0, 0(0, 0)
Procedure:
1. Calculate the maximum response in each observation:
(6)
2. Update parameter 0(0, 0) to t(t, t) with :
(7)

5
And
End
2.2. Proposed EM
Based on the idea of installing EM algorithm, the proposed method is implemented as the
following figure 1:
Figure 1. Proposed method
We have LiDAR point cloud P = {P1, P2, …, Pn} with each point Pi in point cloud there is a set
of values (Xi, Yi, Zi) show location (X, Y) and elevation (Z) of point. The value that author use
to do classification problems is the value of the height Zi of the point (classifying points by
height). In the point cloud data set with N points, the width of the data D = 3. The
implementation steps of the method are explained as follows:
a. Step 1: Extraction LiDAR point elevation
From the point set in the point cloud, take the point height value as input for the later
computation steps, which means we now reduce the width of the data D = 3 → D = 1.
b. Step 2: Initialization model parameter
Suppose there is a data set Z = {Z1, Z2, …, Zn} with Z is the point elevation. We have a
probability model for the Gaussian distribution shown as follows:
(8)
In which, μ - average value of the data set, C - mixing coefficient, σ - standard deviation, β -
scheduling parameter (reduces the convergence time of the algorithm EM) [14], [16].
- Initialization μ, β:
With the value μ after extracting the elevation information of the point in step 1, we will
calculate μ by taking the average value of the height of all points in the point cloud according to
the formula (2):
(9)

6
β is initialized with a value of approximately 0. The parameter β can roughly be interpreted as
the inverse of temperature [16]. At each β value algorithm iterates E step and M step until
convergence.
- Initialization C, δ
The mixing coefficient is the probability to determine the proportion of data belonging to the k
component and must ensure conditions:
(10)
Determining the mixing coefficients for k components of GMM model is calculated by the formula:
(11)
With is the total number of points in the dataset belongs to the kth component. In
this study, author initiates a mixing coefficient for all components of the model to be equal.
There, σ is standard deviation which is calculated by the formula:
(12)
With
c. Step 3: Split LiDAR point cloud vertically
With a very large point dataset in the point cloud, updating parameters and converging
calculations takes a lot of time. Therefore, the graduate student divides the point cloud vertically
based on the height of the point. However, choosing the number of parts to split to prevent the
algorithm from falling into a never-ending state requires calculation.
To solve this problem, author took an example on a given point data set (number <3,000), the
algorithm converges very well. Therefore, it is recommended that the number of points to be
divided should be less than 10.
d. Step 4: Using EM estimate model parameter
- Increase value β by 1 unit for each loop (value of β is not greater than 1, βmax = 1). If β > 1,
reset β value = 1.
- At step E of the algorithm calculates the probability that a point Zi belongs to the kth component
by the formula:
(13)
With, (14)
And (15)
And the Gaussian distribution function is calculated according to the formula:
(16)
- At step M, update the parameters of the model with the values of μ, Σ, Ϲ, σ respectively

7
according to the following formulas:
(17)
(18)
(19)
(20)
- Algorithm convergence: the algorithm converges when the new parameter does not change
compared to the parameter that exceeds the threshold value ε, with ε calculated by the following
formula:
(21)
e. Step 5: LiDAR point cloud classification
Based on the probability of a point belonging to a certain k class in the model, we can proceed to
classify points. Based on observed data, if P (k | Zi) ≥ 0.5 then conclude that Zi belongs to class k.
f. Step 6: Precision evaluating
To evaluate the accuracy of the algorithm, author uses the following measurements: accuracy, recall
and F1 measurement of the improved EM algorithm with the basic EM algorithm and the basic EM
algorithm.
Pseudocode of algorithm can be performance below:
Begin
Input: LiDAR point cloud P = {P1, P2, …, Pn}
Output: Point cloud classified
Initialization: the mean - μ, covariance matrix- Σ, mixing coefficient - C, standard
deviation - σ, scheduling parameter - β
Procedure:
for i = 1 to n
for i = 1 to k
Increase β up to 1 unit
If β < 1 continue E step else set β = 1
E – step
M- step

8
if (convergence) stop loop
else repeat E and M step
if P(k|Zi) > = 0.5 Zi belongs to class k
else Zi not belongs to class k
Precision evaluating
End
2.3. Experiment
2.3.1. Experiment Database
Dataset use in paper is LiDAR point cloud data. Point clouds are a powerful and dynamic
information storage technology. By representing spatial data as a collection of coordinates, they can
handle large datasets for a wide array of downstream processing. Primarily, in this instance, they are
used as a middleman to turn the raw data collected by LiDAR processes into 3D models.
With the LiDAR data was investigated in Dien Chau, Nghe An Province. This is an area with
mountainous terrain that occupies two thirds of the area and hilly slopes inclining from the north to
the south. Data were collected with the full waveform airborne LiDAR survey. Data was measured
from the 29st
of September, 2016 to the 29h
of September, 2016, with an area of 102 km2
, 0.25
pulse/m2
, and a density of 4.18 points/m2
. The data include 934.843 points, the format of data is
‘.las’ version 1.2. Each point in the point cloud has the properties shown in Figure 3. However, the
author uses the coordinates and the height of the point as the input of the classification problem and
displays the 3D point cloud. In order to easily read and understand data from .las files, the author has
saved it in .txt format. In .txt file the author only displays the x, y, and z values of the point The
values of the points are shown in Figure 4. LiDAR point cloud 3D is showed in figure 2. Each point
in the point cloud is represented by 3 values of x, y and z, where, the value of height z will be
considered as the input value of the classification problem.
Figure 2. Attribute of each point

9
Figure 3. X, Y, Z value of points
2.3.2. Results and Discussion
The proposed method tested with the above data set starts with the extraction of point
height data. The accuracy of the algorithm will be assessed on accuracy, recall and F1
measurement.
First step, extract the elevation of the point in the LiDAR point cloud to reduce the dimension of
data D = 1. The extracted results are shown in Figure 4.
Figure 4. Extraction of point elevation
Display points in height and divide the point cloud vertically with 9 parts (because the number
of points is not too large, but the height distribution of points is uneven, there is a large
elevation difference, so the selection of part numbers need to divide by 9 to avoid uneven height
distribution between the parts to help the algorithm on each part converge quickly, then the
point cloud is represented as follows:

10
Figure 5. The following point cloud is divided into 9 parts
The model parameters are initialized on the Z value as follows:
μ = 94.38,
Σ = 1584.07,
δ = 39.80,
β = 10-8,
To determine the number of components of the model, the author used principal component
analysis to find k. The finished matrix is shown in Figure 5, 6 and 7. Using the EM algorithm to
recalculate the model's parameters on each part, with 25 iterations, the algorithm converges with
the parameters changed in figure 8.
Figure 6. Correlation between data sets
Figure 7. principal component analysis of dataset
Figure 8. Component matrix

11
We can see in the total covariance table that the cumulative value is greater than 50%, so we can
choose the number of components for the model k = 2. The mixing coefficient is initialized
equally with 2 components C1 = C2 = 0.5.
Figure 9. Model parameters are updated through each iteration of steps E and M
Conducting classification with the condition of classification P (k | Zi) ≥ 0.5, conclude that Zi
score belongs to class k (k = 2). We have a class 1 score of 768.970 points, a class 2 score of
165.873 points. We have the distribution of points into class 1 and 2 after the classification
shown in Figure 10 and correlation between frequency of distribution and cumulative
percentage of data is shown in table 1.
Table. 1. Correlation between frequency of distribution and cumulative percentage of data
Bin Frequency
Cumulative
% Bin Frequency
Cumulative
%
1 0 0.00% More 934842 100.00%
2 0 0.00% 1 0 100.00%
More 934843 100.00% 2 0 100.00%
Figure 10. Histogram of point distribution after classification

12
To evaluate the accuracy of the proposed EM algorithm, the author compares the results with
the results of basic EM. The results are shown in Table 2 with precision, recall and F1
measurements. The result is shown in table 2.
Table. 1 Compare result proposed EM with basic EM
Precision Recall F1 Running Time Convergence
Proposed EM 92.03% 92.06% 0.92045 102.3s 0.0013
Basic EM 91.80% 91.79% 0.91795 210.25s 0.00021
Through testing and comparing the evaluation of the proposed algorithm with the original EM
algorithm with improved accuracy and faster running time by using scheduling parameters to
help the algorithm converge better. However, due to the condition if β > 1 consider the
condition for β = 1, the condition for the convergence algorithm is not as good as the original
EM algorithm. With the proposed EM algorithm classification results, it meets the requirements
of LiDAR point cloud classification, but more research is needed to make the algorithm's
convergence conditions better and improved the accuracy.
3. CONCLUSIONS
EM algorithm is a popular algorithm in data mining and is used in many different problems.
With the LiDAR point classification cloud problem, with the idea of classification based on a
one-point posterior probability belonging to the classification class, EM algorithm proposed
classification with more than 92% accuracy to meet the requirements of establishment. With the
proposed EM algorithm based on the initialization parameter initialized with very small price
and vertical point cloud subdivision to help the algorithm converge faster, while reducing the
running time of the algorithm compared to Original EM algorithm. However, the algorithm
needs further development when only classifying the cloud into 2 classes: first pulse and last
pulse. While the number of unlabelled points is very large, this is a useful amount of
information to study the nature and morphology of the object.
With the data set measured in Nghe An province, Vietnam, the algorithm gives satisfactory
results and the running time with an acceptable convergence. However, more data sets need to
be tested to verify the accuracy.
ACKNOWLEDGEMENTS
This research is funded by the Hanoi University of Mining and Geology under project number
T19-03 and Ministry of Education and Training number CT.2019.01.07
REFERENCES
[1] Nallig Leal, Esmeide Leal, Sanchez Torres German A linear programming approach for 3D point
cloud simplification.
[2] Zhenyang Hui, Dajung Li, Shuanggen Jin, Yao Yevenyo Ziggah, Leyang Wang (2019) Automatic
DTM extraction from Airborne LiDAR based on expectation - maximization, Optics and Laser
Technology, vol.112, pp. 43-55
[3] Kun Zhang, Weihong Bi, Xiaoming Zhang, Xinghu Fu, Kunpeng Zhu, Li Zhu (2015) A new kmeans
clustering algorithm for point cloud, International Journal of Hybrid Information Technology, vol.
vol. 8, no. 9, pp. 157-170.

13
[4] Keng FanLin, Chi Pei Wang, Pai Hui Su (2012) Object-based classification for LiDAR point cloud.
[5] Chao Luo, Gunho Sohn (2013) Line-based classification of terrestrial laser scanning data using
conditional random field, International Archives of the Photogrammetry, Remote Sensing and
Spatial Information Sciences, Vol XL-7/W2, ISPRS2013.
[6] Xiao Liu, Congyin Han, Tiande Guo (2018) A robust point sets matching method [Online].
Available: https://arxiv.org/ftp/arxiv/papers/1411/1411.0791.pdf. [Accessed 9 10 2019].
[7] Z.Hui, P.Cheng, Y.Y. Ziggah, Y.Nie (2018) A threshold-free filtering algorithm for airborne LiDAR
point clouds based on Expectation Maximization, The International Archives of the
Photogrammetry, RS and Spatial Information Sciences, vol. XLII-3.
[8] Suresh Lodha, David P.Helmbold, Darren M.Fitzpatrick (2007) Aerial LiDAR data classification
using expectation – maximization, Research Gate.
[9] Yang HongLei, Peng JunHuan, Zhang DingXuan (2013) An Improved Em algorithm for remote
sensing classification, Chinese Science Bullentin, vol. 58, no. 9, pp. 1060-1071.
[10] Xiao Liu, Congying han, Tiande Guo (2014), "arXiv”. [Online]. Available:
https://arxiv.org/ftp/arxiv/papers/1411/1411.0791.pdf. [Accessed 7 10 2019].
[11] Yu-chuan Chang, Ayman F.habib, Dong Cheon Lee, Jae Hong Yom (2008) Automatic classification
of LiDAR data into Ground and non-Ground points, The International Archives of the
Photogrammetry, RS and Spatial Information Sciences, vol. XXXVII, no. B4, pp.457-46.
[12] Borja Rodriguez - Cuenca, Silverio Garcia Cortes, Celestino Ordonez, Maria C.Alonso (2015)
Automatic detection and classification of pole-like objects in urban point cloud data using an
anomaly detection algorithm, Remote Sensing, vol. 7, pp.12680-12703.
[13] Serez Kutluk, Koray Kayabol, Aydin Akan (2016) Classification of Hyperspectral Images using
Mixture of Probabilistics PCA models, European signal processing conference, pp. 1568-1572.
[14] Iftekhar Naim, Daniel Gildea (2012) Convergence of EM algorithm for GMM with unbalanced
Mixing coefficients, International Conference on Machine Learning.
[15] Lawrence H. Cox, Marco Better (2009). Sampling from discrete distributions: Application to an
editing problem, Research Gate.
[16] Naonori Ueda, Ryohei Nakano (1998) Deterministic Annealing Variant of the EM algorithm,
[Online]. Available: https://papers.nips.cc/paper/941-deterministic-annealing-variant-of-the-em-
algorithm.pdf. [Accessed 6 3 2020].
[17] S. Borman (2006) The expectation maximization algorithm a short tutorial.
AUTHOR
Born in 1985 in Ninh Binh province, Vietnam Country Graduated from the
University of Mining and Geology University in 2008. Graduated Master of
Science at the University of Natural Sciences - Vietnam National University,
Hanoi in 2012. Author is currently a PhD student specialized in Information
systems at the Institute of Information Technology, Vietnam Academy of
Science and Technology. Currently working at: Faculty of Information
Technology, University of Mining and Geology Research interests:
Information System, Database, Data Mining, Geoinformatics.

LIDAR POINT CLOUD CLASSIFICATION USING EXPECTATION MAXIMIZATION ALGORITHM

More Related Content

What's hot (20)

Similar to LIDAR POINT CLOUD CLASSIFICATION USING EXPECTATION MAXIMIZATION ALGORITHM (20)

Recently uploaded (20)

LIDAR POINT CLOUD CLASSIFICATION USING EXPECTATION MAXIMIZATION ALGORITHM