You are on page 1of 17

Machine Vision and Applications (2023) 34:9

https://doi.org/10.1007/s00138-022-01358-y

ORIGINAL PAPER

Automatic cables segmentation from a substation device based on 3D


point cloud
Qianjin Yuan1 · Jing Chang1 · Yong Luo1 · Tianlei Ma1 · Dongshu Wang1

Received: 30 August 2021 / Revised: 21 November 2022 / Accepted: 22 November 2022 / Published online: 9 December 2022
© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2022

Abstract
The point cloud segmentation of a substation device attached with cables is the basis of substation identification and recon-
struction. However, it is limited by a number of factors including the huge amount of point cloud data of a substation device,
irregular shape, unclear feature distinction due to the auxiliary point cloud data attached to the main body of a device. There-
fore, the segmentation efficiency of a substation device is very low. In order to improve the accuracy and efficiency of the
point cloud segmentation, this paper proposes a method to segment the attached cables point cloud of a substation device by
using the shape feature of point cloud. Firstly, according to the spatial position of the point cloud of a substation device, octree
is used to conduct voxelization of the point cloud, and the point cloud resampling is operated according to point cloud density
of each voxel, so as to reduce original point cloud data and improve computing efficiency. Then Mean Shift algorithm is used
to locate the center axis of the point cloud, and cylinder growth method is used to initially segment cables data and locate the
end of each cable. Finally, points of the end are used as seed points to carry out a region growth based on shape feature of the
point cloud to realize effective segmentation of cables data. In the experiment, 303 sets of point cloud of devices are selected,
including circuit breaker, voltage transformer, transformer, etc. The final result shows that the successful segmentation rate
of this method reaches 95.34%, which effectively proves the feasibility of this method.

Keywords Point cloud · 3D object segmentation · Cable segmentation · Mean shift · Region growth

1 Introduction When modeling a substation based on a 3D point cloud, the


first step is to identify the specific model of the equipment of
The development of computer vision technology has greatly substation, but whether the electrical equipment is manually
promoted the digital transformation of traditional manufac- segmented or automatically segmented from the substation
turing industry and brought great convenience to people’s scene, it inevitably comes with cables data.
production and life, especially the emergence of new tech- The existence of them will seriously affect the result
nologies such as digital city and virtual reality technology, of feature extraction, then influence equipment recognition
which greatly promotes the intelligence of cities and fac- accuracy. So before recognizing a substation device, the
tories. As the basis for realizing digital city and virtual point cloud of auxiliary facilities needs to be segmented and
reality, 3D reconstruction technology gradually becomes a removed from the original point cloud of equipment, then the
key research focus for researchers. points remained can be seen as the main body of the substa-
The booming development of 3D modeling technology tion equipment and can be used to recognize and reconstruct
also promotes the construction of intelligent substations, substation equipment.
which makes them more intelligent and informative. At the 3D data segmentation refers to the clustering problem of
same time, the complex substation scenes have put forward dividing logical whole 3D data into several separate groups
higher requirements for the reconstruction of 3D models. [1]. The current methods for handling 3D point cloud seg-
mentation include two main categories. The first one is the
B Yong Luo traditional segmentation method of point cloud data by using
luoyong@zzu.edu.cn the characteristics of point cloud data itself, which can be
divided into the following three types: edge-based, region-
1 School of Electrical Engineering, Zhengzhou University,
growth-based and model-fitting-based methods.
Zhengzhou, People’s Republic of China

123
9 Page 2 of 17 Q. Yuanet al.

The edge-based segmentation method is to directly apply also be considered as the methods of shape detection. There
2D image segmentation method to 3D point cloud. The edge are two main model fitting algorithms most commonly used:
points are detected by the surface attributes of the point cloud, the Hough transform algorithm (HT) and the random sam-
then the edge points are grouped, so as to achieve the orig- pling consistency algorithm (RANSAC). The HT algorithm
inal point cloud segmentation. Jiang et al. proposed a point is a classical feature detection algorithm in the field of image
cloud segmentation based on scan lines. In this method, scan processing, which was proposed by Richard Duda and Peter
lines in distance image are taken as the primitive segmenta- Hart in 1972. Its core idea is to map samples in original
tion and the 3D model is segmented by clustering scanned space to discrete parameter space, then vote each input sam-
lines [2]. This method reduces the amount of calculation, ple through accumulator. Zhang et al. successfully segmented
but it is only applicable to distance image. Sappa et al. pro- pipeline point cloud in a process industry plant by fitting a
posed a method to extract point cloud boundary from binary column using the HT of 3D [8]. Han proposed a fast Hough
edge graph, which can segment point cloud data with inho- transform based on scanned line to achieve the separation
mogeneous density quickly [3]. The edge-based method can of spherical buildings from building roofs [9]. RANSAC is
achieve fast segmentation of point cloud data, and can achieve another model fitting method, and the algorithm mainly con-
better results when the segmentation scene is relatively sim- sists of two steps. The first step is to generate hypotheses
ple, but the method is sensitive to noise and density, and is from random samples, and the second step is to verify the
not suitable for dense or large area point cloud data sets. hypotheses. Jun proposed a GPU-based RANSAC algorithm,
However, since the wire is attached to the equipment body, which can segment planar model quickly and accurately [10].
the edge-based approach cannot perform effective segmen- However, the composition of substation equipment is com-
tation. plex, so the method based on model-fitting can not achieve
Region growth-based methods obtain multiple segmen- better results.
tation regions which differ significantly from each other by The second category is based on deep learning methods,
merging points or regions with similar features. Compared point cloud segmentation methods based on deep learning are
with edge segmentation methods, region growth methods widely used in unmanned vehicles, smart cities, etc. A classic
perform better for scenes with more noise. Besl first pro- method is to train a convolutional neural network using a 3D
posed the region growth segmentation algorithm in 1988 [4]. filter [11]. It can obtain good segmentation effect, however,
This algorithm mainly consists of two steps: firstly, the curva- this method needs data training, and the convolution kernel
ture information of each point in point cloud is extracted, and transforms from 2 to 3D, which requires high computing
a seed point is determined, and a curvature threshold is set, capacity and will greatly affect the segmentation efficiency.
then a selected region is grown from the seed point according Charles proposed the first neural network algorithm, Point-
to the region growth criterion. However, this method is noise- Net, for processing point cloud data [12]. It firstly projects
sensitive and time-consuming. To solve the problem of long point cloud from some specific perspectives instead of pro-
time-consuming of the algorithm, Yang Lin et al. proposed to cessing them directly. Secondly, it violently pools all points
voxelization point cloud and improved the growth efficiency into a global feature, and the local information is not learned
by representing points in point cloud with voxels during by the network of PointNet. The upgraded network, Point-
the region growth process [5]. However, this segmentation Net++, decreases the local information missing of Point Net,
method cannot segment the edge of point cloud. Vo et al. and solves the deficiency of Point Net extracting local fea-
used octree to organize point cloud and proposed a region tures [13].
growth algorithm based on octree, successfully segmenting The point cloud data of substation is huge, and the shape
building point cloud [6], but the algorithm is limited to the of a substation device is irregular. Moreover, the point cloud
objects with sharp edges and angles. Ning et al. divided seg- of auxiliary devices, such as cables and bases, is attached
mentation process into two steps. In the rough segmentation to the body of a device, which will make feature distinction
step, the algorithm estimated all points normal vectors and unclear. The methods mentioned above have a problem that
extracted seed points, and preliminarily segmented a point they cannot extract the features of auxiliary devices effec-
cloud of scene and got several planes based on the region tively. In order to solve the problem, this paper proposes a
growth method. The second step is finely segmentation based method to segment the point cloud of the cables attached to
on the residuals of points [7]. The method of region growth a substation device by using two kinds of features, one is the
is highly dependent on selected seed points, and inaccurate spatial location information of the device and the cable, and
selection of seed points will affect the segmentation result the other is the shape features of the cable itself.
and may lead to under- or over-segmentation.
The core idea of model-fitting-based segmentation meth-
ods is to fit point cloud to different geometric models, such as
cylindrical, square, spherical and planar. These methods can

123
Automatic cables segmentation from a substation device based … Page 3 of 17 9

Fig. 1 Substation equipment point cloud

2 Proposed method local features of the point cloud (see details in 3.2.4). Only
linear points will be split off in this step. Combining these
The data obtained by LiDAR scanner is a collection of mas- two steps, the complete segmentation of the auxiliary cables
sive points with three-dimensional coordinate information point cloud of the substation device is realized.
(x, y, z) [14], the collection of these points is called point
cloud, point cloud is a mass point collection of target surface
characteristics, as shown in Fig. 1. According to the char-
acteristics of point cloud, the original point cloud set of a
3 Cable segmentation
substation device can be cut into disjoint subsets.
3.1 Cable existence form
The spatial characteristics of the cables attached to sub-
station equipment are analyzed as the following. In the
The substation device point cloud data used in the paper was
substation scene, the position of the substation equipment
provided by Henan Teng long Information Engineering Co.
is fixed, and the equipment connects with each other by the
Ltd. The point cloud data obtained by a three-dimensional
cables. The cables mainly play the role of connecting equip-
laser scanner is shown in Fig. 3. By observation, it can be
ment, generally distributed above the equipment insulators.
found that there is a big difference in spatial location between
What’s more, the cables are far from ground. In addition, it
the cable point cloud and device point cloud, and the points
is not difficult to find that the cables have distinctive prop-
contained in the point cloud of the cable are linear points, and
erties, comparing with substation equipment, which include:
the points of the main device point cloud are mostly surface
(1) the cable structure is smaller than that of the substation
points and body points. (See details in Sect. 3.2.4).
equipment; (2) the cables are linearly distributed in three-
dimensional space, generally straight or curved.
According to the above-mentioned properties of cables 3.2 Cable segmentation process
distinguished from the body of substation equipment, a seg-
mentation method based on point cloud shape features is According to the segmentation method based on point cloud
proposed in this paper. The flowchart of the segmentation shape features proposed in this paper, the segmentation
algorithm of this paper is shown in Fig. 2. First, taking into process is mainly divided into two stages: the coarse seg-
account the huge and disorderly characteristics of cloud data mentation and the fine segmentation stage.
of a substation device, the paper applies octree coding to the The main task of coarse segmentation stage is to remove
original point cloud, so that it can be simplified and denoised the wire point cloud data away from substation equipment
to improve segmentation efficiency. The method divides the body and determine the wire endpoints. This stage is mainly
segmentation process into two steps: in the rough segmen- divided into three steps: (1) Determine the central axis posi-
tation step, according to Mean Shift and cylinder growth tion l of equipment based on Mean Shift algorithm; (2)
algorithms, the simplified point cloud is roughly segmented, Cylinder h is initialized with l as the center and an ini-
part of which is removed, and the end of a cable attached tial radius, r. The increment of radius is r . Recording the
to the substation device is located. In the fine segmentation change ratet of the number of points falling into the space
step, this algorithm combines the shape features of the point between two adjacent cylinders when the radius r increases,
cloud and the improved region growth algorithm, the point and finding the initial cutting position of the device point
cloud been roughly segmented is finely segmented. Unlike cloud according to t; (3) Use DBSCAN to cluster the cables
rough segmentation step, fine segmentation stage considers endpoints.

123
9 Page 4 of 17 Q. Yuanet al.

Fig. 2 Segmentation process

Fig. 3 Existence forms of cables of a substation device point cloud

To achieve complete segmentation of the cable data, the small cubes, each of which corresponds to a unique code. To
fine segmentation stage first performs principal component denoise, for each small cube, if the number of points in the
analysis on all candidate points to obtain the linear points cube is less than 6, all points in the cube are deleted. If the
from points of cables terminal, then uses the linear points of number of points in the cube is greater than 6 and less than
the cables terminal as the initial seed points, and the shape 15, all points will be retained. If the number of points in a
characteristics of point cloud as the region growth condition, cube is greater than 15, only the points closest to the center
to further segment the wire point cloud using the improved of the cube will be retained.
region growth algorithm. The steps of point cloud segmenta-
tion algorithm for the auxiliary cables of a substation device
are shown in Fig. 4.

3.2.2 Rough segmentation of cables


3.2.1 Data preprocessing
Locate the central axis of a device based on Mean Shift
The point cloud data obtained by LiDAR scanning is huge algorithm Before removing the part of cable point cloud
and contains a large number of noise points, which seri- far from the main body of a device, it is necessary to locate
ously affects subsequent processing efficiency and features the position of the central axis of the device. Therefore, the
extraction. This paper uses octree method [15–18] to encode origin cloud data is firstly stratified in the direction of z-axis,
original point cloud data set, and sets the number of encod- then the average value of all point coordinates of each layer
ing layers to 5. The point cloud is divided into a number of is taken as the center of gravity Q i of point cloud of each

123
Automatic cables segmentation from a substation device based … Page 5 of 17 9

Fig. 6 Equipment center movement process (n = 50)

device and attached cables, the projection centers of gravity


points of all layers are not completely coincident. Mean Shift
algorithm can converge at the maximum local density point,
so it is used to locate the center axis position of the device.
Fig. 4 Flow chart of the overall segmentation algorithm The main idea of Mean Shift algorithm is to calculate the
offset mean vector of current coordinate points, then move to
a new clustering center according to the vector, and it is taken
as a new starting point to operate continuous iteration until
the required conditions are met, which is shown in Fig. 6.
The O1 in Fig. 6 represents the starting point to search
the target, O3, which is the xoy projection of the central
axis of a device, and C1 and C2 are the monitoring frames
to search the target. The process of searching the target is
as follows: First, the average vector of the vectors from O1
to the points within the monitoring frame is calculated, M,
which is the drift direction of searching. Then, a new drift
position is calculated by O2 = O1 + M, and the iteration is
carried out continuously until it is close to O3. The specific
steps include:

1) Calculate the mean of the offset vector M. As shown in


Fig. 5 The layered center of gravity projection of equipment (n = 50) Fig. 6, in a two-dimensional circular region, if set the
radius of a circle to r, there will be N target points in this
circle. Then select any point x in the region, the formula
layer,
of Mean Shift vector is defined as follows:

1 
ni
Qi = (x j , y j , z j ), (1) 1 
ni Mh = (xi − x). (2)
j=1 k
xi ∈Sk

where i is the number code of layers, ith, and n i corresponds 2) Introduce kernel function. The Gaussian kernel function
to the number of points of the ith layer, and (x j , y j , z j ) is is the most commonly used density function, and the
the coordinate of a point in the point cloud. specific definition formulation is
After locating the center of gravity Q i of each layer, all the
centers of gravity points are projected to xoy plane, which
is shown in Fig. 5. Due to the peculiar shape of a substation K (x) = ck, d k(x2 ). (3)

123
9 Page 6 of 17 Q. Yuanet al.

Fig. 8 Cylinder expansion in a two-dimensional plane


Fig. 7 Cylinder growth process

distribution is more uniform compared with that of the device


3) Substitute Formula (3) into Formula (2). The calculation body points. The cutting position, D, as shown in Fig. 8,
formula of the Mean Shift algorithm offset vector mean which is the space between h i and h i+1 is the starting position
is as follows: from which all the points far away from the center axis of
the device are removed. In order to preliminarily locate the
  
ck, d n  x − xi 2 cutting position, an array A is used to store the number, n i ,
Mh = k 
 h  .
 (4) the ith of the radius r of the cylinder increasing, i = 0,1….
nh d i=1
Then the rate of change of the points falling into the space
between h i+1 and h i , t, is counted as follows:
By Gaussian kernel function transformation, it can be seen
that the target will move to the region with rising probability
n i+1 − n i
density, and the specific location of the cluster center as O3 ti = × 100%. (5)
is found finally. N

Where N is the size of point cloud. Use an array B to store ti ,


Locate cutting position based on cylinder growth After
and find the rough cutting point position by the change rate ti .
using Mean Shift algorithm to find the center axis of the
If the change rate ti is less than the threshold ω, what’s more,
main body of a device, the cylinder growth method is used to
if the successive change rates after the ith rate are all less
remove the cable points data far from the center axis of the
than the threshold ω, the initial substation equipment point
device. The process of cylinder growth is as follows. First,
cloud is preliminarily segmented at the cylinder h i according
take the center axis of the device, l, as the center axis of a
to this judgment. Namely, if the distance between a point and
cylinder and mark the cylinder as h 0 . The initial radius of
the center axis is more than the ri , the point will be removed,
the cylinder h 0 is set as r0 = 0.1, and the radius increment
and the cutting position, D, is the shaded part between h i and
of the cylinder is r . Second, the radius of the cylinder h 0
h i+1 .
increases with the increment, r , successively. Namely, the
radius of the (i + 1)th increasing cylinder h i+1 can be got
by that of h i : ri+1 = ri +r . Then, record the number of the 3.2.3 DBSCAN algorithm finds the number of cables
device data points falling into the space between h i and h i+1 ,
n. At the same time, if the difference between the number of Due to the wide variety of substation equipment, the loca-
device data points falling into the cylinder, h i , and that of all tion and number of cables attached to equipment are different
data points of the device is less than the threshold, the radius from each other. As shown in Fig. 9, the above rough segmen-
growth of the cylinder ends. Because substation equipment tation process cannot obtain the number of cables attached to
must be installed vertically, and the 3D laser scanner’s coor- the equipment, and the number of cables relies on the number
dinate system is perpendicular to the earth, the central axis of cable ends, so the cable ends number needs to be got after
of the cylinder is perpendicular to the x–y plane. Figure 7 locating rough cutting position.
shows the process of cylinder growth. In order to find the number of cables ends attached to a
After the expansion of the cylinder is completed, due to device, DBSCAN algorithm is used to cluster the rough cut
the large difference of point density between the cable and ends of the cables attached to the device.
the device body, the number of the cable points that fall into DBSCAN is an unsupervised clustering algorithm based
the space between the cylinder h i and h i+1 is fewer and the on density space without knowing the label of data sample

123
Automatic cables segmentation from a substation device based … Page 7 of 17 9

Fig. 9 Equipment with different numbers of cables

values, and intensity values. According to the different shapes


of point cloud, they can be classified into three categories:
line point cloud, area point cloud, and body point cloud [20,
21], and the point of these point clouds are called linear point,
area point, and body point, respectively. As shown in Fig. 11
To classify a point p based on PCA, the K neighborhood
of the point p should be obtained first, then the covariance
matrix, M, should be constructed:


k
M= (xi − p)(xi − p)T , (6)
i=1

Fig. 10 Initial location of equipment cutting position by cylindrical where xi = (xi yi z i ), i = 1, 2, 3, . . . , k is the k-
growth method neighborhood of the point p.
Next, the eigenvalues of the covariance matrix, λ1 ≥ λ2 ≥
λ3 ≥ 0, is calculated, as wellas its corresponding eigenvec-
in advance [19]. The algorithm takes a region with suffi- v1 , −
tors −
→ →
v2 , −

v3 . Let σ j = λ j , ∀ j ∈ [1, 3] represent the
cient density as a distance center, and continuously grows standard deviation. PCA can retrieve the three main direc-
the region, then connects adjacent regions with sufficient tions of the k-neighborhood of the point p, and their sizes
density. It can effectively process abnormal data, which is can be described by the eigenvalues. The eigenvalues are
mainly used for clustering spatial data. mainly used to derive three geometric features:
After using Mean Shift algorithm and the cylindrical
growth method for preliminary segmentation, the points in σ1 −σ2 σ2 −σ3 σ3
the rough cutting position, D, are clustered in some clusters a1D = , a2D = , a3D = , (7)
μ μ μ
of points. A cluster of points can be seen as a cable end, so
the number of the clusters of points is equal to that of the where μ = σ1 , and a1D , a2D , a3D represents the probability
cables. As shown in Fig. 10, the device contains 4 clusters of that the point p belongs to the three shapes feature, respec-
points, and there is no adhesion between them, so DBSCAN tively, and a1D , a2D , a3D ∈ [0, 1]. From Formula (7), the
algorithm can be used to find the number of cable ends. shape attribute of point p can be further deduced as:

3.2.4 Shape characteristics of point cloud d p = arg max[ad D ], d ∈ [1, 3]. (8)

A point cloud is a data set of points in a specific coordinate If σ1  σ2 and σ3 ≈ 0,a1D is greater than a2D and a3D .
system. The points contain a wealth of information, including The shape labeling d p is approximately equal to 1, and the
three-dimensional coordinates X, Y, Z, colors, classification linear shape will be universal. If σ1 , σ2  σ3 ≈ 0, the point

123
9 Page 8 of 17 Q. Yuanet al.

Fig. 11 Different spatial structures of point clouds

p is a plane point. What’s more, σ1 ≈ σ2 ≈ σ3 implies the 1. First, an empty seed stack A is established. For device
area which is a scattered area. point cloud B, its corresponding cable end is di, i = 1, 2,
The main purpose of Sect. 3 is to remove the point cloud 3, . . . , n, where n is the number of cables. Select the
of cable, and the most points of cable point cloud are linear device data point closest to the cable end di as the initial
points, so the point which is linear point can be used as a seed point, P(x0 , y0 , z 0 ), and push P into the stack.
feature for fine segmentation. 2. Pull point P out of the seed stack A, and obtain each point
Pi in the neighborhood of the point P. If Pi is a linear
3.2.5 Fine segmentation based on improved region growth point based on the shape characteristics of point cloud,
algorithm push it into the seed stack A. Pi continues to grow as a
child seed point, and vice versa.
The rough segmentation cannot guarantee that the device 3. Traverse seed points in the seed stack A. Repeat Step (2)
body does not contain any cable point cloud. The cylin- until no seed points appear or the seed stack A is empty.
der growth method can effectively remove some cable point 4. Repeat the above steps until the cable data points near
cloud data quickly, but the cylinder still contains part of the each cable end is removed.
cable point cloud data. The existence of this part of cable The above steps can be shown in Fig. 12.
point cloud will affect the subsequent identification of equip-
ment, so it is necessary to separate this part of point cloud
from the equipment body. Compared with rough segmen-
tation, fine segmentation performs further segmentation of 4 Experimental result and analysis
equipment to ensure that the cable point cloud is completely
removed, which is a more challenging fine-grained task. The experiment of this paper was performed in MAT-
The initial seed point selection and regional growth rules LAB2020a. To validate the automatic segmentation results,
of the traditional model-based regional growth method are the point cloud data of substation equipment were provided
generally based on the geometric information of point cloud by Henan Tenglong Information Engineering Co. Ltd. There
data, such as the normal vector and curvature of data points were 303 test samples, consisting of more than 10 types of
[18, 22]. However, due to the irregular shape of substation electrical devices, such as circuit breaker, transformer and
equipment, it is difficult to obtain good segmentation effect voltage transformer.
by using normal vector or curvature as regional growth con- The experiment is divided into three modules as follows:
ditions, unlike tables and chairs which have obvious plane
information. Therefore, this paper proposes an improved 1. Point cloud data reducing and denoising.
regional growth algorithm to further remove cable point 2. Use Mean Shift and cylinder growth algorithms to per-
cloud. form rough segmentation.
When performing fine segmentation of the point cloud 3. Calculate the number of cables ends, and use the shape
of the auxiliary cables of substation equipment, the selec- feature of the point cloud as the condition for region
tion of seed points directly affects the segmentation effect. growth to perform fine segmentation.
Generally, there are two ways to select seed points: random
selection and manual selection, which have a great impact on
results. Because point cloud data contains many noise points, 4.1 Evaluation criteria
the results are difficult to converge. Therefore, the point cloud
data, which is closest to rough cutting position and belongs to In this paper, the algorithm running time, segmentation accu-
linear point, is selected as the initial seed point in the exper- racy and precision are used as evaluation criteria for wire
iment. The main steps to improve regional growth algorithm segmentation effect. The running time of algorithm can be
are as follows: completed by a timer; and the essence of wire segmentation

123
Automatic cables segmentation from a substation device based … Page 9 of 17 9

Table 1 Comparison of Mean Shift and RANSAC algorithms

Equipment type Locate the center axis


effectively

Mean shift RANSAC

kV500_CP_C_1 Yes Yes


kV500_CP_D_7 Yes Yes
kV500_LA_E_1 Yes No
kV500_PT_E_1 Yes No
kV500_CP_D_5 Yes Yes
kV500_DS_3D_2 Yes No
5kV500_DS_3DD_2 Yes No
… … …
Correct/total 50/50 24/50

along Z-axis. Then the gravity center of each layer is calcu-


lated and projected onto a X–Y dimensional plane. By a lot
of experiments, the center axis of the device can be located
accurately by Mean Shift algorithm if the number of layers,
n, equals 50.
RANSAC algorithm is a universal method used in point
cloud, and can be chosen to locate the center axis of a device.
The paper’s method was compared with RANSAC algorithm,
and the results are shown in Fig. 13.
The comparison results of RANSAC and Mean Shift algo-
rithms are shown in Table 1. It shows that the accuracy of
Mean Shift algorithm can reach 100%, and Mean Shift algo-
rithm obtain the best result. Therefore, Mean Shift algorithm
Fig. 12 Flow chart of improved region growth algorithm is used to locate the center axis of a device in the paper.
The accuracy of RANSAC algorithm is only 48%.
RANSAC is an uncertain algorithm, and it has good results
is to remove the wire point cloud from substation equipment.
if it is used for fitting cylinder to locate the center axis of a
According to the essence, this paper defines the segmentation
device, which has only one single base. But a device usually
accuracy and segmentation precision as shown in Eqs. (9) and
has more than one base, RANSAC algorithm can’t effectively
(10).
recognize the number of bases. Therefore, the position of the
TP + TN center axis of a device cannot be effectively located, which
accuracy = × 100%, (9) leads to a low accuracy.
TP + FN + FP + FN
After the center axis of a device is located, the cylinder
TP growth method is used to locate the rough cutting position.
precision = × 100%, (10)
TP + FN First, the rate of change of the number of point cloud data
points, ti , is obtained by counting the number of the points
where TP, TN, FP, and FN denote True Positive, True falling into the space between adjacent cylinders, h i and h i+1 ,
Negative, False Positive, and False Negative, respectively when radius grows. Secondly, if the values of ti of sequential
TP + FN + FP + FN = N . and N is the total number of indexes are less than the threshold ω, the cylinder radius
samples [23]. corresponding to the first index is taken as the initial cutting
position, ri . kV500_DS_3D_2 was taken as an example, and
4.2 Rough segmentation experiment comparison the results are shown in Fig. 14.
Mean Shift algorithm and the cylinder growth method can
The experimental objects are 50 electrical devices attached remove part of the point cloud data of cables rapidly and accu-
with cables. First, a point cloud data of a device is layered rately. After extensive experimental verification, threshold ω

123
9 Page 10 of 17 Q. Yuanet al.

Fig. 13 Comparison of RANSAC and Mean Shift algorithms

Table 2 Rough segmentation result feature distinction is not obvious, therefore, the rough seg-
mentation algorithm of this paper cannot remove any of the
Equipment type Segment the cable effectively
point cloud data of the cable. Since the number of a device
kV500_CP_C_1 Yes with cables located in the center of it is very few, the algo-
kV500_CP_D_7 Yes
rithm can remove the cable data around the device effectively
with the correct segmentation rate of 98%.
kV500_LA_E_1 Yes
kV500_PT_E_1 Yes
kV500_CP_D_5 Yes 4.3 Fine segmentation experiment
kV500_DS_3D_2 No
5kV500_DS_3DD_2 Yes The shape information of point cloud is used as the condition
… … for region growth. In determining the shape characteristics
Correct/total 49/50 of point p, it is necessary to obtain its k-neighborhood points.
Different values of k may lead to different results. This paper
selects different values of k for experiments so as to obtain the
optimal value of k. Figure 16 shows the results variation with
the k-selection changing for different substation equipment,
is set to 0.16%. The experiment used 50 groups of data for and the y-axis represents the number of the moved point of
testing, and the result is shown in Table 2. cables. By compared with manually segmented wire points,
The result shows that only kV500_DS_3D_2 was not suc- 35 is selected as the optimal k value.
cessfully segmented, as shown in Fig. 15. It is because that CP_C_3 is taken as an example to show the fine segmen-
the position of the cables is in the center of the device, and the tation process, and the result is shown in Fig. 17.

123
Automatic cables segmentation from a substation device based … Page 11 of 17 9

Fig. 14 Rough Segmentation Process

Fig. 15 Invalid segmentation

Fig. 16 The number changing of the moved point of cables with


k increasing
It can be seen from Fig. 15 that the rough segmentation
algorithm is accurate in locating the cables ends of the equip-
ment point cloud CP_C_3, and has achieved very good results

123
9 Page 12 of 17 Q. Yuanet al.

Fig. 17 Fine Segmentation Process

Table 3 Statistics of cable remove result

No Equipment type Remove successful Points number Execution time Accuracy Precision

1 kV500_CP_C_1 Y 1521 20.38 s 99.93% 99.61%


2 kV500_CP_D_7 Y 641 7.35 s 97.42% 96.80%
3 kV500_LA_E_1 Y 3815 17.23 s 98.36% 97.85%
4 kV500_PT_E_1 Y 1106 1.7 s 100% 100%
5 kV500_CP_D_5 Y 1771 3.7 s 97.24% 96.74%
6 kV500_PT_E_6-4 Y 1943 35.03 s 99.01% 96.25%
7 5kV500_DS_3DD_2 Y 2585 16 s 97.99% 96.97%
… … … … …
Correct/total 289/303

123
Automatic cables segmentation from a substation device based … Page 13 of 17 9

Table 4 Comparison of the performance of different segmentation methods

Method Criterion Equipment type

CP_C_3 CP_D_6 ES_A_1 PT_E_1

Segmentation based on line features [25] Precision 80.26% 89.00% 91.17% 100%
Accuracy 97.70% 92.23% 95.81% 98.75%
Time consumption 26.25 s 10.91 s 34.72 s 13.94 s
Segmentation based on conical search [24] Precision 89.00% 90.73% 92.81% 99.12%
Accuracy 98.66% 94.66% 94.28% 98.56%
Time consumption 3.78 s 8.63 s 46.88 s 13.26 s
Method of this paper Precision 99.61% 96.80% 97.85% 100%
Accuracy 99.93% 97.42% 98.36% 100%
Time consumption 2.56 s 2.4 s 31.19 s 1.7 s

in the segmentation of substation equipment point cloud by to compare the effects of the three methods. The experi-
using shape features as the conditions for region growth. A ment takes segmentation accuracy, segmentation precision
total of 303 groups of equipment containing cables were and segmentation time as the evaluation criteria. Some results
selected for segmentation experiments, and the segmenta- are shown in Table 4, as well as in Fig. 18.
tion accuracy can reach 95.34%, the result is shown in Table
3.
The experimental result shows that the fine segmentation
stage can achieve complete segmentation of the equipment- 4.5 The necessity of splitting cables
attached wire point cloud. The algorithm segmentation
success rate can reach 95.34%, and the segmentation pre- This Section focuses on the effect of the presence of wires
cision of each substation equipment-attached wires is above on identification of substation equipment. In this paper, a
96%. point cloud recognition method based on plane features is
used. In the recognition process, the optimal planar features
and key point information of templates are firstly obtained to
establish a template feature library. Then the optimal planar
4.4 Comparative test features of equipment extracted by RANSAC algorithm are
pre-selected by Umeyama algorithm to realize preliminary
In order to prove the advancement of the proposed wire seg- selecting of possible corresponding templates of the equip-
mentation algorithm, we compared the paper’s algorithm ment. Finally, the key point information of the equipment
with the algorithm based on conical search proposed by point cloud is precisely matched by improved ICP algorithm
Wang [24] and the algorithm based on line features proposed to realize precise recognition of the equipment. A total of
by Arastounia [25]. Using the algorithm based on conical 303 groups of substation equipment of different types and
search to segment wire, it first needs to continuously rotate models containing wires were selected for testing, and the
substation equipment to detect the position of the equipment- experimental results are shown in Table 5.
attached wires ends. After the ends are detected, the wires From Table 5, it can be seen that the equipment identifi-
points are clustered along a forward search direction using the cation effect is not good without removing wire point cloud,
K-Means algorithm. At the same time, the search direction 42 substation devices will not be correctly identified, and the
needs to be adjusted. The clustering process will terminate if recognition rate is only 86.14%. After removing the wires,
the equipment body is searched. As for the segmentation the recognition rate is improved obviously. The main reason
method proposed by Arastounia, the point cloud in local is that the presence of wires will affect the acquisition of key
neighborhood is analyzed first. If the points in local neigh- point information. When acquiring the key point information
borhood are linearly distributed, they are considered to be of point cloud, the key point extractor will extract some points
wire points. Finally, the wires are segmented after all the from the wires at the same time, which will cause a big dif-
wire points are obtained by traversing the equipment point ference between the key point information of test equipment
cloud. In this paper, the 303 sets of test equipment provided and corresponding template, thus causing identification fail-
by the cooperative company and the 64 sets of equipment ure. Therefore, it is very necessary and essential to segment
separated from substation site scenic point cloud are used attached wires before substation equipment identification.

123
9 Page 14 of 17 Q. Yuanet al.

Fig. 18 The visual performance of different segmentation methods. a, b Original equipment, c, d The segmentation effect based on this paper’s
method, e, f The segmentation effect based on conical search, g, h The segmentation effect based on line features

123
Automatic cables segmentation from a substation device based … Page 15 of 17 9

Table 5 Unsegmented wire and


segmented wire recognition No Template Test samples Recognition successfully
effect comparison
Unsplit wire Split wire

1 CB_C CB_C_1 Y Y
9 ES_A ES_A_2 N Y
17 CP_D CP_D_4 Y Y
26 LA_E LA_E_3 Y Y
35 LA_G LA_G_2 N Y
76 CP_DDD CP_DDD_1 N Y
90 CP_E CP_E_1 Y Y
131 CP_I CP_I_3 N N
237 CT_AA CT_AA_5 Y Y
300 ZUBOQI_D ZUBOQI_D_10 Y Y
303 DS_1B DS_1B_1 Y Y
Correct 261 280

5 Conclusion 5. Yang, L., Zhai, R.: Segmentation of plant organs point clouds
through super voxel-based region growing methodology. Comput.
Eng. Appl. 55(16), 197–203 (2019)
In this paper, Mean Shift and a cylinder growth algorithm
6. Vo, A.V., et al.: Octree-based region growing for point cloud seg-
are combined to segment the point cloud of substation equip- mentation. ISPRS J. Photogramm. Remote. Sens. 104, 88–100
ment initially, and an improved region growth algorithm is (2015)
proposed to realize the fine segmentation of the point cloud of 7. Ning, X., Zhang, X., Wang, Y., Jaeger, M.: Segmentation of archi-
tecture shape information from 3D point cloud. In: Proceedings of
the cables attached to a substation device. First, octree coding
the 8th International Conference on Virtual Reality Continuum and
is used to realize reduction and denoising of point cloud data. Its Applications in Industry, pp. 127–32 (2009)
Secondly, Mean Shift and the cylinder growth algorithms are 8. Zhang, J., Zhuheng, Lu.: Detection and recognition of pipeline
used for rough segmentation, and density clustering is used to point cloud based on feedback Hough transform. Comput. Mod-
ernization 08, 6–11 (2019)
obtain the number of cables ends. Finally, the region growth
9. Han, T.: Extraction of LiDAR dome houses based on Hough trans-
algorithm uses the shape features of point cloud as the condi- form. Sci. Technol. 2016;26(10):122–125+146
tion to finely segment the remaining point cloud set. By the 10. Lan, J.: Research and Implementation of Real-Time Plane Detec-
experiment of 303 different electrical devices, it comes to a tion Algorithms in LiDAR Point Clouds. North China University
of Technology, Beijing (2019)
conclusion that the algorithm of this paper can realize the 11. Rethage, D., Wald, J., Sturm, J., et al.: Fully-convolutional point
segmentation of cables effectively. It greatly reduces labor networks for large-scale point clouds. In: 2018 European Confer-
force and improves the recognition accuracy of a device, and ence on Computer Vision, pp. 625–640. Springer, Munich (2018).
plays an important role in the reconstruction of substation. 12. Charles, R.Q., Hao, S., Kaichun, M., et al.: PointNet: deep learn-
ing on point sets for 3D classification and segmentation. In: 2017
Acknowledgements This work was supported by National Natural Sci- IEEE Conference on Computer and Pattern Recognition, pp. 77–85
ence Funds of China (62173309). (2017). IEEE, Honolulu
13. Qi, C., Li, Y., Su, H., Guibas, L.: PointNet : Deep Hierarchical
Feature Learning on Point Sets in a Metric Space. ArXiv.org (2017)
14. Xiaochao, W., Xiuping, L., Baojun, Li., Shaoguang, Z.: Feature
detection on point cloud via local reconstruction. J. Comput.-Aided
References Des. Comput. Graph. 25(05), 659–665 (2013)
15. Min, L., Yun, M.: Study on point cloud management strategy based
1. Sipiran, I., Bustos, B., Schreck, T.: Data-aware 3D partitioning for on Octree-like index. Laser & Optoelectronics Progress, pp. 1–17
generic shape retrieval. Comput. Graph. 37(5), 460–472 (2013) [2020-10-29].
2. Jiang, X.Y., Meier, U., Bunke, H.: Fast range image segmentation 16. Yubing, W.: Scattered point cloud data compression based on octree
using high-level segmentation primitives. In: Proceeding. Third coding. Eng. J. Wuhan Univ. 53(08), 734–739 (2020)
IEEE Workshop on Applications of Computer Vision. WACV’96, 17. Kang, Yi., Yuting, Z., Jian, P.: Implementation of 3D object classifi-
pp. 83–88 (1996). cation algorithm based on the octree representation. Telecommun.
3. Sappa, A.D., Devy, M.: Fast range image segmentation by an edge Inf. 04, 38–41 (2019)
detection strategy. In: Proceedings Third International Conference 18. Lv, H., Fu, J., Wang, H., et al.: Fast retrieval method of three-
on 3-D Digital Imaging and Modeling, pp. 292–299 (2001) dimensional indoor map data based on octree. J. Comput. Appl.
4. Besl, P.J., Jain, R.C.: Segmentation through variable-order surface 39(01), 82–86 (2019)
fitting. IEEE Trans. Pattern Anal. Mach. Intell. 10(2), 167–192
(1988)

123
9 Page 16 of 17 Q. Yuanet al.

19. Sun H., Sun X., Li H.: 3D point cloud model segmentation based Jing Chang was born in January
on K-means cluster analysis. Comput. Eng. Appl. 42(10), 42–45 12, 1995, in Zhengzhou, Henan
(2006) Province. She received the mas-
20. Achanta, R., Shaji, A., Smith, K., et al.: SLIC superpixels compared ter’s degree in Control Science
to state-of-the-art super pixel methods. IEEE Trans Pattern Anal. and Engineer from Zhengzhou
Mach. Intell. 34(11), 2274–2282 (2012) University in 2022. Her research
21. Demantké, J., Mallet, C., David, N., et al.: Dimensionality based interests include point cloud
scale selection in 3D LIDAR point clouds. ISPRS Int. Arch. Pho- process, segmentation, and recog-
togramm. Remote Sens. Spatial Inf. Sci. 38(5), 97–102 (2012) nition of electrical equipment.
22. Martin, E., Hans-Peter, K., Jörg, S., et al.: Adensity-based algo-
rithm for discovering clusters in large spatial databases with noise.
In: Proceedings of the Second International Conference on Knowl-
edge Discovery and Data Mining (KDD-96), pp. 226–231. AAAI
Press. CiteSeerX 10.1.1.121.9220. ISBN 1-57735-004-9
23. Fawcett, T.: An introduction to ROC analysis. Pattern Recogn. Lett.
27(8), 861–874 (2006)
24. Wang, P.: Research on Segmentation Algorithm of Auxiliary Facil- Yong Luo received the bachelor’s
ities of Substation Equipment Based on 3D Point Cloud. [Master’s degree in process and equipment
thesis]. Zhengzhou University, Zhengzhou (2019) of machinery manufacturing and
25. Arastounia, M., Lichti, D.D.: Automatic object extraction the MS degree in mechatronic
from electrical substation point clouds. Remote Sens. 7(11), engineering from Yanshan Uni-
15605–15629 (2015) versity, Qinhuangdao, China, in
1999 and 2002, respectively, and
the PhD degree from Nankai
Publisher’s Note Springer Nature remains neutral with regard to juris- University, Tianjin, China, in
dictional claims in published maps and institutional affiliations. 2005. He was a visiting scholar
in UWM, Milwaukee, WI, USA,
Springer Nature or its licensor (e.g. a society or other partner) holds from 2015 to 2016. He is cur-
exclusive rights to this article under a publishing agreement with the rently a professor in Zhengzhou
author(s) or other rightsholder(s); author self-archiving of the accepted University, Zhengzhou, China.
manuscript version of this article is solely governed by the terms of such His research interests include
publishing agreement and applicable law. image process, computer control and management, and automation of
electrical equipment.

Qianjin Yuan was born in Tianlei Ma received his


February 1, 1995, in Zhengzhou, BE degree in automation
Henan Province. He received from Zhengzhou University,
the undergraduate degree from Zhengzhou, China, in 2010. He
Henan University of Technology received his PhD degree in Uni-
in 2013 and the master’s degree versity of Chinese Academy of
in Control Science and Engineer Sciences, Beijing, China. He is
from Zhengzhou University in currently an associate profes-
2021, respectively. His research sor in the School of Electrical
interests include point cloud and Information Engineering,
process, 3D modeling, segmenta- Zhengzhou University, China. His
tion, and recognition of electrical current research interests include
equipment. infrared dim target detection,
image processing, and pattern
recognition.

123
Automatic cables segmentation from a substation device based … Page 17 of 17 9

Dongshu Wang received his


bachelor’s degree in Mechani-
cal Manufacture Technique and
Equipment in 1996, master’s
degree in Mechanical Manufac-
ture and Automation in 2002,
and PhD in Control Theory and
Control Engineering in 2006
from Northeastern University,
China. His research domain is
autonomous mental development,
artificial intelligence, machine
learning, robot intelligent control.
Currently, he is an associate
professor in School of Electrical
Engineering, Zhengzhou University, Zhengzhou, China.

123

You might also like