You are on page 1of 5

High-speed Railway Rod-insulator Detection

Using Segment Clustering and Deformable Part


Models
1 1 2 1 3
Ye Han , Zhigang Liu , Dah-Jye Lee , Guinan Zhang , and Miao Deng

1
Southwest Jiaotong University, Chengdu City, Sichuan Province, China
2
Dept. of Electrical and Computer Eng., Brigham Young University
Provo, Utah, USA
3
State Grid Sichuan Electric Power Company, Chengdu City,
Sichuan Province, China
Abstract— Catenary system maintenance is an important task must be performed to avoid the interruption of railway services
to the operation of a high-seed railway system. Currently, the and prevent devastating damages to the railway system.
inspection of damaged parts in the catenary system is performed Methods and tools used to detect faulty insulators include
manually, which is often slow and unreliable. This paper proposes corona cameras, thermal imaging cameras [1], ultrasonic
a method to detect and locate the rod-insulators in the image sensors, electric field measurement and leakage current
taken from the high-speed railway catenary system. Sub-images
detection [2]. Many image processing-based methods have
containing bar-shaped devices such as cantilever, strut, rod, and
pole are first extracted from the image. Rod-insulator is then been developed. For the detection of the insulators, Wu et al.
recognized and detected from these bar-shaped sub-images by proposed an active contour model based on texture distribution
using deformable part models and latent SVM. Experimental for insulator segmentation [3]. Their method achieves great
results show that the proposed method is able to locate segmentation result but is relatively slow. The processing of a
rod-insulators accurately from the catenary image for the 560 × 201 image takes 30.85 seconds [3]. Oberweger et al.
subsequent detect inspection process. The robustness of this presented an insulator detection method based on local
method ensures its performance in different imaging conditions. gradient-based descriptors and a voting scheme for insulator
localization [4]. Liao et al. proposed a multi-scale and
Index Terms—high-speed railway, rod insulators, segment multi-feature local feature descriptor, which combines SPIN
clustering, deformable part models,
descriptor [5] and RIFT descriptor [6] in a unified framework to
handle the scale variation and rotation of the insulator [7]. Li et
I. INTRODUCTION
al. used the improved MPEG-7 EHD (edge histogram method)
The catenary system that supplies electric power to technique to recognize insulator [8]. Zhao et al. adopted
locomotive is a crucial component of a high-speed railway. The modified Markov Random Field model to extract the repetitive
maintenance of the catenary system is traditionally carried out geometric structure of the insulator [9]. Cui et al. proposed a
by railway workers who manually search along the railway for method of segmenting the infrared insulator image using
damaged parts that need to be replaced. Systems have been improved Unit-linking PCNN algorithm [10]. Li et al. proposed
developed to record the video of the catenary system along the an insulator detection method based on profile projection [11].
railway to be reviewed manually. This manual inspection Pillars and poles form the main structure of the catenary
process is not efficient and does not meet the demand of rapidly system. In this work, the poles and pillars are detected using
expanding high-speed railway networks. The recognition of segment detection and segment clustering. The input image is
damaged catenary parts heavily depends on human observation. then decomposed into several bar-shaped sub-images for
Non-contact inspection methods based on image processing insulator detection based on the distribution of the pillar and
have drawn attention from railway companies in the past few pole. As for the detection of insulator, histogram of orientated
years. Computer is used to analyze images of the catenary gradients (HOG) feature is considered. HOG feature is a dense
system captured by HD digital cameras to recognize and locate local descriptive feature first proposed by Dalal and Triggs [12]
damaged parts. and has shown great effectiveness in pedestrian detection,
Rod-insulator is an important part of a catenary system. It is vehicle detection, and frontal face detection. However the
used to support other mechanical and electrical components detection rate of HOG feature drops when the object is
and insulate the strut from high voltage. Some occluded, which is very common in the catenary system. The
safety-threatening conditions may lead to insulation deformable part models, first proposed by Felzenszwalb at al
deterioration and raise the possibility of insulation breakdown based on HOG features. [13], on the other hand is capable of
in humid weather. For example, the ceramic discs of the handling intra-class variability by combining several
insulator may be broken and foreign matters may be stuck sub-models to allow a wider range of variations in object
between ceramic discs. Regular inspection and maintenance appearance. It performs well with occlusion.

‹,(((  ,&,3


The remainder of the paper is organized as follows. In
Section II, we briefly introduce how the catenary images are Bar-shaped support devices show up in the image as long
captured. In Section III, the image decomposition process is white lines. Hough transform was tested to detect these devices.
described. In Section IV, the deformable part model is It worked well but missed most thin poles. A five-step
discussed. In Section V, the detection accuracy is presented. algorithm as shown in Fig. 3 is developed to address this
challenge. The catenary image is first binarized by
Finally, we conclude this paper in Section VI.
thresholding. This is done without too much challenge because
the image contrast is fairly high. In the second step, the
II.IMAGE ACQUISITION
segments in the binary images are detected. The contours of the
Images of the catenary suspension system are captured by binary image are detected first to recognize the segments in the
HD cameras mounted on the top of the inspection vehicle for binary image. By using the topological analysis method
analysis. As the vehicle runs along the railway, the cameras are proposed in [14], sequences of the coordinates of the pixels of
triggered to capture images when a catenary pillar is detected. the contours can be derived. The inclination-difference of each
Fig. 1(a) illustrates the image acquisition setup. pixel in the contour sequence γk is defined as
In order to minimize the interference of the background and
simplify the process of image, catenary images are captured at γ k = sk + L − sk (1)
night. LED spotlights are utilized to provide proper ⎡y -y ⎤
illumination. The image resolution is 2448×2048 pixels, which where sk = arctan ⎢ k + H k ⎥ , ( xk , yk ) is the coordinate of the
is more than adequate for image processing. ⎣ xk + H - xk ⎦
k-th pixel of the contour sequence. If γk is below a threshold
value T, the k-th pixel of the contour sequence will be detected
as part of a segment. In Fig.4, the segments detected in the
image are marked with blue lines.
The spot and HD camera
Catenary Image

Step 1 Image Binaryzation

Border Following
Step 2
Segment Detection

Segment Clustering based on inclination


Fig. 1. The sketch of the imaging system. angle
Step 3
Segment Clustering based on midpoint
III. IMAGE DECOMPOSITION Connection

The catenary parts that need to be detected or localized Select one Cluster
include insulator, clevis, diagonal tube, and bracket. As shown
Minimal Bounding Rectangle Computng
in Fig.2, they are all installed at the end of the bar-shaped
support devices such as cantilever, strut, rod, and pole. The Step 4 Main Orientation Detection
locations of the catenary parts can be expected to be along the Obtain the Bar-shaped Sub-image by
straight lines representing these bar-shaped devices. In this Extending the Bounding Rectangle on the
paper, the bar-shaped sub-images corresponding to the Main Orientation to the boundary of the
Image.
bar-shaped devices are first detected and extracted from the
input image and are rotated to the horizontal orientation. The Sub-image Saving
program then searches for the insulators along the sub-images No
in the horizontal direction. This allows the classifier to only Step 5 No More Clusters?

recognize insulators in the horizontal direction, which Yes

improves the detection efficiency and accuracy. End

Fig.3 The main steps of the proposed method

Fig.2 The location of the bar-shaped devices (marked with green rectangles)
and catenary parts (four insulators are marked with red rectangles, a clevis is
marked with an orange rectangle). Only a few visible parts are labeled due to Fig. 4. Segment Detection Result
space limitation.


The segments are then clustered. Each cluster of the clustering is shown in Fig.7. The segments that correspond to
segments corresponds to a bar-shaped devise. There are two the same bar-shaped device are marked with the same color.
steps to cluster the segments. First, the segments are clustered The detail procedure of the mid-point connection clustering
based on their inclination angles. The detail clustering is shown in Fig.8.
procedure is shown in Fig.5. The result of the inclination angle
clustering is shown in Fig.6. The segments that have similar 1. Obtain the coordination of the mid-point of every
inclination angles are clustered into one cluster, and are marked segment.
with the same color. 2. Equally divide the connection lines between the
mid-points of every two segments into 20 parts
1. Compute the inclination angles of all the segments. 3. Repeat for ( n is the number of
2. Compute the absolute distances between any two segments)
segments ( , is the Create segment set
inclination angle of segment i). Repeat for
3. Repeat for , Set
Create segment set Repeat for
If the corresponding point of the end point of the
Repeat for ,
j-th unit in the binary image is non-zero
If (typically T is set to 5 degrees),
( represents the i-th If
segment)
4. Repeat for
Fig. 8. Detail procedure of the mid-point connection clustering
Repeat for , B = false.
If ,
, B =true
If B = true,

5. Repeat for
Eliminate the repeated elements in
Fig. 5. Detail procedure of the inclination angle clustering
Fig. 9 Minimal bounding rectangles of the segment clusters

A minimal bounding rectangle for each cluster of segments is


computed as shown in Fig.9. To obtain the bar-shaped
sub-images, the minimal bounding rectangles are extended
along their main directions to the edge of the image. The
extended minimal bounding rectangles are then cropped after
the image is rotated to make each of them in a horizontal
position to generate the sub-images. Four example sub-images
are shown in Fig.10.
Fig. 6. Inclination angle clustering results

Fig. 10. Samples of the bar-shaped sub-images

IV. INSULATOR EXTRACTION


. Deformable Part Models trained by latent SVM is used to
Fig. 7. Mid-point connection clustering result recognize and locate the insulator.
The segments in each cluster generated by inclination angle A.Training Samples
clustering are further clustered based on mid-point connection. Both the positive and negative samples are cropped from the
This is to determine which segments correspond to a certain images captured by the inspection vehicle. Each positive
bar-shaped device. The result of the mid-point connection


sample contains an instance of the insulator. The insulator is ψ ( H ,γ ) = ( φ ( H , p0 ),φ ( H , p1 ),...,φ ( H , pn ),
placed in the center of the sample image. Insulators in all (4)
possible lighting conditions are included for training. Each x1 , y 1 , x 12 , y 12 ,..., xn , y n , x n2 , y n2 )
negative sample randomly contains a non-insulator object. Four
positive and negative samples are shown in Fig. 11). The model of the object class is trained using latent SVM,
proposed in [11]. 180 positive and 120 negative samples are
used for the training process.

V.EXPERIMENT RESULTS
The result of rod-insulator detection in the bar-shaped
sub-images are shown in Fig.12.
Fig. 11. Positive (top row) and negative (bottom row) samples of insulators.

B.Deformable Part Models


Deformable Part Models (DPMs) are seen as the extension of
the histograms of oriented gradients (HOG) features. Unlike
Dalal and Triggs’s method [10] using a single filter to model
the appearance of the object, DPMs are composed of a root
filter and several part filters. The root filter covers the entire Fig. 12. Detection Results of the deformable part models
object, while the part filters cover smaller parts of the object. Our result shows that the rod-insulators can be accurately
An HOG feature pyramid is defined by computing HOG detected from the sub-images using deformable part models.
features on each level of a standard image pyramid. The part The detection result is not affected by the variance of scale,
filters are generated and applied to the HOG feature pyramid lighting conditions and the texture caused by different
that is one level lower than the root filter level. This allows the illumination angles. A testing image dataset containing 452
part filters to capture finer gradient information of the object. catenary images was created to determine the detection rate of
The part model of an object includes not only the the proposed method. The testing result is shown in Table 1.
corresponding part filter, but also three other parameters, vi, ai
TABLE1. TEST RESULTS TABLE
and bi. vi is a two-dimensional vector defining the ideal position Number of Detected
of the i-th part relative to the position of the root filter. ai and bi Total Image False alarms
Insulators Insulators
are both two-dimensional vectors specifying coefficients of a 452 462 413 5
quadratic function measuring a score for each possible
Total 462 insulators were included in the dataset. 413
placement of the i-th part that deviates from the ideal position.
insulators were successfully detected (89.4%). Of the 44
The score of an object hypothesis is formulated as
insulators that were misclassified, some were incomplete in the
n n
image as shown in Fig.13 and some were a result of bad
∑ Fi ⋅ φ( H , pi )+ ∑ ai ⋅( xi , yi ) + bi ⋅( x i2, y i2 ) (2) segment detection. The algorithm also detected five that are
i=0 i=1 actually not insulators. Bad localization as shown in Fig. 14
n
was the cause of these false alarms. They will be discarded in
∑ F ⋅ φ( H , p ) is the total score of the n filters. Fi is the part
i i the subsequent diagnosis of safety-threatening condition.
i=0
filter for the i-th part, F0 is the root filter, ϕ(H, pi) is the HOG
feature of the i-th part (H represents the HOG pyramid
Fig.13 Incomplete insulator.
mentioned above, pi is the i-th part’s location).
n
2 2
∑ a ⋅( x , y ) + b ⋅( x , y
i i i i i i
) is the deformation score of the
i=1 Fig.14 Example of false alarm.
hypothesis. ( xi , y i ) is the coordination of the i-th part relative
VI. CONCLUSIONS
to the root location derived from vi.
The score of a hypothesis γ can be expressed in terms of the A method for rod-insulator detection in the catenary
dot product, βΨ(H, γ), between a vector of model parameters β system is proposed in this paper. The segments detected in the
and a vector Ψ(H, γ)β is a vector of model parameters, γ is a set image are clustered to determine a region of interest according
of latent values. The definition of β and γ are shown in (3) and to their angle and mid-point connection. Deformable part
(4). In the process of object detection, the hypothesis score is models are used to recognize and localize the insulators in the
computed in the sliding window. The object is localized when image. Our experiments show that the localization result is
the hypothesis score reaches the maximum. good and the detection rate of the insulators is high. The
proposed method shows high reliability that will make the
β = ( F0 ,...,Fn ,a1 ,b1 ...,an ,bn ) (3)
process of fault detection much easier. Based on the work
presented in this paper, developing a reliable fault detection
method will be the next step.


REFERENCES
[1] H. Ha, S. Han, J. Lee, “Fault detection on transmission lines using a
microphone array and an infrared thermal imaging camera,”
Instrumentation and Measurement, IEEE Transactions on., vol. 61, no. 1,
pp. 267-275, Jan. 2012.

[2] S. C. Oliveira, E. Fontana, “Optical detection of partial discharges on


insulator strings of high-voltage transmission lines,” Instrumentation and
Measurement, IEEE Transactions on., vol. 58, no. 1, pp. 2328-2334, JUL.
2009.

[3] Q. Wu, J. An, “An active contour model based on texture distribution
for extracting inhomogeneous insulators from aerial images,” Geoscience
and Remote Sensing, IEEE Transactions on., vol. 52, no. 6, pp. 3613-3626,
Jun. 2014.

[4] M. Oberweger, A. Wendel, H. Bischof, “Visual recognition and fault


detection for power line insulators,” in Proceedings of 19th Computer
Vision Winter Workshop., Feb. 2014, pp. 1–8.

[5] S. Lazebnik, C. Schmid, J. Ponce, “A sparse texture representation


using affine-invariant regions,” in Proceedings of IEEE Conference on
Computer Vision and Pattern Recognition., June. 2003, pp. 319–324.

[6] S. Lazebnik, C. Schmid, J. Ponce, “A sparse texture representation


using local affine regions,” Pattern Analysis and Machine Intelligence,
IEEE Transactions on., vol. 27, no. 8, pp. 1265–1278, Aug. 2005.

[7] S. Liao, J An, “A Robust Insulator Detection Algorithm Based on


Local Features and Spatial Orders for Aerial Images,” IEEE Geoscience
and Remote Sensing Letters., vol. 12, no. 5, pp. 963-967, May. 2015.

[8] W. G. Li, G. Ye, F. Huang, S. Wang, and W. Z. Chang, “Recognition


of insulator based on developed MPEG-7 texture feature,” in Proceedings
of 3rd International Conference on Image and Signal Processing., June.
2010, pp. 265–268.

[9] J. Zhao, X. Liu, J. Sun, and L. Lei, “Detecting insulators in the image
of overhead transmission lines,” in Proceedings of 8th International
Conference., July 2012, pp. 442-450.

[10] K. Cui, B. Li, J. Yuan, and P. Wang, “An Improved Unit-Linking


PCNN for Segmentation of Infrared Insulator Image,” Applied
Mathematics & Information Sciences, vol. 8, no. 6, pp. 2997-3004, Feb.
2014.

[11] B. Li, D. Wu, Y. Cong, Y. Xia, and Y. Tang, “A method of insulator


detection from video sequence,” in Proceedings of IEEE International
Symposium on Information Science and Engineering, ISISE., Dec 2012,
pp. 386-389.

[12] N. Dalal, B. Triggs, “Histograms of oriented gradients for human


detection,” in Proceedings of IEEE Computer Society Conference on
Computer Vision and Pattern Recognition, CVPR., June 2005, pp. 886–
893.

[13] P. Felzenszwalb, R. Girshick, D. McAllester, and D. Ramanan,


“Object detection with discriminatively trained part based models,” IEEE
Transactions on Pattern Analysis and Machine Intelligence., vol. 32, no. 9,
pp. 1627-1645, Sep. 2010.

[14] S. Suzuki, “Topological Structural Analysis of Digitized Binary


Images by Border Following,” Computer Vision, Graphics, and Image
Processing., vol. 30, no. 1, pp. 32-46, Apr. 1985.



You might also like