CC BY
17Image Processing Algorithms for High Voltage
Power Line Detection
Batbayar Battseren, Uranchimeg Tudevdagva, Wolfram Hardt Department of Computer Science, Chemnitz University of Technology, Chemnitz, Germany
Abstract—this paper describes an image processing for detection of high voltage power lines. Different algorithms were applied for detection of lines from images which made by unmanned aerial vehicle-based inspection. The purpose of detection is to find out the heading direction, while the unmanned aerial vehicle position and camera view angle not defined. The proposed method started with edge detection and followed by a line detection algorithm to obtain the lines in the frame. Then find all possible intersection points, which located inside and outside of the original image. Based on these intersection points, the power lines are refined. In this paper, we used real images from inspection flight of high voltage transmission lines in Mongolia.
Keywords— power-line inspection, aerial inspection, image processing, line detection, intersection point, unmanned aerial vehicle
INTRODUCTION
Electricity is one of the key element of infrastructure in any countries. On demands of high technology development in many fields of human life increase role of sustainable electricity in countries. Production and distribution of electricity consists of three major parts, which are the generation, transmission and distribution. High voltage transmission line (HVTL) is a structure that used for transmitting the electric energy along a long distance. According to the importance of HVTL reliability, the inspection and maintenance procedures are one of the crucial work in the power sector.
In recent years, HVTL inspection process greatly altered due to the benefit of the unmanned aerial vehicle (UAV) development and capabilities. Now, the vertical take-off UAVs are able to carry a wide variety of camera setups and fly more than 15 minutes [1]. Technicians are using this kind of technologies in all sorts of aerial inspection use cases such as a substation, wind turbine, chimney, bridge, construction heat-loss inspection and more. UAV based power line inspection has numbers of benefits. For instance, it significantly decreased the risk, cost, time consumption of conventional power line inspection, and provided possibilities to carry out the inspection without shutting it down.
Due to the rapid development of the artificial intelligence, embedded system, and robotic technology, there are numerous researches have been done on UAV based automated power line inspection system. For instance, power line tracking or damaged detection methods [2], [3], [4]. One challenging task of this automated system is a computer-vision based object detection or visual navigation since the ordinary positioning sensors are not working sufficient near to strong electromagnetic field, and lack of long-range environment sensing solution.
According to the overhead power line feature, most algorithms for the power line detection based on analytic methods. These methods are mainly start with the edge detection and followed by line detection. In the work of Alexander and Ivan [5], a circle-based approach defined, where authors are used a Steerable filter for the edge detection and followed by line detection. Likewise, in the work of Yuee Liu [6], author started with the Steerable filter for edge detection and followed by ridge point detector and Hough line detectors. Also in this work, researchers compared their result with edge drawing algorithm (EDLines) and line segment detector (LSD). In the research of Koshelev V.I [7], authors recommended to use a priori knowledge of power line position, canny edge detection, and morphological filters.
1. PROPOSED METHOD
In this paper, we present an intersection-point based HVTL detection and localization image-processing algorithm [8]. The proposed image-processing algorithm presented in (Fig. 1). There are three main steps in our concept.
• The first step is an edge detection and filtering, which detects power line edges.
• The second step is a line detection and clustering, which obtains the candidate lines in the image.
• The last step is an intersection point algorithm, which uses the intersecting points of the detected lines to estimate the heading direction of the power line.
End ^ Power line detection algorithm
Same as most of the power line detection methods, here we start with edge detection and followed by line detection. In line detection, we faced with two main problems.
1. First, due to the environmental complexity, not all linear-edges are the power line. Therefore, we intended to eliminate the border edges of the solid object at the beginning.
2. Then the second problem is the double or parallel edges of the single line, which caused by image resolution or width of the line and causes multiple candidate line detection for a single line. Our proposed solution for this issue is to group the lines with clustering technic.
3. The last step is to verify the power line and find heading direction. According to the position and view angle of the camera, the power line orientation and the angular difference could vary a lot. To solve this problem, we used the intersection points of these lines to find out the heading direction. All main steps explained in detail in the following subsections.
Edge Detection and Filtering
Fig. 2 shows the proposed algorithm of edge detection in steps. This method used to highlight a thin line and overcome the solid-object linear edge problem.
First, the source image filtered to remove noise and the power lines. In this step, the Median filtering technique is used (1), which preserves edges while removing the salt-and-pepper noises effectively [9]. Due to the narrow width of the power line, the Median filter considers the power line as a noise. Formula (1) used for Median filter calculation.
/Of, >0 = median [g (s, t}} where: (s, t) e S^
(1)
According to the environment (background) variation, the color based feather detection is not possible. Therefore, both source and filtered images were converted into the grayscale images (Fig. 2. Step: Grayscale S, F) by RGB to Grayscale conversion function (2) [10]. After this, in other to highlight the thin lines, we found the absolute difference (Fig. 2. Step: Grayscale Difference) of these two images with formula (3), and amplified. This step highlights the all power lines in the bright and dark background (Fig. 3).
Y 0.299 'R + 0.537 ■ G + 0.114 ■B A Y = | Yt - Y2
(2) (3)
Edge detection and filtering algorithm
Power line highlight
Then, the canny edge detection (CED) technic applied on the difference image and filtered image (Fig. 2. Step: Edge Detection (D, F)), which consists of several sequential steps [11]. First, it will apply the vertical and horizontal convolution masks (4), and then find the gradient strength (5) and orientations (6). The non-maximum suppression applied to end the edge detection by removing non-edge pixels.
C, =
-1 0 1 -2 0 2 -1 0 1
(4)
(5)
(6)
Lastly, in order to remove the solid object edge (power pole, insulators, building, road and more), the filtered image edge pixels are dilated and subtracted from the edge pixels of the difference image (Fig. 2. Step: Edge Subtraction).
Line Detection and Clustering
Line detection and clustering algorithm
Hough transform (HT) used for line detection (Fig. 4. Step: Line detection) in our approach [12]. HT commonly used in image processing to detect straight lines, circles or ellipses while requiring less computational resource. We consider that previous step (Edge Detection and Filtering) eliminated the solid objects edge pixels to reduce the false negative detection rate.
HT uses 2-dimensional (for r and 8 parameters) array to detect lines described by formula (7). Each edge pixel represented as curve (sinusoid) in the Hough space, the lines can be detected by finding the number of intersections of the curves. HT returns r and 8 values for each line in Hesse normal form (7).
r = j: cos 8 + y sin 8
(7)
However, one drawback of the HT is the result contains multiple candidate lines for each line due to the double edge or noise edge pixels. With the intention of filtering unnecessary lines, we used a clustering technique based on the position (r) and angular (& ) parameters of the candidate lines (Fig. 5).
The first element (line) will be the first cluster. If r and 8 parameters of one line are both under a specific range, that line is grouped in the current cluster. If not, that line will be group as unclustered elements. It will be executed on every line. If there is any unclustered element left, new cluster will be defined and it will loop again. The total number of the cluster is undefined.
Line and angle clustering algorithm
A. Intersection Point Algorithm
In this part, the power line orientation will be calculate based on the intersecting points as shown in Fig. 6.
Intersection point algorithm
The orientation and angular difference of the power line could vary due to the UAV position and camera view angle. However, HVTL is a group of lines, which are heading to one place and intersecting in one direction. Hence, all possible intersecting points will be find at first (Fig. 6. Intersection Point) with formula (8) in matrix form.
(8)
Then, based on these point coordinates, it obtains the most intense area, which outlines the heading direction of power lines (Fig. 6. Point Clustering). Two kinds of clustering methods applied on the points to find the concentrated area. First, an angle based clustering is implemented on intersection points, which found outside of the frame (Fig. 7.a). The same clustering method is use as line clustering. If there is no intersection point outside of the frame, the K-mean clustering applied on the points (Fig. 7.b) [13]. In both cases, the cluster, which contains a maximum number of elements, will be select as the heading direction of the power line.
Point clustering. a. angle clustering; b. K-mean clustering
If there is any line, which not intersected in the area that will not considered as a power line (Fig. 6. Line Filter) and will be deleted. In other words, the lines that intersected in this area is use for next calculations.
In order to find region of interest (ROI), we used the triangle shape. Three points needed to make triangle RIO. The first point found by the point clustering, which indicates the heading direction (Fig. 8. Point A). If the K-mean clustering used, the selected cluster coordinate will be this point. If angle clustering used, the farthest point will be used as this point.
Power line Region of Interest
The borderlines used to find out the next two points of ROI (Fig. 8. Point B and C). Each borderline crosses with the frame two times. The point located on the opposite side of the frame will selected as the next points of the ROI. After all, three points found, the ROI can be visualized.
Finally, the power line orientation found by the AD median of the ROI triangle. The power line heading direction will be point A (Fig. 8).
2. EXPERIMENTAL RESULTS
Edge Detection and Filtering
The Edge detection step result is presenting in Fig. 9. The image shows comparison between direct result of CED on the source image (Fig. 9.b) and our proposed method result (Fig. 9.c).
' r ■ x
• * rW ■ ■"* : -• y, -V bStka^'J .. <• A. ■ r-
v \ ' : - V- '
Edge detection and filtering solution. a. source image; b. CED output; c. proposed method
Line Detection and Clustering
Fig. 10 presents the result of Line Detection and Clustering step. As shown in the figure, the line clustering step groups detected multiple lines (b: left - 20 lines, right - 13 lines) into fewer clusters (c: left - 5 clusters, right - 6 clusters).
(is Iii • 1
20 13
Line detection and clustering result. a. source image; b. Hough line detection result; c. line clustering result
B. Intersection Point Algorithm
The result of the power-line detection algorithm shown in Fig. 11. In the image, the ROI area is highlighting and the power line heading direction is illustrated with a yellow arrowed line.
Intersection point algorithm result
3. DISCUSSION
There are few weak points in our method. First, if a single line detected, the intersection point algorithm is not able to carry out. In addition, due to the different contrast and width of the line, our method is no detecting all individual lines. Therefore, our next step is to improve the positive detection rate of each individual lines
However, there are some main benefits of the intersection point algorithm. It finds the power line area (ROI) and heading direction in any orientation from any position with any number of lines.
The further research will mainly focus on the real-time onboard video processing. In order to fulfill this goal, performance improvement, frame-to-frame data processing (result inheritance in following frames), embedded platform selection and implementation studies will be executed.
CONCLUSION AND FUTURE WORK
Under the scope of power line detection, we attempted to solve three different problems. The first problem is an accurate line detection method. To solve this problem, first, we highlighted the thin lines and then removed the solid object edges. The second problem is multiple candidate lines for a single line. This problem solved with the clustering technique, which groups the lines by r and 8 parameters. The cluster number not defined and depended on the line parameters. The third one is the power line orientation and heading direction finding. In our case, the UAV position and camera view angle is not defined and not stationary. Therefore, a fixed or predefined orientation based method is not appropriate. With
the intention of solving this problem, we used the intersection point coordinates.
The experimental result shows this method is able to detect multiple power lines in any orientation and able to obtain out heading direction independently from the detected line number.
REFERENCES
[1] J. F. McEvoy, G. P. Hall and P. G. McDonald, "Evaluation of unmanned aerial vehicle shape, flight path and camera type for waterfowl surveys: disturbance effects and species recognition," PeerJ, 2016.
[2] U. Tudevdagva, B. Battseren, W. Hardt, S. Blokzyl and M. Lippmann, "UAV Based Fully Automated Inspection System for High Voltage Transmission Lines," 12th International Forum on Strategic Technology (IFOST), 2017.
[3] R. Ishino and F. Tsutsumi., "Detection system of damaged cables using video obtained from an aerial inspection of transmission lines," IEEE Power Engineering Society General Meeting, pp. 1857-1862, 2004.
[4] G. Zhou, J. Yuan, I.-L. Yen and F. Bastani, "Robust Real-Time UAV Based Power Line Detection and Tracking," IEEE International Conference on Image Processing (ICIP), pp. 744748, 2016.
[5] A. Ceron, I. F. M. B and F. Prieto, "Power line detection using a circle based search with UAV images," International Conference on Unmanned Aircraft Systems (ICUAS), pp. 632639, 2014.
[6] Y. Liu and L. Mejias, "Real-time Power Line Extraction from," in 2nd International Conference on Applied Robotics for the Power Industry, Zurich, 2012.
[7] K. V.I and K. D.N, "Wire Recognition in Image within Aerial Inspection," 4th Mediterranean Conference on Embedded Computing , pp. 159-162, 2015.
[8] B. Battseren, U. Tudevdagva, W. Hardt and e. Patchipala, "Intersection Point Based Power Lines Detecting and Tracking Algorithm," IBS International Summerschool on Computer Science, Computer Engineering and Education Technology, pp. 55-57, 2017.
[9] T. S. Huang, G. J. Yang and G. Y. Tang, "A fast two-dimensional median filtering algorithm," IEEE Trans. Acoust., Speech, Signal Processing, vol. 27, pp. 13-18, 1979.
[10] C. Poynton, Digital video and HDTV algorithms and interfaces, San Francisco: Morgan Kaufmann, 2003.
[11] J. Canny, "A Computational Approach to Edge Detection," IEEE Transactions on Pattern Analysis and Machine Intelligence, Vols. PAMI-8, pp. 679-698, 1986.
[12] P. E. H. Richard O. Duda, "Use of the Hough Transformation to Detect Lines and Curves in Pictures," Communications of the ACM, vol. 15, pp. 11-15, 1972.
[13] J. B. MacQueen, "Some Methods for classification and Analysis of Multivariate Observations," Proceedings of 5th Berkeley Symposium on Mathematical Statistics and Probability, pp. 281297, 1967.
Batbayar Battseren received the B.S. degree in Electronic Engineering in 2009 and M.S. degree in Electronic Engineering in 2011 from Mongolian University of Science and Technology (MUST). He is working as lecturer in MUST since 2011. Since 2017, he is working as researcher and studying Ph.D. at Chemnitz University of Technology in Germany.
Uranchimeg Tudevdagva received the B.S. degree in Computer Engineering from Electro Technical University of Novosibirsk, Russia in 1992 and M.S. degree in Electrical Engineering from Mongolian University of Science and Technology (MUST), Mongolia in 1997, respectively. Later, she received Ph.D. in Computer Science from Novosibirsk State Technical University, Russia in 2004 and ScD (Dr.-Ing.habil) degree from Chemnitz University of Technology, Germany in 2014. Since 1992, she is working in MUST, Mongolia. She made long careers from assistant lecturer to Professor and Researcher in MUST. She is the expert in man-machine systems, human computer interaction, e-learning and distance learning, evaluation theory and evaluation model. She now with Chemnitz University of Technology, Germany.
Wolfram Hardt received the Diploma degree in Computer Science in 1991 and the Dr.-rer. nat. degree in Computer Science in 1996 from University of Paderborn. Later in 2000, he was entitled with Habilitation on the topic: Integration von Verzögerungszeit-Invarianz in den Entwurf eingebetteter Systeme at University of Paderborn. He was working as a Chair of the Computer Science and Process Laboratory of the University of Paderborn since 2000 to 2002, and as a Chair (procuration) of the Operating Systems Department of faculty for Elektrotechnik / Informatik, University of Kassel since 2002 to 2003. He is working as a Chair of the Computer Engineering Department of the faculty for Computer Science since 2003, and as a Dean of the Faculty for Computer Science since 2006 in Technische Universität Chemnitz.
