Abstract: Power transmission line inspection robots must detect and identify various obstacles when crawling along the phase line and plan obstacle-crossing behaviors according to the obstacle type. For this purpose, based on the structural characteristics of a 220 kV overhead transmission line, a structure-constrained obstacle recognition algorithm is designed using vision sensors. The algorithm utilizes edge information from the image, employs an improved circle/ellipse detection method based on existence probability maps, and a hierarchical decision-making mechanism to reduce the impact of lighting changes and robot motion in the natural environment on the recognition quality, thus meeting the real-time obstacle-crossing requirements of the power transmission line inspection robot. Laboratory simulation and actual line experiments show that the algorithm can reliably identify obstacles such as vibration dampers, suspension clamps, and tension clamps in complex backgrounds. Keywords: obstacle recognition; circle detection; power transmission line inspection robot; Hough transform; power transmission line The algorithm uses an improved existence-probability-map-based circle/ellipse detection method and a hierarchical decision mechanism to reduce the effects of illumination variation and robot motion on obstacle recognition quality, which satisfies the needs of real-time obstacle negotiation for inspection robots. Experimental results with simulation and real transmission lines show that the algorithm can reliably recognize obstacles such as counterweights, strain clamps, and suspension clamps from cluttered backgrounds. Keywords: obstacle recognition; circle detection; inspection robot; Hough transform; power transmission line 1 Introduction Overhead transmission line inspection robots can crawl along phase or ground lines while the line is energized, crossing obstacles such as vibration dampers, tension clamps, suspension clamps, and towers. Using onboard detection instruments, they replace humans in close-range inspections of towers, conductors, lightning protection wires, insulators, line hardware, and line corridors. This operational method will greatly reduce the labor intensity of transmission line maintenance personnel, improve inspection efficiency and quality, and is of great significance for ensuring the safe and reliable operation of the transmission system. Most line inspection robots use wheeled walking mechanisms. The anti-vibration hammers, insulators, suspension clamps, tension clamps and other pole support accessories installed on the phase line and ground line of the overhead transmission line become obstacles affecting the walking of the line inspection robot. Therefore, the line inspection robot must use sensors to detect and identify these obstacle types, and then plan obstacle crossing behavior according to the obstacle type in order to cross the pole and walk autonomously along the transmission line over a large area. Commonly used sensors include proximity sensors (infrared sensors, ultrasonic sensors) and vision sensors (cameras). Peters et al. [1] configured a sensor array consisting of up to 34 proximity sensors on a line inspection robot, and designed an obstacle classifier using a rough set-based neural computing method. The obstacle identification and classification were performed based on the measurement data of the sensor array, and then the obstacle crossing behavior of the robot was planned. Peungsungwal et al. used a camera to identify overhead line obstacles and completed the identification of insulators at both ends of the experimental line. Based on the structural characteristics of the 220kV overhead transmission line, this paper uses a vision sensor, an improved geometric detection method and a hierarchical decision mechanism to design an obstacle detection and identification algorithm based on structural constraints. 2 Analysis of 220kV Power Transmission Line Inspection Robot and its Working Environment The inspection robot adopts a double-arm wheel structure and rolls along the 220kV phase line. Figure 1 is a schematic diagram of the structure of a 220kV transmission line. There are three typical obstacles that the inspection robot needs to cross: (1) anti-vibration hammer; (2) suspension clamp and single string insulator; (3) tension clamp, double string insulator and current jumper. The anti-vibration hammer is fixed on both sides of the suspension clamp and tension clamp with clamps. It is the first obstacle that the inspection robot needs to cross when approaching the tower. The above three typical obstacles have different structural components and spatial layouts. The inspection robot needs to adopt corresponding obstacle crossing strategies when crossing obstacles. The camera is installed at the front end of the inspection robot body. The angle between the optical axis and the robot's forward direction is about 30°, as shown in Figure 1. In this configuration, the background is the sky, which reduces background complexity. The projections of the vibration damper end face, suspension clamp supporting insulator, and tension clamp supporting insulator onto the camera's imaging plane are typical geometric shapes such as straight lines, circles, and ellipses. The environmental image captured by the camera is analyzed and understood to identify whether specific obstacle targets are present in the image. 3. Obstacle Recognition Algorithm Principle and Implementation Vibration dampers, suspension clamps, tension clamps, and insulator strings lack rich surface textures and distinct color characteristics, and are hinged together, making it difficult to segment them into individual areas. Based on the appearance, shape, and structural characteristics of obstacles, we select relatively simple geometric primitives such as straight lines, circles (arcs), and ellipses as clues to the existence of obstacles. Combined with the constraints of the transmission line structure on the positions of these geometric primitives, we identify different types of obstacles. Since there are significant differences among the three types of obstacles, we simplify the problem of identifying the three types of obstacles into three separate classification problems: distinguishing the vibration damper, suspension clamp, tension clamp, and insulator string from the background, thereby simplifying the design of the classifier. The two key problems of the algorithm are: (1) quickly and reliably extracting graphic primitives such as straight lines, circles (arcs), and ellipses from the edge image; (2) determining the type of obstacle from which the detected graphic primitives originate under structural constraints. 3.1 Identification and positioning of phase lines Vibration dampers, suspension clamps, tension clamps, and insulator strings are fixed to phase lines in different ways, so the identification, positioning, and tracking of phase lines are crucial for obstacle identification. First, a Gaussian filter is used to eliminate high-frequency noise and the image is moderately blurred to reduce local texture edges and improve the detection rate of straight lines at the edges of the wires; then, the edge map of the image is extracted using the Canny operator, and the image size is 320×240; finally, the Progressive Probabilistic Hough Transform (PPHT) [3] is used to extract all straight line segments in the edge map with a length greater than a given threshold T. Each straight line L[sub]i[/sub] is represented by the following parameters: θ, tilt angle, the angle between the straight line and the positive direction of the axis, as shown in Figure 2; (x[sub]1[/sub], y[sub]1[/sub]), the coordinates of the upper endpoint; (x[sub]2[/sub], Y[sub]2[/sub]), the coordinates of the lower endpoint. When the line-following robot walks along the phase line, the projection of the phase line in the image plane is a rod-shaped projection that tilts downward from the upper border of the image plane. The robot's shaking has a certain impact on the projection position and tilt angle of the phase line in the central axis direction of the image plane. Experiments show that the changes in the projection position and tilt angle of the phase line are within a certain small range. Therefore, we use a rule-based classifier to classify the detected line segments and find the two edge lines L[sub]L[/sub] and L[sub]R[/sub] corresponding to the phase line. Let E=；L[sub]i[/sub], i=1,2,…,n＝ represent the set of lines detected by PPHT, and C represent the candidate set of edge lines of the phase line. The classification rule of the classifier is: If any line segment L[sub]i[/sub]∈E satisfies the following conditions: (1) The tilt angle θ satisfies 70[sup]<θ<110[sup]; (2) The upper endpoint coordinate Y-axis component y[sup]1[/sup] x[sub]1i[/sub]，i≠j); T[sub]1[/sub]、T[sub]2[/sub] are determined according to the projection width of the phase line, then L[sub]i[/sub]、L[sub]j[/sub] are the left and right edge lines L[sub]L[/sub]、L[sub]R[/sub] of the phase line, respectively. In this way, the phase line identification and positioning are completed, and the area where the obstacle is located is roughly determined, laying the foundation for subsequent obstacle identification. 3.2 Identification of the anti-vibration hammer The anti-vibration hammer is the first obstacle encountered by the line inspection robot when it approaches the tower. It is composed of two hollow cylinders connected together and fixed to the lower part of the phase line with a clamp, as shown in Figure 6. The identification of the anti-vibration hammer is based on two assumptions: Assumption 1 There is a hole on each of the two visible end faces of the anti-vibration hammer, which is larger at the front and smaller at the back. When projected into the image, it is a local dark area. The hole area of ​​the front hole is smaller and circular; the hole area of ​​the rear hole is larger and crescent-shaped. Therefore, the adaptive threshold segmentation method can be used to segment the above two hole regions as one of the evidences of the existence of the vibration damper. Assumption 2 The edge of the front face of the vibration damper presents an irregular upper semicircle with the center of the hole as the center. This semicircle is the second piece of evidence of the existence of the vibration damper. Based on the above assumptions, we designed a vibration damper identification algorithm, the steps of which are as follows: (1) Low-level image processing: According to the identified phase line position and width, determine the image window W[sub]r[/sub] containing the vibration damper. The Otsu[4] optimal threshold algorithm is used to perform binary segmentation on the image window W[sub]r[/sub], and the resulting binary image is subjected to morphological opening operation filtering to eliminate image noise. Then, the centroid (x[sub]c[/sub], y[sub]c[/sub]), area and contour of each region are extracted. (2) Hole candidate region determination: Based on the position constraints of the phase line and the vibration damper, the hole candidate region representing the existence of the vibration damper is determined. Let θ[sub]L[/sub], θ[sub]R[/sub] (x[sub]1L[/sub], y[sub]1L[/sub]) and (x[sub]1R[/sub], y[sub]1R[/sub]) represent the inclination angle and upper endpoint coordinates of the left and right edge lines of the phase line, respectively; Q is any region segmented in step (1), with centroid coordinates (x[sub]c[/sub], y[sub]c[/sub]), then the horizontal distance from the centroid to the left and right edge lines of the phase line and the width of the phase line at the centroid are respectively: If equation (2) is satisfied, then region Q is a candidate region for end face hole L. (3) Determination of end face hole region: Apply end face geometric constraints to the candidate region. The area of ​​the dark hole on the rear face is large, and the lower part of the boundary contour of the segmented area should be a semi-circular arc. Therefore, the contour of the largest area in the candidate area is detected as a semi-circular arc. If the lower contour of the area is a semi-circular arc, it further supports the hypothesis that the area is the rear face hole area. The front face is circular, and the upper semi-circular arc discontinuous edge concentric with the end face hole can be extracted in the edge map. Therefore, for the remaining candidate areas, it is detected in the edge map whether there is an upper semi-circular arc with the centroid of the candidate area as the center. If it exists, it is further determined that the area is the front face hole area. (4) If both hypotheses about the front and rear face hole areas are true, it is confirmed that the vibration damper has been detected, and the circular parameters of the front and rear faces are estimated at the same time. How to detect whether the irregular edge is a circle or a semi-circular arc in the sense of perception is the key to realizing the vibration damper recognition algorithm. The specific principle will be introduced in detail in Section 3.4. 3.3 Suspension clamp recognition The suspension clamp clamps the phase wire and hangs it under the single insulator string, so that the phase wire is suspended in the air. Its own volume is small and its features are difficult to extract. Under the camera configuration shown in Figure 1, the bottom of the single insulator string appears as an ellipse in the image, with the principal axis basically parallel to the axis. This configuration is unique among the three types of obstacles, thus allowing us to transform the problem of identifying the suspension clamp into an ellipse detection problem. Reference [5] provides an effective ellipse detection method, but it involves a large amount of computation. Based on this, we have improved the ellipse detection algorithm in Reference [5] according to the characteristics of insulator imaging. Describing an arbitrary ellipse requires five unknown parameters: the center coordinates (x₀, y₀), the major semi-axis a, the minor semi-axis b, and the rotation angle α. The corresponding parametric equations are as follows: As shown in Figure 3, we assume that the edge point pair (x₁, y₁) and (x₂, y₂) are the two vertices of the principal axis of the ellipse to be tested. Then, the four parameters of the ellipse to be tested can be estimated according to equation (4): Let (x, y) be a point on the ellipse to be tested. The minor semi-axis b can be calculated using equation (5): Next, we use a voting method similar to the Hough transform (HT) to estimate the minor semi-axis b of the ellipse to be tested. Let A(i) be a one-dimensional accumulator array (min_b≤i≤max_b, min_b and max_b are the minimum and maximum values ​​of the preset b, respectively). W_s is the smallest rectangular region containing the ellipse to be tested in the edge image. For each edge point (x_i, y_i), calculate the corresponding minor axis b_i using equation (5) and round it. If min_b≤b_i≤max_b, then increment the accumulator array unit A(b_i) by 1, that is, the edge point (x_i, y_i) is the minor axis of the ellipse to be tested. After the voting is completed, the maximum value of the accumulator array is calculated and set as A(b_m), and the corresponding minor axis b_m. Then the probability that there is an ellipse in the image described by the estimated 5 parameters is: Take a threshold for P. If P>T_e, then there is an ellipse in the image described by the estimated 5 parameters; otherwise, there is no ellipse. The size of T_e depends on the quality of the extracted ellipse edge and is determined by experiment. The steps for identifying suspension clamps based on ellipse detection are as follows: (1) Extract the edge image and try to eliminate the trivial edge points formed by local textures to improve the ellipse detection rate. (2) Search for the ellipse from top to bottom along the phase line, and determine the width of the search area according to the position constraints of the phase line and the suspension clamp and the possible size of the insulator projection ellipse, as shown in Figure 4. During the search, find the leftmost and rightmost edge points in the long rectangular frame and record them as (x1, y1) and (x2, y2) respectively. Then, calculate the four parameters of the hypothetical ellipse according to equation (4) and determine the smallest rectangular window Ws that encloses the ellipse. Traverse all edge points within Ws and estimate the minor semi-axis b and the existence probability P of the hypothetical ellipse according to equations (5) and (6). Save the ellipse parameters and corresponding existence probabilities in order: {(x0, y0), a, b, α, P}. (3) Calculate the maximum existence probability Pm. If Pm > Te, it indicates that the hanging clamp has been identified from the image, and its position is given by the ellipse parameter corresponding to P. 3.4 Tension Clamp Identification The bottom of the two strings of insulators supporting the tension clamp appears as two circles separated by the phase lines on the left and right in the image. Similar to the identification of suspension clamps, we transform the identification problem of tension clamps into the detection of circles. Any circle can be described by three parameters: the center coordinates (x₀, y₀) and the radius r, i.e.: The Hough Transform (HT) is widely used in geometric shape detection. References [6-8] propose several circle detection methods based on the Hough Transform. For cases where the circular edge of the insulator bottom is incomplete or irregular, this paper proposes a circle detection method based on an existence probability map, which can effectively detect circles with unclear edges. Let f[sub]e[/sub](x,y) represent the edge image extracted by the Canny operator; P(x,y) is a two-dimensional array that stores the probability map of circle existence. The value of cell (x,y) represents the probability that there exists a circle with center (x,y) and radius Rx,y in the edge image f[sub]e[/sub](x,y), where R[sub]x,y[/sub] = R(x,y); R(x,y) is a two-dimensional array, where cell (x,y) stores the radius of the possible circle at point (x,y) in the edge image (x,y); Rmin represents the minimum allowable radius of the circle to be detected, which depends on the specific problem; A(i) is a one-dimensional accumulator array. The algorithm principle is as follows: (1) Select the center (u, v) ∈ f[sub]e[/sub](x, y), as shown in Figure 5, determine the possible maximum radius Rmax of the circle located at this point, and calculate the minimum rectangular region enclosing the circle with center (u, v) and radius Rmax. (2) For each edge point (x[sub]i[/sub], y[sub]i[/sub]) in W[sub]r[/sub], calculate the Euclidean distance D from the point to the center (u, v). If R[sub]min[/sub] If Tp is found, it is assumed that there is a circle in the image f[sub]e[/sub](x,y) with the center (x[sub]1[/sub], y[sub]1[/sub]) and radius R(x[sub]1[/sub]). Then, the probability value of the point (x[sub]1[/sub], y[sub]1[/sub]) and its neighborhood in the circle existence probability map P(x,y) is set to 0. The size of the neighborhood depends on the edge quality and the distance between adjacent circles in the image. Repeat the above peak detection process until N circles are found, or the remaining peak value is less than the given threshold. When identifying tension clamps, first use the above circle detection method to detect all possible circles in the image, and then perform structural constraint checks on the detected circles. If there is a pair of circles that meet the following two conditions, it is considered that the tension clamp has been identified from the image: (1) The radii of the two circles are close; (2) The centers of the two circles are distributed horizontally on the left and right sides of the phase line. Because the algorithm needs to traverse all image points, the computational load is large. Considering that the identification of tension clamps does not require precise estimation of circle parameters, we used the nearest neighbor sampling method to reduce the original image from 320×240 to 80×60 to improve the computation speed. However, this will also cause the image edge quality to decrease and affect the circle detection quality. Nevertheless, the algorithm still shows relatively stable performance. 4 Experimental results We identified obstacles on the laboratory simulated line and the actual line respectively. Figure 6 (1) shows the identification results of the vibration damper, suspension clamp and tension clamp on the simulated line; Figure 6 (2) shows the identification results of the suspension clamp and tension clamp on the actual line. It can be seen from the figure that although the extracted obstacle edges are irregular and incomplete due to factors such as viewing angle, cluttered background and weak contrast, the algorithm can still detect the graphic primitives such as circles/ellipses in the image. After the test and classification under structural constraints, the various obstacles are finally reliably identified. 5 Conclusion The edge points extracted from the image contain all the information of the original image under certain conditions, and the target of interest often appears at the edge of the image. Therefore, the algorithm makes full use of image edge information, uses an improved circle/ellipse detection method to extract typical geometric shapes that constitute obstacles, and uses the structural constraints of the line-crawling robot's operating environment to make hierarchical decisions, reduce the search space, and make the algorithm simple and effective. The circle/ellipse detection method based on the existence probability map transforms the "voting" result of edge points into the relative probability of the existence of circles/ellipses in the parameter space, which improves the detection rate of circles/ellipses of unclear edges. The obstacle visual recognition algorithm for overhead high-voltage transmission line line-crawling robots designed in this paper basically meets the obstacle crossing requirements of line-crawling robots. Since the decision method based on threshold is adopted, although the algorithm shows stable and consistent performance over a wide range, it also has the problem of uncertainty in adaptability to environmental changes. References [1] Peters JF, Ahn TC, Borkowskii M. Obstacle classification by a line-crawling robot: a rough neurocomputing approach [A]. Pro-ceedinss 0f the Third International Conference on Rough Sets and Current Tren (Is in Computing — Lecture Notes in Artificial Intelli• genee [C]. London, UK: Springer-Verlag, 2002. 594-601. Peungsungwal S, Pung~fi B, Chamnongthai K, et a1. Autonomous [2] robot for a power transmission line inspection[A]．Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Systems [C]．Piscataway, NJ, USA: 2001．121-124． Pattern Re— cognition[C]. Alamitos, CA, USA: IEEE Computer Society, 1999. 554-560. [4]Otsu N, A threshold selection method from gray-level histogram[J]. IEEE Transactions on System, Man, and Cybernetics, 1979, 9 (1): 62-66. [5]Xie YH, Ji Q, A new efficient ellipse detection method [A], Proceedings of the 16th International Conference on Pattern Recognition [C]. Los Alamitos, CA, USA: IEEE Computer Society, 2002, 957-960. [6]Duda R 0, Hart PE. Use of the Hough transformation to detect lines and curves in pictures[J]． Communications of the ACM, 1972, 15(1): 11-15. [7]Kimme C, Ballard DH, Sklansky J. Finding circles by an array of accumulators[J]． Communicatim~s of theACM, 1975, 18(2): 120~l22. [8] Leavem VF. Shape Detection in Computer Vision Using Hough Transform[M]． London: Springer-Verlag, 1992.