Method and System for Point Cloud Data Mosaic of Pantograph Slide Based on Calibration Block and Improved ICP Algorithm

Abstract

The present invention discloses a method and system for splicing the point cloud data of a pantograph slider based on a calibration block and an improved ICP algorithm, specifically as follows: First, a laser displacement sensor is used to continuously output multiple frames of slider contour data, which is expanded into three-dimensional point cloud data; then, a denoising algorithm of hybrid filtering is used to process the slider data to obtain a slider contour data set, and the two sets of laser point cloud data collected by the two laser displacement sensors are converted into the same coordinate system to form the complete point cloud data of the slider surface; then, a data rapid splicing method based on the calibration block is adopted for the point cloud data of the slider surface to complete rough registration; finally, an improved ICP registration algorithm is adopted to complete the rapid fine registration of the point cloud data to obtain the complete point cloud data of the slider surface. The present invention has both high precision and real-time performance, can overcome the interference caused by the surface texture, color of the pantograph slider and the dynamic measurement environment, and provides real-time data support for the intelligent operation and maintenance of rail transit.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of urban rail transit, and particularly to a method and system for splicing point cloud data of a pantograph slider based on a calibration block and an improved ICP algorithm. Background Art

[0002] With the rapid development of urban rail transit, the problem of train operation safety has become increasingly prominent. As a key component in the pantograph that directly contacts the catenary wire, the pantograph slider undertakes the important responsibilities of stable power transmission and bearing friction and wear, ensuring the reliability and safety of the train's power supply. Once the pantograph slider malfunctions, it may lead to unstable power transmission, poor contact, arc generation, equipment damage, and even train power outage or operation failure, seriously threatening the train operation safety and operation efficiency. Therefore, it is necessary to regularly detect the pantograph, and by reconstructing the point cloud data of the slider surface, more comprehensive and accurate information can be provided for surface defect detection, so as to timely discover the wear condition of the slider and take corresponding maintenance measures to prevent train suspension and safety accidents caused by slider failures. In addition, regular detection not only helps to optimize the maintenance plan, but also can extend the service life of the equipment and reduce the operation and maintenance costs. Therefore, accurate detection of the slider is of crucial significance for ensuring the safe and efficient operation of urban rail transit trains.

[0003] Currently, for the detection of pantograph sliders, technologies such as binocular stereo vision and structured light are mainly used to obtain the point cloud data of the slider surface. However, limited by the texture, color of the slider surface and the dynamic measurement environment, the accuracy of such methods needs to be improved, and there are problems such as large computational amount, poor real-time performance, and high requirements for equipment. In this case, how to improve the real-time performance of the detection algorithm while ensuring the accuracy of the point cloud reconstruction of the slider surface has become an urgent and challenging problem.

[0004] Patent CN117808782A discloses a method for detecting surface defects of a pantograph slider. This method constructs a selection box through straight lines for the collected image to frame the slider area, effectively suppressing the interference of the surrounding background and debris on the slider surface in the image, and accurately locating the slider area; then, by performing connected component extraction in the slider area, calculating the area of the connected component, the contrast and gray level fluctuation between the connected component and the surrounding area, it is judged whether the connected component is a defect area, and defects are detected within the slider area, improving the detection efficiency and accuracy. This method mainly relies on image recognition and segmentation technologies and may not comprehensively cover all types of pantograph failures.

[0005] Patent CN117611525A discloses a vision detection method and system for the wear of pantograph carbon strips. It completes the stitching of the captured images of the left and right pantographs by means of feature point calibration and matching, locates the pantograph carbon strip area using a target detection algorithm, combines image morphological processing and an improved edge detection algorithm to refine the extraction of the carbon strip contour, and then converts the sub-pixel coordinates of the image into world coordinates to obtain the wear value by calculating the distance between the fitting lines of the upper contour and the lower contour, ensuring the accuracy and reliability of the calculation results and improving the accuracy of the wear detection of the pantograph carbon strip. This method requires high computing resources, places high demands on the hardware performance of on-vehicle equipment, and increases the system cost. Summary of the Invention

[0006] The purpose of the present invention is to provide a method and system for stitching point cloud data of a pantograph carbon strip based on a calibration block and an improved ICP algorithm, which has high accuracy in reconstructing the point cloud data of the pantograph carbon strip surface, strong real-time performance, and can improve the detection accuracy and efficiency of the pantograph carbon strip.

[0007] The technical solution for achieving the purpose of the present invention is as follows: A method for stitching point cloud data of a pantograph carbon strip based on a calibration block and an improved ICP algorithm includes the following steps:

[0008] Step 1: Arrange a on-site acquisition module to scan the carbon strip contour data in real time during the train operation;

[0009] Step 2: Use two groups of laser displacement sensors to continuously output multiple frames of valid data and expand it into three-dimensional point cloud data;

[0010] Step 3: Process the carbon strip data using a denoising algorithm of hybrid filtering to obtain a carbon strip contour data set;

[0011] Step 4: Use the carbon strip contour data set to convert the point cloud data collected by the two groups of laser displacement sensors into the same coordinate system;

[0012] Step 5: Complete rough registration of the point cloud data of the carbon strip surface using a data stitching method based on a calibration block;

[0013] Step 6: Use an improved ICP registration algorithm to complete the fine registration of the point cloud data and obtain the complete point cloud data of the carbon strip surface.

[0014] A system for stitching point cloud data of a pantograph carbon strip based on a calibration block and an improved ICP algorithm is used to implement the method for stitching point cloud data of a pantograph carbon strip based on a calibration block and an improved ICP algorithm. The system includes a first unit to a sixth unit, and the functions of each unit are as follows:

[0015] The first unit scans the carbon strip contour data in real time during the train operation through a on-site acquisition module;

[0016] The second unit uses two sets of laser displacement sensors to continuously output multiple frames of valid data, which are expanded into three-dimensional point cloud data;

[0017] The third unit processes the skateboard data using a denoising algorithm with hybrid filtering to obtain a skateboard contour data set;

[0018] The fourth unit uses the skateboard contour data set to transform the point cloud data collected by the two sets of laser displacement sensors into the same coordinate system;

[0019] The fifth unit uses a data stitching method based on a calibration block for the point cloud data on the skateboard surface to complete rough registration;

[0020] The sixth unit uses an improved ICP registration algorithm to complete the fine registration of the point cloud data and obtain the complete point cloud data of the skateboard surface.

[0021] Compared with the prior art, the significant advantages of the present invention are as follows: (1) While ensuring the accuracy of the point cloud reconstruction of the pantograph skateboard surface, the real-time performance of the detection algorithm is improved; (2) A pantograph skateboard point cloud filtering algorithm based on the slice contour is adopted, and the interference of non-pantograph skateboard data is filtered through hybrid filtering, improving the detection accuracy; (3) Combining a data fast reading and stitching method based on a calibration block and an improved ICP registration algorithm for fast and fine registration of the pantograph skateboard point cloud data, realizing fast and accurate stitching of the pantograph skateboard laser point cloud data. Description of the Drawings

[0022] Figure 1 is a flowchart of the method for stitching the pantograph skateboard point cloud data based on a calibration block and an improved ICP algorithm.

[0023] Figure 2 is a schematic diagram of the system equipment layout of the on-site acquisition module in the embodiment of the present invention.

[0024] Figure 3 is a schematic diagram of the sensor installation of the on-site acquisition module in the embodiment of the present invention.

[0025] Figure 4 is a schematic diagram of the change trend of valid data points in the embodiment of the present invention.

[0026] Figure 5 is a point cloud diagram after hybrid filtering and denoising of the first set of laser displacement sensors in the embodiment of the present invention.

[0027] Figure 6 is a point cloud diagram after hybrid filtering and denoising of the second set of laser displacement sensors in the embodiment of the present invention.

[0028] Figure 7 is an original data diagram collected by the two sets of laser displacement sensors in the embodiment of the present invention.

[0029] Figure 8 It is the transformation result diagram of transforming the data of the first group of laser displacement sensors into the coordinate system of the second group of laser displacement sensors in the embodiment of the present invention.

[0030] Figure 9 It is the schematic diagram of the skateboard surface point cloud data obtained by the rough registration of the skateboard point cloud in the embodiment of the present invention.

[0031] Figure 10 It is the schematic diagram of the skateboard surface point cloud data obtained by the fine registration of the skateboard point cloud in the embodiment of the present invention. Specific Embodiments

[0032] As Figure 1 shown, a method for splicing pantograph skateboard point cloud data based on a calibration block and an improved ICP algorithm of the present invention includes the following steps:

[0033] Step 1, arrange a field acquisition module to scan the skateboard contour data in real time during the train operation;

[0034] Step 2, use two groups of laser displacement sensors to continuously output multiple frames of valid data and expand it into three-dimensional point cloud data;

[0035] Step 3, use a denoising algorithm of hybrid filtering to process the skateboard data to obtain a skateboard contour data set;

[0036] Step 4, use the skateboard contour data set to convert the point cloud data collected by the two groups of laser displacement sensors into the same coordinate system;

[0037] Step 5, use a data splicing method based on a calibration block for the skateboard surface point cloud data to complete rough registration;

[0038] Step 6, use an improved ICP registration algorithm to complete the fine registration of the point cloud data and obtain the complete skateboard surface point cloud data.

[0039] As a specific example, for the arrangement of the field acquisition module in Step 1 to scan the skateboard contour data in real time during the train operation, it is as follows:

[0040] The field acquisition module includes a first group of laser displacement sensors, a second group of laser displacement sensors, a first group of photoelectric sensors, and a second group of photoelectric sensors, where:

[0041] The first group of laser displacement sensors and the second group of laser displacement sensors are installed on the same platform above the catenary. The coordinate axes of the two groups of laser displacement sensors are parallel to each other, and the laser detection surfaces are in the same plane, vertically downward to collect the upper surface contour of the skateboard; the upper surface of the skateboard is within the effective measurement range of the detection surfaces of the two groups of laser displacement sensors, and a set overlapping area is reserved;

[0042] The distance between the first group of optoelectronic sensors and the second group of optoelectronic sensors is , and the transmitting end and the receiving end of each group of optoelectronic sensors are respectively installed on both sides of the pantograph. The two groups of optoelectronic sensors are used to collect the skateboard speed in real time.

[0043] As a specific example, the use of two groups of laser displacement sensors to continuously output multiple frames of valid data and expand them into three-dimensional point cloud data in step 2 is specifically as follows:

[0044] Step 2.1: Use two groups of laser displacement sensors for continuous sampling and output multiple frames of valid data;

[0045] Step 2.2: Use two groups of optoelectronic sensors to collect the skateboard speed in real time, combine the sampling frequency to realize the acquisition of the third-dimensional data, use the train forward direction as the Y-axis, expand the two-dimensional contour data into three-dimensional point cloud data, and multiple groups of three-dimensional contour point cloud data form the upper surface of the skateboard. Calculate the frame spacing according to the following formula Calculation:

[0046]

[0047] In the formula, , are respectively the moments when the skateboard passes through the first group of optoelectronic sensors and the second group of optoelectronic sensors, is the sampling frequency of the first group of laser displacement sensors and the second group of laser displacement sensors.

[0048] As a specific example, the use of a denoising algorithm with hybrid filtering to process the skateboard data in step 3 to obtain a skateboard contour data set is specifically as follows:

[0049] Step 3.1: The collected data includes target point cloud data and non-target point cloud data. The skateboard point cloud data is composed of multiple groups of two-dimensional scanning contours. Each contour is equivalent to a slice of the point cloud data. Transfer the filtering of the three-dimensional point cloud data to the two-dimensional space to obtain the slice contour data;

[0050] Step 3.2: Use a denoising algorithm with hybrid filtering to filter the slice contour data, filter out the non-target point cloud data, and obtain a skateboard contour data set.

[0051] As a specific example, the use of a denoising algorithm with hybrid filtering to filter the slice contour data in step 3.2, filter out the non-target point cloud data, and obtain a skateboard contour data set is specifically as follows:

[0052] Step 3.2.1: Continuously number all the collected data frames in sequence according to the sampling time. During the detection process, there are two peaks in the curve of the effective number of collected points output by the laser displacement sensor, corresponding to the effective number of data points when the two sliding plates of the pantograph pass through the detection area. The effective data frames of the sliding plate are screened by judging whether the effective number of data points output by the laser displacement sensor is greater than a fixed threshold.

[0053] Step 3.2.2: Use a radius filter to filter out discrete points. Determine whether a point is an effective point by comparing whether the number of points within the detection radius of each data point in the point cloud data is less than the set threshold. If it is less than the threshold, it is confirmed as an outlier and the outlier is removed; otherwise, it is determined as an effective point.

[0054] Step 3.2.3: Calculate the Euclidean distance between two adjacent data points in the point cloud data, and compare the Euclidean distance with the threshold to segment the point cloud data: if it is less than the threshold, add the subsequent data points to the current same data point set; if it is greater than the threshold, use the current data point as the segmentation point, search backward from the adjacent data points, and take the largest data point set as the sliding plate contour data set.

[0055] As a specific example, for the step 4 of using the sliding plate contour data set to convert the point cloud data collected by two groups of laser displacement sensors into the same coordinate system, the specific method is as follows:

[0056] Step 4.1: The coordinate systems of the first group of laser displacement sensors and the second group of laser displacement sensors are respectively and , assume that the position coordinate of point in is . After rotation and translation transformation, the position coordinate of point in is , then there is:

[0057]

[0058] In the formula, is a 3 3 rotation matrix, is the translation vector, is the component of the translation vector on the axis, axis, axis;

[0059] Rotate the coordinate system around the axis, axis, axis by the rotation angles , , The obtained rotation matrix is:

[0060]

[0061]

[0062]

[0063] Combining the three rotation matrices to obtain the rotation transformation matrix axis, axis, axis and rotated by , , angles respectively, as follows: , as shown in the following formula:

[0064]

[0065] Step 4.2, Solve the specific values of the parameters , , , , and to splice the data collected by the two groups of laser displacement sensors to obtain the complete point cloud data of the skateboard surface.

[0066] As a specific example, the data splicing method based on the calibration block for the point cloud data of the skateboard surface described in step 5 is completed as follows:

[0067] Step 5.1, Use a rectangular metal calibration block with a size of to calibrate the spatial position relationship of the coordinate system. Place the calibration block within the effective measurement range of the laser displacement sensor, and adjust the installation parameters of the laser displacement sensor according to the positions of the laser contour lines emitted by the two groups of laser displacement sensors on the calibration block until the two laser contour lines coincide finally; , , respectively represent the length, width, and height of the calibration block;

[0068] The angles of the coordinate system around the axis and the axis are close to 0, and at the same time, the translation parameter in the axis direction is also close to 0, that is, the parameters , and are approximately 0, and the registration parameters are reduced to , and ;

[0069] Step 5.2: Taking the coordinate system of the second group of laser displacement sensors as the target coordinate system, rotate and translate the coordinate system of the first group of laser displacement sensors to . The conversion formula is:

[0070]

[0071] Project it onto plane, then there is:

[0072]

[0073] Solve the parameters , and , that is, complete the splicing of skateboard data based on the calibration block;

[0074] Step 5.3: Calculate the rotation angle according to the feature that the two laser lines irradiate on the same straight line on the calibration block. Perform linear fitting on the data collected by the two groups of laser displacement sensors. During the calibration process, the straight lines fitted by the detection data of the two groups of laser displacement sensors are , respectively. The slopes of the straight lines are , respectively. The inclination angles are , respectively. Then the rotation angle is:

[0075]

[0076] Make the laser irradiate on the same straight line of the calibration block, and each of the two groups of laser displacement sensors detects a boundary point of the calibration block. During the calibration process, the coordinates of the boundary point of the calibration block in the coordinate systems of the first group and the second group of laser displacement sensors are , respectively. After rotating the rotation angle , the rotated coordinates of the point are . The conversion formula is:

[0077]

[0078] After the coordinate system of the first group of laser displacement sensors is rotated, the coordinate axes are parallel to those of the second group of laser displacement sensors, but there is a spatial position difference in the coordinate origin. According to the feature of the fixed length of the calibration block, the translation distance of the coordinate origin in the axis direction and the translation distance in the Perform calculations, and the calculation formula is:

[0079]

[0080] To improve the accuracy of rough registration and reduce the influence of random errors, repeat steps 5.3 to 5.4 for multiple calibrations, and use the mean value as the final calibration value.

[0081] As a specific example, the fine registration of point cloud data is completed using the improved ICP registration algorithm in step 6 to obtain the complete point cloud data of the skateboard surface, as follows:

[0082] Step 6.1: The data of the first group of laser displacement sensors constitute the source point cloud dataset , and the data of the second group of laser displacement sensors constitute the target point cloud dataset , represents the three-dimensional real number set, and respectively represent the sizes of the two point cloud datasets corresponding to the first and second groups of laser displacement sensors, , respectively represent the data point numbers in the two point cloud datasets corresponding to the first and second groups of laser displacement sensors; the source point cloud dataset and the target point cloud dataset find matching point pairs according to the set constraint conditions , represents the matching point pair number;

[0083] Use the least squares algorithm to calculate the optimal rotation matrix , translation matrix , so that the registration error function is minimized, and the error function is defined as follows:

[0084]

[0085] In the formula, , is the number of matching point pairs;

[0086] The ultimate goal of point cloud registration is to solve the rotation matrix and translation matrix , so that the two groups of laser point clouds after registration form the complete point cloud of the skateboard surface;

[0087] First, calculate the centroids and of , , and the covariance matrix , of :

[0088]

[0089]

[0090] Then, for the covariance matrix perform singular value decomposition , , is a 3×3 orthogonal matrix, is a non - negative diagonal matrix composed of the eigenvalues of the covariance matrix, then the rotation matrix and the translation matrix are calculated as follows:

[0091]

[0092] Step 6.2. Since there must be an inclusion relationship between the target point cloud and the source point cloud, and the judgment of corresponding points depends on the Euclidean distance, which may lead to incorrect matching or getting stuck in local minimum problems. Combining the characteristics of the pantograph slider and the sensor scanning characteristics, the ICP algorithm is optimized from three aspects: matching point set, removal of wrong points, and construction of error function, to improve the matching speed and accuracy, as follows:

[0093] Step 6.2.1. Matching point set: Calculate the overlapping area of the two groups of laser point clouds, and use the data in the overlapping area for registration, instead of making all points participate in the operation;

[0094] The two groups of laser point clouds are partially overlapping. Incorporating the data in the non - overlapping area into the matching process will not only greatly increase the computational complexity of the algorithm, but also provide wrong matching point pairs that affect the registration accuracy of the algorithm. Therefore, extract the overlapping area of the two groups of point clouds for point cloud registration to reduce the computational complexity of the algorithm. Set the minimum value of the first group of laser displacement sensor point cloud data in the axis direction as , and the maximum value of the second group of laser displacement sensor point cloud data in the axis direction as , then the overlapping area in the X - axis direction of the two is ;

[0095] Step 6.2.2. Removal of wrong points: In actual situations, due to the influence of the slider structure characteristics and noise points, there will inevitably be wrong matching point pairs. Sort all the matching point pairs in ascending order of distance, and then use the fixed - ratio method to retain the first point pairs for the calculation of the transformation matrix, , and remove the remaining point pairs; where is the set fixed ratio, is the number of retained point pairs;

[0096] Step 6.2.3, Error Function Construction: Optimize the error function selection strategy in the iterative stage, combine the point-to-point error function and the point-to-plane error function. The construction method of the point-to-point error function means constructing the error function with the sum of the squares of the distances between all corresponding points in two point clouds. The construction method of the point-to-plane error function means constructing the error function with the sum of the squares of the distances from the points in the source point cloud to the tangent planes of their corresponding points. In the previous multiple iterations where the error is greater than the threshold, the point-to-point error function is selected, and in the subsequent iterations, the point-to-plane error function is selected;

[0097] Point-to-plane error function As shown in the following formula:

[0098]

[0099] In the formula, is the point in , is 's three-axis coordinate components; is the point in 's corresponding point, is 's three-axis coordinate components; is 's normal vector, , is 's three-axis coordinate components; is 's rigid body transformation matrix, as shown in the following formula:

[0100]

[0101] When the rotation angle approaches 0, the point-to-plane error function is approximately optimized into a linear least squares problem. Therefore, when , , approach 0, the rotation matrix is approximately :

[0102]

[0103] The rigid body transformation matrix is approximately expressed as :

[0104]

[0105] Point-to-plane error function is approximately expressed as :

[0106]

[0107] For each point pair in the above formula can all be written as an expression containing parameters 、 、 、 、 and :

[0108]

[0109] For the matching point pairs, the following expressions are obtained:

[0110]

[0111] where

[0112]

[0113]

[0114]

[0115] and are constant terms; in , , , ;

[0116] The optimal solution of the transformation parameter is:

[0117]

[0118] The above formula is a standard linear least squares problem and is solved by SVD.

[0119] The present invention also provides a pantograph slider point cloud data stitching system based on a calibration block and an improved ICP algorithm. This system is used to implement the pantograph slider point cloud data stitching method based on the calibration block and the improved ICP algorithm. The system includes a first unit to a sixth unit, and the functions of each unit are as follows:

[0120] The first unit, through the on-site acquisition module, scans the slider contour data in real time during the train operation;

[0121] The second unit, uses two groups of laser displacement sensors to continuously output multiple frames of valid data and expands it into three-dimensional point cloud data;

[0122] The third unit processes the skateboard data using a denoising algorithm with hybrid filtering to obtain a skateboard contour data set;

[0123] The fourth unit uses the skateboard contour data set to convert the point cloud data collected by two groups of laser displacement sensors into the same coordinate system;

[0124] The fifth unit completes rough registration of the skateboard surface point cloud data using a data stitching method based on a calibration block;

[0125] The sixth unit uses an improved ICP registration algorithm to complete the fine registration of the point cloud data and obtains the complete skateboard surface point cloud data.

[0126] The following further elaborates on the present invention in detail in conjunction with the accompanying drawings and specific embodiments.

[0127] Embodiment

[0128] Combined with Figure 1 , a method for stitching the pantograph skateboard point cloud data based on a calibration block and an improved ICP algorithm provided in this embodiment includes the following steps:

[0129] Step 1: Arrange a field acquisition module to scan the skateboard contour data in real time during the train operation, specifically as follows:

[0130] Arrange a field acquisition module, including two groups of 2D laser displacement sensors and two groups of photoelectric sensors. The 2D laser displacement sensors are of the Keyence LJ-X890 type, and the photoelectric sensors are of the Keyence PZ-G52CP photoelectric sensors;

[0131] The two groups of 2D laser displacement sensors are installed symmetrically about the center line of the rail and collect skateboard data vertically downward. The installation schematic is as shown in Figure 2 , Figure 3 . The two groups of 2D laser displacement sensors are installed above the contact wire and collect the upper surface contour of the skateboard vertically downward. The laser detection surfaces of the two groups of 2D laser displacement sensors are in the same plane. The upper surface of the skateboard is within the effective measurement range of the laser detection surface, and a certain overlapping area is reserved. Moreover, the two laser sensors are installed on the same platform, and the coordinate axes are parallel to each other; the distance between the two groups of photoelectric sensors is d, and the transmitting end and receiving end of the photoelectric sensors are respectively installed on both sides of the pantograph. Two groups of photoelectric sensors are used to realize the real-time acquisition of the skateboard speed.

[0132] Step 2: Use the laser displacement sensor to continuously output multiple frames of valid data and expand it into three-dimensional point cloud data, specifically as follows:

[0133] Step 2.1: Use the laser displacement sensor to perform continuous sampling and output multiple frames of valid data;

[0134] Step 2.2: Use a photoelectric sensor to collect the skateboard speed in real time, and combine the collection frequency to achieve the collection of the third-dimensional data. Take the train's forward direction as the Y-axis, expand the two-dimensional contour data into three-dimensional point cloud data. Multiple groups of three-dimensional contour point clouds form the upper surface of the skateboard. Calculate the frame spacing according to the following formula: Calculation:

[0135]

[0136] In the formula, is the distance between the first group of photoelectric sensors and the second group of photoelectric sensors, , are the moments when the skateboard passes through the first group of photoelectric sensors and the second group of photoelectric sensors respectively, is the collection frequency of the first group of laser displacement sensors and the second group of laser displacement sensors.

[0137] Step 3: Use a denoising algorithm of hybrid filtering to process the skateboard data to obtain a skateboard contour data set, specifically as follows:

[0138] Step 3.1: The data collected by the system includes point cloud target data and non-target point cloud data. The skateboard point cloud data consists of multiple groups of two-dimensional scanning contours. Each contour is equivalent to a slice of the point cloud. Transfer the filtering of the three-dimensional point cloud data to the two-dimensional space and perform filtering processing on the slice contour data;

[0139] Step 3.2: Use a denoising algorithm of hybrid filtering to process the skateboard data, filter out the non-target object point cloud data, and obtain a skateboard contour data set, specifically as follows:

[0140] Step 3.2.1: Take the single-side sensor system as an example. When a complete pantograph enters the detection area and then leaves the detection area, the change trend of the number of valid points output by the laser displacement sensor is as Figure 4 shown. When the pantograph is not in the measurement area, the number of valid data points output by the laser displacement sensor is 0; when the pantograph enters the measurement area, the number of valid data points output by the laser displacement sensor will increase significantly. During the detection process, there are two peaks in the curve of the number of valid acquisition points output by the laser displacement sensor, corresponding to the number of valid data points when the two skateboards of the pantograph pass through the detection area;

[0141] Step 3.2.2: Use a radius filter to remove discrete points. The principle of radius filtering is that each data point in the point cloud has at least a specified number of neighborhood points in the specified radius neighborhood. By comparing whether the number of points within its detection radius of each point is less than the set threshold to determine whether it is a valid point. If it is less than the threshold, it is identified as an outlier and removed; otherwise, it is determined as a valid point;

[0142] Step 3.2.3: There is a certain gap between the pantograph and the skateboard assembly. The laser displacement sensor is installed vertically downward, and the value of the gap data in the Z-axis direction has a certain amplitude mutation compared with the data at other positions. Based on this feature, the data points are segmented.

[0143] Start searching in order from the starting point and calculate the Euclidean distance between adjacent two points , and compare it with the threshold . If it meets , then add the subsequent points to the same point set ; if , then use the point as the segmentation point, and continue to search for the next point set starting from . After the data segmentation is completed, count the number of data points in each point set , and take the point set with the largest number of data points as the skateboard data. The filtered data is as shown in Figure 5 , Figure 6 .

[0144] Step 4: Use the skateboard contour data set to convert the two sets of laser point cloud data collected by the two laser displacement sensors into the same coordinate system to form the complete skateboard surface point cloud data, specifically as follows:

[0145] Step 4.1: The coordinate systems of the first group of laser displacement sensors and the second group of laser displacement sensors are respectively and . Let the position coordinates of the point in be . After rotation and translation transformation, the position coordinates of the point in are , then there is:

[0146]

[0147] In the formula, is a 3 × 3 rotation matrix, is the translation vector, is the component of the translation vector on the axis, axis, axis;

[0148] Rotate the coordinate system around the axis, axis, axis by the rotation angle , , The obtained rotation matrix is:

[0149]

[0150]

[0151]

[0152] Combining the three rotation matrices to obtain the rotation transformation matrix axis, axis, axis respectively rotated , , angle, as follows: , as shown in the following formula:

[0153]

[0154] Step 4.2, Solve the specific values of the parameters , , , , and so that the data collected by the two sets of laser displacement sensors are stitched to obtain the complete point cloud data of the skateboard surface.

[0155] Step 5, Complete the rough registration of the point cloud data of the skateboard surface by using the data fast stitching method based on the calibration block, specifically as follows:

[0156] Step 5.1, Calibrate the space of the two sets of laser displacement sensors by using a calibration block with a size of ;

[0157] Step 5.2, Place the calibration block in the effective measurement area of the system. First, fine-tune the position of the sensor according to the laser contour line so that the two laser contour lines coincide. The original data collected by the laser displacement sensor is as Figure 7 shown. Linearly fit the surface contour of the calibration block and extract the coordinate values of the two side endpoints to obtain the rough registration parameters , , , . Rotate the data of the first set of laser displacement sensors to the coordinate system of the second set of laser displacement sensors. The transformation result is Figure 8 shown.

[0158] Step 5.3: Based on the rough registration parameters of the sensors, rotate and transform the surface point cloud collected by the first group of laser displacement sensors to the coordinate system of the second group of laser displacement sensors to complete the rough registration of the skateboard point cloud. The rough registration result is as shown in Figure 9 shown.

[0159] Step 6: Use the improved ICP registration algorithm to complete the fast fine registration of the point cloud data and obtain the complete skateboard surface point cloud data, which is specifically as follows:

[0160] Step 6.1: Select the minimum value of the first group of laser point cloud data in the axis direction as , and the maximum value of the second group of laser point cloud data in the axis direction as . The area in the X-axis direction of the two is used as the overlapping area of the two groups of point clouds;

[0161] Step 6.2: Sort all the matched points according to the distance according to Step 6.1, and then use the fixed ratio method to remove a part of the point pairs with too large distances. The remaining point pairs are used for the calculation of the transformation matrix.

[0162] Step 6.3: Combine the "point-to-point" and "point-to-plane" error functions, set the rotation angle between the two groups of point clouds to be very small, and approximate and optimize the formula into a linear least squares problem. The registration visual effect is as shown in Figure 10 shown.

[0163] It can be seen from the figure that the present invention adopts a pantograph skateboard point cloud filtering algorithm based on the slice contour, filters out the interference of non-pantograph skateboard data through hybrid filtering, combines the data fast reading and stitching method based on the calibration block and the improved ICP registration algorithm for the fast fine registration of the pantograph skateboard point cloud data, improves the efficiency of the point cloud data stitching at the same time, and improves the real-time performance of the detection algorithm.

[0164] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Therefore, any modifications, equivalent changes and modifications made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A method for splicing the point cloud data of a pantograph slider based on a calibration block and an improved ICP algorithm, characterized in that, It includes the following steps: Step 1: Arrange the on-site acquisition module to scan the profile data of the skateboard in real time during the train operation; Step 2: Use two sets of laser displacement sensors to continuously output multiple frames of valid data and expand it into three-dimensional point cloud data; Step 3: Process the skateboard data using a denoising algorithm with hybrid filtering to obtain a skateboard profile data set, specifically as follows: Step 3.1: The collected data includes target point cloud data and non-target point cloud data. The skateboard point cloud data consists of multiple sets of two-dimensional scan profiles, and each profile is equivalent to a slice of the point cloud data. Transfer the filtering of the three-dimensional point cloud data to the two-dimensional space to obtain slice profile data; Step 3.2: Use a denoising algorithm with hybrid filtering to filter the slice profile data to filter out non-target point cloud data and obtain a skateboard profile data set, specifically as follows: Step 3.2.1: Continuously number all the collected data frames in sequence according to the sampling time. During the detection process, there are two peaks in the curve of the effective acquisition point number output by the laser displacement sensor, corresponding to the effective data point numbers of the two skateboards of the pantograph passing through the detection area. Screen the valid data frames of the skateboard by judging whether the effective data point number output by the laser displacement sensor is greater than a fixed threshold; Step 3.2.2: Use a radius filter to filter out discrete points. Determine whether a point is a valid point by comparing whether the number of points within the detection radius of each data point in the point cloud data is less than the set threshold. If it is less than the threshold, it is confirmed as an outlier and the outlier is removed. Otherwise, it is determined as a valid point; Step 3.2.3: Calculate the Euclidean distance between two adjacent data points in the point cloud data, compare the Euclidean distance with the threshold, and segment the point cloud data: if it is less than the threshold, add the subsequent data points to the current same data point set; if it is greater than the threshold, use the current data point as the segmentation point and search backward from the adjacent data points, and take the largest data point set as the skateboard profile data set; Step 4: Use the skateboard profile data set to convert the point cloud data collected by the two sets of laser displacement sensors into the same coordinate system; Step 5: Complete the rough registration of the point cloud data on the skateboard surface using a data stitching method based on a calibration block; Step 6: Use an improved ICP registration algorithm to complete the fine registration of the point cloud data and obtain the complete point cloud data of the skateboard surface, specifically as follows: Step 6.1: The data of the first group of laser displacement sensors constitutes the source point cloud data set , and the data of the second group of laser displacement sensors constitutes the target point cloud data set , represents the three-dimensional real number set, and represent the sizes of the two point cloud data sets corresponding to the first and second groups of laser displacement sensors respectively, , represent the data point numbers in the two point cloud data sets corresponding to the first and second groups of laser displacement sensors respectively; the source point cloud data set and the target point cloud data set find matching point pairs according to the set constraint conditions , represents the matching point pair number; Calculate the optimal rotation matrix using the least squares algorithm , translation matrix , so that the registration error function is minimized. The error function is defined as follows: ； In the formula, , is the number of matching point pairs; The ultimate goal of point cloud registration is to solve the rotation matrix and the translation matrix so that the two sets of registered laser point clouds form a complete point cloud of the skateboard surface; First, calculate and centroids 、 , and 、 covariance matrices : ； ； Then, for the covariance matrix perform singular value decomposition , , is a 3×3 orthogonal matrix, is a non-negative diagonal matrix composed of the eigenvalues of the covariance matrix, then the rotation matrix and the translation matrix are calculated as follows: ； Step 6.2: Optimize the ICP algorithm from three aspects: matching point set, error point removal, and error function construction.

2. The method for splicing the pantograph slider point cloud data based on the calibration block and the improved ICP algorithm according to claim 1, wherein The arrangement of the on-site acquisition module described in Step 1 to scan the profile data of the skateboard in real time during the train operation is specifically as follows: The on-site acquisition module includes a first set of laser displacement sensors, a second set of laser displacement sensors, a first set of photoelectric sensors, and a second set of photoelectric sensors, where: The first set of laser displacement sensors and the second set of laser displacement sensors are installed on the same platform above the catenary. The coordinate axes of the two sets of laser displacement sensors are parallel to each other, and the laser detection surfaces are in the same plane, vertically downward to collect the profile of the upper surface of the skateboard; the upper surface of the skateboard is within the effective range of the detection surfaces of the two sets of laser displacement sensors, and a set overlap area is reserved; The first group of optoelectronic sensors and the second group of optoelectronic sensors are spaced apart by , and the transmitting end and the receiving end of each group of optoelectronic sensors are respectively installed on both sides of the pantograph. The two groups of optoelectronic sensors are used to collect the skateboard speed in real time.

3. The method for splicing the pantograph slider point cloud data based on the calibration block and the improved ICP algorithm according to claim 2, characterized in that, The continuous output of multiple frames of valid data by using two sets of laser displacement sensors described in Step 2 is expanded into three-dimensional point cloud data, specifically as follows: Step 2.1: Continuously sample using two sets of laser displacement sensors and output multiple frames of valid data; Step 2.2: Use two sets of photoelectric sensors to collect the skateboard speed in real time, combine the collection frequency to achieve the collection of the third-dimensional data, take the train forward direction as the Y-axis, expand the two-dimensional contour data into three-dimensional point cloud data, and multiple sets of three-dimensional contour point cloud data constitute the upper surface of the skateboard. Calculate the frame spacing according to the following formula Calculation: ； Wherein, and are the moments when the skateboard passes through the first group of photoelectric sensors and the second group of photoelectric sensors respectively, is the acquisition frequency of the first group of laser displacement sensors and the second group of laser displacement sensors.

4. The method for splicing the pantograph slider point cloud data based on the calibration block and the improved ICP algorithm according to claim 3, characterized in that, The conversion of the point cloud data collected by two sets of laser displacement sensors into the same coordinate system by using the skateboard contour data set described in Step 4 is as follows: Step 4.1, the coordinate systems of the first group of laser displacement sensors and the second group of laser displacement sensors are respectively and , let the point be in with the position coordinate being , after rotation and translation transformation, the position coordinate of the point in is , then there is: ； Wherein, is the rotation matrix of 3 3, is the translation vector, is the component of the translation vector on the axis, axis, axis components; Rotate the coordinate system around the axis, axis, axis by the angles , , respectively. The resulting rotation matrix is: ； ； ； Combining three rotation matrices to obtain the rotation transformation matrix axis, axis, axis, respectively rotated by , , angles, as shown in the following formula: , as follows: ； Step 4.2, Solve the parameters , , , , and to splice the data collected by the two sets of laser displacement sensors to obtain the complete point cloud data of the skateboard surface.

5. The method for splicing the pantograph slider point cloud data based on the calibration block and the improved ICP algorithm according to claim 4, characterized in that, The rough registration of the point cloud data on the skateboard surface by using the data stitching method based on the calibration block described in Step 5 is as follows: Step 5.

1. Use a rectangular parallelepiped metal calibration block with a size of to calibrate the spatial position relationship of the coordinate system. Place the calibration block within the effective measurement area of the laser displacement sensor, and adjust the installation parameters of the laser displacement sensor based on the positions of the laser contour lines emitted by the two groups of laser displacement sensors on the calibration block, and finally adjust the two laser contour lines to coincide; , , respectively represent the length, width, and height of the calibration block; The coordinate system rotates around axis and the angle of the axis is close to 0, and at the same time the translation parameter in the direction is also close to 0, that is, the parameters and are approximately 0, reducing the registration parameters to 、 and ; Step 5.2: Using the coordinate system of the second group of laser displacement sensors as the target coordinate system, rotate and translate the coordinate system of the first group of laser displacement sensors to . The conversion formula is as follows: ； Projected onto the plane, we have: ； Solve parameters , and , and the splicing of skateboard data based on the calibration block is completed; Step 5.

3. Calculate the rotation angle based on the feature that two laser lines irradiate on the same straight line of the calibration block. Perform linear fitting on the data collected by the two groups of laser displacement sensors. During the calibration process, the straight lines fitted by the detection data of the two groups of laser displacement sensors are respectively , , . The slopes of the straight lines are respectively , . The inclination angles are respectively , . Then the rotation angle is: ； Step 5.4: Irradiate the laser on the same straight line of the calibration block, and each of the two groups of laser displacement sensors detects a boundary point of the calibration block. During the calibration process, the coordinates of the boundary point of the calibration block in the coordinate systems of the first group and the second group of laser displacement sensors are respectively , . Rotate the angle . The rotated coordinates of the point are . The conversion formula is: ； After the coordinate system of the first group of laser displacement sensors is rotated, the coordinate axes are parallel to those of the second group of laser displacement sensors, but there is a spatial position difference between the coordinate origins. According to the characteristic of the fixed length of the calibration block, the translation distances of the coordinate origin in the axis direction and axis direction are calculated. The calculation formula is: ； Repeat Steps 5.3 to 5.4 for multiple calibrations, and take the mean value as the final calibration value.

6. The method for splicing the pantograph slider point cloud data based on the calibration block and the improved ICP algorithm according to claim 5, wherein The optimization of the ICP algorithm from three aspects of matching point set, error point removal, and error function construction described in Step 6.2 is as follows: Step 6.2.1: Matching point set: Calculate the overlapping area of two sets of laser point clouds and perform registration using the overlapping area data; Step 6.2.2, Error Point Removal: For all matching point pairs sort them in ascending order of distance, and then use the fixed ratio method to retain the first point pairs for the calculation of the transformation matrix, , and remove the remaining point pairs; where is the set fixed ratio, is the number of retained point pairs; Step 6.2.3: Error function construction: Optimize the error function selection strategy in the iterative stage, combine the point-to-point error function and the point-to-plane error function. The construction method of the point-to-point error function means constructing the error function with the sum of the squares of the distances between all corresponding points in two pieces of point clouds. The construction method of the point-to-plane error function means constructing the error function with the sum of the squares of the distances from the points in the source point cloud to the tangent planes of their corresponding points. In the previous multiple iterations where the error is greater than the threshold, the point-to-point error function is selected, and the point-to-plane error function is selected in the subsequent iterations; Point-to-plane error function As shown in the following equation: ； wherein, is a point in is the three-axis coordinate components of is the point in the corresponding point in is the three-axis coordinate components of is the normal vector of , is the three-axis coordinate components of is the rigid body transformation matrix of , as shown in the following formula: ； When the rotation angle approaches 0, the point-to-plane error function is approximately optimized into a linear least squares problem. Therefore, when and and approach 0, the rotation matrix is approximately : ； Rigid body transformation matrix Approximately represented as : ； Point-to-plane error function Approximately expressed as : ； For each point pair in the above formula can all be written as an expression containing the parameters , , , , and : ； For For the matching point pairs, the following expression is obtained: ； Where ； ； ； and is a constant term; in , , , ; Optimal solution of transformation parameters is as follows: ； The above formula is a standard linear least squares problem and is solved by SVD.

7. A pantograph slider point cloud data stitching system based on a calibration block and an improved ICP algorithm, characterized in that, This system is used to implement the method for stitching the point cloud data of the pantograph skateboard based on the calibration block and the improved ICP algorithm described in any one of Claims 1 to 6. The system includes a first unit to a sixth unit, and the functions of each unit are as follows: The first unit, through the on-site acquisition module, scans the skateboard contour data in real time during the train operation; The second unit, uses two sets of laser displacement sensors to continuously output multiple frames of valid data and expands it into three-dimensional point cloud data; The third unit, processes the skateboard data by using a denoising algorithm of hybrid filtering to obtain the skateboard contour data set; The fourth unit, uses the skateboard contour data set to convert the point cloud data collected by two sets of laser displacement sensors into the same coordinate system; The fifth unit, completes the rough registration of the point cloud data on the skateboard surface by using the data stitching method based on the calibration block; The sixth unit, uses the improved ICP registration algorithm to complete the fine registration of the point cloud data and obtains the complete point cloud data of the skateboard surface.

Images