A train wheel defect detection method based on three-dimensional information

Abstract

The application discloses a train wheel defect detection method based on three-dimensional information, which comprises data acquisition, data preprocessing, laser line extraction, laser coordinate transformation, three-dimensional wheel splicing and defect detection. The application adopts the combination of multiple linear array cameras and 3D laser scanner modules to generate multiple two-dimensional image data and laser scanning data with depth information, extracts wheel laser center line data through a complex image processing algorithm, reconstructs the three-dimensional profile of the wheel with depth information, splices the three-dimensional information of the complete wheel, compares the local depth information, finds out the area with large depth change, finally detects the defects of the tread and the rim surface of the complete wheel, obtains the information such as the area, position and depth of the defect area, and dynamically generates a detection report for the customer to review and repair in time.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of rail transit, and particularly relates to a train wheel defect detection method based on three-dimensional information and applied to a train wheel set online detection product. BACKGROUND

[0002] With the rapid development of rail transit in China, traditional manual wheel defect detection cannot meet the daily operation requirements, and the emergence of high-precision intelligent detection equipment ensures the safe operation of trains. In train operation, the wheel is a key component of driving safety, and the wheel in motion may have defects such as wear, scratches, and peeling. These serious defects may cause train derailment accidents, so it is necessary to dynamically detect the train wheel.

[0003] At present, most wheel detection equipment uses two-dimensional cameras to obtain wheel image data. The tread image obtained by the two-dimensional camera cannot detect the depth information of the defect area, and the image quality is low and is greatly affected by environmental factors, which is not conducive to the analysis and processing of the wheel image. Laser scanners with depth information are gradually used in wheel online detection systems.

[0004] Chinese patent ZL202110047893 discloses a wheel tread defect three-dimensional detection method and system. The method uses a grating projection method to obtain three-dimensional data of the wheel tread, and then reconstructs the three-dimensional point cloud image of the wheel. Then, the three-dimensional image data of the standard wheel tread is compared to obtain the defect area of the wheel tread. This method only completes the three-dimensional reconstruction of part of the wheel area and does not reconstruct the complete three-dimensional wheel data, so it cannot detect defects in all directions of the wheel. Moreover, the method uses a comparison method to detect defects, which consumes more storage space. Chinese patent ZL201410798858 discloses a tread defect information detection system and method. The method obtains multiple curved surface image information of the wheel tread to construct the spatial curved surface information of the train wheel tread, and compares the constructed spatial curved surface information with the preset tread curved surface information to obtain the defect information of the tread of the train wheel. The tread defect detection module detects the tread defects by comparison. This method consumes more storage space and more processing time. SUMMARY

[0005] In view of the shortcomings and deficiencies of the prior art, the present application aims to provide a train wheel defect detection method based on three-dimensional information.

[0006] To solve the above technical problems, the present application adopts the following technical solutions:

[0007] A train wheel defect detection method based on three-dimensional information, comprising:

[0008] Image acquisition and preprocessing to obtain laser line image;

[0009] Laser center line extraction: extract the center line of the laser image, two-dimensional coordinates are (x, y);

[0010] Laser coordinate transformation: convert the (x, y) of the image point set generated by the extracted laser line on the wheel surface into the real world coordinate system coordinates (X w , Y w , Z w );

[0011] Three-dimensional wheel data splicing: taking the coordinate system of a certain acquisition device as the reference and defining it as the world coordinate system, the coordinate systems of other acquisition devices are subjected to affine transformation, and the overlapping areas are fused to obtain the three-dimensional information of the complete wheel;

[0012] Defect detection: according to the local depth information of the wheel, a multi-level judgment strategy is used to complete defect detection.

[0013] Further, the defect detection: first, take a row of data along the wheel tread direction, and compare the depth information in the column direction with the adjacent row, mark the points with larger values as 1 and the points with smaller values as 0; Then traverse the entire wheel to find the area with large changes, that is, all areas marked as 1 are suspicious defect positions; Finally, the local clustering processing is carried out on the areas marked as 1, and the areas after clustering are compared one by one to find all areas that meet the conditions, which are defects. The area, position and maximum depth information of the defect area can be calculated by counting the defect area.

[0014] Further, the laser center line extraction: the preprocessed laser line image is adjusted and enlarged by several times to achieve a sub-pixel precision image; the edge points of the enlarged laser line image are detected by using an adaptive edge detection algorithm, and the edge detection image is subjected to morphological dilation processing; the adaptive threshold is calculated by counting the histogram of the enlarged laser line image; the enlarged laser line image is segmented by using the adaptive threshold, and the segmented image is subjected to thinning processing; the dilated points are merged by region growing processing, and finally the center line of the laser image is extracted, and the center line coordinates are (x, y).

[0015] Further, the image acquisition: a plurality of acquisition devices are distributed on the track side to form an acquisition device array, and the acquisition device includes a linear array camera and a laser scanner.

[0016] Further, the laser coordinate transformation: the relationship between (x, y) and the world coordinate system coordinates (X w , Y w , Z w ) is as follows:

[0017]

[0018]

[0019]

[0020] wherein f is the focal length of the linear array camera, (X c , Y c , Z c ) is the coordinate in the linear array camera coordinate system, and H is the parameter obtained by calibration of the linear array camera; the corresponding relationship between the two-dimensional image point coordinate and the corresponding target three-dimensional coordinate can be known through the above formula.

[0021] Further, in the world coordinate system O-X w Y w Z w , the three-dimensional wheel information is corrected and processed, so as to facilitate the splicing and reconstruction of the wheel data: the point cloud data of the non-wheel part is segmented through the point cloud clustering algorithm DNSCN, and the point cloud data of only the wheel information is extracted, the arc surface wheel is corrected into a regular rectangular wheel through the coordinate mapping method, and the coordinate transformation and corrected laser line data are obtained.

[0022] Further, the three-dimensional wheel data is spliced: the camera coordinate system of a certain acquisition device is taken as a reference and defined as a world coordinate system, the external parameter rotation matrix R0 and the translation vector T0 of the other acquisition devices relative to the reference are obtained through measurement calibration, R0 and T0 are taken as initial values, the transformation parameters between the adjacent point cloud data are calculated through the NICP algorithm, more accurate rotation parameters R and translation parameters T are obtained, and finally R and T are subjected to affine transformation on the respective data, the overlapping area is fused, and the splicing of the wheel data is completed.

[0023] Further, the preprocessing: the original data is subjected to noise reduction processing through Gaussian filtering, the relatively smooth laser line image is obtained, the definition of the laser line image is improved through contrast stretching, and finally the interference data irrelevant to the laser line is deleted through the contour extraction judgment mode. The train wheel defect detection method based on three-dimensional information adopts the combination of multiple linear array cameras and 3D laser scanner modules, generates multiple two-dimensional image data and laser scanning data with depth information, extracts the wheel laser center line data through complex image processing algorithms, reconstructs the three-dimensional contour of the wheel with depth information, splices the three-dimensional information of the complete wheel, compares the local depth information, finds out the area with large depth change, finally detects the defects of the tread and the flange surface of the complete wheel, obtains the area, position and depth information of the defect area, and dynamically generates a detection report for the customer to review and repair. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 is the overall flow chart of the present application;

[0025] Fig. 2 is a schematic diagram of the collection device of the present application: (a) a schematic diagram of the collection angle of the collection device; (b) a schematic diagram of the arrangement of multiple collection devices;

[0026] Figure 3 is the original wheel laser line diagram of the present application;

[0027] Figure 4 is the laser line extraction diagram of the present application;

[0028] Figure 5 is the three-dimensional coordinate transformation diagram of the present application;

[0029] Figure 6 is the three-dimensional wheel splicing diagram of the present application. DETAILED DESCRIPTION

[0030] The present application is further described below in conjunction with the accompanying drawings and examples.

[0031] Referring to Figure 1 , it is the overall flow chart of the train wheel defect detection method based on three-dimensional information of the present embodiment.

[0032] Fig. 2(a) is a schematic diagram of the collection of wheel data by a collection device of the present application, and Fig. 2(b) is a schematic diagram of the arrangement of the collection device of the present application.

[0033] As can be seen from Figure 1 , the train wheel defect detection method based on three-dimensional information of the present embodiment has six main implementation steps, namely data collection, data preprocessing, laser line extraction, laser coordinate transformation, three-dimensional wheel splicing and defect detection, and the specific implementation of each step is as follows:

[0034] I. Data Collection

[0035] The collection device of the present embodiment is installed on both sides of the track at the throat section of the train in and out of the warehouse, and the number of data collection devices to be installed will be selected according to the size of the wheel. The size of the wheel collected by each collection device is certain, and the larger the wheel, the more collection devices are needed. The collection device includes a laser scanner, a linear array camera, a light source, a transmitter, a receiver and a temperature control, etc. In the present embodiment, ten collection devices are installed in the EMU detection shed, five of which are installed on one side of the track, and the device arrangement is shown in Fig. 2(b), which shows the device arrangement on the left side of the track. The arrangement on the right side is the same as that on the left side. In addition, the collection device is installed on the track side and maintains a safe distance from the track side, cannot be higher than the track surface, and the angle between the collection device and the track is about 5 degrees, and the distance between the collection devices is about 600 mm.

[0036] Due to the characteristics of the wheel image acquisition device fixed imaging, when the wheel passes, each wheel has thousands of scans, capturing the detailed state of each wheel, Fig. 2(a) shows the device acquisition wheel data schematic diagram, the collected wheel images are numbered according to the layout and shooting order of the acquisition device, and transmitted to the server through TCP for storage, which is convenient for subsequent data processing and analysis.

[0037] II. Data preprocessing:

[0038] As Figure 3 shown is a part of the original laser image of the wheel collected by a certain acquisition device, due to the interference of light source, noise and other factors in the obtained data, the subsequent data processing and measurement accuracy will be affected.

[0039] In this embodiment, Gaussian filter is used for noise reduction processing of the original data to obtain a relatively smooth laser line image, and then the contrast stretching is used to improve the clarity of the laser line image, and finally the contour extraction judgment method is used to delete the interference data irrelevant to the laser line. The preprocessed laser line image is denoted as Img.

[0040] III. Laser center line extraction

[0041] For the laser line image extracted in step two, the center of the laser line needs to be further extracted to improve the data processing accuracy.

[0042] The specific steps are as follows:

[0043] (1) adjust and enlarge Img by 2 times to obtain a sub-pixel accuracy image, denoted as ImgResize;

[0044] (2) use adaptive edge detection algorithm to detect the edge points of the enlarged laser line image ImgResize, and then perform morphological dilation processing on the edge detection image, denoted as ImgDilate;

[0045] (3) calculate the adaptive threshold value according to the histogram information of ImgResize;

[0046] (4) use the adaptive threshold value obtained in step (3) to segment the image ImgResize, and perform thinning processing on the segmented image, denoted as ImgThin;

[0047] (5) perform region growing on ImgThin to process and merge the points of ImgDilate, and finally extract the center line of the laser image, as shown in Figure 4 , denoted as the center line coordinates (x, y).

[0048] IV. Laser line coordinate transformation

[0049] The extracted image point set (x, y) of the wheel surface laser line is converted into real world coordinate system coordinates (Xw, Yw, Zw), i.e. the data coordinate conversion shown in Figure 4 is converted into the data shown in Figure 5 According to the principle of photogrammetry, the relationship between them is as follows:

[0050]

[0051]

[0052]

[0053] where f is the focal length of the camera, (X c , Y c , Z c ) is the coordinate in the camera coordinate system, and H is the parameter obtained by calibration. Through the above formula, the corresponding relationship between the two-dimensional image point coordinates and the corresponding target three-dimensional coordinates can be known.

[0054] In this embodiment, the installation positions of the laser scanner and the line array camera in the collection device are relatively fixed, the camera is calibrated in advance through the black and white checkerboard, and the transformation parameter H of the calibrated camera is obtained.

[0055] Since the laser line image coordinates and the calibration object positions are known, the laser scanner is calibrated again, the laser plane equation a·X w +b·Y W +c·Z w +d=0 is written as AX=B, where a 2 +b 2 +c 2 =1, A=[a, b, c], X=[X w Y w Z ww ] T , B=-d, and the parameter A of the calibrated laser scanner is obtained through the SVD (Singular Value Decomposition) decomposition method.

[0056] Once the laser scanner and the line array camera are calibrated, any laser point on the two-dimensional image of the wheel can be converted into the position of the wheel in the world coordinate system through the above coordinate conversion (the installation positions of the line array camera and the laser scanner are relatively fixed, the camera is calibrated first, then the laser scanner is calibrated, the position of the laser scanner relative to the camera is obtained, and the data of the laser scanner is reflected through the camera imaging, so that the two-dimensional laser image data can be converted into the three-dimensional world coordinate system).

[0057] Since the distance between the wheel and the laser scanner is changing during the data acquisition, i.e. the laser line data collected at a close distance is wider, and the laser line data collected at a far distance is narrower, the collected laser line is in the shape of an arc surface in the world coordinate system, so it is necessary to correct the three-dimensional wheel information in the world coordinate system O-X w Y w Z w , so as to facilitate the subsequent splicing and reconstruction of the wheel data. The point cloud data of the non-wheel part is segmented by the point cloud clustering algorithm DNSCN (Density-Based Spatial Clustering of Applications with Noise), and the point cloud data of only the wheel information is extracted. Finally, the arc wheel is corrected into a regular rectangular wheel by a coordinate mapping method, as shown in Figure 5 The laser line data after coordinate transformation correction is shown.

[0058] Five, three-dimensional data splicing

[0059] In this embodiment, multiple acquisition devices are used to acquire data of the same wheel. The acquisition devices are arranged in a horizontal manner, and the data collected by adjacent acquisition devices overlap. Therefore, it is necessary to splice the data of all acquisition devices to construct complete three-dimensional information of the wheel.

[0060] In step four, the data acquired by each laser scanner has been completed coordinate transformation, but there is still an affine transformation between the laser scanners. To complete the three-dimensional wheel data splicing, it is necessary to convert the laser data acquired by all acquisition devices to the same world coordinate system. In this embodiment, the camera coordinate system of the first acquisition device is defined as the world coordinate system, as shown in Fig. 2(b). Taking acquisition device 1 and acquisition device 2 as examples for illustration, the two groups of data to be spliced are denoted as data1 and data2. The specific steps are as follows:

[0061] (1) The rotation matrix R0 and the translation vector T0 of the acquisition device 2 relative to the acquisition device 1 can be obtained by measurement and calibration. R0 and T0 are used as initial values, and the registration transformation parameters between the adjacent point cloud data are calculated by the NICP (Normal Iterative Closest Point) algorithm to obtain more accurate rotation parameters R and translation parameters T;

[0062] (2) R and T are used to process the data data2 to complete the calculation of affine transformation, and the transformed data is denoted as data2c;

[0063] (3) The overlapping area of data1 and data2c can be calculated through step (2), the overlapping area is superimposed and fused, and the fused area is filtered through voxel filtering to remove redundant point cloud data, and in the embodiment, the voxel size is set to 2*2*2 mm, and the wheel data splicing is completed through the above steps.

[0064] Similarly, the data of other adjacent acquisition devices can be processed and spliced according to the above method, for example, the data of the acquisition device 2 is processed and spliced according to the above method. Figure 6 As shown in the figure, it is the three-dimensional data of the wheel after splicing of the five acquisition devices.

[0065] Six, wheel defect detection

[0066] Through step five, the three-dimensional information of the complete wheel can be obtained, and in the embodiment, the multi-level judgment strategy is adopted according to the local depth information of the wheel to complete the defect detection. The specific steps are as follows:

[0067] (1) A certain row of data is taken along the direction of the wheel tread, and is recorded as R i Another adjacent row of data is taken, and is recorded as R i+1 The depth information of R i and R i+1 is compared in the column direction, the point with a larger value is marked as 1, and the point with a smaller value is marked as 0, wherein i=1, 2, …, N, and N is the total number of rows of wheel data.

[0068] (2) The entire wheel is traversed in turn and processed according to step (1), and the position of the area with larger changes is found, that is, all areas marked as 1 are the suspected defect positions;

[0069] (3) The areas marked as 1 are clustered and processed, all points meeting the conditions are found, and the clustered area is recorded as ROI;

[0070] (4) The ROI area and the adjacent area are compared again in terms of depth average, the area meeting the condition is extracted and determined as a defect, and the area of the defect is counted to calculate the area, position and depth information of all defect areas of the wheel.

[0071] The present application has been disclosed with the preferred embodiments as above, but they are not intended to limit the present application, and any skilled person in the art can make various changes or modifications without departing from the spirit and scope of the present application, therefore, the protection scope of the present application should be determined by the protection scope of the claims of the present application.

Claims

1. A method for detecting defects of a train wheel based on three-dimensional information, characterized by, The method comprises: Image acquisition and preprocessing by a linear array camera and a laser scanner to obtain a laser line image; Laser center line extraction: the center line of the laser image is extracted, and the two-dimensional coordinates are (x, y); the laser center line extraction: the preprocessed laser line image is adjusted and enlarged by several times to achieve a sub-pixel precision image; an adaptive edge detection algorithm is used to detect the edge points of the enlarged laser line image, and the edge detection image is then subjected to morphological dilation processing; an adaptive threshold is calculated by counting the histogram of the enlarged laser line image; the enlarged laser line image is segmented using the adaptive threshold, and the segmented image is subjected to thinning processing; the dilated points are merged through region growing processing, and the center line of the laser image is finally extracted, and the center line coordinates are (x, y); Laser coordinate transformation: converting the extracted image point set generated by the wheel surface laser line, i.e. (x, y), into real world coordinate system coordinates ; in the world coordinate system , correcting three-dimensional wheel information: through the point cloud clustering algorithm DBSCAN, the non-wheel part point cloud data is segmented, and the point cloud data with only wheel information is extracted, and through the coordinate mapping method, the cambered wheel is corrected into a regular rectangular wheel, and the corrected laser line data after coordinate transformation is obtained; Three-dimensional wheel data splicing: the coordinate system of a certain acquisition device is taken as a reference and defined as a world coordinate system, the coordinate systems of other acquisition devices are subjected to affine transformation, the overlapping areas are subjected to fusion processing, and the three-dimensional information of a complete wheel is obtained: the camera coordinate system of a certain acquisition device is taken as a reference and defined as a world coordinate system, the external parameter rotation matrix R0 and the translation vector T0 of other acquisition devices relative to the reference are obtained through measurement calibration, R0 and T0 are taken as initial values, the transformation parameters between adjacent point cloud data are calculated through the NICP algorithm, more accurate rotation parameters R and translation parameters T are obtained, and finally R and T are subjected to affine transformation on the respective data, the overlapping areas are subjected to fusion processing, and the splicing of the wheel data is completed; Defect detection: according to the local depth information of the wheel, a multi-level judgment strategy is used to complete defect detection; the defect detection: first, a row of data is taken along the wheel tread direction, and the depth information in the column direction is compared with the adjacent row, the points with a larger value are marked as 1, and the points with a smaller value are marked as 0; then, the entire wheel is traversed, and the area position with a larger change is found, that is, all areas marked as 1 are suspicious defect positions; finally, the areas marked as 1 are subjected to local clustering processing, and the areas after clustering are compared one by one, and all areas meeting the conditions are found as defects, and the area, position and maximum depth information of the defect area are calculated.

2. The method for detecting defects of a train wheel based on three-dimensional information according to claim 1, wherein, The image acquisition: a plurality of acquisition devices are distributed on the track side to form an acquisition device array.

3. The method of claim 1, wherein the method comprises: The laser coordinate transformation: the and the relationship between the coordinates in the world coordinate system is as follows: ， Wherein, f is the focal length of the linear array camera, is the coordinate in the linear array camera coordinate system, and H is the parameter obtained by the linear array camera through calibration; the corresponding relationship between the two-dimensional image point coordinate and the corresponding target three-dimensional coordinate can be known through the above formula.

4. The method of claim 1, wherein the method is characterized by: The preprocessing: Gaussian filtering is used to perform noise reduction processing on the original data to obtain a relatively smooth laser line image, the contrast stretching is used to improve the definition of the laser line image, and finally the contour extraction judgment method is used to delete the interference data irrelevant to the laser line.

Images