An intelligent detection method for the dimensions of high-speed train running gear

Abstract

The invention discloses an intelligent detection method for the size of a high-speed train running gear, comprising the following steps: building an image acquisition platform; obtaining a corresponding relationship between image pixels and actual sizes; using light source hardware information to call a camera to acquire and store images; performing image grayscale processing to convert a color image into a grayscale image, eliminating image hue and saturation information while retaining brightness to achieve a grayscale processing function by converting an RGB image or a color image into a grayscale image; performing smoothing filtering processing on the grayscale image by using an adaptive median filter to reduce image noise; performing sharpening filtering processing on the image; performing edge detection on the image by using an improved Canny operator; using Hough transform to fit the discontinuity boundary shape and calculate the size parameters of the straight line, circle and arc shape; having better denoising and anti-noise capabilities, better retaining edge details and achieving accurate edge positioning.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent inspection, and in particular relates to an intelligent detection method for the dimensions of a running gear of a high-speed railway EMU. Background Art

[0002] With the rapid development of my country's high-speed rail industry, the deformation of parts size will bring a series of safety hazards to high-speed rail, and accurate detection of parts size has become particularly important. As an important part of the train, the running gear of the high-speed rail train needs to be frequently inspected, but the traditional manual inspection is inefficient and has poor accuracy. At this stage, digital images based on machine vision have become the mainstream method of dimensional measurement due to their high processing accuracy and convenient and fast processing process. At present, the accuracy of dimensional measurement is affected by hardware such as camera lens resolution and image acquisition card acquisition speed. This application proposes an automatic measurement system for the cross-sectional dimensions of running gear components. Based on the adaptive filtering denoising algorithm and the improved Canny operator sharpening algorithm digital image, the running gear component cross-sectional dimension measurement software is designed to connect the camera to realize the automatic and accurate measurement of the cross-sectional dimensions of the running gear components. It can effectively identify the dimensional deformation of the running gear components and provide a strong guarantee for the safe operation of the train. Summary of the invention

[0003] In view of the above defects or improvement needs of the prior art, the present invention aims to solve the problem that the prior art high-speed train running gear detection adopts manual detection, which has the problems of low efficiency and poor accuracy.

[0004] In order to achieve the above object, the present invention relates to a method for intelligently detecting the dimensions of a high-speed train running gear, comprising the following steps:

[0005] Step 1: Build an image acquisition platform;

[0006] Step 2: Get the correspondence between image pixels and real size;

[0007] Step 3: Use the light source hardware information to call the camera to collect and store images;

[0008] Step 4: grayscale the image, convert the color image into a grayscale image, eliminate the image hue and saturation information while retaining the brightness to achieve the function of converting the RGB image or color image into a grayscale image to achieve grayscale processing;

[0009] Step 5: Use adaptive median filtering to smooth the grayscale image and reduce the noise of the image;

[0010] Step 6: Perform sharpening filtering on the image obtained in step 5;

[0011] Step 7: Use the improved Canny operator to perform edge detection on the image in step 6;

[0012] Step 8: Use Hough transform to fit the boundary shape of the discontinuity points, and combine steps 5, 6, and 7 to calculate the size parameters of the straight line, circle, and arc shapes;

[0013] Step 9: Get the size measurements.

[0014] Furthermore, the hardware equipment used to build the image acquisition platform includes: a light source, a camera, a lens, an image acquisition card and a computer; here, due to actual needs, three forms of LED surface light source, strip light source and ring light source are used to meet the lighting requirements under different working conditions, and the light source is projected in the form of direct light.

[0015] Furthermore, the specific method of calibrating the camera in step 2 is as follows: in camera calibration, three coordinate systems are included, namely the world coordinate system, the camera coordinate system and the image coordinate system. The world coordinate system describes the position of the target object in the real world, the camera coordinate system describes the position of the object from the perspective of the camera, and the image coordinate system describes the projection transmission relationship of the object from the camera coordinate system to the image coordinate system during the imaging process, and obtains the coordinates in the pixel coordinate system; in the world coordinate system o w - x w y w z w The coordinates in are P ( x w , y w , z w ), which is the same as in the camera coordinate system o c - x c y c z c The corresponding coordinates in P ( x c , y c , z c ) is:

[0016]

[0017] In the formula, R is a 3×3 orthogonal matrix, and R The determinant value is equal to 1, T is the translation vector; let , ,passR and T The direction and position of the industrial camera relative to the object space coordinate system can be determined; the camera coordinate system o c - x c y c z c In P Point coordinates P ( x c , y c , z c ) and the image coordinate system O 1- X 1 Y 1 Corresponding point coordinates ( X , Y ) is:

[0018]

[0019] in, , , a is the scale factor, f is the camera focal length, P is the perspective projection matrix; the combined equations (1) and (2) can be obtained in the world coordinate system P Point coordinates( x w , y w , z w ) and its projection point ( u , v ) relationship.

[0020] Furthermore, the specific method of step 5 is: the adaptive median filtering algorithm also selects a m × n Rectangular window s xy Use the filter window as a noise detection window on the image; however, when the adaptive median filter performs the filtering operation, if the filter window affects the filtering operation, the algorithm will automatically adjust the filter window size according to the pre-set conditions, otherwise, do not adjust; when outputting the filtering result, first determine whether the median value of the pixel in the current traversal window is a noise point, if the median value in the window is a noise point, then increase the window, determine if the median value is not a noise point, and then determine whether the current pixel is a noise point. If it is, use the median value instead of the output, if not, output directly;

[0021] The following symbols are defined:

[0022] Z min —Current filter window S xy The minimum gray value in ; Z max —Current filter window S xy The maximum gray value in ; Z med —Current filter window S xy The median of the grayscale values ​​in ; Z xy -coordinate( x , y ) at the gray value; S max —Current Window S xy Maximum size allowed;

[0023] The adaptive median filtering algorithm is divided into two processes, represented as process A and process B:

[0024] Process A:

[0025] A1=Z med -Z min (4)

[0026] A2=Z med -Z max (5)

[0027] Among them, A 1、 A2 For Process A Two criteria for judging If A1>0 and A2<0, go to process B, otherwise increase the window size; if the window size ≦ Smax, repeat process A; otherwise output Z med ;

[0028] Process B:

[0029] B1=Z xy -Z min ( 6)

[0030] B2=Z xy -Z max ( 7)

[0031] Among them, B 1、 B2 For Process B Two criteria for judging If B1>0 and B2<0, then output Z xy ; Otherwise output Z med .

[0032] Furthermore, the specific method of step 7 is:

[0033] (1) Use selective blurring with edge preservation to replace Gaussian blurring: The smoothing method with edge preservation function does not allow all pixels within the domain to participate in the calculation of the smoothing method, but sets a threshold to allow pixels whose grayscale difference with the central pixel is less than this threshold to participate in the calculation; in this way, pixels that are too different from the central pixel are considered to carry valid information rather than noise, and will not participate in the smoothing calculation, thereby retaining these useful high-frequency signals, and the edge signal is naturally also within the scope of preservation;

[0034] (2) The improved Canny operator uses a 3×3 Sobel convolution template instead of a 2×2 domain to calculate the gradient; from the analysis of the Sobel operator, it can be seen that the Sobel operator uses a 3×3 convolution template to calculate the gradient. The 3x3 convolution model is the smallest size that can capture the eight-neighborhood information of a pixel. The limited receptive field of two 3x3 stacked convolution layers is 5x5; the receptive field of three 3x3 stacked convolution layers is 7x7, so the large-size convolution layer can be replaced by stacking a small-size convolution layer, and the receptive field size remains unchanged; multiple 3x3 convolution layers have more nonlinearity than a large-size convolution layer, making the decision function more decisive; multiple 3x3 convolution layers have fewer parameters than a large-size convolution layer. Assuming that the feature map size of the input and output of the convolution layer is the same as C, then the number of parameters of the three 3x3 convolution layers is 27C. 2 ; The number of parameters of a 7x7 convolutional layer is 49C 2 ; Three 3x3 convolutional layers are regarded as the decomposition of a 7x7 convolutional layer. The middle layer has a nonlinear decomposition and plays an implicit regularization role.

[0035] Furthermore, the method of fitting the discontinuity boundary shape by using Hough transform in step 8 is: x i , y i ) and dot ( x j , y j ) in the parameter space a - b There is a straight line on each of them, and these straight lines intersect at point ( a 0, b 0), and a 0. b 0 is the image coordinate space x - y midpoint( x i ,y i ) and dot ( x j , y j ); conversely, all lines corresponding to the same point in parameter space have collinear points corresponding to them in image coordinate space; based on this property, given some edge points in image coordinate space, the equation of the line connecting these points can be determined by Hough transform.

[0036] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0037] 1. The intelligent detection method for the running gear size of a high-speed train EMU of the present invention can automatically complete image acquisition, camera calibration, image preprocessing, image segmentation and running gear component cross-sectional size detection, etc., meeting the requirements of online detection and real-time performance;

[0038] 2. The intelligent detection method for the running gear size of a high-speed train of the present invention proposes an improved Canny operator edge extraction algorithm; it has better denoising and anti-noise capabilities, better retains edge details and achieves precise edge positioning; it uses Hough transform to combine edge pixels into complete edges to obtain clear and continuous edges in the image. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a schematic diagram of the Zhang Zhengyou calibration method of a preferred embodiment of the present invention;

[0040] Figure 2 A flow chart of edge detection steps in a preferred embodiment of the present invention;

[0041] Figure 3 A Hough transform diagram in the rectangular coordinate system of a preferred embodiment of the present invention;

[0042] Figure 4 A comparison diagram of the improved canny edge extraction effect diagram of the present invention according to a preferred embodiment of the present invention (the right side is the extraction effect diagram using the improved Canny operator edge extraction algorithm, and the left side is before the improvement);

[0043] Figure 5 This is a diagram showing the automatic and accurate measurement results of the cross-sectional dimensions of the running gear components of the preferred embodiment of the present invention; DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0045] Please refer to Figure 1-Figure 3 The present invention relates to a method for intelligently detecting the dimensions of a high-speed train running gear, comprising the following steps:

[0046] Step 1: Build an image acquisition platform;

[0047] Step 2: Get the correspondence between image pixels and real size;

[0048] Step 3: Use the hardware information such as light source to call the camera to collect and store images;

[0049] Step 4: Grayscale the image, convert the color image into a grayscale image, eliminate the image hue and saturation information while retaining the brightness to realize the function of converting the RGB image or color image into a grayscale image to realize grayscale processing; (The color value of each pixel on the grayscale image is also called grayscale, which refers to the color depth of the point in the black and white image, generally ranging from 0 to 255, white is 255, and black is 0. The so-called grayscale value refers to the degree of color intensity, and the grayscale histogram refers to the number of pixels with each grayscale value in a digital image. Image grayscale saves storage space and improves computing efficiency. Grayscale processing can be used as a preprocessing step for image processing when the color requirements are not high)

[0050] Step 5: Use adaptive median filtering to smooth the grayscale image and reduce the noise of the image;

[0051] Step 6: Perform sharpening filtering on the image obtained in step 5 (the sharpening filtering method highlights the edge of the image, and the sharpening filtering is to determine the change of grayscale in a certain point area to better detect the boundary);

[0052] Step 7: Use the improved Canny operator to perform edge detection on the image in step 6; (Use the improved Canny operator method to better remove noise and resist noise, better retain edge details and achieve accurate edge positioning);

[0053] Step 8: Use Hough transform to fit the boundary shape of the discontinuity point, and combine steps 5, 6, and 7 to calculate the size parameters of the shape such as the straight line, circle, and arc; the method of fitting the boundary shape of the discontinuity point using Hough transform is: x i ,y i ) and dot ( x j , y j ) in the parameter space a - b Each of them corresponds to a straight line (see Figure 3 ), these lines intersect at point ( a 0, b 0), and a 0. b 0 is the image coordinate space x - y midpoint( x i , y i ) and dot ( x j , y j ) is the parameter of the line determined by the Hough transform. Conversely, all lines in the parameter space that are compared to the same point have collinear points corresponding to them in the image coordinate space. Based on this property, given some edge points in the image coordinate space, the equation of the line connecting these points can be determined by Hough transform;

[0054] Step 9: The interface results are output and the size measurement results are obtained.

[0055] The hardware equipment used to build the image acquisition platform includes: light source, camera, lens, image acquisition card and computer; the light source adopts LED surface light source, strip light source and ring light source, and the projection mode of the light source adopts direct light (specifically including one MV-EGIGE image acquisition card; MV-EM120M monocular camera and its bracket; one FH-RI9030W ring light source, one FH-FL230230W surface light source, two FH-BL0010W strip light sources, and one FH-AP4CH-24 controller).

[0056] The specific method of calibrating the camera in step 2 is as follows: In camera calibration, three coordinate systems are included, namely the world coordinate system, the camera coordinate system and the image coordinate system. The world coordinate system describes the position of the target object in the real world, the camera coordinate system describes the position of the object from the perspective of the camera, and the image coordinate system describes the projection transmission relationship of the object from the camera coordinate system to the image coordinate system during the imaging process, and the coordinates in the pixel coordinate system are obtained; in the world coordinate system o w - x w y w z w The coordinates in areP ( x w , y w , z w ), which is the same as in the camera coordinate system o c - x c y c z c The corresponding coordinates in P ( x c , y c , z c ) is:

[0057]

[0058] In the formula, R is a 3×3 orthogonal matrix, and R The determinant value is equal to 1, T is the translation vector; let , ,pass R and T The direction and position of the industrial camera relative to the object space coordinate system can be determined; the camera coordinate system o c - x c y c z c Coordinates in P Point coordinates P ( x c , y c , z c ) and the image coordinate system O 1- X 1 Y 1 Corresponding point coordinates ( X , Y ) is:

[0059]

[0060] in, , , a is the scale factor, f is the camera focal length, Pis the perspective projection matrix; the combined equations (1) and (2) can be obtained in the world coordinate system P Point coordinates( x w , y w , z w ) and its projection point ( u , v ) relationship.

[0061] Please refer to Figure 4-Figure 5 The specific method of step 5 is: the adaptive median filtering algorithm also selects a m × n Rectangular window s xy Use it as a filter window to detect noise on the image. However, when the adaptive median filter performs filtering operations, if the filter window affects the filtering operation, the algorithm will automatically adjust the filter window size according to the pre-set conditions, otherwise, do not adjust it. When outputting the filtering results, first determine whether the median of the pixel in the current traversal window (i.e., the filter window) is a noise point. If the median in the window is a noise point, increase the window, determine if the median is not a noise point, and then determine whether the current pixel is a noise point. If so, use the median instead of the output. If not, output directly. The adaptive median filtering algorithm is divided into two processes, represented as process A and process B:

[0062] Process A:

[0063] A1=Z med -Z min (4)

[0064] A2=Z med -Z max (5)

[0065] If A1>0 and A2<0, go to process B; otherwise increase the window size; if the window size ≦ Smax, repeat process A; otherwise output Z med .

[0066] Process B:

[0067] B1=Z xy -Z min (6)

[0068] B2=Z xy -Z max (7)

[0069] If B1>0 and B2<0, then output Z xy ; Otherwise output Z med .

[0070] The specific method of step 7 is:

[0071] (1) Use selective blurring with edge preservation to replace Gaussian blurring: The basic idea of ​​the smoothing method with edge preservation function is not to allow all pixels within the domain to participate in the calculation of the smoothing method, but to set a threshold and only allow pixels whose grayscale difference with the center pixel is less than this threshold to participate in the calculation. In this way, pixels that are too different from the center pixel are considered to have valid information, not noise, and will not participate in the smoothing calculation, thereby preserving these useful high-frequency signals, and the edge signal is naturally also within the scope of preservation;

[0072] (2) The improved Canny operator uses a 3×3 Sobel convolution template instead of a 2×2 domain to calculate the gradient; from the analysis of the Sobel operator, it can be seen that the Sobel operator uses a 3×3 convolution template to calculate the gradient. The 3x3 convolution model is the smallest size that can capture the eight-neighborhood information of a pixel. The limited receptive field of two 3x3 stacked convolution layers is 5x5; the receptive field of three 3x3 stacked convolution layers is 7x7, so the large-size convolution layer can be replaced by stacking small-size convolution layers, and the receptive field size remains unchanged; multiple 3x3 convolution layers have more nonlinearities (more layers of nonlinear functions) than a large-size convolution layer, making the decision function more decisive. Multiple 3x3 convolution layers have fewer parameters than a large-size filter. Assuming that the feature map size of the input and output of the convolution layer is the same as C, then the number of parameters of the three 3x3 convolution layers is 3x(3x3xCxC)=27C. 2 ; A 7x7 convolutional layer has 49C parameters 2 ; So the three 3x3 convolution layers can be regarded as the decomposition of a 7x7 convolution layer (the middle layer has nonlinear decomposition and plays an implicit regularization role);

[0073] The 3×3 Sobel convolution template calculates the gradient relative to the 2×2 field, which can more effectively highlight the edge and suppress noise; it can be seen that compared with Gaussian blur, at the same radius, it can be seen that the selective blur with edge preservation function can obviously retain more edge details, so that relatively weak edges can be preserved (see Figure 4 ).

[0074] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent detection method for the dimensions of high-speed train running gear. It is characterized in that The steps include: Step 1: Build an image acquisition platform; Step 2: Get the correspondence between image pixels and real size; Step 3: Use the light source hardware information to call the camera to collect and store images; Step 4: grayscale the image, convert the color image into a grayscale image, eliminate the image hue and saturation information while retaining the brightness to achieve the function of converting the RGB image or color image into a grayscale image to achieve grayscale processing; Step 5: Use adaptive median filtering to smooth the grayscale image and reduce the noise of the image; Step 6: Perform sharpening filtering on the image obtained in step 5; Step 7: Use the improved Canny operator to perform edge detection on the image in step 6; the specific method of step 7 is: (1) Use selective blurring with edge preservation to replace Gaussian blurring: The smoothing method with edge preservation function does not allow all pixels within the domain to participate in the calculation of the smoothing method, but sets a threshold to allow pixels whose grayscale difference with the central pixel is less than this threshold to participate in the calculation; in this way, pixels that are too different from the central pixel are considered to carry valid information rather than noise, and will not participate in the smoothing calculation, thereby retaining these useful high-frequency signals, and the edge signal is naturally also within the scope of preservation; (2) The improved Canny operator uses a 3×3 Sobel convolution template instead of a 2×2 domain to calculate the gradient; from the analysis of the Sobel operator, it can be seen that the Sobel operator uses a 3×3 convolution template to calculate the gradient. The 3x3 convolution model is the smallest size that can capture the eight-neighborhood information of a pixel. The limited receptive field of two 3x3 stacked convolution layers is 5x5; the receptive field of three 3x3 stacked convolution layers is 7x7, so the large-size convolution layer can be replaced by stacking a small-size convolution layer, and the receptive field size remains unchanged; multiple 3x3 convolution layers have more nonlinearity than a large-size convolution layer, making the decision function more decisive; multiple 3x3 convolution layers have fewer parameters than a large-size convolution layer. Assuming that the feature map size of the input and output of the convolution layer is the same as C, then the number of parameters of the three 3x3 convolution layers is 27C. 2 ; The number of parameters of a 7x7 convolutional layer is 49C 2 ; Three 3x3 convolution layers are considered as a decomposition of a 7x7 convolution layer. The middle layer has a nonlinear decomposition and plays an implicit regularization role; Step 8: Use Hough transform to fit the boundary shape of the discontinuity points, and combine steps 5, 6, and 7 to calculate the size parameters of the straight line, circle, and arc shapes; Step 9: Get the size measurements.

2. According to claim 1, the intelligent detection method for the running gear size of a high-speed train set, It is characterized in that The hardware equipment used to build the image acquisition platform includes: a light source, a camera, a lens, an image acquisition card and a computer; here, due to actual needs, three forms of LED surface light source, strip light source and ring light source are used to meet the lighting requirements under different working conditions, and the projection mode of the light source adopts direct light.

3. According to claim 1, the intelligent detection method for the running gear size of a high-speed train set, Features: The specific method of calibrating the camera in step 2 is as follows: In camera calibration, three coordinate systems are included, namely the world coordinate system, the camera coordinate system and the image coordinate system. The world coordinate system describes the position of the target object in the real world, the camera coordinate system describes the position of the object from the perspective of the camera, and the image coordinate system describes the projection transmission relationship of the object from the camera coordinate system to the image coordinate system during the imaging process, and the coordinates in the pixel coordinate system are obtained; in the world coordinate system o w - x w y w z w The coordinates in are P ( x w , y w , z w ), which is the same as in the camera coordinate system o c - x c y c z c The corresponding coordinates in P ( x c , y c , z c ) is: In the formula, R is a 3×3 orthogonal matrix, and R The determinant value is equal to 1, T is the translation vector; let , ,pass R and T The direction and position of the industrial camera relative to the object space coordinate system can be determined; the camera coordinate system o c - x c y c z c In P Point coordinates P ( x c , y c , z c ) and the image coordinate system O 1 - X 1 Y 1 The coordinates of the corresponding points in X , Y ) is: in, , , a is the scale factor, f is the camera focal length, P is the perspective projection matrix; Combining equations (1) and (2) gives P Point coordinates( x w , y w , z w ) and its projection point ( u , v ) relationship.

4. According to claim 1, the intelligent detection method for the running gear size of a high-speed train set, It is characterized in that The specific method of step 5 is: the adaptive median filtering algorithm also selects a m × n Rectangular window s xy Use the filter window as a noise detection window on the image; however, when the adaptive median filter performs the filtering operation, if the filter window affects the filtering operation, the algorithm will automatically adjust the filter window size according to the pre-set conditions, otherwise, do not adjust; when outputting the filtering result, first determine whether the median value of the pixel in the current traversal window is a noise point, if the median value in the window is a noise point, then increase the window, determine if the median value is not a noise point, and then determine whether the current pixel is a noise point. If it is, use the median value instead of the output, if not, output directly; The following symbols are defined: Z min —Current filter window S xy The minimum gray value in ; Z max —Current filter window S xy The maximum gray value in ; Z med —Current filter window S xy The median of the grayscale values ​​in ; Z xy -coordinate( x , y ) at the gray value; S max —Current Window S xy Maximum size allowed; The adaptive median filtering algorithm is divided into two processes, represented as process A and process B: Process A: And 1 =Z med -Z min (4) And 2 =Z med -Z max (5) Among them, A 1、 A 2 For Process A Two criteria for judging If A 1 >0 and A 2 <0, go to process B, otherwise increase the window size; if the window size ≦ Smax, repeat process A; otherwise output Z med ; Process B: B 1 =Z xy -Z min (6) B 2 =Z xy -Z max (7) Among them, B 1、 B 2 For Process B Two criteria for judging If B 1 >0 and B 2 <0, then output Z xy ; Otherwise output Z med .

5. According to claim 1, the intelligent detection method for the running gear size of a high-speed train set, It is characterized in that The method of fitting the discontinuity boundary shape by using Hough transform in step 8 is: x i , y i ) and dot ( x j , y j ) in the parameter space a - b There is a straight line on each of them, and these straight lines intersect at point ( a 0 , b 0 ),and a 0 , b 0 It is the image coordinate space x - y midpoint( x i , y i ) and dot ( x j , y j ); conversely, all lines corresponding to the same point in parameter space have collinear points corresponding to them in image coordinate space; based on this property, given some edge points in image coordinate space, the equation of the line connecting these points can be determined by Hough transform.

Images