Obstacle detection method of magnetic suspension detection vehicle

Abstract

The invention discloses an obstacle detection method for a magnetic suspension detection vehicle, and the method comprises the steps: extracting a track template image corresponding to a to-be-detected track from an initial position to an end position, and taking an initial comparison image corresponding to the initial position of the track template image as a reference, sequentially performing feature extraction on the acquired track real-time image set and the initial comparison image, and screening out a first base layer image serving as a detection initial end based on feature geometric distortion of the track real-time image set at an initial position relative to the initial comparison image; acquiring feature geometric distortion of a track real-time image set shot by the shooting system every time relative to a previous base layer image, performing geometric correction on selected image splicing objects, and sequentially splicing the image splicing objects to form a track detection image; performing calendar comparison on the track detection image and the track template image to identify an obstacle; according to the invention, the accuracy of one-to-one comparison of the pixels in the track detection image and the track template image is improved, so that the robustness of obstacle recognition is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to an obstacle detection method of a magnetic levitation detection vehicle. Background Art

[0002] Current track inspection technologies, including chord measurement, total station inspection, and inertial reference methods, are mainly used to measure the smoothness and damage of rails. They can only be used to detect the geometric smoothness of the track. In essence, they are geometric state detectors. Maglev vehicles use permanent magnet tracks, which are prone to adsorbing some metal materials. To improve the safety of permanent magnet tracks, it is necessary to regularly inspect the permanent magnet guide rails for metal obstacles and promptly remove any metal obstacles on the permanent magnet guide rails.

[0003] In the existing technology, an inspection vehicle is used to collect images and image recognition technology is used to identify metal obstacles on the permanent magnetic guide rail. However, since the permanent magnetic guide rail is relatively long, it is necessary to splice the track images collected each time and compare them with the track template image without obstacles to identify metal obstacles on the permanent magnetic guide rail.

[0004] However, since the structural shapes of permanent magnet guide rails are mostly the same, if a small number of features are used to match the two images to be stitched, it is very likely that pixel points with matching pixel values ​​but mismatched positions will be found in the real-time track image set, resulting in inaccurate calculation of image geometric distortion and errors in geometric distortion calculation. Summary of the Invention

[0005] The object of the present invention is to provide an obstacle detection method for a magnetic levitation inspection vehicle to solve the technical problem in the prior art that the method for calculating image geometric distortion is inaccurate and errors in geometric distortion calculation occur.

[0006] In order to solve the above technical problems, the present invention specifically provides the following technical solutions:

[0007] An obstacle detection method for a magnetic levitation inspection vehicle comprises the following steps:

[0008] Step 100: Extract the track template image corresponding to the starting position to the ending position of the track to be inspected, and use the camera system on the inspection vehicle to capture images of the track to be inspected from the starting position to the ending position, and obtain a real-time track image set corresponding to each capture;

[0009] Step 200: Using the starting reference image corresponding to the starting position of the track template image as a reference, feature extraction is performed on the track real-time image set at the starting position and the starting reference image in sequence. Based on the characteristic geometric distortion of the track real-time image set at the starting position relative to the starting reference image, a first base layer image as the starting point of detection is selected from multiple images captured in a single shot of the track real-time image set.

[0010] Step 300: Obtain characteristic geometric distortion of each track real-time image set captured by the camera system relative to the previous base layer image, select the image with the smallest geometric distortion from the track real-time image set as the image stitching object, perform geometric correction on the image stitching object, and stitch the overlapping areas in sequence to form a track detection image from the starting position to the ending position;

[0011] Step 400: Perform a calendar comparison between the track detection image and the track template image to determine obstacles in the track detection image based on the comparison result.

[0012] As a preferred solution of the present invention, in step 100, a two-dimensional coordinate system is constructed on the track template image, wherein the starting position of the track to be detected is used as the origin, the length direction of the track to be detected is used as the X-axis, and the width direction of the track to be detected is used as the Y-axis;

[0013] Determine a pixel value array of a horizontal edge of the track template image along the X-axis direction, and a pixel value array of a vertical edge of the track template image along the Y-axis direction.

[0014] As a preferred solution of the present invention, the camera system uses a single-shot multi-exposure camera mode to capture multiple images in a single shot. In step 200, feature extraction of the initial comparison image is performed as follows:

[0015] A starting comparison image is intercepted from the track template image within the shooting range of the camera system, the starting comparison image is binarized, and a pixel value distribution matrix of the starting comparison image is constructed.

[0016] Based on the pixel value distribution matrix of the starting comparison image, a first regional feature for feature comparison is selected, and a second regional feature of a row corresponding to the first regional feature in the pixel value distribution matrix is ​​obtained, and a third regional feature of a column corresponding to the first regional feature in the starting comparison image is obtained.

[0017] Obtain pixel value distribution queues corresponding to the first region feature, the second region feature, and the third region feature respectively.

[0018] As a preferred solution of the present invention, the method for implementing feature extraction for each image of the track real-time image set at the starting position is:

[0019] Based on each vertical pixel value distribution queue of the first regional feature and the second regional feature, screening out a first feature matching region identical to each vertical pixel value distribution queue of the first regional feature and the second regional feature from each image of the track real-time image set;

[0020] Based on each horizontal pixel value distribution queue of the first regional feature and the third regional feature, screening out a second feature matching region identical to each horizontal pixel value distribution queue of the first regional feature and the third regional feature from each image of the track real-time image set;

[0021] Generate a row-wise regional feature pixel value distribution matrix corresponding to the first feature matching region and a column-wise regional feature pixel value distribution matrix corresponding to the second feature matching region in each image of the track real-time image set.

[0022] As a preferred solution of the present invention, a method for determining the characteristic geometric distortion of each image in the track real-time image set relative to the initial comparison image is as follows:

[0023] Constructing a row-wise regional feature pixel value distribution matrix using the first regional feature and the second regional feature, determining the minimum coordinates (a1, b1) and the maximum coordinates (a2, b2) of the pixel points in the row-wise regional feature pixel value distribution matrix, and fitting and calculating the lateral deformation θ of each image in the track real-time image set relative to the starting comparison image;

[0024] Where, tanθ=|b1-b2| / |a1-a2|;

[0025] Constructing a column-wise regional feature pixel value distribution matrix using the first regional feature and the third regional feature, determining the minimum coordinates (a1, b1) and the maximum coordinates (a3, b3) of the pixel points in the column-wise regional feature pixel value distribution matrix, and fitting and calculating the column-wise deformation α of each image in the track real-time image set relative to the starting reference image;

[0026] Wherein, tanα=|b1-b3| / |a1-a3|.

[0027] As a preferred solution of the present invention, the characteristic geometric distortion ɡ of each image of the track real-time image set shot in a single shot relative to the initial reference image is calculated based on the lateral deformation θ and the column deformation α;

[0028] ɡ=θ*W1+α*W2;

[0029] Among them, W1 is the assignment weight of the lateral deformation θ, and W2 is the assignment weight of the column deformation α. ​​W1 has a positive exponential relationship with the lateral deformation θ, and W2 has a positive exponential relationship with the column deformation α.

[0030] As a preferred solution of the present invention, the characteristic geometric distortions ɡ corresponding to all images of a single-shot track real-time image set are compared, and the image with the smallest characteristic geometric distortion ɡ is selected as the first base image. The first base image is geometrically corrected so that the pixel distribution of the first base image is the same as the pixel distribution of the starting comparison image. The specific geometric correction method is:

[0031] According to the tilt angle of the lateral deformation relative to the initial comparison image, distinguish the positive and negative values ​​of the lateral deformation θ, and perform reverse correction on a single row of pixels of the first base layer image;

[0032] According to the tilt angle of the column-wise deformation relative to the initial comparison image, the positive and negative values ​​of the column-wise deformation α are distinguished, and the single column of pixels of the first base layer image are reversely corrected.

[0033] As a preferred solution of the present invention, in step 300, the track detection image from the starting position to the ending position is formed as follows:

[0034] Obtaining a pixel value distribution queue corresponding to an overlapping area between the first base layer image and a set of track real-time images captured in a single shot;

[0035] respectively obtaining pixel value distribution queues corresponding to overlapping areas of a set of track real-time images shot in a single shot;

[0036] Extracting features from the overlapping area, sequentially determining the lateral deformation and vertical deformation of each image in the track real-time image set captured in a single shot, and selecting a single image with the smallest characteristic geometric distortion for image stitching based on the characteristic geometric distortion composed of the lateral deformation and the vertical deformation;

[0037] Performing geometric correction on the single image and then splicing it with the first base image;

[0038] Repeat the above steps to select a single image with the smallest characteristic geometric distortion compared with the previous stitched image from the multiple images of the orbital real-time image set taken in a single shot as the stitched image;

[0039] The stitched image is geometrically corrected and then stitched with the previous base image to form a track detection image from the starting position to the ending position.

[0040] As a preferred solution of the present invention, in step 400, the calendar comparison of the track detection image and the track template image is implemented as follows:

[0041] Taking the pixel value array data of the horizontal edge of the track template image along the X-axis direction as a reference value, finding the pixel point distribution horizontal axis corresponding to the same pixel value array data in the track detection image, segmenting the track detection image according to the pixel point distribution horizontal axis, and retaining the two images within the pixel point distribution horizontal axis;

[0042] Using the pixel value array data of the vertical edge of the track template image along the Y-axis as a reference value, finding the pixel point distribution vertical axis corresponding to the same pixel value array data in the track detection image, segmenting the track detection image along the pixel point distribution vertical axis, and retaining the two images within the pixel point distribution vertical axis;

[0043] The pixel values ​​of all pixels of the track detection image and the track template image are compared one by one, and the pixel points corresponding to the different pixel values ​​between the track detection image and the track template image are marked as obstacles.

[0044] As a preferred solution of the present invention, the location of the obstacle is determined based on the moving position of the magnetic levitation detection vehicle and the specific marking time point of the obstacle. The specific implementation method is as follows:

[0045] When the magnetic levitation inspection vehicle moves, it captures track images through a camera system, splices the track images in real time and compares them with the track template image. Based on the moving speed of the magnetic levitation inspection vehicle and the corresponding moving time when the obstacle is detected, the distance of the obstacle relative to the starting position of the track inspection is determined to mark the location of the obstacle.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The present invention first captures a comparison reference image based on a track template image, selects an image with the smallest geometric distortion compared with the comparison reference image from multiple images formed by a single shot at the track starting position as the first base image, then selects an image with the smallest geometric distortion compared with the previous base image from multiple images formed by each shot as a stitched image, and stitches the selected stitched images in sequence according to an image stitching method. The stitched track detection image has the smallest geometric distortion compared with the track template image, thereby improving the accuracy of the one-to-one comparison of pixels in the track detection image and the track template image, thereby improving the robustness of obstacle recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other implementation drawings based on the provided drawings without inventive effort.

[0049] Figure 1 Schematic diagram of the flow of the obstacle detection method according to an embodiment of the present invention; DETAILED DESCRIPTION

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] like Figure 1 As shown, the present invention provides an obstacle detection method for a magnetic levitation detection vehicle, comprising the following steps:

[0052] Step 100: Extract the track template image corresponding to the track starting position to the end position, and use the camera system on the inspection vehicle to capture images of the track to be inspected from the starting position to the end position to obtain a real-time track image set corresponding to each shooting.

[0053] In step 100, a two-dimensional coordinate system is constructed on the track template image, wherein the starting position of the track to be detected is the origin, the length direction of the track to be detected is the X-axis, and the width direction of the track to be detected is the Y-axis.

[0054] Determine the pixel value array of the horizontal edge of the track template image along the X-axis direction, and the pixel value array of the vertical edge of the track template image along the Y-axis direction.

[0055] In this embodiment, the track template image can be the most recent track acquisition image in which no obstacles are detected, or it can be the original track acquisition image. After determining the pixel value array corresponding to the horizontal edge of the track template image and the pixel value array corresponding to the vertical edge of the track template image, the track real-time image set can be cut based on the above two pixel value arrays, and the corresponding image is retained, so that the cut image can be compared one by one with the track template image.

[0056] This embodiment uses the camera system on the inspection vehicle to collect images of the track to be inspected from the starting position to the ending position. Each shot forms a corresponding real-time track image set. The camera system uses a single-shot multi-exposure shooting method to capture multiple images in a single shot, and the multiple images captured in a single shot constitute the real-time track image set. This embodiment selects the image with the smallest geometric distortion compared to the previous base image from the multiple images captured each time and stitches them together, thereby improving the robustness of the track detection image and further improving the accuracy of obstacle detection.

[0057] Step 200: Taking the starting reference image corresponding to the starting position of the track template image as a reference, feature extraction is performed on the track real-time image set at the starting position and the starting reference image in sequence. Based on the characteristic geometric distortion of the track real-time image set at the starting position relative to the starting reference image, the first base image as the starting end of the detection is screened from multiple images taken in a single shot of the track real-time image set.

[0058] During the movement of the detection vehicle in this embodiment, a single image with the smallest geometric distortion is selected from the track real-time image set as the first base image, and the track real-time image formed by each shooting is concentrated, and a single image with the smallest geometric distortion compared with the previous base image is selected in sequence. The selected images are geometrically corrected and spliced ​​in sequence to form a track detection image, and the track detection image is compared with the track template image in a calendar manner to determine the obstacles in the track detection image based on the comparison results.

[0059] Among them, the pixel distribution of the first base image determines the screening conditions for screening images when subsequently stitching images. Therefore, in this embodiment, in order to screen out the first base image, this embodiment screens out a starting comparison image from the starting position of the track template image. The interception range of the starting comparison image is the same as the shooting range of a single shot by the photographic system, that is, the pixel range within the starting comparison image is the same as the pixel range corresponding to the image shot by the photographic system at the starting position of the track.

[0060] To extract the geometric distortion of each image in the orbital real-time image set from the starting comparison image, it is necessary to first perform feature extraction on the starting comparison image, then find the corresponding feature in each image in the orbital real-time image set, and determine the geometric distortion corresponding to the feature. Based on the size of the geometric distortion, the image with the smallest geometric distortion is selected as the first base image.

[0061] The feature extraction method for the initial comparison image is as follows:

[0062] A starting comparison image is intercepted from the track template image within the shooting range of the camera system, the starting comparison image is binarized, and a pixel value distribution matrix of the starting comparison image is constructed.

[0063] Based on the pixel value distribution matrix of the initial comparison image, a first regional feature for feature comparison is selected, and a second regional feature of the first regional feature in a corresponding row of the pixel value distribution matrix is ​​obtained, and a third regional feature of the first regional feature in a corresponding column in the initial comparison image is obtained.

[0064] Obtain pixel value distribution queues corresponding to the first region feature, the second region feature, and the third region feature respectively.

[0065] To determine the geometric distortion of each image in the real-time track image set, this embodiment uses the pixel values ​​of regional pixels as the screening object. The advantages of this method are:

[0066] Based on the pixel value distribution matrix of the initial comparison image, the matrix units therein are selected as the first regional features for feature comparison. The selection condition of the first regional feature is that the pixel value domain within the first regional feature is stable, such as the pixel values ​​of most pixels within the first regional feature are 0 or 255, and the pixel values ​​of pixels outside the area surrounded by the first regional feature are 255 or 0. Therefore, the first regional feature with stable pixel values ​​and significant difference from the surrounding pixels can be screened out.

[0067] All pixels in the row where the pixel in the first regional feature is located are used as the second regional feature, and all pixels in the column where the pixel in the first regional feature is located are used as the third regional feature.

[0068] Based on the first region features, second region features and third region features extracted in the starting comparison image, feature extraction is performed on each image of the track real-time image set at the starting position, and the corresponding first region features, second region features and third region features are screened out from each image of the track real-time image set. The lateral deformation θ of each image in the track real-time image set is determined according to the pixel offsets corresponding to the first region features and the second region features, and the column deformation α of each image in the track real-time image set is determined according to the pixel offsets corresponding to the first region features and the third region features.

[0069] The method for implementing feature extraction for each image of the track real-time image set at the starting position is as follows:

[0070] Based on each vertical pixel value distribution queue of the first regional feature and the second regional feature, screening out a first feature matching region identical to each vertical pixel value distribution queue of the first regional feature and the second regional feature from each image of the track real-time image set;

[0071] Based on each horizontal pixel value distribution queue of the first regional feature and the third regional feature, screening out a second feature matching region identical to each horizontal pixel value distribution queue of the first regional feature and the third regional feature from each image of the track real-time image set;

[0072] In each image of the track real-time image set, a row-wise regional feature pixel value distribution matrix corresponding to the first feature matching region and a column-wise regional feature pixel value distribution matrix corresponding to the second feature matching region are generated.

[0073] As a preferred embodiment of this invention, the pixel position offset of all pixel points corresponding to the regional features is used to determine the lateral deformation θ and the column deformation α, which can improve the accuracy of screening out the corresponding regional features from each image of the track real-time image set. If a small number of features are used for matching and calculating the offset, it is very likely that pixel points with matching pixel values ​​but mismatched positions will be found in the track real-time image set, resulting in inaccurate calculated offsets and errors in offset calculation.

[0074] For example, if a feature set is selected for comparison in the initial comparison image, it is very likely that more than two feature sets will be found in each image of the orbital real-time image set. If the image geometric distortion in the orbital real-time image set is relatively large, it is impossible to use the feature pixel position to limit and further select the matching feature set, which will lead to errors in the calculation of geometric distortion.

[0075] This embodiment uses the first region feature, the second region feature and the third region feature as the basis for feature extraction of the image of the track real-time image set. Even if the lateral deformation θ is relatively large, the pixel value distribution of each vertical pixel value distribution queue of the first region feature and the second region feature is the same as that in the initial comparison image. Based on the set formed by each vertical pixel value distribution queue, a part of the features can be extracted from the image of the track real-time image set, and then based on the lateral pixel point position coordinates of the two adjacent vertical pixel value distribution queues, the same first region feature and the second region feature can be extracted from the image of the track real-time image set.

[0076] Similarly, even if the column-wise deformation α is relatively large, the pixel value distribution of each horizontal pixel value distribution queue of the first region feature and the third region feature is the same as that in the initial comparison image. Based on the set formed by each horizontal pixel value distribution queue, a part of the features can be extracted from the image of the track real-time image set, and then based on the vertical pixel point position coordinates of the two adjacent vertical pixel value distribution queues, the same first region feature and third region feature can be extracted from the image of the track real-time image set.

[0077] Due to the above-mentioned feature extraction process, the accuracy of extracting the same features from each image in the track real-time image set is improved, thereby improving the accuracy of calculating the geometric distortion of each image and the accuracy of the first base image screened out. As a result, the first base image has lower geometric distortion than the starting comparison image in the track template image. When image stitching is performed, the other images used for stitching are screened out through the above-mentioned feature screening and feature matching process, and the geometric distortion of the stitched image is also smaller. As a result, the track detection image formed by the entire image stitching has lower geometric distortion than the track template image, thereby improving the accuracy of obstacle detection.

[0078] Furthermore, the method for determining the characteristic geometric distortion of each image in the orbit real-time image set relative to the initial reference image is as follows:

[0079] Construct a row-wise regional feature pixel value distribution matrix using the first regional feature and the second regional feature, determine the minimum coordinates (a1, b1) and maximum coordinates (a2, b2) of the pixel points in the row-wise regional feature pixel value distribution matrix, and calculate the lateral deformation θ of each image in the track real-time image set relative to the starting reference image by fitting;

[0080] Where, tanθ=|b1-b2| / |a1-a2|;

[0081] Constructing a column-wise regional feature pixel value distribution matrix using the first and third regional features, determining the minimum coordinates (a1, b1) and maximum coordinates (a3, b3) of the pixel points in the column-wise regional feature pixel value distribution matrix, and fitting and calculating the column-wise deformation α of each image in the track real-time image set relative to the initial reference image;

[0082] Wherein, tanα=|b1-b3| / |a1-a3|.

[0083] Calculate the characteristic geometric distortion ɡ of each image of the single-shot orbital real-time image set relative to the initial reference image based on the lateral deformation θ and the column deformation α;

[0084] ɡ=θ*W1+α*W2;

[0085] Among them, W1 is the assignment weight of the lateral deformation θ, and W2 is the assignment weight of the column deformation α. ​​W1 has a positive exponential relationship with the lateral deformation θ, and W2 has a positive exponential relationship with the column deformation α.

[0086] In this embodiment, the lateral deformation θ and the columnar deformation α respectively have set thresholds. When the lateral deformation θ exceeds the set threshold, W1 has a positive exponential relationship with the lateral deformation θ. When the columnar deformation α exceeds the set threshold, W2 has a positive exponential relationship with the columnar deformation α. ​​At this time, the characteristic geometric distortion ɡ increases rapidly. Therefore, when the lateral deformation θ or the columnar deformation α exceeds the threshold and does not exceed the threshold, the difference between the characteristic geometric distortion ɡ is very large, which facilitates the screening of the first base image as the starting point of detection based on the characteristic geometric distortion ɡ of each image in the track real-time image set.

[0087] Compare the characteristic geometric distortion ɡ of all images in the single-shot orbital real-time image set, select the image with the smallest characteristic geometric distortion ɡ as the first base image, and perform geometric correction on the first base image so that the pixel distribution of the first base image is the same as that of the starting comparison image. The specific geometric correction method is:

[0088] According to the tilt angle of the lateral deformation relative to the starting reference image, the positive and negative values ​​of the lateral deformation θ are distinguished, and a single row of pixels of the first base image is reversely corrected;

[0089] According to the tilt angle of the column-wise deformation relative to the initial comparison image, the positive and negative values ​​of the column-wise deformation α are distinguished, and the single column of pixels of the first base image are reversely corrected.

[0090] Step 300: Obtain the characteristic geometric distortion of the track real-time image set taken by the camera system each time relative to the previous base layer image, select the image with the smallest geometric distortion from the track real-time image set as the image stitching object, perform geometric correction on the image stitching object, and stitch the overlapping areas in sequence to form a track detection image from the starting position to the ending position.

[0091] According to the above workflow, the first base image is screened out, and the image with the smallest geometric deformation compared with the first base image is screened out from the next set of real-time track images as the stitching image. Similarly, with each shooting work of the inspection vehicle, the track inspection image is stitched together until the overall track inspection image from the starting position to the ending position is formed.

[0092] In step 300, the track detection image from the starting position to the ending position is formed as follows:

[0093] Obtaining a pixel value distribution queue corresponding to an overlapping area between the first base image and a single-shot track real-time image set;

[0094] respectively obtaining pixel value distribution queues corresponding to overlapping areas of a set of track real-time images shot in a single shot;

[0095] Extract features from the overlapping areas, determine the lateral and vertical deformations of each image in the single-shot orbital real-time image set, and select a single image with the smallest characteristic geometric distortion for image stitching based on the characteristic geometric distortions composed of the lateral and vertical deformations;

[0096] Perform geometric correction on the single image and then splice it with the first base image;

[0097] Repeat the above steps to select a single image with the smallest feature geometric distortion compared with the previous stitched image from multiple images of the orbital real-time image set taken in a single shot as the stitched image;

[0098] The stitched image is geometrically corrected and then stitched with the previous base image to form a track detection image from the starting position to the ending position.

[0099] Step 400: Perform a calendar comparison between the track detection image and the track template image to determine obstacles in the track detection image based on the comparison result.

[0100] In step 400, the track detection image is compared with the track template image in a calendar manner as follows:

[0101] Taking the pixel value array data along the X-axis direction of the horizontal edge of the track template image as a reference value, find the pixel point distribution horizontal axis corresponding to the same pixel value array data in the track detection image, split the track detection image according to the pixel point distribution horizontal axis, and retain the image within the two pixel point distribution horizontal axes;

[0102] Taking the pixel value array data along the Y-axis of the vertical edge of the track template image as a reference value, find the vertical axis of pixel point distribution corresponding to the same pixel value array data in the track detection image, split the track detection image by the vertical axis of pixel point distribution, and retain the image within the two vertical axes of pixel point distribution;

[0103] The pixel values ​​of all pixels in the track detection image and the track template image are compared one by one, and the pixels corresponding to the different pixel values ​​between the track detection image and the track template image are marked as obstacles.

[0104] In this embodiment, the track image captured by the camera system is subjected to geometric distortion screening, geometric correction, and image stitching to form a track detection image from the starting position to the ending position of the track. The track detection image contains images corresponding to areas outside the track, so the finally generated track detection image needs to be segmented and cropped. Therefore, before performing a calendar comparison between the track detection image and the track template image, this embodiment uses the pixel value array data of the horizontal edge of the track template image along the X-axis direction and the pixel value array data of the vertical edge of the track template image along the Y-axis direction as the cutting basis, finds the same pixel value array data from the track detection image, uses the above pixel value array data as the dividing line, and cuts the track detection image, and retains the image with the above pixel value array data as the inner side of the dividing line as the track detection image.

[0105] Furthermore, based on the moving position of the magnetic levitation detection vehicle and the specific marking time of the obstacle, the location of the obstacle is determined. The specific implementation method is as follows:

[0106] When the maglev inspection vehicle moves, it uses a camera system to capture track images, splices the track images in real time, and compares them with the track template images. Based on the moving speed of the maglev inspection vehicle and the corresponding moving time when the obstacle is detected, the distance of the obstacle relative to the starting position of the track inspection is determined to mark the location of the obstacle and facilitate subsequent maintenance by staff.

[0107] In this embodiment, a comparison reference image is first captured based on the track template image, and an image with the smallest geometric distortion compared with the comparison reference image is selected from multiple images formed by a single shot at the track starting position as the first base image. Subsequently, an image with the smallest geometric distortion compared with the previous base image is screened out from the multiple images formed by each shot as a stitched image, and the screened stitched images are stitched together in sequence according to the image stitching method. The stitched track detection image has the smallest geometric distortion compared with the track template image, thereby improving the accuracy of the one-to-one comparison of pixels in the track detection image and the track template image, thereby improving the robustness of obstacle recognition.

[0108] The above embodiments are merely exemplary embodiments of the present application and are not intended to limit the scope of the present application. The scope of protection of the present application is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present application within the essence and scope of protection of the present application, and such modifications or equivalent substitutions shall also be deemed to fall within the scope of protection of the present application.

Claims

1. An obstacle detection method for a magnetic levitation inspection vehicle, characterized in that: The following steps are involved: Step 100: Extract the track template image corresponding to the starting position to the ending position of the track to be inspected, and use the camera system on the inspection vehicle to capture images of the track to be inspected from the starting position to the ending position, and obtain a real-time track image set corresponding to each capture; Step 200: Using the starting reference image corresponding to the starting position of the track template image as a reference, feature extraction is performed on the track real-time image set at the starting position and the starting reference image in sequence. Based on the characteristic geometric distortion of the track real-time image set at the starting position relative to the starting reference image, a first base layer image as the detection starting point is selected from multiple images captured in a single shot of the track real-time image set. Step 300: Obtain characteristic geometric distortion of each track real-time image set captured by the camera system relative to the previous base layer image, select the image with the smallest geometric distortion from the track real-time image set as the image stitching object, perform geometric correction on the image stitching object, and stitch the overlapping areas in sequence to form a track detection image from the starting position to the ending position; Step 400: Perform a calendar comparison between the track detection image and the track template image to determine obstacles in the track detection image based on the comparison result.

2. The obstacle detection method for a magnetic levitation inspection vehicle according to claim 1, characterized in that: In step 100, a two-dimensional coordinate system is constructed on the track template image, wherein the starting position of the track to be detected is the origin, the length direction of the track to be detected is the X-axis, and the width direction of the track to be detected is the Y-axis; Determine a pixel value array of a horizontal edge of the track template image along the X-axis direction, and a pixel value array of a vertical edge of the track template image along the Y-axis direction.

3. The obstacle detection method for a magnetic levitation inspection vehicle according to claim 2, characterized in that: The photographing system uses a single-shot multi-exposure photographing mode to capture multiple images in a single shot. In step 200, feature extraction of the initial comparison image is performed as follows: A starting comparison image is intercepted from the track template image within the shooting range of the camera system, the starting comparison image is binarized, and a pixel value distribution matrix of the starting comparison image is constructed. Based on the pixel value distribution matrix of the starting comparison image, a first regional feature for feature comparison is selected, and a second regional feature of a row corresponding to the first regional feature in the pixel value distribution matrix is ​​obtained, and a third regional feature of a column corresponding to the first regional feature in the starting comparison image is obtained. Obtain pixel value distribution queues corresponding to the first region feature, the second region feature, and the third region feature respectively.

4. The obstacle detection method for a magnetic levitation inspection vehicle according to claim 3, characterized in that: The method for implementing feature extraction for each image of the track real-time image set at the starting position is as follows: Based on each vertical pixel value distribution queue of the first regional feature and the second regional feature, screening out a first feature matching region identical to each vertical pixel value distribution queue of the first regional feature and the second regional feature from each image of the track real-time image set; Based on each horizontal pixel value distribution queue of the first regional feature and the third regional feature, screening out a second feature matching region identical to each horizontal pixel value distribution queue of the first regional feature and the third regional feature from each image of the track real-time image set; Generate a row-wise regional feature pixel value distribution matrix corresponding to the first feature matching region and a column-wise regional feature pixel value distribution matrix corresponding to the second feature matching region in each image of the track real-time image set.

5. The obstacle detection method for a magnetic levitation inspection vehicle according to claim 4, characterized in that: The method for determining the characteristic geometric distortion of each image in the track real-time image set relative to the initial comparison image is as follows: Constructing a row-wise regional feature pixel value distribution matrix using the first regional feature and the second regional feature, determining the minimum coordinates (a1, b1) and the maximum coordinates (a2, b2) of the pixel points in the row-wise regional feature pixel value distribution matrix, and fitting and calculating the lateral deformation θ of each image in the track real-time image set relative to the starting comparison image; Where, tanθ=|b1-b2| / |a1-a2|; Constructing a column-wise regional feature pixel value distribution matrix using the first regional feature and the third regional feature, determining the minimum coordinates (a1, b1) and the maximum coordinates (a3, b3) of the pixel points in the column-wise regional feature pixel value distribution matrix, and fitting and calculating the column-wise deformation α of each image in the track real-time image set relative to the starting comparison image; Wherein, tanα=|b1-b3| / |a1-a3|.

6. The obstacle detection method for a magnetic levitation inspection vehicle according to claim 5, characterized in that: Calculate the characteristic geometric distortion ɡ of each image of the track real-time image set shot in a single shot relative to the initial reference image according to the lateral deformation θ and the column deformation α; ɡ=θ*W1+α*W2; Among them, W1 is the assignment weight of the lateral deformation θ, and W2 is the assignment weight of the column deformation α. ​​W1 has a positive exponential relationship with the lateral deformation θ, and W2 has a positive exponential relationship with the column deformation α.

7. The obstacle detection method for a magnetic levitation inspection vehicle according to claim 6, characterized in that: Compare the characteristic geometric distortions ɡ corresponding to all images in the single-shot orbital real-time image set, select the image with the smallest characteristic geometric distortion ɡ as the first base image, and perform geometric correction on the first base image so that the pixel distribution of the first base image is the same as the pixel distribution of the starting comparison image. The specific geometric correction method is: According to the tilt angle of the lateral deformation relative to the initial comparison image, distinguish the positive and negative values ​​of the lateral deformation θ, and perform reverse correction on a single row of pixels of the first base layer image; According to the tilt angle of the column-wise deformation relative to the initial comparison image, the positive and negative values ​​of the column-wise deformation α are distinguished, and the single column of pixels of the first base layer image are reversely corrected.

8. The obstacle detection method for a magnetic levitation inspection vehicle according to claim 5, characterized in that: In step 300, the track detection image from the starting position to the ending position is formed by: Obtaining a pixel value distribution queue corresponding to an overlapping area between the first base layer image and a set of track real-time images captured in a single shot; respectively obtaining pixel value distribution queues corresponding to overlapping areas of a set of track real-time images shot in a single shot; Extracting features from the overlapping area, sequentially determining the lateral deformation and vertical deformation of each image in the track real-time image set captured in a single shot, and selecting a single image with the smallest characteristic geometric distortion for image stitching based on the characteristic geometric distortion composed of the lateral deformation and the vertical deformation; Performing geometric correction on the single image and then splicing it with the first base image; Repeat the above steps to select a single image with the smallest characteristic geometric distortion compared with the previous stitched image from the multiple images of the orbital real-time image set taken in a single shot as the stitched image; The stitched image is geometrically corrected and then stitched with the previous base image to form a track detection image from the starting position to the ending position.

9. The obstacle detection method for a magnetic levitation inspection vehicle according to claim 2, characterized in that: In step 400, the track detection image and the track template image are compared in a calendar manner as follows: Taking the pixel value array data of the horizontal edge of the track template image along the X-axis direction as a reference value, finding the pixel point distribution horizontal axis corresponding to the same pixel value array data in the track detection image, segmenting the track detection image according to the pixel point distribution horizontal axis, and retaining the two images within the pixel point distribution horizontal axis; Using the pixel value array data of the vertical edge of the track template image along the Y-axis as a reference value, finding the pixel point distribution vertical axis corresponding to the same pixel value array data in the track detection image, segmenting the track detection image along the pixel point distribution vertical axis, and retaining the two images within the pixel point distribution vertical axis; The pixel values ​​of all pixels of the track detection image and the track template image are compared one by one, and the pixel points corresponding to the different pixel values ​​between the track detection image and the track template image are marked as obstacles.

10. The obstacle detection method for a magnetic levitation inspection vehicle according to claim 9, characterized in that: Based on the moving position of the magnetic levitation inspection vehicle and the specific marking time point of the obstacle, the location of the obstacle is determined. The specific implementation method is as follows: When the magnetic levitation inspection vehicle moves, it captures track images through a camera system, splices the track images in real time and compares them with the track template image. Based on the moving speed of the magnetic levitation inspection vehicle and the corresponding moving time when the obstacle is detected, the distance of the obstacle relative to the starting position of the track inspection is determined to mark the location of the obstacle.

Images