Catenary tracking parameter dynamic control method based on vanishing point geometric constraint

Abstract

The invention belongs to the field of railway contact network intelligent detection, and relates to a contact network tracking parameter dynamic control method based on vanishing point geometric constraint, which comprises the following steps: carrying out vanishing point calculation on a track area in track video data to obtain a vanishing point calculation result; the vanishing point calculation result comprises vanishing point coordinates; generating a dynamic parameter of the target track based on the vanishing point calculation result; the dynamic parameter is related to the bending degree of the track; dynamically correcting the tracking parameters based on the vanishing point calculation result and the dynamic parameters; and tracking the target track in the track video data based on the corrected tracking parameter to obtain a track tracking result, so as to solve the core defects of the traditional algorithm in continuous video stream processing.

Description

Technical Field

[0001] The present invention relates to the field of intelligent detection of railway catenaries, and specifically discloses a dynamic control method for catenary tracking parameters based on vanishing point geometric constraints. Background Art

[0002] The detection and tracking of catenary systems are important parts of rail transit automation systems. Especially in the monitoring and maintenance of power systems, accurately identifying and tracking catenary targets is crucial for ensuring railway transportation safety. Currently, object detection technologies have been widely applied to catenary target detection. However, when faced with continuous video streams, the repeated detection of the same target has become a problem. Especially in dynamic scenarios, the switching between the target and the background and the rapid movement of the target often lead to false detections. In traditional object tracking, algorithms based on deep learning (such as Kalman filtering, SORT, etc.) are usually used for object matching and duplicate removal. However, these algorithms often have problems such as large computational complexity and poor real-time performance. Especially in long video streams, the tracking accuracy and stability are relatively poor. At the same time, in the curved section, due to the height difference inside and outside the track, the camera coordinate system is tilted, which affects the accuracy of image-based detection and analysis, facing severe challenges.

[0003] Currently, existing technical solutions perform object matching by combining object detection and tracking algorithms, such as the object tracking method based on Kalman filtering and the object detection method based on deep learning (such as the YOLO series, etc.). These methods can solve the tasks of object detection and tracking to a certain extent, but still face the following problems: I. The computational complexity of the tracking algorithm is large, and the real-time processing ability is weak. The catenary target tracking has the characteristics of fast non-reciprocating movement. Directly adopting traditional tracking algorithms will lead to large consumption of system resources and long calculation time, unable to meet the real-time requirements in practical applications. II. Many parameters need to be manually set during the tracking process. The adjustment of these parameters often depends on experience, resulting in unstable effects in different environments. Especially in the dynamic scale change scenario where the target moves from far to near, common measurement parameters (such as intersection over union distance, area, etc.) cannot automatically adapt to the changes of the target, and may lead to tracking failure when the threshold is set inappropriately. III. For the processing of catenary targets in dynamic scenarios, existing technologies have not effectively considered perspective changes and spatial geometric parameters. In the scenarios of curved sections and long-distance fields of view, due to the lack of utilization of the perspective geometric constraints of the track vanishing point, the success rate of target tracking drops by 40%-60%. At the same time, in the case of the tilt of the track plane, traditional image analysis methods cannot directly obtain the tilt parameters of the track plane, resulting in an error of ≥5° in the detection of catenary support tilt, affecting the accurate assessment of the catenary state.

[0004] Aiming at the key technical bottlenecks in the field of intelligent detection of railway catenaries, the present invention aims to construct a dynamic control method for catenary tracking parameters based on vanishing point geometric constraints to solve the core defects existing in the processing of continuous video streams by traditional algorithms. Based on the unique vanishing point geometric features of railway scenes and the dynamic perspective law of catenary targets, a three-dimensional geometric space constraint model is established. Summary of the Invention

[0005] The purpose of the present invention is to provide a dynamic control method for catenary tracking parameters based on vanishing point geometric constraints. The specific scheme is as follows: A dynamic control method for catenary tracking parameters based on vanishing point geometric constraints includes: performing vanishing point calculation on the track area in the track video data to obtain the vanishing point calculation result; the vanishing point calculation result includes vanishing point coordinates; generating dynamic parameters of the target track based on the vanishing point calculation result; the dynamic parameters are related to the bending degree of the track; dynamically correcting the tracking parameters based on the vanishing point calculation result and the dynamic parameters; and tracking the target track in the track video data based on the corrected tracking parameters to obtain the track tracking result.

[0006] Further, it also includes: processing the tilt warning threshold of the target track based on the corrected tracking parameters to obtain a dynamic tilt threshold; the expression of the dynamic tilt threshold is: ; where represents the dynamic tilt threshold; represents the normalized bend offset; represents the fixed tilt threshold.

[0007] Further, obtaining the vanishing point calculation result includes: performing normalization processing on the track video data to obtain normalized video data; performing ballast segmentation on the normalized video data to obtain a ballast mask area; performing track fitting on the ballast mask area to obtain multiple groups of track lines; screening the multiple groups of track lines to obtain suspected tracks; and calculating the vanishing point coordinates based on the suspected tracks to obtain the vanishing point coordinates.

[0008] Further, obtaining the ballast mask area includes: segmenting the ballast in the normalized video data to obtain a binary mask area of the ballast; performing morphological opening operation on the binary mask area using a rectangular kernel and filling the semantic segmentation holes inside the ballast using a cross kernel; and removing the non-ballast areas in the processed binary mask area to obtain the ballast mask area.

[0009] Further, obtaining the suspected tracks includes: scanning the ballast mask area and recording the left and right boundary points of the track; the calculation formula for the left and right boundary points is: ; ; Among them, represents the left boundary point; min represents taking the minimum value; x represents the horizontal axis variable; y represents the vertical axis variable; represents the ballast mask area; represents the right boundary point; max represents taking the maximum value; Scan the target track based on the left and right boundary points of the track. When the boundary points cannot be scanned for a continuous preset number of rows, obtain the left and right track boundary point sets; the expression of the left and right track boundary point sets is; ; ; Among them, represents the left and right track boundary point set; i represents the vertical axis variable; represents the left track boundary point set; represents the right track boundary point set; represents the value of the horizontal axis coordinate of the (y + 1)-th row in the left track boundary point set; represents the value of the horizontal axis coordinate of the y-th row in the left track boundary point set; represents the value of the horizontal axis coordinate of the (y + 1)-th row in the right track boundary point set; represents the value of the horizontal axis coordinate of the y-th row in the right track boundary point set; Based on the geometry of the left and right boundary points, use the least squares method to fit multiple groups of fitting lines in the preset height area. Among them, the expression of the preset height area is: ; Among them, represents the height of the image; The expressions of multiple groups of lines are: ; ; Among them, represents the vertical axis value of the left line; represents the slope of the left line; x represents the horizontal axis variable; represents the intercept of the left line; represents the vertical axis value of the right line; represents the slope of the right line; represents the intercept of the right line; Obtain the average slope of each group of fitting lines, and use the group of fitting lines with the minimum average slope as the suspected track; the expression of the average slope is: ; Among them, represents the average slope.

[0010] Further, obtaining the vanishing point coordinates includes: calculating the geometric perspective vanishing point of the suspected orbit to obtain the initial geometric perspective vanishing point coordinates; the expression of the initial geometric perspective vanishing point coordinates is: ; where x and y respectively represent the horizontal axis variable and the vertical axis variable; represents the initial geometric perspective vanishing point coordinates; calculating the orbit extension vanishing point of the suspected orbit to obtain the orbit extension vanishing point coordinates; the expression of the orbit extension vanishing point coordinates is: ; where, represents the orbit extension vanishing point coordinates; represents the maximum value of the vertical axis; represents the column where the maximum value of the vertical axis of the left straight line is located; represents the column where the maximum value of the vertical axis of the right straight line is located; When the distance between the initial geometric perspective vanishing point coordinates in adjacent frames is greater than the preset distance threshold, assign the initial geometric perspective vanishing point coordinates of the latter frame image to the initial geometric perspective vanishing point coordinates of the previous frame image to obtain the final geometric perspective vanishing point coordinates.

[0011] Further, generating the dynamic parameters of the target orbit includes: judging the orbit type based on the difference of the vanishing point coordinates; the orbit type includes a straight track and a curved track; when the orbit type is a straight track, cache the vertical midpoint value of the straight track; when the orbit type is a curved track, calculate the curve offset of the curved track.

[0012] Further, the calculation formula of the vertical midpoint value is: ; ; The calculation formula of the curve offset is: ; ; where, represents the vertical midpoint value of the straight track; represents the smoothing coefficient; represents the vertical midpoint value of the previous frame; represents the average vertical midpoint value of the current frame; H represents the pixel height of the image; i represents the vertical axis variable; represents the i-th row of the vertical axis; represents the horizontal axis coordinate of the left straight line in the i-th row of the vertical axis; represents the vertical axis coordinate of the right straight line in the i-th row of the vertical axis; represents the normalized curve offset; Represents the curve offset; Represents the width of the image.

[0013] Furthermore, dynamically correct the tracking parameters, including: determining the vanishing point vector based on the vanishing point coordinates; decomposing the displacement of the center point coordinates of the bounding box of the target track recognized in adjacent frame images into a depth component along the vanishing point depth direction and a horizontal component along the vanishing point horizontal direction; based on the depth component and the horizontal component, perform target measurement on the bounding box to be tracked and matched to obtain the dynamically corrected tracking parameters.

[0014] Furthermore, the expression of the vanishing point vector is: ; ; where ImageCenter represents the center point of the image; Represents the width of the image; Represents the height of the image; Represents the vanishing point vector; Represents the initial geometric perspective vanishing point coordinates; The expressions of the depth component and the horizontal component are respectively: ; ; where, Represents the depth component; Represents the center point coordinates of the bounding box of the target track in the subsequent frame image; Represents the center point coordinates of the bounding box of the target track in the previous frame image; Represents taking the absolute value; Represents the horizontal component; The expression of the dynamically corrected tracking parameters is: ; ; ; ; where, Represents the intersection over union of the bounding boxes; Represents the bounding box of the target track in the previous frame image; Represents the bounding box of the target track in the subsequent frame image; Represents the intersection of the bounding boxes in the previous frame image and the subsequent frame image; Represents the union of the bounding boxes in the previous frame image and the subsequent frame image; Represents the central distance between two targets in the approximate 3D space; Represents the maximum distance between two targets in an approximate 3D space; Represents the smoothing coefficient, ; VP-DIoU represents a target box similarity metric improved based on the geometric perspective vanishing point.

[0015] The present invention has the following advantages and beneficial effects: Dynamic parameter adaptive optimization: In the traditional tracking algorithm based on a fixed threshold, when the catenary target moves from far to near (the target area change range reaches 300% - 500%), due to scale sensitivity, the mismerging rate of near-view targets and the missing tracking rate of far-view targets increase. The present invention constructs approximate 3D information in the vanishing point coordinate system, creates VP-DIoU to adapt to the target relationship measurement under straight and curved roads, reduces the target ID switching rate, and improves the target tracking rate.

[0016] Deep coupling of geometric information: In the existing methods, due to ignoring the perspective constraint of the track vanishing point, the target space positioning error in the curved road scene expands to ±15px. The present invention innovatively combines the track segmentation result with the vanishing point solution, establishes a mapping model between the vanishing point - target coordinates, and real-time calculates the inclination angle of the track plane based on the left and right track lines, providing a geometric reference for the catenary inclination defect detection.

[0017] Lightweight real-time processing: Aiming at the computing power bottleneck caused by traditional ReID feature extraction, the present invention proposes a dual metric reduction architecture, which only retains the VP-DIoU distance and the area change rate as the association basis, and combines the dynamic generation mechanism of the vanishing point constraint parameters to meet the requirements of real-time indicators. Brief Description of the Drawings

[0018] Figure 1 Is an exemplary flowchart of a method for dynamically controlling catenary tracking parameters based on vanishing point geometric constraints of the present invention. Detailed Embodiment

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0020] Overhead Catenary Inspection Device (2C Device): The 2C device refers to a safety monitoring equipment that collects images of the status of the overhead catenary and the external environment through a portable video acquisition device temporarily installed in the driver's cab of an operating multiple unit train (or locomotive), and guides the operation and maintenance of the overhead catenary by analyzing the technical state of the overhead catenary. The video data captured by it is continuous. In video analysis, continuous shooting can be carried out for the target to be detected, but only one archiving is expected for the same target during data archiving. Therefore, it is necessary to filter continuous targets. At the same time, when shooting, the rotation of the camera coordinate system will be caused when the train turns, resulting in an increase in the measurement error of the overhead catenary tilt and an increase in false alarms. To solve the above technical problems, the present invention proposes a dynamic control method for overhead catenary tracking parameters based on vanishing point geometric constraints. Figure 1 The following is an exemplary flowchart of a dynamic control method for overhead catenary tracking parameters based on vanishing point geometric constraints according to the present invention. As Figure 1 shown, the dynamic control method for overhead catenary tracking parameters based on vanishing point geometric constraints includes the following: Step 110, perform vanishing point calculation on the track area in the track video data to obtain the vanishing point calculation result; the vanishing point calculation result includes the vanishing point coordinates. The track video data can be the video data obtained by collecting images of the status of the overhead catenary and the external environment through a video acquisition device installed in the driver's cab of an operating multiple unit train (or locomotive). The vanishing point calculation result includes the coordinates of the geometric perspective vanishing point (Vanishing Point, VP) and the track extension vanishing point (Track Extension Point, TEP).

[0021] In some embodiments, obtaining the vanishing point calculation result includes: Perform normalization processing on the track video data to obtain normalized video data. Normalize the input video frame in terms of size, downsample the abscissa to less than 1024 pixels, and maintain the aspect ratio.

[0022] Perform ballast segmentation on the normalized video data to obtain the ballast mask region. The ballast mask region refers to the binary image of the segmented ballast region. In some embodiments, obtaining the ballast mask region includes: segmenting the ballast in the normalized video data to obtain the binary mask region of the ballast. The binary mask region can be obtained through model detection. For example, use a deep learning segmentation model to segment the ballast region. The segmentation network can use semantic segmentation models such as Unet and BiseNetV2. The network outputs a binary mask (mapping the track region to 255 and the background to 0). Perform morphological opening on the binary mask region using a rectangular kernel, and use a cross kernel to fill the semantic segmentation holes inside the ballast. For example, perform morphological opening on the mask region using a 3×3 rectangular kernel to eliminate discrete noise, and use a 5×5 cross kernel to fill the poor semantic segmentation holes inside the ballast. Remove the non-ballast regions within the processed binary mask region to obtain the ballast mask region. For example, remove the non-ballast regions based on the "Conditional Random Field" (CRF). Perform track fitting on the ballast mask region to obtain multiple sets of track lines. Screen the multiple sets of track lines to obtain suspected tracks. In some embodiments, obtaining the suspected tracks includes: scanning the ballast mask region column by column and recording the left and right boundary points of the track; the calculation formulas for the left and right boundary points are: ; ; where, represents the left boundary point; min represents taking the minimum value; x represents the horizontal axis variable; y represents the vertical axis variable; represents the ballast mask region; represents the right boundary point; max represents taking the maximum value.

[0023] For the target track, scan the target track row by row from bottom to top based on the left and right boundary points of the track. When the boundary points cannot be scanned in a continuous preset number of rows, obtain the left and right track boundary point sets; where the preset number can be N, 3 < N ≤ 10. The expression for the left and right track boundary point sets is; ; ; where, represents the left and right track boundary point sets; i represents the vertical axis variable; represents the left track boundary point set; represents the right track boundary point set; represents the value of the horizontal axis coordinate of the (y + 1)-th row in the left track boundary point set; represents the value of the horizontal axis coordinate of the y-th row in the left track boundary point set; represents the value of the horizontal axis coordinate of the (y + 1)-th row in the right track boundary point set; Represents the value of the horizontal axis coordinate of the y-th row in the set of right track boundary points.

[0024] Based on the geometry of the left and right boundary points, multiple sets of fitting lines are obtained by the least squares method for a preset height region. Among them, each set of fitting lines includes two lines, and the expression of the preset height region is: ; Among them, Represents the height of the image.

[0025] The expressions of multiple sets of lines are: ; ; Among them, Represents the vertical axis value of the left line; Represents the slope of the left line; x represents the horizontal axis variable; Represents the intercept of the left line; Represents the vertical axis value of the right line; Represents the slope of the right line; Represents the intercept of the right line.

[0026] Calculate the average slope of each set of the above fitting lines, and take the set of fitting lines with the minimum average slope as the suspected track; the expression of the average slope is: ; Among them, Represents the average slope.

[0027] Based on the suspected track, calculate the vanishing point coordinates to obtain the vanishing point coordinates. In some embodiments, obtaining the vanishing point coordinates includes: calculating the geometric perspective vanishing point of the suspected track to obtain the initial geometric perspective vanishing point coordinates; the expression of the initial geometric perspective vanishing point coordinates is: ; Among them, x and y respectively represent the horizontal axis variable and the vertical axis variable; Represents the initial geometric perspective vanishing point coordinates.

[0028] Calculate the track extension vanishing point of the suspected track to obtain the track extension vanishing point coordinates; the expression of the track extension vanishing point coordinates is: ; Among them, Represents the track extension vanishing point coordinates; Represents the maximum value of the vertical axis; Represents the column where the maximum value of the vertical axis of the left line is located; Represents the column where the maximum value of the vertical axis of the right line is located.

[0029] When the distance between the initial geometric perspective vanishing point coordinates in adjacent frames is greater than a preset distance threshold, assign the initial geometric perspective vanishing point coordinates of the subsequent frame image to the initial geometric perspective vanishing point coordinates of the previous frame image to obtain the final geometric perspective vanishing point coordinates. For example, when calculating for the previous frame image and the current frame image the distance exceeds 5px, assign it to , and use the track information calculated from the previous frame image.

[0030] Step 120, generate dynamic parameters of the target track based on the vanishing point solution result; the dynamic parameters are related to the curvature of the track. In some embodiments, generating the dynamic parameters of the target track includes: Based on the difference in vanishing point coordinates, determine the track type; the track type includes straight tracks and curved tracks. The judgment formula is as follows: ; When the track type is a straight track, for the track boundary point set , in the region of y ∈ [H−H / 5, H], obtain the average vertical midpoint value of the current frame and cache the vertical midpoint value of the straight track (the maximum cache amount is 50 frames). In some embodiments, the calculation formula for the vertical midpoint value is: ; ; When the track type is a curved track, calculate the curve offset of the curved track.

[0031] ; ; Among them, represents the vertical midpoint value of the straight track; represents the smoothing coefficient; represents the vertical midpoint value of the previous frame; represents the average vertical midpoint value of the current frame; H represents the pixel height of the image; i represents the vertical axis variable; represents the i-th row of the vertical axis; represents the horizontal axis coordinate of the left line of the i-th row of the vertical axis; represents the vertical axis coordinate of the right line of the i-th row of the vertical axis; represents the normalized curve offset; represents the curve offset; represents the width of the image.

[0032] Step 130, dynamically correct the tracking parameters based on the vanishing point solution result and the dynamic parameters. In some embodiments, dynamically correcting the tracking parameters includes: Based on the geometric perspective vanishing point , determine the vanishing point vector. In some embodiments, the expression of the vanishing point vector is: ; ; where ImageCenter represents the center point of the image; represents the width of the image; represents the height of the image; represents the vanishing point vector; represents the coordinates of the initial geometric perspective vanishing point.

[0033] Decompose the displacement of the center point coordinates of the bounding box of the target track recognized in adjacent frame images into a depth component along the vanishing point depth direction and a horizontal component along the vanishing point horizontal direction. In some embodiments, the expressions of the depth component and the horizontal component are respectively: ; ; where, represents the depth component; represents the center point coordinates of the bounding box of the target track in the subsequent frame image; represents the center point coordinates of the bounding box of the target track in the previous frame image; represents taking the absolute value; represents the horizontal component.

[0034] Based on the depth component and the horizontal component, use VP-DIoU for target measurement of the bounding box to be tracked and matched, and obtain the dynamically corrected tracking parameters. In some embodiments, the expression of the dynamically corrected tracking parameters is: ; ; ; ; where, represents the intersection over union of the bounding boxes; represents the bounding box of the target track in the previous frame image; represents the bounding box of the target track in the subsequent frame image; represents the intersection of the bounding boxes in the previous frame image and the subsequent frame image; represents the union of the bounding boxes in the previous frame image and the subsequent frame image; represents the central distance between two targets in the approximate 3D space; represents the maximum distance between two targets in the approximate 3D space; represents the smoothing coefficient, ; VP-DIoU represents the similarity metric for object bounding boxes improved based on the geometric perspective vanishing point.

[0035] Step 140: Track the target track in the orbital video data based on the corrected tracking parameters to obtain the track tracking result. Achieve efficient object tracking and real-time calculation of the tilt angle of the orbital plane and compensation for geometric deformation based on dynamic parameters. For example, based on object detection algorithms such as yolov11, detect the target of interest with tilt information; measure two objects of the same type based on VP-DIoU, and solve the optimal matching based on the Hungarian algorithm.

[0036] In some embodiments, it further includes: processing the tilt warning threshold of the target track based on the corrected tracking parameters to obtain a dynamic tilt threshold. For example, quantify the track tilt based on the bend offset, compensate for the turning curvature, and correct the fixed threshold for the tilt warning of the target of interest to a dynamic threshold. The expression of the dynamic tilt threshold is: ; where, represents the dynamic tilt threshold; represents the normalized bend offset; represents the fixed tilt threshold.

[0037] The present invention uses the geometric relationship between the orbital vanishing point and the extended orbital vanishing point to model the depth information and relative position of the target movement, thereby realizing the adaptive adjustment of dynamic parameters. Solve the problem that the fixed threshold fails when the target moves from far to near, reduce the risk of false merging and missed tracking, and improve the tracking stability.

[0038] The present invention combines a deep learning segmentation network with morphological processing, extracts the orbital boundary through boundary tracking and least squares straight line fitting, and then calculates the geometric perspective vanishing point and the extended orbital vanishing point.

[0039] In the bend and long-distance scene, the present invention effectively improves the object positioning accuracy by using the vanishing point geometric constraint, making up for the deficiency that the traditional method fails to fully utilize the perspective information.

[0040] Based on the discrimination between straight sections and bends, the present invention respectively takes measures such as caching the vertical midpoint of the straight section and calculating the bend offset to realize the real-time correction and optimization of the tracking parameters. The dynamically generated parameters can quickly respond to changes in the target scale and perspective, significantly improving the robustness and accuracy of the tracking algorithm in complex scenarios.

[0041] The present invention uses the bend offset and the orbital geometric characteristics to calculate the tilt angle of the orbital plane in real time and perform corresponding geometric compensation to minimize the tilt angle error of the catenary pole during the detection process, providing a more accurate evaluation basis for catenary safety monitoring.

[0042] Each module of the present invention is interconnected to form a complete, real-time and efficient dynamic control system for catenary targets, which not only solves the problems of fixed parameters and missing geometric information in traditional methods, but also takes into account the lightweight and real-time performance of the system.

[0043] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

[0044] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for dynamic control of contact network tracking parameters based on vanishing point geometric constraints, characterized in that: include: Performing vanishing point calculation on the track area in the track video data to obtain a vanishing point calculation result; The vanishing point calculation result includes vanishing point coordinates; Generate dynamic parameters of the target track based on the vanishing point solution results; The dynamic parameter is related to the curvature of the track; Based on the vanishing point solution results and dynamic parameters, the tracking parameters are dynamically corrected; The target track in the track video data is tracked based on the corrected tracking parameters to obtain a track tracking result.

2. The method for dynamic control of contact network tracking parameters based on vanishing point geometric constraints according to claim 1 is characterized in that: Also includes: The tilt warning threshold of the target orbit is processed based on the corrected tracking parameters to obtain a dynamic tilt threshold; The expression of the dynamic tilt threshold is: ； in, represents the dynamic tilt threshold; Indicates the normalized curve offset; Indicates the fixed tilt threshold.

3. The method for dynamic control of contact network tracking parameters based on vanishing point geometric constraints according to claim 1, characterized in that: Get the vanishing point solution results, including: Normalizing the track video data to obtain normalized video data; Performing trackbed segmentation on the normalized video data to obtain trackbed mask area; Perform track fitting on the mask area of ​​the ballast bed to obtain multiple sets of track straight lines; Screen multiple groups of track lines to obtain suspected tracks; The vanishing point coordinates are calculated based on the suspected orbit to obtain the vanishing point coordinates.

4. The method for dynamic control of contact network tracking parameters based on vanishing point geometric constraints according to claim 3 is characterized in that: Get the ballast mask area, including: Segment the roadbed in the normalized video data to obtain a binary mask area of ​​the roadbed; Use a rectangular kernel to perform morphological opening on the binary mask area, and use a cross kernel to fill the semantic segmentation holes inside the roadbed; The non-ballast area in the binary mask area after the processing is eliminated to obtain the ballast mask area.

5. The method for dynamic control of contact network tracking parameters based on vanishing point geometric constraints according to claim 3, characterized in that: Get suspected tracks, including: Scan the ballast mask area and record the left and right boundary points of the track; the calculation formula for the left and right boundary points is: ； ； in, represents the left boundary point; min represents the minimum value; x represents the horizontal axis variable; y represents the vertical axis variable; Indicates the ballast mask area; Indicates the right boundary point; max indicates the maximum value; The target track is scanned based on the left and right boundary points of the track. When no boundary point is found after a preset number of consecutive scans, a set of left and right track boundary points is obtained. The expression of the left and right track boundary point set is: ； ； in, represents the set of left and right track boundary points; i represents the y-axis variable; Represents the set of left track boundary points; represents the set of right track boundary points; Represents the value of the horizontal axis coordinate of the y+1th row in the left track boundary point set; Represents the value of the horizontal axis coordinate of the y-th row in the left track boundary point set; Represents the value of the horizontal axis coordinate of the y+1th row in the right track boundary point set; Represents the value of the horizontal axis coordinate of the y-th row in the right track boundary point set; Based on the geometry of the left and right boundary points, the preset height area is fitted by the least squares method to obtain multiple sets of fitting straight lines, where the expression of the preset height area is: ； in, Indicates the height of the image; The expressions for multiple sets of straight lines are: ； ； in, Indicates the vertical axis value of the left straight line; represents the slope of the straight line on the left; x represents the horizontal axis variable; represents the intercept of the left line; Indicates the vertical axis value of the right straight line; represents the slope of the line on the right; represents the intercept of the right line; The average slope of each group of fitted straight lines is calculated, and the group of fitted straight lines with the smallest average slope is taken as the suspected track; the expression of the average slope is: ； in, represents the average slope.

6. The method for dynamic control of contact network tracking parameters based on vanishing point geometric constraints according to claim 3, characterized in that: Get the vanishing point coordinates, including: The geometric perspective vanishing point of the suspected track is calculated to obtain the initial geometric perspective vanishing point coordinates; the expression of the initial geometric perspective vanishing point coordinates is: ； Among them, x and y represent the horizontal axis variable and the vertical axis variable respectively; Represents the initial geometric perspective vanishing point coordinates; The track extension vanishing point of the suspected track is calculated to obtain the track extension vanishing point coordinates; the expression of the track extension vanishing point coordinates is: ； in, Indicates the coordinates of the vanishing point of the track extension; Indicates the maximum value of the vertical axis; Indicates the column where the maximum value of the vertical axis of the left straight line is located; Indicates the column where the maximum value of the vertical axis of the right straight line is located; When the distance between the initial geometric perspective vanishing point coordinates in adjacent frames is greater than a preset distance threshold, the initial geometric perspective vanishing point coordinates of the subsequent frame image are assigned to the initial geometric perspective vanishing point coordinates of the previous frame image to obtain the final geometric perspective vanishing point coordinates.

7. The method for dynamic control of contact network tracking parameters based on vanishing point geometric constraints according to claim 1, characterized in that: Generate dynamic parameters of target track, including: Based on the difference in vanishing point coordinates, determining the track type; the track type includes a straight track and a curved track; When the track type is a straight track, the vertical midpoint value of the straight track is cached; When the track type is curve, calculate the curve offset of the curve.

8. The method for dynamic control of contact network tracking parameters based on vanishing point geometric constraints according to claim 7, characterized in that: The vertical midpoint value is calculated as: ； ； The calculation formula for the curve offset is: ； ； in, Indicates the vertical midpoint value of the straight line; represents the smoothing coefficient; Indicates the vertical midpoint value of the previous frame; Indicates the average vertical midpoint value of the current frame; H indicates the pixel height of the image; i indicates the vertical axis variable; Represents the i-th row of the vertical axis; Represents the horizontal coordinate of the straight line to the left of the i-th row of the vertical axis; Represents the vertical coordinate of the straight line on the right side of the i-th row of the vertical axis; Indicates the normalized curve offset; Indicates the curve offset; Indicates the width of the image.

9. The method for dynamic control of contact network tracking parameters based on vanishing point geometric constraints according to claim 1, characterized in that: Dynamically modify tracking parameters, including: Based on the vanishing point coordinates, determine the vanishing point vector; Decomposing the displacement of the center point coordinates of the bounding box of the target track identified in the adjacent frame images into a depth component along the depth direction of the vanishing point and a horizontal component along the horizontal direction of the vanishing point; Based on the depth component and the horizontal component, the bounding box to be tracked and matched is measured to obtain the dynamically corrected tracking parameters.

10. The method for dynamic control of contact network tracking parameters based on vanishing point geometric constraints according to claim 9, characterized in that: The expression of the vanishing point vector is: ； ； Among them, ImageCenter represents the midpoint of the image; Indicates the width of the image; Indicates the height of the image; represents the vanishing point vector; Represents the initial geometric perspective vanishing point coordinates; The expressions for the depth component and the horizontal component are: ； ； in, represents the depth component; Represents the center point coordinates of the bounding box of the target track in the next frame image; Represents the center point coordinates of the bounding box of the target track in the previous frame image; Indicates taking the absolute value; represents the horizontal component; The expression of the tracking parameter after dynamic correction is: ； ； ； ； in, represents the intersection-over-union ratio of bounding boxes; Represents the bounding box of the target track in the previous frame image; Represents the bounding box of the target track in the next frame image; Represents the intersection of the bounding boxes in the previous frame and the next frame; Represents the union of the bounding boxes in the previous frame and the next frame; Represents the center distance between two targets in approximate 3D space; Indicates the maximum distance between two targets in approximate 3D space; represents the smoothing coefficient, VP-DIoU represents the target box similarity measurement index based on the improved geometric perspective vanishing point.

Images