Rail disease automatic identification system and method based on image recognition

Abstract

The application discloses an image recognition-based track disease automatic identification system and method, belongs to the track detection technical field, and synchronously collects 2D texture images, 3D point cloud images and encoder mileage signals of a track surface, establishes a correlation index, constructs an initial coordinate mapping relationship, and forms a multi-modal basic data package; a fixed reference object of a positioning track is detected, individual space offset amounts of 2D mapping positions and 3D actual positions are calculated, outlier reference objects are identified based on offset amount distribution and a proportion is counted, and overall offset features are extracted; the deviation source type is discriminated according to the outlier proportion, the initial space corresponding relationship is corrected or reviewed, the 2D pixels and the 3D point cloud are accurately matched based on the corrected mapping relationship, the texture features and the geometric features are fused and input into a disease identification model, the track disease type and the space position are automatically identified, and the automatic detection of the track disease is realized.

Description

Technical Field

[0001] This invention relates to the field of track inspection technology, specifically to an automatic track defect identification system and method based on image recognition. Background Technology

[0002] Automatic track defect identification systems typically utilize a 2D linear array texture camera and a 3D structured light camera. These two types of cameras simultaneously trigger data acquisition, obtaining 2D texture images and 3D point cloud images of the same spatial location on the track surface. The 2D texture image provides grayscale and texture details of the track surface, while the 3D point cloud image provides spatial geometry and depth information. The two are complementary at the feature level. In practical applications, spatial alignment of the two types of data is achieved by mapping the pixel coordinates of the 2D texture image to the 3D point cloud spatial coordinate system, enabling feature-level fusion and defect identification. Compared to a single imaging method, this approach can acquire more comprehensive defect information.

[0003] However, factors such as train body vibration during operation, multi-camera installation and assembly deviations, and encoder mileage accumulation errors can cause the 2D pixel coordinates and 3D point cloud spatial coordinates corresponding to the same track area to fail to match accurately, resulting in spatial coordinate offset. Current common multimodal registration methods typically assume that the two types of data are accurately aligned after one registration and then directly enter the fusion and recognition process, lacking verification of the validity of the spatial correspondence after registration. In practical applications, deviations from different sources exhibit different spatial distributions. Overall deviations affect all locations, while local deviations only affect individual areas. Severe interference may lead to complete registration failure. However, existing methods do not further distinguish the types of deviations based on their spatial distribution characteristics and use a uniform correction logic for all. This means that deviations may still remain in the corrected spatial correspondence, and these deviations will be passed on to subsequent fusion and recognition stages, affecting the accuracy of defect identification. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic track defect identification system and method based on image recognition, so as to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: an automatic identification method for track defects based on image recognition, comprising: S1: Simultaneously acquire 2D texture images and 3D point cloud images of the same spatial location on the track surface, simultaneously collect encoder odometer signals to establish a spatial association index, map the pixel coordinates of the 2D texture images to the 3D point cloud spatial coordinate system to generate an initial spatial correspondence, and associate and integrate the 2D texture images, 3D point cloud images and odometer signals to form a basic data package corresponding to the spatial location. S2: Detect and locate the track-fixed reference objects in the 2D texture image and 3D point cloud image of the basic data package respectively. Calculate the individual spatial offset of each reference object between the 2D mapped position and the actual 3D position according to the initial spatial correspondence. Construct an offset set based on the individual spatial offsets of all reference objects. Identify outlier reference objects in the offset set and count the proportion of outlier reference objects. At the same time, extract the overall offset features based on the offsets of non-outlier reference objects. S3: Determine the type of deviation source based on the proportion of outlier reference objects, including global spatial deviation, local deformation deviation, and registration datum failure; and correct or verify the initial spatial correspondence based on the type of deviation source. S4: Establish the association mapping between 2D pixels and 3D point clouds based on the corrected spatial correspondence. After extracting and fusing 2D texture features and 3D geometric features, input them into the trained disease identification model to automatically identify the type of track disease and its corresponding spatial location.

[0006] S1 specifically includes: S1-1: A 2D linear scan camera and a 3D structured light camera installed on the bottom of the track inspection vehicle synchronously acquire image data of the track surface during the vehicle's movement. The 2D linear scan camera, triggered by encoder pulse signals, scans the track surface line by line at preset spatial sampling intervals, acquiring continuous multi-line grayscale data and stitching them together to form a 2D texture image. The 3D structured light camera, synchronously triggered by encoder pulse signals, projects structured light stripes onto the track surface, and generates a 3D point cloud image by acquiring the deformation patterns formed by the structured light on the track surface. The mileage signal is collected in real time by an encoder. The mileage signal records the mileage information of the detected vehicle. The encoder outputs pulse signals during the rotation of the vehicle's wheels. The real-time mileage of the detected vehicle is determined based on the cumulative number of pulse signals. The mileage signal corresponds one-to-one with the scan lines of the 2D texture image and the acquisition frames of the 3D point cloud image in time sequence. An association index is established with the real-time mileage of the track as the main key and the 2D scan line data and 3D point cloud frame data as the mapping content. S1-2: Preprocess the acquired 2D texture image and 3D point cloud image respectively; synchronize and align the preprocessed 2D texture image and 3D point cloud image according to the encoder mileage signal.

[0007] S1-3: By using the rotation matrix and translation vector between the pre-calibrated 2D linear scan camera and the 3D structured light camera, establish the initial coordinate mapping relationship between the pixel coordinate system of the 2D texture image and the spatial coordinate system of the 3D point cloud, map the pixel coordinates of each pixel in the 2D texture image to the spatial coordinate system of the 3D point cloud, and generate the initial mapped spatial coordinates corresponding to each pixel. The rotation matrix and translation vector are obtained through a joint calibration process. A calibration target with known spatial coordinates is placed in the common field of view of the 2D and 3D cameras, and corresponding 2D images and 3D point clouds are acquired respectively. The pixel coordinates and spatial coordinates of each feature point on the calibration target are extracted. Based on the projection relationship between the three-dimensional spatial coordinates and two-dimensional pixel coordinates of multiple sets of feature points, a set of coordinate transformation equations is constructed, and the specific values ​​of the rotation matrix and translation vector are obtained by minimizing the reprojection error. S1-4: The 2D texture image, 3D point cloud image and corresponding encoder odometer signal that have completed coordinate mapping are associated and packaged to form a basic data packet for the corresponding spatial location; the basic data packet includes the odometer identifier of the corresponding spatial location, the pixel coordinates and corresponding pixel values ​​of the 2D texture image, the spatial coordinates of each point in the 3D point cloud image and the corresponding initial mapped spatial coordinates, and the initial correspondence index between the 2D pixel coordinates and the 3D spatial coordinates.

[0008] S2 specifically includes: S2-1: Detect track-fixed reference objects in 2D texture images, and pre-build and store standard template images of track-fixed reference objects; The track fixing reference objects are structural components in the track structure that are spatially fixed and arranged sequentially at fixed intervals along the track direction, including but not limited to fasteners, elastic strips, and sleeper edges; The standard template image is a standard image region that is pre-extracted from a historical 2D texture image and contains a complete track-fixed reference object and its local neighborhood; the standard image region records the grayscale distribution features and geometric contour features of the track-fixed reference object; During the detection process, the standard template image is traversed across the entire 2D texture image using a sliding window. The correlation coefficient between the standard template image and the corresponding image region at each window position is calculated. The window position where the correlation coefficient is greater than a preset threshold and the correlation coefficient reaches its maximum value within a preset neighborhood is determined as the reference center. The center pixel coordinates of each track-fixed reference object in the pixel coordinate system are obtained. For the i-th track-fixed reference object, its center pixel coordinates are (ui, vi). Locate the corresponding track-fixed reference objects in the 3D point cloud image, determine the track mileage segment corresponding to the current data packet based on the encoder mileage signal, and determine the target spatial region of each reference object in the point cloud based on the layout spacing of the track-fixed reference objects; perform point cloud segmentation and clustering processing within the target spatial region, extract the point cloud subset corresponding to each track-fixed reference object, and take the geometric center of each point cloud subset as the actual spatial position of each track-fixed reference object to obtain the center spatial coordinates of each track-fixed reference object in the 3D spatial coordinate system; for the i-th track-fixed reference object, its center spatial coordinates are denoted as (Xi, Yi, Zi).

[0009] S2-2: For the i-th orbit-fixed reference object, map its center pixel coordinate (ui, vi) in the 2D texture image to the 3D point cloud space coordinate system through the initial spatial correspondence, to obtain the 2D mapped spatial coordinate (Xim, Yim, Zim) of the i-th orbit-fixed reference object; where Xim, Yim, Zim represent the horizontal spatial coordinate, vertical spatial coordinate, and height-direction spatial coordinate of the i-th orbit-fixed reference object after 2D mapping; Calculate the individual spatial offset di of the i-th orbit-fixed reference object: di=||(Xim,Yim,Zim)-(Xi,Yi,Zi)||; S2-3: Form an offset set D={d1,d2,...,dN} from the individual spatial offsets of all orbit-fixed reference objects; where dN represents the individual spatial offset of the N-th orbit-fixed reference object; sort all individual spatial offsets in the offset set D by numerical value, and take the value at the middle position after sorting as the offset median dm; compare the difference between the individual spatial offset di of each reference object and the median dm one by one, when |di-dm| exceeds the preset outlier determination threshold T, identify the corresponding orbit-fixed reference object as an outlier reference object, count the total number Nl of outlier reference objects, and calculate the outlier reference object proportion R=Nl / N; the outlier determination threshold T is determined by counting the offset differences of historical normal registration samples and selecting the 95th percentile.

[0010] S2-4: Retain the individual spatial offsets of non-outlier reference objects in the offset set D to form a non-outlier offset set Df={df1,df2,...,dfq}; where dfq represents the individual spatial offset of the q-th non-outlier reference object; calculate the vector mean of all individual spatial offsets in the non-outlier offset set to obtain the overall translation offset; meanwhile, for the i-th non-outlier reference object, construct a spatial offset vector , project all offset vectors into the orbit horizontal plane, count the azimuth distribution of each vector, and take the central angle corresponding to the angle interval with the highest occurrence frequency as the overall rotation offset; Output the overall translation offset and the overall rotation offset together as the overall offset feature.

[0011] S3 specifically includes: S3-1: Obtain the outlier reference object proportion R obtained in S2, compare R with the preset first outlier proportion threshold R1 and second outlier proportion threshold R2, where 0<R1<R2<1; when R≤R1, determine that the deviation source type is global spatial deviation; when R1<R<R2, determine that the deviation source type is local deformation deviation; when R≥R2, determine that the deviation source type is registration reference failure; S3-2: Based on the overall offset features extracted in S2, including the overall translation offset and the overall rotation offset, each 2D pixel in the initial spatial correspondence is corrected. When the source of the deviation is a global spatial deviation, for the j-th 2D pixel in the initial spatial correspondence, let its initial mapped spatial coordinates be Pj=(Xj,Yj,Zj); using the overall translation offset ΔT=(ΔX,ΔY,ΔZ) and the overall rotation offset ΔR as correction parameters, the initial mapped spatial coordinates of the j-th 2D pixel are transformed by translation and rotation. The corrected spatial coordinates are Pj'=ΔR×Pj+ΔT. When the source of the deviation is local deformation deviation, the outlier references identified in S2 are removed from the reference set. The overall translational offset ΔT' and overall rotational offset ΔR' are recalculated based only on the individual spatial offsets of the non-outlier references. The initial spatial correspondence is corrected using ΔT' and ΔR' as correction parameters. At the same time, the spatial location of the outlier reference is marked as a suspected state change and recorded in the metadata. When the source of the deviation is registration benchmark failure, the initial spatial correspondence is not corrected, and the corresponding basic data package is transferred to the verification channel.

[0012] S3-3: For the basic data packet transferred to the verification channel, retrieve the outlier reference ratio and deviation source type discrimination results of several basic data packets at adjacent positions along the track direction. Compare the outlier reference ratio of the current basic data packet with the outlier reference ratio of adjacent positions: If the outlier reference ratio of adjacent positions is also greater than or equal to R2 and the deviation source type is registration benchmark failure, it is determined that the acquisition environment is abnormal; if the outlier reference ratio of adjacent positions is less than R1 and the deviation source type is global spatial deviation or local deformation deviation, it is determined that the single-point data is abnormal; if the outlier reference ratio and deviation source type discrimination results of adjacent positions do not completely meet any of the above conditions, it is marked as needing further verification, and the deviation feature data of the current and adjacent positions are pushed to the background; the judgment result is recorded in the metadata of the current basic data packet.

[0013] S4 specifically includes: S4-1: Based on the corrected spatial correspondence, the pixel coordinates of each pixel in the 2D texture image are mapped to the 3D point cloud spatial coordinate system, and an association mapping table between 2D pixels and 3D spatial points is established. S4-2: Based on the association mapping table, the texture features of each pixel in the 2D texture image are concatenated with the geometric features of the corresponding spatial point in the 3D point cloud to form a fused feature vector for each pixel. The fused feature vectors of all pixels constitute a fused feature set. The texture features include the gray value, local gradient direction, and magnitude distribution of the pixel. The geometric features include the three-dimensional spatial coordinates, local normal vector, and local curvature of the spatial point. The feature concatenation is to concatenate the texture feature vector of each pixel with the geometric feature vector of the corresponding 3D spatial point according to the dimension to form a joint feature vector. S4-3: Input the fused feature set Ff into the pre-trained disease identification model; the disease identification model is a pixel semantic segmentation model built on a deep convolutional neural network. It takes the fused feature map composed of the fused feature vectors of each pixel arranged in spatial position as input, performs forward propagation calculation on the fused feature vectors of each pixel, outputs the probability distribution of each type of disease corresponding to each pixel, and takes the category with the highest probability as the disease type determination result of the corresponding pixel; at the same time, the spatial coordinates of the corresponding pixel in the 3D point cloud spatial coordinate system are used as the spatial position of the disease; all pixel positions determined to have diseases and their corresponding disease types and spatial positions are summarized to form a disease identification result list. S4-4: Sort and integrate the disease identification results list according to the track mileage order, merge the judgment results that are adjacent in spatial location and have the same disease type, and use the boundary of the merged area as the actual spatial range of the corresponding disease. Finally, output each disease type and its corresponding spatial location and spatial range.

[0014] An image recognition-based automatic track defect identification system includes a data acquisition module, an offset analysis module, a correction processing module, and a defect identification module. The data acquisition module synchronously acquires 2D texture images and 3D point cloud images of the same spatial location on the track surface, synchronously acquires encoder odometer signals to establish a spatial association index, maps the pixel coordinates of the 2D texture images to the 3D point cloud spatial coordinate system to generate an initial spatial correspondence, and integrates the 2D texture images, 3D point cloud images and odometer signals to form a basic data package corresponding to the spatial location. The offset analysis module detects and locates track-fixed reference objects in the 2D texture image and 3D point cloud image of the basic data package, respectively. Based on the initial spatial correspondence, it calculates the individual spatial offset of each reference object between the 2D mapped position and the actual 3D position. The offset set is constructed based on the individual spatial offsets of all reference objects. The outlier reference objects in the offset set are extracted and the proportion of outlier reference objects is counted. At the same time, the overall offset features are extracted based on the offset of non-outlier reference objects. The correction processing module determines the type of deviation source based on the proportion of outliers, including global spatial deviation, local deformation deviation, and registration datum failure; and performs correction or verification processing on the initial spatial correspondence based on the type of deviation source. The disease identification module establishes a mapping between 2D pixels and 3D point clouds based on the corrected spatial correspondence. After extracting and fusing 2D texture features and 3D geometric features, it inputs them into the trained disease identification model to automatically identify the type of track disease and its corresponding spatial location.

[0015] Compared with the prior art, the beneficial effects achieved by the present invention are: This invention detects the spatial positional differences of fixed track reference objects in 2D texture images and 3D point cloud images, calculates the individual spatial offset of each reference object and constructs an offset set, identifies outlier reference objects and counts their proportion, and determines the source type of deviation based on the proportion of outlier reference objects and performs corresponding correction or verification processing. This effectively improves the pertinence and accuracy of the correction, enabling the corrected spatial correspondence to accurately reflect the actual spatial relationship between 2D and 3D, effectively reducing false alarms and missed alarms caused by spatial correspondence deviations, and improving the overall reliability of automatic track defect identification. Furthermore, this invention establishes a precise association mapping between 2D pixels and 3D spatial points through the corrected spatial correspondence, and inputs the 2D texture features and 3D geometric features into the defect identification model for joint discrimination after feature concatenation, improving the accuracy and positioning precision of defect identification. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an image recognition-based automatic identification method for track defects. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This invention provides a technical solution: an automatic identification method for track defects based on image recognition, specifically including: S1: Simultaneously acquire 2D texture images and 3D point cloud images of the same spatial location on the track surface, simultaneously collect encoder odometer signals to establish a spatial association index, map the pixel coordinates of the 2D texture images to the 3D point cloud spatial coordinate system to generate an initial spatial correspondence, and associate and integrate the 2D texture images, 3D point cloud images and odometer signals to form a basic data package corresponding to the spatial location. S2: Detect and locate the track-fixed reference objects in the 2D texture image and 3D point cloud image of the basic data package respectively. Calculate the individual spatial offset of each track-fixed reference object between the 2D mapped position and the actual 3D position according to the initial spatial correspondence. Construct an offset set based on the individual spatial offsets of all reference objects. Identify outlier reference objects in the offset set and count the proportion of outlier reference objects. At the same time, extract the overall offset features based on the offsets of non-outlier reference objects. S3: Determine the type of deviation source based on the proportion of outlier reference objects, including global spatial deviation, local deformation deviation, and registration datum failure; and correct or verify the initial spatial correspondence based on the type of deviation source. S4: Establish the association mapping between 2D pixels and 3D point clouds based on the corrected spatial correspondence. After extracting and fusing 2D texture features and 3D geometric features, input them into the trained disease identification model to automatically identify the type of track disease and its corresponding spatial location.

[0019] S1 specifically includes: S1-1: A 2D linear scan camera and a 3D structured light camera installed on the bottom of the track inspection vehicle synchronously acquire image data of the track surface during the vehicle's movement. The 2D linear scan camera, triggered by encoder pulse signals, scans the track surface line by line at preset spatial sampling intervals, acquiring continuous multi-line grayscale data and stitching them together to form a 2D texture image. The 3D structured light camera, synchronously triggered by encoder pulse signals, projects structured light stripes onto the track surface. By acquiring the deformation patterns formed by the structured light on the track surface, the spatial depth value of each point on the track surface is calculated based on the triangulation principle to generate a 3D point cloud image. The encoder collects mileage signals in real time and establishes a time-series correspondence between them and the scan lines of the 2D texture image and the acquisition frames of the 3D point cloud image. The real-time mileage of the track is used as the main key, and the 2D scan line data and 3D point cloud frame data are used as the mapping content. The mileage signal records the mileage information of the detected vehicle. The encoder outputs pulse signals during the rotation of the vehicle's wheels. The real-time mileage of the detected vehicle is determined based on the cumulative number of pulse signals. The specific process is as follows: The encoder rotates synchronously with the wheels of the track detection vehicle. When the wheel completes one full rotation, the encoder outputs G pulse signals. Based on the wheel circumference L, the pulse equivalent s is calculated: s = L / G; where the pulse equivalent s represents the distance the vehicle travels when the encoder outputs a single pulse. During the inspection process, the track inspection vehicle continuously collects the encoder output pulse signals, counts and accumulates the pulse signals in real time to obtain the total number of pulses Y; based on the pulse equivalent s and the cumulative number of pulses Y, the vehicle's current total distance traveled S is calculated in real time: S = Y × s; S1-2: Median filtering is used to remove impulse noise from the acquired 2D texture image, and histogram equalization is used to stretch the grayscale distribution of the 2D texture image to improve image contrast. Statistical filtering is used to detect and remove outliers from the acquired 3D point cloud image. Outliers are points with sparse spatial distribution and whose distance from neighboring point clouds exceeds a preset multiple of the standard deviation. A voxel grid filter is used to downsample the 3D point cloud image, dividing it into several three-dimensional voxel grids. The centroid of all point clouds within each voxel grid is used to replace all point clouds within that voxel grid. The preprocessed 2D texture image and 3D point cloud image are synchronized and aligned according to the encoder odometer signal. S1-3: By using the rotation matrix and translation vector between the pre-calibrated 2D linear scan camera and the 3D structured light camera, establish the initial coordinate mapping relationship between the pixel coordinate system of the 2D texture image and the spatial coordinate system of the 3D point cloud, map the pixel coordinates of each pixel in the 2D texture image to the spatial coordinate system of the 3D point cloud, and generate the initial mapped spatial coordinates corresponding to each pixel. The rotation matrix and translation vector are obtained through a joint calibration process: a calibration target with known spatial coordinates is placed in the common field of view of the 2D and 3D cameras, and the corresponding 2D images and 3D point clouds are acquired respectively. The pixel coordinates and spatial coordinates of each feature point are extracted. Based on the projection relationship between the three-dimensional spatial coordinates and two-dimensional pixel coordinates of multiple feature points, a set of coordinate transformation equations is constructed, and the specific values ​​of the rotation matrix and translation vector are obtained by minimizing the reprojection error. The 2D texture image, 3D point cloud image and corresponding encoder odometer signal that have completed coordinate mapping are associated and packaged to form a basic data packet for the corresponding spatial location. The basic data packet includes the odometer identifier of the spatial location, the pixel coordinates and corresponding pixel values ​​of the 2D texture image, the spatial coordinates of each point in the 3D point cloud image and the corresponding initial mapped spatial coordinates, and the initial correspondence index between the 2D pixel coordinates and the 3D spatial coordinates.

[0020] S2 specifically includes: S2-1: Detect track-fixed reference objects in 2D texture images, and pre-build and store standard template images of track-fixed reference objects; The track fixing reference objects are structural components in the track structure that are spatially fixed and arranged sequentially at fixed intervals along the track direction, including but not limited to fasteners, elastic strips, and sleeper edges; The standard template image is a standard image region that is pre-extracted from a historical 2D texture image and contains a complete track-fixed reference object and its local neighborhood; the standard image region records the grayscale distribution features and geometric contour features of the track-fixed reference object; During the detection process, the standard template image is traversed across the entire 2D texture image using a sliding window. The correlation coefficient between the standard template image and the corresponding image region at each window position is calculated. The window position where the correlation coefficient is greater than a preset threshold and the correlation coefficient reaches its maximum value within a preset neighborhood is determined as the reference center. The center pixel coordinates of each track-fixed reference object in the pixel coordinate system are obtained. For the i-th track-fixed reference object, its center pixel coordinates are (ui, vi). Locate the corresponding track-fixed reference objects in the 3D point cloud image, determine the track mileage segment corresponding to the current data packet based on the encoder mileage signal, and determine the target spatial region of each reference object in the point cloud based on the layout spacing of the track-fixed reference objects; perform point cloud segmentation and clustering processing within the target spatial region, extract the point cloud subset corresponding to each track-fixed reference object, and take the geometric center of each point cloud subset as the actual spatial position of each track-fixed reference object to obtain the center spatial coordinates of each track-fixed reference object in the 3D spatial coordinate system; for the i-th track-fixed reference object, its center spatial coordinates are denoted as (Xi, Yi, Zi). S2-2: For the i-th track-fixed reference object, its center pixel coordinates (ui, vi) in the 2D texture image are mapped to the 3D point cloud spatial coordinate system through the initial spatial correspondence to obtain the 2D mapped spatial coordinates (Xim, Yim, Zim) of the i-th track-fixed reference object; where Xim, Yim, and Zim represent the horizontal spatial coordinates, vertical spatial coordinates, and height spatial coordinates of the i-th track-fixed reference object after 2D mapping; Calculate the individual spatial offset di of the i-th orbital fixed reference: di=||(Xim,Yim,Zim)-(Xi,Yi,Zi)||; S2-3: Form an offset set D={d1,d2,...,dN} from the individual spatial offsets of all track fixed reference objects, where dN represents the individual spatial offset of the Nth track fixed reference object; sort all individual spatial offsets in the offset set D by numerical value, and take the value at the middle position after sorting as the offset median dm; compare the difference between the individual spatial offset di of each reference object and the median dm one by one, when |di-dm| exceeds the preset outlier determination threshold T, identify the corresponding track fixed reference object as an outlier reference object, count the total number of outlier reference objects Nl, and calculate the outlier reference object ratio R=Nl / N; the outlier determination threshold T is determined by counting the offset differences of historical normal registration samples and selecting the 95th percentile.

[0021] S2-4: Keep the individual spatial offsets of non-outlier reference objects in the offset set D to form a non-outlier offset set Df={df1,df2,...,dfq}, where dfq represents the individual spatial offset of the qth non-outlier reference object; calculate the vector mean of all individual spatial offsets in the non-outlier offset set to obtain the overall translation offset; meanwhile, for the ith non-outlier reference object, construct a spatial offset vector , project all offset vectors into the track horizontal plane, count the azimuth distribution of each vector, and take the central angle corresponding to the angle interval with the highest occurrence frequency as the overall rotation offset; Output the overall translation offset and the overall rotation offset together as the overall offset feature.

[0022] S3 specifically includes: S3-1: Obtain the outlier reference object ratio R obtained in S2, compare R with the preset first outlier ratio threshold R1 and second outlier ratio threshold R2, where 0<R1<R2<1; when R≤R1, determine that the deviation source type is global spatial deviation, in this case, the individual spatial offsets of the vast majority of reference objects in the offset distribution are concentrated near the median value, and the overall offset feature can effectively characterize the spatial offset situation of all reference objects; when R1<R<R2, determine that the deviation source type is local deformation deviation, in this case, there is a certain proportion of outlier reference objects in the offset distribution, but the individual spatial offsets of most reference objects are still concentrated near the median value, and the offset distribution presents the feature of coexistence of central aggregation and local dispersion; when R≥R2, determine that the deviation source type is registration reference failure, in this case, the offset distribution is divergent as a whole, no statistically significant central trend can be extracted from the offset distribution, and the overall offset feature can no longer effectively characterize the spatial offset situation of the reference objects; S3-2: Based on the overall offset features extracted in S2, including the overall translation offset and the overall rotation offset, each 2D pixel in the initial spatial correspondence is corrected. When the source of the deviation is a global spatial deviation, for the j-th 2D pixel in the initial spatial correspondence, let its initial mapped spatial coordinates be Pj=(Xj,Yj,Zj); using the overall translation offset ΔT=(ΔX,ΔY,ΔZ) and the overall rotation offset ΔR as correction parameters, the initial mapped spatial coordinates of the j-th 2D pixel are transformed by translation and rotation. The corrected spatial coordinates are Pj'=ΔR×Pj+ΔT. When the source of the deviation is local deformation deviation, the outlier reference objects identified in S2 are removed from the reference object set. The overall translational offset ΔT' and overall rotational offset ΔR' are recalculated based only on the individual spatial offsets of the non-outlier reference objects. The initial spatial correspondence is corrected using ΔT' and ΔR' as correction parameters. At the same time, the spatial location of the outlier reference object is marked as a suspected state change and recorded in the metadata. When the source of the deviation is registration benchmark failure, the initial spatial correspondence is not corrected, and the corresponding basic data package is transferred to the verification channel. S3-3: For the basic data packet transferred to the verification channel, retrieve the outlier reference ratio and deviation source type discrimination results of several basic data packets at adjacent positions along the track direction. Compare the outlier reference ratio of the current basic data packet with the outlier reference ratio of adjacent positions: If the outlier reference ratio of adjacent positions is also greater than or equal to R2 and the deviation source type is registration benchmark failure, it is determined that the acquisition environment is abnormal; if the outlier reference ratio of adjacent positions is less than R1 and the deviation source type is global spatial deviation or local deformation deviation, it is determined that the single-point data is abnormal; if the outlier reference ratio and deviation source type discrimination results of adjacent positions do not completely meet any of the above conditions, it is marked as needing further verification, and the deviation feature data of the current and adjacent positions are pushed to the background; the judgment result is recorded in the metadata of the current basic data packet.

[0023] S4 specifically includes: S4-1: Based on the spatial correspondence after correction in S3, establish the association mapping between the 2D texture image pixel coordinate system and the 3D point cloud spatial coordinate system; for the k-th pixel in the 2D texture image, let its pixel coordinates in the 2D texture image pixel coordinate system be (uk, vk). According to the corrected spatial correspondence, map it to the 3D point cloud spatial coordinate system to obtain the corresponding mapped spatial coordinates Pk'=(Xk', Yk', Zk'). Establish the final mapping table Map={(uk, vk)→Pk'|k=1,2,...,M} between 2D pixels and 3D spatial points, where M represents the total number of pixels in the 2D texture image; S4-2: Based on the association mapping table, the texture features of each pixel in the 2D texture image are fused with the geometric features of the corresponding spatial points in the 3D point cloud. The texture features include the gray value, local gradient direction, and amplitude distribution of the pixel. The gray value is directly read from each pixel in the 2D texture image. The local gradient direction and amplitude distribution are obtained by calculating the gray value change rate of each pixel in the horizontal and vertical directions. The normal vector and curvature are obtained by performing principal component analysis on the local neighborhood point cloud of each spatial point. The geometric features include the three-dimensional spatial coordinates, local normal vector, and local curvature of the spatial point. The local normal vector is obtained by performing principal component analysis on the local neighborhood point cloud of each spatial point. The local curvature is obtained by calculating the rate of change of the local normal vector. For the k-th pixel and its corresponding 3D spatial point, its texture features and geometric features are concatenated to form a fused feature vector Ff,k=(F2D,k,F3D,k), where F2D,k represents the texture feature vector of the k-th pixel and F3D,k represents the geometric feature vector of the k-th 3D spatial point. The fused feature vectors of all pixels constitute the final fused feature set Ff={Ff,k|k=1,2,...,M}. The feature concatenation is to concatenate the texture feature vector of each pixel with the geometric feature vector of the corresponding 3D spatial point according to the dimension to form a fused feature vector. S4-3: Input the fused feature set Ff into the pre-trained disease identification model; the disease identification model is a pixel semantic segmentation model built on a deep convolutional neural network, calling the UNet or DeepLabV3+ network architecture, setting the number of channels to h, where h is the dimension of the fused feature vector, and outputting the probability distribution of each pixel corresponding to various diseases. The disease identification model pre-extracts the fused feature vectors of each pixel position from the registered 2D texture image and 3D point cloud data, and labels the corresponding disease category of each pixel position to construct a training dataset; the disease category includes at least one or more of the following: cracks, peeling, and wear; based on the standard UNet / DeepLabV3+ architecture, the number of input convolutional layer channels is modified from 3 to h; during model training, with the goal of minimizing the cross-entropy loss between the pixel-by-pixel classification result of the model output and the labeled label, the backpropagation algorithm is used to iteratively optimize the model parameters until the classification accuracy of the model on the validation set reaches the preset requirement; After the model is trained, the fused feature map, which is composed of the fused feature vectors of each pixel arranged in spatial position, is used as the input of the disease identification model. The disease identification model performs forward propagation calculation on the fused feature vectors of each pixel and outputs the probability distribution of each type of disease for each pixel. The category with the highest probability is taken as the disease type determination result for the corresponding pixel. At the same time, the spatial coordinates of the corresponding pixel in the 3D point cloud spatial coordinate system are used as the spatial location of the disease. All pixel locations that are determined to have diseases and their corresponding disease types and spatial locations are summarized to form a disease identification result list Result={(Pk',Ck)|k=1,2,...,M}, where Ck represents the disease category determined at the k-th pixel location. S4-4: Sort and integrate the disease identification results list Result according to the track mileage order, merge the judgment results that are adjacent in spatial location and have the same disease type, and use the boundary of the merged area as the actual spatial range of the corresponding disease. Finally, output each disease type and its corresponding spatial location and spatial range.

[0024] This embodiment uses a track inspection vehicle equipped with a 2D linear scan camera and a 3D structured light camera to inspect a section of track at a speed of 60 km / h. The encoder triggers a synchronous acquisition every 1 meter of travel to achieve continuous acquisition of track surface data. The 2D camera scans line by line to form a track texture image, which retains clear textures of fasteners, sleepers, etc. after noise reduction and enhancement preprocessing. The 3D camera acquires track point cloud data by projecting structured light, and removes outliers after filtering and noise reduction. An initial mapping relationship between pixel coordinates and 3D spatial coordinates is established through a pre-calibrated camera rotation matrix and translation vector. The single-frame 2D image, 3D point cloud, and corresponding mileage information are associated and integrated to form a basic data packet with 1 meter as the unit. The data packet contains mileage markers, image pixel values, point cloud spatial coordinates, and the initial mapping relationship. Using rail fasteners as reference objects for track fixing, a pre-constructed standard fastener template image is called. A sliding window approach is used to traverse and match the 2D texture image, with a window sliding step size of 2 pixels. The normalized correlation coefficient between the standard template image and the corresponding image region is calculated window by window, resulting in a correlation coefficient response map for the entire image. A threshold filter is applied to the response map, setting the correlation coefficient threshold to 0.7, retaining only window positions with a correlation coefficient greater than 0.7. Then, neighborhood non-maximum suppression is applied to the retained positions, setting the neighborhood window to 50×50 pixels. Only the window position with the highest correlation coefficient within the neighborhood is retained as the final detected reference center. A total of 10 fastener pixel coordinates are located within a single frame image. Simultaneously... Based on mileage information and standard fastener spacing, corresponding spatial regions are delineated in the 3D point cloud. Point cloud subsets corresponding to each fastener are extracted through point cloud segmentation and clustering, with the geometric center as the actual spatial coordinate. The 2D center coordinates of each fastener are initially mapped to 3D space, and the Euclidean distance between them and the actual spatial coordinates is calculated to obtain a set of 10 individual spatial offsets. After sorting the set, the median value is taken, and the outlier threshold is set using the 95th percentile of historical normal data. One offset exceeding the threshold is identified as an outlier reference, and the outlier ratio is calculated to be approximately 10%. The overall translational offset is obtained by calculating the vector mean based on the remaining effective offsets. The overall rotational offset is determined by statistical analysis of the horizontal plane azimuth angle, forming a complete overall offset feature. The deviation type is determined based on the outlier percentage and two preset thresholds. If the current percentage is lower than the first threshold, it is determined to be a global spatial deviation. The overall translational and rotational offsets are used as correction parameters to perform rigid body transformation correction on the initial spatial correspondence and update the mapping spatial coordinates of all pixels. If it is determined to be a local deformation deviation, the outlier references are removed, the offset features are recalculated and corrected, and the suspected deformation positions are marked. If the percentage exceeds the second threshold, it is determined to be a registration reference failure. The process is then directly transferred to the review process, and adjacent data packets are retrieved for horizontal comparison to distinguish between environmental interference anomalies and single-point data anomalies. The judgment results are recorded in the data packet metadata. Based on the corrected and accurate spatial correspondence, a one-to-one mapping table between 2D pixels and 3D point clouds is established. Texture features such as grayscale value and gradient direction are extracted pixel by pixel, while geometric features such as spatial coordinates and normal vectors of the corresponding 3D points are extracted. The two types of feature vectors are directly concatenated according to their dimensions to form a fusion feature vector, which constitutes the fusion feature set of the entire image. The fusion feature set is input into the trained disease identification model, and the disease category probability is output pixel by pixel. The maximum probability is taken as the identification result, and the pixels with diseases and their spatial locations are summarized. Then, the results of diseases that are spatially adjacent and of the same type are sorted by mileage, and the regions are merged to determine the disease boundary and spatial range. Finally, the disease type and its corresponding spatial coordinates, corresponding mileage and spatial range are output, realizing the automated and high-precision identification and location of track diseases.

[0025] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An automatic identification method for track defects based on image recognition, characterized in that: The method includes: S1: Simultaneously acquire 2D texture images and 3D point cloud images of the same spatial location on the track surface, simultaneously collect encoder odometer signals to establish a spatial association index, map the pixel coordinates of the 2D texture images to the 3D point cloud spatial coordinate system to generate an initial spatial correspondence, and associate and integrate the 2D texture images, 3D point cloud images and odometer signals to form a basic data package corresponding to the spatial location. S2: Detect and locate the track-fixed reference objects in the 2D texture image and 3D point cloud image of the basic data package respectively. Calculate the individual spatial offset of each reference object between the 2D mapped position and the actual 3D position according to the initial spatial correspondence. Construct an offset set based on the individual spatial offsets of all reference objects. Identify outlier reference objects in the offset set and count the proportion of outlier reference objects. At the same time, extract the overall offset features based on the offsets of non-outlier reference objects. S3: Determine the type of deviation source based on the proportion of outlier reference objects, including global spatial deviation, local deformation deviation, and registration datum failure; and correct or verify the initial spatial correspondence based on the type of deviation source. S4: Establish the association mapping between 2D pixels and 3D point clouds based on the corrected spatial correspondence. After extracting and fusing 2D texture features and 3D geometric features, input them into the trained disease identification model to automatically identify the type of track disease and its corresponding spatial location.

2. The automatic identification method for track defects based on image recognition according to claim 1, characterized in that: S1 specifically includes: S1-1: A 2D linear scan camera and a 3D structured light camera installed on the bottom of the track inspection vehicle synchronously acquire image data of the track surface during the vehicle's movement. The 2D linear scan camera, triggered by encoder pulse signals, scans the track surface line by line at preset spatial sampling intervals, acquiring continuous multi-line grayscale data and stitching them together to form a 2D texture image. The 3D structured light camera, synchronously triggered by encoder pulse signals, projects structured light stripes onto the track surface, and generates a 3D point cloud image by acquiring the deformation patterns formed by the structured light on the track surface. The mileage signal is collected in real time by an encoder. The mileage signal records the mileage information of the detected vehicle. The encoder outputs pulse signals during the rotation of the vehicle's wheels. The real-time mileage of the detected vehicle is determined based on the cumulative number of pulse signals. The mileage signal corresponds one-to-one with the scan lines of the 2D texture image and the acquisition frames of the 3D point cloud image in time sequence. An association index is established with the real-time mileage of the track as the main key and the 2D scan line data and 3D point cloud frame data as the mapping content. S1-2: Preprocess the acquired 2D texture image and 3D point cloud image respectively; synchronize and align the preprocessed 2D texture image and 3D point cloud image according to the encoder mileage signal.

3. The automatic identification method for track defects based on image recognition according to claim 2, characterized in that: S1 further includes: S1-3: By using the rotation matrix and translation vector between the pre-calibrated 2D linear scan camera and the 3D structured light camera, establish the initial coordinate mapping relationship between the pixel coordinate system of the 2D texture image and the spatial coordinate system of the 3D point cloud, map the pixel coordinates of each pixel in the 2D texture image to the spatial coordinate system of the 3D point cloud, and generate the initial mapped spatial coordinates corresponding to each pixel. The rotation matrix and translation vector are obtained through a joint calibration process. A calibration target with known spatial coordinates is placed in the common field of view of the 2D and 3D cameras, and corresponding 2D images and 3D point clouds are acquired respectively. The pixel coordinates and spatial coordinates of each feature point on the calibration target are extracted. Based on the projection relationship between the three-dimensional spatial coordinates and two-dimensional pixel coordinates of multiple sets of feature points, a set of coordinate transformation equations is constructed, and the specific values ​​of the rotation matrix and translation vector are obtained by minimizing the reprojection error. S1-4: The 2D texture image, 3D point cloud image and corresponding encoder odometer signal that have completed coordinate mapping are associated and packaged to form a basic data packet for the corresponding spatial location; the basic data packet includes the odometer identifier of the corresponding spatial location, the pixel coordinates and corresponding pixel values ​​of the 2D texture image, the spatial coordinates of each point in the 3D point cloud image and the corresponding initial mapped spatial coordinates, and the initial correspondence index between the 2D pixel coordinates and the 3D spatial coordinates.

4. The automatic identification method for track defects based on image recognition according to claim 1, characterized in that: S2 specifically includes: S2-1: Detect track-fixed reference objects in 2D texture images, and pre-build and store standard template images of track-fixed reference objects; The track fixing reference objects are structural components in the track structure that are spatially fixed and arranged sequentially at fixed intervals along the track direction, including but not limited to fasteners, elastic strips, and sleeper edges; The standard template image is a standard image region that is pre-extracted from a historical 2D texture image and contains a complete track-fixed reference object and its local neighborhood; the standard image region records the grayscale distribution features and geometric contour features of the track-fixed reference object; During the detection process, the standard template image is traversed across the entire 2D texture image using a sliding window. The correlation coefficient between the standard template image and the corresponding image region at each window position is calculated. The window position where the correlation coefficient is greater than a preset threshold and the correlation coefficient reaches its maximum value within a preset neighborhood is determined as the reference center. The center pixel coordinates of each track-fixed reference object in the pixel coordinate system are obtained. For the i-th track-fixed reference object, its center pixel coordinates are (ui, vi). Locate the corresponding track-fixed reference objects in the 3D point cloud image, determine the track mileage segment corresponding to the current data packet based on the encoder mileage signal, and determine the target spatial region of each reference object in the point cloud based on the layout spacing of the track-fixed reference objects; perform point cloud segmentation and clustering processing within the target spatial region, extract the point cloud subset corresponding to each track-fixed reference object, and take the geometric center of each point cloud subset as the actual spatial position of each track-fixed reference object to obtain the center spatial coordinates of each track-fixed reference object in the 3D spatial coordinate system; for the i-th track-fixed reference object, its center spatial coordinates are denoted as (Xi, Yi, Zi).

5. The automatic identification method for track defects based on image recognition according to claim 4, characterized in that: S2 further includes: S2-2: For the i-th orbit-fixed reference object, map its center pixel coordinates (ui, vi) in the 2D texture image to the 3D point cloud space coordinate system through the initial spatial correspondence, to obtain the 2D mapped spatial coordinates (Xim, Yim, Zim) of the i-th orbit-fixed reference object; wherein Xim, Yim, Zim respectively represent the horizontal spatial coordinate, vertical spatial coordinate, and height-direction spatial coordinate of the i-th orbit-fixed reference object after 2D mapping; Calculate the individual spatial offset di of the i-th orbit-fixed reference object: di=||(Xim,Yim,Zim)-(Xi,Yi,Zi)||; S2-3: Form an offset set D={d1,d2,...,dN} from the individual spatial offsets of all orbit-fixed reference objects, where dN represents the individual spatial offset of the N-th orbit-fixed reference object; sort all individual spatial offsets in the offset set D by numerical value, and take the value at the middle position after sorting as the offset median dm; compare the difference between the individual spatial offset di of each reference object and the median dm one by one, when |di-dm| exceeds the preset outlier determination threshold T, identify the corresponding orbit-fixed reference object as an outlier reference object, count the total number of outlier reference objects Nl, and calculate the outlier reference object proportion R=Nl / N; the outlier determination threshold T is determined by counting the offset difference values of historical normal registration samples and selecting the 95th percentile.

6. The automatic identification method for track defects based on image recognition according to claim 5, characterized in that: The step S2 further comprises: S2-4: Retain the individual spatial offsets of non-outlier reference objects in the offset set D to form a non-outlier offset set Df={df1,df2,...,dfq}; where dfq represents the individual spatial offset of the q-th non-outlier reference object; calculate the vector mean of all individual spatial offsets in the non-outlier offset set to obtain the overall translation offset; simultaneously, construct the spatial offset vector for the i-th non-outlier reference object. Project all offset vectors onto the horizontal plane of the track, count the azimuth angle distribution of each vector, and take the center angle corresponding to the angle interval with the highest frequency as the overall rotation offset. Output the overall translation offset and the overall rotation offset together as an overall offset feature.

7. The automatic identification method for track defects based on image recognition according to claim 1, characterized in that: The step S3 specifically comprises: S3-1: Acquire the outlier reference object proportion R obtained in S2, compare R with the preset first outlier proportion threshold R1 and the preset second outlier proportion threshold R2, where 0<R1<R2<1; when R≤R1, determine that the deviation source type is global spatial deviation; when R1<R<R2, determine that the deviation source type is local deformation deviation; when R≥R2, determine that the deviation source type is registration benchmark failure; S3-2: Perform correction processing on each 2D pixel in the initial spatial correspondence according to the overall offset feature extracted in S2, which includes the overall translation offset and the overall rotation offset; when the deviation source type is global spatial deviation, for the j-th 2D pixel in the initial spatial correspondence, let its initial mapped spatial coordinates be Pj=(Xj,Yj,Zj); use the overall translation offset ΔT=(ΔX,ΔY,ΔZ) and the overall rotation offset ΔR as correction parameters to perform translation transformation and rotation transformation on the initial mapped spatial coordinates of the j-th 2D pixel, and the corrected spatial coordinates Pj'=ΔR×Pj+ΔT; When the source of the deviation is local deformation deviation, the outlier references identified in S2 are removed from the reference set. The overall translational offset ΔT' and overall rotational offset ΔR' are recalculated based only on the individual spatial offsets of the non-outlier references. The initial spatial correspondence is corrected using ΔT' and ΔR' as correction parameters. At the same time, the spatial location of the outlier reference is marked as a suspected state change and recorded in the metadata. When the source of the deviation is registration benchmark failure, the initial spatial correspondence is not corrected, and the corresponding basic data package is transferred to the verification channel.

8. The automatic identification method for track defects based on image recognition according to claim 7, characterized in that: S3 further includes: S3-3: For the basic data packet transferred to the verification channel, retrieve the outlier reference ratio and deviation source type discrimination results of several basic data packets at adjacent positions along the track direction. Compare the outlier reference ratio of the current basic data packet with the outlier reference ratio of adjacent positions: If the outlier reference ratio of adjacent positions is also greater than or equal to R2 and the deviation source type is registration benchmark failure, it is determined that the acquisition environment is abnormal; if the outlier reference ratio of adjacent positions is less than R1 and the deviation source type is global spatial deviation or local deformation deviation, it is determined that the single-point data is abnormal; if the outlier reference ratio and deviation source type discrimination results of adjacent positions do not completely meet any of the above conditions, it is marked as needing further verification, and the deviation feature data of the current and adjacent positions are pushed to the background; the judgment result is recorded in the metadata of the current basic data packet.

9. The automatic identification method for track defects based on image recognition according to claim 1, characterized in that: S4 specifically includes: S4-1: Based on the corrected spatial correspondence, the pixel coordinates of each pixel in the 2D texture image are mapped to the 3D point cloud spatial coordinate system, and an association mapping table between 2D pixels and 3D spatial points is established. S4-2: Based on the association mapping table, the texture features of each pixel in the 2D texture image are concatenated with the geometric features of the corresponding spatial point in the 3D point cloud to form a fused feature vector for each pixel. The fused feature vectors of all pixels constitute a fused feature set. The texture features include the gray value, local gradient direction, and magnitude distribution of the pixel. The geometric features include the three-dimensional spatial coordinates, local normal vector, and local curvature of the spatial point. The feature concatenation is to concatenate the texture feature vector of each pixel with the geometric feature vector of the corresponding 3D spatial point according to the dimension to form a joint feature vector. S4-3: Input the fused feature set Ff into the pre-trained disease identification model; the disease identification model is a pixel semantic segmentation model built on a deep convolutional neural network. It takes the fused feature map composed of the fused feature vectors of each pixel arranged in spatial position as input, performs forward propagation calculation on the fused feature vectors of each pixel, outputs the probability distribution of each type of disease corresponding to each pixel, and takes the category with the highest probability as the disease type determination result of the corresponding pixel; at the same time, the spatial coordinates of the corresponding pixel in the 3D point cloud spatial coordinate system are used as the spatial position of the disease; all pixel positions determined to have diseases and their corresponding disease types and spatial positions are summarized to form a disease identification result list. S4-4: Sort and integrate the disease identification results list according to the track mileage order, merge the judgment results that are adjacent in spatial location and have the same disease type, and use the boundary of the merged area as the actual spatial range of the corresponding disease. Finally, output each disease type and its corresponding spatial location and spatial range.

10. An automatic track defect identification system based on image recognition, characterized in that: The system includes a data acquisition module, an offset analysis module, a correction processing module, and a defect identification module; The data acquisition module synchronously acquires 2D texture images and 3D point cloud images of the same spatial location on the track surface, synchronously acquires encoder odometer signals to establish a spatial association index, maps the pixel coordinates of the 2D texture images to the 3D point cloud spatial coordinate system to generate an initial spatial correspondence, and integrates the 2D texture images, 3D point cloud images and odometer signals to form a basic data package corresponding to the spatial location. The offset analysis module detects and locates track-fixed reference objects in the 2D texture image and 3D point cloud image of the basic data package, calculates the individual spatial offset of each reference object between the 2D mapped position and the 3D actual position according to the initial spatial correspondence, and constructs an offset set based on the individual spatial offsets of all reference objects; identifies outlier reference objects in the offset set and counts the proportion of outlier reference objects, and extracts the overall offset features based on the offsets of non-outlier reference objects. The correction processing module determines the type of deviation source based on the proportion of outlier reference objects, including global spatial deviation, local deformation deviation, and registration benchmark failure; and performs correction or verification processing on the initial spatial correspondence based on the type of deviation source. The disease identification module establishes a correlation mapping between 2D pixels and 3D point clouds based on the corrected spatial correspondence. After extracting and fusing 2D texture features and 3D geometric features, it inputs them into the trained disease identification model to automatically identify the type of track disease and its corresponding spatial location.

Images