A method for quickly detecting geometric size of track slab based on binocular camera

Abstract

The present application relates to a kind of track board geometry fast detection method based on binocular camera, comprising: using binocular camera to collect track board point cloud data, local geometric center point is obtained by carrying out voxel filtering processing to point cloud data, and the position of rail platform is detected using geometric feature;Construct GeoTransformer network, introduce geometric self-attention mechanism and cross attention mechanism, fuse the distance and angle information between point pair, generate hybrid feature, based on the feature, construct Gaussian correlation matrix, and combine double normalization constraint to inhibit fuzzy matching, realize coarse registration by RANSAC algorithm, with the initial value of coarse registration result, carry out fine registration using ICP algorithm, output final transformation matrix, obtain complete track board point cloud model;Finally, based on the track board point cloud model obtained, extract track board length, width, height and other geometric dimensions, and compare with standard value, realize track board geometry fast detection.

Description

Technical Field

[0001] This invention relates to the field of rail transit infrastructure inspection technology, and in particular to a rapid detection method for track slab geometry based on a binocular camera. Background Technology

[0002] As the fundamental structure of the track system, the geometric state of the track slab directly affects the stability and safety of train operation. Traditional track slab measurement methods mainly rely on manual operation and tools such as total stations and vernier calipers, which suffer from low detection efficiency, limited accuracy, and susceptibility to human interference. Point cloud technology can effectively characterize the geometric features and spatial position of the target, avoiding errors caused by contact measurement. However, track slabs have a long strip-like structure, and the acquisition range of binocular cameras is limited, requiring the stitching of multiple point cloud acquisitions to obtain a complete model. In addition, track slab features are similar and sparse, making complete modeling difficult. Traditional registration methods such as ICP and its variants are sensitive to initial pose and are prone to getting trapped in local optima. In scenarios with similar structures, repetitive textures, and low overlap, the registration accuracy and robustness decrease significantly. Deep learning-based point cloud registration methods (such as PointNetLK and PCRNet) improve matching performance through end-to-end feature learning, but their feature extraction modules often ignore local geometric relationships, and still suffer from mismatches, high computational complexity, and insufficient generalization ability under complex conditions such as noise interference, low overlap, and repetitive textures. To address this, the present invention proposes a point cloud registration method that possesses geometric structure perception capabilities, is invariant to rigid transformations, and is applicable to complex track slab scenarios. Furthermore, it rapidly obtains the geometric dimensions of the track slab based on the registered complete point cloud. Summary of the Invention

[0003] The purpose of this invention is to propose a rapid detection method for the geometric dimensions of track slabs based on a binocular camera. This method solves the problem that current point clouds with low overlap and many repetitive textures in track slabs have insufficient feature extraction and difficulty in balancing registration accuracy and efficiency during the stitching process. It achieves rapid stitching of track slab point clouds and rapid extraction of geometric dimensions.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] A method for rapid detection of track slab geometry based on a binocular camera includes:

[0006] Step S1: Use a binocular camera to collect point cloud data of the track slab in a preset manner;

[0007] Step S2: Filter the point cloud data of the track slab and detect the track support area using geometric features;

[0008] Step S3: Based on the local center point coordinates output after filtering, map them to a high-dimensional feature space to construct an initial feature vector;

[0009] Step S4: Construct a GeoTransformer model. The GeoTransformer model encodes and embeds the initial feature vector using a geometric self-attention mechanism and a cross-attention mechanism to extract highly discriminative hybrid features.

[0010] Step S5: Construct a Gaussian correlation matrix based on the hybrid features, use double normalization constraints to suppress fuzzy matching, and combine the RANSAC algorithm to estimate the initial transformation matrix;

[0011] Step S6: Using the initial transformation matrix as the initial value, perform fine registration using the ICP algorithm to obtain a complete track slab point cloud model;

[0012] Step S7: Based on the point cloud model of the track slab, extract the geometric dimensions of the track slab and compare them with the standard values ​​to complete the rapid detection of the geometric dimensions of the track slab.

[0013] Optionally, the preset method in step S1 includes: collecting samples at equal intervals along the longitudinal direction of the track plate several times, and ensuring that adjacent parts have a preset overlapping area.

[0014] Optionally, the filtering process for the track slab point cloud data in step S2 includes:

[0015] Outliers generated during the acquisition of point clouds on the track slab were removed using statistical filtering methods.

[0016] Based on the point cloud after removing outliers, voxel filtering is used for downsampling to construct a uniform voxel grid. The mean coordinates of all points within each voxel are calculated to obtain the local geometric center point representing the local features of the point cloud.

[0017] Optionally, the detection of the rail support platform area based on a three-layer constraint method of elevation slicing, local flatness, and lateral constraints includes:

[0018] Elevation slice constraint: The filtered point cloud data is divided into n equal-thickness horizontal slices along the vertical direction with equal step size. The thickness of each slice is adaptively determined by the global elevation range.

[0019] Local flatness constraint: For the point set p within the k-th slice k, The elevation standard deviation is calculated as a local flatness index. When the elevation standard deviation is less than the preset value, the k-th slice is determined to meet the geometric prior conditions of the rail support surface.

[0020] Lateral constraint: Calculate the circumscribed rectangle size of the slice in the horizontal plane that meets the geometric prior conditions, wherein the circumscribed rectangle size must match the actual width range of the support platform;

[0021] Slices that simultaneously satisfy elevation slice constraints, local flatness constraints, and lateral constraints are marked as rail support area.

[0022] Optionally, in step S3, the local center point coordinates output after filtering are mapped to a 256-dimensional initial feature vector through three fully connected layers with hidden dimensions of 64, 128, and 256 respectively, and each layer is followed by a BatchNorm and ReLU activation function.

[0023] Optionally, step S4 includes:

[0024] The geometry embedding module in GeoTransformer is constructed, and the geometric attention mechanism is used to establish the geometric relationship between points within a single point cloud. The Euclidean distance and the included angle of the triplet between local center points are encoded by the sinusoidal position function to generate a geometrically embedded structure that is invariant to rigid transformation, and the corresponding self-attention matrix is ​​output.

[0025] Based on the self-attention matrix, the geometric consistency across point clouds is learned using the cross-attention mechanism, the cross-attention matrix is ​​output, and a correspondence is established between the source point cloud and the target point cloud.

[0026] The self-attention matrix and the cross-attention matrix are fused by residuals and then fed forward to output a hybrid feature with high discriminative power.

[0027] Optionally, generating a rigid transformation-invariant geometric embedding includes:

[0028] Pairwise Euclidean distance embedding: by calculating two local centers p i and p j The Euclidean distance between them is then mapped to a fixed-dimensional distance embedding vector using a sinusoidal position encoding function;

[0029] Triplet angle embedding: with any local center point p i Centered on the target, select the k nearest neighbors in its neighborhood to form a triplet, calculate the included angle, and encode the angle value into an angle embedding vector.

[0030] Euclidean distance embedding and triplet angle embedding are jointly encoded as geometric structure embedding.

[0031] Optionally, step S5 includes:

[0032] Step S5.1. Based on the hybrid features, calculate the similarity between the two feature sets using the Gaussian kernel function, construct the Gaussian correlation matrix, and filter high-confidence initial matching point pairs through row and column normalization as input to the RANSAC algorithm;

[0033] Step S5.2. Randomly select 3 point pairs, estimate the rigid transformation matrix through singular value decomposition, and use the rigid transformation matrix to transform the source point cloud;

[0034] Step S5.3. Calculate the Euclidean distance between the transformed point pairs. If the distance is less than the set distance threshold, it is considered an interior point; otherwise, it is an exterior point. Count the number of interior points. If the number of interior points is insufficient, the model estimation is considered inaccurate, and you need to return to step S5.2. Here, the model refers to the model fitted using interior points.

[0035] Step S5.4. Iteratively execute the above process, dynamically update the transformation model, until the maximum number of iterations or the interior point proportion converges, and output the final initial transformation matrix.

[0036] Optionally, step S6 includes:

[0037] The initial transformation matrix is ​​used as the starting parameter of the ICP algorithm, applied to the source point cloud of the track board, and the nearest corresponding point is found in the target point cloud.

[0038] Construct the least-squares optimization objective function for the corresponding points, calculate the corresponding transformation matrix through singular value decomposition, and update the pose of the source point cloud.

[0039] The algorithm iteratively executes the corresponding point search-transformation solution-point cloud update process until the distance between the point clouds of two adjacent iterations is less than a set threshold, at which point the algorithm is considered to have converged.

[0040] The rotation and translation matrices from the last iteration are output as the final result, thus completing the registration between the source and target point clouds.

[0041] Optionally, in step S7, the minimum circumscribed hexahedron of the track slab point cloud is constructed using the bounding box method, its geometric dimensions are quickly extracted, and compared with standard values ​​to achieve rapid measurement of the track slab's geometric dimensions.

[0042] The beneficial effects of this invention are as follows:

[0043] This invention provides a fast geometric dimension detection method for track slab point clouds based on a binocular camera. The method uses a binocular camera to acquire track slab point cloud data, performs voxel filtering on the original point cloud to obtain several geometric center points, and uses these geometric features to detect the track support platform. Subsequently, a GeoTransformer network is constructed, incorporating geometric self-attention and cross-attention mechanisms to fuse distance and angle information between point pairs, generating a highly discriminative hybrid feature with rigid transformation invariance. Based on this feature, a Gaussian correlation matrix is ​​constructed, and fuzzy matching is suppressed using dual normalization constraints. The RANSAC algorithm is then used for coarse registration. Finally, using the coarse registration result as initial values, the ICP algorithm is used for fine registration, outputting the final transformation matrix to complete the track slab point cloud. This method, by introducing a GeoTransformer to explicitly encode the distance and angle relationships between points, significantly improves the matching discriminative power in regions with repetitive textures. The combination of a coarse-fine two-level registration strategy effectively avoids ICP getting trapped in local optima, improves registration accuracy under low overlap rates, enhances model controllability, and has relatively low memory requirements. This method enables high-precision and robust stitching of point clouds for low-overlap, weak-texture track slabs, and has significant engineering application value. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic flowchart of a method for rapid detection of track slab geometry based on a binocular camera, according to an embodiment of the present invention.

[0046] Figure 2 This is a schematic diagram of the GeoTransformer feature extraction process according to an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram illustrating the calculation of relative differences in RANSAC according to an embodiment of the present invention;

[0048] Figure 4 This is a point cloud stitching diagram of two adjacent module track slabs according to an embodiment of the present invention;

[0049] Figure 5 This is the final stitched image of the point cloud of the track board according to an embodiment of the present invention. Detailed Implementation

[0050] 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.

[0051] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] like Figure 1 As shown, this embodiment proposes a rapid detection method for track slab geometry based on a binocular camera, including:

[0053] Step S1: Use a binocular camera to collect point cloud data of the track slab in a preset manner;

[0054] Step S2: Filter the point cloud data of the track slab and use geometric features to detect the track support area;

[0055] Step S3: Based on the local center point coordinates output after filtering, map them to a high-dimensional feature space to construct an initial feature vector;

[0056] Step S4: Construct a GeoTransformer model. The GeoTransformer model encodes and embeds the initial feature vector using geometric self-attention and cross-attention mechanisms to extract highly discriminative hybrid features.

[0057] Step S5: Construct a Gaussian correlation matrix based on the hybrid features, use double normalization constraints to suppress fuzzy matching, and combine the RANSAC algorithm to estimate the initial transformation matrix;

[0058] Step S6: Using the initial transformation matrix as the initial value, perform fine registration using the ICP algorithm, output the final transformation parameters, and obtain the complete track slab point cloud model;

[0059] Step S7: Based on the point cloud model of the track slab, extract the geometric dimensions of the track slab and compare them with the standard values ​​to complete the rapid detection of the geometric dimensions of the track slab.

[0060] Specifically, in step S1 of this embodiment, a binocular camera is used to collect data five times at equal intervals along the longitudinal direction of the track slab, ensuring that adjacent parts have a 50% overlap area, so as to obtain point cloud data of each part of the track slab.

[0061] Specifically, in step S2 of this embodiment, based on the obtained track slab point cloud data, outliers generated during the track slab point cloud acquisition process are removed using statistical filtering. Subsequently, a uniform voxel grid is constructed, and the mean coordinate of all points within each voxel is calculated to obtain the local geometric center point representing the local features of the point cloud. Finally, the track support area is detected based on a three-layer constraint method of elevation slicing, local flatness, and lateral constraints, providing an effective target area for the registration step.

[0062] In this embodiment, the statistical filtering process for the track slab point cloud is specifically as follows:

[0063] Statistical filtering is used to remove outliers generated during the acquisition of the track slab point cloud. Outliers are sparsely distributed in space. Therefore, by calculating the average distance from each point to its nearest n points, the distances of all points in the point cloud will eventually form a Gaussian distribution.

[0064] For point clouds , where any point p i Define the set of points within its neighborhood as ,in Let be the Euclidean distance to the nth nearest neighbor.

[0065] Then the mean neighborhood distance:

[0066]

[0067] The standard deviation of the neighborhood distance is:

[0068]

[0069] If point p i satisfy ,in , Here, represents the mean and standard deviation of the global distance distribution, and α is a threshold value representing a multiple of the standard deviation (usually taken as...). In this embodiment, point 2) is considered an outlier.

[0070] In this embodiment, the voxel filtering process for the track slab point cloud is specifically as follows:

[0071] Local geometric center point S under voxel k From the following formula, we get

[0072]

[0073] In the formula, i is point p i coordinates Let k be the kth voxel.

[0074] Points after statistical filtering and voxel sampling are called local geometric center points. The set of geometric center points formed by all representative points of the source and target point clouds under voxel filtering is denoted as . and .

[0075] In this embodiment, detecting the position of the rail support platform using geometric features specifically involves:

[0076] S2.1. Layered slicing: Divide the vertical direction (Z-axis) into n equal-thickness horizontal slices with equal step size ∆Z, where the thickness ∆Z of each slice is adaptively determined by the global elevation range;

[0077]

[0078] S2.2. Local flatness determination, for the point set p within the k-th slice k, Calculate its elevation standard deviation σ z As a local flatness index, it represents the degree of dispersion of the point cloud in the vertical direction; when σ z When the elevation change is less than 2cm, it indicates that the elevation change within the layer is minimal, which meets the geometric a priori conditions of the track support surface.

[0079] S2.3. Lateral Spread Constraints: Calculate the dimensions Δx and Δy of the bounding rectangle of the slice in the horizontal plane to ensure that the candidate region has sufficient lateral span and avoid false detection of narrow strip noise points. The experiment sets Δx≥0.10m and Δy≥0.10m to match the representation of the actual track platform width (0.5–0.7m) in the point cloud.

[0080] Only when all three constraints are met is the slice marked as a candidate rail support platform, and the corresponding point set C is assigned. k Inclusion in the candidate set C = {C1, C2, ..., C} k The final output C is the effective target area for subsequent key point extraction and registration steps.

[0081] Specifically, in step S3 of this embodiment, the local center point coordinates output after filtering are mapped to a 256-dimensional initial feature vector through three fully connected layers with hidden dimensions of 64, 128, and 256 respectively. BatchNorm and ReLU activation functions are then applied to each layer to enhance nonlinear expressive power and accelerate model convergence.

[0082] Specifically, step S4 in this embodiment includes:

[0083] The geometry embedding module in GeoTransformer is constructed, and the geometric attention mechanism is used to establish the geometric relationship between points within a single point cloud. The Euclidean distance and the included angle of the triplet between local center points are encoded by the sinusoidal position function to generate a geometrically embedded structure that is invariant to rigid transformation, and the corresponding self-attention matrix is ​​output.

[0084] Based on the self-attention matrix, the geometric consistency across point clouds is learned using the cross-attention mechanism, and the cross-attention matrix is ​​output to establish a correspondence between the source point cloud and the target point cloud.

[0085] The self-attention matrix and cross-attention matrix are fused by residuals and then fed forward to output a hybrid feature with high discriminative power.

[0086] Figure 2 The present invention provides a flowchart of the Geotransformer feature extraction process.

[0087] In this embodiment, the specific method for generating rigid transformation-invariant geometric structure embeddings using the geometric attention mechanism is as follows:

[0088] Pairwise Euclidean distance embedding, by calculating two local centers p i and p j Euclidean distance between Subsequently, it is mapped to a fixed-dimensional distance embedding vector using a sinusoidal position encoding function. .

[0089]

[0090] In the formula, σ d To control the sensitivity to changes in distance.

[0091] Triple angle embedding, with arbitrary local center point p i Centered on a given point, select a triplet consisting of the k nearest neighbors within its neighborhood. Calculate the included angle. The angle value is encoded into an angle embedding vector.

[0092]

[0093] In the formula, It is the sensitivity to changes in angle.

[0094] Finally, the Euclidean distance embedding and the triplet angle embedding are jointly encoded as a geometric structure embedding.

[0095]

[0096] Specifically, in this embodiment, generating the cross-attention matrix using the cross-attention mechanism is as follows:

[0097] The cross-attention mechanism focuses on the feature interactions between the source and target point clouds of the track slab, establishing a bidirectional correspondence between the two point clouds. For any local geometric center points p and q in the source and target point clouds, the corresponding geometric self-attention matrix Y is input. p and Y q The cross-attention matrix of the output point p is denoted as... .

[0098]

[0099] Specifically, step S5 in this embodiment includes:

[0100] S5.1. Based on the hybrid features, the similarity between the two feature sets is calculated using the Gaussian kernel function, a Gaussian correlation matrix is ​​constructed, and high-confidence initial matching point pairs are selected through row and column normalization as input to the RANSAC algorithm;

[0101] S5.2. Randomly select a set of 3 point pairs, estimate the rigid transformation matrix through singular value decomposition, and use the rigid transformation matrix to transform the source point cloud;

[0102] S5.3. Calculate the Euclidean distance between transformed point pairs. If the distance is less than the set distance threshold, it is considered an interior point; otherwise, it is an exterior point. Count the number of interior points. If the number of interior points is insufficient, the model estimation is considered inaccurate, and it is necessary to return to S5.2. Here, the model refers to the model fitted using interior points.

[0103] S5.4. Iteratively execute the above process, dynamically update the transformation model, until the maximum number of iterations or the interior point proportion converges, and output the final initial transformation matrix.

[0104] In step S5.1 of this embodiment, the correlation matrix is ​​calculated using the Gaussian kernel function. Specifically:

[0105]

[0106] In the formula, S i,j This represents the similarity between the local geometric center of the i-th source point cloud and the geometric center of the j-th target point cloud. The value is between 0 and 1, with the closer to 1 indicating a higher similarity between the two centers. It represents the Euclidean distance between two local geometric centers.

[0107] In step S5.2 of this embodiment, the specific process of randomly selecting a set of 3 point pairs and estimating the rigid transformation matrix through singular value decomposition is as follows:

[0108] S5.2.1. Corresponding point confirmation: Randomly select three geometric center points from the source point cloud P. Using a KD-tree, a nearest neighbor search is performed on the target point cloud Q to obtain the corresponding points. To initially suppress mismatches, the Gaussian matrices of the feature encoding and position encoding parts are weighted and calculated in subsequent processes.

[0109] S5.2.2. Difference vectors for corresponding side lengths, such as Figure 3 As shown, for the three point pairs selected in S5-2-1, calculate the lengths of the three sides in the two point sets respectively:

[0110] Side length of the triangle in the source point cloud:

[0111] , , ,

[0112] Similarly, the side length of the triangle in the target point cloud is

[0113] , , .

[0114] Then, the relative difference vector is calculated based on this.

[0115]

[0116] when If the condition is met, the triplet is retained; otherwise, return to S5-2-1 to select a new point.

[0117] S5.2.3. Assuming model estimation, for the corresponding geometric center point pairs that meet the requirements, the corresponding transformation matrix Ti is obtained through singular value decomposition.

[0118] In step S5.4 of this embodiment, the number of iterations k of RANSAC is continuously updated and changed when it is less than the maximum number of iterations. The calculation formula is as follows:

[0119]

[0120] In the formula, p is the assumed success rate, ω represents the proportion of inliers, and m represents the minimum number of samples used in the model calculation, which is taken as 3 here.

[0121] Specifically, in this embodiment, step S6, fine registration, includes:

[0122] S6.1. Use the initial transformation matrix as the starting parameter of the ICP algorithm, apply it to the source point cloud of the track board, and find the nearest corresponding point in the target point cloud;

[0123] S6.2. Construct the least squares optimization objective function for the corresponding points, calculate the corresponding transformation matrix through singular value decomposition, and update the pose of the source point cloud;

[0124] S6.3. Iteratively execute the corresponding point search-transform solution-point cloud update process until the distance between two adjacent iterations of the point cloud is less than the set threshold, and determine that the algorithm has converged;

[0125] S6.4. Output the rotation and translation matrices from the last iteration as the final result, and then complete the registration between the source point cloud and the target point cloud.

[0126] Specifically, in step S6-1 of this embodiment, for each point (p) in the source point cloud after coarse registration... i Find the nearest neighbor (q) in the target point cloud (Q). j This minimizes the Euclidean distance between two points, thus establishing a geometric set of corresponding point pairs, denoted as C = {p i, q j )}.

[0127] In step S6-2 of this embodiment, the specific relationship between determining the minimum objective function and updating the point cloud pose is as follows:

[0128] The optimal rotation matrix R and translation matrix t are obtained by minimizing the objective function.

[0129] Wherein, the objective function is minimized:

[0130]

[0131] Applying the optimal rotation and translation matrices obtained from the transformation estimate to the source point cloud yields the transformed point cloud. The specific transformation relationship is as follows:

[0132]

[0133] In step S6-4 of this embodiment, the rotation and translation matrices from the last iteration are output as the final result, thereby completing the registration between the source point cloud and the target point cloud. Figure 4 This is a stitched image showing the effect of connecting the point clouds of two adjacent track slabs. Figure 5 This is the final point cloud mosaic image of the track slab.

[0134] Under the same conditions, the algorithm proposed in this invention is compared with the traditional ICP algorithm, Predator algorithm and PointNetLk algorithm to evaluate the performance of each algorithm. The specific results are shown in Table 1.

[0135] Table 1

[0136] Algorithm Name Root mean square error / m Registration time / s Registration success rate ICP 0.0156 12.3 85.2 Predator 0.0124 28.7 89.1 PointNetLk 0.0098 35.2 91.3 This paper's algorithm 0.0089 22.1 94.8

[0137] In step S7 of this embodiment, based on the obtained track slab point cloud model, a minimum circumscribed hexahedron is formed outside the track slab point cloud by using the bounding box method to quickly extract the geometric dimensions of the track slab point cloud and compare them with the standard values. The specific results of the comparison with the standard values ​​are shown in Table 2.

[0138] Table 2

[0139] Testing items Extraction size / m Standard size / m Difference track slab length 3.500 3.500 0.000 track slab width 2.198 2.200 -0.002 Track slab height 0.344 0.345 -0.001

[0140] This invention has been described in detail through specific embodiments, illustrating the principles and implementation methods of the invention. These embodiments are merely illustrative of the core ideas and methods of the invention and do not constitute a limitation thereof. Furthermore, those skilled in the art, upon grasping the basic ideas of this invention, may make appropriate adjustments to the implementation methods and application scope according to specific circumstances. Therefore, the content of this specification should not be considered a limitation on the implementation of this invention.

[0141] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for rapid detection of track slab geometry based on a binocular camera, characterized in that, include: Step S1: Use a binocular camera to collect point cloud data of the track slab in a preset manner; Step S2: Filter the point cloud data of the track slab and detect the track support area using geometric features; Step S3: Based on the local center point coordinates output after filtering, map them to a high-dimensional feature space to construct an initial feature vector; Step S4: Construct a GeoTransformer model. The GeoTransformer model encodes and embeds the initial feature vector using a geometric self-attention mechanism and a cross-attention mechanism to extract highly discriminative hybrid features. Step S5: Construct a Gaussian correlation matrix based on the hybrid features, use double normalization constraints to suppress fuzzy matching, and combine the RANSAC algorithm to estimate the initial transformation matrix; Step S6: Using the initial transformation matrix as the initial value, perform fine registration using the ICP algorithm to obtain a complete track slab point cloud model; Step S7: Based on the point cloud model of the track slab, extract the geometric dimensions of the track slab and compare them with the standard values ​​to complete the rapid detection of the geometric dimensions of the track slab. The preset method in step S1 includes: collecting data at equal intervals along the longitudinal direction of the track plate several times, and ensuring that adjacent parts have a preset overlapping area; The detection of the rail support platform area based on a three-layer constraint method of elevation slicing, local flatness, and lateral constraints includes: Elevation slice constraint: The filtered point cloud data is divided into n equal-thickness horizontal slices along the vertical direction with equal step size. The thickness of each slice is adaptively determined by the global elevation range. Local flatness constraint: For the point set p within the k-th slice k The elevation standard deviation is calculated as a local flatness index. When the elevation standard deviation is less than the preset value, the k-th slice is determined to meet the geometric prior conditions of the rail bearing platform surface. Lateral constraint: Calculate the circumscribed rectangle size of the slice in the horizontal plane that meets the geometric prior conditions, wherein the circumscribed rectangle size must match the actual width range of the support platform; Slices that simultaneously satisfy elevation slice constraints, local flatness constraints, and lateral constraints are marked as rail support platform areas. In step S3, the local center point coordinates output after filtering are mapped to a 256-dimensional initial feature vector through three fully connected layers with hidden dimensions of 64, 128, and 256 respectively, and BatchNorm and ReLU activation functions are applied after each layer. Step S4 includes: The geometry embedding module in GeoTransformer is constructed, and the geometric attention mechanism is used to establish the geometric relationship between points within a single point cloud. The Euclidean distance and the included angle of the triplet between local center points are encoded by the sinusoidal position function to generate a geometrically embedded structure that is invariant to rigid transformation, and the corresponding self-attention matrix is ​​output. Based on the self-attention matrix, the geometric consistency across point clouds is learned using the cross-attention mechanism, the cross-attention matrix is ​​output, and a correspondence is established between the source point cloud and the target point cloud. The self-attention matrix and the cross-attention matrix are fused by residuals and then fed forward to output a hybrid feature with high discriminative power. Generating rigid transformation invariant geometric embeddings includes: Pairwise Euclidean distance embedding: by calculating two local centers p i and p j The Euclidean distance between them is then mapped to a fixed-dimensional distance embedding vector using a sinusoidal position encoding function; Triplet angle embedding: with any local center point p i Centered on, select its domain k The angle between the three nearest neighbors is calculated, and the angle value is encoded into an angle embedding vector. Euclidean distance embedding and triplet angle embedding are jointly encoded as geometric structure embedding; Step S5 includes: Step S5.

1. Based on the hybrid features, calculate the similarity between the two feature sets using the Gaussian kernel function, construct the Gaussian correlation matrix, and filter high-confidence initial matching point pairs through row and column normalization as input to the RANSAC algorithm; Step S5.

2. Randomly select 3 point pairs, estimate the rigid transformation matrix through singular value decomposition, and use the rigid transformation matrix to transform the source point cloud; Step S5.

3. Calculate the Euclidean distance between the transformed point pairs. If the distance is less than the set distance threshold, it is considered an interior point; otherwise, it is an exterior point. Count the number of interior points. If the number of interior points is insufficient, the model estimation is considered inaccurate, and the process should return to step S5.

2. Here, "model" refers to the model fitted using interior points. Step S5.

4. Iteratively execute the above process, dynamically update the transformation model, until the maximum number of iterations or the interior point proportion converges, and output the final initial transformation matrix.

2. The method for rapid detection of track slab geometry based on a binocular camera according to claim 1, characterized in that, Step S2, which involves filtering the point cloud data of the track slab, includes: Outliers generated during the acquisition of point clouds on the track slab were removed using statistical filtering methods. Based on the point cloud after removing outliers, voxel filtering is used for downsampling to construct a uniform voxel grid. The mean coordinates of all points within each voxel are calculated to obtain the local geometric center point representing the local features of the point cloud.

3. The method for rapid detection of track slab geometry based on a binocular camera according to claim 1, characterized in that, Step S6 includes: The initial transformation matrix is ​​used as the starting parameter of the ICP algorithm, applied to the source point cloud of the track board, and the nearest corresponding point is found in the target point cloud. Construct the least squares optimization objective function for the corresponding points, calculate the corresponding transformation matrix through singular value decomposition, and update the pose of the source point cloud. The algorithm iteratively executes the corresponding point search-transformation solution-point cloud update process until the distance between the point clouds of two adjacent iterations is less than a set threshold, at which point the algorithm is considered to have converged. The rotation and translation matrices from the last iteration are output as the final result, thus completing the registration between the source and target point clouds.

4. The method for rapid detection of track slab geometry based on a binocular camera according to claim 1, characterized in that, In step S7, the minimum circumscribed hexahedron of the track slab point cloud is constructed using the bounding box method, its geometric dimensions are quickly extracted, and compared with the standard values ​​to achieve rapid measurement of the track slab's geometric dimensions.

Images