Railway sleeper spacing automatic calculation and anomaly detection method and system based on three-dimensional point cloud

Abstract

A method and system for automatic calculation and anomaly detection of sleeper spacing based on 3D point clouds are disclosed. This method acquires high-precision point cloud data of railway tracks using 3D laser or mobile scanning equipment. It employs a deep learning-driven structure-guided point cloud segmentation network to achieve robust automatic extraction of sleeper point clouds in complex environments. For different structural sections such as ordinary lines, bridges, and tunnels, the system integrates a parameter-self-tuning DBSCAN clustering algorithm to dynamically adapt to the point cloud density and spatial distribution characteristics of the section, automatically segmenting single sleeper point cloud clusters and further subdividing span-abnormal clusters to ensure clustering accuracy. The system supports spacing statistics and anomaly detection for the entire line or specific sections. Based on the mean, standard deviation, and distribution range, it automatically identifies abnormal sleeper spacing and outputs anomaly distribution maps, anomaly lists, and automatic detection reports. This invention can provide reliable data support and technical assurance for multi-scenario, batch-based intelligent inspection and maintenance management of railway tracks.

Description

Technical Field

[0001] This invention relates to the field of railway track inspection and intelligent maintenance technology, specifically to a method and system for automatic calculation of sleeper spacing and anomaly detection based on three-dimensional point clouds, which belongs to the related technologies of digital inspection and intelligent operation and maintenance of railway infrastructure. Background Technology

[0002] In recent years, with the rapid development of the railway transportation industry and the continuous extension of operating lines, the health status of railway infrastructure has received increasing attention. As a key load-bearing and positioning component of the track structure, the arrangement and spacing of sleepers have a significant impact on the overall stability of the track structure, train operation safety, and smoothness. Abnormal sleeper spacing is often related to track settlement, structural deformation, or construction deviations. If these are not detected and corrected in a timely manner, they can easily lead to early damage to the track structure and even affect train operation safety. Therefore, conducting efficient and accurate detection and anomaly early warning of sleeper spacing is an important foundation for the digitalization and intelligentization of railway operation and maintenance.

[0003] Currently, the detection of sleeper spacing on railway sites still mainly relies on traditional methods such as manual spot checks, handheld measuring tools, or two-dimensional image-assisted analysis. These methods are limited by manual operation, long inspection cycles, and incomplete spatial information, making it difficult to meet the needs of modern railways for large-scale, automated, and precise inspection. In recent years, spatial digital sensing technologies such as LiDAR scanning and 3D mobile measurement systems have begun to be applied in the field of railway facility inspection. Through the acquisition of large-scale, high-density point cloud data, three-dimensional reconstruction of track structures and acquisition of spatial features can be achieved, providing basic data support for the automated detection of sleeper spacing.

[0004] However, existing three-dimensional point cloud sleeper detection methods still have the following prominent problems in practical applications: (1) Automatic identification of sleeper point clouds depends on the spatial density and geometric features of the point cloud. Due to stray noise, point cloud occlusion and changes in sleeper shape, false detection, missed detection or insufficient identification accuracy often occur; (2) Spatial clustering and segmentation of sleepers depends on manual experience parameter settings, which is difficult to adapt to the point cloud distribution characteristics of different line types (such as ordinary lines and structural sections such as bridges and tunnels), which can easily lead to segmentation errors such as aggregation of adjacent sleepers or splitting of single sleepers; (3) The calculation of sleeper spacing is often based on two-dimensional projection or simple Euclidean distance, which fails to fully consider the dynamic changes of the dominant direction of track space. Especially in special scenarios such as curves, turnouts and bridges, the accuracy of spacing calculation is difficult to guarantee; (4) There is a lack of efficient spacing anomaly statistics and intelligent early warning mechanism, which makes it difficult to realize automatic anomaly detection and maintenance decision support for a large range of tracks.

[0005] With the development of digital operation and maintenance and intelligent inspection technologies for railways, there is an urgent need for a 3D point cloud analysis method and system that is adaptable to multiple scenarios and structural types, possessing capabilities such as high-precision automatic extraction of sleeper point clouds, adaptive clustering and segmentation, self-tuning of the track's dominant direction, automatic calculation of sleeper spacing across the entire line, and intelligent anomaly detection. This type of method should be able to effectively adapt to different track structures (including ordinary lines, bridges, tunnels, etc.), improve the accuracy, efficiency, and intelligence level of sleeper spacing detection, and provide a solid data foundation and technical support for the full life-cycle maintenance and safety management of railway facilities. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies in automatic sleeper spacing detection, such as low efficiency of manual inspection and limited automation and accuracy, this invention proposes an automatic sleeper spacing calculation and anomaly detection method and system based on three-dimensional point cloud. It can be applied to various track structures such as ordinary lines, bridges, and tunnels, and realize fully automatic high-precision detection of sleeper spacing and intelligent anomaly identification, effectively improving the level of refined inspection and intelligent maintenance of track facilities.

[0007] This invention is achieved through the following technical solution:

[0008] An automatic calculation and anomaly detection method for sleeper spacing based on 3D point cloud, characterized in that the method includes the following steps:

[0009] (1) For the preprocessed point cloud data, a structure-guided point cloud segmentation network based on deep learning is used to accurately extract the sleeper point cloud;

[0010] (2) Based on the spatial distribution characteristics of sleeper point clouds in different sections (ordinary lines, bridges, tunnels, etc.), section adaptive point cloud density analysis is performed, clustering parameters are dynamically adjusted, and the DBSCAN intelligent clustering algorithm with parameter self-tuning is used to automatically cluster the sleeper point clouds into single sleeper point cloud clusters. The clusters with abnormal spans after the initial clustering are automatically subdivided to ensure that the clustering granularity is consistent with the actual sleeper structure.

[0011] (3) For each effective sleeper point cloud cluster, extract its principal axis direction and major and minor axis feature values ​​by principal component analysis (PCA) method, and use the centroid of the point cloud or the midpoint of the principal axis as the sleeper center coordinates to record the number of cluster points, spatial distribution and morphological parameters.

[0012] (4) Based on all valid sleeper center points, the dominant direction of track space is automatically estimated by methods such as covariance matrix eigenvalue decomposition, and segmented dominant direction estimation and correction are supported in sections with structural changes such as curves and bridges, thereby improving the spatial consistency of the dominant direction and the accuracy of spacing calculation.

[0013] (5) Project all sleeper center points according to the dominant direction and automatically sort them according to the projection results to form a spatially ordered sleeper center sequence.

[0014] (6) Based on the sleeper center sequence, the projection spacing of each adjacent sleeper in the dominant direction is automatically calculated to obtain the sleeper spacing sequence;

[0015] (7) Perform statistical analysis on the sleeper spacing sequence, including mean, standard deviation and distribution range, set anomaly detection threshold, automatically identify and mark abnormal spacing, support cluster parameter review and spacing correction suggestions for abnormal sections, and finally output an abnormal sleeper spacing distribution map, anomaly list and detection report.

[0016] Preferably, the semantic segmentation of the sleeper point cloud employs a deep learning model or a structure-guided point cloud segmentation network that combines multi-scale geometric and reflection features, specifically including:

[0017] A-1. Multi-scale geometric features (such as local point density, principal curvature, rate of change of normal vector, etc.) and reflection features are used to automatically extract features and perform multi-scale neighborhood analysis on the normalized original point cloud.

[0018] A-2. Using graph convolutional networks (such as DGCNN / GACNet) and K-nearest neighbor graph structures, sleeper point clouds are accurately extracted, achieving highly robust segmentation under stray noise, occlusion, and structural deformation scenarios. The high-dimensional feature vector of each point includes:

[0019] Preferably, the parameter self-tuning clustering method specifically includes:

[0020] B-1. Estimate the average nearest neighbor distance of the segment point cloud through k-nearest neighbor analysis or minimum distance statistics, and adaptively set the neighborhood radius and minimum number of samples for DBSCAN clustering in combination with the actual spatial scale of the sleepers;

[0021] B-2. To address the differences in point cloud distribution and sleeper spacing between different sections such as bridges and ordinary tracks, clustering parameters are set separately, and a segmented processing strategy is adopted when necessary to achieve adaptive single sleeper extraction in complex scenarios. For the initial clustering results, in order to further improve the accuracy of single sleeper point cloud extraction, the system performs span anomaly detection on each cluster.

[0022] Preferably, the extraction of the sleeper center and main shaft direction attributes includes:

[0023] C-1. Calculate the three-dimensional centroid of each sleeper point cloud cluster as the preliminary center coordinates;

[0024] C-2. Principal component analysis is used to construct a covariance matrix for point cloud clusters. After eigenvalue decomposition, the principal axis direction and the eigenvalues ​​of the major and minor axes are obtained. Combined with geometric parameters such as aspect ratio and number of points, effective sleeper clusters are screened.

[0025] Preferably, the estimation steps for the dominant direction of orbital space are as follows:

[0026] D-1. Summarize all sleeper center points, calculate their mean coordinates segment by segment or along the entire line, and use the covariance matrix eigenvalue decomposition method to extract the eigenvector corresponding to the largest eigenvalue as the dominant direction.

[0027] D-2. When there are abrupt changes in spatial structure (such as curves or bridge sections), the dominant direction is corrected according to the segment to improve the spatial accuracy of spacing calculation.

[0028] Preferably, the automatic calculation of sleeper spacing includes:

[0029] The center point of the sleeper is projected onto the dominant direction and automatically sorted according to the projection value to form an ordered sequence of sleeper centers. The difference in the projection coordinates of two adjacent sleeper centers in the dominant direction is used as the actual spatial spacing of the sleepers to obtain a complete sleeper spacing sequence, which supports section / full-line statistics and personalized processing in special scenarios.

[0030] Preferably, the sleeper spacing anomaly detection includes:

[0031] F-1. Calculate the mean, standard deviation and distribution range of all sleeper spacing data, set the anomaly identification threshold, automatically mark abnormal sleeper spacing pairs and output the abnormal sleeper spacing list;

[0032] F-2. For abnormal spacing segments, it supports automatic triggering of cluster parameter review, segmentation adjustment or suggestion for manual review, and supports cluster recalculation and spacing correction for abnormal segments.

[0033] F-3. Generates a histogram of spacing distribution, anomaly distribution, and an automatic detection and analysis report.

[0034] Preferably, it is applicable to various types of track structures such as ordinary lines, bridges, and tunnels, and ensures the accuracy and robustness of sleeper spacing detection in complex scenarios through adaptive switching of section parameters.

[0035] This invention also discloses an automatic sleeper spacing calculation and anomaly detection system based on three-dimensional point clouds, characterized in that:

[0036] A-1. Point cloud data acquisition and management module, used for 3D point cloud acquisition, storage and spatial registration;

[0037] A-2. Sleeper point cloud semantic segmentation module, which enables automatic extraction of sleeper point clouds;

[0038] A-3. Parameter self-tuning clustering and sleeper cluster extraction module, realizes section adaptive single sleeper segmentation, and supports multi-level clustering and parameter optimization;

[0039] A-4. Sleeper center and dominant direction attribute extraction module, which completes the calculation and screening of sleeper center of gravity and main axis direction;

[0040] A-5. Track space dominant direction estimation and sleeper sorting module, which realizes automatic extraction of segment dominant directions and sleeper center sorting;

[0041] A-6. Automatic Spacing Calculation and Anomaly Detection Module: This module enables the calculation, statistical analysis, and anomaly identification of spacing in the dominant direction.

[0042] A-7. Result visualization and automatic reporting module, which supports 3D sleeper distribution visualization, spacing analysis chart generation, anomaly detection report export, and sleeper sequence management for the entire line.

[0043] Preferably, the system supports automatic switching of parameters for multiple structural sections and can call dedicated clustering and anomaly detection rules for different track types to achieve automatic consistency detection and anomaly identification of sleeper spacing across the entire line.

[0044] Preferably, the result visualization and reporting module supports functions such as 3D sleeper distribution point cloud display, spacing distribution histogram, automatic highlighting of abnormal sleepers, and export of abnormal sections and raw data, which facilitates the analysis and application of detection results.

[0045] Beneficial effects

[0046] 1. It can achieve fully automated and high-precision detection of sleeper spacing in different sections of railway tracks (including ordinary lines, bridges, tunnels, etc.), significantly reducing the need for manual intervention and improving the standardization and objectivity of detection.

[0047] 2. By integrating multi-source 3D point clouds, parameter self-tuning clustering, and dynamic analysis of principal axis direction, it can effectively address issues such as point cloud density variations, complex structures, stray noise, and occlusion interference in real-world scenarios, exhibiting strong detection robustness and high adaptability.

[0048] 3. The detection process can automatically analyze the statistical distribution of sleeper spacing, and use indicators such as mean, standard deviation and distribution range to quickly and intelligently identify and mark abnormal sleepers, providing data support for maintenance personnel to accurately locate abnormal sections and make correction suggestions.

[0049] 4. The entire process supports seamless integration of steps such as data acquisition, preprocessing, sleeper point cloud extraction, clustering and segmentation, attribute calculation, dominant direction estimation, spacing statistics and anomaly detection, which greatly improves detection efficiency and is suitable for rapid batch detection and periodic inspection of a wide range of railway lines.

[0050] 5. The test results support multiple output methods, such as 3D visualization, spacing distribution histogram, abnormal section highlighting, and automatic generation of test reports, which helps maintenance personnel to intuitively analyze, compare, and archive the results.

[0051] 6. It can provide scientific and batch-based basic data for the health monitoring, refined maintenance and condition assessment of railway track infrastructure, significantly improving the level of intelligent operation and maintenance and safety management of railway facilities.

[0052] 7. The method is highly versatile and easy to integrate with existing railway 3D laser / mobile scanning systems. It can be widely applied in practical engineering projects, providing reliable technical support for the rail transit industry. Attached Figure Description

[0053] Figure 1 is a block diagram of the overall system architecture;

[0054] Figure 2 is a schematic diagram of point cloud data acquisition;

[0055] Figure 3 is a schematic diagram of the semantic segmentation principle of sleeper point cloud;

[0056] Figure 4 shows the semantic segmentation results of multi-segment sleepers.

[0057] Figure 5 is a schematic diagram of multi-segment parameter self-tuning clustering;

[0058] Figure 6 is a schematic diagram of the extraction of the main axis direction and center of the sleeper point cloud cluster;

[0059] Figure 7 is a schematic diagram of automatic estimation and segmented correction of the dominant direction of the orbital space;

[0060] Figure 8 is a flowchart of sleeper center point projection sorting and sleeper spacing calculation.

[0061] Figure 9 A visual diagram illustrating the clustering and detection results. Detailed Implementation

[0062] The first aspect of this invention discloses an automatic calculation and anomaly detection method for sleeper spacing based on three-dimensional point clouds, the method comprising the following steps:

[0063] (1) For the preprocessed point cloud data, a structure-guided point cloud segmentation network based on deep learning is used to accurately extract the sleeper point cloud;

[0064] (2) Based on the spatial distribution characteristics of sleeper point clouds in different sections (ordinary lines, bridges, tunnels, etc.), section adaptive point cloud density analysis is performed, clustering parameters are dynamically adjusted, and the DBSCAN intelligent clustering algorithm with parameter self-tuning is used to automatically cluster the sleeper point clouds into single sleeper point cloud clusters. The clusters with abnormal spans after the initial clustering are automatically subdivided to ensure that the clustering granularity is consistent with the actual sleeper structure.

[0065] (3) For each effective sleeper point cloud cluster, extract its principal axis direction and major and minor axis feature values ​​by principal component analysis (PCA) method, and use the centroid of the point cloud or the midpoint of the principal axis as the sleeper center coordinates to record the number of cluster points, spatial distribution and morphological parameters.

[0066] (4) Based on all valid sleeper center points, the dominant direction of track space is automatically estimated by methods such as covariance matrix eigenvalue decomposition, and segmented dominant direction estimation and correction are supported in sections with structural changes such as curves and bridges, thereby improving the spatial consistency of the dominant direction and the accuracy of spacing calculation.

[0067] (5) Project all sleeper center points according to the dominant direction and automatically sort them according to the projection results to form a spatially ordered sleeper center sequence.

[0068] (6) Based on the sleeper center sequence, the projection spacing of each adjacent sleeper in the dominant direction is automatically calculated to obtain the sleeper spacing sequence;

[0069] (7) Perform statistical analysis on the sleeper spacing sequence, including mean, standard deviation and distribution range, set anomaly detection threshold, automatically identify and mark abnormal spacing, support cluster parameter review and spacing correction suggestions for abnormal sections, and finally output an abnormal sleeper spacing distribution map, anomaly list and detection report.

[0070] A. In this embodiment, the semantic segmentation of the sleeper point cloud preferably adopts a structure-guided point cloud segmentation network based on deep learning, and combines a multi-scale geometric feature and reflection feature fusion method to improve the automatic recognition accuracy of sleeper point cloud in complex scenes.

[0071] The specific steps are as follows:

[0072] A-1. Normalize and preprocess the original orbital region point cloud. For each point p i Construct a multi-scale neighborhood and extract the following features:

[0073] A-1.1 Selection of Multi-Scale Radius:

[0074] To fully adapt to point cloud density and spatial structure, the neighborhood radius r i A parameterless adaptive setting based on local point cloud density is adopted, and the specific calculation formula is as follows:

[0075]

[0076] r i For point p i The adaptive neighborhood radius; k is the number of points contained in the target neighborhood; ρ(p i Let p be the local point cloud density. i The number of points within a unit volume centered on the center; π is the ratio of pi to circumference.

[0077] A-1.2 Multi-scale geometric features:

[0078] (1) Local point density

[0079]

[0080] Where, N r (p i ) indicates p i Let r be the set of points in the neighborhood centered at r with radius r.

[0081] (2) Principal curvature (minimum eigenvalue of PCA)

[0082]

[0083] Where, λ min λ is the minimum eigenvalue of the PCA decomposition. k There are three eigenvalues.

[0084] (3) Rate of change of normal vector

[0085]

[0086] Where, n i , n j Points p i , p j The normal vector.

[0087] A-1.3 Reflection characteristics:

[0088] Mean and variance of reflection intensity:

[0089]

[0090]

[0091] The above features are concatenated to form a high-dimensional feature vector for each point.

[0092] A-2. A structure-guided point cloud segmentation network with feature vector input. The network structure adopts a deep network based on graph convolution (such as DGCNN / GACNet) and is constructed using a K-nearest neighbor graph to ensure that the point cloud structure conforms to the physical constraints of the railway sleepers. The sleeper point cloud is obtained through the structure-guided point cloud segmentation network.

[0093] "Structure guidance" refers to explicitly introducing prior knowledge of the physical structure of the sleepers and topological relationship information of the point cloud spatial neighborhood (such as the connectivity of the K-nearest neighbor graph and the expression of spatial structure by EdgeConv / Graph Attention) into the network feature input, neighborhood construction, and attribute learning stages. This structure guidance mechanism enables the network to take into account both local geometric features and global structural consistency during point cloud segmentation, thereby effectively improving the automatic recognition accuracy of sleeper point clouds and the segmentation robustness in complex scenarios.

[0094] A-2.1 takes the 3D coordinates, multi-scale geometry, and reflection intensity of the point cloud as input. It dynamically establishes a neighborhood for each point using K-nearest neighbor search, and aggregates and updates the features between each point and its neighbors using EdgeConv or graph attention mechanisms to recursively extract spatial association information. The entire network consists of four stacked EdgeConv (or Graph Attention) layers, each extracting spatial features at different scales. The output features of each layer are 64, 64, 128, and 256 dimensions, respectively. Subsequently, a fully connected layer (MLP) and global pooling operations are used to fuse global and local features, ultimately outputting the probability that each point belongs to the sleeper or non-sleeper category.

[0095] A-2.2 The training samples are derived from self-collected 3D laser scanning data. The collected data are manually labeled with categories. All training data retain attributes such as spatial coordinates, reflection intensity, and multi-scale geometric features of the points to realistically reflect the spatial distribution and physical characteristics of the site.

[0096] A-2.3 This network uses the cross-entropy loss function as the sole loss during model training to measure the difference between the network output and the true segmentation labels. The cross-entropy loss function is defined as follows:

[0097]

[0098] C represents the number of segmentation categories; y i p represents the true category label for the i-th point; i,c Let be the probability that the network predicts the i-th point as belonging to the c-th class.

[0099] B. In this embodiment, the parameter self-tuning clustering method aims to automatically determine the optimal clustering parameters based on the point cloud density and spatial distribution of different sections, thereby achieving efficient and accurate clustering of sleeper point clouds. This method mainly includes the following steps:

[0100] B-1. First, perform k-nearest neighbor distance analysis or minimum distance statistics on the sleeper point cloud samples within the section to be processed. Specifically, the system randomly selects several sampling points from the section point cloud, and for each sampling point p... iCalculate the Euclidean distance d from the nearest neighbor. i The average nearest neighbor distance of all sampled points is taken as the average nearest neighbor distance for that segment.

[0101] According to the parameter self-tuning strategy of the present invention, the neighborhood radius ε of the DBSCAN clustering algorithm is adaptively set as follows:

[0102]

[0103] Where, ε max As a preset upper limit, a value range of 0.8-1.5 is recommended, but in actual bridge sections, a value of 1.0-1.3 is often used. avg It is the average nearest neighbor distance. The minimum number of samples, min_samples, is also adaptively set based on the total number of points in the current segment's point cloud.

[0104]

[0105] Where, N min b2 is the lower limit, and b2 is an empirical coefficient, with a recommended value range of 0.002~0.01. For bridge sections, a value of 0.004~0.008 is commonly used. N all This represents the total number of points in the segment.

[0106] Through the aforementioned parameter self-tuning mechanism, DBSCAN can automatically adapt to different point cloud densities and structure distributions, effectively avoiding missed segmentation or erroneous merging caused by clustering parameter mismatch.

[0107] B-2. Segment Differentiation Parameter Configuration and Segmentation Strategy

[0108] This invention further addresses the differences in sleeper distribution characteristics and point cloud acquisition density across different structural sections such as bridges, ordinary railway lines, and tunnels by employing segment-adaptive clustering parameter configuration. The system automatically selects the optimal ε and min_samples parameters based on segment attributes (such as section type and sleeper spacing standards). For sections with complex structures or significant scene changes (such as bridgeheads, curves, and transition zones in the track structure), the system supports grouping by segment or clustering by segment separately.

[0109] When necessary, a segmented processing strategy can be adopted. Based on the line design ledger or GNSS positioning information, the entire track can be divided into multiple sections according to fixed mileage (such as every 50 meters, 100 meters or according to the length of construction units). Parameter self-tuning and clustering operations can be performed separately to prevent the mis-merging or extraction omission of local sleeper point clouds due to parameter uniformity, and to achieve adaptive and accurate clustering of sleepers in complex scenarios.

[0110] To further improve the accuracy of single-rail sleeper point cloud extraction based on the initial clustering results, the system performs span anomaly detection for each cluster. The specific criterion is: for each cluster C...j Calculate the maximum Euclidean distance D between all points within the cluster. max (C j ).like

[0111]

[0112] Then the cluster is identified as a span anomalous cluster, where L sleeper The design length of the sleeper is γ, and the anomaly determination coefficient is γ (recommended value 1.3, adjustable range 1.2-1.5). The system will automatically subdivide the clusters determined to have an abnormal span using stricter clustering parameters (such as reducing the DBSCAN neighborhood radius and increasing the minimum number of samples) until the maximum distance between points in all clusters meets the requirements.

[0113] C. In this embodiment, the steps for extracting the sleeper center and main shaft direction attributes are as follows:

[0114] C-1. For each sleeper point cloud cluster that has been clustered, C j The three-dimensional coordinates of all points within the cluster are counted, and the centroid coordinates (i.e., geometric center) are calculated as the preliminary sleeper center of the cluster.

[0115] C-2. To further ensure the spatial accuracy of sleeper point cloud clustering, this invention uses the following alignment criterion to screen the validity of clusters based on the spatial relationship between the principal axis direction and the dominant track direction. Principal component analysis is used to construct the covariance matrix for the point cloud clusters. The formula for calculating the covariance matrix of the point cloud in this region is:

[0116] Where M j For point cloud cluster C j Number of points in G j For point cloud cluster C j The three-dimensional centroid coordinate vector is given, with the superscript T denoteing the transpose of the vector. Eigenvalue decomposition of the covariance matrix yields three eigenvalues ​​λ1 ≥ λ2 ≥ λ3 and their corresponding eigenvectors. The eigenvector corresponding to the largest eigenvalue λ1 is used as the principal axis direction of the sleeper, and the major axis... and minor axis eigenvalues They can be represented as:

[0117]

[0118] Combine the number of points M of the clusters j Geometric parameters such as aspect ratio and principal axis direction are used to screen the validity of all sleeper point cloud clusters.

[0119] • The number of points must be greater than the preset lower limit to filter out small clusters of noise;

[0120] • The aspect ratio should conform to the sleeper structure (e.g., greater than 2).

[0121] For each sleeper point cloud cluster, the principal axis direction vector is obtained through principal component analysis. And take into account its relationship with the dominant direction of the orbit. The included angle θ:

[0122]

[0123] If the included angle θ does not satisfy 95°≥θ≥85°, it is determined to be an invalid sleeper cluster and is automatically removed or marked as abnormal.

[0124] Through the above steps, the spatial center and main axis direction of each sleeper are automatically obtained, laying the foundation for subsequent estimation of the dominant direction, sleeper sorting, and automatic calculation of spacing.

[0125] D. In this embodiment, the estimation of the dominant direction in orbital space includes the following specific steps:

[0126] D-1. Automatic estimation of the dominant direction of the entire line or segments

[0127] All valid sleeper center points (obtained from the centroids of each sleeper point cluster) are spatially aggregated. Based on the line's structural characteristics and actual testing requirements, the dominant direction analysis can be performed either along the entire line or by section (e.g., at regular intervals or by structural zones). For each section, the set of all sleeper center points for that section is denoted as {C}. k First, calculate the mean coordinates (center point).

[0128]

[0129] Where N is the number of center points of the sleepers in this section.

[0130] With C mean Based on this, the centers of each sleeper are decentralized, and a three-dimensional covariance matrix S is constructed. C :

[0131]

[0132] For S C Eigenvalue decomposition is performed to obtain the eigenvector corresponding to the largest eigenvalue, which is the dominant spatial direction of the track in that segment.

[0133] D-2. Correction of the dominant direction in abrupt changes in spatial structure

[0134] When encountering track sections with abrupt changes in spatial structure, such as curves, bridges, or other abrupt bridges, it is considered that the dominant track direction may experience localized changes or bends. In this case, a segmented analysis strategy is adopted to localize the dominant direction estimation window. For example, by selecting several adjacent sleeper center points according to a segment or sliding window, the above-mentioned D-1 covariance eigenvalue decomposition method is repeated to calculate the dominant direction vector for each segment separately.

[0135] For areas where there are significant deviations in the dominant direction between sections (i.e., the angle between adjacent dominant direction vectors is greater than 10°; this threshold can be adjusted within the range of 5° to 20° according to actual engineering needs), further optimization of the dominant direction's continuity or spatial consistency can be achieved using methods such as linear interpolation, weighted smoothing, or manual correction. This ensures the accuracy of the dominant direction estimation in structurally changing sections, thereby improving the spatial accuracy and robustness of subsequent sleeper sequencing and spacing calculations. The above steps automatically output the dominant direction for each section or sliding window, providing a reliable spatial reference for subsequent processing such as sleeper center sequencing, spacing calculation, and anomaly analysis.

[0136] E. In this embodiment, the automatic calculation of sleeper spacing includes the following steps:

[0137] For all valid sleeper center points that have been screened, under the estimated dominant direction, first, each sleeper center point C... k Projecting onto the dominant direction, calculate the projection value S. k :

[0138]

[0139] Where C0 represents the mean coordinates of all sleeper center points in this section. Then, all sleeper center points are projected according to the value S. k The sleeper centers are automatically sorted in ascending order to form a spatially ordered sequence. Based on this sequence, the difference in the projected coordinates of the dominant direction of adjacent sleeper centers is calculated as the actual spatial spacing, specifically:

[0140]

[0141] Among them, D j Let be the distance between the centers of the j-th pair of adjacent sleepers. The system iterates through all pairs of sleeper centers and outputs a complete sleeper spacing sequence. This method not only supports spacing statistics for the entire line, but also allows for flexible setting of spacing analysis windows for specific sections or special structural sections (such as curves, bridges, tunnels, etc.) according to engineering needs, enabling personalized spacing calculation and result output, and improving the spatial adaptability and accuracy of sleeper spacing detection in complex scenarios.

[0142] F. In this embodiment, the abnormal sleeper spacing detection includes the following steps:

[0143] F-1. First, all sleeper spacing data D obtained through calculation in the dominant direction are... j Perform statistical analysis. The system automatically calculates the mean μ, standard deviation σ, and interval distribution of the interval data. Commonly used calculation formulas are as follows:

[0144]

[0145]

[0146] Based on this, an anomaly identification threshold is set (e.g., a spacing greater than 1.5 times the average, or less than 0.5 times). For sleeper spacing exceeding the threshold, the system automatically marks the corresponding sleeper center pair as "abnormal spacing" and outputs the abnormal sleeper spacing number, specific value, and spatial location information to form an abnormal sleeper spacing list, which is convenient for subsequent manual review or automatic processing.

[0147] F-2. For detected abnormal spacing sections, the system supports automatically triggering a clustering parameter review process. Based on the actual distribution of the abnormal points in the section, it can dynamically adjust the neighborhood radius and minimum sample size of clustering algorithms such as DBSCAN to re-segment the sleeper point cloud of that section; or output segmentation adjustment suggestions for manual review to ensure that the clustering granularity is consistent with the actual structure. Simultaneously, the system supports secondary segmentation and spacing correction of the clustering results for abnormal sections to ensure the accuracy and reliability of the final detection results.

[0148] F-3. The detection results can automatically generate visual charts such as spacing distribution histograms and abnormal spacing distribution maps, intuitively displaying the overall distribution, mean, abnormal intervals, and abnormal points of sleeper spacing within the entire line or section. The system can also output automatic detection analysis reports, which include statistical parameters of sleeper spacing for the entire line or section, abnormal detection results, suggested measures, and original detection data, facilitating operation and maintenance management and decision support.

[0149] This embodiment discloses an automatic sleeper spacing calculation and anomaly detection system based on three-dimensional point clouds, including the following functional modules:

[0150] A-1. Point Cloud Data Acquisition and Management Module

[0151] This module is responsible for acquiring 3D point cloud data of the railway line. The system integrates multiple sensors, including a high-precision 3D laser scanner, GNSS positioning, and inertial navigation units, enabling simultaneous acquisition of high-density point cloud data of the track and its surrounding structures. The acquired data is stored, indexed, and uniformly spatially registered through the module's built-in storage and management mechanisms, ensuring that the point cloud data for the entire line has absolute spatial reference coordinates, providing a high-quality data foundation for subsequent analysis and processing.

[0152] A-2. Sleeper point cloud semantic segmentation module

[0153] This module enables automatic semantic segmentation of collected railway track point cloud data. Based on a deep learning-guided point cloud segmentation network that integrates multi-scale geometric and reflection features, it can accurately distinguish and automatically extract sleeper point cloud regions even in complex environments with stray noise, occlusion, and structural variations. The segmentation output is a labeled subset of sleeper point clouds, significantly improving the automation level of sleeper detection.

[0154] A-3. Parameter Self-Tuning Clustering and Sleeper Cluster Extraction Module

[0155] This module, based on the point cloud density and structural characteristics of different track sections, automatically estimates parameters such as the neighborhood radius and minimum sample size for clustering algorithms like DBSCAN through k-nearest neighbor distance analysis or minimum distance statistics, achieving adaptive, segmented single-sleeper clustering and segmentation. For different sections such as bridges and ordinary tracks, the system can flexibly adjust clustering parameters and supports segmented multi-level clustering and parameter optimization. After clustering, it automatically subdivides abnormal large clusters and screens valid sleeper point cloud clusters using multi-dimensional indicators such as span along the principal axis and morphological characteristics.

[0156] A-4. Sleeper Center and Main Axis Direction Attribute Extraction Module

[0157] This module calculates the 3D centroid of each sleeper point cloud cluster as the sleeper center and extracts the principal axis direction, major axis, and minor axis feature values ​​of the point cloud cluster through principal component analysis (PCA). Combining attributes such as aspect ratio, number of points, and the angle between the principal axis direction and the track's principal direction, all sleeper point cloud clusters are screened for validity, retaining only those that conform to the physical sleeper shape, thus providing data support for dominant direction analysis and subsequent processing.

[0158] A-5. Track Spatial Dominant Direction Estimation and Sleeper Sequencing Module

[0159] This module automatically estimates the dominant spatial direction of the track for effective sleeper center points within a section or the entire line using the covariance matrix eigenvalue decomposition method. It also supports the use of local windowing or sliding segmentation strategies to calculate and correct the dominant direction in sections with abrupt structural changes such as curves, bridges, and tunnels. Based on the dominant direction, all sleeper center points are projected and sorted to form a spatially ordered sequence of sleeper center points, providing a spatial reference for spacing calculation.

[0160] A-6. Automatic Spacing Calculation and Anomaly Detection Module

[0161] This module automatically calculates the spacing between adjacent sleeper centers in the dominant direction based on the sorted sleeper center sequence and outputs the complete spacing sequence. The system performs statistical analysis on all spacing data, automatically calculates the mean, standard deviation, and distribution range, and automatically marks and outputs abnormal sleeper spacings based on preset anomaly detection thresholds. For abnormal sections, it can trigger cluster parameter review, secondary segmentation, or manual verification suggestions to ensure the accuracy of anomaly detection and the usability of the project.

[0162] A-7. Results Visualization and Automatic Reporting Module

[0163] This module supports visualization and report export of various inspection results. It includes intuitive graphical outputs such as 3D sleeper distribution point cloud display, spacing distribution histogram, and highlighted abnormal sleepers. It can also automatically generate inspection and analysis reports, covering sleeper spacing statistics, anomaly detection lists, spatial distribution maps, and original point cloud data indexes, providing efficient and convenient technical support for on-site review, maintenance management, and data archiving.

[0164] C. In this embodiment, the system supports automatic switching of parameters for multiple structural sections and can call dedicated clustering and anomaly detection rules for different track types to achieve automatic consistency detection and anomaly identification of sleeper spacing across the entire line. Specifically, when processing 3D point cloud data, the system first automatically divides the line into different structural sections (such as ordinary lines, bridges, tunnels, etc.) based on methods such as line design ledgers, GNSS positioning, manual marking, or automatic algorithm identification. The system has built-in point cloud characteristic templates and sleeper distribution standards for each structural section, and configures corresponding clustering parameters (such as DBSCAN neighborhood radius, minimum number of samples, sleeper template aspect ratio threshold, etc.) and anomaly detection thresholds for each type of section.

[0165] During automatic processing, the system can identify the attributes of the current point cloud segment and dynamically switch to parameters and specific rules matching that segment to segment the sleeper point cloud for segmentation, clustering, sorting, and spacing detection. For transitional segments with significant structural changes, the system can automatically invoke multi-level clustering or sliding window strategies to ensure that clustering and anomaly detection can adaptively achieve optimal accuracy and spatial consistency in different segments. This function effectively avoids false detections, missed detections, or decreased segment detection accuracy caused by parameter fixation, realizing automatic consistency detection and intelligent anomaly identification of sleeper spacing across the entire line.

[0166] D. In this embodiment, the result visualization and reporting module supports intuitive display and data export of various detection results. Based on 3D point cloud data and detection analysis results, the system can automatically generate a 3D sleeper distribution visualization map of the entire line or sections, displaying the actual spatial distribution of the track and each sleeper. Simultaneously, the system supports generating a histogram of sleeper spacing distribution, statistically analyzing the mean, standard deviation, and anomaly distribution of the spacing sequence. For detected abnormal sleepers or sections, the module can automatically highlight them, facilitating quick location and comparison of anomalies by the user.

[0167] Furthermore, the system supports exporting the inspection results and raw point cloud data of selected sections or the entire line in a standard format, including a list of abnormal sleepers, a spatial index of abnormal sections, a 3D distribution point cloud, and related statistical reports. These functions provide efficient and convenient technical support and an intuitive application interface for subsequent analysis of inspection results, on-site review, track maintenance management, and data archiving.

[0168] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for automatic calculation and anomaly detection of sleeper spacing based on 3D point cloud, characterized in that, Includes the following steps: (1) Preprocess point cloud data and extract sleeper point cloud based on deep learning guided point cloud segmentation network; (2) Based on the spatial distribution characteristics of sleeper point clouds in different sections, perform section adaptive point cloud density analysis, dynamically adjust clustering parameters, and use parameter self-tuning clustering algorithm to automatically cluster sleeper point clouds into single sleeper point cloud clusters. Automatically perform secondary subdivision on clusters with abnormal spans after the initial clustering. The parameter self-tuning clustering algorithm includes: estimating the average nearest neighbor distance of the segment point cloud through k-nearest neighbor analysis or minimum distance statistics, and adaptively setting the neighborhood radius and minimum sample number of DBSCAN clustering in combination with the actual spatial scale of the sleepers; setting clustering parameters for the differences in point cloud distribution and sleeper spacing in different sections of the bridge and line to achieve adaptive single sleeper extraction in complex scenarios; and performing span anomaly detection for each cluster based on the initial clustering results. (3) For each effective sleeper point cloud cluster, extract its main axis direction and major and minor axis feature values, and use the point cloud centroid or the midpoint of the main axis as the sleeper center coordinates to record the number of cluster points, spatial distribution and morphological parameters. (4) Based on all valid sleeper center points, the dominant direction of track space is automatically estimated, and segmented dominant direction estimation and correction are supported in structural change sections. (5) Project all sleeper center points according to the dominant direction, and automatically sort them according to the projection results to form a spatially ordered sleeper center sequence. (6) Based on the sleeper center sequence, the projection spacing of each adjacent sleeper in the dominant direction is automatically calculated to obtain the sleeper spacing sequence; (7) Perform statistical analysis on the sleeper spacing sequence, set anomaly detection threshold, automatically identify and mark abnormal spacing, support cluster parameter review and spacing correction suggestions for abnormal sections, and finally output an abnormal sleeper spacing distribution map, anomaly list and detection report.

2. The method according to claim 1, characterized in that: Step (1) specifically includes: Multi-scale geometric and reflection features are used to automatically extract features and perform multi-scale neighborhood analysis on the normalized original point cloud. By using graph convolutional networks and K-nearest neighbor graph structures, sleeper point clouds are extracted, achieving highly robust segmentation in scenarios with stray noise, occlusion, and structural deformation.

3. The method according to claim 1, characterized in that, The extraction of the sleeper center and main shaft direction attributes includes: (1) Calculate the three-dimensional centroid of each sleeper point cloud cluster as the preliminary center coordinates; (2) The principal component analysis method is used to construct the covariance matrix of the point cloud cluster. After eigenvalue decomposition, the principal axis direction and the eigenvalues ​​of the major axis and minor axis are obtained. The effective sleeper clusters are screened by combining the aspect ratio and point number geometric parameters.

4. The method according to claim 1, characterized in that, The estimation of the dominant direction of orbital space includes the following: (1) Sum the center points of all sleepers, calculate their mean coordinates by segment or the whole line, and use the covariance matrix eigenvalue decomposition method to extract the eigenvector corresponding to the largest eigenvalue as the dominant direction. (2) When there are abrupt changes in spatial structure, the dominant direction is corrected according to the segment to improve the spatial accuracy of the spacing calculation.

5. The method according to claim 1, characterized in that, The automatic calculation of sleeper spacing includes the following: projecting the center point of the sleeper onto the dominant direction, automatically sorting according to the projection value to form an ordered sequence of sleeper centers; using the difference in projection coordinates between two adjacent sleeper centers in the dominant direction as the actual spatial spacing of the sleepers to obtain a complete sleeper spacing sequence, supporting personalized processing under section / full line statistics.

6. The method according to claim 1, characterized in that, The abnormal sleeper spacing detection includes: (1) Calculate the mean, standard deviation and distribution range of all sleeper spacing data, set the anomaly identification threshold, automatically mark abnormal sleeper spacing pairs and output the abnormal sleeper spacing list. (2) For abnormal spacing segments, it supports automatic triggering of cluster parameter review, segmentation adjustment or suggestion for manual review, and supports cluster recalculation and spacing correction for abnormal segments; (3) Generate spacing distribution histogram, anomaly distribution map and automatic detection and analysis report.

7. A system employing the automatic sleeper spacing calculation and anomaly detection method based on three-dimensional point clouds as described in any one of claims 1-6, characterized in that, include: (1) Point cloud data acquisition and management module, used for 3D point cloud acquisition, storage and spatial registration; (2) Sleeper point cloud semantic segmentation module, which realizes automatic extraction of sleeper point cloud; (3) Parameter self-tuning clustering and sleeper cluster extraction module, realizes section adaptive single sleeper segmentation, and supports multi-level clustering and parameter optimization; (4) Sleeper center and main axis direction attribute extraction module, completes the calculation and screening of sleeper center of gravity and main axis direction; (5) Track space dominant direction estimation and sleeper sorting module, which realizes automatic extraction of segment dominant direction and sleeper center sorting; (6) Automatic spacing calculation and anomaly detection module, which realizes spacing calculation, statistical analysis and anomaly identification in the dominant direction; (7) Results visualization and automatic reporting module, which supports three-dimensional sleeper distribution visualization, spacing analysis diagram generation, anomaly detection report export and sleeper sequence management for the entire line.

8. The system according to claim 7, characterized in that, The system supports automatic switching of parameters for multiple structural sections and calls clustering and anomaly detection rules for different track types to achieve automatic consistency detection and anomaly identification of sleeper spacing across the entire line.

9. The system according to claim 7, characterized in that, The results visualization and reporting module supports 3D sleeper distribution point cloud display, spacing distribution histogram, automatic highlighting of abnormal sleepers, and export of abnormal sections and raw data, which facilitates the analysis and application of detection results.

10. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein the program, when executed, controls the device where the non-volatile storage medium is located to perform the method described in any one of claims 1 to 6.

11. A terminal device, characterized in that, The terminal device includes: a processor, a memory, a communication interface, and a bus; the processor, the memory, and the communication interface are connected through the bus and communicate with each other; the memory stores executable program code; the processor reads the executable program code stored in the memory to run a program corresponding to the executable program code, so as to perform the method as described in any one of claims 1-6 above.

Images