Railway emergency scene point cloud relative pose estimation method and system

Abstract

The present application provides a railway emergency scene point cloud relative pose estimation method and system, constructs a direction alignment voxel coding network and a geometric consistency constraint fusion registration framework, combines local structure standardization modeling and graph structure consistency screening mechanism, and realizes high-precision point cloud alignment in complex railway scene. The method includes three main steps: firstly, the rotation invariant voxel feature is constructed, the spatial local structure is extracted through the direction alignment matrix, local voxel division and multi-layer perception machine mapping, and the multi-scale voxel expression is constructed by using three-dimensional convolution; then, the geometric compatibility graph is constructed based on the distance difference and the direction difference, the initial edge weight is determined by the first-order compatibility constraint, the global structure consistency is enhanced by using the second-order weight matrix, and the false matching points are removed by using the maximum group search; finally, the singular value decomposition is performed on the matching point pairs in the maximum group, the rotation matrix and the translation vector are respectively solved, and the final rigid body transformation is selected according to the group score, so as to realize the accurate pose calculation of the template point cloud and the on-site emergency point cloud.

Description

Technical Field

[0001] This invention relates to the fields of emergency intelligent sensing, computer vision and three-dimensional spatial information processing technology, and in particular to a method and system for estimating the relative pose of point clouds in railway emergency scenarios. Background Technology

[0002] When railways encounter sudden disasters such as landslides, floods, collapses, and debris impacts, the damaged areas often exhibit typical "unstructured" point cloud characteristics, including dense noise, structural defects, severe occlusion, and irregular deformation. Emergency equipment, during rapid data acquisition, is affected by factors such as attitude changes, lighting disturbances, and differences in flight altitude, resulting in significant differences in scale, density, rotation angle, and geometric consistency between point clouds acquired at different times and by different equipment. Traditional point cloud registration methods rely on manual features, iterative optimization, or rigid assumptions. When the noise ratio exceeds 15% or the viewing angle difference exceeds 30°, the matching error rate increases significantly, and rotation and displacement errors multiply, making it difficult to meet the requirements of rapid, robust, and real-time computation in emergency scenarios. Furthermore, because disaster point clouds are often accompanied by large-scale missing data, the initial matching contains a large number of spurious correspondences, making the overall solution process prone to erroneous convergence or inability to solve, severely impacting the accuracy of post-disaster assessment and 3D reconstruction.

[0003] To address the issues of inconsistent point cloud rotation and large pose differences, a point cloud description method based on rotation-invariant feature extraction is proposed. This type of method projects the point cloud onto a space with a normalized structure, such as a spherical or cylindrical coordinate system, by standardizing the local geometry, and then combines this with a deep convolutional network to extract rotation-robust local features. This approach can maintain feature consistency even with significant differences in device pose and random target orientation, mitigating the vulnerability of traditional features to rotational changes and providing a foundation for stable matching between point clouds.

[0004] To address common false matches and noise interference in disaster point clouds, an anomaly matching removal mechanism based on maximal cliques or second-order geometric compatibility graphs is proposed. This method constructs a geometric relationship graph between matching point pairs and searches for highly consistent maximal cliques within it, thereby eliminating a large number of erroneous matching points that do not meet geometric constraints. Since maximal cliques inherently possess the constraint property of "global consistency," they exhibit higher robustness in point cloud data with high noise levels, numerous missing data points, and significant deformation, providing a reliable matching basis for solving rotation and translation matrices. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for relative pose estimation of point clouds in railway emergency scenarios, which is used to solve the problems existing in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution.

[0007] A method for relative pose estimation of point clouds in railway emergency scenarios includes: S1. The template point cloud and the on-site emergency point cloud are centered, regionally clipped and voxelized. The local point cloud is mapped to spherical or cylindrical coordinate system, and rotation-invariant voxel features are constructed based on the orientation alignment matrix. S2. Three-dimensional convolution and multilayer perceptron are used to encode rotation-invariant pixel features, extract cross-view stable deep local structure features, and generate initial deep local structure feature matching pairs based on feature similarity. S3. Construct a geometric compatibility map based on distance consistency and orientation consistency using template point cloud and on-site emergency point cloud. Generate a weight matrix through first-order compatibility constraints and further calculate a second-order compatibility matrix to strengthen global structural constraints. S4. Perform a search for maximal cliques based on the compatibility graph, score the consistency of all deep local structural feature matching pairs within the clique, eliminate false matches that do not meet geometric constraints, and obtain a highly reliable matching set. S5. After performing step S4, use the singular value decomposition method to obtain the rotation and translation matrix for each maximal clique, and select the final pose result according to the maximal clique score to obtain the optimal rigid body transformation of the template point cloud relative to the emergency point cloud.

[0008] Preferably, the construction of the rotationally invariant chromosome in step S1 includes: S11. Spatial division of template point cloud and on-site emergency point cloud, mapping local areas to spherical coordinates or cylindrical coordinates, and voxelization processing according to radius, angle and height dimensions; S12. Construct a local orientation alignment matrix for each voxel to unify the orientation of the point cloud inside the voxel and improve the consistency of voxel features across viewpoints. S13. Use a multilayer perceptron with shared weights to extract local features of point clouds within voxels, and obtain fixed-dimensional voxel descriptors through max pooling.

[0009] Preferably, the feature encoding process in step S2 includes: S21. All voxel features are combined according to spatial index to construct a four-dimensional voxel feature tensor, which is used to maintain the spatial topological relationship of the three-dimensional structure. S22. A three-dimensional convolutional network along the radial, pitch, and azimuth directions is used to extract deep geometric features. This three-dimensional convolutional network is implemented using the formula...

[0010] Perform convolutional updates; where j represents the voxel index in the radial direction, k represents the voxel index in the height or axial direction, l represents the voxel index in the azimuth or orientation direction, and s represents the network layer index, used to distinguish the feature maps of layer s from those of layer (s-1); superscript The input features from the previous layer are represented by d, which represents the feature channel index, i.e., the d-th feature dimension, used to describe the specific channel in the multi-channel feature mapping. d′ represents the channel index of the output feature after the current convolution or mapping operation, used to distinguish between the input and output feature dimensions. r represents the radial offset index of the convolution kernel, and its value range is determined by the radial convolution kernel size. The value of x is determined by the kernel width parameter, which describes the local receptive range within the radial neighborhood. The value of y represents the offset index of the convolution kernel in the height direction (axial direction), and its range is determined by the convolution kernel height parameter. The decision is as follows: D represents the total number of channels in the input feature, i.e., the dimension of the previous layer feature map in the channel dimension; F represents the feature mapping, which specifically includes: This represents the input feature of the d-th channel in the (s-1)-th layer. This indicates the position of the s-th layer at the voxel. The output feature value of the d-th channel, where w represents the learnable weight parameters of the 3D cylindrical convolution kernel, corresponding to the convolution weights for different channels and different spatial offset positions. These represent the receptive range or size parameters of the s-th layer 3D cylindrical convolution kernel in the radial, height, and angular dimensions, respectively; S23. Enhance the structural representation ability of key parts in template point cloud and on-site emergency point cloud by multi-scale regional feature aggregation operation, so as to improve the stability and noise resistance of the initial matching pair.

[0011] Preferably, the compatibility graph construction process in step S3 includes: S31. Based on the distance consistency as the edge weight, establish a first-order compatibility graph according to the distance difference of the deep local structural feature matching pairs. S32. Perform multiple matrix multiplication operations on the weight matrix of the first-order compatibility graph to obtain a second-order compatibility weight matrix, which is used to enhance global geometric consistency. S33. Based on the structural distribution of the second-order compatibility graph, remove deep local structural feature matching pairs with weak geometric constraints to obtain a set of candidate deep local structural feature matching pairs with strong geometric constraints.

[0012] Preferably, the maximal cluster screening process in step S4 includes: S41. Search for maximal cliques that satisfy complete connectivity in the second-order compatibility graph and use them as a set of candidate reliable matches. S42. Accumulate the edge weights of the maximal clique to obtain the geometric consistency score of the maximal clique, which is used to measure its matching quality. S43. Based on the normal vector consistency constraint, perform secondary screening of matching points inside the maximal clique to eliminate false matching points with inconsistent local directions.

[0013] Preferably, the pose calculation process in step S5 includes: S51. For each pair of matching points inside a maximal clique after substep S43, use singular value decomposition to obtain the rotation matrix and translation vector. S52. Select the clique with the highest geometric consistency score based on the maximal clique as the source of the final pose output.

[0014] Secondly, the present invention provides a point cloud relative pose estimation system for railway emergency scenarios, comprising: Rotation-insensitive voxel encoding module: Maps the input point cloud to a cylindrical coordinate system, performs spatial voxel division and orientation alignment, and generates voxel-level feature representations with rotation insensitivity through multilayer perceptron and 3D convolution; Geometric compatibility graph and maximal clique selection module: Initial matching is constructed based on voxel features, geometric compatibility graph is constructed using distance difference and orientation difference, first-order and second-order compatibility weights are calculated, and maximal clique search is used to obtain a highly consistent subset of matches; The SVD-based pose solving and registration result evaluation module performs rigid body transformation on the matched point pairs within the maximal clique to obtain the rotation matrix and translation vector, and visualizes and evaluates multiple sets of experimental data to verify the effectiveness of the method of this invention.

[0015] As can be seen from the technical solutions provided by the embodiments of the present invention above, the present invention provides a method and system for relative pose estimation of point clouds in railway emergency scenarios. It constructs a registration framework that integrates a direction-aligned voxel encoding network with geometric consistency constraints, and combines local structure standardization modeling and graph structure consistency screening mechanisms to achieve high-precision point cloud alignment in complex railway scenarios. This method includes three main steps: First, it constructs rotation-invariant voxel features, extracts spatial local structures through direction alignment matrices, local voxel partitioning, and multilayer perceptron mapping, and constructs multi-scale voxel representations using 3D convolution; then, it constructs a geometric compatibility graph based on distance difference and direction difference, determines initial edge weights through first-order compatibility constraints, enhances global structure consistency using second-order weight matrices, and eliminates false matching points using maximal clique search; finally, it performs singular value decomposition on matching point pairs within maximal cliques, calculates rotation matrices and translation vectors respectively, and selects the final rigid body transformation based on the clique score to achieve accurate pose calculation of the template point cloud and the on-site emergency point cloud. The advantages and positive effects of the present invention are: (1) The rotation-invariant feature extraction module proposed in this invention can effectively eliminate rotational differences caused by UAV attitude changes, heterogeneous acquisition by multiple devices, and disaster disturbances, so that the point cloud maintains a consistent structural expression under different perspectives. Compared with traditional manual features, this method significantly improves matching robustness and enhances feature stability under random rotation conditions, meeting the rapid response requirements of sudden disaster scenarios.

[0016] (2) The maximal clique geometric consistency screening method proposed in this invention can effectively remove a large number of false matching point pairs caused by noise, missing or deformed points in disaster point clouds. Experiments show that when the noise ratio exceeds 15%, the maximal clique-based removal mechanism can significantly reduce the number of mismatches, significantly improve the accuracy of subsequent pose solving, and ensure the stability of registration results in complex environments.

[0017] (3) While maintaining high registration accuracy, this invention has optimized the overall computation process to ensure that the algorithm can run in real time on UAV edge computing units, emergency terminals, or on-site command equipment. Compared with traditional point cloud registration methods, this method has a shorter average inference time in complex environments, which can meet the actual engineering needs of "rapid acquisition - rapid processing - rapid decision-making" in railway emergency scenarios.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0020] Figure 1 This is a flowchart of a point cloud relative pose estimation method for railway emergency scenarios provided by the present invention; Figure 2 This is a schematic diagram of shallow feature extraction for a point cloud relative pose estimation method in a railway emergency scenario provided by the present invention; Figure 3 This is a compatibility diagram of the construction of a point cloud relative pose estimation method for railway emergency scenarios provided by the present invention; Figure 4 This is a comparative experimental result diagram of a point cloud relative pose estimation method for railway emergency scenarios provided by this invention; Figure 5 This is a logic block diagram of a point cloud relative pose estimation system for railway emergency scenarios provided by the present invention. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0023] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0024] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0025] This invention provides a method for relative pose estimation of point clouds in railway emergency scenarios based on rotation invariance and maximal clique screening, to address the following technical problems existing in the prior art: (1) Existing point cloud registration techniques are mostly based on iterative nearest point (ICP) or local geometric feature description, usually assuming that the initial attitude error is small and the structure is relatively intact. When UAVs collect data on the same railway equipment multiple times at different altitudes, routes, and perspectives, there are often large rotational differences and perspective changes between point clouds. In addition, the damage to components, occlusion, and missing point clouds caused by disasters make it difficult for traditional methods to obtain a stable initial correspondence, and it is easy to fail to converge or have large deviations in the registration results. Especially in railway emergency scenarios, it is necessary to align targets such as support devices and monitoring equipment with high precision. Existing methods have obvious shortcomings in terms of rotation invariance and structural adaptability.

[0026] (2) Multi-view acquisition leads to inconsistent point cloud orientations, making it difficult for traditional methods to establish stable local structural representations. In emergency scenarios, point clouds acquired by UAVs, mobile platforms, or from different operating angles exhibit significant rotational differences. Traditional registration methods based on points, edges, or geometric descriptors struggle to construct features stable to rotational changes, often resulting in feature mismatches and distorted attitude estimation. To address this issue, this invention introduces an orientation-aligned voxel encoding mechanism. By standardizing local coordinates and modeling with voxels, rotation-insensitive features are generated, fundamentally solving the problem of unifying the representation of multi-pose point clouds.

[0027] (3) Complex noise and local missing interference lead to a large number of false matches, and traditional matching strategies are difficult to effectively eliminate false matches. Railway disaster sites often have complex interference such as vegetation obstruction, equipment damage, and structural gaps. Traditional matching strategies based on neighborhood similarity or local geometry are not stable enough in high-noise environments and are prone to forming a large number of false correspondences. This invention constructs a geometric compatibility graph based on distance difference and direction difference to achieve global consistency constraints between matching point pairs, providing a reliable structural foundation for subsequent screening.

[0028] (4) Existing graph matching methods rely only on local consistency, which makes it difficult to reflect the overall geometric structure, resulting in unstable matching points. Many methods only use local geometric relationships for matching and filtering, lacking modeling of the overall structural consistency, which easily leads to matching that is locally consistent but globally incorrect. This invention uses first-order and second-order compatibility matrices to fuse local and global geometric relationships, and strengthens the connectivity of corresponding points of the real structure through a global weight diffusion mechanism, which significantly improves the stability of the matching point set.

[0029] (5) The maximal clique screening strategy lacks a dynamic consistency metric, making it impossible to accurately identify matching subsets that truly satisfy structural constraints. Traditional maximal clique methods often rely solely on connectivity judgments, lacking a quantitative measure of the strength of consistency within the clique, leading to unstable result quality. This invention proposes a clique consistency scoring mechanism based on edge weight accumulation, and combines it with normal vector direction constraints to re-verify points within the clique, effectively eliminating false matches caused by local directional anomalies or structural mutations.

[0030] (6) Existing pose estimation methods cannot robustly select the optimal solution among multiple candidate matching structures, leading to the final result easily deviating from the true pose. Faced with multiple possible candidate structures, traditional SVD cannot determine which set of matches is most meaningful. This invention solves the rotation and translation matrices for each maximal clique separately and selects the final solution through the maximal clique consistency score, making the pose estimation more robust and stable, and better adaptable to structural deficiencies and noise characteristics in emergency scenarios.

[0031] (7) Traditional pose solving methods based on singular value decomposition typically perform least squares optimization on a single matching set. When this set still contains a certain proportion of mismatches or structural inconsistencies, the obtained rigid body transformation will be significantly disturbed by outliers, resulting in deviations in rotation and translation estimations. At the same time, when there are multiple possible geometric correspondences, conventional methods lack a mechanism to select the optimal solution from multiple candidate solutions, making it difficult to ensure that the output corresponds to the real target structure. On the other hand, although some registration algorithms using global search or complex energy optimization have high theoretical accuracy, they have high computational cost and slow convergence speed, making it difficult to meet the real-time requirements in railway emergency scenarios that require rapid response.

[0032] See Figure 1 This invention provides a method for estimating the relative pose of point clouds in railway emergency scenarios, comprising the following steps: S1. The template point cloud and the on-site emergency point cloud are centered, regionally clipped and voxelized. The local point cloud is mapped to spherical or cylindrical coordinate system, and rotation-invariant voxel features are constructed based on the orientation alignment matrix. S2. Three-dimensional convolution and multilayer perceptron are used to encode voxel features, extract cross-view stable deep local structural features, and generate initial matching pairs based on feature similarity. S3. Construct a geometric compatibility graph based on distance consistency and direction consistency, generate a weight matrix through first-order compatibility constraints, and further calculate a second-order compatibility matrix to strengthen global structural constraints. S4. Search for maximal cliques based on compatibility graphs, score the consistency of all matching point pairs within the clique, and remove false matches that do not meet geometric constraints to obtain a highly reliable matching set. S5. Use the singular value decomposition method to obtain the rotation and translation matrix for each maximal clique, and select the final pose result based on the maximal clique score to obtain the optimal rigid body transformation of the template point cloud relative to the emergency point cloud.

[0033] In the preferred embodiment provided by the present invention, the specific execution process of each step is as follows.

[0034] In the point cloud rotation-invariant feature construction stage, this embodiment proposes a feature generation method based on orientation alignment and spatial voxelization encoding for 3D point clouds of emergency facilities collected from different perspectives and equipment in railway emergency scenarios, in order to obtain structural representations with high stability and rotation insensitivity. For example... Figure 1 The step shown first performs spatial normalization on the input point cloud, dividing the global point cloud into local spatial regions centered on the target device, and initially partitioning the point cloud using spherical or cylindrical coordinates. Based on this, the continuous three-dimensional space is discretized along the three directions of radius, pitch angle, and azimuth angle. Each voxel unit corresponds to a local point cloud within a fixed spatial range. When the drone's shooting angle, equipment orientation, or platform attitude changes, the above spatial division ensures that local point clouds under different acquisition conditions can obtain a consistent encoding entry point within the same spatial grid, preventing structural anomalies caused by differences in observation angles.

[0035] After completing voxel segmentation, this embodiment performs a local alignment operation on the point cloud within each voxel to eliminate the influence of different device perspectives and on-site acquisition postures on the local geometry. Specifically, for the point set near the voxel center... Construct a rotation matrix with the voxel direction as a reference. The matrix is ​​then applied to the local point cloud, aligning it spatially to a unified reference coordinate system. The aligned voxel point cloud exhibits significant directional consistency in its geometric structure, ensuring comparability of the point cloud layout within each voxel even when the same emergency equipment is collected at different flight paths, orientations, or lighting conditions. To compress the point cloud from a variable-length point set representation into a fixed-dimensional description vector, this embodiment uses a multilayer perceptron with shared weights for feature mapping within the aligned voxels and extracts the most discriminative local point features for each voxel using a max-pooling operator. This process can be represented as follows:

[0036] in, Nonlinear encoding is performed on the local point cloud. Aggregation is achieved at the voxel level, giving each voxel a stable and compact local geometry descriptor.

[0037] After obtaining the primary features of all voxels, this embodiment combines them according to spatial index order into a dimension of A four-dimensional feature tensor is used. This tensor fully covers the local structure of the emergency equipment in the spatial dimension and contains high-dimensional geometric responses of different voxels in the channel dimension, forming the input for subsequent deep encoding. Based on this, this embodiment employs a spatially continuous three-dimensional convolutional network to further enhance the voxel features, extracting higher-level geometric semantic information through convolutional accumulation along the radial, pitch, and azimuth directions. For the convolutional layers located at spatial indices... Channel Index Features, such as Figure 2 As shown, its update form can be expressed as

[0038] In this model, the three dimensions of the convolution kernel correspond to the neighborhood relationships in the radial, height, and circumferential directions, respectively. By sliding the three-dimensional convolution kernel along these dimensions, the overall structural distribution of the emergency equipment point cloud in the local space can be fully captured. This convolution method has natural rotational equivariance in cylindrical or spherical spaces, ensuring that features maintain a consistent response under rotational changes, thereby significantly improving feature stability. In the formula, j, k, and l represent the three-dimensional discrete spatial indices in the voxelized feature space, where: j represents the voxel index in the radial direction; k represents the voxel index in the height or axial direction; and l represents the voxel index in the angular or azimuth direction. These three together determine the voxel position in a cylindrical coordinate system. s represents the layer index of the network, used to distinguish the feature maps of layer s from those of layer (s-1); the superscript... This represents the input features from the previous layer. `d` represents the feature channel index, i.e., the `d`-th feature dimension, used to describe the specific channel in the multi-channel feature mapping. `d′` (or denoted as the upper-layer correspondence of `d`) represents the channel index of the output feature after the current convolution or mapping operation, used to distinguish between the input and output feature dimensions; in specific implementations, `d′` can be the same as or different from `d`, depending on the network channel transformation settings. `r` represents the radial offset index of the convolution kernel, its value range being determined by the radial convolution kernel size. The value is determined to describe the local receptive range within the radial neighborhood. x represents the offset index of the convolution kernel in the width direction (angular or lateral), and its value range is determined by the convolution kernel width parameter. Determined. y represents the offset index of the convolution kernel in the height direction (axial direction), and its value range is determined by the convolution kernel height parameter. The decision is made. D represents the total number of channels in the input feature map, i.e., the dimensionality of the previous layer's feature map in terms of channels. F represents the feature map, which specifically includes: The input feature of the d-th channel in the (s-1)-th layer; : The s-th layer at the voxel position The output feature value of the d-th channel. w represents the learnable weight parameters of the 3D cylindrical convolution kernel, corresponding to the convolution weights of different channels and different spatial offset positions. These represent the receptive range or size parameters of the s-th layer 3D cylindrical convolution kernel in the radial, height, and angular dimensions, respectively. Through the accumulation of multiple convolutions, the feature tensor aggregates local information from different voxels layer by layer, gradually deepening the overall features of the emergency equipment in both spatial and semantic dimensions. When the local structure is affected by noise or partial occlusion, the convolutional features can still compensate for the missing regions by relying on information from adjacent voxels, maintaining the robustness of the features. After the convolutional features are generated, pooling, normalization, and necessary linear mapping can be used to further obtain the rotation-invariant descriptors used in the matching stage, ensuring that the same equipment maintains a consistent high-level feature expression under different acquisition conditions.

[0039] Based on the aforementioned features, an initial matching relationship is established between the template point cloud and the emergency site point cloud. By constructing a geometric compatibility graph and employing a maximal clique search mechanism, anomalous matching pairs are eliminated, thereby obtaining a robust set of corresponding points. Specifically, firstly, using the voxel-level or point-level feature descriptors obtained in step one, nearest neighbor feature matching is performed on the candidate emergency equipment targets in the template scene and the site scene, resulting in several candidate matching pair sets. Since railway disaster site point clouds often contain a large amount of environmental noise, occluded areas, and non-target facilities, erroneous correspondences inevitably mix into these candidate matching pairs. If directly used for pose calculation, this will cause the rotation and translation parameters to deviate significantly from their true values. Therefore, it is necessary to further introduce strict geometric consistency constraints to screen and optimize these initial matching pairs.

[0040] First, construct as follows Figure 3 The order compatibility graph shown denotes each candidate matching pair as a node, and quantifies the spatial geometric relationship between any two matching pairs. For the source point in the template point cloud... Target points in the on-site point cloud Define their distance difference as

[0041] In the above formula, : Indicates a pair of matching nodes. These are nodes in the compatibility graph, where each node represents a hypothetical matching relationship between the source point cloud and the target point cloud. : indicates a matching pair and The coordinates of points in the associated source point cloud. (Superscript) It represents Source. : indicates a matching pair and The coordinates of points in the associated TargetPoint Cloud. (Superscript) It represents Target. : Represents the distance difference metric between two matching pairs.

[0042] If the difference is close to zero, it indicates that the two pairs of matches maintain good rigidity consistency on a spatial scale. Based on this, to distinguish between geometrically close and loosely matched pairs, this embodiment further constructs a compatibility score.

[0043] in, For distance scale parameters, Here, "nodes" represents the vertices of a graph. In the context of constructing a first-order graph (FOG), a node... Typically, it is a data pair , representing points in the source point cloud Points in the target point cloud were matched. , and It refers to any two distinct matching hypotheses selected in a graph network. The subscript 'dist' is an abbreviation for Distance. This represents the "dissimilarity" or "error value" calculated based on geometric distance differences. The subscript cmp is an abbreviation for Compatibility. This represents the final calculated "compatibility score". It is a parameter that controls compatibility sensitivity (similar to the standard deviation in a Gaussian function). This is used to adjust the tolerance for distance errors. If... If the score exceeds a preset threshold, an edge connection is established between the corresponding nodes in the first-order compatibility graph, and the score is used as the edge weight. After this processing, a weighted graph structure reflecting the first-order geometric relationship between each matching pair can be obtained. True matching pairs often form dense connections, while incorrect matching pairs usually have only weak connections or almost no connections with a few nodes.

[0044] Considering that relying solely on first-order geometric relationships may still retain some incorrect matches in emergency scenarios with high noise and severe local deficiencies, this embodiment further constructs a second-order compatibility graph through weight matrix operations to strengthen global consistency constraints. Let the weight matrix of the first-order compatibility graph be... Then the weight matrix of the second-order compatibility graph It can be expressed in the form of multiple matrix multiplication as

[0045] This process is equivalent to comprehensively considering multi-hop adjacency relationships in the graph, significantly amplifying the weight of correspondences that are widely supported by other matching pairs in the global structure, while weakening or even reducing the weight of spurious matching pairs that only locally and accidentally satisfy first-order constraints. The resulting second-order compatibility graph can more effectively highlight the structural clusters formed by true matchings, providing a clearer graph structure foundation for subsequent maximal clique extraction.

[0046] After obtaining the second-order compatibility graph, this embodiment extracts the most compatible subset of matches from the graph based on the concept of maximal cliques in graph theory. A maximal clique is a complete subgraph in which every two nodes are connected by an edge, characterized by all matching pairs within the clique being geometrically identical. This embodiment addresses each candidate clique... The weights of all edges within the clique are summed to obtain the overall weight of the clique.

[0047] in Let the set of edges within the group be . The weights represent the corresponding edges. Larger weights indicate stronger geometric consistency between matching pairs within a clique, meaning the clique is more likely to correspond to a real emergency equipment structure. This embodiment selects several maximal cliques with the largest weights as the final effective matching set. Furthermore, normal vector direction constraints can be introduced to verify the normal relationships of the neighborhood of matching points within the clique, retaining only those that satisfy the constraints.

[0048] The matching pairs are further eliminated to remove outlier points with significant inconsistencies in local geometric orientation. Through joint constraints of distance consistency, global weight accumulation, and normal angle consistency, a set of matching point pairs that are highly consistent in spatial scale, global structure, and local orientation is finally obtained. Where, Defined as two points in the source point cloud. and Corresponding normal vector and The angle between them. Defined as two corresponding points in the target point cloud. and Corresponding normal vector and The angle between them. : This is a manually set threshold.

[0049] After screening maximal cliques, the corresponding rotation and translation matrices are obtained using the matching point pairs within each maximal clique. Since each maximal clique consists of matching points that strictly satisfy geometric consistency, singular value decomposition can be applied to each maximal clique to calculate the rigid body transformation, resulting in a set of candidate rotation matrices and translation vectors. To evaluate the reliability of the transformation matrices generated by different maximal cliques, this embodiment evaluates them based on the consistency score of the matching point pairs within the maximal cliques. Let the set of effective maximal cliques be... The quantity is Each of the maxima Each corresponds to a set of matching points and a rotation / translation matrix obtained from singular value decomposition. A score is calculated for each maximal clique. This score comprehensively reflects the geometric consistency between matching point pairs within a clique, and can be calculated based on indicators such as mean absolute error, mean square error, and the number of internal points. Based on this, the rotation matrix and translation vector corresponding to the maximal clique with the highest score are selected from all candidate transformations as the final relative pose solution. The above selection process can be expressed as follows:

[0050] In the formula, This represents the final preferred rotation matrix. This represents the final preferred translation vector.

[0051] This invention also provides an embodiment to demonstrate the beneficial effects of the method provided by this invention. This experiment mainly evaluates the method from aspects such as registration accuracy, success rate, and comparison with existing typical algorithms. This experiment verifies the point cloud registration method based on rotation invariance and maximal clique selection proposed in this invention on point cloud datasets collected from 3DMtch, ETH Zurich labs, and railway field UAVs. The experiment uses rotation error (RE) and translation error (TE) as quantitative evaluation indicators.

[0052] In terms of experimental setup, this invention selects multiple representative 3D point cloud pairs as test samples, including device structure point clouds under different viewpoints, acquisition heights, and occlusion levels. Evaluation metrics include rotation error (RE) and translation error (TE): rotation error measures the angular difference between the estimated rotation matrix and the reference rotation, while translation error is the Euclidean distance between the estimated translation vector and the reference translation vector. Successful registration is determined by RE being less than a given angle threshold and TE being less than a given distance threshold. By statistically analyzing the number of successful registrations and error distribution across all samples, the registration performance of this invention's method in different scenarios can be comprehensively reflected.

[0053] For benchmark comparison, this invention selects several typical point cloud registration algorithms as references, including Fast Global Registration (FGR), globally optimized Go-ICP, and PointDSC based on consistency graphs. All methods are compared under the same dataset and evaluation metrics to ensure the comparability of experimental conclusions. For each set of point clouds, this invention and the comparison algorithm output the corresponding rotation and translation matrices, and calculate the corresponding RE and TE. Simultaneously, visualizations of the registration before and after are plotted on representative samples, displaying the template point cloud and the transformed field point cloud in superimposed red and green colors to intuitively demonstrate the registration quality.

[0054] To verify the model's generalization ability using the ETH dataset, registration success rate (SR) was chosen as the metric. Experiments were conducted on four different scenarios within the ETH dataset, and the results are shown in the table below: Table 1 Test results for the ETH dataset

[0055] Experimental results on the ETH dataset show that the model in this chapter maintains a high registration success rate of over 90% in four unseen scenarios, proving the effectiveness of the registration method.

[0056] The point cloud registration method proposed in this invention is compared with the classic point cloud registration algorithms FGR, Go-ICP, and PointDSC. Rotation error (RE) and translation error (TE) are used as quantitative evaluation indicators. Registration experiments are conducted using contact wire poles as a typical representative of railway emergency scenarios. The experimental results are as follows: Figure 4 As shown.

[0057] The comparison results are shown in Table 2.

[0058] Table 2 Comparison of experimental results

[0059] As shown in the table, the proposed method outperforms FGR, Go-ICP, and PointDSC in both RE and TE across all test scenarios. Particularly under complex structures and local occlusion conditions, the proposed method exhibits smaller rotation errors and more stable translation estimations. Compared to other algorithms that show significant rotational deviations or translational shifts in some scenarios, this method maintains stable registration accuracy across all experimental groups. This fully demonstrates that the rotation invariance and maximal clique screening mechanism proposed in this invention can more effectively suppress erroneous matches, improve the overall robustness and geometric consistency of registration, and has significant practical application value.

[0060] In summary, experimental results show that even in samples with large-angle viewing differences, missing local structures, and significant noise interference, the method of this invention maintains relatively small rotation and translation errors, and the registration results of multiple arrays meet the preset success criteria. In contrast, some traditional methods exhibit large rotational deviations or alignment failures under the same conditions. Quantitative comparisons with FGR, Go-ICP, and PointDSC demonstrate that this invention exhibits better stability in terms of average RE, TE, and successful registration rate, particularly in experimental scenarios with high noise levels and partially obscured structures. Furthermore, visualization results on actual railway UAV-collected data show that multiple flight point clouds registered using the method of this invention highly overlap at key structural locations, providing a reliable foundation for subsequent deformation analysis and displacement monitoring.

[0061] Secondly, the present invention provides a point cloud relative pose estimation system for railway emergency scenarios, comprising: Rotation-insensitive voxel encoding module 501: maps the input point cloud to a cylindrical coordinate system, performs spatial voxel division and orientation alignment, and generates voxel-level feature representations with rotation insensitivity through multilayer perceptron and 3D convolution; Module 502 for Geometric Compatibility Graph and Maximal Clique Selection: Constructs initial matching based on voxel features, builds a geometric compatibility graph using distance difference and orientation difference, calculates first-order and second-order compatibility weights, and uses maximal clique search to obtain a highly consistent subset of matches; The SVD-based pose solving and registration result evaluation module 503 performs rigid body transformation on the matched point pairs within the maximal clique to obtain the rotation matrix and translation vector, and visualizes and evaluates multiple sets of experimental data to verify the effectiveness of the method of the present invention.

[0062] In a preferred embodiment provided by the present invention, the working process of this system is as follows: First, the two point cloud frames are respectively fed into the rotation-invariant voxel encoding module 501. For each point cloud frame, this invention first divides and aligns the space around the target based on cylindrical coordinates, then extracts the local geometric response through three-dimensional cylindrical convolution, and then uses a multilayer perceptron and max pooling to generate a fixed-dimensional voxel-level feature vector, resulting in the feature bar representation shown on the right side of the figure. The green and red channels correspond to the feature encoding results of the two point cloud frames, respectively. This stage completes the rotation-invariant feature modeling of the original point cloud.

[0063] Subsequently, the voxel features of the two point clouds were used to establish an initial set of matching pairs through nearest neighbor search or similarity measurement, and the matching relationships were connected in the form of lines. Figure 1 The diagram is shown in the upper right corner. These lines represent candidate corresponding points between the template point cloud and the actual point cloud, but they contain a certain number of incorrect matches, which need to be further filtered using geometric constraints.

[0064] In the purple box on the right, this invention utilizes the matching results to construct a geometric compatibility graph and performs maximal clique filtering on the graph structure. Specifically, each matching pair is considered a node in the graph, and edges are established between different nodes based on distance and direction differences, forming the compatibility graph structure shown in the upper part of the figure. By applying first-order and second-order geometric constraints to the compatibility graph, nodes and edges with weak connectivity or low weights are gradually eliminated, retaining structurally stable subgraphs. Finally, several maximal cliques are extracted from the graph, each representing a set of matching points that are highly consistent in geometric relationships. Figure 1 The middle section illustrates the process of gradually shrinking from the original diagram to the maximal clique structure.

[0065] Within the selected maximal cliques, singular value decomposition is used to obtain the rotation matrix R and translation vector t. The template point cloud is then transformed to the coordinate system of the actual point cloud, resulting in the registration point cloud shown in the lower right corner. This completes the entire process, from original point cloud input, rotation-invariant feature extraction, global geometric consistency screening to rigid body pose solving, providing a reliable foundation for 3D reconstruction and equipment offset analysis in railway emergency scenarios.

[0066] like Figure 2As shown, this invention first performs cylindrical voxel division and voxel feature encoding on point cloud data in railway emergency scenarios. Considering that railway equipment (such as support devices, monitoring poles, emergency communication towers, etc.) has obvious longitudinal and circumferential characteristics in space, this invention selects a cylindrical coordinate system to model the space surrounding the target.

[0067] Specifically, taking the geometric center of the target device to be aligned or a designated reference point as the axis of the cylinder, the cylinder is discretized in the radial direction, vertical direction, and circumferential direction, forming a structure as follows: Figure 1 The "cylindrical element stacking structure" shown on the left. Among them: J represents the number of discrete layers along the radial direction, used to characterize the radial structural changes of the device from the center to the outer space; K represents the number of layers along the height direction, used to describe the differences in the longitudinal structure of the equipment; L represents the number of subdivisions along the circumferential direction, used to capture geometric details in different orientations.

[0068] After the above division, the three-dimensional continuous point cloud space can be discretized into J×K×L cylindrical voxel units of the same size, each voxel containing a local set of points within that spatial region. To eliminate the influence of the overall rotation of the point cloud on the representation of local structures under different acquisition postures, this invention performs orientation alignment processing on the point cloud within each voxel. That is, based on the relationship between the axis and the local normal of the cylindrical coordinate system, the points inside the voxel are rotated and normalized, so that the same structure acquired from different viewpoints has a similar spatial distribution after alignment.

[0069] After completing voxel segmentation and orientation alignment, this invention employs multilayer perceptrons (MLPs) with shared parameters to encode the features of the local point set within each voxel. The points within a voxel are first normalized to a local coordinate system and then input into several fully connected layers. Structural features such as geometric shape and curvature changes are extracted through nonlinear mapping. Subsequently, max pooling is performed within the voxel to compress the variable-length point set into a fixed-dimensional feature vector. Figure 2 The multiple "cylindrical layers" shown on the right represent D-dimensional voxel features stacked on a J×K×L structure, where D is the number of voxel feature channels.

[0070] Through the above processing, this invention transforms the original point cloud into a four-dimensional feature tensor of size J×K×L×D. This tensor preserves the local geometric structure and reduces the sensitivity to the acquisition pose through cylindrical coordinates and orientation alignment, providing a robust rotation-invariant variogram representation for subsequent deep feature extraction and matching construction based on three-dimensional convolution.

[0071] In the geometric compatibility graph and maximal clique screening module 502, after obtaining the voxel features, the present invention models the candidate correspondence between the template point cloud and the field point cloud as a geometric compatibility graph. Each circular node C1 to C7 in the graph represents a pair of candidate matching pairs (i.e., a point in the template point cloud and a point in the field point cloud), and the lines connecting the nodes indicate that the two pairs of matching satisfy certain compatibility conditions in terms of geometric structure.

[0072] In its specific implementation, this invention first performs an initial match between the template and the on-site point cloud using voxel features or keypoint descriptors, obtaining a set of several candidate point pairs. For any two pairs of matched points... c i and c j This invention calculates a compatibility score by measuring geometric quantities such as distance and orientation differences between the template point cloud and the actual point cloud. When the distance between two pairs of matching points changes little and their orientations are similar, they are considered to be consistent under rigid body transformation. Figure 2 Add an edge between the two corresponding nodes and use the compatibility score as the edge weight.

[0073] After constructing the first-order compatibility graph, this invention further obtains a second-order compatibility weight matrix by performing multiple matrix multiplications on the first-order weight matrix. This matrix is ​​used to characterize the global consistency of matching point pairs over a larger scope. In this way, multiple sets of local compatibility relationships can be accumulated into the graph structure, making the real matching points present a highly connected cluster structure in the graph, while the edge weights of pseudo-matching points with other nodes are small and the connections are sparse.

[0074] On a geometric compatibility graph, this invention employs a maximal clique search algorithm to find fully connected subgraphs within the graph. Figure 2 A substructure consisting of several tightly connected nodes is a candidate maximal clique. Every two nodes within a maximal clique are connected by an edge, meaning these matching point pairs are geometrically compatible and can jointly interpret the same rigid body transformation. For each maximal clique, this invention further scores it based on edge weight accumulation and normal consistency, using the clique with higher scores and more stable structures as input for subsequent pose calculations. This effectively eliminates isolated nodes and weakly connected matching points, significantly improving the robustness of registration.

[0075] In the SVD-based pose solving and registration result evaluation module 503, after constructing a geometric compatibility graph and filtering maximal cliques, this invention obtains several sets of highly consistent matching point pairs. Each maximal clique represents a set of geometrically highly compatible correspondences. To calculate the rigid body transformation of the template point cloud relative to the field point cloud from these correspondences, a pose solving process based on singular value decomposition is adopted.

[0076] Specifically, for each maximal clique, the coordinates of all matching points within the clique are first extracted in the template point cloud and the actual point cloud. The centroids of the two sets of points are calculated, and the point coordinates are decentered so that the rotation solution depends only on the relative structure between the matching points. Then, a matrix for SVD is constructed using the covariance relationship between the matching points. This matrix is ​​then subjected to singular value decomposition to obtain a set of rotation matrices satisfying rigid body constraints and their corresponding translation vectors, thus obtaining the candidate rigid body transformations for that maximal clique. Repeating the above process for all maximal cliques yields multiple sets of candidate rotation and translation parameters.

[0077] This invention calculates a score for each maximal clique based on information such as edge weights and normal consistency in the aforementioned geometric compatibility graph. This score reflects the overall consistency and stability of the matching points within the clique. Based on this, the group with the highest score among all maximal cliques is selected as the final pose output. The selection process can be expressed as follows:

[0078] By employing this "cluster-by-cluster solution + scoring and selection" approach, this invention maintains the high efficiency of SVD closed-form solution while significantly reducing the impact of outliers and local mismatches on the final rotation and translation estimation, making the registration results obtained in railway emergency scenarios more stable and reliable.

[0079] In summary, this invention provides a method and system for relative pose estimation of point clouds in railway emergency scenarios. The method includes: mapping the input point cloud to a cylindrical coordinate system, performing spatial voxel partitioning and orientation alignment, generating voxel-level feature representations with rotation insensitivity through a multilayer perceptron and 3D convolution; constructing initial matching based on voxel features, building a geometric compatibility graph using distance difference and orientation difference, calculating first-order and second-order compatibility weights, and using maximal clique search to obtain highly consistent matching subsets; solving for rigid body transformations on matching point pairs within the maximal clique to obtain rotation matrices and translation vectors, and visualizing and evaluating multiple sets of experimental data to verify the effectiveness of the method. The method and system provided by this invention divide the point cloud around railway emergency equipment into cylindrical voxels, normalize the pose of the point cloud within the voxels using an orientation alignment matrix, and then extract voxel-level features using a multilayer perceptron and max pooling. This achieves rotation-invariant modeling of point clouds acquired from multiple views and with different poses, laying the foundation for subsequent stable matching. Point pairs obtained from voxel feature matching are treated as graph nodes. Edge weights are defined based on the distance difference and normal direction difference between point pairs, forming a first-order geometric compatibility graph. A second-order compatibility graph is obtained through multiple multiplications of the weight matrix, achieving unified modeling of the local and global geometric consistency between matched point pairs. A maximal clique satisfying complete connectivity is searched on the geometric compatibility graph. The edge weights within the clique are accumulated and combined with a normal consistency criterion to form a clique score, which is used to measure the geometric stability of the matched point set. This automatically filters out the matching subset with the highest structural consistency from a large number of candidate matches, significantly suppressing false matches and outliers. For the matched point set within each maximal clique, the rotation matrix and translation vector are solved using SVD, and the clique score function is utilized. The optimal rigid body transformation is selected from multiple candidate rigid body transformations, and the highest-scoring maximal clique is selected. As the final output, the robustness of pose estimation to mismatches and local missing values ​​is improved at the algorithm level.

[0080] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0081] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0082] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0083] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for estimating the relative pose of point clouds in a railway emergency scenario, characterized in that, include: S1. The template point cloud and the on-site emergency point cloud are centered, regionally clipped and voxelized. The local point cloud is mapped to spherical or cylindrical coordinate system, and rotation-invariant voxel features are constructed based on the orientation alignment matrix. S2. Three-dimensional convolution and multilayer perceptron are used to encode rotation-invariant pixel features, extract cross-view stable deep local structure features, and generate initial deep local structure feature matching pairs based on feature similarity. S3. Construct a geometric compatibility map based on distance consistency and orientation consistency using template point cloud and on-site emergency point cloud. Generate a weight matrix through first-order compatibility constraints and further calculate a second-order compatibility matrix to strengthen global structural constraints. S4. Perform a search for maximal cliques based on the compatibility graph, score the consistency of all deep local structural feature matching pairs within the clique, eliminate false matches that do not meet geometric constraints, and obtain a highly reliable matching set. S5. After performing step S4, use the singular value decomposition method to obtain the rotation and translation matrix for each maximal clique, and select the final pose result according to the maximal clique score to obtain the optimal rigid body transformation of the template point cloud relative to the emergency point cloud.

2. The method according to claim 1, characterized in that, The construction of rotationally invariant chromatids in step S1 includes: S11. Spatial division of template point cloud and on-site emergency point cloud, mapping local areas to spherical coordinates or cylindrical coordinates, and voxelization processing according to radius, angle and height dimensions; S12. Construct a local orientation alignment matrix for each voxel to unify the orientation of the point cloud inside the voxel and improve the consistency of voxel features across viewpoints. S13. Use a multilayer perceptron with shared weights to extract local features of point clouds within voxels, and obtain fixed-dimensional voxel descriptors through max pooling.

3. The method according to claim 1, characterized in that, The feature encoding process in step S2 includes: S21. All voxel features are combined according to spatial index to construct a four-dimensional voxel feature tensor, which is used to maintain the spatial topological relationship of the three-dimensional structure. S22. A three-dimensional convolutional network along the radial, pitch, and azimuth directions is used to extract deep geometric features. This three-dimensional convolutional network is implemented using the formula... Perform convolutional updates; where: j represents the voxel index in the radial direction, k represents the voxel index in the height or axial direction, l represents the voxel index in the azimuth or orientation direction, and s represents the network layer index, used to distinguish the feature maps of layer s from those of layer (s-1); superscript The input features from the previous layer are represented by d, which represents the feature channel index, i.e., the d-th feature dimension, used to describe the specific channel in the multi-channel feature mapping. d′ represents the channel index of the output feature after the current convolution or mapping operation, used to distinguish between the input and output feature dimensions. r represents the radial offset index of the convolution kernel, and its value range is determined by the radial convolution kernel size. The value of x is determined by the kernel width parameter, which describes the local receptive range within the radial neighborhood. The value of y represents the offset index of the convolution kernel in the height direction (axial direction), and its range is determined by the convolution kernel height parameter. The decision is as follows: D represents the total number of channels in the input feature, i.e., the dimension of the previous layer feature map in the channel dimension; F represents the feature mapping, which specifically includes: This represents the input feature of the d-th channel in the (s-1)-th layer. Indicates the position of the s-th layer at the voxel. The output feature value of the d-th channel, where w represents the learnable weight parameters of the 3D cylindrical convolution kernel, corresponding to the convolution weights for different channels and different spatial offset positions. These represent the receptive range or size parameters of the s-th layer 3D cylindrical convolution kernel in the radial, height, and angular dimensions, respectively; S23. Enhance the structural representation ability of key parts in template point cloud and on-site emergency point cloud by multi-scale regional feature aggregation operation, so as to improve the stability and noise resistance of the initial matching pair.

4. The method according to claim 3, characterized in that, The compatibility graph construction process in step S3 includes: S31. Based on the distance consistency as the edge weight, establish a first-order compatibility graph according to the distance difference of the deep local structural feature matching pairs. S32. Perform multiple matrix multiplication operations on the weight matrix of the first-order compatibility graph to obtain a second-order compatibility weight matrix, which is used to enhance global geometric consistency. S33. Based on the structural distribution of the second-order compatibility graph, remove deep local structural feature matching pairs with weak geometric constraints to obtain a set of candidate deep local structural feature matching pairs with strong geometric constraints.

5. The method according to claim 4, characterized in that, The maximal clique screening process in step S4 includes: S41. Search for maximal cliques that satisfy complete connectivity in the second-order compatibility graph and use them as a set of candidate reliable matches. S42. Accumulate the edge weights of the maximal clique to obtain the geometric consistency score of the maximal clique, which is used to measure its matching quality. S43. Based on the normal vector consistency constraint, perform secondary screening of matching points inside the maximal clique to eliminate false matching points with inconsistent local directions.

6. The method according to claim 5, characterized in that, The pose calculation process in step S5 includes: S51. For each pair of matching points inside a maximal clique after substep S43, use singular value decomposition to obtain the rotation matrix and translation vector. S52. Select the clique with the highest geometric consistency score based on the maximal clique as the source of the final pose output.

7. A point cloud relative pose estimation system for railway emergency scenarios, characterized in that, include: Rotation-insensitive voxel encoding module: Maps the input point cloud to a cylindrical coordinate system, performs spatial voxel division and orientation alignment, and generates voxel-level feature representations with rotation insensitivity through multilayer perceptron and 3D convolution; Geometric compatibility graph and maximal clique selection module: Initial matching is constructed based on voxel features, geometric compatibility graph is constructed using distance difference and orientation difference, first-order and second-order compatibility weights are calculated, and maximal clique search is used to obtain a highly consistent subset of matches; The SVD-based pose solving and registration result evaluation module performs rigid body transformation on the matched point pairs within the maximal clique to obtain the rotation matrix and translation vector. It also visualizes and evaluates multiple sets of experimental data to verify the effectiveness of the method of this invention.

Images