Nonlinear registration optimization method for fastener point cloud

Abstract

Disclosed in the present invention is a nonlinear registration optimization method for a fastener point cloud, comprising the following steps: S1, collecting raw data; S2, preprocessing the data; S3, transforming frame images into a deformation graph; S4, performing registration between the frame images; and S5, simplifying corrected point cloud data. The method of the present invention improves the accuracy and confidence score of raw point cloud data, and provides a more reliable data input for subsequent point cloud processing and higher-level interpretation processes, thereby ensuring the fulfillment rate of point cloud processing objectives. The point cloud optimization process of the present invention exhibits high robustness, preserves effective features of the raw point cloud data as much as possible while removing noise points and outliers, and highly preserves fine details. Additionally, the nonlinear optimization process avoids getting trapped in local minima, exhibits high computing efficiency, demonstrates high universal applicability, and has broad potential for practical application.

Description

A nonlinear registration optimization method for fastener point clouds Technical Field

[0001] This invention belongs to the field of rapid integrated inspection of rail transit, and in particular relates to a nonlinear registration optimization method for fastener point clouds. Background Technology

[0002] Track fasteners are a crucial component of rail transit engineering, and their condition directly impacts train operational safety. Over long-term operation and maintenance, the service condition of fasteners changes, affecting the line's geographical location and smoothness requirements. With the development of high-speed rail and urban rail transit, the need for detecting abnormal fastener conditions is increasing. However, during fastener inspection, the point cloud data, the most critical input, often suffers from deviations exceeding the typical redundancy range due to environmental noise, hardware distortion, human error, partial fastener failure, and complex target area conditions. This directly affects the accuracy of fastener condition assessment and the timeliness of maintenance decisions, ultimately impacting railway safety. Therefore, data correction and feature optimization of the collected fastener point clouds are urgently needed.

[0003] There are three main types of existing targeted point cloud optimization algorithms:

[0004] (1) Point cloud optimization algorithm based on mathematical statistics concepts;

[0005] (2) Point cloud optimization algorithm based on features of each dimension;

[0006] (3) Point cloud optimization algorithm based on spatial domain state.

[0007] The above method has the following problems:

[0008] (1) The algorithm has a single optimization dimension, weak descriptive coverage, and is prone to introducing additional errors or mistakes.

[0009] (2) The algorithm process is poorly coupled with the target point cloud and has poor targeting.

[0010] (3) The solution process is repetitive and wastes a lot of computing power, resulting in low time efficiency and a large workload.

[0011] To address the aforementioned problems and weaknesses, there is an urgent need to develop a nonlinear registration optimization technique for fastener point clouds. This technique can comprehensively, efficiently, and accurately process and optimize the collected fastener point cloud data, providing more reliable and trustworthy data input for subsequent point cloud understanding and advanced processing. Summary of the Invention

[0012] To solve the problems in the prior art, the application provides a nonlinear registration optimization method for fastener point clouds, which can comprehensively, efficiently and accurately process and optimize the collected fastener point cloud data, and provide more reliable and credible data input for subsequent point cloud understanding and advanced processing.

[0013] To this end, the application adopts the following technical solutions:

[0014] 1. A nonlinear registration optimization method for fastener point clouds, characterized by comprising the following steps:

[0015] S1, original data acquisition: using a collection device to collect point cloud data containing spatial coordinates and timestamp information in the active axial direction, the depth axial direction and the offset axial direction; the active axial direction is the tangent direction of the curve of the movement of the collection device, and the active axial direction, the depth axial direction and the offset axial direction are perpendicular to each other;

[0016] S2, data preprocessing: dividing the point cloud data collected in S1 into a plurality of frame images according to the timestamp information in the form of spatial slices according to the time axis;

[0017] S3, converting the frame images into deformation maps: S i ={p i ,χ i};

[0018] Wherein, S i is the frame image of the i-th frame, which is composed of a point set p i and an edge set χ i ; n i is the total number of points in the frame image of the i-th frame, m i is the total number of edges in the frame image of the i-th frame, i is the number of frame images, i∈[1,I], and I is the total number of frame images;

[0019] S4, registration between frame images:

[0020] First, a registration mapping relationship E wicp between frame images is constructed, and E wicp =min(E pr +β wicp E nreg +γ wicp E lor );

[0021] Wherein: E pr is an alignment description item; E nreg is a regularization description item; E lor is a local transformation item; β wicp and γ wicp are index weights, and the βwicp and the value of said γ wicp depends on the degree of differentiation between said frame images, said β wicp and γ wicp have a value range of (0, 1),

[0022] Then, the best registration relationship is obtained according to the registration mapping relationship E wicp , and all frame images obtained by S2 are registered according to the best registration relationship, so as to obtain the modified point cloud data.

[0023] S5, simplifying the modified point cloud data, comprising the following steps:

[0024] S51, establishing an initial ellipsoid space:

[0025] First, a point p tag in the modified point cloud data obtained by S4 is taken as the ellipsoid center of the ellipsoid space; the initial value of the length of the long axis of the ellipsoid space is twice the distance between the ellipsoid center and the nearest neighbor point in the projection of the offset axis, and the initial value of the length of the short axis of the ellipsoid space is twice the distance between the ellipsoid center and the nearest neighbor point in the projection of the depth axis.

[0026] Wherein, the nearest neighbor point is the point in the modified point cloud data with the smallest Euclidean distance from the ellipsoid center;

[0027] Then, the direction of the long axis of the ellipsoid is parallel to the offset axis, and the direction of the short axis is parallel to the depth axis, and the two end points of the long axis are O au and O ad respectively; the two end points of the short axis are O bu and O bd respectively, to obtain the initial ellipsoid space.

[0028] Finally, an ellipsoid space set is constructed, and the initial ellipsoid space is added as the first element in the set.

[0029] S52, ellipsoid space diffusion: the ellipsoid spaces in the ellipsoid space set are added or reconstructed, and the S4 modified point cloud data covered by the ellipsoid spaces are simplified.

[0030] S53, repeating step S52 until the ellipsoid spaces in the ellipsoid space set cover all the S4 modified point cloud data, and the obtained point cloud is output as the high-quality point cloud data after nonlinear registration.

[0031] The alignment description item E pr in S4 is calculated by the following method:

[0032] Wherein, fwel (x) is a Welsch function; is a point in the frame image of the i-th frame, y is the number of the point in the frame image of the i-th frame; y ∈ [1, n i ]; is the nearest target point in the frame image to be registered of , the frame to be registered is set as the j-th frame, j ∈ [1, I]; c wel is a constant greater than 0 and making the f wel monotonically increasing on the positive half-axis.

[0033] The regularization description item E nreg is calculated by the following method:

[0034] wherein c′ wel is the updated constant item in the Welsch function, c′ wel is a constant greater than 0 and making the f wel monotonically increasing on the positive half-axis. denotes the set of adjacent points of the point in the frame image to be registered, z is the number of the point in the frame image of the j-th frame, z ∈ [1, n j ], n j is the total number of points in the frame image of the j-th frame. is the transformation form of each point in the deformation graph considering the spatial cross coupling.

[0035] The regularization description item is calculated by the following method:

[0036] wherein, is the affine rotation matrix of the embedded deformation graph; and are the affine translation vectors of the embedded deformation graph corresponding to the i-th frame and the j-th frame respectively.

[0037] The local transformation item E lor is calculated by the following method:

[0038] wherein η proj is a modification factor considering the local transformation intensity, and its value range is (0, 1); the operator symbol ‖·‖ F is the F norm solution considering the differential local rigidity state.

[0039] S52 specifically comprises the following steps: S521, for each ellipsoid space in the ellipsoid space set, a point p k is randomly determined outside the ellipsoid space, and the p kFor each point p in the modified point cloud data obtained in S4 k S522 is executed when it satisfies the following formula, otherwise S523 is executed,

[0040] wherein, is the normal vector of the plane formed by the long axis and the short axis of the ellipse and is the distance threshold of point p dis to the projection plane, ξ k is the vector deviation threshold; the values of said ξ ang and ξ dis are determined according to the fastener to be registered, and the projection plane is p ang the corresponding ellipse center O k of the ellipsoid space; k the plane where the corresponding frame is located;

[0041] S522, reconstructing the ellipsoid space:

[0042] integrating point p k into its corresponding ellipsoid space, and the ellipse center O k remains unchanged, the length of the long axis of the corresponding ellipsoid space is changed to a' el , the length of the short axis is changed to b' el , and the reconstructed ellipsoid space is obtained, which has:

[0043] wherein, O kau is the long axis endpoint of the ellipsoid space with O k as the center, and O kbu is the short axis endpoint of the ellipsoid space with O k as the center;

[0044] all point cloud data covered in the reconstructed ellipsoid space are changed to the ellipse center point O k , the two endpoints of the long axis of the reconstructed ellipsoid space, the two endpoints of the short axis of the reconstructed ellipsoid space, and the p k ;

[0045] Step S53 is executed;

[0046] S523, constructing an ellipsoid space:

[0047] First, the diffusion direction of the ellipsoid space is determined according to the position of the point p k and the ellipse center O k ;

[0048] Then, the nearest neighbor point p k of the point p w is determined in the diffusion direction; the pw As the ellipse center of the ellipsoid space, an additional ellipsoid space is added, the length of the long axis of the ellipse of the additional ellipsoid space is the length of the long axis of the ellipse of the point p k The length of the long axis of the ellipse of the additional ellipsoid space is the length of the long axis of the ellipse of the point p k The length of the long axis of the ellipse of the additional ellipsoid space is the length of the long axis of the ellipse of the point p

[0049] The additional ellipsoid space is added to the ellipsoid space set, and step S53 is performed.

[0050] The diffusion direction in S523 is determined by the following formula:

[0051] Wherein, qua(p k ) represents the quadrant where the point p k is located.

[0052] Preferably, the acquisition device in S1 is a self-moving rail transit mobile three-dimensional scanning system.

[0053] Compared with the prior art, the present application has the following beneficial effects:

[0054] 1. The present application is directed to the original point cloud quality problem which is ignored in the current fastener evaluation work, and a nonlinear registration optimization algorithm is proposed, the data correction process of the point cloud is quantized into three types of description items, and the data accuracy and confidence of the original point cloud are further improved, more reliable data input is provided for the subsequent point cloud processing and advanced understanding process, and then the realization rate of the point cloud processing target is ensured.

[0055] 2. The method of the present application combines the time and space continuity of the collected point cloud data, and mainly considers the local difference and global uniformity, the detail rigidity distinction and the overall non-rigid structure, proposes a nonlinear target optimization equation, uses three types of description items, and further improves the accuracy, coverage and confidence of the optimization process.

[0056] 3. The method of the present application designs an ellipsoid containing space, realizes point cloud simplification and optimization in the form of ellipse diffusion, further improves the availability, simplicity and description of the point cloud data, and matches the needs of the subsequent point cloud processing and advanced understanding process; The point cloud optimization process of the present application has high robustness, and on the basis of removing noise points and random points, the effective features of the original point cloud data are preserved as much as possible, and the detail retention is strong. At the same time, the nonlinear optimization process avoids falling into the local optimization trap, has high calculation efficiency, strong universality, and high application value. BRIEF DESCRIPTION OF DRAWINGS

[0057] Fig. 1 is a flowchart of the present application;

[0058] Fig. 2 is a schematic diagram of point cloud data composed of each subdivided frame data of the data preprocessing process in the present application;

[0059] Fig. 3 is a schematic diagram of ellipsoid space set in the present application. DETAILED DESCRIPTION

[0060] The technical solutions of the present application are described in detail below in combination with the drawings and examples.

[0061] In the process of collecting digital point cloud, a point cloud coordinate system is defined, which includes three coordinate axes, namely depth axis direction, offset axis direction and active axis direction, wherein the active axis direction is the curve tangent direction of the motion of the collection system, the depth axis direction and the offset axis direction are perpendicular to the active axis direction, and the depth axis direction and the offset axis direction are also in a perpendicular relationship.

[0062] As shown in Fig. 1, the nonlinear registration optimization method of the fastener point cloud of the present application includes the following steps:

[0063] S1, original data collection: using a self-moving rail transit mobile three-dimensional scanning system, or using other collection systems and devices to collect and obtain point cloud data containing spatial coordinates and timestamp information under the aforementioned point cloud coordinate system.

[0064] S2, data preprocessing: according to the timestamp information obtained in S1, the complete point cloud is divided into several frame images in the form of spatial slices according to the time axis, and the plane composed of the depth axis and the offset axis is the plane of the point cloud slice as shown in Fig. 2.

[0065] S3, converting the frame image into a deformation graph: introducing an embedded deformation graph in graph theory into the frame image, that is, the frame image S i is described as a set composed of a point set p i and an edge set χ i , n i , m i are the maximum number of points and the maximum number of edges of the frame image of the i-th frame, respectively, i is the frame number in the frame image, i∈[1, I], I is the total number of frame images, then: S i ={p i , χ i};

[0066] S4, registration between frame images: registering each frame image divided in S2 in a nonlinear mapping relationship, considering the spatial continuity and the subtle degree of apparent feature differentiation in the data collection process, a registration process and optimization target with higher difference tolerance should be considered, including the following steps:

[0067] S41, establishing a registration mapping equation: a registration mapping relationship E is established between the frame images of the i-th frame and the j-th frame wicp , the j-th frame is a frame to be registered of the i-th frame, and there is: wicp pr +β wicp E nreg +γ wicp E lor ;

[0068] wherein E pr is an alignment description item; E nreg is a regularization description item; E lor is a local transformation item; β wicp and γ wicp are index weights, and the values are determined according to the degree of difference between the frame images, the value range of β wicp and γ wicp is (0, 1), and j ∈ [1, I];

[0069] S42, calculating the alignment description item E pr , and there is:

[0070] wherein f wel (x) is a Welsch function; is a point in the frame image of the i-th frame, y is the number of the point in the frame image of the i-th frame, and y ∈ [1, n i ]; is the nearest target point in the frame image to be registered of ; c wel is a constant greater than 0 and making f wel monotonically increasing on the positive half axis;

[0071] S43, calculating the regularization description item E nreg , and there is:

[0072] wherein c′ wel is a set constant item in the Welsch function, c′ wel is a constant greater than 0 and making f wel monotonically increasing on the positive half axis; indicates a set of adjacent points of the point in the frame image to be registered, z is the number of the point in the frame image of the j-th frame, z ∈ [1, n j ], and n j is the total number of the points in the frame image of the j-th frame; is a transformation form of each point in the deformation graph considering spatial cross coupling;

[0073] wherein,​ affine rotation matrix of the embedded morphing map; and and

[0074] S44, calculating the local transformation term E lor , where:

[0075] where η proj is a modification factor considering the local transformation intensity, and its value range is (0, 1); the operator symbol ‖·‖ F is the F-norm solution considering the differential local rigidity state;

[0076] S45, solving the nonlinear optimization formula: according to the registration mapping relationship E wicp of step S41, solving the best registration relationship between all the to-be-registered frame images of the highest confidence fastener point cloud, and correcting the positions of the adjacent frame images according to the relationship; obtaining the corrected point cloud data.

[0077] S5, simplifying the optimized point cloud: after the nonlinear optimization of the fastener point cloud as described in step S4, additional point cloud topological structures, point-line-edge relationships, and possible spatial outliers, data jitter, and local burrs will be introduced, which will affect the accuracy of subsequent point cloud processing, so the point cloud result processed in S4 needs to be simplified using an ellipsoid space. The spatial dimension optimization process of the point cloud data is converted into a growth process of multiple base spaces into a feature-inclusive space, and generalized elliptical parameter diffusion is performed, including the following steps:

[0078] S51, establishing an initial ellipsoid space: taking a point p tag in the corrected point cloud data in S4 as the center of the ellipsoid space; the initial values of the lengths of the major axis and the minor axis of the ellipsoid space are twice the distances between the center of the ellipsoid space and the nearest neighbor points in the offset axis direction and the depth axis direction, respectively, and the nearest neighbor point is the point in the point cloud data that has the smallest Euclidean distance from the center of the ellipsoid. The major axis direction of the ellipsoid in the ellipsoid space is parallel to the offset axis direction, and the minor axis direction is parallel to the active axis direction. The two end points of the major axis of the ellipsoid space are O au and O ad , respectively; the two end points of the minor axis of the ellipsoid space are O bu and O bd , respectively; and an initial ellipsoid space is obtained.

[0079] An ellipsoid space set is constructed, and the initial ellipsoid space is added as the first element in the set.

[0080] S52, ellipsoid space diffusion, including the following steps:

[0081] S521, randomly determine point p within the point cloud data processed by S4. k , and p k It is a point outside each ellipsoidal space in the ellipsoidal space set, for each point p k If the following condition is met, execute S522; otherwise, execute S523.

[0082] in, The normal vector of the plane formed by the major and minor axes of the ellipse and ξ dis For p k The distance threshold to the projection surface, ξ ang ξ is the threshold for vector sway. dis and ξ ang The value is determined based on the fastener to be registered, and the projection plane is p. k The center O of the ellipse in the corresponding ellipsoidal space k The plane containing the corresponding frame; p k The actual center O of the ellipse k If there are multiple ellipsoidal spaces, then a point p must be taken outside each ellipsoidal space. k Therefore, each p here k Each has a corresponding ellipse center O k .

[0083] S522, Reconstructing the ellipsoidal space:

[0084] Point p k It is integrated into its corresponding ellipsoidal space, and the center of the ellipse O k Remaining unchanged, the lengths of the major and minor axes of the corresponding ellipsoidal space are changed to point p. k The length a′ projected onto the major axis of the ellipse el and the length b′ projected onto the minor axis el The reconstructed ellipsoidal space is obtained; where,

[0085] O kau For O k O is the endpoint of the major axis of the ellipsoid centered at a circle. kbu For O k The endpoint of the minor axis of the ellipsoid centered at the circle;

[0086] For the reconstructed ellipsoidal space, p k Located on its surface, all point cloud data within the reconstructed ellipsoidal space are changed to the ellipse's center point O. k The two endpoints of the major axis of the reconstructed ellipsoidal space, the two endpoints of the minor axis of the reconstructed ellipsoidal space, and p k .

[0087] performing step S53;

[0088] S523, building an ellipsoid space;

[0089] According to the position of point p k , the diffusion direction of the ellipsoid space is determined:

[0090] wherein, qua(p k ) represents the quadrant in which point p k is located.

[0091] The nearest neighbor point p k of p w is determined in the obtained diffusion direction; p w is taken as the ellipsoid center of the ellipsoid space to build an additional ellipsoid space, the lengths of the major axis and the minor axis of the ellipsoid of the additional ellipsoid space are the lengths of the major axis and the minor axis of the ellipsoid of the point p k , the lengths of the major axis and the minor axis of the additional ellipsoid space are projected, and all point cloud data in the additional ellipsoid space are retained;

[0092] The additional ellipsoid space is added to the ellipsoid space set, and step S53 is performed.

[0093] S53, repeating step S52 until the ellipsoid spaces in the ellipsoid space set cover all the point cloud data modified in S4, and outputting the obtained point cloud as the high-quality point cloud data after nonlinear registration. The process of reconstruction or building of the ellipsoid space is shown in FIG. 3.

[0094] The method of the present application realizes the registration and alignment of the local details of the collected point cloud and the average of the data errors through the establishment of three types of description items under the perspective of global optimization. On the one hand, the collected point cloud is modified by the multiple frame point clouds segmented therefrom, avoiding the blurring of local details and the appearance of data holes, and improving the data accuracy of the collected point cloud; on the other hand, the error average between the multiple frame point cloud data is realized in the point cloud registration process guided by the nonlinear optimization idea, improving the data confidence of the collected point cloud.

[0095] In the process of nonlinear optimization, the collected point cloud is regarded as a finite subdivided complex rigid body set composed of multiple simple structure rigid bodies by referring to the spatiotemporal continuity of the point cloud data in the collection process. Therefore, in view of the two major contradictions of local difference and global uniformity, detail rigidity and overall non-rigidity, the problems of (1) data difference tolerance difference introduced by measurement error, (2) low gradient descent resolution introduced by single optimization perspective, and (3) weak spatial structure similarity introduced by point cloud sparsity are mainly considered in the optimization process, and the to-be-fitted term in the nonlinear optimization objective equation is further decomposed and quantified into three types of description items, respectively improving the accuracy, coverage ability and confidence index in the optimization process.

[0096] The algorithm establishes a deep coupling relationship between the collected point cloud and the target point cloud through the three types of description items designed in the optimization process, and the mutual independent optimization structure ensures the robustness of the algorithm. At the same time, the algorithm retains as much as possible the effective features of the collected point cloud data on the basis of removing noise and disorder points, and has strong detail retention. In addition, the traditional single equation gradient descent process is analyzed into a three-way calculation process in the iterative optimization process of the algorithm, and the intermediate results can be reused, further improving the calculation efficiency and avoiding falling into the local optimization trap. The above processes are point cloud feature representation and optimization description, and there is no additional requirement and limitation on the basic attributes and data distribution of the point cloud, so the universality is strong, and it can be further applied to the non-rigid target registration scene, realizing human body reconstruction, high-resolution reconstruction, high-detail virtual reality, etc.

Claims

1. A method of nonlinear registration optimization of a point cloud of a fastener, comprising: The method comprises the following steps: S1, original data acquisition: using a collection device to collect point cloud data containing spatial coordinates and timestamp information in the active axial, depth axial and offset axial; the active axial is the curve tangent direction of the movement of the collection device, and the active axial, the depth axial and the offset axial are perpendicular to each other in pairs; S2, data preprocessing: dividing the point cloud data collected in S1 into several frame images according to the timestamp information in the form of spatial slices according to the time axis; S3, converting the frame images into deformation maps: S i = {p i ,χ i}; wherein S i is a frame image of the i-th frame, composed of a point set p i and an edge set χ i ; n i is the total number of points in the frame image of the i-th frame, m i is the total number of edges in the frame image of the i-th frame, i is the number of the frame image, i∈[1, I], and I is the total number of frame images. S4, registration between frame images: First, the registration mapping relationship E between frame images is constructed wicp There are: E wicp = min(E pr + β wicp E nreg + γ wicp E lor ); wherein: E pr is an alignment description term; E nreg is a regularization description term; E lor is a local transformation term; β wicp and γ wicp are index weights, the values of said β wicp and said γ wicp depending on the degree of differentiation between the frame images, said β wicp and γ wicp ranging between (0, 1), Then, according to the registration mapping relationship E wicp An optimal registration relationship is obtained, and all frame images obtained in S2 are registered according to the optimal registration relationship, to obtain corrected point cloud data. S5, simplifying the corrected point cloud data, comprising the following steps: S51, establishing an initial ellipsoid space: First, a point p in the modified point cloud data obtained by S4 tag an ellipsoid space; a length of an ellipsoid long axis of the ellipsoid space is initially set as twice a distance between the ellipsoid center and a nearest neighbor point projected in the offset axis direction, and a length of an ellipsoid short axis of the ellipsoid space is initially set as twice a distance between the ellipsoid center and the nearest neighbor point projected in the depth axis direction; Wherein, the nearest neighbor point is the point in the corrected point cloud data with the minimum Euclidean distance from the center of the ellipse; Then, the direction of the long axis of the ellipse is parallel to the shift axis, and the direction of the short axis is parallel to the depth axis, and the two endpoints of the long axis of the ellipse are O au and O ad ; the two endpoints of the short axis of the ellipse are O bu and O bd ; to obtain an initial ellipsoid space; Finally, an ellipsoid space set is constructed, and the initial ellipsoid space is added as the first element in the ellipsoid space set; S52, ellipsoid space diffusion: adding or reconstructing the ellipsoid space in the ellipsoid space set, and simplifying the S4 corrected point cloud data covered by the ellipsoid space; S53, repeating step S52 until the ellipsoid space in the ellipsoid space set covers all the S4 corrected point cloud data, and outputting the obtained point cloud as high-quality point cloud data after nonlinear registration.

2. The method of claim 1, wherein: The alignment description item E in S4 pr This is calculated by the following method: wherein f wel (x) is the Welsch function; for a point within the frame image of the ith frame, y is the number of the point in the frame image of the ith frame; y e [1, n] i ]; ​ is the nearest target point in the frame image to be registered, setting the frame to be registered as the jth frame, j ∈ [1, I]; c wel is greater than 0 and such that the f wel is a constant that is monotonically increasing on the positive half-axis.

3. The method of claim 2, wherein: The regularization description term E in S4 nreg The following method is used for calculation: where c' is a constant term in the Welsch function updated as wel c' = c' + 0.1 wel is a constant greater than 0 and such that the f wel is monotonically increasing on the positive half-axis; representing a point in the frame image to be registered a set of neighboring points of the point z, z is a point number in the frame image of the jth frame, z e [1, n j ], n j is a total number of points in the frame image of the jth frame; The transformation form of each point in the deformation map is considered for spatial cross coupling.

4. The nonlinear registration optimization method for the point cloud of the fastener according to claim 3, characterized in that: The regularization description term The following methods were used to calculate: wherein affine rotation matrix for an embedded morphed image; and The affine translation vectors of the embedded deformation maps corresponding to the i-th frame and the j-th frame, respectively.

5. The method of claim 4, wherein: The local transformation term E lor The local transformation term E is calculated by the following method: where η proj is a modification factor considering the local transformation severity, which takes a value in the range (0, 1); the operator sign ||·||F F is the F-norm solution considering the differential local rigid state.

6. The method of claim 1, wherein: S52 specifically comprises the following steps: S521, for each ellipsoid space in the ellipsoid space set, randomly determining a point p outside the ellipsoid space k , the p k is a point in the modified point cloud data obtained in S4, for each point p k , when it satisfies the following formula, S522 is performed, otherwise S523 is performed, wherein, a normal vector to a plane formed by the major axis and the minor axis of the ellipse and ξ dis for a point p k distance threshold to the projection plane, ξ ang for a vector yaw threshold; the ξ dis and ξ ang values are determined according to the fastener to be registered, the projection plane being p k the corresponding ellipsoid space's ellipsoid center O k the plane in which the corresponding frame lies; S522, reconstructing the ellipsoid space: The point p k is merged into its corresponding ellipsoidal space, and the ellipsoidal space has an ellipsoidal center O k unchanged, the length of the long axis of the corresponding ellipsoidal space is changed to a' el , and the length of the short axis of the ellipsoid is changed to b' el , to obtain the reconstructed ellipsoidal space, and has: wherein O kau is the long axis endpoint of the ellipsoidal space with O k as the center, O kbu is the short axis endpoint of the ellipsoidal space with O k as the center; Change all the point cloud data covered in the reconstructed ellipsoid space to the ellipsoid center point O k , the two end points of the long axis of the reconstructed ellipsoid space, the two end points of the short axis of the reconstructed ellipsoid space and the p k ; Step S53 is performed; S523, adding the ellipsoid space: First, the position of the point p k with the center O of the ellipse k of the ellipsoid space: Then, a nearest neighbor point p k of the point p w is determined in the diffusion direction; the point p w is added as an ellipsoid space of an ellipsoid center to build an additional ellipsoid space, an ellipsoid major axis of the additional ellipsoid space has a length of a length of the point p k projected on an ellipsoid major axis of the additional ellipsoid space; and an ellipsoid minor axis of the additional ellipsoid space has a length of a length of the point p k projected on an ellipsoid minor axis of the additional ellipsoid space, all point cloud data in the additional ellipsoid space is reserved. The additional ellipsoid space is added to the ellipsoid space set, and step S53 is performed.

7. The method of claim 6, wherein: The diffusion direction described in S523 is determined by the following equation: wherein qua(p k ) denotes the quadrant in which point p k lies.

8. The method of claim 1, wherein: The collection device in S1 is a self-moving rail transit mobile three-dimensional scanning system.

Images