Disordered multi-view three-dimensional point cloud fine registration method based on recursive element shape generation

Abstract

The embodiment of the invention relates to the technical field of computer vision, and discloses a disordered multi-view three-dimensional point cloud fine registration method based on recursive element shape generation, which comprises the following steps: S1, combining all initial point cloud views into an unupdated element shape; s2, entering an iteration process of calling element shape-multi-view updating, and determining the number of iteration rounds included in the calling; s3, updating the meta shape which is not updated; s4, updating all the initial point cloud views; s5, combining all the updated point cloud views into a new element shape; s6, repeating the steps S3 to S5 until the maximum iteration round number corresponding to the current call is reached; s7, taking the updated meta shape of the current call, still taking all the initial point cloud views as non-updated point cloud views, starting a new call, and repeating the steps S2 to S6 until the maximum call frequency is reached; and S8, outputting all the point cloud views updated by the last call. According to the method, stable error convergence is realized, and local optimum is avoided.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and more particularly to a method for generating an unordered multi-view image based on a recursive shape element. Figure 3 dimensional point cloud precise registration method. Background Art

[0002] Compared with two-dimensional image data, three-dimensional point cloud data is directly sampled by three-dimensional sensors (such as lidar), and has the following advantages: First, with such a direct three-dimensional representation method of point coordinates, the sampling point group can express the three-dimensional shape and geometric relationship of objects in space. Two-dimensional images can only express the projected shape of objects through limited optical information and cannot restore the three-dimensional geometric form of objects; second, it has strong adaptability in complex environments. Even in poor light or severe weather conditions such as storms, high-quality three-dimensional point cloud data can still be obtained through lidar, while the imaging quality of two-dimensional images obtained under the same conditions is greatly reduced.

[0003] Disordered Multi-View Figure 3 The 3D point cloud registration technology is a technology that aligns or splices point clouds sampled from multiple perspectives under the condition of unknown LiDAR camera pose and sampling order to reconstruct a complete object or scene point cloud. Because this technology does not require information such as LiDAR pose and sampling sequence, it can more efficiently and simply perform 3D point cloud splicing and reconstruction. Figure 3 Dimensional point cloud registration technology is widely used in many application scenarios such as digital protection of cultural relics, autonomous driving, and remote sensing stereo maps, and has high practical value.

[0004] Disordered Multi-View Figure 3 The dimensional point cloud registration technology can be divided into two different technical routes, namely the traditional method based on spanning trees and graphs and the method based on deep learning. The traditional method based on spanning trees and graphs mainly evaluates the composition basis such as the overlap score between views, represents each point cloud view with a node, establishes a spanning tree or graph structure for all point cloud views to characterize the relationship between point cloud views, and then registers them in a certain order. This method relies on the effect of two-view registration in the composition process and subsequent registration links. There are problems such as poor registration effect and long composition time due to the influence of low overlap views. The method based on deep learning extracts and matches local and global features of point clouds through the network. Although it has a good registration effect with the help of the network's high-dimensional feature extraction and attention mechanisms, it requires a large amount of data training for specific scenes. When the scene scale changes and the geometric distribution is different from the training set, its registration success rate is greatly reduced. How to improve the generalization of the method based on deep learning has become an urgent problem to be improved. Summary of the invention

[0005] In view of this, the embodiment of the present application proposes an unordered multi-view image processing method based on recursive element shape generation. Figure 3 The 3D point cloud registration method recursively calls the meta-shape-multi-view iterative process, and timely widens the distance between the meta-shape and the point cloud view while gradually updating the meta-shape and the point cloud view, so as to achieve stable error convergence and avoid falling into the local optimum, thereby effectively improving the disordered multi-view Figure 3 The success rate and registration accuracy of 3D point cloud registration.

[0006] To achieve the above objectives, the embodiment of the present application proposes an unordered multi-view image processing method based on recursive element shape generation. Figure 3 The method for precise registration of dimensional point clouds comprises the following steps: S1, merging all initial point cloud views into an unupdated meta-shape; wherein the initial point cloud view is a roughly aligned point cloud view after rough registration; S2, entering an iterative process of calling meta-shape-multi-view update, and determining the number of iterative rounds of meta-shape-multi-view update included in this call according to the accumulated number of calls; S3, randomly sampling the unupdated meta-shape, selecting seed points, and updating the meta-shape based on the selected seed points to obtain an updated meta-shape; S4, performing point cloud registration on all initial point cloud views and the updated meta-shape in parallel, and All initial point cloud views are aligned to the updated meta-shape to complete the update of this round of point cloud views; S5, all updated point cloud views are merged into new meta-shapes to complete the update of this round of meta-shapes, and the complete meta-shape-multi-view update of this round is completed; S6, repeat S3 to S5 until the maximum number of iterations corresponding to this call is reached; S7, take the meta-shape updated in this call, still use all initial point cloud views as the unupdated point cloud views, start a new call, repeat S2 to S6, until the maximum number of calls is reached; S8, output all point cloud views updated after the last call, and complete the disordered multi-view Figure 3 dimensional point cloud precise registration.

[0007] Optionally, S2 specifically includes: S21, entering the iterative process of calling meta-shape-multi-view update, and calculating the number of iterative rounds of meta-shape-multi-view update included in this call based on the initial number of iterations, the cumulative number of calls and the incremental step of the number of iterations; S22, if this call is the first call, use the unupdated meta-shape to enter this call, if this call is not the first call, use the meta-shape updated in the previous call to enter this call; S23, after entering the meta-shape-multi-view update iteration, if this round of iteration is the first round of iteration, use all initial point cloud views to enter this round of iteration, if this round of iteration is not the first round of iteration, use all point cloud views updated in the previous round of iteration to enter this round of iteration.

[0008] Optionally, S21 is specifically implemented by the following formula: Among them, Iinit represents the initial iteration number, j represents the cumulative number of calls including this call, ΔI represents the incremental step of the iteration number, Indicates the number of iterations of the meta-shape-multi-view update included in this call.

[0009] Optionally, S3 specifically includes: S31, based on the number of iterations of this round of iterations, calculating the number of sampling points corresponding to this round of iterations, and randomly sampling the element shapes used in this round of iterations according to the number of sampling points corresponding to this round of iterations, and selecting seed points; the formula for calculating the number of sampling points corresponding to this round of iterations is: Among them, i represents the number of iterations in this round, N all represents the total number of points in all point cloud views, λ is the preset hyperparameter for controlling the sampling ratio, represents the number of sampling points corresponding to the current iteration calculated; S32, in the meta-shape used in the current iteration, a kNN nearest neighbor search is performed on each selected seed point, and k i The nearest neighbor points, k i =k init +(i-1)×Δk,k init represents the initial k value, Δk represents the increment step of the nearest neighbor point; S33, based on the k selected for each seed point i The nearest neighbor points constitute the local point cloud P local , the local point cloud P local The centroid point of the updated seed point is the updated meta-shape P. meta .

[0010] Optionally, S4 specifically includes: S41, parallelly comparing all point cloud views used in this round of iteration with the updated element shape P meta , use the ICP algorithm to perform precise point cloud registration, align all point cloud views to the updated primitive shape, and obtain the updated rotation parameters and translation parameters of each point cloud view; S42, adjust the posture of each point cloud view based on the updated rotation parameters and translation parameters of each point cloud view, and complete the update of this round of point cloud views.

[0011] Optionally, the rotation parameters and translation parameters of the point cloud view after update are expressed by the following formula:

[0012]

[0013] Among them, R represents the rotation parameter, t represents the translation parameter, and p z represents a point z in the point cloud view, N represents the total number of points in the point cloud view, and p meta Represents point z in the updated primitive shape P meta The corresponding points in .

[0014] Optionally, the maximum number of calls is 30.

[0015] To achieve the above objectives, the embodiment of the present application also proposes an unordered multi-view image processing method based on recursive element shape generation. Figure 3 The invention discloses a point cloud precision registration system, comprising: a preparation module M1, used to merge all initial point cloud views into an unupdated meta-shape, wherein the initial point cloud view is a roughly aligned point cloud view after rough registration; an iteration number determination module M2, used to enter the iterative process of calling meta-shape-multi-view update, and determine the number of iteration rounds of meta-shape-multi-view update included in this call according to the accumulated number of calls; a meta-shape first update module M3, used to randomly sample the unupdated meta-shape, select seed points, and update the meta-shape based on the selected seed points to obtain the updated meta-shape; a point cloud view update module M4, used to perform point cloud registration on all initial point cloud views and the updated meta-shape in parallel, register all initial point cloud views to the updated meta-shape, and complete the update of the point cloud view of this round; a meta-shape second ...5, used to randomly select seed points from the unupdated meta-shape, select seed points from the selected seed points, and update the meta-shape based on the selected seed points to obtain the updated meta-shape; a point cloud view update module M6, used to perform point cloud registration on all initial point cloud views and the updated meta-shape in parallel, register all initial point cloud views to the updated meta-shape, and complete the update of the point cloud view of this round; a meta-shape second update module M7, used to randomly select seed points from the unupdated meta-shape, select seed points from the selected seed points, and update the meta-shape based on the selected seed points to obtain the updated meta-shape; a meta-shape second update module M8, used to randomly select seed points from the unupdated meta-shape, select seed points The update module M5 is used to merge all updated point cloud views into a new meta-shape, complete the update of this round of meta-shapes, and thus complete the meta-shape-multi-view update of this round; the first discrimination module M6 is used to determine whether the maximum number of iterations corresponding to this call has been reached. If so, enter the second discrimination module M7, otherwise, return to the meta-shape first update module M3; the second discrimination module M7 is used to determine whether the maximum number of calls has been reached. If so, enter the final output module M9, otherwise, enter the recursive call module M8; the recursive call module M8 is used to take the meta-shape updated by this call, still use all initial point cloud views as the unupdated point cloud views, start a new call, and return to the iteration number determination module M2; the final output module M9 is used to output all point cloud views updated by the last call to complete the disordered multi-view. Figure 3 dimensional point cloud precise registration.

[0016] To achieve the above-mentioned purpose, an embodiment of the present application further proposes an electronic device, comprising: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the disordered multi-view image generation based on recursive element shapes as described above. Figure 3 dimensional point cloud precise registration method.

[0017] To achieve the above object, the embodiment of the present application further proposes a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the disordered multi-view image generated based on recursive element shapes as described above can be realized. Figure 3 dimensional point cloud precise registration method.

[0018] The embodiment of the present application provides an unordered multi-view image processing method based on recursive element shape generation. Figure 3 The method is a three-dimensional point cloud precision registration method. It merges all the roughly aligned initial point cloud views after coarse registration into unupdated meta-shapes, and indirectly eliminates the errors between all point cloud views through gradually growing meta-shapes, which increases the overlap rate between views and registration objects during point cloud registration, thereby improving the success rate and registration accuracy of view point cloud registration. It can still eliminate the errors between point cloud views relatively stably even when facing a low overlap rate between views, and does not rely on auxiliary structures such as spanning trees or compositions. It has strong robustness and universality. The method adopts a scheme of large-loop recursive calling and small-loop meta-shape-multi-view iterative updating. Through the small-loop meta-shape-multi-view iterative updating, the errors of all point cloud views are continuously eliminated, while the large-loop recursive calling only updates the meta-shapes, which timely widens the distance between the point cloud views and the meta-shapes, avoiding the problem of all point cloud views being lost. Figure 1 The vertical element shape registration falls into the dilemma of local optimality, which effectively improves the disordered multi-view Figure 3 In addition, the entire point cloud registration process does not involve neural networks and does not require any learning-based training steps, which effectively controls the computational cost and achieves a good balance between benefit and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the related technologies, the drawings required for use in the embodiments of the present application or the related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 An unordered multi-view image generated based on recursive element shapes is provided in one embodiment of the present application. Figure 3 Flowchart of the 3D point cloud precise registration method;

[0021] Figure 2 are all point cloud views with residuals provided as input in one embodiment of the present application;

[0022] Figure 3 is a schematic diagram of a meta-shape gradually growing with iterations provided in one embodiment of the present application;

[0023] Figure 4 is a schematic diagram of a point cloud view error being gradually eliminated with iterations provided in an embodiment of the present application;

[0024] Figure 5It is a multi-view point cloud effect diagram with error elimination output provided in one embodiment of the present application;

[0025] Figure 6 It is a comparison diagram of all point cloud views of a satellite simulation data set input and a multi-view point cloud effect diagram of error elimination of a satellite simulation data set output provided in one embodiment of the present application;

[0026] Figure 7 Another embodiment of the present invention provides an unordered multi-view image generated based on recursive element shapes. Figure 3 Schematic diagram of the structure of the 3D point cloud precise registration system;

[0027] Figure 8 It is a structural schematic diagram of an electronic device provided by another embodiment of the present application. DETAILED DESCRIPTION

[0028] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the embodiments of the present application will be described in detail below in conjunction with the accompanying drawings. However, it will be appreciated by those skilled in the art that in the present application, many technical details are proposed in order to enable the reader to better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical scheme claimed in the present application can also be implemented. The division of the following embodiments is for the convenience of description, and the specific implementation of the present application should not be construed as any limitation, and the various embodiments can be combined and referenced with each other under the premise of no contradiction.

[0029] In order to solve the problems existing in the traditional methods based on spanning trees and graphs and the methods based on deep learning, an embodiment of the present application proposes an unordered multi-view method based on recursive meta-shape generation. Figure 3 The method for accurate registration of a dimensional point cloud is applied to an electronic device, wherein the electronic device can be a terminal or a server. In this embodiment and the following embodiments, the electronic device is described by taking a server as an example. The following is an unordered multi-view image based on recursive element shape generation proposed in this embodiment. Figure 3 The following content is only for the convenience of understanding the implementation details provided, and is not necessary for the implementation of this solution.

[0030] This embodiment proposes a disordered multi-view based on recursive element shape generation Figure 3 The specific process of the dimensional point cloud precise registration method can be as follows Figure 1 As shown, including:

[0031] S1, merge all initial point cloud views into unupdated primitive shapes, where the initial point cloud views are roughly aligned point cloud views after coarse registration.

[0032] In the specific implementation, the server first needs to obtain all the initial point cloud views. The initial point cloud views are roughly aligned point cloud views after rough registration, but all the initial point cloud views still have residuals and disordered multi-view. Figure 3 The ultimate goal of 3D point cloud registration is to eliminate these residuals. After obtaining all the initial point cloud views, the server needs to merge all the initial point cloud views into an unupdated primitive shape, which is an imperfect and rough primitive shape.

[0033] In one example, all point cloud views with residuals in the input can be Figure 2 As shown, the server merges it into an unupdated metashape

[0034] S2, entering the iterative process of calling the meta-shape-multi-view update, and determining the number of iterative rounds of the meta-shape-multi-view update included in this call according to the accumulated number of calls.

[0035] In the specific implementation, after the server obtains the unupdated metashape, it can enter the iterative process of calling the metashape-multi-view update. This first requires determining the number of iterative rounds of metashape-multi-view update included in this call based on the accumulated number of calls. As the number of calls increases, the number of iterative rounds of metashape-multi-view update also increases accordingly.

[0036] In one example, S2 specifically includes:

[0037] S21, entering the iterative process of calling the meta-shape-multi-view update, and calculating the number of iterative rounds of the meta-shape-multi-view update included in this call based on the initial number of iterations, the accumulated number of calls and the incremental step of the number of iterations.

[0038] The number of iterations corresponding to this call can be calculated using the following formula:

[0039]

[0040] Among them, I init represents the initial iteration number, j represents the cumulative number of calls including this call, ΔI represents the incremental step of the iteration number, Indicates the number of iterations of the meta-shape-multi-view update included in this call.

[0041] I init It can be set to 5, ΔI can be set to 5. For the first call, a total of 5 iterations of meta-shape-multi-view update are required. For the second call, a total of 10 iterations of meta-shape-multi-view update are required.

[0042] S22, if this call is the first call, use the unupdated meta-shape to enter this call, if this call is not the first call, use the meta-shape updated in the previous call to enter this call.

[0043] For the first call, it is naturally necessary to use the unupdated meta-shape to enter the first call, and for the j-th call, it is necessary to use the meta-shape updated in the j-1-th call to enter the j-th call.

[0044] S23, after entering the meta-shape-multi-view update iteration, if this round of iteration is the first round of iteration, all initial point cloud views are used to enter this round of iteration; if this round of iteration is not the first round of iteration, all point cloud views updated in the previous round of iteration are used to enter this round of iteration.

[0045] After entering a called meta-shape-multi-view update iteration, for the first round of iteration, it is naturally necessary to use all initial point cloud views to enter the first round of iteration, and for the i-th round of iteration, it is necessary to use all point cloud views updated in the i-1th round of iteration to enter the i-th round of iteration.

[0046] S3, randomly sampling the unupdated meta-shapes, selecting seed points, and updating the meta-shapes based on the selected seed points to obtain updated meta-shapes.

[0047] In a specific implementation, after the server enters the first round of iteration, it first updates the unupdated meta-shapes, which requires random sampling of the unupdated meta-shapes, selecting seed points, and updating the meta-shapes based on the selected seed points to obtain updated meta-shapes.

[0048] In one example, S3 specifically includes:

[0049] S31, based on the number of iterations of this round, the number of sampling points corresponding to this round is calculated, and the element shapes used in this round are randomly sampled according to the number of sampling points corresponding to this round, to select seed points.

[0050] The formula for calculating the number of sampling points corresponding to this round of iteration is:

[0051]

[0052] Among them, i represents the number of iterations in this round, N all represents the total number of points in all point cloud views, λ is the preset hyperparameter for controlling the sampling ratio, Indicates the calculated number of sampling points corresponding to this round of iteration.

[0053] Generally speaking, λ can be set to 200.

[0054] S32, in the meta-shape used in this round of iteration, a kNN nearest neighbor search is performed on each selected seed point, and k i The nearest neighbor points, k i =k init +(i-1)×Δk,k init represents the initial k value, and Δk represents the incremental step size of the nearest neighbor point.

[0055] Generally speaking, k init It can be set to 3, and Δk can be set to 2. That is to say, in the first round of iteration, each seed point needs to select 3 nearest neighbor points, and in the second round of iteration, each seed point needs to select 5 nearest neighbor points.

[0056] S33, based on k selected for each seed point i The nearest neighbor points constitute the local point cloud P local , the local point cloud P local The centroid point of the updated seed point is the updated meta-shape P. meta .

[0057]

[0058] S4, perform point cloud registration on all initial point cloud views and the updated primitive shapes in parallel, register all initial point cloud views to the updated primitive shapes, and complete the update of this round of point cloud views.

[0059] In the specific implementation, for the first round of iterations, the server performs point cloud registration on all initial point cloud views with the updated meta-shapes in parallel, taking each initial point cloud view as the source point cloud and the updated meta-shape as the target point cloud, and registers all initial point cloud views to the updated meta-shapes to complete the update of this round of point cloud views.

[0060] In one example, S4 specifically includes:

[0061] S41, parallelize all point cloud views used in this iteration with the updated primitive shape P meta ,The ICP algorithm is used to perform precise point cloud registration, register all point cloud views to the updated element shape, and obtain the updated rotation and translation parameters of each point cloud view.

[0062] Among them, the ICP algorithm is used with a matching distance screening threshold of 4 times the point cloud resolution, the closest matching, and a maximum number of iterations not exceeding 20.

[0063] S42, adjusting the posture of each point cloud view based on the updated rotation parameters and translation parameters of each point cloud view, to complete the update of this round of point cloud views.

[0064] The rotation and translation parameters of the point cloud view after updating can be expressed by the following formula:

[0065]

[0066] Among them, R represents the rotation parameter, t represents the translation parameter, and p z represents a point z in the point cloud view, N represents the total number of points in the point cloud view, and p meta Represents point z in the updated primitive shape P meta The corresponding points in .

[0067] S5, merge all updated point cloud views into a new meta-shape, complete the meta-shape update of this round, and thus complete the meta-shape-multi-view update of this round.

[0068] In the specific implementation, after completing the update of this round of point cloud views, the server needs to update the meta-shape for the second time, that is, merge all updated point cloud views into new meta-shapes to complete the update of this round of meta-shapes, thus completing this round of complete meta-shape-multi-view update.

[0069] S6, repeat S3 to S5 until the maximum number of iterations corresponding to this call is reached.

[0070] In the specific implementation, after the server completes a round of iteration, it determines whether the accumulated number of iterations has reached the maximum number of iterations corresponding to this call. If it has reached it, the call is completed. Otherwise, it returns to S3 and repeats S3 to S5 until the maximum number of iterations corresponding to this call is reached.

[0071] In an example, Figure 3 , Figure 4 As shown, as the meta-shape-multi-view update is performed over multiple iterations, the meta-shape gradually grows with each iteration, and the errors between all point cloud views are gradually eliminated with each iteration.

[0072] S7, taking the updated primitive shape of this call, still taking all the initial point cloud views as the unupdated point cloud views, starting a new call, and repeating S2 to S6 until the maximum number of calls is reached.

[0073] In the specific implementation, after the current call is completed, the server needs to determine whether the cumulative number of calls has reached the preset maximum number of calls (generally set to 30 times). If it has reached it, the recursive call ends, otherwise, it returns to S2, takes the updated primitive shape of this call, still uses all the initial point cloud views as the unupdated point cloud views, starts a new call, and repeats S2 to S6 until the maximum number of calls is reached.

[0074] S8, output all point cloud views updated after the last call, and complete the disordered multi-view Figure 3dimensional point cloud precise registration.

[0075] In the specific implementation, after the recursive call is completed, the server outputs all point cloud views updated after the last call (such as Figure 5 As shown), the disordered multi-view is now complete Figure 3 dimensional point cloud precise registration.

[0076] In an example, the method proposed in this embodiment is applied to the actual multi-view point cloud precise registration, and the effect is as follows: Figure 6 As shown, from the performance of the method proposed in this embodiment on the satellite simulation data set, it can be seen that the method proposed in this embodiment can eliminate the residuals existing between multiple input point cloud views, and obtain a high-precision multi-view point cloud after alignment.

[0077] In this embodiment, all the roughly aligned initial point cloud views after rough registration are merged into unupdated meta-shapes, and the errors between all point cloud views are indirectly eliminated through the gradually growing meta-shapes, which increases the overlap rate between the view and the registration object during the point cloud registration process, thereby improving the success rate and registration accuracy of the view point cloud registration. Even when the overlap rate between views is low, the errors between point cloud views can still be eliminated relatively stably, and it does not rely on auxiliary structures such as spanning trees or compositions, and has strong robustness and universality. A scheme of large-loop recursive calls and small-loop meta-shape-multi-view iterative updates is adopted. Through the small-loop meta-shape-multi-view iterative updates, errors are continuously eliminated for all point cloud views. The large-loop recursive calls only update the meta-shapes, which timely widens the distance between the point cloud views and the meta-shapes, avoiding all point cloud views. Figure 1 The vertical element shape registration falls into the dilemma of local optimality, which effectively improves the disordered multi-view Figure 3 In addition, the entire point cloud registration process does not involve neural networks and does not require any learning-based training steps, which effectively controls the computational cost and achieves a good balance between benefit and efficiency.

[0078] The step division of the above methods is only for clear description. When implemented, they can be combined into one step or some steps can be split and decomposed into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this application; adding insignificant modifications to the algorithm or process or introducing insignificant designs without changing the core design of the algorithm and process are all within the scope of protection of this application.

[0079] Another embodiment of the present application proposes an unordered multi-view based on recursive element shape generation. Figure 3 The following is a random multi-view point cloud registration system based on recursive primitive shape generation proposed in this embodiment. Figure 3The following content is only the relevant implementation details provided for the convenience of understanding and is not necessary for the implementation of this solution.

[0080] The proposed unordered multi-view based on recursive element shape generation in this embodiment Figure 3 The specific structure of the 3D point cloud precise registration system can be described as follows: Figure 7 As shown, it includes: a preparation module M1, an iteration number determination module M2, a meta-shape first update module M3, a point cloud view update module M4, a meta-shape second update module M5, a first discrimination module M6, a second discrimination module M7, a recursive call module M8 and a final output module M9.

[0081] The preparation module M1 is used to merge all initial point cloud views into unupdated primitive shapes, wherein the initial point cloud views are roughly aligned point cloud views after rough registration.

[0082] The iteration round number determination module M2 is used to enter the iterative process of calling the meta-shape-multi-view update, and determine the number of iteration rounds of the meta-shape-multi-view update included in this call according to the accumulated number of calls.

[0083] The first meta-shape updating module M3 is used to randomly sample the meta-shapes that have not been updated, select seed points, and update the meta-shapes based on the selected seed points to obtain updated meta-shapes.

[0084] The point cloud view updating module M4 is used to perform point cloud registration on all initial point cloud views and the updated primitive shapes in parallel, register all initial point cloud views to the updated primitive shapes, and complete the update of this round of point cloud views.

[0085] The second meta-shape updating module M5 is used to merge all updated point cloud views into a new meta-shape to complete the meta-shape update of this round, thus completing the meta-shape-multi-view update of this round.

[0086] The first determination module M6 is used to determine whether the maximum number of iterations corresponding to this call has been reached. If so, it enters the second determination module M7; otherwise, it returns to the first meta-shape updating module M3.

[0087] The second determination module M7 is used to determine whether the maximum number of calls has been reached. If so, it enters the final output module M9; otherwise, it enters the recursive call module M8.

[0088] The module M8 is recursively called to obtain the updated primitive shape of this call, still using all initial point cloud views as unupdated point cloud views, start a new call, and return to the iteration round number determination module M2.

[0089] The final output module M9 is used to output all point cloud views updated after the last call, completing the disordered multi-view Figure 3 dimensional point cloud precise registration.

[0090] It is worth mentioning that all modules involved in this embodiment are logic modules. In practical applications, a logic unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, in order to highlight the innovative part of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed by this application, but this does not mean that there are no other units in this embodiment.

[0091] It is not difficult to find that this embodiment is a system embodiment corresponding to the above method embodiment, and this embodiment can be implemented in conjunction with the above method embodiment. The relevant technical details and technical effects mentioned in the above embodiments are still valid in this embodiment, and in order to reduce repetition, they are not repeated here. Accordingly, the relevant technical details mentioned in this embodiment can also be applied in the above embodiments.

[0092] Another embodiment of the present application provides an electronic device, the structure of which is as follows: Figure 8 As shown, it includes: at least one processor C1; and a memory C2 that is communicatively connected to the at least one processor C1; wherein the memory C2 stores instructions that can be executed by the at least one processor C1, and the instructions are executed by the at least one processor C1, so that the at least one processor C1 can perform a disordered multi-view image generated based on recursive element shapes as described in the above-mentioned method embodiments. Figure 3 dimensional point cloud precise registration method.

[0093] Among them, the memory and the processor can be connected in a bus manner, and the bus can include any number of interconnected buses and bridges, and the bus connects various circuits of one or more processors and memories together. The bus can also connect various other circuits such as peripherals, voltage regulators, and power management circuits together, which are well known in the art and are therefore not further described herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be one element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices on a transmission medium. The data processed by the processor is transmitted on a wireless medium via an antenna, and further, the antenna also receives data and transmits the data to the processor.

[0094] The processor is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory can be used to store data used by the processor when performing operations.

[0095] Another embodiment of the present application relates to a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the computer program can implement the disordered multi-view image generated based on recursive element shapes described in the above method embodiments. Figure 3 dimensional point cloud precise registration method.

[0096] That is, those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program is stored in a storage medium, including a number of instructions to enable a device (which can be a single-chip microcomputer, chip, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as: ROM), random access memory (Random Access Memory, referred to as: RAM), disk or optical disk and other media that can store program codes.

[0097] Those skilled in the art will appreciate that the above embodiments are specific embodiments for implementing the present application, and in actual applications, various changes may be made thereto in form and detail without departing from the spirit and scope of the present application.

Claims

1. A method for accurate registration of unordered multi-view 3D point clouds based on recursive primitive shape generation, characterized in that: include: S1, merge all initial point cloud views into an unupdated primitive shape; wherein the initial point cloud view is a roughly aligned point cloud view after coarse registration; S2, entering the iterative process of calling the meta-shape-multi-view update, and determining the number of iterative rounds of the meta-shape-multi-view update included in this call according to the accumulated number of calls; S3, randomly sampling the unupdated meta-shapes, selecting seed points, and updating the meta-shapes based on the selected seed points to obtain updated meta-shapes; S4, performing point cloud registration on all initial point cloud views and the updated primitive shapes in parallel, registering all initial point cloud views to the updated primitive shapes, and completing the update of the point cloud views in this round; S5, merging all updated point cloud views into a new meta-shape, completing the meta-shape update of this round, and thus completing the meta-shape-multi-view update of this round; S6, repeat S3 to S5 until the maximum number of iterations corresponding to this call is reached; S7, taking the updated primitive shape of this call, still taking all the initial point cloud views as the unupdated point cloud views, starting a new call, and repeating S2 to S6 until the maximum number of calls is reached; S8, outputs all point cloud views updated after the last call, and completes the precise registration of the unordered multi-view 3D point cloud.

2. The method for accurate registration of unordered multi-view 3D point clouds based on recursive primitive shape generation according to claim 1, characterized in that: S2 specifically includes: S21, entering the iterative process of calling the meta-shape-multi-view update, and calculating the number of iterative rounds of the meta-shape-multi-view update included in this call based on the initial number of iterations, the accumulated number of calls and the incremental step of the number of iterations; S22, if this call is the first call, then use the unupdated meta-shape to enter this call; if this call is not the first call, then use the meta-shape updated in the last call to enter this call; S23, after entering the meta-shape-multi-view update iteration, if this round of iteration is the first round of iteration, all initial point cloud views are used to enter this round of iteration; if this round of iteration is not the first round of iteration, all point cloud views updated in the previous round of iteration are used to enter this round of iteration.

3. The method for accurate registration of unordered multi-view 3D point clouds based on recursive primitive shape generation as claimed in claim 2, characterized in that: S21 is specifically implemented through the following formula: Among them, I init represents the initial iteration number, j represents the cumulative number of calls including this call, ΔI represents the incremental step of the iteration number, Indicates the number of iterations of the meta-shape-multi-view update included in this call.

4. The method for accurate registration of unordered multi-view 3D point clouds based on recursive primitive shape generation as claimed in claim 2, characterized in that: S3 specifically includes: S31, based on the number of iterations of this round, the number of sampling points corresponding to this round is calculated, and the element shape used in this round is randomly sampled according to the number of sampling points corresponding to this round, and the seed point is selected; the formula for calculating the number of sampling points corresponding to this round is: Among them, i represents the number of iterations in this round, N all represents the total number of points in all point cloud views, λ is the preset hyperparameter for controlling the sampling ratio, Indicates the calculated number of sampling points corresponding to this round of iteration; S32, in the meta-shape used in this round of iteration, a kNN nearest neighbor search is performed on each selected seed point, and k i The nearest neighbor points, k i =k init +(i-1)i×Δk, init represents the initial k value, Δk represents the incremental step size of the nearest neighbor point; S33, based on k selected for each seed point i The nearest neighbor points constitute the local point cloud P local , the local point cloud P local The centroid point of the updated seed point is the updated meta-shape P. meta .

5. The method for accurate registration of unordered multi-view 3D point clouds based on recursive primitive shape generation as claimed in claim 4, characterized in that: S4 specifically includes: S41, parallelize all point cloud views used in this iteration with the updated primitive shape P meta , use the ICP algorithm to perform point cloud precise registration, register all point cloud views to the updated primitive shape, and obtain the updated rotation and translation parameters of each point cloud view; S42, adjusting the posture of each point cloud view based on the updated rotation parameters and translation parameters of each point cloud view, to complete the update of this round of point cloud views.

6. The method for accurate registration of unordered multi-view 3D point clouds based on recursive primitive shape generation as claimed in claim 5, characterized in that: The updated rotation and translation parameters of the point cloud view are expressed by the following formula: Among them, R represents the rotation parameter, t represents the translation parameter, and p z represents a point z in the point cloud view, N represents the total number of points in the point cloud view, and p meta Represents point z in the updated primitive shape P meta The corresponding points in .

7. A method for accurate registration of unordered multi-view 3D point clouds based on recursive primitive shape generation according to any one of claims 1 to 6, characterized in that: The maximum number of calls is 30.

8. A system for accurately registering unordered multi-view 3D point clouds based on recursive primitive shape generation, characterized in that: include: A preparation module M1, used for merging all initial point cloud views into unupdated primitive shapes, wherein the initial point cloud views are roughly aligned point cloud views after coarse registration; The iteration round number determination module M2 is used to enter the iterative process of calling the meta-shape-multi-view update, and determine the number of iteration rounds of the meta-shape-multi-view update included in this call according to the accumulated number of calls; The first meta-shape updating module M3 is used to randomly sample the meta-shapes that have not been updated, select seed points, and update the meta-shapes based on the selected seed points to obtain updated meta-shapes; The point cloud view updating module M4 is used to perform point cloud registration on all initial point cloud views and the updated primitive shapes in parallel, register all initial point cloud views to the updated primitive shapes, and complete the update of the point cloud views in this round; The second meta-shape updating module M5 is used to merge all updated point cloud views into a new meta-shape to complete the meta-shape update of this round, thus completing the meta-shape-multi-view update of this round; The first determination module M6 is used to determine whether the maximum number of iterations corresponding to this call has been reached. If so, it enters the second determination module M7; otherwise, it returns to the first meta-shape update module M3; The second determination module M7 is used to determine whether the maximum number of calls has been reached. If so, it enters the final output module M9, otherwise, it enters the recursive call module M8; Recursively call module M8 to obtain the updated primitive shape of this call, still use all initial point cloud views as unupdated point cloud views, start a new call, and return to the iteration round number determination module M2; The final output module M9 is used to output all the point cloud views updated by the last call, and complete the precise registration of the disordered multi-view 3D point cloud.

9. An electronic device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a method for precise registration of unordered multi-view three-dimensional point clouds based on recursive primitive shape generation as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it is possible to implement a method for precise registration of unordered multi-view three-dimensional point clouds based on recursive primitive shape generation as described in any one of claims 1 to 7.

Images