A method, system and device for constructing a three-dimensional model of a multi-level structure of a pump house

Abstract

The present application relates to the field of image processing, and more particularly to a kind of pump house multi-level structure three-dimensional model construction method, system and equipment, comprising: point cloud set is divided into point cloud grouping using preset window size, point cloud grouping of target point cloud set is aligned to point cloud grouping of reference point cloud set, the preferred window size of target point cloud set;All point cloud sets are divided into several groups for fusion and three-dimensional model construction, the preferred window size of different point cloud sets in each group has maximum difference, and point cloud set has no aligned point cloud grouping, is contained in the point cloud grouping of other point cloud set alignment.This application avoids the problem that different positions of point cloud cannot be accurately aligned and fused in the presence of shielding in pump house, and further cannot restore the details of pump equipment structure.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a method, system and device for constructing a multi-level structure three-dimensional model of a pump house. BACKGROUND

[0002] When upgrading or implementing intelligent management on an existing pump house, a BIM model of the pump house needs to be constructed, and constructing a three-dimensional model of the pump house is an important part. A fast three-dimensional model construction method is to use a laser radar to scan each position in the existing pump house to obtain a three-dimensional point cloud of the pump house, and then construct a three-dimensional model of the pump house based on the three-dimensional point cloud.

[0003] The obtained three-dimensional model not only needs to include the building model of the pump house, but also needs to include the models of water pumps, pipelines, electrical equipment and the like, so as to ensure that the three-dimensional model of the pump house is multi-level; in addition, the three-dimensional model of the pump house needs to be able to interact with the user, for example, the model needs to be independently highlighted, the model needs to be bound with animation or special effects, real-time display of device parameters (such as water pump power, pipeline pressure, power line voltage, etc.), interactive actions in response to user mouse events, etc., so the three-dimensional model needs to include various details, such as the details of the water pump and the pipeline, the connection details between the water pump and the pipeline and the pipeline and the pipeline, the details of the electrical equipment (such as the buttons, switches and wiring of the power cabinet), etc.

[0004] However, the devices such as pipelines and water pump units in the pump house are severely obstructed, which makes it impossible to accurately align and fuse the point clouds collected at all positions together; or the aligned point clouds have a large error that cannot be ignored, especially when considering that the three-dimensional model needs to retain many details, there may be a problem that the structure formed by the error point cloud is misjudged as a three-dimensional model detail structure, which is not conducive to the accurate construction of the three-dimensional model of the pump house. SUMMARY

[0005] To solve the above problems, the present application provides a method, system and device for constructing a multi-level structure three-dimensional model of a pump house.

[0006] The method, system and device for constructing a multi-level structure three-dimensional model of a pump house provided by the present application adopt the following technical solutions:

[0007] An embodiment of the present application provides a method for constructing a multi-level structure three-dimensional model of a pump house, which comprises the following steps:

[0008] Point cloud sets are collected at different positions of the pump house, and a point cloud set is designated as a reference point cloud set; any point cloud set other than the reference point cloud set is recorded as a target point cloud set;

[0009] The target point cloud set and the reference point cloud set are divided into several point cloud groups by using a preset window size, the point cloud groups of the target point cloud set are aligned into the point cloud groups of the reference point cloud set by using a transformation matrix, and the aligned point cloud groups and the non-aligned point cloud groups of the target point cloud set under the window size are obtained; wherein the difference between the alignment error of the non-aligned point cloud groups and the alignment error of the aligned point cloud groups is recorded as the alignment index under the window size; the window size when the alignment index is maximum is recorded as the optimal window size of the target point cloud set;

[0010] All the point cloud sets are divided into several groups, the optimal window sizes of different point cloud sets in each group have maximum difference, and any one point cloud set in each group has no aligned point cloud group under the optimal window size, and other point cloud sets outside the point cloud set contained in each group have aligned point cloud groups under the optimal window size;

[0011] All the point cloud sets in each group are fused together and three-dimensional model construction is performed.

[0012] Preferably, the specific steps of aligning the point cloud groups of the target point cloud set into the point cloud groups of the reference point cloud set by using a transformation matrix, and obtaining the aligned point cloud groups and the non-aligned point cloud groups of the target point cloud set under the window size, include the following steps:

[0013] All the point cloud groups of the target point cloud set and all the point cloud groups of the reference point cloud set are matched to obtain all the matching pairs; the parameters of translation and rotation transformation when the point cloud groups from the target point cloud set are translated and rotated to the point cloud groups from the reference point cloud set in each matching pair are recorded as the transformation parameters of each matching pair;

[0014] Among all the transformation parameters of the matching pairs, the transformation parameters that have minimum difference between several transformation parameters constitute a target category; the mean value of the transformation parameters in the target category is recorded as a target transformation parameter; the point cloud groups in the matching pairs corresponding to all the transformation parameters in the target category are the aligned point cloud groups, and the point cloud groups outside the aligned point cloud groups are the non-aligned point cloud groups;

[0015] The target transformation parameter is the parameter of the translation matrix and the rotation matrix used when the point cloud groups in the target point cloud set are aligned into the point cloud groups in the reference point cloud set.

[0016] Preferably, the specific steps of obtaining the alignment error of the non-aligned point cloud groups and the alignment error of the aligned point cloud groups include the following steps:

[0017] For any one aligned or non-aligned point cloud group in the target point cloud set, a translation matrix and a rotation matrix corresponding to the target transformation parameter are used to perform translation and rotation transformation on all point clouds in the point cloud group to obtain transformed point clouds; the nearest distance between the transformed point clouds and the point clouds in the reference point cloud set is recorded as the average first distance of the point cloud group; for all aligned or non-aligned point cloud groups in the target point cloud set, the average of the average first distances of all aligned or non-aligned point cloud groups is recorded as the alignment error of the aligned or non-aligned point cloud group.

[0018] Preferably, the all point cloud sets are divided into several groups, the preferred window size of different point cloud sets in each group has the maximum difference, and any one point cloud set in each group has no aligned point cloud group at the preferred window size, and the other point cloud sets outside the point cloud set in each group are included in the aligned point cloud group at the preferred window size. The specific steps include the following:

[0019] D1: performing initial classification on all point cloud sets to obtain all initial categories, for each point cloud set x in the same initial category, and for each point cloud set y outside x in the same initial category; the point clouds in the non-aligned point cloud group of x at the preferred window size are included in the aligned point cloud group of y at the preferred window size;

[0020] D2: for any one initial category, performing KMeans clustering on the preferred window size of all point cloud sets in the initial category to obtain all first categories;

[0021] D3: for any one initial category corresponding to all first categories, selecting and removing one preferred window size from each first category respectively, so that the variance of the selected preferred window sizes in all first categories is maximum, and the point cloud set corresponding to the selected preferred window size in all first categories is taken as a group;

[0022] D4: repeating the process in D3 until the selected preferred window size in all first categories is less than two, and obtaining several groups.

[0023] Preferably, the all point cloud sets are classified initially to obtain all initial categories, and the specific steps include the following:

[0024] Traverse all point cloud sets, and record the traversed point cloud set as s1, and put s1 into an empty set S;

[0025] Traverse all point cloud sets other than s1, and record the i-th point cloud set as ; judge whether the point clouds in the non-aligned point cloud group of each point cloud set in S at the preferred window size are respectively included in the aligned point cloud group of the point cloud set at the preferred window size. In the point cloud group aligned under the preferred window size, if yes, Put into the set S; when all point cloud sets except s1 are traversed, set S as an initial class;

[0026] When all point cloud sets s1 are traversed, all initial classes are obtained.

[0027] Preferably, the specific steps of the step of respectively merging all cloud sets in each group together include the following steps:

[0028] D5: For all point cloud sets of each group, first merge the first point cloud set into the reference point cloud set to obtain a merged point cloud set; then merge the second point cloud set into the merged point cloud set to obtain the merged point cloud set again; then merge the third point cloud set into the merged point cloud set to obtain the merged point cloud set again; and so on until all point cloud sets of each group are merged into the merged point cloud set, and the obtained merged point cloud set is recorded as the point cloud merging result of each group;

[0029] D6: The point cloud merging results of all groups are regarded as a group, and the cloud merging results of all groups are merged together again according to the method in D5 to obtain the final point cloud merging result for constructing a three-dimensional model.

[0030] Preferably, the specific steps of the step of merging the first point cloud set into the reference point cloud set to obtain a merged point cloud set include the following steps:

[0031] The first point cloud set is divided into a plurality of point cloud groups by using the preferred window size of the first point cloud set, and the reference point cloud set is also divided into a plurality of point cloud groups by using the preferred window size of the first point cloud set; the point cloud groups of the first point cloud set are aligned to the point cloud groups of the reference point cloud set by using the transformation matrix; and the aligned first point cloud set and the reference point cloud set are merged as a merged point cloud set; and in the merged point cloud set, the point clouds with a Euclidean distance less than a preset alignment threshold are merged.

[0032] Preferably, the specific steps of the step of selecting the transformation parameter with the smallest difference between the plurality of transformation parameters as the target class include the following steps:

[0033] The transformation parameters of all matching pairs are normalized, the normalized transformation parameters of all matching pairs are subjected to mean shift clustering, all classes are obtained, the class with the most transformation parameters is recorded as the target class of all matching pairs, the Euclidean distance between all normalized transformation parameters in the target class is calculated, and the mean value is obtained, and the transformation error of all matching pairs is obtained.

[0034] For all matching pairs of which the transformation error is the smallest, the target category of all matching pairs of which the transformation error is the smallest is taken as the target category of the transformation parameters that have the smallest difference.

[0035] Another embodiment of the present application provides a system for constructing a three-dimensional model of a multi-level structure of a pump house, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes all steps of the above-mentioned method for constructing a three-dimensional model of a multi-level structure of a pump house when executing the computer program.

[0036] Still another embodiment of the present application provides a device for constructing a three-dimensional model of a multi-level structure of a pump house, which comprises a drone carrying a laser radar for collecting a point cloud set at each position in the pump house, and a computer for reading the point cloud set at each position, and the computer creates a three-dimensional model by executing the above-mentioned system for constructing a three-dimensional model of a multi-level structure of a pump house after reading the point cloud set at all positions.

[0037] The technical scheme of the present application has the following beneficial effects:

[0038] The present application aligns the point cloud groups of the target point cloud set to the point cloud groups of the reference point cloud set by using the transformation matrix, and obtains the point cloud groups of the target point cloud set that are aligned and the point cloud groups of the target point cloud set that are not aligned under the window size; the difference between the alignment error of the point cloud groups that are not aligned and the alignment error of the point cloud groups that are aligned is recorded as the alignment index under the window size; the window size under which the alignment index is the largest is recorded as the preferred window size of the target point cloud set. The preferred window size can reasonably divide the target point cloud set and the reference point cloud set, and the point cloud groups divided under the window size can accurately align the same pump equipment, avoid the mixing of different pump equipment or different parts of the same pump equipment, and help restore the detailed structure of the pump equipment under the occlusion scene.

[0039] Further, the application divides all point cloud sets into several groups, the preferred window size of different point cloud sets in each group has the maximum difference, and any point cloud set in each group has no aligned point cloud group at the preferred window size, and the point cloud group aligned at the preferred window size of other point cloud sets in each group is contained in the point cloud group. On the one hand, when the corresponding multiple preferred window sizes in the group have large differences, the diversity of the window size can be ensured, so that the point cloud groups divided based on different scales and different levels of window can be aligned and matched from different detail levels of the pump house equipment, and the error of the point cloud sets in different positions during alignment and matching is reduced; on the other hand, the point cloud group that cannot be aligned in any point cloud set in the same group is contained in the aligned point cloud group of other point cloud sets, so that the details of the pump house equipment that cannot be collected at certain positions are contained in the point cloud data collected at other positions, so that the point cloud group that cannot be aligned can also participate in alignment and matching in other point cloud sets, avoiding the problem that the point cloud sets in different positions cannot be effectively aligned and matched due to the shielding problem, especially reducing the error of the pump house equipment detail area affected by the distance or the angle during alignment and matching.

[0040] In summary, the application divides the point cloud sets by using appropriate windows, and aligns and matches the point clouds in appropriate positions respectively and fuses them, so that the problem that the point clouds in different positions cannot be accurately aligned and fused due to the shielding in the pump house, and the details of the pump house equipment cannot be restored, is avoided. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0042] Figure 1 A step flow chart of a pump house multi-level structure three-dimensional model construction provided by an embodiment of the application. DETAILED DESCRIPTION

[0043] In order to further illustrate the technical means and effects adopted by the application to achieve the predetermined application purpose, the following will combine the drawings and the preferred embodiments to specifically describe the construction method, system and device of the pump house multi-level structure three-dimensional model according to the application, the specific implementation, structure, features and effects thereof in detail. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0044] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0045] The application provides a pump house multi-level structure three-dimensional model construction method, system and device.

[0046] Embodiment one:

[0047] Please refer to Figure 1 , which shows a step flow chart of a pump house multi-level structure three-dimensional model construction method provided by an embodiment of the application, and the method comprises the following steps:

[0048] Step S101, respectively collecting point cloud sets at different positions of the pump house.

[0049] In this embodiment, considering that the equipment such as pipelines and water pump units in the pump house is seriously shielded, a staff member needs to hold a laser radar (or a laser radar carried by a drone according to a set trajectory) to scan at different positions in the pump house, so as to avoid areas that cannot be scanned due to shielding.

[0050] Specifically, when the laser radar moves to scan, the pump house space is scanned once every 0.5 seconds, and a point cloud set at the position of the laser radar is obtained. After the laser radar moves to scan, all position point cloud sets are obtained. In this embodiment, all point cloud coordinates in each point cloud set are normalized. As an example, the normalization method is that the three-dimensional coordinates of each point cloud are divided by the length, width and height of the pump house. In addition, in order to facilitate description, each object in the pump house is recorded as a pump house equipment (including building entities and water pumps, pipelines, electrical equipment and the like).

[0051] Step S102, for the reference point cloud set and the target point cloud set, the target point cloud set and the reference point cloud set are divided into a plurality of point cloud groups by using a preset window size, the point cloud groups of the target point cloud set are aligned into the point cloud groups of the reference point cloud set by using a transformation matrix, and the point cloud groups that are aligned and the point cloud groups that are not aligned of the target point cloud set under the window size are obtained.

[0052] A conventional method of the application is to align all position point cloud sets together by matrix transformation, and to obtain the point cloud distribution of all positions in the pump house.

[0053] For example, in a comparative embodiment, all point cloud sets are respectively transformed and aligned into a fixed position point cloud set, and then the point cloud distribution of all positions in the pump house is obtained.

[0054] However, this method has the following problems: on the one hand, due to the occlusion problem, the point cloud sets at different positions may contain point clouds of different pump house equipment or different parts of the same pump house equipment; on the other hand, in order to obtain the details on the three-dimensional model, the point cloud distribution density on each pump house equipment is large (that is, the point cloud distribution is affected by the detailed structure of the pump house equipment), and the point cloud distribution of the same pump house equipment at different positions (such as different distances or different angles) may have large point cloud distribution differences. The existence of the above two aspects leads to the problem that the point clouds cannot be accurately transformed and aligned.

[0055] For another example, in another comparative example, the point clouds at two adjacent positions are transformed and aligned together, the transformed and aligned point clouds are transformed and aligned with the point clouds at the next adjacent position, and so on, so that all the point clouds at different positions are matched together.

[0056] In this example, although the point cloud distribution difference caused by large changes in distance or angle at different positions can be avoided to a certain extent, error accumulation may occur in this example, and finally the problem of inaccurate alignment and matching of point clouds still exists under the influence of the above two aspects.

[0057] In summary, due to the occlusion problem and the need to consider the details of the three-dimensional model, the conventional method cannot accurately align all the point cloud sets at different positions together, or the point clouds aligned together have large errors, especially in this embodiment, since the three-dimensional model contains many details, the structure formed by the error point clouds may be misjudged as a detail structure of the three-dimensional model, which is not conducive to the accurate construction of the three-dimensional model of the pump house.

[0058] In order to solve the above problems, in this embodiment, first, a point cloud set collected at a position is specified as a reference point cloud set; then, a point cloud set at any position is recorded as a target point cloud set (not including the reference point cloud set), the target point cloud set and the reference point cloud set are divided into a plurality of point cloud groups by using a preset window, and the point cloud groups of the target point cloud set are aligned into the point cloud groups of the reference point cloud set by using a transformation matrix.

[0059] The purpose of obtaining the reference point cloud set is to facilitate the subsequent alignment and matching of all the point cloud sets at different positions to the reference point cloud set. The purpose of grouping the point cloud sets is that for two point cloud sets at different positions, the pump house equipment contained in the two point cloud sets may be different, and the two point cloud sets cannot be aligned and matched as a whole, and then the alignment and matching of the local point clouds are performed by window division in this embodiment.

[0060] As an example, one of the point cloud sets at all positions is artificially specified as the reference point cloud set.

[0061] As another example, the method for obtaining the reference point cloud set is:

[0062] For all point clouds in any point cloud set, DBSCAN clustering is performed on the point clouds to obtain all classes of the point cloud set. Each class represents a structure on the pump house equipment. In the point cloud sets at all positions, the point cloud set with the largest number of classes is obtained, which contains most of the structures in all pump house equipment, and is taken as the reference point cloud set. The DBSCAN clustering is a known technology, and the present embodiment is not specifically limited.

[0063] As an example, the target point cloud set and the reference point cloud set are respectively divided into a plurality of point cloud groups using a preset window, including the following method:

[0064] For any point cloud set in the target point cloud set and the reference point cloud set, a minimum cuboid bounding box of all point clouds in the point cloud set is obtained, i.e. all point clouds are contained in the bounding box. The bounding box is equally divided into a plurality of regular cubes, and each regular cube is used as a window. In the present embodiment, the side length of each window is set to 0.1. It should be noted that the dimension and order of magnitude of the window size are not considered in the present embodiment because the point clouds in the cloud set have been normalized in advance. Specifically, when the bounding box cannot be equally divided, i.e. the area divided at the boundary of the bounding box is not a regular cube of the window size, the area divided at the boundary of the bounding box is still regarded as a window. Specifically, when the number of point clouds in the window is less than 15, the window is deleted.

[0065] All point clouds in each window are used as a point cloud group, and the window length used when dividing the window is used as the window size when dividing the point cloud group.

[0066] As an example, the point cloud groups of the target point cloud set are aligned to the point cloud groups of the reference point cloud set using a transformation matrix, including the following method:

[0067] (1) Randomly match all point cloud groups of the target point cloud set and all point cloud groups of the reference point cloud set to obtain all matching pairs corresponding to the matching method. Each matching pair contains two point cloud groups, one from the target point cloud set and the other from the reference point cloud set. Meanwhile, when a point cloud group is in a matching pair, the point cloud group does not appear in other matching pairs.

[0068] For any one matching pair, where the point cloud group from the target point cloud set is denoted as GA, where the point cloud group from the reference point cloud set is denoted as GB, the translation matrix and the rotation matrix when all the point clouds in GA are transformed to all the point clouds in GB are solved by using the RANSAC algorithm, and the translation matrix and the rotation matrix are respectively flattened into one-dimensional vectors and spliced together, denoted as the parameter of the transformation matrix, and denoted as the transformation parameter for short.

[0069] (2) The transformation parameters of all the matching pairs are normalized (specifically, the same dimension of all the transformation parameters is linearly normalized), the mean shift clustering is performed on the normalized transformation parameters of all the matching pairs, all the categories are obtained, the transformation parameters in the same category are the same or similar, the category containing the most transformation parameters is obtained, denoted as the target category, and the target category indicates that most of the matching pairs have the same or similar transformation method (or have the same or similar transformation matrix).

[0070] The mean shift clustering is a known technology, and the bandwidth used in the mean shift clustering is set to 0.2 in this embodiment.

[0071] For all the normalized transformation parameters in the target category, the Euclidean distance between the transformation parameters is calculated, and the mean of the obtained Euclidean distance is denoted as the transformation error of all the matching pairs. The larger the transformation error is, the more difficult it is to align all the matching pairs using a unified transformation parameter, and thus it is more difficult to accurately align all the matching pairs.

[0072] (3) At this point, a transformation error is obtained after the point cloud groups of the target point cloud set and the point cloud groups of the reference point cloud set are matched by using a matching method. The matching method of the point cloud groups of the target point cloud set and the point cloud groups of the reference point cloud set is obtained by using the genetic algorithm when the transformation error is the smallest. For the target category corresponding to all the matching pairs under the matching method, the mean of all the (unnormalized) transformation parameters contained in the target category is denoted as the target transformation parameter. For the point cloud groups in the matching pairs corresponding to all the transformation parameters in the target category, the point cloud groups are aligned point cloud groups, and the point cloud groups other than the aligned point cloud groups are non-aligned point cloud groups.

[0073] The aligned point cloud groups refer to the point clouds in the target point cloud set and the reference point cloud set that can describe the same structure of the pump house equipment. The non-aligned point cloud groups refer to the point clouds at different pump house equipment or different structures of the same pump house equipment due to the difference between the corresponding positions of the target point cloud set and the reference point cloud set.

[0074] So far, the above obtains the transformation parameters used when the point cloud group of the target point cloud set is aligned to the point cloud group of the reference point cloud set under a given window size (i.e. the target transformation parameters), and the aligned point cloud group and the non-aligned point cloud group. The translation matrix and the rotation matrix corresponding to the target transformation parameters can transform the point cloud in the target point cloud set to the reference point cloud set, thereby realizing the alignment of the target point cloud set and the reference point cloud set.

[0075] In step S103, the difference between the alignment error of the non-aligned point cloud group and the alignment error of the aligned point cloud group is recorded as the alignment index under the window size; the window size when the alignment index is maximum is recorded as the preferred window size of the target point cloud set.

[0076] The alignment error C1 of the aligned point cloud group is obtained, and the alignment error C2 of the non-aligned point cloud group is obtained. The larger C1 is, the more difficult it is to accurately align the target point cloud set and the reference point cloud set on the same pump house equipment by dividing the point cloud group according to the above window size; the smaller C1 is, the more easily it can be aligned. The larger C2 is, the more it can ensure that the target point cloud set and the reference point cloud set do not overlap on different pump house equipment by dividing the point cloud group according to the above window size, and can avoid that the point clouds of different pump house equipment or different parts of the same pump house equipment are mixed together.

[0077] In this embodiment, the difference between C2 and C1 is taken as the alignment index when aligning under the above window size. The larger the alignment index is, the more the non-overlapping situation of the point clouds between the target point cloud set and the reference point cloud set on different pump house equipment or different parts of the same pump house equipment is far greater than the point cloud pairing error of the same part of the same pump house equipment when aligning the point cloud group divided according to the above window size, at this time it is explained that the purpose of "the aligned point cloud can be accurately aligned, and the non-aligned point cloud is not aligned" can be achieved.

[0078] As an example, the alignment error C1 of the aligned point cloud group is obtained by the following method:

[0079] For any one aligned point cloud group in the target point cloud set, the translation matrix and the rotation matrix corresponding to the target transformation parameters are used to perform translation and rotation transformation on all point clouds in the point cloud group, to obtain the transformed point clouds; for any one of the transformed point clouds, and for the point cloud in the reference point cloud set closest to the point cloud in the Euclidean distance, the Euclidean distance between the two point clouds is recorded as the first distance, the average of the first distances corresponding to all transformed point clouds is calculated, and the average first distance of any one aligned point cloud group is recorded. For all aligned point cloud groups in the target point cloud set, the average of the average first distances of all aligned point cloud groups is recorded as the alignment error C1.

[0080] The alignment error C2 of the point cloud groups without alignment is calculated.

[0081] As an example, the alignment index is equal to C2-C1.

[0082] So far, after grouping the point clouds with a given window size, the alignment index of the target point cloud set under the window size is obtained.

[0083] Further, the embodiment needs to be explained that when the window size is not appropriate, the alignment process above will be affected, resulting in too small alignment index. Specifically, when the window size is too large, each point cloud group divided from the target point cloud set compared with each point cloud group divided from the reference point cloud set, may result in that part of the point clouds in the former corresponding to the pump house equipment structure does not appear in part of the point clouds in the latter, while another part in the former corresponding to the pump house equipment structure appears in part of the point clouds in the latter, which leads to difficulty in aligning the target point cloud set and the reference point cloud set together according to the divided point cloud groups, that is, the target transformation parameter obtained above is inaccurate. When the window size is too small, on the one hand, too much attention is paid to the details of the pump house equipment structure, and the point cloud error collected at the structure details between different positions is large (for example, the point cloud error of the same detail structure under different distances and angles is large), on the other hand, there are not enough point cloud samples participating in the RANSAC algorithm; these two aspects lead to that the target transformation parameter may also be inaccurate.

[0084] Based on this, in the embodiment, different window sizes are used to perform the steps S102 and S103 above respectively, so as to obtain the alignment index of the same target point cloud set under each window size, and the window size when the alignment index is maximum is taken as the preferred window size of the target point cloud set.

[0085] The preferred window size describes that the window size can reasonably divide the target point cloud set and the reference point cloud set, the point cloud groups divided under the window size can accurately align the same pump house equipment, avoid the case that different pump house equipments or different parts of the same pump house equipment are mixed together, and help to restore the detail structure of the pump house equipment under the occlusion scene.

[0086] As an example, the values of different window sizes are respectively: 0.1, 0.15, …, 0.95, 1.0.

[0087] Step S104, divide all point cloud sets into several groups, the preferred window size of different point cloud sets in each group has the maximum difference, and any one point cloud set in each group under the preferred window size has no aligned point cloud group, and the other point cloud sets except the point cloud set in each group are included in the aligned point cloud group under the preferred window size.

[0088] For each point cloud set of each position, the preferred window size of each point cloud set is obtained according to the above method respectively, and the target transformation parameter, the aligned group and the non-aligned group of each point cloud set under the preferred window size are obtained correspondingly.

[0089] In this embodiment, the point cloud sets of all positions are divided into a plurality of groups, and each group contains a plurality of point cloud sets.

[0090] The preferred window sizes of different point cloud sets in each group have the maximum difference, and for any one point cloud set g in each group, the non-aligned point cloud group of g under the preferred window size is contained in the aligned point cloud groups of other point cloud sets except g in each group.

[0091] In this embodiment, the point cloud sets of all positions are divided into a plurality of groups, and the purpose is to subsequently align and match the plurality of point cloud groups in each group together, avoiding the cumulative error when the point cloud sets of all positions are aligned together in sequence.

[0092] In this embodiment, the preferred window size has the maximum difference, and the non-aligned point cloud group is contained in the aligned point cloud group, and the purpose is that, on the one hand, when the corresponding plurality of preferred window sizes in the group have large differences, the diversity of the window size can be ensured, so that the point cloud groups divided based on windows of different scales and different levels can be aligned and matched from different detail levels of the pump house equipment, reducing the error when the point cloud sets of different positions are aligned and matched; on the other hand, the non-aligned point cloud group in any one point cloud set in the same group is contained in the aligned point cloud group of other point cloud sets, which ensures that the details of the pump house equipment that cannot be collected at some positions are contained in the point cloud data collected at other positions, so that the non-aligned point cloud group can also participate in the alignment and matching in other point cloud sets, avoiding the problem that the point cloud sets of different positions cannot be effectively aligned and matched due to the shielding problem, especially reducing the error of the pump house equipment detail area affected by the distance or the angle of view during the alignment and matching.

[0093] As an example, the point cloud sets of all positions are divided into a plurality of groups, including the following methods:

[0094] First of all, it should be pointed out that all positions described in this example do not include the position corresponding to the reference point cloud set.

[0095] (1) Firstly, all the point clouds of all positions are initially classified to obtain all initial classes, and multiple point clouds are contained in the same initial class; for each point cloud set x in the same initial class and for each point cloud set y outside x in the same initial class; the point clouds of x in the non-aligned point cloud group under the preferred window size are contained in the aligned point cloud group of y under the preferred window size.

[0096] (2) For any one initial class, the preferred window size of all point cloud sets in the initial class is clustered by KMeans, and the number of classes of KMeans clustering is set to 5 in this embodiment. All classes obtained by clustering are referred to as first classes; the preferred window sizes in the same first class are the same or similar.

[0097] (3) For all first classes corresponding to any one initial class, a preferred window size is selected and removed from each first class, so that the variance of the selected preferred window sizes in all first classes is maximum, and at this time the point cloud set corresponding to the selected preferred window size in all first classes is taken as a group. The purpose of the maximum variance is to ensure that the multiple preferred window sizes are significantly different, thereby ensuring the diversity of window sizes.

[0098] (4) Repeat the process in (3) until the selected preferred window size of all first classes is less than two, at which time all point cloud sets of all positions are divided into several groups.

[0099] Step S105, respectively fuse all cloud sets in each group together and construct a three-dimensional model.

[0100] For all point cloud sets of each group, arrange them in the order of the collection time of the point cloud set. First, fuse the first point cloud set into the reference point cloud set to obtain a fused point cloud set; then fuse the second point cloud set into the fused point cloud set to obtain the fused point cloud set again, and then fuse the third point cloud set into the fused point cloud set to obtain the fused point cloud set again; and so on, until all point cloud sets of each group are fused into the fused point cloud set, at which time the obtained fused point cloud set is referred to as the point cloud fusion result of each group.

[0101] The point cloud fusion results of all groups and the point cloud sets not in all groups are also regarded as a group, and all the cloud fusion results of all groups and all point cloud sets not in all groups are fused together again according to the method described above in this step to obtain a final point cloud fusion result. The final point cloud fusion result contains all pump house equipment and all details of the same pump house equipment. In this embodiment, the three-dimensional model of the pump house is constructed by using the final point cloud fusion result.

[0102] This embodiment ends.

[0103] In this embodiment, the point cloud set is divided by using a suitable window, and the point clouds at suitable positions are respectively aligned, matched and fused, so that the problem that the point clouds at different positions cannot be accurately aligned and fused due to the existence of occlusion in the pump house, and the details of the structure of the pump house equipment cannot be restored, is avoided.

[0104] Embodiment two:

[0105] In step S104, the point cloud set at all positions is initially classified to obtain all initial categories. As an example, the process specifically includes:

[0106] According to the collection sequence of the point cloud set at each position, the point cloud set at each position is traversed, and the traversed point cloud set at the i-th position is recorded as si. The si is put into the empty set S.

[0107] Each position except si is traversed, and the traversed point cloud set at the i-th position is recorded as .

[0108] It is judged whether the point clouds in the point cloud group that is not aligned under the preferred window size of each point cloud set in S are respectively contained in the point cloud group aligned under the preferred window size, if yes, the is put into the set S, and if no, the is not put into the set S.

[0109] Then each position except si is traversed, and the traversed point cloud set at the i+1-th position is recorded as .

[0110] It is judged whether the point clouds in the point cloud group that is not aligned under the preferred window size of each point cloud set in S are contained in the point cloud group aligned under the preferred window size, if yes, the is put into the set S, and if no, the is not put into the set S.

[0111] By analogy, when all positions except si are traversed, the obtained set S is taken as an initial category.

[0112] Thus, an initial category is obtained for the point cloud set si at any one position and the point cloud set outside si. All initial categories can be obtained by traversing the point cloud set at all positions. In this embodiment, only the initial category containing more than 10 point cloud sets is retained.

[0113] As an example, for each point cloud set in S, whether the point clouds in the point cloud group that is not aligned at its preferred window size are respectively contained in In the point cloud group that is aligned at its preferred window size, the method includes:

[0114] For any one point cloud set in S, the point clouds in the point cloud group that is not aligned at its preferred window size (see S102 in detail), and for In the point cloud group that is aligned at its preferred window size (see S102 in detail), the point clouds in the point cloud group that is not aligned are translated and rotated into the reference point cloud set by using the target transformation parameters corresponding to any one point cloud set in S (see S102 in detail), to obtain a first point cloud, and the point clouds in the point cloud group that is aligned are also translated and rotated into the reference point cloud set by using the target transformation parameters, to obtain a second point cloud. In the point cloud group that is aligned at its preferred window size, the method includes:

[0115] For any one first point cloud, if the Euclidean distance between the first point cloud and the nearest second point cloud is less than the alignment threshold, it is determined that the first point cloud is contained in the second point cloud. When more than 30% of the first point clouds in all first point clouds are contained in the second point clouds, the any one point cloud set in S is marked as containing point clouds, otherwise it is marked as not containing point clouds.

[0116] When all point cloud sets in S are marked as containing point clouds, it is determined that: the point clouds in the point cloud group that is not aligned at its preferred window size of each point cloud set in S are respectively contained in In the point cloud group that is aligned at its preferred window size. When there is a mark of not containing point clouds in all point cloud sets in S, it is determined that: the point clouds in the point cloud group that is not aligned at its preferred window size of each point cloud set in S are not contained in In the point cloud group that is aligned at its preferred window size.

[0117] In step S105, the first point cloud set is fused into the reference point cloud set to obtain a fused point cloud set. As an example, the specific process includes:

[0118] The first point cloud set is divided into a plurality of point cloud groups by using the preferred window size of the first point cloud set, and the reference point cloud set is also divided into a plurality of point cloud groups by using the preferred window size of the first point cloud set; by using the transformation matrix, the point cloud groups of the first point cloud set are aligned to the point cloud groups of the reference point cloud set according to the method of step S102, and the aligned first point cloud set and the reference point cloud set are merged as a fused point cloud set. In the fused point cloud set, the point clouds with a Euclidean distance less than the alignment threshold are considered as aligned or matched together, and the aligned or matched point clouds are merged.

[0119] The step S105 further includes: fusing the second point cloud set into the fused point cloud set, and obtaining the fused point cloud set again. The process is the same as the above.

[0120] The second point cloud set is divided into a plurality of point cloud groups by using the preferred window size of the second point cloud set, and the fused point cloud set is also divided into a plurality of point cloud groups by using the preferred window size of the second point cloud set. According to the method of step S102, the point cloud groups of the second point cloud set are aligned into the point cloud groups of the fused point cloud set by using the transformation matrix. The aligned second point cloud set and the fused point cloud set are merged to be the fused point cloud set again. In the fused point cloud set, the point clouds with the Euclidean distance less than the alignment threshold are regarded as the point clouds aligned or matched together, and the point clouds aligned or matched together are merged.

[0121] As an example, in the fused point cloud set, the point clouds with the Euclidean distance less than the alignment threshold are regarded as the point clouds aligned or matched together, and the point clouds aligned or matched together are merged, including the following method:

[0122] For the origin of the coordinate system in which all the point clouds in the fused point cloud set are located, each point cloud in the fused point cloud set is traversed in the order of the Euclidean distance from the origin from small to large. When the jth point cloud is traversed, all the point clouds with the Euclidean distance less than the alignment threshold from the jth point cloud are obtained. These point clouds (including the jth point cloud) form a merging set. After all the point clouds in the fused point cloud set are traversed, a plurality of merging sets are obtained. If the point clouds contained in the plurality of merging sets are completely the same, only one of the merging sets is retained. The mean value of the coordinates of all the point clouds in each merging set is calculated, all the point clouds in each merging set are deleted, and a point cloud is generated at the mean value position.

[0123] This embodiment is described taking the alignment threshold equal to 0.01 as an example.

[0124] The step S104 includes: as an example, a preferred window size is selected and removed from each first category respectively, so that the variance of the selected preferred window sizes in all the first categories is maximum. As an example, the specific process is as follows:

[0125] (1) A preferred window size is randomly selected from each first category respectively, and the variance of the selected preferred window sizes in all the first categories is obtained.

[0126] (2) The selection method of all the selected window sizes in all the first categories is traversed, and the window sizes selected by all the first categories when the variance is maximum are obtained. These window sizes are removed from the first categories respectively.

[0127] In other examples, when the number of selected window sizes in the first category is large, the process in (2) can be replaced by the following method to reduce the amount of calculation:

[0128] Repeat (1) for several times (for example, 100 times), obtain a variance after each repetition of (1), and obtain the window size selected by the first category at the maximum method. Remove these window sizes from the first category.

[0129] Embodiment three:

[0130] As an example, a three-dimensional model of the pump house is constructed using the final point cloud fusion result, including the following methods:

[0131] The final point cloud fusion is converted into an initial model of the pump house using three-dimensional processing software such as Autodesk ReCap and Blender, and then the initial model is manually fine-tuned and repaired in the three-dimensional processing software to obtain the final three-dimensional model of the pump house. Subsequently, the three-dimensional model is added with materials, drawn with maps, bound with animations and special effects, which will not be specifically described in this embodiment.

[0132] In other embodiments, other structures are manually created in the final three-dimensional model, such as water and electricity lines, pipeline lines and steel reinforcement structures in the walls created according to the construction drawings of the pump house.

[0133] Embodiment four:

[0134] The embodiment provides a three-dimensional model construction system of a multi-level structure of a pump house, which includes a memory, a processor and a computer program stored in the memory and executable on the processor. The processor executes all steps of all embodiments described above when executing the computer program.

[0135] Embodiment five:

[0136] The embodiment provides a three-dimensional model construction device of a multi-level structure of a pump house, which includes a drone carrying a laser radar for collecting point cloud sets at each position in the pump house. The device also includes a computer for reading the point cloud sets at each position. After reading the point cloud sets at all positions, the computer creates a three-dimensional model by running the three-dimensional model construction system of a multi-level structure of a pump house described above.

[0137] The above only describes the preferred embodiments of the present application and does not limit the present application. Any modification, equivalent replacement, improvement, etc. within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for constructing a three-dimensional model of a multi-story structure of a pump house, characterized by, The method comprises the following steps: Collecting point cloud sets at different positions of the pump house respectively, and designating a point cloud set as a reference point cloud set; any point cloud set other than the reference point cloud set is recorded as a target point cloud set; Dividing the target point cloud set and the reference point cloud set into a plurality of point cloud groups by using a preset window size, aligning the point cloud groups of the target point cloud set to the point cloud groups of the reference point cloud set by using a transformation matrix, and obtaining point cloud groups that are aligned and point cloud groups that are not aligned of the target point cloud set under the window size; wherein the difference between the alignment error of the point cloud groups that are not aligned and the alignment error of the point cloud groups that are aligned is recorded as an alignment index under the window size; the window size when the alignment index is maximum is recorded as an optimal window size of the target point cloud set; Dividing all the point cloud sets into a plurality of groups, the optimal window size of different point cloud sets in each group having maximum difference, and any point cloud set in each group having no point cloud group that is not aligned under the optimal window size, other point cloud sets outside the point cloud set contained in each group having point cloud groups that are aligned under the optimal window size; Fusing all the point cloud sets in each group together respectively and constructing a three-dimensional model; The specific steps of the above-mentioned aligning the point cloud groups of the target point cloud set to the point cloud groups of the reference point cloud set by using a transformation matrix, and obtaining point cloud groups that are aligned and point cloud groups that are not aligned of the target point cloud set under the window size, comprise the following steps: Matching all the point cloud groups of the target point cloud set and all the point cloud groups of the reference point cloud set to obtain all matching pairs; the parameters of translation and rotation transformation when the point cloud group from the target point cloud set is translated and rotated to the point cloud group from the reference point cloud set in each matching pair are recorded as the transformation parameters of each matching pair; Among the transformation parameters of all the matching pairs, the transformation parameters when the difference between a plurality of transformation parameters is minimum constitute a target category; the mean value of the transformation parameters in the target category is recorded as a target transformation parameter; the point cloud groups in the matching pairs corresponding to all the transformation parameters in the target category are the point cloud groups that are aligned, and the point cloud groups other than the point cloud groups that are aligned are the point cloud groups that are not aligned; The target transformation parameter is the parameter of the translation matrix and the rotation matrix used when the point cloud groups in the target point cloud set are aligned to the point cloud groups in the reference point cloud set.

2. The method of claim 1, wherein the method further comprises: The specific steps of obtaining the alignment error of the point cloud groups that are not aligned and the alignment error of the point cloud groups that are aligned are as follows: For any aligned or not aligned point cloud group in the target point cloud set, all the point clouds in the point cloud group are translated and rotated by using the translation matrix and the rotation matrix corresponding to the target transformation parameter to obtain transformed point clouds; the nearest distance between the transformed point clouds and the point clouds in the reference point cloud set is recorded as the average first distance of the point cloud group; for all the aligned or not aligned point cloud groups in the target point cloud set, the mean value of the average first distances of all the aligned or not aligned point cloud groups is recorded as the alignment error of the aligned or not aligned point cloud groups.

3. The method of claim 1, wherein the method further comprises: The step of dividing all point cloud sets into several groups, wherein the preferred window size of different point cloud sets in each group has the maximum difference, and any one point cloud set in each group has no aligned point cloud group at the preferred window size, comprises the following specific steps: D1: performing initial classification on all point cloud sets to obtain all initial categories, wherein for each point cloud set x in the same initial category and for each point cloud set y other than x in the same initial category, the point cloud of x in the point cloud group having no alignment at the preferred window size is contained in the point cloud group aligned at the preferred window size of y; D2: performing KMeans clustering on the preferred window size of all point cloud sets in any initial category to obtain all first categories; D3: selecting and removing a preferred window size from each first category respectively, so that the variance of the selected preferred window sizes in all first categories is maximum, and the point cloud set corresponding to the selected preferred window size in all first categories is taken as a group; D4: repeating the process in D3 until the selected preferred window sizes in all first categories are less than two, and obtaining several groups.

4. The method of claim 3, wherein the method further comprises: The step of performing initial classification on all point cloud sets to obtain all initial categories comprises the following specific steps: traversing all point cloud sets, taking the traversed point cloud set as s1, and putting s1 into an empty set S; traverse all point cloud sets except s1, where the i-th point cloud set is denoted as ; determine whether the point clouds in the point cloud set that are not aligned in the preferred window size in S are respectively contained in the point cloud set that is aligned in the preferred window size; if yes, put into set S; after traversing all point cloud sets except s1, set S is taken as an initial class; after traversing all point cloud sets s1, obtaining all initial categories.

5. The method of claim 1, wherein the method further comprises: The step of fusing all point cloud sets in each group together comprises the following specific steps: D5: for all point cloud sets in each group, first fuse the first point cloud set into a reference point cloud set to obtain a fused point cloud set; then fuse the second point cloud set into the fused point cloud set to obtain the fused point cloud set again, and then fuse the third point cloud set into the fused point cloud set to obtain the fused point cloud set again; and so on, until all point cloud sets in each group are fused into the fused point cloud set, and the obtained fused point cloud set is taken as the point cloud fusion result of each group; D6: taking the point cloud fusion results of all groups as a group, and fusing all point cloud fusion results of all groups together again according to the method in D5 to obtain the final point cloud fusion result for constructing a three-dimensional model.

6. The method of claim 5, wherein the method further comprises: The step of fusing the first point cloud set into the reference point cloud set to obtain the fused point cloud set comprises the following specific steps: divide the first point cloud set into several point cloud groups by using the preferred window size of the first point cloud set, and divide the reference point cloud set into several point cloud groups by using the preferred window size of the first point cloud set, align the point cloud groups of the first point cloud set to the point cloud groups of the reference point cloud set by using a transformation matrix, and combine the aligned first point cloud set and reference point cloud set as the fused point cloud set; In the fusion point cloud set, the point cloud sets with a Euclidean distance less than a preset alignment threshold are merged.

7. The method of claim 1, wherein the method further comprises: The transformation parameters of the all matching pairs are normalized, and the normalized transformation parameters of the all matching pairs are subjected to mean shift clustering to obtain all categories, and a category with the most transformation parameters is recorded as a target category of the all matching pairs, and the Euclidean distances between all the normalized transformation parameters in the target category are calculated, and then the mean value is obtained to obtain a transformation error of the all matching pairs. For the all matching pairs with the minimum transformation error, a target category of the all matching pairs with the minimum transformation error is the target category of the transformation parameters with the minimum difference. The processor executes all steps of the construction method of the multi-level structure three-dimensional model of the pump house according to any one of claims 1-7 when the computer program is executed.

8. A system for constructing a three-dimensional model of a multi-layered pump house structure, the system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, After the computer reads the point cloud sets of all positions, a three-dimensional model is created by executing the construction system of the multi-level structure three-dimensional model of the pump house according to claim 8.

9. A device for constructing a three-dimensional model of a multi-story structure of a pump house, the device comprising a drone carrying a laser radar for collecting a point cloud set at each position in the pump house, the device further comprising a computer configured to read the point cloud set at each position, characterized in that, ​

Images