An automated detection method for pose anomalies and registration of models

Abstract

The present invention discloses an automated detection method for model pose anomalies and registration, which is applied to the technical field of three-dimensional urban modeling. The method includes the following steps: converting the model into derivative point cloud data, constructing a shared octree between the derivative point cloud obtained from the model conversion and the reference point cloud, calculating the exclusive-or exponent h, detecting the degree of pose anomaly of the original model through the exclusive-or exponent h, establishing corresponding evaluation indicators, and outputting the overall analysis and detection results. The result of judging the anomaly is registered using the ICP algorithm. Through the format conversion of the OBJ model and the analysis of the double-buffer mechanism of the octree, the present invention can replace the existing method for detecting pose anomalies based on manual visual interpretation, has higher detection accuracy, greatly improves the detection efficiency, and automatically adjusts the model with pose anomalies, saving considerable time and labor costs for the modeling unit.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional city modeling, and more particularly to a method for automatically detecting abnormalities in model posture and alignment. Background Art

[0002] Highly detailed 3D city models are a crucial foundation for digital city construction and have widespread applications in areas such as smart cities, autonomous driving, and intelligent transportation. During the modeling process, prefabricated models from a template library must be placed at corresponding locations. However, there are cases where the models do not fully match the position, tilt angle, and rotation of road ancillary facilities in the real scene. These non-compliant models must be detected so that they can be subsequently matched to the real scene and corrected for angles and postures. Existing screening methods rely primarily on manual visual inspection, where each model is scaled and rotated, and then repeatedly compared with point cloud and oblique imagery data collected in the field. This entire screening process is inefficient, susceptible to human error, and subject to significant individual variability, making it difficult to achieve consistently stable results and resulting in a minimal degree of automation. Some semi-automated software can only guarantee high accuracy at the bottom or top positions, but cannot precisely guarantee a complete match of the model's overall posture information. Therefore, after modeling is complete, significant labor and time are required to inspect the entire model's posture information, identify and correct any abnormal postures, and apply corrections.

[0003] Point cloud registration based on local features calculates and matches the source and target point clouds at local key points and their corresponding feature descriptors, then uses the corresponding feature points to solve the pose transformation matrix. However, due to factors such as the lack of global information in the feature point description data, local feature-based registration does not include an optimal solution for all points. Alternatively, there are fast global registration methods that use features and alternating optimization techniques to accelerate registration, improving registration speed. While robust to singular values, these methods still suffer from initialization sensitivity and high network complexity. Existing deep learning-based point cloud registration methods, such as PointNetLK and DCP, are still affected to some extent by factors such as noise and singular values, and their registration rates need to be improved. Furthermore, they are affected by the number of training samples. The higher the model complexity, the lower the detection accuracy. Furthermore, different projects require different models, and establishing a training sample library is time-consuming.

[0004] Therefore, how to provide an automated modeling posture anomaly detection and alignment method is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0005] In view of this, the present invention provides a method for automatically detecting the pose abnormality and registration of a model, which eliminates the existing manual visual screening process through the research of an automatic detection algorithm, and improves the quality screening efficiency and automation level of urban three-dimensional models in the process of large-scale refined urban modeling.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A method for automatically detecting the pose abnormality and registration of a model includes the following steps:

[0008] S1. Obtain a reference point cloud, and perform filtering, denoising, and sampling processing on the reference point cloud;

[0009] S2. Use the PointNet++ network to classify the reference point cloud, and group the reference point cloud according to a preset category;

[0010] S3. Obtain an urban three-dimensional model, and convert the original urban three-dimensional model into a derived point cloud;

[0011] S4. Perform octree encoding on the derived point cloud and the reference point cloud to obtain the octree leaf nodes of the reference point cloud;

[0012] S5. Perform differential encoding on the derived point cloud to obtain the octree leaf nodes of the derived point cloud;

[0013] S6. Calculate the exclusive-or exponent of the octree leaf nodes of the reference point cloud and the derived point cloud;

[0014] S7. Establish a pose abnormality evaluation criterion for the model, and output a detection result based on the exclusive-or exponent calculation result;

[0015] S8. For the derived point cloud with an abnormal pose in the output detection result, calculate a spatial transformation matrix for overall registration;

[0016] S9. Perform clustering segmentation on the derived point cloud, and perform local registration on the segmented point cloud.

[0017] Optionally, S3 is specifically as follows:

[0018] S31. Open the original three-dimensional urban model file in ASCII encoding form, and classify it based on the prefix keyword of each line of data according to the basic principle of whether the information of this line is valuable for generating the corresponding derived point cloud;

[0019] S32. Traverse each line of data in the original three-dimensional urban model file, obtain the corresponding keyword prefix and perform screening, delete the worthless data content, and obtain the derived point cloud;

[0020] S33. Set the header file parameters of the derived point cloud;

[0021] S34. Use the coordinate system of the 3D city model itself as the coordinate system for the corresponding derived point cloud, and assign the vertex coordinates of the 3D city model to the data part of the derived point cloud one by one.

[0022] Optionally, S4 is specifically as follows:

[0023] S4 is specifically as follows:

[0024] S41. Denote the original point cloud obtained by field surveying means as the reference point cloud;

[0025] S42. Establish a complete spatial index for the reference point cloud using an octree-based data structure. Each leaf node corresponds to a voxel, and allocate memory space A in the computer memory to store all voxels in the reference point cloud;

[0026] S43. Uniquely encode the octree leaf nodes of the reference point cloud, and establish a unique code for each voxel where each point in the reference point cloud is located;

[0027] S44. Copy the spatial index and the corresponding codes of the reference point cloud, implement the derived point cloud and the reference point cloud sharing the same octree structure in the memory space A, and uniquely encode each octree leaf node of the derived point cloud.

[0028] Optionally, S5 is specifically as follows:

[0029] S51. Starting from the root node, calculate the mapping relationship between the voxels corresponding to each octree leaf node of the reference point cloud and the voxels corresponding to each octree leaf node of the derived point cloud, and set the initial value of the counter to 0;

[0030] S52. If there are points in the derived point cloud that fall into the voxel space of the corresponding reference point cloud, share the code of the reference point cloud at this position with the derived point cloud, and increase the counter by 1;

[0031] S53. If there are no points in the derived point cloud that fall into the voxel space of the corresponding reference point cloud, the code and the counter remain unchanged;

[0032] S54. Repeat S52 - S53 until all leaf nodes in the reference point cloud are traversed, and record the final value of the counter.

[0033] Optionally, in S6, the ratio of the number of codes shared between the derived point cloud and the reference point cloud to the total number of codes of all octree leaf nodes of the derived point cloud is used as the XOR index of the derived point cloud.

[0034] Optionally, the model pose anomaly evaluation criterion in S7 is:

[0035] When h = 0, the city 3D model and the real scene are completely matched; when 0 <h<Δ时,城市三维模型存在局部偏差;当Δ<h<3Δ时,城市三维模型在位置和姿态上存在偏差;当3Δ<h<5Δ时,城市三维模型存在位置和姿态异常;当h=100时,城市三维模型的位置和姿态完全错误;其中,h为异或指数,0<Δ<20。

[0036] Optionally, S8 is:

[0037] S81. Take point set a in the derived point cloud i , and find the corresponding point b in the reference point cloud i , so that a i -b i =min;

[0038] S82. Calculate the rotation matrix R and the translation matrix T so that the objective function E takes the minimum value:

[0039]

[0040] S83, performing rotation and translation transformation on the derived point cloud to update and obtain a new point cloud dataset A′;

[0041] S84. Calculate the distances between the new point cloud dataset A′ and all corresponding points of the reference point cloud and perform normalization processing:

[0042]

[0043] S85. If d is greater than the preset threshold, repeat S81-S84, otherwise go to S9.

[0044] Optional, S9 specifically:

[0045] S91, clustering the new point cloud dataset A′ by using the DBSCAN clustering algorithm, and cropping the reference point cloud based on the clustered new point cloud dataset A′;

[0046] S92, calculating the XOR index of the clustered new point cloud dataset A′ and the cropped reference point cloud to obtain an outlier;

[0047] S93. When the outlier value exceeds the preset threshold, manual inspection and intervention are performed; if there is an anomaly and the outlier value is less than the preset threshold, precise registration is performed; if there is no anomaly, the new point cloud dataset A′ after clustering is directly output.

[0048] As can be seen from the above technical solutions, compared with the prior art, the present invention provides an automated method for detecting model pose anomalies and registration, which has the following beneficial effects: The present invention can automatically detect and register model pose anomalies, breaking through the limitations of the existing manual visual judgment method. Compared with the direct registration algorithm, it shortens the retrieval time and improves the model detection efficiency; it quantifies the pose anomaly situation based on the exclusive OR index, with higher reliability, and can save considerable time and labor costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on the provided drawings without creative efforts.

[0050] Figure 1 It is a flowchart for detecting and registering model pose anomalies of the present invention;

[0051] Figure 2 It is a schematic diagram of the coordinate system of the original OBJ model of the present invention itself;

[0052] Figure 3 It is a schematic diagram of the principle of the complete octree index of the reference point cloud of the present invention;

[0053] Figure 4 It is a schematic diagram of the principle of the DBSCAN clustering algorithm of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0055] The embodiments of the present invention disclose an automated method for detecting model pose anomalies and registration, as Figure 1 shown, including the following steps:

[0056] S1. Obtain the reference point cloud, and perform filtering, denoising, and sampling processing on the reference point cloud;

[0057] S2. Use the PointNet++ network to classify the reference point cloud, and group the reference point cloud according to preset categories;

[0058] S3. Obtain the 3D city model and convert the original 3D city model into a derived point cloud;

[0059] S4. Perform octree encoding on the derived point cloud and the reference point cloud to obtain the octree leaf nodes of the reference point cloud;

[0060] S5. Perform differential encoding on the derived point cloud to obtain the octree leaf nodes of the derived point cloud;

[0061] S6. Perform XOR exponent calculation on the octree leaf nodes of the reference point cloud and the derived point cloud;

[0062] S7. Establish a model pose anomaly evaluation criterion and output a detection result based on the XOR exponent calculation result;

[0063] S8. For the derived point cloud with an abnormal pose in the output detection result, calculate the spatial transformation matrix for global registration;

[0064] S9. Perform clustering segmentation on the derived point cloud and perform local registration on the segmented point cloud.

[0065] Further, in an embodiment of the present invention, the original point cloud obtained by using field surveying means such as a high-precision terrestrial laser scanning system or a vehicle-mounted laser scanning system is denoted as the reference point cloud t is the number of laser points in the reference point cloud. The reference point cloud Compared with the corresponding 3D city model, it has higher position accuracy and more accurate pose information. Therefore, it can be regarded as an accurate portrayal of the physical entity corresponding to the 3D city model in the real scene and used as a reference benchmark for subsequent 3D city model pose anomaly detection. For the reference point cloud After filtering, denoising, and sampling processing, a network training sample library and a validation library are established. The sample library is segmented into 10m×10m×10m and input into the PointNet++ network for network training. The network parameters are continuously adjusted through the backpropagation algorithm to improve the classification accuracy. Through methods such as cross-validation, the trained neural network is evaluated and optimized to improve the robustness and generalization ability of the classification. Use the trained PointNet++ network to classify the reference point cloud Classify the reference point cloud Group according to the preset categories.

[0066] Further, S3 is specifically:

[0067] S31. Open the original 3D city model file in ASCII encoding form and classify it based on the prefix keyword of each line of data according to the basic principle of whether the information in this line is valuable for generating the corresponding derived point cloud;

[0068] S32. Traverse each line of data in the original 3D city model file, obtain the corresponding keyword prefix and perform screening, delete the worthless data content, and obtain the derived point cloud;

[0069] S33. Set the header file parameters of the derived point cloud;

[0070] S34. Use the coordinate system of the 3D city model itself as the coordinate system of the corresponding derived point cloud, and assign the vertex coordinates of the 3D city model to the data part of the derived point cloud one by one.

[0071] In an embodiment of the present invention, the 3D city model is an OBJ model. The OBJ model is a file format based on geometric definitions, which consists of basic information such as model vertex coordinates, vertex normal vectors, surface elements, and their connection relationships. In addition, each original OBJ model file also contains a large amount of unstructured annotation information, which causes serious interference to subsequent analysis and processing. Therefore, it is necessary to clean the redundant data of the original OBJ model file based on the requirements of pose anomaly detection. In the embodiment of the present invention, the classification principle is shown in Table 1:

[0072] Table 1

[0073]

[0074] In S32, delete the worthless data content according to the classification categories in Table 1, and only retain the valuable data content, that is, the lines corresponding to the keywords prefixed with "v", denoted as n is the number of vertex coordinates in the original OBJ model, x i 、y i 、z i are the coordinate values of each OBJ model vertex in its own space rectangular coordinate system. The coordinate system of the original OBJ model is as Figure 2 shown; the corresponding PCD format point cloud generated by converting the vertex coordinates of each OBJ model is denoted as the derived point cloud m is the number of points contained in the derived point cloud; the PCD format, as a standard point cloud data format, contains two parts: a header file and data. The header file part has a fixed format and contains 10 keywords. In S33, set the header file parameters of the derived point cloud according to Table 2.

[0075] Table 2

[0076] Serial number Parameter name Value Remarks 1 VERSION 0.7 PCD file version 2 FIELDS xyz Dimensions of each point 3 SIZE 444 Memory occupied by each data 4 TYPE FFF Data type 5 COUNT 111 Number of elements in each dimension 6 WIDTH n Width of data points 7 HEIGHT 1 Height of data points 8 VIEWPOINT 0001000 Viewpoint of data acquisition 9 POINTS n Total number of points 10 DATA ascii Data storage type

[0077] After completing the parameter setting of the header file part, use the coordinate system of each original OBJ model itself as the coordinate system for the corresponding derived point cloud and assign the vertex coordinates of each original OBJ model Assign one-to-one values to the derived point cloud according to the parameter settings in Table 2 In the data part, the number of points of the two are consistent, that is:

[0078]

[0079] Complete batch conversion of original OBJ models to PCD format derived point clouds.

[0080] Furthermore, S4 is specifically:

[0081] S41. Recording the original point cloud obtained by field surveying and mapping as a reference point cloud;

[0082] S42. Use an octree-based data structure to establish a complete spatial index for the reference point cloud. Each leaf node corresponds to a voxel. Memory space A is opened in the computer memory to store the reference point cloud. All voxels in Figure 3 As shown;

[0083] S43. Reference point cloud The octet tree leaf nodes are uniquely coded, with reference to the point cloud Each voxel where a point is located in the voxel is uniquely coded, denoted as

[0084]

[0085] S44. Copy reference point cloud Spatial index and corresponding encoding Derived point cloud is realized in memory space A and reference point cloud Share the same octree structure and derive point clouds Each octet tree leaf node is uniquely coded and recorded as

[0086] Furthermore, S5 is specifically:

[0087] S51. Starting from the root node, calculate the reference point cloud The voxel and derived point cloud corresponding to each octree leaf node The mapping relationship between the voxels corresponding to each octree leaf node sets the initial value of the counter num to 0;

[0088] S52, if the point cloud is derived The middle point falls into the corresponding reference point cloud In the voxel space of The code at this location is shared with the derived point cloud The counter increases by 1, i.e.

[0089]

[0090] num = num + 1

[0091] S53. If there is no point in the derived point cloud falling into the voxel space of the corresponding reference point cloud , the encoding and the counter num remain unchanged;

[0092] S54. Repeat S52 - S53 until all leaf nodes in all reference point clouds are traversed, and record the final value of the counter num.

[0093] Furthermore, in S6, the ratio of the number of encodings shared between the derived point cloud and the reference point cloud to the total number of encodings of all octree leaf nodes of the derived point cloud is used as the exclusive - or index h of the derived point cloud , that is:

[0094]

[0095] Furthermore, the model pose anomaly evaluation criterion in S7 is:

[0096] When h = 0, the urban three - dimensional model and the real scene are completely matched; when 0 < h < Δ, there are local deviations in the urban three - dimensional model; when Δ < h < 3Δ, there are deviations in the position and pose of the urban three - dimensional model; when 3Δ < h < 5Δ, there are position and pose anomalies in the urban three - dimensional model; when h = 100, the position and pose of the urban three - dimensional model are completely wrong; where 0 < Δ < 20.

[0097] Furthermore, S8 is specifically as follows:

[0098] S81. Take a point set a i from the derived point cloud, and find the corresponding point b i in the reference point cloud, so that a i - b i = min;

[0099] S82. Calculate the rotation matrix R and the translation matrix T to minimize the objective function E:

[0100]

[0101] S83. Perform rotation and translation transformation on the derived point cloud to update and obtain a new point cloud dataset A';

[0102] S84. Calculate the distances between all corresponding points of the new point cloud dataset A' and the reference point cloud, and perform normalization processing:

[0103]

[0104] S85. If d is greater than the preset threshold, repeat S81 - S84; otherwise, proceed to S9.

[0105] Further, S9 is specifically as follows:

[0106] S91. Cluster the new point cloud dataset A′ using the DBSCAN clustering algorithm, denoted as cluster_i. The DBSCAN clustering algorithm is as Figure 4 shown; crop the reference point cloud based on the clustered new point cloud dataset A′.

[0107] S92. Calculate the exclusive - OR index h between the clustered new point cloud dataset A′ and the cropped reference point cloud to obtain outliers.

[0108] S93. When the outliers exceed the preset threshold, perform manual inspection and intervention; if there are outliers and the outliers are less than the preset threshold, perform precise registration; if there are no outliers, directly output the derived point cloud cluster_i.

[0109] Furthermore, in the embodiment of the present invention, after the registration is completed, fuse the registered clustered derived point cloud cluster_i, replace the vertex coordinate values in the urban three - dimensional model with the coordinate values of the registered derived point cloud, and obtain a model without outliers and with registration.

[0110] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0111] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An automated detection method for the abnormal pose and registration of a model, characterized in that, It includes the following steps: S1. Obtain the reference point cloud and perform filtering, denoising, and sampling processing on the reference point cloud; S2. Use the PointNet++ network to classify the reference point cloud and group the reference point cloud according to preset categories; S3. Obtain the 3D urban model and convert the original 3D urban model into a derivative point cloud; S4. Perform octree encoding on the derivative point cloud and the reference point cloud to obtain the octree leaf nodes of the reference point cloud; S5. Perform differential encoding on the derivative point cloud to obtain the octree leaf nodes of the derivative point cloud; S6. Perform XOR exponent calculation on the octree leaf nodes of the reference point cloud and the derivative point cloud; S7. Establish a model pose anomaly evaluation criterion and output a detection result based on the XOR exponent calculation result; S8. For the derivative point cloud with an abnormal pose in the output detection result, calculate the spatial transformation matrix for overall registration; S9. Perform clustering segmentation on the derivative point cloud and perform local registration on the segmented point cloud; Specifically, S3 is as follows: S31. Open the original 3D urban model file in ASCII encoding form and classify it based on the prefix keyword of each line of data according to the basic principle of whether the information in this line is valuable for generating the corresponding derivative point cloud; S32. Traverse each line of data in the original 3D urban model file, obtain the corresponding keyword prefix and perform screening, delete the worthless data content, and obtain the derivative point cloud; S33. Set the header file parameters of the derivative point cloud; S34. Use the coordinate system of the 3D urban model itself as the coordinate system of the corresponding derivative point cloud, and assign the vertex coordinates of the 3D urban model to the data part of the derivative point cloud one by one; Specifically, S5 is as follows: S51. Starting from the root node, calculate the mapping relationship between the voxel corresponding to each octree leaf node of the reference point cloud and the voxel corresponding to each octree leaf node of the derivative point cloud, and set the initial value of the counter to 0; S52. If there is a point in the derivative point cloud that falls into the voxel space corresponding to the reference point cloud, share the encoding of the reference point cloud at this position with the derivative point cloud, and increase the counter by 1; S53. If there is no point in the derivative point cloud that falls into the voxel space corresponding to the reference point cloud, the encoding and the counter remain unchanged; S54. Repeat S52 - S53 until all leaf nodes in the reference point cloud are traversed, and record the final counter value; Specifically, S8 is as follows: S81. Take point set a from the derived point cloud i , and find the corresponding point b in the reference point cloud i , such that a i - b i = min; S82. Calculate the rotation matrix R and the translation matrix T to minimize the objective function E: S83. Perform rotation and translation transformation on the derivative point cloud to update and obtain a new point cloud dataset A'; S84. Calculate the distances between all corresponding points of the new point cloud dataset A' and the reference point cloud and perform normalization processing; S85. If d is greater than the preset threshold, repeat S81 - S84, otherwise enter S9.

2. The method for automatically detecting pose abnormality and registration of a detection model according to claim 1, characterized in that Specifically, S4 is as follows: S41. Denote the original point cloud obtained by field surveying means as the reference point cloud; S42. Use an octree-based data structure to establish a complete spatial index for the reference point cloud. Each leaf node corresponds to a voxel, and allocate memory space A in the computer memory to store all voxels in the reference point cloud; S43. Uniquely encode the octree leaf nodes of the reference point cloud, and establish a unique encoding for each voxel where a point in the reference point cloud is located; S44. Copy the spatial index and corresponding encoding of the reference point cloud, and implement sharing the same octree structure between the derived point cloud and the reference point cloud in memory space A, and uniquely encode each octree leaf node of the derived point cloud.

3. The automated detection model pose anomaly and registration method according to claim 1, wherein In S6, the ratio of the number of encodings shared between the derived point cloud and the reference point cloud to the total number of encodings of all octree leaf nodes of the derived point cloud is used as the exclusive-or index of the derived point cloud.

4. An automated detection method for pose abnormality and registration of a model according to claim 1, characterized in that The model pose abnormality evaluation criterion in S7 is: When h = 0, the urban 3D model and the real scene are completely matched; when 0 < h < Δ, there are local deviations in the urban 3D model; when Δ < h < 3Δ, there are deviations in the position and pose of the urban 3D model; when 3Δ < h < 5Δ, there are position and pose abnormalities in the urban 3D model; when h = 100, the position and pose of the urban 3D model are completely wrong; where h is the exclusive-or index and 0 < Δ < 20.

5. An automated detection method for pose abnormality and registration of a model according to claim 1, characterized in that S9 is specifically as follows: S91. Cluster the new point cloud dataset A' through the DBSCAN clustering algorithm, and crop the reference point cloud based on the clustered new point cloud dataset A'; S92. Calculate the exclusive-or index of the clustered new point cloud dataset A' and the cropped reference point cloud to obtain an abnormal value; S93. When the abnormal value exceeds the preset threshold, perform manual inspection and intervention; if there are abnormalities and the abnormal value is less than the preset threshold, perform precise registration; If there are no abnormalities, directly output the clustered new point cloud dataset A'.

Images