Urban scene-oriented three-dimensional change point cloud data set construction method and system

Abstract

The invention discloses a three-dimensional change point cloud data set construction method and system for an urban scene, and belongs to the technical field of three-dimensional GIS (Geographic Information System), and the method comprises the steps: collecting multi-time-sequence three-dimensional change point cloud data for the urban scene, and carrying out the preprocessing; performing coarse screening on the preprocessed multi-time-sequence three-dimensional change point cloud data by adopting an octree algorithm; and based on the coarsely screened multi-time-sequence three-dimensional change point cloud data, performing manual labeling and cross validation to obtain finally labeled multi-time-sequence point cloud data, and completing the construction of the three-dimensional change point cloud data set. According to the method, through a semi-automatic method of collecting the real multi-time-sequence three-dimensional change point cloud data facing the urban scene, coarsely screening the change information by adopting the octree algorithm and manually marking, the efficiency of constructing the three-dimensional change point cloud data set is improved, so that an efficient data support is provided for constructing the three-dimensional change detection method.

Description

Technical Field

[0001] The present invention belongs to the technical field of 3D GIS, and particularly relates to a method and system for constructing a 3D change point cloud data set for urban scenes. Background Technique

[0002] Cities are the main places for human activities and carry the main social production activities of humans. In modern times when technology is gradually developing, cities are constantly in dynamic change, such as urban expansion, deforestation, changes in lake areas, and changes in landforms after natural disasters. In order to accurately grasp these changes and achieve precise urban management, change detection is required. Change detection locates the changed areas and mines key information such as the degree and type of changes by comparing data of the same area at different times. 3D change detection is to obtain the changes in the geospatial at different time series by comparing 3D data, such as point clouds, DEMs, and DSMs. Compared with 2D change detection, 3D change detection can detect changes in the vertical direction and the detected changes are more comprehensive and accurate.

[0003] The 3D change detection method based on deep learning is currently the mainstream. Compared with algebraic difference and machine learning, it has better stability, stronger robustness, and higher accuracy. The 3D change detection based on deep learning requires the support of a 3D change detection data set. In order to achieve urban-level 3D change detection, a 3D change detection data set for urban scenes needs to be adopted. Since point cloud data is easy to collect, has high accuracy, and a simple data structure, most 3D change detection data sets are constructed based on point clouds. In order to construct a 3D change detection data set for urban scenes, currently, the differences between point clouds at different time series are mainly compared, and the change information in multi-temporal point clouds is labeled point by point. This method has extremely high accuracy, but the characteristics of multi-phase data and point-by-point comparison greatly increase the workload of data annotation and significantly increase the cost of data set production.

[0004] In order to improve the data set production speed and reduce the production cost, currently, a semi-automated method is mainly used for data production. First, multi-temporal virtual point cloud data is generated through a simulation engine, and then changes are created between multi-temporal data according to the established change rules or manually, so as to form rich 3D point cloud change data. Although this method can quickly obtain a large amount of 3D change data, due to the simple geometric shape and low authenticity of texture materials of virtual data, it cannot accurately describe the changes in urban real scenes, so the data still has certain limitations.

[0005] In summary, the 3D change detection dataset is a necessary condition for constructing a 3D change detection method based on deep learning. Although the current construction methods of 3D change detection datasets can provide rich and accurate 3D change data, there are still the following problems: 1) Although the method of manual annotation can achieve extremely high data accuracy, the cost and difficulty of data annotation are too high, which greatly hinders the production and extension of 3D change detection datasets. 2) Although the semi-automated method of generating virtual data through a simulation engine can effectively improve the speed of dataset production and reduce the cost of dataset production, the authenticity of virtual data is poor and it cannot accurately describe the changes in urban real scenes, which will greatly limit the implementation of large-scale urban 3D change detection. Therefore, it is urgent to study an efficient method for constructing a 3D change detection dataset for urban scenes.

[0006] Therefore, it is necessary to design a method and system for constructing a 3D change point cloud dataset for urban scenes to address the above problems. Summary of the Invention

[0007] The purpose of the present invention is to address the problems existing in the prior art and provide a method and system for constructing a 3D change point cloud dataset for urban scenes. Through methods such as data cleaning and point cloud registration, the data preprocessing is completed. Then, the octree algorithm is used to roughly screen the change information of multi-temporal point cloud data, and the manually marked and corrected the roughly screened change information, overcoming the problems of too high cost and difficulty of data annotation, obtaining a real dataset that describes the changes in urban real scenes, and realizing the construction of a 3D change point cloud dataset.

[0008] According to one aspect of this specification, a method for constructing a 3D change point cloud dataset for urban scenes is provided, including:

[0009] Step 1: Collect multi-temporal point cloud data for urban scenes and perform preprocessing;

[0010] Step 2: Use the octree algorithm to roughly screen the preprocessed multi-temporal point cloud data, specifically including:

[0011] First, according to the spatial range of the point cloud data, construct an octree for the point cloud data of each time series.

[0012] Secondly, extract features for each node in the octree, and the features are the number of nodes and the node density.

[0013] Finally, based on the number of nodes and the node density, compare the feature differences of each node in the octrees of different time series to determine the point cloud data that has changed and the point cloud data that has not changed.

[0014] Step 3: Based on the multi-temporal 3D change point cloud data after rough screening, perform annotation and cross-validation to obtain the finally annotated multi-temporal point cloud data, and obtain the constructed 3D change point cloud dataset for urban scenes.

[0015] Further, the construction of the octree includes:

[0016] Determine the cubic space represented by the root node of the octree according to the spatial range of the point cloud data;

[0017] Divide the cubic space to obtain eight sub-cubes, and determine that each sub-cube reaches the predetermined maximum recursion depth or the number of point clouds in each sub-cube satisfies a certain threshold to obtain the constructed octree.

[0018] Further, determining the point cloud data that has changed and has not changed includes:

[0019] Use the nearest neighbor algorithm to determine the corresponding positions of each node in the octrees of different time series, and determine the threshold according to the point cloud density;

[0020] The octrees of different time series form voxels by dividing the space. According to the number of nodes and the volume of the voxels, calculate the node density, and use the node density as the feature difference for comparison;

[0021] When the feature difference exceeds the threshold, determine it as the point cloud data that has changed. When the feature difference does not exceed the threshold, determine it as the point cloud data that has changed.

[0022] Further, performing annotation and cross-validation includes:

[0023] Annotate the multi-temporal point cloud data after rough screening, and correct the unchanged data into changed data;

[0024] Perform cross-validation on the data after different manual corrections to obtain the finally annotated multi-temporal point cloud data.

[0025] Further, after performing preprocessing, it also includes performing point cloud registration:

[0026] Select multiple feature points to perform preliminary alignment on the multi-temporal point cloud data, and use the iterative closest point algorithm to perform fine registration on the preliminarily aligned point cloud data.

[0027] According to one aspect of the present specification, a 3D change point cloud dataset construction system for urban scenes is provided, including:

[0028] A data acquisition module for acquiring multi-temporal 3D change point cloud data for urban scenes and performing preprocessing;

[0029] Coarse screening module, which is used to coarsely screen the preprocessed multi-temporal point cloud data by using the octree algorithm, specifically including:

[0030] First, according to the spatial range of the point cloud data, an octree is constructed for the point cloud data of each time series.

[0031] Secondly, features are extracted for each node in the octree, and the features are the number of nodes and the node density.

[0032] Finally, based on the number of nodes and the node density, the feature differences of each node in the octrees of different time series are compared to determine the point cloud data that has changed and the point cloud data that has not changed.

[0033] Dataset construction module, which is used to perform annotation and cross-validation based on the coarsely screened multi-temporal three-dimensional change point cloud data to obtain the finally annotated multi-temporal point cloud data, and obtain the constructed three-dimensional change point cloud dataset for urban scenes.

[0034] According to one aspect of this specification, an electronic device is provided, including a memory and a processor, the memory stores a computer program, and it is characterized in that when the processor executes the computer program, the steps of the method for constructing a three-dimensional change point cloud dataset for urban scenes are implemented.

[0035] According to one aspect of this specification, a computer-readable storage medium is provided, on which a computer program is stored, and it is characterized in that when the computer program is executed by a processor, the steps of the method for constructing a three-dimensional change point cloud dataset for urban scenes are implemented.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] 1. The present invention completes the preprocessing of data through methods such as data cleaning and point cloud registration, then coarsely screens the change information of multi-temporal point cloud data by using the octree algorithm, manually annotates and corrects the coarsely screened change information, and performs cross-validation on the annotated data to ensure the accuracy of the data, so as to efficiently and accurately construct a three-dimensional change point cloud dataset.

[0038] 2. Aiming at the difficulties of annotating and identifying multi-temporal data in a three-dimensional change dataset, by collecting real multi-temporal three-dimensional change point cloud data for urban scenes and using the octree algorithm to coarsely screen the change information, a semi-automatic method of manual annotation is adopted, which helps to improve the efficiency of constructing a three-dimensional change point cloud dataset, thereby providing efficient data support for the construction of a three-dimensional change detection method. Description of the Drawings

[0039] 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 in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0040] Figure 1 It is a flowchart of a method for constructing a three-dimensional change point cloud data set for an urban scenario according to an embodiment of the present invention;

[0041] Figure 2 It is a flowchart of data preprocessing according to an embodiment of the present invention;

[0042] Figure 3 It is a semi-automatic annotation flowchart according to an embodiment of the present invention;

[0043] Figure 4 It is a sample diagram of a three-dimensional change detection data set constructed according to an embodiment of the present invention. Detailed implementation manners

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all 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.

[0045] An embodiment of the present invention provides a method for constructing a three-dimensional change point cloud data set for an urban scenario, as Figure 1 shown, including:

[0046] Step 1: Collect multi-temporal point cloud data for an urban scenario and perform preprocessing;

[0047] Step 2: Coarsely screen the preprocessed multi-temporal point cloud data using the octree algorithm, specifically including:

[0048] First, construct an octree for the point cloud data of each time series;

[0049] Second, extract features for the point cloud region corresponding to each node in the octree;

[0050] Finally, based on the extracted features, compare the feature differences of each node in the octrees of different time series to determine the point cloud data that has changed and the point cloud data that has not changed;

[0051] Step 3: Based on the multi-temporal 3D change point cloud data after rough screening, perform manual annotation and cross-validation to obtain the finally annotated multi-temporal point cloud data, thus completing the construction of the 3D change point cloud dataset for urban scenes.

[0052] Specifically, the embodiments of the present invention also provide the acquisition of multi-temporal point clouds. The urban-level 3D change detection dataset requires being oriented to urban scenes and needs to use airborne equipment for data acquisition. In addition, change detection requires observing the same area multiple times at different time series. When collecting data at different times, it is necessary to ensure that the performance and parameters of the equipment remain consistent as much as possible each time data is collected, and try to avoid false change samples caused by the acquisition equipment to ensure the accuracy of the change data. The specific implementation process is described as follows:

[0053] First, for UAV data acquisition, it is necessary to select the specific range of the area where data is to be collected, and determine the relevant flight parameters of the UAV according to the surface undulation characteristics of the area, namely flight altitude, flight speed, aerial survey tilt angle, etc. According to the principle of full coverage, delimit the UAV flight path trajectory, and determine whether it is necessary to adopt cross flight paths according to the complexity of the area. After the aerial survey task is completed, it is necessary to clean the image data or LiDAR point cloud to ensure the correctness of the experimental data. Finally, according to the principle of single variable, repeat the acquisition of the same area at different time intervals to obtain multi-temporal data. Then, to solve the problem that it is time-consuming and laborious to obtain data by UAVs, a method of multi-UAV collaborative acquisition is adopted to obtain experimental data. However, when multiple UAVs obtain data simultaneously, there will be problems with data ovules. Therefore, a data fusion method for multi-UAVs is proposed, which will fuse the data of multiple UAVs with different starting points based on ground control points, and filter the point cloud boundary through statistical data to eliminate the influence brought by different data boundaries. Finally, according to the oblique photography images of the relevant area, the image pose parameters and camera parameters obtained by aerial triangulation are used to obtain the 3D point cloud of the relevant area through dense image matching. Currently, 3D point cloud reconstruction is a relatively mature engineering technology, which can be achieved through algorithm design or through automatic 3D modeling software, such as Smart3D, Photoscan, etc. After reconstruction, the multi-temporal real-scene point cloud data for urban scenes can be obtained, providing a rich data basis for subsequent data annotation.

[0054] Specifically, the implementation plan for the acquisition of multi-temporal point clouds provided by the embodiments of the present invention is described as follows: Taking the acquisition of the 3D real-scene point cloud of the Information School of Wuhan University as an example, taking the DJI M300 RTK + Seer 102S as an example. First, according to the average surface height of the Information School of Wuhan University and the requirements of the actual size of each pixel, determine the ground resolution to be 1.5 cm, the flight altitude to be 96 meters, the flight speed to be 11.9 m / s, and the operation area to be 0.35 km 2, the workload per day is 5.3 km 2 , the oblique photography image acquisition of the Information School of Wuhan University can be completed within one day. If multi-aircraft joint acquisition is required, it should be noted that during the flight line design process, the forward overlap rate should generally be 60% - 80%, and the minimum should not be less than 53%; the side overlap rate should generally be 15% - 60%, and the minimum should not be less than 8%. When dealing with areas with sparse buildings, for the survey area of the UAV, the forward and side overlap rates should be at least not less than 70%; while in the face of areas with dense buildings, the image overlap rate can be designed up to 80% - 90%. After the acquisition is completed, the three-dimensional point cloud reconstruction can be carried out. Based on the Smart3D algorithm, a control interval can be given from the perspective of the final aerial triangulation feature point cloud, that is, a control point is arranged every 20,000 to 40,000 pixels. If there is differential POS data (which belongs to a relatively accurate initial value), then this interval can be relaxed to 40,000 pixels; if there is no differential POS data, then at least one control point should be arranged every 20,000 pixels. And, this rule should also be flexibly applied according to the actual terrain and ground features of each task. For example, in the case of extremely large terrain undulations, large areas of vegetation, and extremely few feature points in planar water areas, the number of control points should be appropriately increased as the situation requires, and the control points are measured by the method of connecting traverse survey to obtain high-precision position information.

[0055] Specifically, the embodiment of the present invention also provides the preprocessing of the three-dimensional point cloud. Before data annotation, it is also necessary to perform data preprocessing on the multi-temporal point cloud data collected in the above steps, mainly including data cleaning and data registration. Data cleaning is mainly to remove the noise points and outliers contained in the data, reduce the influence of noise data on the main data, and improve the quality and accuracy of the data. Data registration is to unify the multi-temporal point cloud data into the same coordinate system, clarify the corresponding relationship between different temporal point clouds, and facilitate the subsequent change annotation. The overall process is as Figure 2 shown. The specific implementation process is described as follows:

[0056] First, data cleaning is performed. Statistical filtering is mainly used to clean the point cloud data, removing the included noise points and outliers, reducing the impact of noise data on the main data, and improving the quality and accuracy of the data. Statistical filtering reduces the impact of noise on the main data by analyzing the statistical characteristics in the point cloud and improving the quality of the point cloud data. It is necessary to determine the neighborhood range of each point in the given input point cloud. For each point in the point cloud, a suitable neighborhood range is set. For example, with a certain point as the center, a spherical neighborhood with a certain radius is determined, or a cubic neighborhood is delimited according to the spatial coordinate range, etc. The setting of the size of this neighborhood range will be comprehensively determined based on the characteristics of the point cloud data, the scale of the target object, and the noise situation, etc. The purpose is to prepare for subsequent statistics of the relevant characteristics of the points within the neighborhood. Within the determined neighborhood, some key characteristics of each point are statistically analyzed. Common characteristics include the distance from this point to other points within the neighborhood, etc. For example, calculate the distance values from a certain point in the point cloud to all other points within its spherical neighborhood, and then statistically analyze statistics such as the mean and standard deviation of these distance values. Then, according to the statistically obtained characteristic statistics and the cleaning effect that the point cloud data wants to achieve, corresponding thresholds are set. For example, a multiple of the distance mean is set as the distance threshold. If the distance from a certain point to other points within its neighborhood is greater than this set threshold, it is initially determined that this point may be a noise point. The setting of the threshold usually requires certain experiments and adjustments, and different point cloud data may have different suitable threshold ranges. Finally, according to the set threshold, the points marked as noise points or abnormal points are deleted from the original point cloud data, thus obtaining point cloud data that has been cleaned and has relatively less noise. Then, point cloud registration is performed, including two steps: manual rough alignment and automatic fine registration. Since there may be large differences in the coordinate systems of point clouds at different time series, resulting in a large distance between multi-temporal point clouds, it is necessary to first manually select multiple feature points, and based on the corresponding relationships between these feature points, complete the preliminary alignment of multi-temporal point clouds. On this basis, the Iterative Closest Point algorithm is used to achieve the precise registration of multi-temporal point clouds.

[0057] Specifically, the implementation scheme of the preprocessing of the three-dimensional point cloud provided by the embodiment of the present invention is as follows:

[0058] Taking the real - scene point cloud of the Information School of Wuhan University obtained above as an example. First, data cleaning is carried out. This method judges whether a point is a noise point by calculating the distance from each point to the underlying surface of the point. If a point is too far from the fitted plane, it is judged as a noise point, which can be regarded as a low - pass filter. When estimating the distance from the fitted plane, it is first necessary to determine the neighborhood points of each point. This method determines the neighborhood set of any point in the point cloud set through K - nearest neighbors. According to the density of the point cloud, the number of nearest neighbors is determined to be 6 for K - nearest neighbor judgment. In addition, an input maximum error also needs to be determined during the judgment to determine whether the input point is judged as a noise point. The error can be a relative error or an absolute error. In this embodiment, the relative error is adopted, and the input maximum error is set to 1.0. Then, data registration is carried out. When manually aligning roughly, 3 feature points are selected in this embodiment. Try to select feature - obvious points such as the inflection points of buildings for rough alignment. Then the Iterative Closest Point algorithm is adopted, and the RMS difference is set to 1.0×10 -5 , and finally the coverage rate reaches 95%. The average distance between multi - temporal point clouds after registration is small, reaching 0.2m, effectively realizing the correspondence of multi - temporal data.

[0059] Specifically, the embodiment of the present invention also provides a method for roughly screening change information using the octree algorithm. Based on the output multi - temporal point cloud pair, the change information is labeled. In order to reduce the difficulty and cost of change information labeling as much as possible, a semi - automated labeling method is proposed. For point clouds of different time series, it is necessary to set the resolution of octree detection according to their distribution characteristics, balance the false changes caused by different detection accuracies and densities, and screen out as many true changes as possible. The process of semi - automated labeling is as Figure 3 shown. The specific implementation process is described as follows:

[0060] The octree algorithm can effectively screen out possible changes in multi-temporal point clouds, providing an information guidance for subsequent data annotation, thereby reducing the difficulty of data annotation. First, it is necessary to construct an octree for the point cloud of each time series according to parameters such as the set maximum recursion depth. During the construction process, according to the spatial range covered by the point cloud data, determine the cubic space represented by the root node of the octree, and then recursively divide the cubic space into eight sub-cubes until the predetermined maximum depth is reached or the number of point clouds in each sub-cube meets a certain threshold, thus constructing a complete octree structure. Then, for each node of the octree or the point cloud region corresponding to the node, extract relevant features, such as the density, normal vector, curvature, etc. of the point cloud, and these features will be used as the basis for judging changes. Finally, the octree forms voxels by dividing the space, and then for the points located within the voxels, calculate the density according to the number of points and the volume of the voxel (since it is a voxel, the distribution position is ignored), which is used as the comparison feature of the octree node, and make a comparison at the corresponding position to determine the changed area and degree in the three-dimensional space. A certain threshold can be set, and when the feature difference exceeds the threshold, it is determined that a change has occurred.

[0061] Specifically, the implementation scheme for roughly screening change information using the octree algorithm provided by the embodiments of the present invention is as follows:

[0062] Use the Point Cloud Library (PCL) open-source library to complete three-dimensional change detection based on the octree. Use the PointCloud() data structure to store the read multi-temporal point cloud data, and use the make_octreeChangeDetector() function to construct an octree structure for the point cloud data of the first period. According to the density of the input point cloud, set the resolution of the octree to 3.0. Then, read the point cloud data of the second period in the same way, and based on the octree structure of the point cloud data of the first period, input the point cloud of the second period into this octree structure. Use the get_PointIndicesFromNewVoxels() function to obtain the changes in the point cloud of the second period, and similarly, the changes in the point cloud of the first period can be obtained. When outputting, it should be noted that the changed points and unchanged points are output separately to facilitate the subsequent correction of the changed data.

[0063] Specifically, the embodiments of the present invention also provide manual correction of change information. The roughly screened change information obtained in the previous step is not accurate, and it is necessary to manually check this change information, correct false change samples, and supplement true change samples. Then, cross-validate the different manually processed data to ensure the consistency and accuracy of the data. In addition, during the data annotation process, it is also necessary to clarify unified annotation principles to ensure that the marked changes can provide effective, clear, and consistent information support. An example diagram of the marked three-dimensional change data set is as Figure 4As shown below. The specific implementation process is described as follows:

[0064] Manually correcting the changed information mainly involves checking and correcting the coarsely screened changes. To ensure the usability and effectiveness of the data, the data correction needs to meet the following principles: 1) There must be clear changes at corresponding positions in point clouds at different time series, and these changes need to conform to the requirements of social and economic activities, such as building project management, tree monitoring, etc. It should be noted that the inconsistent point cloud density caused by sampling and occlusion cannot be classified as a change. 2) Focus on three types of changes, namely addition, reduction, and replacement. Addition refers to a change from non - existent to existent, such as the addition of a house; reduction refers to a change from existent to non - existent, such as the withering of a tree; replacement refers to a change in the object category at the same position, such as a change from a tree to a house. During the correction process, following the process of first correcting misclassified points and then supplementing the changed points can better complete the data correction.

[0065] Specifically, the implementation scheme for manually correcting the changed information provided by the embodiments of the present invention is as follows:

[0066] Manually correcting the changed information is a data annotation task, and the embodiments of the present invention use Cloud Compare software to implement it. For the multi - time - series real - scene point cloud data of the Information School of Wuhan University, the correction is carried out step by step according to the time series. When processing the point cloud of each time series, first check the point cloud determined to be changed by the octree algorithm, compare the point clouds of different time series, correct the false changed data caused by reasons such as density and occlusion, and classify it as unchanged data. This type of data is mostly found in objects such as buildings and vegetation. Then check the point cloud determined to be unchanged by the octree algorithm, also compare the point clouds of different time series, correct the false unchanged samples caused by improper octree resolution settings, and classify it as changed data. This type of data is mostly found in small - volume objects in the city, such as vehicles, street facilities, etc. Finally, cross - check the changed data annotated by different personnel to further ensure the accuracy of the changed data annotation.

[0067] The implementation basis of each embodiment of the present invention is achieved through programmed processing by a device with processor functions. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the present invention are encapsulated into various modules. Based on this actual situation, on the basis of the above - mentioned embodiments, the embodiments of the present invention provide a fast construction system for a three - dimensional change detection data set for urban scenes, and this system is used to execute the method for constructing a three - dimensional change point cloud data set for urban scenes in the above - mentioned method embodiments.

[0068] The system includes: a data acquisition module, which is used to acquire multi-temporal three-dimensional change point cloud data for urban scenes and perform preprocessing; a rough screening module, which uses the octree algorithm to roughly screen the preprocessed multi-temporal three-dimensional change point cloud data; a data set construction module, which, based on the roughly screened multi-temporal three-dimensional change point cloud data, performs manual annotation and cross-validation to obtain the finally well-annotated multi-temporal point cloud data, and completes the construction of the three-dimensional change point cloud data set.

[0069] The system for quickly constructing a three-dimensional change detection data set for urban scenes provided by the embodiment of the present invention aims at the problems of too high cost and difficulty of data annotation and untrue data. It adopts several modules to complete the preprocessing of data through methods such as data cleaning and point cloud registration, then uses the octree algorithm to roughly screen the change information of multi-temporal point cloud data, manually annotates and corrects the roughly screened change information, and performs cross-validation on the annotated data to ensure the accuracy of the data, so as to efficiently and accurately realize the construction of the three-dimensional change point cloud data set.

[0070] Based on the same inventive concept as the foregoing embodiment, the embodiment of the present invention also provides an electronic device, including a memory and a processor. The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement a method for constructing a three-dimensional change point cloud data set for urban scenes as proposed in the above embodiment.

[0071] The embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it overcomes the problems of too high cost and difficulty of data annotation, obtains a real data set describing the changes in the urban real scene, and realizes the construction of the three-dimensional change point cloud data set. The storage medium can be any non-volatile storage device such as a hard disk, a solid-state drive, a flash drive, an optical disc, etc., which is used to store computer program codes and necessary data files. The stored computer program includes: a data acquisition module, a rough screening module, and a data set construction module.

[0072] Finally, it should be pointed out that the above specific embodiments are only relatively representative examples of the present invention. Obviously, the present invention is not limited to the above specific embodiments and there can be many variations. Any simple modification, equivalent change and modification made to the above specific embodiments according to the technical essence of the present invention should be considered as belonging to the protection scope of the present invention.

Claims

1. A method for constructing a three-dimensional change point cloud dataset for urban scenes, characterized in that: include: Step 1: Collect multi-time point cloud data for urban scenes and perform preprocessing; Step 2: Use the octree algorithm to roughly screen the preprocessed multi-time series point cloud data, including: First, according to the spatial range of the point cloud data, an octree is constructed for each time series of point cloud data. Secondly, extract features for each node in the octree, the features include the number of nodes and node density; Finally, based on the number of nodes and node density, the feature differences of each node in different time series octrees are compared to determine the point cloud data that has changed and has not changed; Step 3: Based on the roughly screened multi-time series 3D change point cloud data, annotation and cross-validation are performed to obtain the final annotated multi-time series point cloud data and the constructed 3D change point cloud data set for urban scenes.

2. The method for constructing a three-dimensional change point cloud dataset for urban scenes according to claim 1, characterized in that: The construction of the octree includes: According to the spatial range of the point cloud data, determine the cube space represented by the root node of the octree; Based on the cubic space, eight sub-cubes are obtained by partitioning, and it is determined that each sub-cube reaches a predetermined maximum recursive depth or the number of point clouds in each sub-cube meets a certain threshold, and the constructed octree is obtained.

3. The method for constructing a three-dimensional change point cloud dataset for urban scenes according to claim 1, characterized in that: Identify point cloud data that has changed and has not changed, including: The nearest neighbor algorithm is used to determine the corresponding position of each node in the octree of different time series, and the threshold is determined according to the point cloud density; Different time series octrees form voxels by dividing the space. The node density is calculated according to the number of nodes and the volume of the voxels, and the node density is used as the feature difference for comparison; When the feature difference exceeds the threshold, it is determined as the point cloud data that has changed; when the feature difference does not exceed the threshold, it is determined as the point cloud data that has changed.

4. The method for constructing a three-dimensional change point cloud dataset for urban scenes according to claim 1, characterized in that: Perform annotation and cross-validation, including: Label the multi-time series point cloud data after rough screening, and correct the unchanged data to the changed data; The data after different manual corrections are cross-validated to obtain the final labeled multi-time series point cloud data.

5. The method for constructing a three-dimensional change point cloud dataset for urban scenes according to claim 1, characterized in that: After preprocessing, point cloud registration is also included: Select multiple feature points to perform preliminary alignment on the multi-time series point cloud data, and use the iterative closest point algorithm to perform precise registration on the preliminarily aligned point cloud data.

6. A system for constructing a three-dimensional change point cloud dataset for urban scenes, characterized in that: include: The data acquisition module is used to collect multi-time series 3D change point cloud data for urban scenes and perform pre-processing; The coarse screening module is used to use the octree algorithm to perform coarse screening on the preprocessed multi-time series point cloud data, including: First, according to the spatial range of the point cloud data, an octree is constructed for each time series of point cloud data. Secondly, extract features from the point cloud area corresponding to each node in the octree, where the features include point cloud density, normal vector, and curvature; Finally, based on the point cloud density, normal vector, and curvature, the feature differences of each node in the octrees of different time series are compared to determine the point cloud data that has changed and has not changed; The data set construction module is used to perform annotation and cross-validation based on the coarsely screened multi-time series three-dimensional change point cloud data, obtain the final annotated multi-time series point cloud data, and obtain a constructed three-dimensional change point cloud data set for urban scenes.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for constructing a three-dimensional change point cloud data set for urban scenes as described in any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for constructing a three-dimensional change point cloud data set for urban scenes as described in any one of claims 1 to 5 are implemented.

Images