Three-dimensional data processing method for VR virtual scene reconstruction

Abstract

The invention relates to the technical field of three-dimensional reconstruction, in particular to a three-dimensional data processing method for VR virtual scene reconstruction. The method comprises the following steps: acquiring three-dimensional point cloud data; obtaining a data feature value of each data point according to the data value of the data point in the neighborhood window of each data point, and screening feature points; obtaining the redundancy probability of the target feature point according to the similarity between the target feature point and the data features of the adjacent points on the first line segment and the second line segment, and screening a final feature point sequence; and determining the structural similarity between the feature points and the feature matching points according to the feature vectors of the feature points and the feature matching points in the matching group so as to obtain the retention probability of the feature points and perform three-dimensional point cloud data registration, thereby completing the three-dimensional reconstruction of the virtual scene. According to the invention, through analyzing and screening the point cloud data, the efficiency and precision of scene three-dimensional reconstruction are improved.

Description

Technical Field

[0001] This invention relates to the field of 3D reconstruction technology, and more specifically to a 3D data processing method for VR virtual scene reconstruction. Background Technology

[0002] VR virtual scene reconstruction transforms real-world objects into digital 3D models, allowing users to interact with each other and with software agents through digital avatars within virtual environments. This significantly enhances the user's immersive experience. Therefore, 3D reconstruction technology, which creates computer-generated 3D models of physical objects, is a crucial research direction for the development of VR technology and plays a vital role in various related application areas such as virtual reality and augmented reality. The quality of the 3D model directly impacts the user experience.

[0003] In 3D reconstruction, point cloud registration is often required. When registering point cloud data using the ICP algorithm, feature point extraction is necessary. Harris 3D feature points are sensitive to rotation and viewpoint changes during feature point extraction, which may lead to inconsistencies in the detected feature points for the same object under different viewpoints. This affects the efficiency and accuracy of subsequent feature point matching, and low registration accuracy directly impacts the efficiency and quality of subsequent 3D modeling. Summary of the Invention

[0004] To address the technical problem of poor matching results caused by unstable feature point extraction, this invention provides a 3D data processing method for VR virtual scene reconstruction. The specific technical solution adopted is as follows: This invention proposes a three-dimensional data processing method for VR virtual scene reconstruction, which includes the following steps: Acquire 3D point cloud data; Based on the data values ​​of data points within the neighborhood window of each data point, obtain the data feature value of each data point, and use the data points whose data feature value is greater than the threshold as feature points; Obtain the first and second main orientations for each set of 3D point cloud data; obtain the redundancy probability of the target feature points based on the similarity of the data features between the target feature points and their adjacent points on the first and second line segments; obtain the final feature point sequence based on the redundancy probability. Feature matching points are obtained from the data points within the neighborhood window of the feature points in the final feature point sequences of different groups; the structural similarity between the feature points and the feature matching points is determined based on the feature vectors of the feature points and the feature matching points within the matching group; the retention probability of the feature points is obtained based on the structural similarity, the first distance, and the first angle difference of the three-dimensional point cloud data of different groups. Based on the retention probability of feature points, 3D point cloud data registration is performed to complete the 3D reconstruction of the virtual scene.

[0005] Preferably, the method for obtaining the data feature value of each data point is as follows: The normal vector of each data point is obtained by least squares plane fitting. The derivatives of the normal vector in different directions are calculated as the projections of the normal vector in different directions. The curvature of each data point is obtained based on the projections. For any given data point, a neighborhood window is constructed with the data point as the center and a preset value as the radius. The absolute value of the difference between the data value of each data point in the neighborhood window and the center data point is calculated and recorded as the first absolute value. The absolute value of the difference between the curvature of each data point in the neighborhood window and the center data point is calculated and recorded as the second absolute value. The product of the first absolute value and the second absolute value is accumulated over all data points in the neighborhood window and multiplied by the curvature of the center data point as the data feature value of the center data point.

[0006] Preferably, the method for obtaining the first main direction and the second main direction of each set of 3D point cloud data is as follows: For each set of 3D point cloud data, the RANSAN algorithm is used to fit each set of 3D point cloud data to obtain a plane. The first principal component and the second principal component are obtained by the PCA algorithm. The feature directions corresponding to the first principal component and the second principal component are denoted as the first principal direction and the second principal direction, respectively. The first principal direction and the second principal direction are perpendicular to each other.

[0007] Preferably, the method for obtaining the redundancy probability of the target feature points is as follows: Any feature point is recorded as the target feature point. Based on the neighboring points determined in the first subject direction and the second subject direction, the degree of feature difference between the target feature point and the neighboring points is obtained. The sum of the degree of feature difference between the target feature point and all its neighboring points, plus the sum of the constant parameter, is used as the denominator; The ratio of the number of adjacent points on the first line segment and the second line segment of the target feature point to the denominator is used as the redundancy probability of the target feature point.

[0008] Preferably, the method for obtaining the degree of feature difference between the target feature point and its neighboring points is as follows: Within the neighborhood window of the target feature point, for each target feature point, a line segment is obtained in the first main direction and the second main direction, and they are respectively denoted as the first line segment and the second line segment. Data points located on the first line segment and the second line segment are all regarded as the neighboring points of the target feature point. The absolute value of the difference between the data feature values ​​of the target feature point and its neighboring points is used as the first difference. The absolute value of the difference between the mean of all data points in the neighborhood window of the target feature point and the mean of all data points in the neighborhood window of the adjacent point is used as the second difference. The product of the first difference and the second difference is used as the degree of feature difference between the target feature point and its neighboring points.

[0009] Preferably, the method for obtaining the final feature point sequence is as follows: Target feature points with a redundancy probability greater than the judgment threshold are removed, while target feature points with a redundancy probability less than or equal to the judgment threshold are retained. The final feature point sequence is composed of all the retained target feature points.

[0010] Preferably, the method for obtaining feature matching points is as follows: In the final feature point sequence, any final feature point is designated as the first feature point. The first feature point and each feature point in the final feature point sequence are combined to form a feature vector, and the feature vector sequence of each first feature point is obtained. Choose any remaining set of 3D point cloud data as the data to be matched for the first feature point, and take any one of the feature points in the final feature point sequence corresponding to the data to be matched as a second feature point. The mean values ​​of all data points within the neighborhood windows of the first feature point and the second feature point are calculated respectively as the first mean and the second mean. The second feature point corresponding to the second mean with the smallest absolute value of the difference between the first and second mean is taken as the feature matching point of the first feature point. The feature matching points of the first feature point in each group of 3D point cloud data are obtained in turn.

[0011] Preferably, the method for determining the structural similarity between feature points and feature matching points is as follows: Calculate the Euclidean distance between the feature vectors of the first feature point and the feature vector sequence of the feature matching point respectively. Take the two feature vectors with the smallest Euclidean distance as a matching group. The number of matching groups is consistent with the minimum number of feature vectors in the two feature vector sequences. Calculate the angle between the two feature vectors in each matching group determined by the first feature point and each feature matching point, and the absolute value of the difference in magnitude between the two feature vectors; The angle between the two feature vectors and the absolute value of the difference between the magnitudes of the two feature vectors are calculated by weighted fusion. The weighted calculation result is accumulated on all matching groups determined by the first feature point and each feature matching point as the structural similarity between the first feature point and each feature matching point.

[0012] Preferably, the method for obtaining the retention probability of feature points is as follows: The absolute value of the difference between the first distance corresponding to the three-dimensional point cloud data where the first feature point is located and the first distance corresponding to the three-dimensional point cloud data where each feature matching point is located is used as the first metric. The absolute value of the difference between the first angle corresponding to the three-dimensional point cloud data where the first feature point is located and the first angle corresponding to the three-dimensional point cloud data where each feature matching point is located is used as the second metric. Use the product of the first and second measures and the sum of the constant parameter as the denominator; The sum of the ratio of the structural similarity between the first feature point and each matching feature point to the denominator over all matching feature points is used as the retention probability of the first feature point.

[0013] Preferably, the method for registering 3D point cloud data based on the retention probability of feature points is as follows: For a set of 3D point cloud data, perform one-dimensional K-means clustering on the retention probability of all feature points. The clustering distance is the absolute value of the difference in the retention probability of feature points. Divide the feature points into two classes, calculate the mean retention probability in each class, and select the feature points contained in the class with the largest mean retention probability as the perfect feature points. The improved feature points are sequentially registered with 3D point cloud data, reconstructed using Poisson data, and then textured to complete the 3D reconstruction.

[0014] The present invention has the following beneficial effects: Addressing the instability of detected feature points under different viewpoints due to the sensitivity to rotation and perspective changes in existing feature point extraction algorithms, the present invention first obtains the initial feature points by utilizing the curvature of each data point. Then, it considers all feature points in the neighboring area of ​​each feature point and performs feature point filtering based on the fact that feature points on building surfaces lie on a single plane. Finally, it analyzes the similarity of multiple sets of pixels collected from different angles and extracts feature points again. This solves the problem of instability of small feature points under different viewpoints. The extracted, better feature points are then used for registration, improving the efficiency and accuracy of registration, thereby enhancing the efficiency of 3D reconstruction of building models in scene reconstruction. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a three-dimensional data processing method for VR virtual scene reconstruction provided in one embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a three-dimensional data processing method for VR virtual scene reconstruction proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] 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 invention pertains.

[0019] An embodiment of a 3D data processing method for VR virtual scene reconstruction: The following description, in conjunction with the accompanying drawings, details a specific scheme for a three-dimensional data processing method for VR virtual scene reconstruction provided by the present invention.

[0020] Please see Figure 1 The diagram illustrates a flowchart of a three-dimensional data processing method for VR virtual scene reconstruction according to an embodiment of the present invention. The method includes the following steps: Step S001: Obtain 3D point cloud data.

[0021] When reconstructing a VR virtual scene, it is necessary to acquire a real-world architectural model, convert it into 3D point cloud data for analysis, and then reconstruct the VR virtual scene based on the analysis results. In this application, firstly, a stereo camera is used to acquire multiple images of the target scene from different angles. Secondly, a depth map is obtained through a stereo matching algorithm, ensuring that the captured images have overlapping parts. Each depth map is converted into a set of 3D point cloud data using existing technology. As an example, this embodiment uses the NumPy library in Python to load and read the depth map, and then uses OpenCV and Open3D to assign the depth map reading information amplitude to the point cloud data points, obtaining a set of corresponding 3D point cloud data. Afterwards, guided filtering is applied to each set of 3D point cloud data to obtain denoised 3D point cloud data. The stereo matching algorithm and guided filtering are commonly used techniques in the field of image processing, and the specific process will not be elaborated further.

[0022] At this point, the denoised 3D point cloud data was obtained.

[0023] Step S002: Obtain the data feature value of each data point based on the data value of the data points in the neighborhood window of each data point, and filter feature points based on the data feature value.

[0024] After denoising each set of 3D point cloud data, the large number of data points results in excessive data to be registered using the ICP algorithm, leading to low registration efficiency and a high risk of misregistration. Low registration accuracy directly impacts the efficiency and quality of subsequent 3D virtual scene modeling. Therefore, it is necessary to analyze the 3D point cloud data and extract its feature points.

[0025] For each set of 3D point cloud data, the normal vector of each data point is obtained through least squares plane fitting. The derivatives of the normal vector in different directions are calculated as the projections of the normal vector in different directions. The rate of change of the projection and the curvature in each direction are calculated to obtain the curvature of each data point. The curvature calculation method is a well-known technique and will not be elaborated here. Any data point is designated as the center data point. A neighborhood window is constructed with the center data point as the center. The radius of the neighborhood window is set to 5 in this embodiment, but this value can be set by the technician according to the scenario. The data feature value of the center data point is obtained by the difference in data value and curvature between the center data point and each data point in its neighborhood window, as shown in the following formula: In the formula, The data value representing the center data point. This represents the data value of the i-th data point within the neighborhood window. Represents the curvature of the center data point. This represents the curvature of the i-th data point within the neighborhood window. This indicates the number of data points within the neighborhood window. The data feature value representing the central data point.

[0026] Specifically, the greater the difference in data value and curvature between the central data point and the other data points in its neighborhood, the more prominent the central data point is, and the more likely it is to be a feature point, with a larger data feature value. Linear normalization is performed on the data feature value of each central data point, and a feature threshold is set. If the data feature value of the central data point is greater than or equal to the feature threshold, then the central data point is considered a feature point. In this embodiment, the feature threshold is set to 0.75.

[0027] At this point, the feature points in each set of 3D point cloud data have been extracted.

[0028] Step S003: Obtain the first main direction and the second main direction of each group of 3D point cloud data; obtain the redundancy probability of the target feature point based on the similarity of the data features of the target feature point and the adjacent points on the first and second line segments; obtain the final feature point sequence based on the redundancy probability.

[0029] For each set of acquired 3D point cloud data, a set of feature point sequences can be extracted from each set of 3D point cloud data. The feature point sequences are sorted in ascending order according to the data feature values ​​of each feature point. Since there are many repetitive structures on the building surface, the feature points of the extracted 3D point cloud data have a certain redundancy probability. When matching, this not only wastes computing resources, but also makes it easier for mismatches to occur.

[0030] For each set of data points, the RANSAN algorithm is used to fit and obtain a plane, which is called the feature plane. The first principal component and the second principal component of all data points in the feature plane are obtained by the PCA algorithm. The feature directions corresponding to the first principal component and the second principal component are denoted as the first principal direction and the second principal direction, respectively. The first principal direction and the second principal direction are perpendicular to each other. The PCA algorithm is a commonly used technique in the field of data processing, and the specific process will not be described in detail.

[0031] Let any feature point be designated as the target feature point. Within the neighborhood window of the target feature point, for each target feature point, obtain a line segment in the first main direction and a line segment in the second main direction, respectively designated as the first line segment and the second line segment. Feature points on the first and second line segments of the target feature point are designated as adjacent points. The redundancy probability of the target feature point is obtained based on the number of feature points in the two lines of the target feature point, the data feature value, and the data value of each data point in the neighborhood window. The specific process is as follows: In the formula, It represents the degree of feature difference between the t-th target feature point and its c-th neighboring point. It is the data feature value of the t-th target feature point. It is the data feature value of the c-th neighboring point. It is the mean of all data points within the neighborhood window of the t-th target feature point. It is the mean of all data points within the neighborhood window of the c-th nearest neighbor.

[0032] Among them, the more similar the data features of the adjacent points on the first and second line segments taken from the t-th target feature point are to the t-th target point, the higher the first difference value. Second difference The smaller the value, the lower the degree of feature difference between the two data points, the more redundant the data features of the t-th target feature point, and the smaller the impact on the scene reconstruction result.

[0033] Then, the redundancy probability of the t-th target feature point is determined based on the degree of feature difference between all t-th target feature points and all their neighboring points: In the formula, It is the redundancy probability of the t-th target feature point. It represents the number of adjacent points on the first line segment and the second line segment of the t-th target feature point. It is the sum of the degree of feature difference between all t-th target feature points and all their neighboring points; It is a constant parameter, a positive number not exceeding 0.01, in this embodiment... Take the empirical value of 0.001.

[0034] The more similar the data features of the t-th target feature point are to the data features of the adjacent points on the first and second line segments, the closer the projections of the t-th target feature point and the adjacent points are to the feature directions corresponding to the first and second principal component components. The more substitutable the data features of the t-th target feature point are to the data features of the adjacent points, the higher the redundancy probability of the t-th target feature point, and the lower its necessity for matching.

[0035] Furthermore, a redundancy threshold of 0.7 is set. If the redundancy probability of a target feature point is greater than the redundancy probability threshold, the point is removed from the feature point sequence; otherwise, the feature point is retained. This completes the filtering of the feature point sequence to obtain the final feature point sequence.

[0036] At this point, the final feature point sequence has been obtained.

[0037] Step S004: Obtain feature matching points based on data points within the neighborhood window of feature points in different groups of final feature point sequences; determine the structural similarity between feature points and feature matching points based on the feature vectors of feature points and their feature matching points within the matching group; obtain the retention probability of feature points based on the structural similarity, the first distance, and the first angle difference of different groups of 3D point cloud data.

[0038] Therefore, a final feature point sequence can be obtained for each set of 3D point cloud data. Since different sets of 3D point cloud data are obtained from the same scene at different angles, the final feature point sequences corresponding to different sets of 3D point cloud data are different. Images obtained by the camera at different angles are not completely consistent, resulting in low matching efficiency and accuracy when performing feature point matching. Furthermore, among different sets of 3D point cloud data obtained at different angles, the closer the angles are, the more likely there will be duplicate feature points. Preset external parameters for camera acquisition, namely the distance between the camera and the scene and the camera angle, are denoted as the first distance and the first angle, respectively. Optimization and filtering are then performed based on different sets of 3D point cloud data obtained by the camera under different external parameters.

[0039] Specifically, for a set of three-dimensional point cloud data, which corresponds to a final feature point sequence, any feature point in the final feature point sequence is recorded as the first feature point. In the final feature point sequence, the first feature point and each feature point in the final feature point sequence are combined to form a feature vector, and the feature vector sequence of each first feature point is obtained.

[0040] Further, a remaining set of 3D point cloud data is randomly selected as the data to be matched for the first feature point. Any feature point in the final feature point sequence corresponding to the data to be matched is taken as a second feature point. Next, the mean values ​​of all data points within the neighborhood windows of the first and second feature points are calculated as the first mean and the second mean, respectively. The second feature point corresponding to the second mean with the smallest absolute difference from the first mean is taken as the feature matching point of the first feature point. Then, the feature matching points of the first feature point in each set of 3D point cloud data are obtained sequentially.

[0041] Since the number of feature vectors in the feature vector sequences corresponding to the first feature point and its feature matching point are not necessarily the same, the Euclidean distance between the feature vector sequences of the first feature point and the feature vector sequences of the feature matching point is calculated respectively. The two feature vectors with the smallest Euclidean distance are taken as a matching group. The number of matching groups is consistent with the minimum number of feature vectors in the two feature vector sequences.

[0042] The structural similarity between the first feature point and its matching feature point is determined based on the matching groups between the feature vector sequences corresponding to the first feature point and its matching feature point. In the formula, represents the structural similarity between the q-th first feature point and the a-th feature matching point, and m is the number of matching groups determined between the feature vector sequence of the first feature point and the feature vector sequence of the a-th feature matching point. It is the angle between the two feature vectors in the determined i-th matching group. It is the absolute value of the difference in magnitude between the two feature vectors in the determined i-th matching group. , The weights are used when weighting and fusing the absolute values ​​of the differences between the included angle and the modulus, satisfying the following conditions: In this embodiment, it is set to .

[0043] Furthermore, the retention probability of each first feature point is obtained based on the structural similarity between the camera extrinsic parameters and the first feature point and its feature matching points in different groups, as shown in the following formula: In the formula, is the retention probability of the q-th first feature point, and M is the number of groups of the 3D point cloud data. It represents the structural similarity between the q-th first feature point and the a-th feature matching point. , These are the first distance and the first angle corresponding to the 3D point cloud data of the q-th first feature point, respectively. , These are the first distance and the first angle corresponding to the 3D point cloud data of the a-th feature matching point, respectively. This is a constant parameter used to prevent the denominator from being zero. The size is taken as an empirical value of 0.001.

[0044] The greater the structural similarity, the greater the contribution of the first feature point to feature point matching, and the more likely that feature point should be retained, with a higher retention probability. Conversely, the greater the difference in camera extrinsic parameters, the higher the first metric. Second metric The larger the value, the greater the difference between the two sets of images, and the smaller the weight. Conversely, the higher the retention probability, the lower the redundancy probability of the feature point. It can effectively reflect the spatial structural features of the building model surface and has a higher probability of successful matching with other sets of feature points. Therefore, the feature point should be retained.

[0045] The above calculation is performed on each feature point in the final feature point sequence to obtain the retention probability of each feature point.

[0046] Thus, the retention probability of each feature point in each set of 3D point cloud data has been obtained.

[0047] Step S005: Perform 3D point cloud data registration based on the retention probability of feature points to complete the 3D reconstruction of the virtual scene.

[0048] One-dimensional K-means clustering is performed on the retention probabilities of all feature points in a set, with K set to 2. The cluster distance is the absolute value of the difference in retention probabilities between feature points, thus dividing the feature points into two classes. The mean retention probability in each class is calculated, and the feature points in the class with the largest mean retention probability are selected as perfect feature points. Perfect feature points are obtained for each set of 3D point cloud data.

[0049] Next, the ICP algorithm was used for 3D point cloud data registration. After registration, the 3D point cloud data was transformed to the same world coordinate system, and Poisson reconstruction was performed using MeshLab. Then, texture mapping was performed using Blender software to obtain the scene reconstruction result. Poisson reconstruction and texture mapping are well-known techniques in the field of point cloud data processing and will not be elaborated upon here.

[0050] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0051] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A three-dimensional data processing method for VR virtual scene reconstruction, characterized in that, The method includes the following steps: Acquire 3D point cloud data; Based on the data values ​​of data points within the neighborhood window of each data point, obtain the data feature value of each data point, and use the data points whose data feature value is greater than the threshold as feature points; Obtain the first and second main orientations for each set of 3D point cloud data; obtain the redundancy probability of the target feature points based on the similarity of the data features between the target feature points and their adjacent points on the first and second line segments; obtain the final feature point sequence based on the redundancy probability. Feature matching points are obtained from the data points within the neighborhood window of the feature points in the final feature point sequences of different groups; the structural similarity between the feature points and the feature matching points is determined based on the feature vectors of the feature points and the feature matching points within the matching group; the retention probability of the feature points is obtained based on the structural similarity, the first distance, and the first angle difference of the three-dimensional point cloud data of different groups. Based on the retention probability of feature points, 3D point cloud data registration is performed to complete the 3D reconstruction of the virtual scene.

2. The three-dimensional data processing method for VR virtual scene reconstruction as described in claim 1, characterized in that, The method for obtaining the data feature value of each data point is as follows: The normal vector of each data point is obtained by plane fitting using the least squares method. The derivatives of the normal vector in different directions are calculated as the projections of the normal vector in different directions. The curvature of each data point is obtained based on the projections. For any given data point, a neighborhood window is constructed with the data point as the center and a preset value as the radius. The absolute value of the difference between the data value of each data point in the neighborhood window and the center data point is calculated and recorded as the first absolute value. The absolute value of the difference between the curvature of each data point in the neighborhood window and the center data point is calculated and recorded as the second absolute value. The product of the first absolute value and the second absolute value is accumulated over all data points in the neighborhood window and multiplied by the curvature of the center data point as the data feature value of the center data point.

3. The three-dimensional data processing method for VR virtual scene reconstruction as described in claim 1, characterized in that, The method for obtaining the first and second main directions of each set of 3D point cloud data is as follows: For each set of 3D point cloud data, the RANSAN algorithm is used to fit each set of 3D point cloud data to obtain a plane. The first principal component and the second principal component are obtained by the PCA algorithm. The feature directions corresponding to the first principal component and the second principal component are denoted as the first principal direction and the second principal direction, respectively. The first principal direction and the second principal direction are perpendicular to each other.

4. The three-dimensional data processing method for VR virtual scene reconstruction as described in claim 1, characterized in that, The method for obtaining the redundancy probability of target feature points is as follows: Any feature point is recorded as the target feature point. Based on the neighboring points determined in the first subject direction and the second subject direction, the degree of feature difference between the target feature point and the neighboring points is obtained. The sum of the degree of feature difference between the target feature point and all its neighboring points, plus the sum of the constant parameter, is used as the denominator; The ratio of the number of adjacent points on the first line segment and the second line segment of the target feature point to the denominator is used as the redundancy probability of the target feature point.

5. A three-dimensional data processing method for VR virtual scene reconstruction as described in claim 4, characterized in that, The method for obtaining the degree of feature difference between the target feature point and its neighboring points is as follows: Within the neighborhood window of the target feature point, for each target feature point, a line segment is obtained in the first main direction and the second main direction, and they are respectively denoted as the first line segment and the second line segment. Data points located on the first line segment and the second line segment are all regarded as the neighboring points of the target feature point. The absolute value of the difference between the data feature values ​​of the target feature point and its neighboring points is used as the first difference. The absolute value of the difference between the mean of all data points in the neighborhood window of the target feature point and the mean of all data points in the neighborhood window of the adjacent point is used as the second difference. The product of the first difference and the second difference is used as the degree of feature difference between the target feature point and its neighboring points.

6. The three-dimensional data processing method for VR virtual scene reconstruction as described in claim 1, characterized in that, The method for obtaining the final feature point sequence is as follows: Target feature points with a redundancy probability greater than the judgment threshold are removed, while target feature points with a redundancy probability less than or equal to the judgment threshold are retained. The final feature point sequence is composed of all the retained target feature points.

7. The three-dimensional data processing method for VR virtual scene reconstruction as described in claim 1, characterized in that, The method for obtaining feature matching points is as follows: In the final feature point sequence, any final feature point is designated as the first feature point. The first feature point and each feature point in the final feature point sequence are combined to form a feature vector, and the feature vector sequence of each first feature point is obtained. Choose any remaining set of 3D point cloud data as the data to be matched for the first feature point, and take any one of the feature points in the final feature point sequence corresponding to the data to be matched as a second feature point. The mean values ​​of all data points within the neighborhood windows of the first feature point and the second feature point are calculated respectively as the first mean and the second mean. The second feature point corresponding to the second mean with the smallest absolute value of the difference between the first and second mean is taken as the feature matching point of the first feature point. The feature matching points of the first feature point in each group of 3D point cloud data are obtained in turn.

8. The three-dimensional data processing method for VR virtual scene reconstruction as described in claim 7, characterized in that, The method for determining the structural similarity between feature points and feature matching points is as follows: Calculate the Euclidean distance between the feature vectors of the first feature point and the feature vector sequence of the feature matching point respectively. Take the two feature vectors with the smallest Euclidean distance as a matching group. The number of matching groups is consistent with the minimum number of feature vectors in the two feature vector sequences. Calculate the angle between the two feature vectors in each matching group determined by the first feature point and each feature matching point, and the absolute value of the difference in magnitude between the two feature vectors; The angle between the two feature vectors and the absolute value of the difference between the magnitudes of the two feature vectors are calculated by weighted fusion. The weighted calculation result is accumulated on all matching groups determined by the first feature point and each feature matching point as the structural similarity between the first feature point and each feature matching point.

9. A three-dimensional data processing method for VR virtual scene reconstruction as described in claim 1, characterized in that, The method for obtaining the retention probability of feature points is as follows: The absolute value of the difference between the first distance corresponding to the three-dimensional point cloud data where the first feature point is located and the first distance corresponding to the three-dimensional point cloud data where each feature matching point is located is used as the first metric. The absolute value of the difference between the first angle corresponding to the three-dimensional point cloud data where the first feature point is located and the first angle corresponding to the three-dimensional point cloud data where each feature matching point is located is used as the second metric. Use the product of the first and second measures and the sum of the constant parameter as the denominator; The sum of the ratio of the structural similarity between the first feature point and each matching feature point to the denominator over all matching feature points is used as the retention probability of the first feature point.

10. A three-dimensional data processing method for VR virtual scene reconstruction as described in claim 1, characterized in that, The method for registering 3D point cloud data based on the retention probability of feature points is as follows: For a set of 3D point cloud data, perform one-dimensional K-means clustering on the retention probability of all feature points. The clustering distance is the absolute value of the difference in the retention probability of feature points. Divide the feature points into two classes, calculate the mean retention probability in each class, and select the feature points contained in the class with the largest mean retention probability as the perfect feature points. The improved feature points are sequentially registered with 3D point cloud data, reconstructed using Poisson data, and then textured to complete the 3D reconstruction.

Images