3D point cloud simplification method combined with human visual perception characteristics

Abstract

The 3D point cloud simplification method combining human visual perception characteristics belongs to the field of 3D point cloud data processing, and aims to solve the problem that the sharp increase of the amount of dense 3D point cloud data aggravates the burden of data processing, storage and transmission in the later stage, the present application provides a kind of 3D point cloud simplification method combining geometric features and human visual perception characteristics;Combined with the geometric features of point cloud, the algorithm establishes a one-way perception sharpness function and a local visibility function to complete the importance evaluation of points, and then formulates different simplification rules according to the importance of points to realize hierarchical simplification of point cloud.In addition, in order to improve the universality of the mixed feature evaluation model, a dynamic optimization strategy for the weight of each evaluation function is established, and the real-time updating of the weight value is realized based on the feature evaluation result.The experiment verifies the effectiveness of the algorithm, compared with the traditional point cloud simplification algorithm, the algorithm can maximize the retention of local details of point cloud while maintaining the overall uniformity of data.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of 3D point cloud data processing, and particularly relates to a 3D point cloud simplification method combining human visual perception characteristics. BACKGROUND

[0002] The rapid development of 3D reconstruction technology lays a foundation for the acquisition of 3D point cloud, and the continuous improvement of 3D point cloud precision and density makes its application in 3D printing, online detection, target recognition and other tasks more extensive. For different application requirements, various types of dense 3D point clouds all have different degrees of information redundancy, and the use of reasonable 3D point cloud simplification technology can effectively improve the processing, storage and transmission efficiency of later 3D data, so the research on 3D point cloud simplification algorithm has become a hot topic in the field of data processing at present.

[0003] Existing 3D point cloud simplification algorithms mainly include two types: grid-based simplification and point-based simplification. The grid-based simplification algorithm first establishes an irregular grid according to the point cloud distribution, and then removes the redundant grid by formulating rules to achieve point cloud simplification. S.-M. Hur et al. use Delaunay triangulation to remove point data and achieve point cloud data reduction. T KDey et al. propose a method to effectively avoid data oversampling by controlling user input density to improve surface fitting accuracy. Sun Feng et al. propose an algorithm driven by shape approximation error measurement for quantitative simplification of the initial intermediate grid. Li Minglei et al. propose a crease contour extraction method based on grid filtering technology, which realizes the reservation of detail features. Although the grid-based simplification can effectively preserve the overall contour and geometric detail features of the point cloud, the construction of the grid structure requires huge computational overhead, which limits its application in practical tasks.

[0004] Point-based simplification algorithm has low computational complexity, and is the mainstream algorithm for point cloud simplification in recent years. When simplifying data, this kind of algorithm should evaluate the importance of each point of the point cloud to determine whether to retain the point. When evaluating the importance of the point cloud, the existing algorithms are mostly based on a single feature evaluation index. For example, Zang Yufu et al. proposed a method of combining local surface changes with adjacent significant point distribution to extract significant points in the point cloud. Wei Xuan et al. proposed a feature evaluation index based on local entropy of normal angle for point cloud simplification algorithm. Gao Yanfeng et al. proposed a point cloud simplification algorithm using octree coding combined with curvature feature evaluation. Since the single evaluation index is mostly only suitable for specific scenarios or specific distribution of 3D point cloud, it limits the universality of the algorithm. In order to adapt to the trend of increasingly diversified morphology of 3D point cloud, point cloud simplification algorithms based on multiple evaluation indexes are proposed in succession. For example, Ji Chunyang et al. proposed a multi-feature evaluation index combining normal vector difference, projection distance, spatial distance and curvature difference. Yang Yang et al. proposed a point cloud data segmentation method based on normal vector, angular entropy, curvature and density information to realize the simplification of point cloud. Leal et al. proposed a dictionary learning method based on normal vector coordinates, position coordinates and surface curvature to realize the simplification of point cloud. Although these algorithms use different feature evaluation indexes, they all focus on preserving the geometric feature significant area of the point cloud model, ignoring the sensitive area of the point cloud to human vision, and the weight of each feature evaluation index is usually set according to experience. When the shape of the point cloud to be simplified changes greatly, the imbalance of the weight value will reduce the application value of the point cloud in the later stage. SUMMARY

[0005] The present application aims at providing a 3D point cloud simplification method combined with human visual perception characteristics to solve the above technical problems. The algorithm of the present application further strengthens the local detail features of the 3D point cloud by the proposed human visual perception evaluation function on the basis of the compromise between the overall geometric contour and the local detail contour of the 3D point cloud, and establishes a dynamic weight optimization strategy for each evaluation function to enhance the universality of the point cloud simplification model.

[0006] To solve the above technical problems, the technical scheme adopted by the present application is as follows: the 3D point cloud simplification method combined with human visual perception characteristics is carried out according to the following steps:

[0007] Step 1) K-neighborhood search is performed on the point cloud;

[0008] Step 2) the one-way perception sharpness, local visibility, curvature, average distance and projection distance value of each point are calculated, and the weight value of each feature is calculated by using the weight dynamic optimization formula. The mixed feature value of each point is obtained by weighted average of different feature values and corresponding weights;

[0009] Step 3) Classify the point cloud according to the mixed eigenvalue, and set the step-by-step simplification rule to realize the down-sampling of the point cloud at each level;

[0010] Step 4) Fuse the down-sampled data at each level to obtain the simplified point cloud.

[0011] Further, the one-way perception sharpness function in step 2) is defined as follows:

[0012]

[0013] Where p represents the current point, p i represents the K-neighborhood point of p, p(x, y, z) is the coordinate value of the current point, p i (x, y, z) is the coordinate value of the neighborhood point of p, ||p(x, y, z)-p i (x, y, z)|| represents the Euclidean distance between p and its neighborhood point, p(o) represents the coordinate value in the direction of maximum variance of x, y or z, |p(o)-p i (o)| represents the coordinate difference between p and its neighborhood point in the direction of maximum variance of x, y or z, p(o) and p i (o) can be calculated by the following formula:

[0014]

[0015] Where δ 2 [p(x)], δ 2 [p(y)] and δ 2 [p(z)] represent the variance of the coordinate value of the point cloud in x, y and z directions, respectively. If δ 2 [p(x)]> δ 2 [p(y)]> δ 2 [p(z)], the value returned by p(o) is the coordinate value in the x direction.

[0016] Further, the function expression of the local visibility in step 2) is as follows:

[0017]

[0018] Where p j is the current point and its K-neighborhood point, n θ (p j ) represents the normal vector angle of point p j , represents the average value of the normal vector angle of the current point p and its K-neighborhood point, and the normal vector angle n θ (p) can be obtained by the following formula:

[0019]

[0020] wherein, n p is the normal vector of the current point p, is the normal vector of the K-neighborhood of point p, the size of the angle between the normal vector of point p and the normal vector of its K-neighborhood points can reflect the steepness of the local region of the point cloud.

[0021] Further, the expression of the mixed feature evaluation model of each point in the step 2) is as follows:

[0022]

[0023] wherein, respectively represent the unidirectional perception sharpness PSSD function and the local visibility LV function based on the human visual perception characteristics; respectively represent the curvature based on the geometric feature evaluation, the average distance from the current point p to its K-neighborhood points, and the projection distance from the current point p to the fitting plane of its K-neighborhood points; w1-w5 represent the weights of different feature evaluation functions;

[0024] In order to reduce the scale difference between the feature values and ensure the sensitivity of each feature function, each feature value needs to be normalized, and the normalized feature peak value needs to be filtered. For the convenience of description, the normalized feature value is denoted as Simplified as F1(p)-F5(p), formula (6) can be simplified as: The weights w1-w5 of different feature evaluation functions are dynamically optimized, and the specific formula is as follows:

[0025]

[0026] The weights w1-w4 monotonically decrease with x(n), and x(n) monotonically increases with the feature value F n (p), so the greater the feature value F n (p) is, the smaller the corresponding weight w n is. The value range of x(n) is [0, +∞), and the value range of w1-w4 is (0, 0.25].

[0027] Further, the step 3) is progressively simplified according to the following rules:

[0028] First, the mixed feature value of each point in the point cloud is calculated according to the mixed feature evaluation model, and the point cloud is divided into I-level feature points, II-level feature points and III-level feature points according to the size of the mixed feature value; then, different downsampling rules are used for the point cloud of each level to simplify it: the I-level features are all retained, the II-level feature points are simplified by using the hierarchical random sampling method, and the III-level feature points are simplified by using the cuboid grid method.

[0029] The advantages and positive effects of the present application are: the algorithm is used for improving the efficiency of 3D printing model data generation. When the point cloud is subjected to regional saliency detection, the unidirectional perception sharpness and local visibility function is established from the human visual perception characteristics, and the sensitivity of the mixed feature evaluation model is improved by combining the geometric feature evaluation function. At the same time, in order to realize the dynamic optimization of the weight value of each feature evaluation function, the weight real-time updating strategy guided by the feature evaluation result is established. The ablation experiment results show that after adding the unidirectional perception sharpness and local visibility function, the point cloud detail feature reservation effect is better. Compared with other simplification algorithms, from the visual effect, the point cloud packaging effect after the simplification algorithm of the present application has fewer holes and better local detail reservation; from the geometric error index and simplification time, under the condition of not significantly increasing the time complexity, the present application algorithm effectively improves the point cloud simplification accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The algorithm flowchart of the present application.

[0031] Figure 2 It is the contrast of the human eye sensitive area based on different point cloud distribution: (a) "book" point cloud distribution diagram, (b) "cone" point cloud distribution diagram, (c) "book" point cloud packaging effect, (d) "cone" point cloud packaging effect.

[0032] Figure 3 It is the unidirectional perception sharpness under different point cloud distribution: (a) visual sensitive area point cloud distribution, (b) flat area point cloud distribution.

[0033] Figure 4 It is the angle diagram between the normal vector of the current point and the normal vector of its neighborhood point.

[0034] Figure 5 It is the influence of different feature evaluation functions on 3D point cloud simplification: (a1)-(a2) original point cloud distribution and packaging effect, (b1)-(b3) point cloud distribution, packaging effect and deviation distribution after simplification based on geometric features, (c1)-(c3) point cloud distribution, packaging effect and deviation distribution after simplification based on geometric features and PSSD function, (d1)-(d3) point cloud distribution, packaging effect and deviation distribution after simplification based on mixed feature evaluation model.

[0035] Figure 6The simplified results of the "elephant" point cloud: (a1)-(a2) distribution and packaging effect of the original point cloud, (b1)-(b3) distribution, packaging effect and deviation distribution of the point cloud after simplification by the DFPSA algorithm, (c1)-(c3) distribution, packaging effect and deviation distribution of the point cloud after simplification by the bounding box method, (d1)-(d3) distribution, packaging effect and deviation distribution of the point cloud after simplification by the normal vector method, (e1)-(e3) distribution, packaging effect and deviation distribution of the point cloud after simplification by the curvature method, (f1)-(f3) distribution, packaging effect and deviation distribution of the point cloud after simplification by the algorithm of the present invention.

[0036] Figure 7 The simplified results of the "gargoyle" point cloud: (a1)-(a2) distribution and packaging effect of the original point cloud, (b1)-(b3) distribution, packaging effect and deviation distribution of the point cloud after simplification by the DFPSA algorithm, (c1)-(c3) distribution, packaging effect and deviation distribution of the point cloud after simplification by the bounding box method, (d1)-(d3) distribution, packaging effect and deviation distribution of the point cloud after simplification by the normal vector method, (e1)-(e3) distribution, packaging effect and deviation distribution of the point cloud after simplification by the curvature method, (f1)-(f3) distribution, packaging effect and deviation distribution of the point cloud after simplification by the algorithm of the present invention.

[0037] Figure 8 Geometric error and reduction time of “gargoyle” at different reduction rates: (a) maximum distance error, (b) average distance error, (c) relative volume error, and (d) reduction time. DETAILED DESCRIPTION

[0038] In order to make the objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below.

[0039] The overall flow chart of the algorithm of the present invention is as follows: Figure 1 As shown in the figure, step 1) performs a K-neighborhood search on the point cloud; step 2) calculates the one-way perception sharpness, local visibility, curvature, average distance and projection distance value of each point, and uses the weight dynamic optimization formula to calculate the weight value of each feature. The mixed eigenvalue of each point is obtained by weighted averaging different eigenvalues ​​and corresponding weights; step 3) classifies the point cloud according to the mixed eigenvalue, and sets a step-by-step simplification rule to realize the downsampling of point clouds at all levels; step 4) the downsampling data at all levels are fused to obtain the simplified point cloud.

[0040] Visual saliency is a key characteristic of the human visual perception system, describing the distribution of attention or eye movements in a specific scene. Detecting visually salient regions is an important research area in computer vision and computer graphics. Currently, most visual saliency detection work focuses on 2D images or videos, with less work on 3D point cloud visual saliency detection. To more effectively streamline 3D point clouds, this paper establishes two 3D point cloud saliency detection functions based on human visual perception characteristics, as follows:

[0041] a. One-way perception sharpness

[0042] In general, the human eye is more sensitive to sharp areas of 3D point clouds than flat areas. Traditional geometric feature evaluation functions such as curvature and projection distance (PD) can be used to judge sharp areas. However, for some special "sharp" areas, it is necessary to combine visual perception characteristics for analysis. Figure 2 (a) and (b) show two different point cloud distributions. It is obvious that the human eye is more sensitive to area A than area B. However, after encapsulating the model, the curvature or projection distance of point p in area A is much smaller than that of point q in area B, as shown in Figure 2. Figure 2 (c) and (d) based on the K-neighbor search, it is found that the position differences between point p and its neighboring points are concentrated in the same direction, while the position differences between point q and its neighboring points are scattered in multiple directions. Based on the above analysis, a single-direction perception sharpness (PSSD) function based on human visual perception characteristics is proposed, which is defined as follows:

[0043]

[0044] Among them, p represents the current point, p i Represents the K neighborhood points of p, p(x,y,z) is the coordinate value of the current point, p i (x,y,z) is the coordinate value of the neighborhood point p, ||p(x,y,z)-p i (x,y,z)|| represents the Euclidean distance between point p and its neighboring points, p(o) represents the coordinate value of the direction with the maximum variance of x, y, and z, |p(o)-p i (o)| represents the coordinate difference between point p and its neighboring points in the direction of maximum x, y or z variance, p(o) and p i (o) can be calculated by the following formula:

[0045]

[0046] Among them, δ 2 [p(x)], δ2 [p(y)] and δ 2 [p(z)] represents the variance of the point cloud coordinates in the x, y, and z directions respectively. If δ 2 [p(x)]>δ 2 [p(y)]>δ 2 The value returned by [p(z)] and p(o) is the coordinate value in the x direction.

[0047] In order to analyze the correctness of formula (1), we take Figure 3 The two extreme cases shown in the figure are discussed. The distribution of the current point p and its K neighboring points is as follows: Figure 3 As shown in (a), the position difference between point p and its neighboring points is concentrated in the x direction, so p(o)=p(x),||p(x,y,z)-p i (x,y,z)||=|p(x)-p i (x)|, so It can take the maximum value 1; when the distribution of point p and its K neighboring points is as follows Figure 3 As shown in (b), the position differences between point p and its neighboring points are scattered in the x, y, and z directions. At this time, p(o) = p(x) = p(y) = p(z). Let p(o) = p(x), then Take the minimum value The above two extreme cases illustrate that the PSSD function proposed in the present invention can reflect the directionality of local position differences in 3D point clouds.

[0048] b. Local visibility

[0049] A change in the value of a pixel in a 2D image can be viewed as a changing signal superimposed on a uniform background. This signal must reach a certain intensity to be visible to the visual system. Based on the nonlinear relationship between the contrast sensitivity threshold and background brightness, Chai Yi et al. proposed the concept of image visibility (VI). This concept takes into account the characteristics of the visual system and uses the contrast of image pixels to measure the change in the signal within the image block. Its definition is as follows:

[0050]

[0051] Where I(x,y) represents the grayscale value of the pixel at position (x,y); M×N represents the size of the image I(x,y); m k is the average intensity value of I(x,y); γ is a visual constant, ranging from 0.6 to 0.7. The larger the value of VI, the higher the image visibility.

[0052] The 2D image visibility function cannot be directly used for 3D point cloud feature evaluation, and the normal vector angle concept of the 3D point cloud is introduced in the present application to establish a 3D point cloud visibility evaluation function based on the human visual system. The normal vector angle represents the average value of the included angle between the normal vector of the current point p and the normal vectors of its K-neighbor points, and the corresponding normal vector is calculated by principal component analysis. The normal vector angle n θ (p) can be obtained by the following formula:

[0053]

[0054] wherein n p is the normal vector of the current point p, is the normal vector of the K-neighbor points of p. The size of the included angle between the normal vector of the current point p and the normal vectors of its K-neighbor points can reflect the steepness of the local region of the point cloud, as shown in the following formula: Figure 4 The included angle θ a , θ a1 between the normal vector of the point p in the steep region and the normal vectors of its neighbor points is obviously larger than the included angle θ a2 , θ b between the normal vector of the point p in the flat region and the normal vectors of its neighbor points. b1 b2 .

[0055] With the normal vector angle as the input, the expression of the 3D point cloud local visibility function (Local visibility, LV) established by the present application is as follows:

[0056]

[0057] wherein p j is the current point and its K-neighbor points, n θ (p j ) represents the normal vector angle of the point p j , represents the average value of the normal vector angles of the current point p and its K-neighbor points.

[0058] The 3D point cloud hybrid feature evaluation model combining the geometric features and the human visual perception characteristics is adopted in step 2), and the PSSD function and the LV function proposed in the present application are used for visual saliency evaluation. When the geometric feature saliency evaluation is performed, the curvature and the projection distance are used as the evaluation indexes to express the details of the point cloud model, and the average distance is used to ensure the uniform distribution of the global points. The expression of the hybrid feature evaluation model is as follows:

[0059]

[0060] wherein, respectively represent the PSSD function and the LV function based on the human visual perception characteristics; ​ respectively represent the curvature based on the geometric feature evaluation, the average distance of the current point p to its K-neighbor points, the projection distance of the current point p to the plane fitted by its K-neighbor points; w1-w5 represent the weights of different feature evaluation functions. In order to reduce the scale difference between the feature values and ensure the sensitivity of each feature function, each feature value needs to be normalized, and then the normalized feature peak value is filtered. For the convenience of description, F1(p)-F5(p) are simplified as F1(p)-F5(p), and formula (6) can be simplified as: The weights w1-w5 of different feature evaluation functions are dynamically optimized, and the specific formula is as follows:

[0061]

[0062] The weights w1-w4 decrease monotonically with x(n), and x(n) increases monotonically with the feature value F n (p), so the greater the feature value F n (p) is, the smaller the corresponding weight w n is. The value range of x(n) is [0, +∞), and the value range of w1-w4 is (0, 0.25]. Formula (7) realizes the dynamic optimization of the weight value in the hybrid feature evaluation model through the continuous updating of F1(p)-F5(p) values of different points.

[0063] According to the above hybrid feature evaluation model, the hybrid feature value of each point in the point cloud can be calculated, and the point cloud can be divided into I-level feature points, II-level feature points and III-level feature points according to the size of the hybrid feature value. The number of points of each level can be set in proportion according to the reduction rate, for example, the I-level feature points account for 10% of the total number of point clouds, the II-level feature points account for 60% of the total number of point clouds, and the III-level feature points account for 30% of the total number of point clouds. The I-level feature points correspond to larger hybrid feature values, and these points are crucial to the expression of the point cloud contour, and the III-level feature points correspond to smaller hybrid feature values, and there is a lot of redundancy in expressing the geometric distribution of the point cloud. Different downsampling rules are used to simplify each level of point cloud: the I-level features are all retained, the II-level feature points are simplified by hierarchical random sampling method, and the III-level feature points are simplified by cuboid grid method.

[0064] The algorithm of the application is described in detail above. In order to verify the effectiveness of the 3D point cloud simplification algorithm proposed in the application, experiments are carried out from three aspects of effectiveness verification of the hybrid feature evaluation model, qualitative evaluation of algorithm performance and quantitative evaluation. The specific experimental results and data analysis are as follows.

[0065] Effectiveness verification of the hybrid feature evaluation model

[0066] ​In order to verify the effectiveness of the mixed feature evaluation model, an ablation experiment is performed on different feature evaluation functions. The "bunny" point cloud is selected as the test object, and the original point cloud distribution and its encapsulation effect are as shown in Figure 5 (a1), 5(a2). First, based on the geometric feature evaluation function, the point cloud is simplified, w1 and w2 in the mixed feature evaluation model of formula (6) are set to zero, the geometric feature index (curvature, average distance, projection distance) is used for feature evaluation, and the point cloud is classified and simplified according to the feature value. The distribution of the simplified 3D point cloud, the encapsulation effect and the deviation distribution are as shown in Figure 5 (b1)-5(b3). Second, w2 in formula (6) is set to zero, the geometric feature and the PSSD function are used for feature evaluation of the point cloud, and the simplified result is as shown in Figure 5 (c1)-5(c3) under the condition that the simplification rate is unchanged. Third, the mixed feature evaluation model of formula (6) is used to simplify "bunny", and the simplified result is as shown in Figure 5 (d1)-5(d3). By comparing the local enlarged area of the simplified point cloud, Figure 5 (b2), the ears of "bunny" have obvious holes, and the corresponding Figure 5 (b3) has a larger deviation in this area; compared with Figure 5 (b2), Figure 5 (c2), the holes in this area are smaller, but there is still a large deviation in Figure 5 (c3); compared with Figure 5 (c2), Figure 6 (d2), the details of the enlarged area of the ears of "bunny" are preserved more completely, and the deviation in this area is significantly reduced in 5(d3). It can be seen that the mixed feature evaluation model can better preserve the detail information of the original point cloud while maintaining the integrity of the data.

[0067] Comparison of visual effects of different simplification algorithms

[0068] In order to qualitatively evaluate the performance of the 3D point cloud simplification algorithm, the detail feature point simplification algorithm (DFPSA), the bounding box method, the normal vector method, the curvature method and the algorithm are used to simplify the point cloud. The "elephant" point cloud is used as the test object, and the distribution of the original point cloud and its encapsulation effect are as shown in Figure 6 (a1), 6(a2). The control simplification rate is 50%, and the distribution of the simplified point cloud obtained by different algorithms is as shown in Figure 6 (b1)-6(f1), the encapsulation effect is as shown in Figure 6 (b2)-6(f2), and the deviation distribution is as shown inFigure 6 (b3)-6(f3) as shown. Comparison Figure 6 (a1)-6(f1) found that five kinds of algorithm can realize the effective simplification of point cloud, wherein the bounding box method focuses on describing the global uniformity of point cloud, and the remaining algorithms focus on retaining the detailed features of point cloud. Comparison Figure 7 (b2)-6(f2), 6(b3)-6(f3), from the overall encapsulation effect of the point cloud model, it can be seen that the tail and the rear paw of the "elephant" in (b2) (dashed box), the rear paw of the "elephant" in (d2) and (e2) (dashed box) all have different degrees of voids; although there is no void in (c2), compared with (a2), there is a loss of detail information, which shows that the simplification algorithm proposed in the application can achieve better encapsulation effect. From the local enlarged area (solid line box) of the encapsulation diagram, it can be seen that the nose of the "elephant" in (b2)-(e2) all has obvious voids, and the deviations in the enlarged area of (b3)-(e3) are all larger; while the information of the nose of the "elephant" in (f2) is retained more completely, and the deviation in this area of (f3) is the smallest, which further shows that the performance of the simplification algorithm proposed in the application is better.

[0069] In order to ensure the repeatability of the experimental results, another "gargoyle" point cloud is taken as the test object for experiment, and the distribution of the original point cloud and its encapsulation effect are as shown in Figure 7 (a1), 7(a2). The control simplification rate is 50%, and the simplification results obtained by using different algorithms are as shown in Figure 7 (b1)-7(f1), Figure 7 (b2)-7(f2), Figure 7 (b3)-7(f3). From the overall encapsulation effect of the simplified point cloud, Figure 7 (b2), Figure 7 (d2), Figure 7 (e2) exist a large number of voids, Figure 7 (c2) has relatively less voids, but the local detail loss exists in the area framed by the dashed line box, and Figure 8 (f2) retains the local details while maintaining the overall encapsulation voids; from the overall deviation distribution, Figure 8 (f3) has the smallest deviation, which shows that the simplification algorithm proposed in the application can achieve better simplification effect for different point clouds.

[0070] Quantitative evaluation of the performance of different simplification algorithms

[0071] In order to objectively evaluate the performance of each simplification algorithm, the geometric error of the simplified point cloud is analyzed, and the geometric error evaluation indexes include maximum distance error, average distance error and relative volume error.

[0072] The expression of the maximum distance error is:

[0073]

[0074] wherein S is the original point cloud; S' is the simplified point cloud; d(p, S') represents the Euclidean distance from a point p in S to the nearest triangular facet on S' after the grid subdivision of S'.

[0075] The expression of the average distance error is:

[0076]

[0077] wherein N is the number of the original point cloud.

[0078] The expression of the relative volume error is:

[0079]

[0080] wherein V S is the volume of the original point cloud; V S′ is the volume of the simplified point cloud.

[0081] Table 1 compares the geometric error and simplification time of different point cloud models under various simplification algorithms (MATLAB, AMD Ryzen 75800H, 3.2GHz, 8 cores, 16GB memory), while ensuring approximately equal simplification rates. In the table, italic bold font indicates the minimum value of the same evaluation metric under different simplification algorithms, while black bold font indicates the next minimum value. The distribution of the minimum values ​​reflects the performance of the different simplification algorithms. The maximum distance error, average distance error, and relative volume error of the proposed algorithm are significantly lower than those of other algorithms. This is because the proposed simplification algorithm uses average distance as a constraint function to preserve the overall characteristics of the point cloud, avoiding the generation of voids during point cloud encapsulation. When preserving local details, it uses geometric features combined with human visual perception characteristics as an evaluation function, increasing the sensitivity of the local saliency evaluation function. Ultimately, it achieves a good compromise between preserving overall uniformity and highlighting local detail features. In terms of simplification time, the bounding box algorithm uses only simple Euclidean distance as a criterion to remove redundant points, and its time complexity is lower than that of the proposed algorithm. Although the feature evaluation function of the proposed algorithm is more complex than that of the curvature method and the normal vector method, the algorithm complexity is reduced during the subsequent step-by-step data simplification, resulting in a shorter overall algorithm runtime. The DFPSA algorithm (Ji Chunyang, Li Ying, Fan Jiahao, Lan Shumei. A Novel Simplification Method for 3D Geometric Point Cloud Based on the Importance of Point[J]. IEEE Access, 2019, 7:129029–129042.) takes the longest time to simplify weak feature points. This is because the algorithm uses an octree algorithm to simplify weak feature points, and the construction of the octree significantly increases the time complexity.

[0082] Table 1 Performance comparison of various simplification algorithms for different point cloud models

[0083]

[0084] In order to further evaluate the advantages of the algorithm of the present invention at different reduction rates, the “gargoyle” point cloud is taken as an example to draw a curve showing the geometric error and reduction time as a function of the reduction rate. Figure 8 Analysis ​ From the overall distribution of curves (a)-8(c), it can be seen that as the reduction rate increases, the geometric errors of each algorithm gradually increase, but the geometric errors of the algorithm proposed in this invention are the smallest at the same reduction rate. In addition, the solid line representing the error of the algorithm of this invention grows the most slowly, indicating that the performance of the algorithm proposed in this invention is the most stable as the reduction rate increases. ​(d)It can be seen that the compression time of each compression algorithm remains basically unchanged at different compression rates, and the algorithm proposed in the application ranks second in time complexity. It is shown that the algorithm of the application effectively improves the accuracy of 3D point cloud compression while not significantly increasing the time complexity of the algorithm.

[0085] The above has described the embodiments of the application in detail, but the application is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the application.

Claims

1. A 3D point cloud simplification method based on human visual perception characteristics, characterized by: Follow these steps: Step 1) Perform K-neighborhood search on the point cloud; Step 2) Calculate the one-way perceptual sharpness, local visibility, curvature, average distance, and projection distance of each point, and use the weighted dynamic optimization formula to calculate the weight value of each feature. The mixed feature value of each point is obtained by weighted averaging different feature values ​​and corresponding weights. Among them, the one-way perceptual sharpness function is defined as follows: （1） Among them, p represents the current point, represents the K neighborhood points of p, is the coordinate value of the current point, is the coordinate value of the p neighborhood point, represents the Euclidean distance between point p and its neighboring points, Represents the coordinate value of the direction of maximum variance of x, y, and z, Indicates the coordinate difference between point p and its neighboring points in the direction of maximum x, y or z variance, and It can be calculated by the following formula: （2） in, 、 and Respectively represent the variance of the point cloud coordinate values ​​in the x, y, and z directions. If , The returned value is the coordinate value in the x direction; The function expression of local visibility is as follows: （5） in, is the current point and its K neighboring points, Representative Points The normal vector angle, Represents the average value of the normal vector angle of the current point p and its K neighboring points. It can be obtained by the following formula: （4） in, is the normal vector of the current point p, is the normal vector of the K neighborhood of point p. The angle between the normal vector of point p and the normal vectors of its K neighborhood points can reflect the degree of steepness of the local area of ​​the point cloud. The expression of the mixed feature evaluation model for each point is as follows: （6） in, 、 They represent the one-way perceptual sharpness PSSD function and local visibility LV function based on human visual perception characteristics; 、 、 They represent the curvature evaluated based on geometric features, the average distance from the current point p to its K neighboring points, and the projection distance from the current point p to the fitting plane of its K neighboring points respectively; Represents the weights of different feature evaluation functions; In order to reduce the scale difference between eigenvalues ​​and ensure the sensitivity of each eigenfunction, each eigenvalue needs to be normalized and then the normalized eigenvalue peak value needs to be filtered. 、 、 、 、 Simplified to , formula (6) can be simplified to: , the weights of different feature evaluation functions Perform dynamic optimization. The specific formula is as follows: （7） Weight Follow Monotonically decreasing, With the eigenvalue Monotonically increasing, so the eigenvalue The larger the value, the corresponding weight The smaller, The value range is , The value range is ; Step 3) Classify the point cloud according to the mixed eigenvalues ​​and set the level-by-level simplification rules to achieve downsampling of each level of point cloud; Step 4) The downsampled data at all levels are fused to obtain a streamlined point cloud.

2. The 3D point cloud simplification method according to claim 1, characterized in that: The step-by-step simplification rules in step 3) are: First, the mixed feature evaluation model is used to calculate the mixed feature value of each point in the point cloud, and the point cloud is divided into Level feature points, Level feature points and Level feature points; Then, different downsampling rules are used to simplify the point clouds at each level: All level features are retained. Level I feature points are simplified by using the hierarchical random sampling method, and level III feature points are simplified by using the rectangular grid method.

Images