Three-dimensional dynamic point cloud denoising method and device based on gradient field

Abstract

The invention relates to a gradient field-based three-dimensional dynamic point cloud denoising method and device. The method comprises the following steps: estimating an initial gradient field for each frame in a dynamic point cloud sequence; based on the estimated initial gradient field, performing time correspondence search on surface blocks of the point cloud through rigid motion to obtain corresponding surface blocks in adjacent frames; constructing a time gradient field by averaging the inverse transform gradient fields of the corresponding surface blocks in the adjacent frames; and based on the constructed time gradient field, denoising the dynamic point cloud through gradient rising. The invention provides a dynamic point cloud denoising normal form based on a gradient field, and denoising is carried out by utilizing a time sequence corresponding relation between adjacent frames and rigid body motion in classical mechanics and by utilizing distribution of noise point clouds; the model provided by the invention is superior to the existing most advanced method under the conditions of synthesizing noise and simulating actual noise.

Description

Technical Field

[0001] The present invention relates to the field of point cloud denoising, and in particular to a three-dimensional dynamic point cloud denoising method and device based on a gradient field. Background Art

[0002] The maturity of laser scanning technology makes it possible to easily acquire 3D dynamic point clouds, which are discrete representations of objects or scenes in motion, with various applications such as autonomous driving, robotics, and immersive telepresence. A dynamic point cloud consists of a series of static point clouds, each of which consists of a set of points irregularly sampled from the continuous surface of an object or scene. However, 3D dynamic point clouds are often affected by noise due to inherent limitations of the acquisition equipment or computational errors in the sampling algorithm, which affects downstream tasks such as recognition and analysis. Therefore, denoising 3D dynamic point clouds is crucial for related 3D applications.

[0003] Recently, many methods have been proposed for denoising static point clouds, which can be divided into two categories: optimization-based methods and deep learning-based methods. Optimization-based methods usually use optimization algorithms to remove noise from point clouds. These methods rely heavily on geometric prior knowledge and sometimes have difficulty in striking a balance between denoising effect and detail preservation. Deep learning-based methods have achieved good denoising performance due to the emergence of neural network structures designed for point clouds. Most deep learning-based methods predict the displacement of each noise point relative to the underlying surface and then apply the inverse displacement to the noisy point cloud, which can lead to shrinkage and outlier problems due to over- or under-estimation of the displacement. In contrast, Luo et al. proposed a method that learns the underlying manifold of the noisy point cloud in a downsampling-upsampling architecture for reconstruction, but this method may lose details in the downsampling stage when the noise level is low. Recently, Luo et al. proposed a score-based point cloud denoising method, in which the position of each point is iteratively updated by gradient ascent from the log-likelihood distribution of the noisy point cloud.

[0004] Although a large number of methods have been proposed for static point cloud denoising, few algorithms are designed for dynamic point cloud denoising. This problem is also challenging because each point cloud frame is irregularly sampled and different frames may exhibit varying sampling patterns with possibly different numbers of points, which means there is no clear spatiotemporal correspondence between points at different time points. Summary of the invention

[0005] In order to solve the above problems and defects of the methods, the present invention proposes a dynamic point cloud denoising paradigm based on gradient field, which denoises the distribution of noise point clouds by utilizing the temporal correspondence between adjacent frames and the rigid body motion in classical mechanics.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A three-dimensional dynamic point cloud denoising method based on gradient field comprises the following steps:

[0008] Estimate the initial gradient field for each frame in the dynamic point cloud sequence;

[0009] Based on the estimated initial gradient field, the surface patches of the point cloud are searched for temporal correspondence through rigid motion to obtain the corresponding surface patches in adjacent frames.

[0010] The temporal gradient field is constructed by averaging the inverse transformed gradient fields of corresponding surface patches in adjacent frames;

[0011] Based on the constructed temporal gradient field, the dynamic point cloud is denoised by gradient ascent.

[0012] Furthermore, the initial gradient field is estimated using a gradient estimation network, which includes a feature extraction unit and a gradient estimation unit; first, the feature extraction unit is used to extract point features of each frame, and then the gradient estimation unit is used to estimate the initial gradient field of each frame from the point features.

[0013] Furthermore, the searching for temporal correspondence of surface blocks of the point cloud by rigid motion includes:

[0014] Given a surface block in a frame, its corresponding surface block in the previous frame is searched, and the surface block is moved to its corresponding position by performing surface block gradient ascent. The surface block gradient ascent process is regarded as a series of rigid body motions under the action of the gradient field, including surface block rotation and surface block translation.

[0015] Furthermore, the surface block rotates relative to the center of the surface block. The rotation angle is:

[0016]

[0017] The superscript (h) indicates the hth iteration, γ (h) is the rotation coefficient, which is used to control the amplitude of the rotation. represents the set of points in the surface block after the hth iteration, x (h) represents the coordinates of the midpoint of the surface block after the hth iteration, represents the average coordinates of all points in the h-th iteration surface block, I is the inertia matrix, Represents the gradient field of the t-1th frame.

[0018] Furthermore, the calculation formula of the surface block translation is as follows:

[0019]

[0020] The superscript (h) indicates the hth iteration. represents the overall translation of the surface block, β (h) is a coefficient used to control the magnitude of the translation, and m represents the number of points in the surface block.

[0021] Further, constructing the temporal gradient field by averaging the inverse transformed gradient fields of corresponding surface blocks in adjacent frames comprises:

[0022] For a surface block P in the tth frame t , and its corresponding surface block in the adjacent frame is represented as and

[0023] Diguo rotation will and The gradient field Transformed to the coordinate system of the tth frame, the transformed gradient field is

[0024] Surface block P t The temporal gradient field is the corresponding surface patch {P t-1 ,P t ,P t+1} is the average value of the gradient on

[0025]

[0026] in, is the gradient field of the tth frame.

[0027] Furthermore, the denoising of the dynamic point cloud by gradient ascent is to denoise each surface block separately, including:

[0028] Initialize an empty collection When a surface block is denoised, add the denoised points to the collection middle;

[0029] Through the FPS algorithm Sampling is performed so that the number of sampling points is the same as the point cloud X t The number of points in is the same, and the denoised point cloud is obtained;

[0030] Each frame is denoised, and finally a denoised point cloud sequence is obtained.

[0031] A three-dimensional dynamic point cloud denoising device based on gradient field, comprising:

[0032] An initial gradient field estimation module, used to estimate the initial gradient field for each frame in a dynamic point cloud sequence;

[0033] A temporal correspondence search module is used to perform temporal correspondence search on surface blocks of the point cloud through rigid motion based on the estimated initial gradient field to obtain corresponding surface blocks in adjacent frames;

[0034] a temporal gradient field construction module for constructing a temporal gradient field by averaging inverse transformed gradient fields of corresponding surface patches in adjacent frames;

[0035] The dynamic point cloud denoising module is used to denoise the dynamic point cloud by gradient ascent based on the constructed temporal gradient field.

[0036] The beneficial effects of the present invention are as follows:

[0037] The present invention proposes a novel dynamic point cloud denoising method and apparatus based on gradient field, which exploits the frame-to-frame-to-time correspondence of dynamic point clouds by estimating the gradient field, which is a fundamental problem in the processing and analysis of dynamic point clouds. The gradient field is the gradient of the logarithmic probability function of the noisy point cloud, and the present invention uses this gradient to perform gradient ascent, so that each point converges to the underlying clean surface. The present invention estimates the gradient of each surface patch and exploits the time correspondence, using rigid body motion in classical mechanics to search for the corresponding surface patch. In particular, the present invention treats each surface patch as a rigid object, moving in the gradient field of adjacent frames until an equilibrium state is reached, that is, the sum of the gradients on the surface patch reaches 0. The smaller the gradient of a point, the closer the point is to the underlying surface, so the equilibrium patch will adapt well to the underlying surface, thereby finding the time correspondence of the point cloud frame-to-frame. Finally, the position of each point in each surface patch is updated along the gradient direction averaged from the corresponding patch in adjacent frames. Experimental results show that the proposed model outperforms the existing state-of-the-art methods under both synthetic noise and simulated real noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 The overall framework of the dynamic point cloud denoising method proposed in this invention, where the surface block shown is the same surface block from the tth frame. The method mainly includes four steps: 1) estimating the initial gradient field; 2) searching for the corresponding surface block in time through rigid motion in a classical mechanical way; 3) constructing the temporal gradient field by averaging the inverse transformed gradient fields of the corresponding surface blocks in adjacent frames; 4) based on the constructed temporal gradient field, denoising the dynamic point cloud by gradient ascent.

[0039] Figure 2 Visual comparison of the proposed method and competitive denoising methods under Gaussian noise.

[0040] Figure 3 A visual comparison of the proposed method and competing denoising methods under simulated LiDAR noise.

[0041] Figure 4 A visualization of the present invention's search process for finding temporal correspondences in adjacent frames. DETAILED DESCRIPTION

[0042] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below through specific embodiments and drawings.

[0043] This paper proposes a dynamic point cloud denoising method based on gradient field, which uses the temporal correspondence between adjacent frames and the rigid body motion in classical mechanics to denoise the distribution of the noise point cloud. The dynamic point cloud consists of discrete points x sampled from the surface of a 3D dynamic object or scene, so it can be modeled as a dynamic 3D distribution p supported by some dynamic 2D manifold. t (x) is a set of samples. When the dynamic point cloud is disturbed by noise, its distribution can be modeled as the convolution of the original distribution and the noise, that is, (p t *n)(x), where n is the noise distribution. Under some mild assumptions about the noise n, (p t The modulus of *n)(x) is a potential clean surface with higher probability than the surrounding space. Therefore, recovering a noisy point cloud is naturally equivalent to maximizing (p t *n)(x). However, it is usually difficult to obtain (p t *n)(x) distribution information. Therefore, the present invention estimates the gradient of the logarithmic probability function to get the surface direction, and then move the perturbed point cloud towards the potential clean surface by gradient ascent.

[0044] The key to dynamic point cloud denoising is to search for temporal consistency between adjacent frames. Inspired by the motion of a free fall from high energy to low potential energy in classical mechanics, the present invention designs a process in which a surface block of a point cloud will spontaneously move to the base surface of its adjacent frame to obtain temporal consistency. Inspired by the fact that the negative gradient of the potential energy of a rigid body in classical mechanics represents the gradient direction force exerted on the rigid body, the present invention simulates the motion of the object in the gradient field to search for temporal correspondence. If the present invention converts the negative logarithmic density -log(p t *n) is defined as the potential energy of the noise point cloud, then the negative gradient of the potential energy It can be regarded as a force that pushes the noise point toward the base surface. Therefore, the present invention allows the surface block to move in the gradient field of adjacent frames to adapt to the local structure in the consecutive frames until a balanced state is reached, that is, when the sum of the gradients on the block reaches 0. Since the point is on the base surface when the gradient is 0, the balanced block will adapt well to the base surface, resulting in temporal consistency. The gradient of the current surface block is estimated as the average of the gradients of the temporally corresponding blocks in the adjacent frames, and the position of each noise point converges to the base surface in the gradient direction.

[0045] Specifically, the present invention first divides each noisy point cloud in the noisy dynamic point cloud sequence into surface patches (i.e., surface blocks), and designs a network to estimate the initial gradient field from the noisy patches in each frame, which is also used for subsequent temporal correspondence searches. Next, the present invention treats each noisy patch as a rigid body and moves them into the gradient field of the adjacent frame to search for corresponding patches in consecutive frames, which essentially performs rotations and translations on the patches. Then, the present invention averages the gradients of the inverse transformed corresponding patches in the adjacent frames as the temporal gradient. Finally, based on the estimated temporal gradient, the dynamic point cloud is denoised by gradient ascent.

[0046] Given a noisy dynamic point cloud sequence containing T frame point clouds in Represents the point cloud of the tth frame. The goal of this invention is to restore the noise-free dynamic point cloud sequence

[0047] like Figure 1 As shown, the method proposed in this patent mainly includes the following four steps (or four modules):

[0048] 1. Initial gradient field estimation. The present invention first uses a feature extraction unit to extract point features for each frame. Then, the present invention estimates the initial gradient field for each frame from the point features, which will be used in the subsequent time correspondence search. This module is trained for each noisy point cloud frame.

[0049] 2. Time correspondence search. The present invention searches for time corresponding patches in adjacent frames by modeling the transformation of patches in the gradient field as rigid body motion in classical mechanics until an equilibrium state is reached, that is, the total gradient is close to zero.

[0050] 3. Temporal gradient field construction: The present invention constructs a temporal gradient field from the searched temporal corresponding patches by inverse transformation.

[0051] 4. Dynamic point cloud denoising: The present invention performs gradient ascent based on the time gradient field to perform dynamic point cloud denoising.

[0052] The present invention will introduce these four steps in detail below.

[0053] 1) Initial gradient field estimation

[0054] The present invention first estimates an initial gradient field for each frame in a dynamic point cloud sequence. The present invention trains a gradient estimation network, which is composed of a feature extraction unit and a gradient estimation unit.

[0055] 1.1) Feature Extraction Unit

[0056] First, for a noisy point cloud where x i represents a point in the point cloud, N represents the number of points, and the present invention learns that for each point x i Contextual features of the surrounding local neighborhood. Specifically, the present invention uses DGCNN (point cloud dynamic graph convolutional network) to construct a series of densely connected edge convolutional layers for contextual feature extraction. DGCNN is a graph convolutional network commonly used in existing denoising models. Each densely connected edge convolutional layer takes the output features of the previous layer as input (for the first layer, the input is a noisy point cloud Q). It then dynamically builds a k-nearest neighbor graph on the input features, where each feature item is regarded as a vertex and is connected to its k nearest neighbors. The densely connected edge convolutional layer extracts rich contextual features through dense connections of local and non-local features. This feature makes the extracted features suitable for point cloud denoising tasks. Specifically, the x learned from the feature extraction unit i The characteristics are:

[0057]

[0058] in Represents the feature extraction unit.

[0059] 1.2) Gradient Estimation Unit

[0060] This unit is a multi-layer perceptron (MLP) that takes a i Nearby 3D coordinates is the input and outputs the gradient

[0061]

[0062] where || is a concatenation operator, Represents a gradient estimation unit. Supervised learning is performed with the true gradient g(x) defined as:

[0063] g(x)=NN(x, Y)-x, (3)

[0064] in is a true noise-free point cloud, NN(x, Y) returns the point closest to x in Y.

[0065] The optimization objective of gradient field estimation is defined as:

[0066]

[0067] in is a center point x i Nearby The distribution of space, is the mathematical expectation. In practice, the present invention generates a i A collection of several coordinates sampled nearby And calculate g(x) and The average difference of:

[0068]

[0069] The overall training loss is formulated as the average loss per point:

[0070]

[0071] Each point The initial gradient field It can be calculated by the following formula:

[0072]

[0073] Where kNN(x) is the k nearest neighbors of x.

[0074] 2) Time Correspondence Search

[0075] In classical mechanics, the negative gradient of the potential energy of a rigid body represents the force exerted on the rigid body along the gradient direction, which inspires the present invention to simulate the motion of an object in a gradient field to search for temporal correspondence. Specifically, the present invention allows surface blocks to move in the gradient fields of adjacent frames to match the local structure in consecutive frames, wherein the moved surface blocks are referred to as "corresponding surface blocks". In the following discussion, the present invention takes the temporal correspondence search between the current t-th frame and the previous frame as an example.

[0076] Based on the gradient field described in the previous section The present invention expresses the gradient field of the t-th frame as t = 1, ..., T. In order to consider the current noisy point cloud X t The present invention first searches for the time correspondence of X t Divide into M overlapping surface blocks, the number of points in each surface block is m, and the center of the surface block is represented by X t A subset of The points in are selected by the farthest point sampling (FPS) algorithm. Next, for each center point c i , the present invention constructs a surface block by determining its (m-1) nearest neighbors to obtain M surface blocks.

[0077] Given a noisy surface patch in frame t The goal of the present invention is to search for its corresponding surface block in the previous frame. The present invention performs surface block gradient ascent to move the surface block to its corresponding position. As mentioned above, the present invention regards the iterative surface block gradient ascent process as a series of rigid body motions under the action of the gradient field. Specifically, the rigid body motion of the moving surface block includes "rotation" and "translation", and then iteratively updates the surface block P t location.

[0078] 2.1) Surface Block Rotation

[0079] In each iteration, the rotation angle is calculated by the following formula:

[0080]

[0081] where the rotation is relative to the block center Calculation. The superscript (h) indicates the hth iteration, γ (h) is the rotation coefficient, which is used to control the amplitude of the rotation. represents the set of points in the surface block after the hth iteration, x (h) represents the coordinates of the midpoint of the surface block after the hth iteration, represents the average coordinates of all points in the h-th iteration surface block.

[0082] In practice, the present invention sets each point to unit mass. Therefore, the inertia matrix is:

[0083]

[0084] where x j ,y j and z j is relative to the block center The x, y, z coordinates of m are the number of points in the surface patch. Then, the rotation angle relative to the center of the patch is:

[0085]

[0086] Therefore, the position of the surface patch after rotation is:

[0087]

[0088] in Convert the angle axis to a rotation matrix. Note that the rotation is relative to the surface patch center. Therefore, the present invention adds the absolute coordinates of the center of the surface block in formula (11).

[0089] 2.2) Surface Block Translation

[0090] The translation of the surface patch is calculated as follows:

[0091]

[0092] The superscript (h) indicates the hth iteration. represents the overall translation of the surface block, β (h) is a coefficient used to control the magnitude of the translation, and m represents the number of points in the surface block.

[0093] With the calculated rotation and translation, the surface patch is transformed in each iteration as follows:

[0094]

[0095] The transformation process is iterated until the sum of the gradients on the surface patch is 0. Then, the transformed surface patch It will fall on the surface block corresponding to the t-1th frame.

[0096] Similarly, the present invention can also search for P in the (t+1)th frame in the same way. t The corresponding surface block.

[0097] 3) Construction of time gradient field

[0098] To construct the temporal gradient field, the present invention first converts the gradient field of the adjacent frames to the target frame, wherein the present invention considers each surface block separately. t , the present invention represents its corresponding surface block in the adjacent frame as and The present invention will and The gradient field is converted to the coordinate system of the tth frame. Specifically, For example, the converted gradient field is:

[0099]

[0100] in is corresponding to rotation. and Have different orientations but the same magnitude, so the inverse transform is just reflected in the rotation.

[0101] Considering the target frame and its previous and next frames, the surface block P t The temporal gradient field is the corresponding surface patch {P t-1 , P t , P t+1} is the average value of the gradient on

[0102]

[0103] 4) Dynamic point cloud denoising

[0104] Based on the estimated temporal gradient field, the present invention uses a gradient ascent algorithm to update the position of each point:

[0105]

[0106] where α (h) is the step size of the hth iteration, and H′ is the total number of iterations.

[0107] In particular, the present invention performs denoising on each surface block separately. Specifically, the present invention first initializes an empty set When the present invention completes the denoising of a surface block, the present invention adds the denoised points to the set Note that the surface patches overlap, so The final point set in X t Therefore, the present invention uses the FPS algorithm to Sampling is performed so that the number of sampling points is the same as X t The number of points in is the same, and the denoised point cloud is obtained The present invention performs denoising on each frame in this way, and finally obtains a denoised point cloud sequence

[0108] Figure 2 This is a visual comparison of the proposed method and the competitive denoising method under Gaussian noise. Points with lighter colors are farther from the true surface of the point cloud.

[0109] Figure 3 This is a visual comparison of our proposed method and competing denoising methods under simulated LiDAR noise. Lighter points are farther from the true surface of the point cloud.

[0110] Figure 4 It is a visualization of the search process of finding temporal correspondence in adjacent frames of the present invention. The surface block in the adjacent frame corresponding to the target surface block in the tth frame (marked in gray) moves under the action of the gradient field in the adjacent frame until an equilibrium state is reached.

[0111] Table 1 is a comparison of the noise reduction effects of the present invention and the competitive noise reduction algorithm under Gaussian noise. 4 , HD multiplied by 10 2 , P2M multiplied by 10 4 Both M2M and the present invention are designed for dynamic point cloud denoising.

[0112] Table 2 is a comparison of the noise reduction effects of the present invention and competitive noise reduction algorithms under simulated LiDAR noise. 4 , HD multiplied by 10 2 , P2M multiplied by 10 4 Both M2M and the present invention are designed for dynamic point cloud denoising.

[0113] Table 1: Comparison of denoising effects of the present invention and competitive denoising algorithms under Gaussian noise

[0114]

[0115] Table 2: Comparison of denoising effects of the present invention and competitive denoising algorithms under simulated LiDAR noise

[0116]

[0117] Another embodiment of the present invention provides a three-dimensional dynamic point cloud denoising device based on a gradient field, comprising:

[0118] An initial gradient field estimation module, used to estimate the initial gradient field for each frame in a dynamic point cloud sequence;

[0119] A temporal correspondence search module is used to perform temporal correspondence search on surface blocks of the point cloud through rigid motion based on the estimated initial gradient field to obtain corresponding surface blocks in adjacent frames;

[0120] a temporal gradient field construction module for constructing a temporal gradient field by averaging inverse transformed gradient fields of corresponding surface patches in adjacent frames;

[0121] The dynamic point cloud denoising module is used to denoise the dynamic point cloud by gradient ascent based on the constructed temporal gradient field.

[0122] The specific implementation process of each module refers to the above description of the method of the present invention.

[0123] Another embodiment of the present invention provides a computer device (computer, server, smart phone, etc.), which includes a memory and a processor, the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing each step in the method of the present invention.

[0124] Another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, magnetic disk, optical disk), wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a computer, the steps of the method of the present invention are implemented.

[0125] The specific embodiments of the present invention disclosed above are intended to help understand the content of the present invention and implement it accordingly. It can be understood by those skilled in the art that various replacements, changes and modifications are possible without departing from the spirit and scope of the present invention. The present invention should not be limited to the contents disclosed in the embodiments of this specification, and the scope of protection of the present invention shall be subject to the scope defined in the claims.

Claims

1. A three-dimensional dynamic point cloud denoising method based on gradient field, characterized in that: The following steps are involved: Estimate the initial gradient field for each frame in the dynamic point cloud sequence; Based on the estimated initial gradient field, the surface patches of the point cloud are searched for temporal correspondence through rigid motion to obtain the corresponding surface patches in adjacent frames. The temporal gradient field is constructed by averaging the inverse transformed gradient fields of corresponding surface patches in adjacent frames; Based on the constructed temporal gradient field, the dynamic point cloud is denoised by gradient ascent.

2. The method according to claim 1, characterized in that The initial gradient field is estimated using a gradient estimation network, which includes a feature extraction unit and a gradient estimation unit. The feature extraction unit is first used to extract point features of each frame, and then the gradient estimation unit is used to estimate the initial gradient field of each frame from the point features.

3. The method according to claim 1, characterized in that: The step of searching for temporal correspondence of surface blocks of the point cloud by rigid motion comprises: Given a surface block in a frame, its corresponding surface block in the previous frame is searched, and the surface block is moved to its corresponding position by performing surface block gradient ascent. The surface block gradient ascent process is regarded as a series of rigid body motions under the action of the gradient field, including surface block rotation and surface block translation.

4. The method according to claim 3, characterized in that The surface block rotates relative to the surface block center The rotation angle is: The superscript (h) indicates the hth iteration, γ (h) is the rotation coefficient, used to control the amplitude of the rotation, P t (h) represents the set of points in the surface block after the hth iteration, x (h) represents the coordinates of the midpoint of the surface block after the hth iteration, represents the average coordinates of all points in the h-th iteration surface block, I is the inertia matrix, Represents the gradient field of the t-1th frame.

5. The method according to claim 4, characterized in that The calculation formula of the surface block translation is as follows: The superscript (h) indicates the hth iteration. represents the overall translation of the surface block, β (h) is a coefficient used to control the magnitude of the translation, and m represents the number of points in the surface block.

6. The method according to claim 1, characterized in that The method of constructing a temporal gradient field by averaging the inverse transformed gradient fields of corresponding surface blocks in adjacent frames comprises: For a surface block P in the tth frame t , and its corresponding surface block in the adjacent frame is represented as and By rotating and The gradient field Transformed to the coordinate system of the tth frame, the transformed gradient field is Surface block P t The temporal gradient field is the corresponding surface patch {P t-1 ,P t ,P t+1 } is the average value of the gradient on in, is the gradient field of the tth frame.

7. The method according to claim 1, characterized in that The denoising of the dynamic point cloud by gradient ascent is to denoise each surface block separately, including: Initialize an empty collection When a surface block is denoised, add the denoised points to the collection middle; Through the FPS algorithm Sampling is performed so that the number of sampling points is the same as the point cloud X t The number of points in is the same, and the denoised point cloud is obtained; Each frame is denoised, and finally a denoised point cloud sequence is obtained.

8. A three-dimensional dynamic point cloud denoising device based on gradient field, characterized in that: include: An initial gradient field estimation module, used to estimate the initial gradient field for each frame in a dynamic point cloud sequence; A temporal correspondence search module is used to perform temporal correspondence search on surface blocks of the point cloud through rigid motion based on the estimated initial gradient field to obtain corresponding surface blocks in adjacent frames; a temporal gradient field construction module for constructing a temporal gradient field by averaging inverse transformed gradient fields of corresponding surface patches in adjacent frames; The dynamic point cloud denoising module is used to denoise the dynamic point cloud by gradient ascent based on the constructed temporal gradient field.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program comprises instructions for executing the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a computer, the method according to any one of claims 1 to 7 is implemented.

Images