Affine motion estimation method for dynamic point cloud

Abstract

The present invention provides an affine motion estimation method for a dynamic point cloud. The method comprises the following steps: (1) point cloud preorder motion analysis: reading a point cloud sequence to be processed, estimating an affine motion field of the current point cloud starting from a third frame, and performing, on a point cloud having undergone preorder decoding, inter-frame geometric motion analysis to obtain motion prior information; (2) point cloud subset division based on the motion prior information; and (3) affine motion estimation of point cloud subsets: establishing a geometric matching relationship between a reference point cloud subset and the current frame, obtaining an affine transformation matrix between a reference frame subset and the current frame on the basis of singular value decomposition, and iterating all the point cloud subsets to complete affine motion estimation from reference frames to the current frame. The method of the present invention further improves the compression performance of a dynamic point cloud.

Description

A method for affine motion estimation of dynamic point clouds Technical Field

[0001] The present invention relates to the field of point cloud data compression, and in particular to a method for affine motion estimation of dynamic point clouds. Background Art

[0002] Recent advances in point cloud capture and reconstruction technology have made significant progress, greatly stimulating the demand for 3D media applications such as immersive telepresence, virtual reality, and free-viewpoint television. Point cloud is a popular data format used to discretize static and dynamic people and objects in three-dimensional space into a set of sparse and unstructured points with position, color, and other information such as reflectivity and transparency. However, with the increasing demand for high-resolution and high-bit-depth point clouds, the large amount of data in the raw point cloud poses a huge challenge to applications that are limited by transmission bandwidth and storage capacity. Therefore, there is an urgent need for effective compression schemes to reduce the spatial and temporal redundancy inherent in point clouds. Motion estimation techniques play a key role in eliminating the temporal redundancy of point cloud sequences.

[0003] Research status of motion estimation: There are mainly the following existing point cloud motion estimation solutions:

[0004] 1. Block-matching-based translational motion estimation: This method first spatially decomposes the point cloud using an octree to generate prediction blocks. It then sets a reference frame search range and search step for each prediction block. It then uses an iterative search method to optimize the rate-distortion function to obtain the minimum-cost local translational motion between the prediction block and the matching block. This method has limited search accuracy, and compression performance needs to be improved.

[0005] 2. Feature-based translational motion estimation: This method first uses an octree to spatially decompose the point cloud, dividing it into specific layers to obtain prediction blocks. Within each non-empty prediction block, a graph is constructed (the nodes of the graph are all the points in the block, and each node is connected to its 26 nearest neighbors in the current block to form the graph's edges). For each node in the graph, the graph spectral features are estimated, and node-to-node matching is completed based on feature matching. The translational motion field between the two frames is then estimated through iterative optimization. This method has higher motion accuracy than the first method, but is computationally more complex.

[0006] 3. ICP-based rigid motion estimation: This method first spatially decomposes the point cloud using an octree to generate macroblocks. The texture variance within each macroblock is calculated. For macroblocks with texture variances below a threshold, a reference block with the same spatial position is selected from the reference frame. ICP is used to establish a matching relationship between the reference block and the current block, and rigid motion is estimated. However, this method's motion search range is too small, resulting in limited search accuracy.

[0007] The difficulty of motion estimation research is as follows: Point cloud motion estimation is a challenging technical task, mainly because the characteristics and complexity of point cloud data lead to the following technical difficulties. (1) Data sparsity and irregularity: Point cloud data usually consists of a large number of discrete points, which may be unevenly distributed and contain a large number of missing values ​​or noise. This makes it more difficult to perform accurate motion estimation on point clouds. (2) Complexity of motion: The motion in point cloud data may involve multiple types, including translation, rotation, scaling, deformation, etc., which increases the complexity of motion estimation. In particular, when the objects in the point cloud have non-rigid motion or complex shape changes, it becomes more difficult to accurately estimate their motion. (3) Establishment of motion model: In point cloud motion estimation, it is necessary to establish an appropriate motion model to describe the motion of point cloud data. However, due to the complexity and uncertainty of point cloud data, designing a suitable motion model is a challenging task. (4) Computational complexity: Point cloud data usually has a large number of points and high-dimensional features. Therefore, motion estimation on point clouds requires processing large-scale data and complex calculations, which increases the computational complexity and running time of the algorithm. (5) Data association and matching: In point cloud motion estimation, point cloud data association and matching across frames is required to determine the motion relationship between corresponding points. However, due to the irregularity and noise of point cloud data, accurate point cloud matching is a challenging problem.

[0008] Research significance of motion estimation: Point cloud motion estimation is of great significance to point cloud compression. Point cloud motion estimation refers to the modeling and prediction of the motion of point cloud data in a time series. In point cloud compression, point cloud data usually moves over time. For example, point cloud data collected by radar may contain moving targets or changes in the environment. By estimating the point cloud motion, the compression efficiency and reconstruction quality can be improved, thereby achieving a higher degree of compression rate and better visual effects. Specifically, point cloud motion estimation can help compression algorithms find spatial and temporal correlations in point cloud data, so as to better utilize these correlations for compression. For example, when an object in a point cloud maintains a stable motion pattern in time, the amount of information that needs to be transmitted or stored can be reduced by estimating and utilizing this motion pattern. In addition, point cloud motion estimation can also help better maintain the geometric and semantic information of point cloud data during the compression and decompression process, thereby improving the accuracy and realism of reconstruction.

[0009] Disclosure of the invention

[0010] In view of the shortcomings of the existing technology, the present invention provides a method for affine motion estimation of dynamic point clouds to further improve the compression performance of dynamic point clouds.

[0011] The technical solutions of the present invention are as follows:

[0012] The method for affine motion estimation of a dynamic point cloud of the present invention comprises the following steps: (1) point cloud pre-sequence motion analysis: reading in a point cloud sequence to be processed, estimating the affine motion field of the current point cloud starting from the third frame; performing inter-frame geometric motion analysis on the point cloud after pre-sequence decoding to obtain motion prior information; (2) point cloud subset division based on motion prior information: performing spatial motion consistency measurement on the point cloud based on motion prior information; performing geometric correlation measurement on the point cloud based on Euclidean distance; obtaining a joint similarity measurement based on the motion consistency measurement and the geometric correlation measurement; dividing the current point cloud into irregular point cloud subsets based on the joint similarity measurement; and (3) affine motion estimation of each point cloud subset: establishing a geometric matching relationship between a subset of the reference point cloud and the current frame; obtaining an affine transformation matrix between the reference frame subset and the current frame based on singular value decomposition; iterating all point cloud subsets to complete affine motion estimation from the reference frame to the current frame.

[0013] Optionally, in the above-mentioned affine motion estimation method of dynamic point cloud, the inter-frame geometric motion analysis of the point cloud after pre-order decoding includes: 1-1) Preparation before motion analysis: Let F t Represents the point cloud frame of the current affine motion to be estimated, t is the index of the current frame, let F t-1 and F t-2 Represents F t The first two frames that have been encoded, for F t-1 and F t-2 Decode and obtain the decoded point cloud as preparation for motion analysis; 1-2) Construct the matching relationship between the frames: In the frame before the decoded current point cloud F t-1 and the first two frames of the current point cloud F t-2 Between, the nearest neighbor search method based on geometric Euclidean distance is used to find the point cloud frame F t-1 All points in the reference frame F t-2 Find matching points in the image and build matching point pairs; 1-3) Inter-frame geometric motion analysis: Based on the geometric coordinates of the built matching point pairs, calculate the spatial displacement between the two frames as F t-1 and F t-2 The translation motion field between the two frames is used as the motion prior information between the two frames.

[0014] Optionally, in the above-mentioned method for affine motion estimation of dynamic point clouds, in step (2), a joint similarity metric is obtained based on the motion consistency metric and the geometric correlation metric, including: 2-1) let p i and p j Represents two points in the current frame, and performs geometric correlation measurement on these two points based on Euclidean distance to obtain the geometric distribution error geo_error(p i ,p j), the geometric correlation metric expression is

[0015] where ||·||2 is the L2 norm, geo_error(p i ,p j ) is larger, indicating that the geometric correlation between the two points is smaller; conversely, the geometric correlation between the two points is greater;

[0016] 2-2) Let mv(i) and mv(j) represent the motion information of two points in the current frame. Using the motion prior information obtained from the inter-frame geometric motion analysis in step (1), we can get the estimated values ​​of mv(i) and mv(j). i and p j The two points are used to measure the motion consistency and obtain the motion error motion_error(p i ,p j ), the geometric correlation metric expression is:

[0017] where motion_error(p i ,p j ) is larger, indicating that the motion consistency between the two points is smaller; conversely, the motion consistency between the two points is greater;

[0018] 2-3) Calculate the geometric distribution error geo_error(p i ,p j ) and motion error motion_error(p i ,p j )’s joint metric total_error(p i ,p j ), whose expression is:

[0019] Where a1 and a2 represent the geometric distribution error geo_error(p i ,p j ) and motion error motion_error(p i ,p j ) in the joint similarity measure total_error(p i ,p j ) in the weight coefficient, total_error(p i ,p j ) is larger, indicating that the similarity between the two points is smaller; conversely, the similarity between the two points is greater.

[0020] Optionally, in the above-mentioned method for affine motion estimation of dynamic point clouds, in step S2, based on the joint similarity measurement, the current point cloud is divided into irregular point cloud subsets, including: a) in the initial stage, the number of subsets k to be divided into the point cloud is set, the maximum number of iterations is set, and k points are sampled in the point cloud to be divided as the center points of the initial subsets; b) in the allocation stage, the joint similarity measurement is performed on the points to be allocated in the current point cloud and the center points of the k initial subsets according to formula (3), and the current point will be allocated to the point cloud subset with the maximum joint similarity. This step is iterated to complete the subset allocation of all points in the current point cloud; c) in the update stage, the center points of each subset are recalculated according to the subset allocation result of step b); d) steps b) and c) are iterated until the subset allocation result of each point in b) no longer changes, or the maximum number of iterations is reached, and the iteration is terminated.

[0021] Optionally, in the above-mentioned method for affine motion estimation of dynamic point clouds, in step (3), the affine motion estimation of each subset of the point cloud includes: 3-1) according to the result of the division of the irregular subsets of the point cloud in step (2), the estimation of the affine motion field from the reference frame to the current frame is converted into a series of estimations of the affine motion field from the irregular subsets of the reference frame to the current frame, and let p and p' represent the points of the reference frame and the current frame respectively. The affine motion process from p to p' can be described as

[0022] Where R is a 3×3 affine rotation matrix with 9 parameters, t is a 3×1 translation vector with 3 parameters, and points p and p′ both have 3×1 geometric coordinates; 3-2) In the initial stage, set the maximum number of iterations and the threshold for the change of the Frobenius norm of the affine rotation matrix; 3-3) In the affine motion estimation stage, for point p in the irregular subset of the reference frame, use the nearest neighbor search based on Euclidean distance to find the corresponding matching point p′ in the current frame, iterate this step to complete the construction of matching point pairs between all points in the irregular subset and the current frame, let n represent the number of points contained in the irregular subset, and let P represent all points in the irregular subset of the reference frame The 4×n matrix is ​​composed of, let P′ represent the matching point in the current frame The estimation process of the affine transformation matrix T from the reference frame point set P to the current frame point set P′ can be described as the following optimization problem: argmin||P′-TP|| Frobenius (5)

[0023] Where T is the 4×4 affine transformation matrix to be estimated, ||·|| Frobeniusis the Frobenius norm of the matrix, which is a norm used to measure the size of the matrix. The calculation method is to take the square root of the sum of the squares of all elements in the matrix, use singular value decomposition to solve the optimization problem of formula (5), and obtain the estimated affine transformation matrix T; 3-4) in the update phase, according to the result of the affine motion estimation in step 3-3), recalculate the coordinates of the reference frame point set after affine motion compensation as the new reference frame point set; 3-5) iterate steps 3-3) and 3-4) until the error of the affine transformation matrix calculated by formula (5) in step 3-3) is less than the threshold of the Frobenius norm set in step 3-2), or when the maximum number of iterations is reached, the iteration terminates; 3-6) iterate steps 3-3) to 3-5) until all irregular subsets in the reference frame complete the estimation of the affine motion field of the current frame.

[0024] According to the technical solution of the present invention, the beneficial effects produced are:

[0025] 1) A motion analysis method based on inter-frame geometric displacement is proposed, which can capture inter-frame motion and generate motion prior information.

[0026] 2) A similarity measure that combines geometric correlation and motion consistency analysis is proposed, which can quickly realize the similarity measurement between point pairs.

[0027] 3) A method for dividing point clouds into irregular subsets based on similarity measurement is proposed, and a point cloud affine motion estimation method based on irregular subset structure is designed.

[0028] BRIEF DESCRIPTION OF THE DRAWINGS

[0029] FIG1 is a flow chart of an algorithm for dynamic point cloud affine motion estimation according to the present invention;

[0030] Figures 2a and 2b are examples of irregular subset partitioning results of point clouds on the test set.

[0031] Best Mode for Carrying Out the Invention

[0032] In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the present invention is described in detail below through specific implementations in conjunction with the accompanying drawings.

[0033] To make the purpose, technical methods and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and specific examples. These examples are merely illustrative and not intended to limit the present invention.

[0034] As shown in FIG1 , the present invention provides an affine motion estimation method for a dynamic point cloud, comprising the following steps:

[0035] (1) Pre-order motion analysis of point cloud: read the point cloud sequence to be processed, and estimate the affine motion field of the current point cloud starting from the third frame; perform inter-frame geometric motion analysis on the point cloud after pre-order decoding to obtain motion prior information.

[0036] The specific process of performing inter-frame geometric motion analysis on the pre-decoded point cloud is as follows:

[0037] 1-1) Preparation before motion analysis: Let F t represents the point cloud frame whose affine motion is to be estimated, and t is the index of the current frame. Let F t-1 and F t-2 Represents F t The first two frames that have been encoded. t-1 and F t-2 Decode and obtain the decoded point cloud as preparation for motion analysis;

[0038] 1-2) Construction of inter-frame geometric matching relationship: In the frame F before the decoded current point cloud t-1 and the first two frames of the current point cloud F t-2 Between, the nearest neighbor search method based on geometric Euclidean distance is used to find the point cloud frame F t-1 All points in the reference frame F t-2 Find matching points in and build matching point pairs;

[0039] 1-3) Inter-frame geometric motion analysis: Based on the geometric coordinates of the constructed matching point pairs, the spatial displacement between the two frames is calculated as F t-1 and F t-2 The translation motion field between the two frames is used as the motion prior information between the two frames.

[0040] (2) Point cloud subset division based on motion prior information: perform spatial motion consistency measurement on the point cloud based on motion prior information; perform geometric correlation measurement on the point cloud based on Euclidean distance; obtain joint similarity measurement based on motion consistency measurement and geometric correlation measurement; and divide the current point cloud into irregular point cloud subsets based on the joint similarity measurement.

[0041] Specifically, based on the motion prior information in the first stage and the geometric distribution characteristics of the point cloud within the frame, a similarity measure that combines the geometric distribution and motion prior information is proposed. Starting from the third frame, the point cloud is divided into irregular subsets based on the similarity measure. The specific process of combining the similarity measure of geometric distribution and motion prior information (i.e., obtaining a joint similarity measure based on the motion consistency measure and the geometric correlation measure) is as follows:

[0042] 2-1) Let p i and p jRepresents two points in the current frame, and performs geometric correlation measurement on these two points based on Euclidean distance to obtain the geometric distribution error geo_error(p i ,p j ). The geometric correlation metric expression is

[0043] where ||·||2 is the L2 norm. geo_error(p i ,p j ) is larger, indicating that the geometric correlation between the two points is smaller; conversely, the geometric correlation between the two points is greater;

[0044] 2-2) Let mv(i) and mv(j) represent the motion information of two points in the current frame. Using the motion prior information obtained from the inter-frame geometric motion analysis in step (1), we can get the estimated values ​​of mv(i) and mv(j). i and p j The two points are used to measure the motion consistency and obtain the motion error motion_error(p i ,p j ). The geometric correlation metric expression is:

[0045] where motion_error(p i ,p j ) is larger, indicating that the motion consistency between the two points is smaller; conversely, the motion consistency between the two points is greater;

[0046] 2-3) Calculate the geometric distribution error geo_error(p i ,p j ) and motion error motion_error(p i ,p j )’s joint metric total_error(p i ,p j ), whose expression is:

[0047] Where a1 and a2 represent the geometric distribution error geo_error(p i ,p j ) and motion error motion_error(p i ,p j ) in the joint similarity measure total_error(p i ,p j ) in the weight coefficient. total_error(p i ,p j ) is larger, indicating that the similarity between the two points is smaller; conversely, the similarity between the two points is greater.

[0048] In addition, an example of the division of irregular subsets of point clouds based on similarity measurement on the test set is shown in Figure 2, where (a) in Figure 2 is the original point cloud and (b) is the result of the division of irregular subsets of point clouds.

[0049] The specific process of dividing irregular subsets of point clouds based on similarity measurement is as follows:

[0050] a) In the initial stage, the number of subsets k of the point cloud to be divided, the maximum number of iterations, and k points are sampled from the point cloud to be divided as the center points of the initial subset;

[0051] b) In the allocation phase, the joint similarity measurement is performed on the current point cloud point to be allocated and the center point of the k initial subsets according to formula (3). The current point will be allocated to the point cloud subset with the maximum joint similarity. This step is iterated to complete the subset allocation for all points in the current point cloud.

[0052] c) In the update phase, the center points of each subset are recalculated according to the subset allocation results of step b);

[0053] d) Iterate steps b) and c) until the subset assignment results for each point in b) no longer change, or the maximum number of iterations is reached. The iteration terminates. At this point, the point cloud irregular subset partitioning process based on similarity measurement is completed, and the current point cloud is partitioned into a series of irregular point cloud subsets.

[0054] (3) Affine motion estimation of each subset of the point cloud: Establish a geometric matching relationship between the subset of the reference point cloud and the current frame; obtain the affine transformation matrix between the subset of the reference frame and the current frame based on singular value decomposition; iterate all subsets of the point cloud to complete the radial motion estimation from the reference frame to the current frame.

[0055] Based on the point cloud subset division results of the second stage, the geometric matching relationship between the reference point cloud subset and the current point cloud is established, and an affine motion estimation scheme describing the motion from the reference frame to the current frame is designed.

[0056] Among them, the specific process of the affine motion estimation scheme for each subset (irregular subset) of the point cloud is as follows:

[0057] 3-1) Based on the result of the division of the point cloud into irregular subsets in step (2), the estimation of the affine motion field from the reference frame to the current frame is converted into a series of estimations of the affine motion field from the irregular subsets of the reference frame to the current frame. Let p and p′ represent the points of the reference frame and the current frame respectively. The affine motion process from p to p′ can be described as

[0058] Where R is a 3×3 affine rotation matrix with 9 parameters, t is a 3×1 translation vector with 3 parameters, and points p and p′ both have 3×1 geometric coordinates;

[0059] 3-2) In the initial stage, set the maximum number of iterations and the threshold for changing the Frobenius norm of the affine rotation matrix;

[0060] 3-3) In the affine motion estimation stage, for a point p in the irregular subset of the reference frame, a nearest neighbor search based on Euclidean distance is used to find the corresponding matching point p′ in the current frame. This step is iterated to complete the construction of matching point pairs between all points in the irregular subset and the current frame.

[0061] Let n represent the number of points contained in the irregular subset, and let P represent all points in the irregular subset of the reference frame. The 4×n matrix is ​​composed of, let P′ represent the matching point in the current frame The estimation process of the affine transformation matrix T from the reference frame point set P to the current frame point set P′ can be described as an optimization problem:

[0062] argmin||P′-TP|| Frobenius (5)

[0063] Where T is the 4×4 affine transformation matrix to be estimated, ||·|| Frobenius is the Frobenius norm of the matrix, which is a measure of the size of a matrix and is calculated by taking the square root of the sum of the squares of all elements in the matrix.

[0064] The optimization problem of formula (5) is solved using singular value decomposition to obtain the estimated affine transformation matrix T. The solution of this process is not limited to the singular value decomposition method;

[0065] 3-4) In the update phase, based on the result of the affine motion estimation in step 3-3), the coordinates of the reference frame point set after affine motion compensation are recalculated as the new reference frame point set;

[0066] 3-5) Iterate steps 3-3) and 3-4) until the affine transformation matrix error of the two estimates (calculated by formula (5) in step 3-3) is less than the threshold of the Frobenius norm set in step 3-2), or when the maximum number of iterations is reached, the iteration terminates, and the affine motion estimation from this irregular subset of the reference frame to the current frame is completed.

[0067] 3-6) Iterate steps 3-3) to 3-5) until all irregular subsets in the reference frame have completed the estimation of the affine motion field to the current frame.

[0068] At this point, the estimation of the affine motion field from the reference point cloud to the current point cloud is completed.

[0069] (4) Geometric compression of point clouds: Using the motion vector obtained in step (3), the best inter-frame reference is selected from the reference frame for the current frame, and a geometric inter-frame context model is constructed for point cloud geometric conditional entropy coding, making full use of inter-frame geometric redundancy to achieve efficient geometric compression of dynamic point clouds.

[0070] To verify the effectiveness of the proposed method for affine motion estimation of dynamic point clouds, tests were conducted using the MPEG 3DG standard dataset 8i point cloud sequence and the JPEG Pleno standard dataset MVUB point cloud sequence. The compression performance comparison results are shown in Table 1. Compared to the current anchor platform, the proposed method achieved a 6.28% bitrate reduction in lossless compression of dynamic point cloud geometry on a 32-frame test, demonstrating significant compression performance improvements.

[0071] Table 1 Comparison of compression performance of the method of the present invention and other mainstream methods under 32-frame test

[0072] The above description is the best embodiment based on the concept and working principle of the invention. The above embodiment should not be understood as limiting the scope of protection of the present claims. Other embodiments and combinations of implementations of the present invention are also within the scope of protection of the present invention. Industrial Applicability

[0073] The present invention provides a method for affine motion estimation from dynamic point clouds. Point clouds are a popular data format for discretizing static and dynamic people and objects in three-dimensional space into a sparse, unstructured set of points that contain position, color, and other information, such as reflectivity and transparency. With the increasing demand for high-resolution and high-bit-depth point clouds, this method is particularly useful in the field of point cloud capture and reconstruction technology, with potential applications in immersive telepresence, virtual reality, free-viewpoint television, and other 3D media.

Claims

1. A method for affine motion estimation of dynamic point clouds, characterized in that: The steps include: (1) Pre-order motion analysis of point cloud: read the point cloud sequence to be processed, estimate the affine motion field of the current point cloud starting from the third frame; perform inter-frame geometric motion analysis on the decoded point cloud to obtain motion prior information; (2) Point cloud subset division based on motion prior information: performing spatial motion consistency measurement on the point cloud based on the motion prior information; performing geometric correlation measurement on the point cloud based on Euclidean distance; Obtaining a joint similarity measure based on the motion consistency measure and the geometric correlation measure; and dividing the current point cloud into irregular point cloud subsets based on the joint similarity measure; as well as (3) Affine motion estimation of each subset of the point cloud: Establishing the geometric matching relationship between the subset of the reference point cloud and the current frame; Based on singular value decomposition, the affine transformation matrix between the reference frame subset and the current frame is obtained; Iterate all subsets of the point cloud to complete the affine motion estimation from the reference frame to the current frame.

2. The method for affine motion estimation of dynamic point clouds according to claim 1, characterized in that: The inter-frame geometric motion analysis of the pre-decoded point cloud includes: 1-1) Preparation before motion analysis: Let F t Represents the point cloud frame of the current affine motion to be estimated, t is the index of the current frame, let F t-1 and F t-2 Respectively represent F t The first two frames that have been encoded, for F t-1 and F t-2 Decode and obtain the decoded point cloud as preparation for motion analysis; 1-2) Construction of inter-frame geometric matching relationship: In the frame F before the decoded current point cloud t-1 and the first two frames F of the current point cloud t-2 Between, the nearest neighbor search method based on geometric Euclidean distance is used to find the point cloud frame F t-1 All points in the reference frame F t-2 Find matching points in and build matching point pairs; 1-3) Inter-frame geometric motion analysis: Based on the geometric coordinates of the constructed matching point pairs, the spatial displacement between the two frames is calculated as F t-1 and F t-2 The translation motion field between the two frames is used as the motion prior information between the two frames.

3. The method for affine motion estimation of dynamic point clouds according to claim 1, characterized in that: In step (2), the joint similarity measure is obtained based on the motion consistency measure and the geometric correlation measure, including: 2-1) Let p i and p j Represents two points in the current frame, and performs geometric correlation measurement on these two points based on Euclidean distance to obtain the geometric distribution error geo_error(p i ,p j ), the geometric correlation metric expression is where ||·||2 is the L2 norm, geo_error(p i ,p j ) is larger, indicating that the geometric correlation between the two points is smaller; conversely, The greater the geometric correlation between two points; 2-2) Let mv(i) and mv(j) represent the motion information of two points in the current frame. Using the motion prior information obtained from the inter-frame geometric motion analysis in step (1), we can get the estimated values ​​of mv(i) and mv(j). i and p j The two points are used to measure the motion consistency and obtain the motion error motion_error(p i ,p j ), the geometric correlation metric expression is: where motion_error(p i ,p j ) is larger, indicating that the motion consistency between the two points is smaller; conversely, the motion consistency between the two points is greater; 2-3) Calculate the geometric distribution error geo_error(p i ,p j ) and motion error motion_error(p i ,p j )'s joint metric total_error(p i ,p j ), whose expression is: Where a1 and a2 represent the geometric distribution error geo_error(p i ,p j ) and motion error motion_error(p i ,p j ) in the joint similarity measure total_error(p i ,p j ) in the weight coefficient, total_error(p i ,p j ) is larger, indicating that the similarity between the two points is smaller; conversely, the similarity between the two points is greater.

4. The method for affine motion estimation of dynamic point clouds according to claim 3, characterized in that: In step S2, based on the joint similarity measure, the current point cloud is divided into irregular point cloud subsets, including: a) In the initial stage, the number of subsets k of the point cloud to be divided, the maximum number of iterations, and k points are sampled from the point cloud to be divided as the center points of the initial subset; b) In the allocation phase, the joint similarity measurement is performed on the point to be allocated in the current point cloud and the center point of the k initial subsets according to the formula (3). The current point will be allocated to the point cloud subset with the maximum joint similarity. This step is iterated to complete the subset allocation for all points in the current point cloud. c) in the update phase, based on the subset allocation results of step b), the center points of each subset are recalculated; d) Iterate steps b) and c) until the subset allocation results of each point in b) no longer change, or the maximum number of iterations is reached, and the iteration terminates.

5. The method for affine motion estimation of dynamic point clouds according to claim 1, characterized in that: In step (3), the affine motion estimation of each subset of the point cloud includes: 3-1) According to the result of the division of the irregular subset of the point cloud in step (2), the estimation of the affine motion field from the reference frame to the current frame is converted into a series of estimations of the affine motion field from the irregular subset of the reference frame to the current frame. Let p and p′ represent the point of the reference frame and the point of the current frame respectively. The affine motion process from p to p′ can be described as Where R is a 3×3 affine rotation matrix with 9 parameters, t is a 3×1 translation vector with 3 parameters, and points p and p′ both have 3×1 geometric coordinates; 3-2) In the initial stage, set the maximum number of iterations and the threshold for changing the Frobenius norm of the affine rotation matrix; 3-3) In the affine motion estimation stage, for a point p in the irregular subset of the reference frame, a nearest neighbor search based on Euclidean distance is used to find the corresponding matching point p′ in the current frame. This step is iterated to complete the construction of matching point pairs between all points in the irregular subset and the current frame. Let n represent the number of points contained in the irregular subset, and let P represent all points in the irregular subset of the reference frame. The 4×n matrix is ​​composed of, let P′ represent the matching point in the current frame The estimation process of the affine transformation matrix T from the reference frame point set P to the current frame point set P′ is described as the following optimization problem: argmin||P ′ -TP|| Frobenius (5) Where T is the 4×4 affine transformation matrix to be estimated, ||·|| Frobenius is the Frobenius norm of the matrix, which is a norm used to measure the size of the matrix and is calculated by taking the square root of the sum of the squares of all elements in the matrix. Use singular value decomposition to solve the optimization problem of formula (5) and obtain the estimated affine transformation matrix T; 3-4) In the update phase, based on the result of the affine motion estimation in step 3-3), the coordinates of the reference frame point set after affine motion compensation are recalculated as the new reference frame point set; 3-5) Iterate steps 3-3) and 3-4) until the affine transformation matrix error calculated by formula (5) in step 3-3) is less than the threshold of the Frobenius norm set in step 3-2), or the iteration is terminated when the maximum number of iterations is reached; 3-6) Iterate steps 3-3) to 3-5) until all irregular subsets in the reference frame have completed the estimation of the affine motion field to the current frame.

Images