Human eye simulation three-dimensional point cloud data compression method

Abstract

The invention discloses a human eye-imitating three-dimensional point cloud data compression method, which comprises the following steps of: acquiring original point cloud data, and performing Gaussian filtering denoising processing on the original point cloud data; performing region-of-interest extraction on the denoised point cloud by using a human eye-imitated target-of-interest extraction algorithm; performing high-resolution retention on the extracted region of interest by using a human eye-imitated interested target compression algorithm, and performing non-uniform compression on a non-interested region; and finally, displaying a compression result by using a display program. The region of interest is obtained according to the human eye retina attention perception mechanism, then high resolution is performed on the region of interest, non-uniform redundancy compression is performed on the non-region of interest, the problem of low efficiency of fixed resolution when a current point cloud data processing algorithm processes the region of interest and the non-region of interest is solved, and the accuracy of point cloud data processing is improved. Therefore, the quality of the three-dimensional point cloud data is effectively improved, and the recognition precision and speed in the field of three-dimensional target recognition are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving data processing, and in particular relates to a human eye-simulating three-dimensional point cloud data compression method. Background Art

[0002] With the continuous progress of optics, computer science, electronic information and automatic control technology, intelligent upgrades are becoming increasingly popular in various fields. In particular, the rise of autonomous driving technology in recent years has put forward more stringent requirements for optoelectronic perception technology in complex environments. This trend not only promotes the development of sensor technology, but also promotes the innovation of data processing and intelligent algorithms to meet the urgent needs of high-dynamic and multi-dimensional scene perception.

[0003] However, existing 3D point cloud processing technologies still face many challenges in practical applications. In terms of efficiency, 3D point cloud data is usually huge, especially in complex scenes, containing hundreds of thousands or even millions of point clouds, which brings a huge computational burden to data processing and analysis. In terms of accuracy, due to the high computational complexity of point cloud data, how to quickly and accurately extract features of interest and identify them has become a difficult problem. In addition, the acquisition of point cloud data is often accompanied by noise and incompleteness, which puts higher requirements on the accuracy of target recognition.

[0004] Among the existing methods, although there are some technologies that can process 3D point cloud data to a certain extent, such as point cloud recognition for autonomous driving, the shortcomings of the existing technologies are mainly reflected in their inefficiency and interference in processing areas of interest and non-interest areas with fixed resolution. Especially in complex scenes, it is difficult to effectively distinguish and process different objects in the point cloud during 3D point cloud recognition, resulting in inaccurate recognition results. For example, PointNet++ mentions the adaptability problem of point sets sampled with different densities, which shows that when processing point clouds with different densities, the fixed resolution method is difficult to effectively capture all important feature details. This inadaptability will lead to inefficiency and decreased accuracy in the recognition process. In addition, in complex scenes, the number of point clouds usually reaches millions. These methods are inefficient in processing large-scale point cloud data and are difficult to meet real-time requirements. For example, in the process of 3D target recognition, VoxelNet first needs to convert each point cloud into voxels, and then extract features from a large amount of voxel data, which makes it difficult for VoxelNet to meet real-time requirements in complex scenes. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a three-dimensional point cloud data compression method that simulates the human eye. The human eye retina has a non-uniform structure, with many central cells and high resolution characteristics, and the surrounding cells gradually decrease. The data of non-interested targets can be compressed, so that clear imaging can be achieved in the central gaze area, and compression and early warning can be performed in the surrounding non-interested areas, achieving rapid imaging, which provides ideas for solving the above problems.

[0006] To achieve the above object, the present invention provides a human eye-simulating three-dimensional point cloud data compression method, comprising:

[0007] Obtain the original 3D point cloud data of autonomous driving, and perform Gaussian filtering and denoising on the original 3D point cloud data;

[0008] The region of interest is extracted from the denoised 3D point cloud data using an algorithm for extracting objects of interest that simulates the human eye.

[0009] The human eye-simulated object of interest compression algorithm is used to retain the high resolution of the extracted region of interest and perform non-uniform compression on the non-interest region.

[0010] Preferably, the process of performing Gaussian filtering and denoising on the original three-dimensional point cloud data includes:

[0011] Estimate the noise level of each point in the original point cloud data;

[0012] Perform a Gaussian fit on the local neighborhood of each point to determine its mean and variance;

[0013] Use Gaussian model to denoise each point;

[0014] Update the position of each point to reflect the denoising result.

[0015] Preferably, the human eye-simulating object of interest extraction algorithm utilizes an encoder-decoder structure to simulate the context information of the point cloud through a three-dimensional point cloud context attention change module encoder and a scene context attention change change decoder.

[0016] Preferably, the process of retaining the extracted region of interest at high resolution by using a human eye-simulating target of interest compression algorithm and performing non-uniform compression on the non-interest region includes:

[0017] Using the three-dimensional retinal variable resolution positive change model, the point cloud data is converted from the Cartesian coordinate system to the polar coordinate system, and the polar coordinates are changed logarithmically;

[0018] Using the 3D retinal variable resolution compression model, the uniform density distribution of point cloud data is achieved through voxel compression point cloud method;

[0019] The point cloud in log polar coordinates is converted back to the point cloud in Cartesian coordinates using the 3D retinal variable resolution inverse change model;

[0020] The point cloud is fused using the 3D retinal variable resolution point cloud fusion model.

[0021] Preferably, the point cloud data is converted from a Cartesian coordinate system to a polar coordinate system, and the formula for logarithmic polar coordinate change includes:

[0022]

[0023] θ=arctan2(y-y0,x-x0)

[0024]

[0025] Where r is the distance from the point to the origin, θ is the angle between the projection of the point on the XY plane and the X-axis, and φ is the angle between the point and the Z-axis;

[0026] The calculation formula for the point cloud in logarithmic polar coordinates is:

[0027] log r =log(r+ε)

[0028] Among them, ∈ is a constant.

[0029] Preferably, the process of achieving uniform density distribution of point cloud data by voxel compression point cloud method using a three-dimensional retinal variable resolution compression model includes:

[0030] Divide the point cloud space into voxel grids of the same size, each voxel is a cubic area, and the voxel size is set according to application requirements;

[0031] All points in each voxel are replaced with representative points, which are the centroid or center point of the points in the voxel;

[0032] The original point cloud is downsampled to fewer points by point replacement, maintaining the overall structure and shape of the point cloud. Preferably, the formula for converting the point cloud in logarithmic polar coordinates back to the point cloud in Cartesian coordinates is:

[0033] x=(e logr -ε)sin(φ)cos(θ)+x0

[0034] y=(e logr -ε)sin(φ)sin(φ)+y0

[0035] z=(e logr -ε)cos(φ)+z0

[0036] Among them, (x, y, z) are the coordinates in the Cartesian coordinate system, and (x0, y0, z0) is the center point of the area of ​​interest.

[0037] Preferably, the process of fusing the point cloud includes:

[0038] The point cloud data of all the targets of interest are superimposed in sequence to form a fused point cloud;

[0039] Establish a KD tree index to organize the spatial data of the fused point cloud to facilitate efficient spatial query;

[0040] Use KD tree to perform radius search to find the neighboring points of each point in the fused point cloud;

[0041] By comparing the indexes of neighboring points, duplicate points in the fused point cloud are removed.

[0042] The present invention also provides a computer storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the human eye-simulating three-dimensional point cloud data compression method is implemented.

[0043] Compared with the prior art, the present invention has the following advantages and technical effects:

[0044] The present invention provides a human eye-simulating three-dimensional point cloud data compression method. First, the original point cloud data is obtained and Gaussian filtering denoising is performed on the original point cloud data; secondly, the denoised point cloud is extracted with an interest target extraction algorithm simulating the human eye; finally, the extracted interest region is retained with high resolution using an interest target compression algorithm simulating the human eye, and non-uniform compression is performed on non-interest regions.

[0045] The present invention obtains the region of interest in the three-dimensional data according to the retinal attention perception mechanism of the human eye, and then performs high resolution on the region of interest of the three-dimensional point cloud target, and performs non-uniform data redundancy compression on the three-dimensional point cloud region of no interest, thereby reducing the amount of data involved in the region of no interest, solving the inefficiency of the current point cloud data processing algorithm in processing the region of interest and the region of no interest, reducing the amount of data involved in the region of no interest, thereby effectively improving the quality of the three-dimensional point cloud data, and improving the recognition accuracy and rate in the field of three-dimensional target recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0047] Figure 1 This is an overall flow chart of the human eye-simulating three-dimensional point cloud data compression method according to an embodiment of the present invention;

[0048] Figure 2 Flow chart of the human eye-simulating interesting target compression algorithm according to an embodiment of the present invention;

[0049] Figure 3 A schematic diagram of a forward transformation and an inverse transformation of a human eye-simulating resolution change according to an embodiment of the present invention;

[0050] Figure 4 This is a schematic diagram of the complete process results of using the human eye-simulating three-dimensional point compression method in the field of autonomous driving according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0052] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0053] Embodiment 1

[0054] like Figure 1 As shown, this embodiment provides a human eye-simulating three-dimensional point cloud data compression method, including:

[0055] S1, obtaining original 3D point cloud data, and performing Gaussian filtering and denoising on the original point cloud data;

[0056] Specifically, the original three-dimensional point cloud data can be obtained from a point cloud file (in formats such as pcd, ply, xyz, etc.) or directly from a lidar device.

[0057] Furthermore, the process of performing Gaussian filtering and denoising on the original three-dimensional point cloud data includes:

[0058] Estimate the noise level of each point in the original point cloud data;

[0059] Perform a Gaussian fit on the local neighborhood of each point to determine its mean and variance;

[0060] Use Gaussian model to denoise each point;

[0061] Update the position of each point to reflect the denoising result.

[0062] Specifically, a point cloud Gaussian filtering algorithm is used to remove noise from point cloud data while retaining the edge and feature information of the point cloud data.

[0063] The basic idea of ​​3D Gaussian denoising is to use the characteristics of Gaussian distribution to estimate the noise level of each point and perform denoising accordingly. In 3D space, the noise of each point can be regarded as a random variable with Gaussian distribution. The probability density function of Gaussian distribution is given by the following formula:

[0064]

[0065] Where μ is the mean of the Gaussian distribution, which indicates the true position of the point, and σ is the standard deviation, which indicates the intensity of the noise. In a three-dimensional point cloud, the noise of each point can be described by its Gaussian distribution on the three coordinate axes. For a point p = (x, y, z), its Gaussian model can be expressed as:

[0066]

[0067] The denoising process usually includes the following steps:

[0068] Noise Estimation: First, the noise level is estimated for each point in the point cloud, which is usually achieved by calculating the variance of the points in a local neighborhood.

[0069] Gaussian Fitting: A Gaussian is fit to the local neighborhood of each point to determine its mean and variance.

[0070] Noise removal: Use a Gaussian model to remove noise from each point, usually through weighted averaging or filtering. The weight is usually inversely proportional to the distance from the point to the center of the neighborhood, determined by the Gaussian kernel.

[0071] Result update: Update the position of each point to reflect the denoising result.

[0072] S2, extracting the region of interest from the denoised three-dimensional point cloud data using a human eye-simulated target of interest extraction algorithm;

[0073] Furthermore, the human eye-simulating object of interest extraction algorithm utilizes an encoder-decoder structure to simulate the context information of the point cloud through a three-dimensional point cloud context attention change module encoder and a scene context attention change change decoder.

[0074] Specifically, the human eye-simulating object of interest extraction algorithm is a point cloud region of interest ROI detection algorithm with a self-attention mechanism. The algorithm uses an encoder-decoder structure and makes full use of the self-attention mechanism to simulate multi-scale and scene context feature information to achieve ROI detection of the point cloud of interest.

[0075] Point cloud ROI detection algorithm with self-attention mechanism: First, the encoder extracts multi-level features of the point cloud, and the decoder gradually restores the resolution of the point cloud. Specifically, given a fixed view of N points V = {v1, v2, ..., vN}, each V is a raw feature (such as XYZ position information and RGB color information), it is then input to the multi-level encoder to understand the semantic information of the entire scene. The encoder extracts local context and simulates global context through two different transformers.

[0076] The decoder decodes the features produced by the encoder through upsampling and transformers, and then integrates the multi-scale features of the encoder through the context aggregation module to learn the global scene context information. Specifically, the decoder upsamples the output value of the encoder, and integrates this value and the output of the encoder to the transformer to form a decoder; the context aggregation module takes the output of the multi-level encoder as input, connects them together as the global context, and forms aggregated information.

[0077] Finally, the result is output to the prediction head, the output features are converted into predictions of interest, and finally the point cloud ROI region of interest is generated. Specifically, the output value of the encoder is fused and spliced ​​with the context aggregation module, and then the features of different sources are fused and converted through the multi-layer perceptron (MLP), and finally the prediction result is output to form the point cloud ROI region of interest.

[0078] S3, using the human eye-simulating target compression algorithm to retain the extracted region of interest at high resolution and perform non-uniform compression on the non-interest region. The flowchart of the human eye-simulating target compression algorithm is as follows: Figure 2 shown.

[0079] Specifically, by studying the principles of retinal imaging in the human eye, it was found that the human retina has a non-uniform structure, with more central cells and high resolution characteristics, and the surrounding cells gradually decrease. The data of targets of non-interest can be compressed, so that clear imaging can be achieved in the central gaze area while early warning can be given in the surrounding non-interest areas, achieving rapid imaging. This lays the foundation for resolving the contradiction between the amount of data and real-time performance of three-dimensional target recognition technology.

[0080] Furthermore, the process of retaining the extracted region of interest at high resolution and performing non-uniform compression on the non-interest region by using the human eye-simulated target of interest compression algorithm includes:

[0081] S31, using a three-dimensional retinal variable resolution positive change model, converting the point cloud data from a Cartesian coordinate system to a polar coordinate system, and performing a logarithmic polar coordinate change;

[0082] The core of the human eye retinal variable resolution sampling model is the logarithmic polar coordinate change, which is the mainstream idea for achieving spatial variable resolution imaging. The human eye sampling model is a geometric image transformation that simulates the mapping relationship between the human eye retina and the brain's visual cortex.

[0083] First, convert the point cloud data from the Cartesian coordinate system to the polar coordinate system. For a point P (x, y, z) in a three-dimensional space, the polar coordinates (r, θ, φ) of the point of interest P0 (x0, y0, x0) can be converted from the Cartesian coordinate system to the polar coordinate system by the following formula:

[0084]

[0085] θ=arctan2(y-y0,x-x0) (4)

[0086]

[0087] Where r is the distance from the point to the origin, θ is the angle between the projection of the point on the XY plane and the X-axis, and φ is the angle between the point and the Z-axis.

[0088] In order to further process the point cloud data in the polar coordinate system, the logarithmic transformation is introduced. The logarithmic transformation can effectively compress the data range and reduce the dynamic range of the data. The logarithmic polar coordinates (logr, θ, φ) can be calculated by the following formula to obtain the point cloud in the logarithmic polar coordinates:

[0089] log r =log(r+ε) (6)

[0090] Here, ε is a small constant used to prevent taking the logarithm of 0.

[0091] S32, using a three-dimensional retinal variable resolution compression model, and achieving uniform density distribution of point cloud data through a voxel compression point cloud method;

[0092] Specifically, in the field of visual perception, the structural characteristics of the human eye's retina are manifested in that the fovea area has a higher visual resolution, and as it transitions to the peripheral area, the resolution gradually decreases. This non-uniform distribution of visual sensitivity inspires the sampling strategy in point cloud data processing. In the logarithmic polar coordinate system, point cloud data can be effectively simulated by parameterizing the polar diameter and angle to simulate this non-uniform resolution characteristic of the human eye. Based on this, this embodiment adopts a voxel compression point cloud method to achieve a uniform density distribution of point cloud data. The specific process is as follows:

[0093] S3201, space division: First, divide the point cloud space into voxel grids of the same size, each voxel is a cubic area. The key to this step is to determine the size of the voxel, which will directly affect the effect and efficiency of downsampling. The size of the voxel is usually set by the user according to application requirements. Smaller voxels can retain more details, but increase the amount of calculation; larger voxels can reduce the amount of calculation, but may lose some details.

[0094] S3202, point replacement: For each voxel, all points in it are replaced with a representative point. This representative point is usually the centroid or center point of these points. The formula for calculating the centroid is as follows

[0095]

[0096] Among them, (x i ,y i ,z i ) is the coordinate of the ith point in the voxel, and m is the total number of points in the voxel.

[0097] S3203, downsampling: Through the above process, the original point cloud is downsampled to fewer points while maintaining the overall structure and shape of the point cloud.

[0098] S33, using a three-dimensional retinal variable resolution inverse change model, converting the point cloud in the logarithmic polar coordinates back to a point cloud in the Cartesian coordinate system;

[0099] In order to facilitate the intuitive understanding of deep learning networks and human visual perception, it is also necessary to convert the point cloud in logarithmic polar coordinates back to the point cloud in Cartesian coordinates, as shown in the formula:

[0100] x=(e logr -ε)sin(φ)cos(θ)+x0

[0101] y=(e logr -ε)sin(φ)sin(φ)+y0

[0102] z=(e logr -ε)cos(φ)+z0 (8)

[0103] Where (x, y, z) is the coordinate in the Cartesian coordinate system, and (x0, y0, z0) is the center point of the region of interest.

[0104] S34. Use the three-dimensional retinal variable resolution point cloud fusion model to fuse the point cloud.

[0105] When there are multiple targets of interest, the point clouds need to be fused. The fusion steps are:

[0106] First, the point cloud is superimposed: all the objects of interest are superimposed in sequence to form a point cloud fusion point cloud, and then the point cloud is deduplicated: first, a KD tree index is established for the point cloud. The KD tree (k-dimensional tree) is a data structure used to organize k-dimensional spatial data, which can efficiently perform spatial queries. In the point cloud deduplication operation, the KD tree can be used to quickly find the neighborhood points of a given point. Then, for each point in the point cloud, the KD tree is used to perform a radius search to find the neighborhood points of each point. Next, the point cloud is deduplicated: duplicate points are removed by comparing the indexes of the neighborhood points. Finally, a compressed point cloud is formed.

[0107] The effect of the human eye-like interesting target compression algorithm is shown in the figure. Figure 3 shown.

[0108] S4. Use a display program to display the compression result.

[0109] like Figure 4 The figure shows the complete flow chart of human-eye-like 3D point cloud data compression in the field of autonomous driving using the public KITTI dataset, as well as its verification in target recognition applications.

[0110] (a) RGB image taken by the visible light camera of the KITTI dataset;

[0111] (b) The point cloud image formed after denoising after the original point cloud data is input into the point cloud preprocessing algorithm;

[0112] (c) Two regions of interest are formed by using the human eye-simulating object of interest extraction algorithm;

[0113] (d) is the point cloud compression data obtained after the human eye-like object compression algorithm;

[0114] (e) The result obtained by feeding the compressed input to the open source PointRCNN 3D object recognition algorithm.

[0115] After compression by the above-mentioned human eye-simulating three-dimensional point cloud data compression method, the number of original point clouds was compressed from 19388 to 798, with a compression ratio of 24.3 times, effectively identifying cars, pedestrians, and cyclists. Compared with the point cloud method formed by traditional imaging, the above scheme has a recognition speed of 14.7ms instead of the original 102ms, which increases the recognition rate by 6.94 times. In addition, in the KITTI data set, pedestrian target recognition is performed, and the recognition target accuracy is averaged in the three modes of simple, medium, and difficult, from the original 52.28% average recognition accuracy to 53.62%, an increase of 1.34% accuracy. The reason for this is that the resolution adaptive processing method has the effect of compressing redundancy, which makes it possible to more fully extract the effective feature information of the point cloud, thereby improving the accuracy of target detection and recognition.

[0116] The present invention obtains the target of interest in a low-resolution image and maps it to a high-resolution image, then changes the resolution of the high-resolution image to simulate the human eye, thereby effectively retaining the high resolution of the attention image and compressing the data in the non-attention area, greatly improving the information utilization efficiency of the system, and finally passing the human eye-like variable resolution image into the target recognition network to recognize the target in the field of view, effectively improving the accuracy of target recognition, and being able to meet the large field of view, high dynamic and complex environment recognition tasks in the field of autonomous driving.

[0117] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A human eye-simulating three-dimensional point cloud data compression method, characterized in that: The following steps are involved: Obtain the original 3D point cloud data of autonomous driving, and perform Gaussian filtering and denoising on the original 3D point cloud data; The region of interest is extracted from the denoised 3D point cloud data using an algorithm for extracting objects of interest that simulates the human eye. The human eye-simulated object of interest compression algorithm is used to retain the high resolution of the extracted region of interest and perform non-uniform compression on the non-interest region.

2. The human eye-simulating three-dimensional point cloud data compression method according to claim 1, characterized in that: The process of Gaussian filtering and denoising the original 3D point cloud data includes: Estimate the noise level of each point in the original point cloud data; Perform a Gaussian fit on the local neighborhood of each point to determine its mean and variance; Use Gaussian model to denoise each point; Update the position of each point to reflect the denoising result.

3. The human eye-simulating three-dimensional point cloud data compression method according to claim 1, characterized in that: The human eye-simulating object of interest extraction algorithm utilizes an encoder-decoder structure to simulate the context information of the point cloud through a three-dimensional point cloud context attention change module encoder and a scene context attention change change decoder.

4. The human eye-simulating three-dimensional point cloud data compression method according to claim 1, characterized in that: The process of using the human eye-simulating object of interest compression algorithm to retain the high resolution of the extracted region of interest and non-uniformly compressing the non-interest region includes: Using the three-dimensional retinal variable resolution positive change model, the point cloud data is converted from the Cartesian coordinate system to the polar coordinate system, and the polar coordinates are changed logarithmically; Using the 3D retinal variable resolution compression model, the uniform density distribution of point cloud data is achieved through voxel compression point cloud method; The point cloud in log polar coordinates is converted back to the point cloud in Cartesian coordinates using the 3D retinal variable resolution inverse change model; The point cloud is fused using the 3D retinal variable resolution point cloud fusion model.

5. The human eye-simulating three-dimensional point cloud data compression method according to claim 4, characterized in that: The formulas for converting point cloud data from Cartesian coordinates to polar coordinates and logarithmic polar coordinates include: θ=arctan2(y-y0,x-x0) Where r is the distance from the point to the origin, θ is the angle between the projection of the point on the XY plane and the X-axis, and φ is the angle between the point and the Z-axis; The calculation formula for the point cloud in logarithmic polar coordinates is: log r =log(r+ε) Among them, ∈ is a constant.

6. The human eye-simulating three-dimensional point cloud data compression method according to claim 4, characterized in that: The process of achieving uniform density distribution of point cloud data by voxel compression point cloud method using the three-dimensional retinal variable resolution compression model includes: Divide the point cloud space into voxel grids of the same size, each voxel is a cubic area, and the voxel size is set according to application requirements; All points in each voxel are replaced with representative points, which are the centroid or center point of the points in the voxel; The original point cloud is downsampled to fewer points through point replacement, maintaining the overall structure and shape of the point cloud.

7. The human eye-simulating three-dimensional point cloud data compression method according to claim 4, characterized in that: The formula for converting the point cloud in logarithmic polar coordinates back to the point cloud in Cartesian coordinate system is: x=(e logr -ε)sin(φ)cos(θ)+x0 y=(e logr -ε)sin(φ)sin(φ)+y0 z=(e logr -ε)cos(φ)+z0 Among them, (x, y, z) are the coordinates in the Cartesian coordinate system, and (x0, y0, z0) is the center point of the area of ​​interest.

8. The human eye-simulating three-dimensional point cloud data compression method according to claim 4, characterized in that: The process of fusing point clouds includes: The point cloud data of all the targets of interest are superimposed in sequence to form a fused point cloud; Establish a KD tree index to organize the spatial data of the fused point cloud to facilitate efficient spatial query; Use KD tree to perform radius search to find the neighboring points of each point in the fused point cloud; By comparing the indexes of neighboring points, duplicate points in the fused point cloud are removed.

9. A computer storage medium, characterized in that The computer storage medium stores computer program instructions, which, when executed by a processor, implement the human eye-simulating three-dimensional point cloud data compression method as described in any one of claims 1 to 8.

Images