Point cloud compression method and system based on depth map and redundant point removal

Abstract

The invention relates to a point cloud compression method based on a depth map and redundant point removal, and belongs to the technical field of data compression, and the method comprises the following steps: S1, converting single-frame point cloud data into a depth map for representation; s2, estimating a horizon line; s3, removing perigees in the depth map; s4, clustering the remaining points of the depth map; s5, redundant points are removed according to the density characteristics of the point cloud clusters; s6, performing intra-frame prediction on the point cloud data; and S7, compressing the obtained point cloud cluster label, the point cloud cluster center value and the prediction residual error of each frame of distance image into a bit stream, and packaging the bit stream into a file.

Description

Technical Field

[0001] The present invention belongs to the technical field of data compression, and relates to a point cloud compression method and system based on depth map and redundant point removal. Background Art

[0002] In recent years, LiDAR (Light Detection and Ranging) has gradually become an indispensable part of many practical applications, such as self-driving cars, mobile robots, and drones. LiDAR point cloud can accurately obtain a three-dimensional representation of the surrounding environment, namely a three-dimensional (3D) LiDAR point cloud. LiDAR point cloud consists of geometric information (three-dimensional coordinates) and corresponding attribute information (reflection intensity) of points, and has the characteristics of sparseness, disorder, and large amount of data. Taking the most common LiDAR as an example, the number of points generated per second is as high as more than two million. The transmission and storage of such a large amount of data will undoubtedly bring huge pressure to the limited network bandwidth and storage devices, thus greatly hindering the development of LiDAR point cloud related applications and the implementation of point cloud related technologies. Therefore, exploring how to efficiently encode LiDAR point clouds has become a problem that needs to be solved by academia and industry, and has important research and application value.

[0003] One of the core issues of LiDAR point cloud coding is how to better remove the spatial redundancy of point cloud geometric information. Due to the sparsity of LiDAR point clouds, especially outdoor scenes, there will be a large number of discontinuous areas, which means that there are no points in a considerable part of the space. Therefore, the existing representation methods are relatively inefficient. In addition, most of the current methods only focus on data fidelity considerations, often ignoring that LiDAR point cloud data is not for human vision, but for machine vision. The ultimate goal of LiDAR point cloud coding is to reduce the amount of data transmitted by point clouds and enable each terminal to complete a series of machine vision tasks with excellent performance at the decoding end. Therefore, when considering the design of the encoder, it is crucial to introduce the performance evaluation of machine perception tasks. Furthermore, some existing methods are complex and real-time encoding is difficult. From a practical point of view, real-time performance helps ensure the realization of many application tasks. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a point cloud compression method and system based on depth map and redundant point removal, which converts the single-frame original point cloud into a depth map for representation so as to structure the sparse and disordered points in the single-frame point cloud data, and clusters the non-ground points in combination with the information of the depth map. In order to better take into account the machine perception application, the redundant points with low density after clustering are eliminated. Under the condition of the same compression rate, more finely quantified high-density objects can be used to improve the performance of machine perception applications.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] In one aspect, the present invention provides a point cloud compression method based on a depth map and redundant point removal, comprising the following steps:

[0007] S1: Convert single-frame point cloud data into depth map representation;

[0008] S2: Estimate the horizon;

[0009] S3: remove the perigee in the depth map;

[0010] S4: cluster the remaining points of the depth map;

[0011] S5: Remove redundant points based on the density characteristics of the point cloud cluster;

[0012] S6: Perform intra-frame prediction on point cloud data;

[0013] S7: compress the point cloud cluster labels, point cloud cluster center values ​​and prediction residuals of each frame of the range image into a bit stream and package them into a file.

[0014] Further, in step S1, the 3D points represented by the Cartesian coordinate system are mapped into the depth map, and the formula for converting the Cartesian coordinate system into the spherical coordinate system is:

[0015]

[0016] For a point P(x,y,z) in the point cloud, it is represented as P′(ρ,φ,θ) in the spherical coordinate system; the angle between P′ and the xy plane is the pitch angle φ, the angle between P′ and the xz plane is the yaw angle θ, and the distance from the origin is ρ;

[0017] The hollow cylinder is then projected onto the xz plane to form a depth map.

[0018] Further, in step S2, points with z component size <1.5 in the original point cloud are selected and put into the point set to be measured, and a suitable number of points are extracted from them to select the ground plane using the Ransic algorithm. The formula of the ground plane to be estimated is as follows:

[0019] Ax+By+Cz+D=0

[0020] A, B, C, and D are the parameters to be found.

[0021] Further, the use of the Ransic algorithm to perform ground plane specifically includes the following steps:

[0022] Randomly select N points from the set of points to be tested;

[0023] Calculate model parameters;

[0024] Take the remaining points and calculate whether the distance to the plane is less than the threshold. If so, record it as an inner point and count it. If not, record it as an outer point and count it.

[0025] When all points are taken, determine whether the number of inliers is greater than that of the previous iteration. If so, update the model parameters. If not, continue iterating until the maximum number of iterations is reached.

[0026] Furthermore, in step S3, the distance from each valid point in the depth map to the fitted ground plane is calculated, and the point whose distance is less than a typical value is set as the perigee.

[0027] Further, step S4 specifically includes the following steps:

[0028] S41: Group the points from the same object in each laser radar scanning channel; check the Euclidean distance between every two adjacent valid points, if the distance is greater than th H , a new channel cluster is created; if the distance is less than th H , then expand the current cluster; check the distance between the last point and the first point of the channel, if they have the continuous property, this will merge the two channel clusters;

[0029] S42: Perform inter-channel grouping and merge the groups on the channels by checking the connection of the filters; connect the filters to determine the connectivity between the current group and the adjacent channel groups; the center point of the filter passes through each valid point on the range image, and during the checking process, the filter checks the distance between the center point and each valid point above the current channel; if the distance is less than th V , then the two point cloud clusters are connected, and the cluster with a large number of labels is merged into the cluster with the smallest number of labels.

[0030] Furthermore, in step S5, redundant points are removed through a redundant point removal mechanism based on cluster density, specifically including: setting a typical density threshold, when the density of a point cloud cluster is lower than the threshold, the cluster is considered to be a low-density cluster, and all points in such clusters are divided into a predefined empty set; point cloud compression is performed using fidelity compression.

[0031] Further, step S6 specifically includes the following steps:

[0032] Divide the pixel block area according to the clustering result in step S5; use the mean value of the pixel block to replace the actual value of all pixels in the block;

[0033] For each point cloud cluster obtained after optimization, calculate its geometric center as the center point;

[0034] For each point in the point cloud cluster, the prediction residual value is calculated and obtained based on its positional relationship relative to the center point;

[0035] The solution formula for the prediction residual value is as follows:

[0036]

[0037] QuantFactor is used to adjust the compression rate of the point cloud to be compressed. Increasing QuantFactor can increase the compression rate of the point cloud to be compressed.

[0038] Further, step S7 specifically includes: using signed data compression for the point cloud cluster label and the prediction residual, using floating point data compression for the floating point data of the point cloud cluster center value, and then encapsulating the data after entropy encoding of the three into a code stream, and outputting a coded bit stream.

[0039] On the other hand, the present invention provides a point cloud compression system based on depth map and redundant point removal, comprising:

[0040] A depth map conversion module, used to obtain depth map data from 3D point cloud data;

[0041] A clustering segmentation module, used to divide each point in the depth map into multiple sub-point cloud clusters; the clustering segmentation module includes a ground plane estimation module, a ground plane removal module and a clustering module;

[0042] The ground plane estimation module is used to extract ground plane information from the point cloud;

[0043] The ground plane removal module is used to remove the perigee in the point cloud;

[0044] The clustering module is used to cluster the points in the depth map according to the horizontal and vertical relationships;

[0045] Redundant point elimination module, used to remove redundant points according to the density characteristics of point cloud clusters;

[0046] The intra-frame prediction module is used to obtain the center point of each point cloud cluster and obtain the prediction residual value of each point in the point cloud cluster relative to the center point of the point cloud cluster;

[0047] The compression coding module is used to compress the point cloud cluster labels, point cloud cluster center values ​​and prediction residuals of each frame of the range image into a bit stream and package them into a file.

[0048] The beneficial effect of the present invention is that a method for removing redundant points based on depth map and point cloud is proposed, which effectively removes redundant points in irregular sparse point cloud data to ensure that high perception capabilities are maintained in downstream machine perception tasks. In addition, the implementation device of the present invention adopts a parallel point clustering module, which can realize real-time point cloud compression processing on embedded or edge computing devices.

[0049] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0051] Figure 1 It is a schematic diagram of a point cloud compression method based on a depth map and redundant point removal according to Embodiment 1 of the present invention;

[0052] Figure 2 is a flow chart of the design of the ground plane estimation method in Embodiment 1 of the present invention;

[0053] Figure 3 is a schematic diagram of a bit stream encoding design in Embodiment 1 of the present invention;

[0054] Figure 4 It is a schematic diagram of the structure of the point cloud compression system described in Example 2 of the present invention. DETAILED DESCRIPTION

[0055] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0056] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0057] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.

[0058] Example 1

[0059] This embodiment is a preferred implementation of the point cloud compression method based on depth map and redundant point removal of the present invention.

[0060] See also Figure 1 ,like Figure 1 As shown, the method of this embodiment includes:

[0061] The single-frame point cloud data is converted into a depth map, that is, the 3D points represented by the Cartesian coordinate system are mapped into the depth map. In this invention, a 64*2000 resolution depth map is selected. The specific resolution used for mapping varies according to different laser radars. The formula for converting the Cartesian coordinate system to the spherical coordinate system is given below:

[0062]

[0063] Take a point P(x,y,z) in the point cloud as an example. It is represented as P′(ρ,φ,θ) in the spherical coordinate system. The angle between P′ and the xy plane is the pitch angle φ, the angle between P′ and the xz plane is the yaw angle θ, and the distance from the origin is ρ. Next, the hollow cylinder is projected onto the xz plane to form a depth map.

[0064] The second step is to estimate the ground plane. A large part of the points in the LiDAR point cloud are on the ground plane. The present invention selects points with z component size <1.5 in the original point cloud and puts them into the point set to be measured. An appropriate number of points are extracted from them and the Ransic algorithm is used to estimate the ground plane. The formula for the ground plane to be estimated is as follows:

[0065] Ax+By+Cz+D=0

[0066] A, B, C, and D are the parameters to be determined. The implementation process of the Ransic algorithm is as follows: Figure 2 As shown, the following steps are included:

[0067] 1) Randomly select N points from the set of points to be tested, and in this embodiment, 10 points are selected;

[0068] 2) Calculate model parameters;

[0069] 3) Take the remaining points and calculate whether the distance to the plane is less than the threshold. If yes, record it as an inner point and count it. If not, record it as an outer point and count it.

[0070] 4) When all points are taken, determine whether the number of inliers is greater than that of the previous iteration. If so, update the model parameters. If not, continue iterating until the maximum number of iterations is reached.

[0071] The third step is to remove the perigee in the depth map. This step calculates the distance from each valid point (i.e., non-zero point) in the depth map to the fitted ground plane. The point whose distance is less than a typical value is set as the perigee. This typical value is set to 0.1m in the present invention.

[0072] The fourth step is to cluster the remaining points in the depth map. As a perception task, point cloud clustering algorithms can be used to segment points into object instances. The first substep is to group points from the same object in each lidar scan channel. We check the Euclidean distance between every two adjacent valid points and if the distance is greater than th H , a new channel cluster is created; if the distance is less than th H , the current cluster is expanded. At the same time, due to the lidar scanning mechanism, each channel is a ring in three-dimensional space. Therefore, it is necessary to check the distance between the last point and the first point of the channel to ensure continuity. If it has a continuous property, the two channel clusters will be merged. The second sub-step is to perform inter-channel grouping. In this part, a filter-like connection check is designed to merge the groups on the channel. The connection filter determines the connectivity between the current group and the adjacent channel group. Due to the missing point problem, points belonging to the same object may be separated in multiple non-adjacent channels, resulting in over-segmentation problems. In this case, the size of the filter can determine the search space of the local range image area, thereby solving this problem. In this sub-step, the center point of the filter will pass through each valid point on the range image. During the inspection process, the filter will check the distance between the center point and each valid point above the current channel. If the distance is less than th V , then the two point cloud clusters are connected, and the cluster with a larger number of labels will be merged into the cluster with the smallest number of labels.

[0073] In the fifth step of the present invention, for the above classification results, the redundant points are further effectively removed according to the density characteristics of the point cloud clusters. Specifically, in the process of the fourth step, we perform preliminary clustering operations on the points in the depth map based on the horizontal and vertical distances to form several point cloud clusters. However, since the point cloud data itself has significant sparsity characteristics, if a too small distance segmentation step is used in the clustering process, a large number of point cloud clusters may be formed containing only a very small number of points, or even clusters with only a few points. Such low-density point cloud clusters will not only significantly reduce the compression efficiency of the point cloud data and increase the amount of data after encoding, but also in the downstream tasks of deep learning, such low-density point cloud clusters carry limited information and have little effect on improving the training and reasoning effects of the model. In order to optimize the compression rate of point cloud data and improve the performance of downstream tasks, this step introduces a redundant point removal mechanism based on cluster density. The specific method is to set a typical density threshold. When the density of a point cloud cluster is lower than the threshold, the cluster is considered to be a low-density cluster, and all points in such a cluster will be divided into a predefined empty set. In this way, the length of the encoded code stream can be effectively reduced, and the storage and transmission costs can be reduced. At the same time, since the key point information in the high-density clusters is retained, it ensures that sufficient feature information is retained during the compression process, so as not to affect the deep learning model's ability to understand and process point cloud data. In addition, in order to meet the task requirements of non-deep learning downstream tasks, this step can be closed by program control, and point cloud compression is performed using fidelity compression.

[0074] In the sixth step of the present invention, based on the aforementioned optimized clustering results, intra-frame prediction of point cloud data is realized. In the entire point cloud compression framework, frames are used as basic coding units for processing. Here, a frame is defined as a point cloud data set generated after one week of LiDAR scanning. The core idea of ​​the proposed intra-frame prediction algorithm is that the pixel values ​​of the distance image reflect the distance characteristics of each point in space to the origin of the LiDAR coordinate system. Specifically, for point groups belonging to the same object, their distance values ​​to the center of the LiDAR are usually similar. Therefore, during the processing, the pixel block area can be divided according to the clustering results in the fifth step. Subsequently, the mean value of the pixel block is used to replace the actual values ​​of all pixels in the block. This operation not only generates more uniform and flat areas, but also effectively reduces the amount of data that needs to be processed during the encoding process, thereby significantly improving the efficiency of subsequent encoding. The specific operation process is as follows:

[0075] First, before intra-frame prediction, the center point of each point cloud cluster needs to be obtained. Specifically, for each point cloud cluster obtained after optimization, its geometric center is calculated as the center point. Then, for each point in the point cloud cluster, based on its positional relationship relative to the center point, the prediction residual value is calculated and obtained. The prediction residual value is calculated by taking the difference between the actual coordinates of each point and the coordinates of the center point as the residual to capture the subtle changes of each point in the point cloud cluster, thereby providing effective prediction information for subsequent compression. The solution formula for the prediction residual value is as follows:

[0076]

[0077] By controlling QuantFactor, the compression rate of the point cloud to be compressed can be controlled. Increasing QuantFactor can increase the compression rate of the point cloud to be compressed.

[0078] After completing the above operations, the "point cloud cluster label", "point cloud cluster center value" (i.e. predicted pixel value) and "prediction residual" of each frame of the distance image can be obtained. The three sets of data obtained above will be further compressed. The specific process of bit stream encoding is as follows: Figure 3 As shown, the signed integer data such as the point cloud cluster label and the prediction residual are compressed by signed data, and the floating point data of the center value of the point cloud cluster is compressed by floating point data. Then, the data after entropy coding of the three are encapsulated into code stream and the coded bit stream is output.

[0079] Example 2

[0080] This embodiment is a preferred implementation of the point cloud compression device based on depth map and redundant point removal of the present invention.

[0081] See also Figure 4 , the device of this embodiment includes:

[0082] A depth map conversion module, used to obtain depth map data from 3D point cloud data;

[0083] The formula for converting the Cartesian coordinate system to the spherical coordinate system is:

[0084]

[0085] Take a point P(x,y,z) in the point cloud as an example. It is represented as P′(ρ,φ,θ) in the spherical coordinate system. The angle between P′ and the xy plane is the pitch angle φ, the angle between P′ and the xz plane is the yaw angle θ, and the distance from the origin is ρ. Next, the hollow cylinder is projected onto the xz plane to form a depth map.

[0086] The clustering segmentation module is used to divide each point in the depth map into multiple sub-point cloud clusters;

[0087] In this specific embodiment, the cluster segmentation module may further include:

[0088] The ground plane estimation module is used to extract the ground plane information in the point cloud; the method is: put the points into a buffer array according to the height information in the space, randomly sample a small number of points from the buffer array, and Figure 2 The flowchart shown performs ground plane estimation.

[0089] The ground plane removal module is used to remove ground points in the point cloud. The method is: calculate the distance from each valid point (i.e., non-zero point) in the depth map to the fitted ground plane. Set the point whose distance is less than a typical value as the perigee. This typical value is set to 0.1m in this device.

[0090] The clustering module clusters the points in the depth map according to their horizontal and vertical relationships. The specific implementation steps are as follows: The first sub-step is to group the points from the same object in each lidar scanning channel. We check the Euclidean distance between every two adjacent valid points. If the distance is greater than th H , a new channel cluster is created; if the distance is less than th H , the current cluster is expanded. At the same time, due to the lidar scanning mechanism, each channel is a ring in three-dimensional space. Therefore, it is necessary to check the distance between the last point and the first point of the channel to ensure continuity. If it has a continuous property, the two channel clusters will be merged. The second sub-step is to perform inter-channel grouping. In this part, a filter-like connection check is designed to merge the groups on the channel. The connection filter determines the connectivity between the current group and the adjacent channel group. Due to the missing point problem, points belonging to the same object may be separated in multiple non-adjacent channels, resulting in over-segmentation problems. In this case, the size of the filter can determine the search space of the local range image area, thereby solving this problem. In this sub-step, the center point of the filter will pass through each valid point on the range image. During the inspection process, the filter will check the distance between the center point and each valid point above the current channel. If the distance is less than th V , then the two point cloud clusters are connected, and the cluster with a larger number of labels will be merged into the cluster with the smallest number of labels.

[0091] The redundant point elimination module follows the cluster segmentation module, which eliminates clusters with smaller density according to the density statistics of each sub-point cloud cluster. These point cloud clusters may not be helpful for machine perception recognition. The specific implementation steps are: statistically divide clusters smaller than the typical value into empty sets, which can reduce the length of the encoded bitstream while retaining sufficient feature information. In addition, in order to meet the task requirements of non-deep learning downstream tasks, this step can be closed by program control, and point cloud compression is performed using fidelity compression.

[0092] The module for eliminating redundant points is followed by the intra-frame prediction module. The specific implementation steps are: first, the center point of each point cloud cluster is obtained, and the prediction residual value of each point in the point cloud cluster is calculated relative to the center point of the point cloud cluster.

[0093] After completing the above operations, the “point cloud cluster label”, “point cloud cluster center value” (ie, predicted pixel value) and “prediction residual” of each frame of the distance image can be obtained.

[0094] The last module required is the compression encoding module, which compresses the "point cloud cluster label", "point cloud cluster center value" (i.e. predicted pixel value) and "prediction residual" of each frame of the distance image obtained by the above modules into a bit stream and packages it into a file.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A point cloud compression method based on depth map and redundant point removal, characterized in that: The following steps are involved: S1: Convert single-frame point cloud data into depth map representation; S2: Estimate the horizon; S3: remove the perigee in the depth map; S4: Cluster the remaining points of the depth map; S5: Remove redundant points based on the density characteristics of the point cloud cluster; S6: Perform intra-frame prediction on point cloud data; S7: compress the point cloud cluster labels, point cloud cluster center values ​​and prediction residuals of each frame of the range image into a bit stream and package them into a file.

2. The point cloud compression method based on depth map and redundant point removal according to claim 1, characterized in that: In step S1, the 3D points represented by the Cartesian coordinate system are mapped to the depth map. The formula for converting the Cartesian coordinate system to the spherical coordinate system is: For a point P(x,y,z) in the point cloud, it is represented as P′(ρ,φ,θ) in the spherical coordinate system; the angle between P′ and the xy plane is the pitch angle φ, the angle between P′ and the xz plane is the yaw angle θ, and the distance from the origin is ρ; The hollow cylinder is then projected onto the xz plane to form a depth map.

3. The point cloud compression method based on depth map and redundant point removal according to claim 1, characterized in that: In step S2, points with z component size <1.5 in the original point cloud are selected and put into the point set to be tested, and a suitable number of points are extracted from them to select the ground plane using the Ransic algorithm. The formula of the ground plane to be estimated is as follows: Ax+By+Cz+D=0 A, B, C, and D are the parameters to be found.

4. The point cloud compression method based on depth map and redundant point removal according to claim 1, characterized in that: The Ransic algorithm is used to estimate the ground plane, specifically including the following steps: Randomly select N points from the set of points to be tested; Calculate model parameters; Take the remaining points and calculate whether the distance to the plane is less than the threshold. If so, record it as an inner point and count it. If not, record it as an outer point and count it. When all points are taken, determine whether the number of inliers is greater than that of the previous iteration. If so, update the model parameters. If not, continue iterating until the maximum number of iterations is reached.

5. The point cloud compression method based on depth map and redundant point removal according to claim 1, characterized in that: In step S3, the distance from each valid point in the depth map to the fitted ground plane is calculated, and the point whose distance is less than a typical value is set as the perigee.

6. The point cloud compression method based on depth map and redundant point removal according to claim 1, characterized in that: Step S4 The specific steps include: S41: Group the points from the same object in each laser radar scanning channel; check the Euclidean distance between every two adjacent valid points, if the distance is greater than th H , a new channel cluster is created; if the distance is less than th H , then expand the current cluster; check the distance between the last point and the first point of the channel, if they have the continuous property, this will merge the two channel clusters; S42: Perform inter-channel grouping and merge the groups on the channels by checking the connection of the filters; connect the filters to determine the connectivity between the current group and the adjacent channel groups; the center point of the filter passes through each valid point on the range image, and during the checking process, the filter checks the distance between the center point and each valid point above the current channel; if the distance is less than th V , then the two point cloud clusters are connected, and the cluster with a large number of labels is merged into the cluster with the smallest number of labels.

7. The point cloud compression method based on depth map and redundant point removal according to claim 1, characterized in that: In step S5, redundant points are removed through a redundant point removal mechanism based on cluster density, which specifically includes: setting a typical density threshold. When the density of a point cloud cluster is lower than the threshold, the cluster is considered to be a low-density cluster, and all points in such clusters are divided into a predefined empty set; and point cloud compression is performed using fidelity compression.

8. The point cloud compression method based on depth map and redundant point removal according to claim 1, characterized in that: Step S6 specifically includes the following steps: Divide the pixel block area according to the clustering result in step S5; use the mean value of the pixel block to replace the actual value of all pixels in the block; For each point cloud cluster obtained after optimization, calculate its geometric center as the center point; For each point in the point cloud cluster, the prediction residual value is calculated and obtained based on its positional relationship relative to the center point; The solution formula for the prediction residual value is as follows: QuantFactor is used to adjust the compression rate of the point cloud to be compressed. Increasing QuantFactor can increase the compression rate of the point cloud to be compressed.

9. The point cloud compression method based on depth map and redundant point removal according to claim 1, characterized in that: Step S7 specifically includes: using signed data compression for the point cloud cluster labels and prediction residuals, using floating-point data compression for the floating-point data of the point cloud cluster center value, and then encapsulating the data after entropy encoding of the three into a bit stream to output a coded bit stream.

10. A point cloud compression system based on depth map and redundant point removal, characterized in that: include: A depth map conversion module, used to obtain depth map data from 3D point cloud data; A clustering segmentation module, used to divide each point in the depth map into multiple sub-point cloud clusters; the clustering segmentation module includes a ground plane estimation module, a ground plane removal module and a clustering module; The ground plane estimation module is used to extract ground plane information from the point cloud; The ground plane removal module is used to remove the perigee in the point cloud; The clustering module is used to cluster the points in the depth map according to the horizontal and vertical relationships; Redundant point elimination module, used to remove redundant points according to the density characteristics of point cloud clusters; The intra-frame prediction module is used to obtain the center point of each point cloud cluster and obtain the prediction residual value of each point in the point cloud cluster relative to the center point of the point cloud cluster; The compression coding module is used to compress the point cloud cluster labels, point cloud cluster center values ​​and prediction residuals of each frame of the range image into a bit stream and package them into a file.

Images