A method and device for adaptive point cloud geometry encoding and decoding

Abstract

The present invention discloses an adaptive point cloud geometry encoding and decoding method and device, the encoding method comprising: obtaining original point cloud data; performing geometric prediction on each point in the original point cloud data to obtain prediction residuals; selectively setting the prediction residuals to zero according to preset rules; encoding the prediction residuals to obtain a geometric information code stream. The adaptive point cloud geometry encoding method provided by the present invention selectively sets the geometric prediction residuals to zero according to certain rules, and does not perform encoding, thereby reducing the geometric code stream while ensuring a certain geometric distortion, thereby improving the geometric encoding efficiency; at the same time, by setting different encoding modes, the optimal encoding mode for each point is adaptively selected, further improving the geometric encoding efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of coding and decoding, and in particular relates to an adaptive point cloud geometry coding and decoding method and device. Background Art

[0002] In the point cloud G-PCC encoder framework, the geometric information of the point cloud and the attribute information corresponding to each point are encoded separately. At present, the geometric coding and decoding of G-PCC can be divided into octree-based geometric coding and decoding and prediction tree-based geometric coding and decoding. The prediction tree-based geometric coding first sorts the input point cloud and establishes the prediction tree structure at the encoding end by using two different methods. Then, based on the structure of the prediction tree, each node in the prediction tree is traversed, and the geometric position information of the node is predicted by selecting different prediction modes to obtain the prediction residual, and the geometric prediction residual is quantized using the quantization parameter. Finally, through continuous iteration, the prediction residual of the prediction tree node position information, the prediction tree structure, and the quantization parameters are encoded to generate a binary code stream.

[0003] At present, the commonly used prediction tree structures mainly include the prediction tree based on KD-Tree (high-latency slow mode) and the prediction tree based on lidar calibration information.

[0004] However, for the existing KD-Tree-based prediction tree structure and the laser radar calibration information-based prediction tree geometric prediction coding scheme, when the quantized point cloud after geometric preprocessing is geometrically predicted, the geometric prediction residual of each node is obtained through the prediction tree structure, and the prediction residual is losslessly coded regardless of whether it is geometrically lossless or lossy. In this way, although the quality of geometric reconstruction can be guaranteed in the case of geometric lossy coding, the impact brought about is that more information needs to be compressed, thereby limiting the geometric coding efficiency based on the prediction tree. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides an adaptive point cloud geometry encoding and decoding method and device. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] An adaptive point cloud geometry encoding method, comprising:

[0007] Get the original point cloud data;

[0008] Performing geometric prediction on each point in the original point cloud data to obtain a prediction residual;

[0009] Selectively setting the prediction residual to zero according to a preset rule;

[0010] The prediction residual is encoded to obtain a geometric information code stream.

[0011] In one embodiment of the present invention, selectively setting the prediction residual to zero according to a preset rule includes:

[0012] Setting the encoding mode includes mode 1 and mode 2;

[0013] Determine the best encoding mode for each point;

[0014] If mode 1 is selected for encoding, the prediction residual is not set to zero; if mode 2 is selected for encoding, the prediction residual is set to zero or part of the prediction residual is set to zero.

[0015] In one embodiment of the present invention, determining the best encoding mode for each point includes:

[0016] If the preset condition is determined to be met, the best coding mode for the current point is determined according to rate-distortion optimization; or

[0017] Determine the best encoding mode for the current point according to the number of reconstructed points; or

[0018] Directly use mode 2 as the best encoding mode for the current point;

[0019] Otherwise, the mode 1 is used as the best encoding mode for the current point.

[0020] In one embodiment of the present invention, determining the best encoding mode for the current point according to rate-distortion optimization includes:

[0021] Constructing a cost factor according to the prediction residual of the current point and the bitstream size under different coding modes;

[0022] Selecting the encoding mode corresponding to the optimal cost factor as the optimal encoding mode for the current point; or

[0023] An average cost factor is calculated according to the cost factors of the reconstructed points, and an optimal encoding mode for the current point is selected according to the average cost factor and a preset threshold.

[0024] In one embodiment of the present invention, determining the best encoding mode of the current point according to the number of reconstruction points includes:

[0025] When the number of reconstructed points is less than a preset value, mode 1 is selected as the best encoding mode for the current point; when the number of reconstructed points is greater than or equal to the preset value, mode 2 is selected as the best encoding mode for the current point.

[0026] In one embodiment of the present invention, before performing geometric prediction on each point in the original point cloud data to obtain the prediction residual, the method further includes:

[0027] The original point cloud data is analyzed and divided into a first type of data and a second type of data.

[0028] In one embodiment of the present invention, determining the optimal encoding mode for each point further includes:

[0029] Mode 1 is used as the optimal encoding mode for each point in the first category of data, and mode 2 is used as the optimal encoding mode for each point in the second category of data.

[0030] Another embodiment of the present invention provides an adaptive point cloud geometry encoding device, comprising:

[0031] A first data acquisition module, used to acquire original point cloud data;

[0032] A first prediction module, used for performing geometric prediction on each point in the original point cloud data to obtain a prediction residual;

[0033] A residual zeroing module, used for selectively setting the prediction residual to zero according to a preset rule;

[0034] The encoding module is used to encode the prediction residual to obtain a geometric information code stream.

[0035] Another embodiment of the present invention provides an adaptive point cloud geometry decoding method, comprising:

[0036] Get the geometric information code stream and decode it to get the parsed data;

[0037] Perform geometric prediction on the current point to obtain the geometric prediction value;

[0038] Determine whether the prediction residual of the current point is set to zero according to the analytical data, and obtain the prediction residual of the current point;

[0039] The geometric point cloud is reconstructed according to the geometric prediction value and the prediction residual of the current point to obtain reconstructed point cloud data.

[0040] Another embodiment of the present invention further provides an adaptive point cloud geometry decoding device, comprising:

[0041] The second data acquisition module is used to acquire the geometric information code stream and decode it to obtain the parsed data;

[0042] The second prediction module is used to perform geometric prediction on the current point to obtain a geometric prediction value;

[0043] A residual analysis module, used to determine whether the prediction residual of the current point is set to zero according to the analysis data, and obtain the prediction residual of the current point;

[0044] The reconstruction module is used to reconstruct the geometric point cloud according to the geometric prediction value and the prediction residual of the current point to obtain reconstructed point cloud data.

[0045] Beneficial effects of the present invention:

[0046] 1. The adaptive point cloud geometric prediction coding method provided by the present invention selectively sets the geometric prediction residual to zero according to certain rules and does not perform coding. Under the condition of ensuring a certain geometric distortion, the geometric code stream is reduced, thereby improving the geometric coding efficiency;

[0047] 2. The present invention sets different coding modes and adaptively selects the optimal coding mode for each point, thereby further improving the geometric coding efficiency;

[0048] 3. The present invention sets a threshold value of the number of reconstruction points to reasonably determine the encoding mode, which can not only improve the encoding efficiency, but also reduce the time complexity of decoding, thereby improving the decoding efficiency;

[0049] 4. When establishing a prediction tree structure based on lidar information, the present invention uses the nearest neighbor interpolation calculation to obtain the r value of the adjusted node, which can further improve the quality of geometric reconstruction, thereby ensuring the geometric coding efficiency.

[0050] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Schematic diagram of an adaptive point cloud geometry encoding method provided by an embodiment of the present invention;

[0052] Figure 2 It is a block diagram of point cloud geometry encoding based on prediction tree provided by an embodiment of the present invention;

[0053] Figure 3 is a schematic diagram of a prediction tree structure of four laser scanners provided in an embodiment of the present invention;

[0054] Figure 4 is a schematic diagram of a two-dimensional regular structure of a point cloud provided by an embodiment of the present invention;

[0055] Figure 5 It is a schematic diagram of the structure of an adaptive point cloud geometry encoding device provided by an embodiment of the present invention;

[0056] Figure 6 Schematic diagram of an adaptive point cloud geometry decoding method provided by an embodiment of the present invention;

[0057] Figure 7 is a block diagram of adaptive point cloud geometry decoding provided by an embodiment of the present invention;

[0058] Figure 8 It is a schematic diagram of the structure of an adaptive point cloud geometry decoding device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0059] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0060] Embodiment 1

[0061] See also Figure 1 , Figure 1 Schematic diagram of an adaptive point cloud geometry prediction coding method provided by an embodiment of the present invention, comprising:

[0062] S1: Get the original point cloud data.

[0063] Specifically, the original point cloud data is usually composed of a set of three-dimensional space points, each of which records its own geometric position information, as well as additional attribute information such as color, reflectivity, and normal. Among them, the geometric position information of the point cloud is generally represented based on a Cartesian coordinate system, that is, it is represented by the x, y, and z coordinates of the point. The original point cloud data can be obtained through 3D scanning equipment such as lidar, or through public data sets provided by various platforms. In this embodiment, it is assumed that the geometric position information of the original point cloud data obtained is represented based on a Cartesian coordinate system. It should be noted that the method of representing the geometric position information of the original point cloud data is not limited to Cartesian coordinates.

[0064] S2: Perform geometric prediction on each point in the original point cloud data to obtain the prediction residual.

[0065] In this embodiment, an octree structure may be used for geometric prediction, or a prediction tree structure may be used for geometric prediction. The following takes the geometric coding based on the prediction tree as an example for detailed description. Figure 2 , Figure 2 It is a block diagram of point cloud geometry encoding based on prediction tree provided by an embodiment of the present invention, which specifically includes:

[0066] 21) voxelize and reorder the original point cloud data to obtain sorted point cloud data.

[0067] 22) Establish a prediction tree structure for the sorted point cloud data.

[0068] 23) Select a certain prediction model based on the prediction tree structure.

[0069] 24) Perform geometric prediction on each node in the prediction tree structure according to the selected prediction mode to obtain the geometric prediction value and prediction residual of the current node.

[0070] 25) Encode the prediction residual to obtain a geometric information code stream.

[0071] Specifically, for the establishment of the prediction tree, the prediction tree structure can be established using KD-tree, and the prediction tree structure can also be established according to the laser radar calibration information. It should be noted that before establishing the prediction tree structure according to the laser radar calibration information, the point cloud data must be converted from the Cartesian coordinate system to the cylindrical coordinate system.

[0072] This embodiment takes the establishment of a prediction tree structure based on KD-tree as an example for detailed introduction.

[0073] First, the original point cloud is quantized and duplicate points are removed, also known as voxelization, to obtain voxelized point cloud data. The voxelized point cloud is reordered to facilitate the subsequent construction of a more efficient prediction tree. Specifically, the order can be random, Morton order, azimuth order, and radial distance order.

[0074] Since there is no need to perform coordinate transformation when establishing a prediction tree structure based on KD-tree, the prediction tree structure based on KD-tree can be directly established according to the sorted point cloud data.

[0075] In the whole process, KD-Tree is used to search for the nearest neighbor. At the beginning, KD-Tree is empty. Each node is visited in turn in the selected order, and the K nearest neighbors are found in KD-Tree using the position of the current node. The nearest neighbors found are used as child nodes of the current point for prediction according to certain criteria. For example, the nearest neighbors can be found as child nodes of the current point for prediction according to the size of the geometric prediction residual of the current node, the number of child nodes of the current node, or the frequency selected by each prediction mode. Next, a new prediction mode is created based on the current node, and the position corresponding to each prediction mode is added to KD-Tree. This process is repeated until all nodes are traversed and the entire prediction tree encoding structure is established.

[0076] Next, the prediction mode is selected.

[0077] Specifically, in this embodiment, a prediction tree is established by depth-first traversal, and each node in the tree can only be predicted by its ancestors.

[0078] Furthermore, this embodiment sets the following four prediction modes:

[0079] Mode0: No prediction

[0080] Mode1: Delta prediction (ie p0)

[0081] Mode2: Linear prediction (i.e. 2p0-p1)

[0082] Mode3: Parallelogram prediction (i.e. p0+p1-p2)

[0083] Among them, p0, p1, and p2 are the positions of the parent node, grandfather node, and great-grandfather node of the current node respectively. The best prediction mode can be selected for the current node according to the reconstruction quality for prediction.

[0084] Finally, prediction is made for each node according to the corresponding prediction mode.

[0085] Specifically, this embodiment uses Cartesian coordinate prediction to implement geometric prediction. If the current node is the root node of the prediction tree, the prediction mode selects Mode0, that is, the cylindrical coordinates of the current node are not predicted, and the corresponding Cartesian coordinate prediction value is (0,0,0). If the current node is not the root node, and the prediction mode selects Mode1, the Cartesian coordinates of the current point are reconstructed through the geometric coordinates of its parent node. If the current node is not the root node and the prediction mode is Mode2 or Mode3, the Cartesian coordinates of the current point are predicted using the corresponding prediction method. The predicted value of the Cartesian coordinates of the current node is

[0086] At this point, the geometric prediction value of the current node is obtained. The geometric position information of the current node is differentially predicted to obtain the prediction residual of the current node (r x ,r y ,r z ).

[0087] The geometric prediction value and prediction residual of the current point together with the prediction mode adopted by the current point, the quantization parameter of the current point, the number of repeated points and the number of child nodes are used as the information to be encoded.

[0088] S3: Selectively set the prediction residual to zero according to preset rules.

[0089] First, set the encoding mode to include mode 1 and mode 2.

[0090] In this embodiment, the method of encoding the prediction residual is defined as mode 1, which is represented by DirectMode. The method of setting the prediction residual to zero, that is, not needing to encode the prediction residual of the current point, is defined as mode 2, which is represented by RegularMode. In the geometric prediction coding based on the KD-Tree structure, the prediction residual (r x ,r y ,r z) is set to zero, that is, the RegularMode coding mode does not encode the prediction residual of the KD-Tree prediction tree structure.

[0091] Then, the best encoding mode for each point is determined. If mode 1 is selected for encoding, the prediction residual is not set to zero; if mode 2 is selected for encoding, the prediction residual is set to zero or part of the prediction residual is set to zero.

[0092] Specifically include:

[0093] If it is determined that the preset condition is met, the optimal coding mode for the current point is determined according to the rate-distortion optimization model; or

[0094] Determine the best encoding mode for the current point according to the number of reconstructed points; or

[0095] Directly use mode 2 as the best encoding mode for the current point;

[0096] Otherwise, the mode 1 is used as the best encoding mode for the current point.

[0097] In this embodiment, a coding switch can be set as a preset condition to determine whether to turn on the Regular Mode mode. When the coding switch is turned on, that is, the preset condition is met, the Regular Mode mode is used as an encoding option, and when the coding switch is turned off, that is, the preset condition is not met, the Direct Mode mode is directly used for encoding.

[0098] Specifically, geom_enable_rdo_flag is introduced in the gps (Geometry parameter set syntax) parameter set to guide whether the geometry turns on the RegularMode encoding algorithm in the entire encoding. When geom_enable_rdo_flag is 1, it means turned on; otherwise it is turned off. See Table 1 for details. Correspondingly, tile_geom_enable_rdo_flag is also introduced in the Tile inventory syntax to determine whether the RegularMode encoding algorithm is turned on for each Tile. When tile_geom_enable_rdo_flag is 1, it means turned on; otherwise it is turned off. See Table 2 for details. In each slice, gsh_geom_enable_rod_flag is also introduced to determine whether the RegularMode encoding algorithm is turned on for each slice. When gsh_geom_enable_rod_flag is 1, it means turned on; otherwise it is turned off. See Table 3 for details.

[0099] In this embodiment, when the encoding switch is turned on, a rate-distortion optimization model can be introduced to determine the best encoding mode at the current point. The details are as follows:

[0100] Construct a cost factor based on the prediction residual of the current point and the bitstream size under different coding modes;

[0101] Selecting the encoding mode that makes the cost factor optimal as the best encoding mode for the current point; or

[0102] An average cost factor is calculated according to the cost factors of the reconstructed points, and an optimal encoding mode for the current point is selected according to the average cost factor and a preset threshold.

[0103] In this embodiment, for the prediction tree based on KD-tree, the first cost factor J1 is constructed i , and its corresponding first rate distortion optimization model is as follows:

[0104] maxJ1 i =PSNR i ((r x ,r y ,r z ))-λ1·R1 i (Mode)(1)

[0105] Among them, J1 i represents the cost of the i-th node, PSNR i Indicates the geometric reconstruction quality of the i-th node in the corresponding coding mode, R1 i Indicates the code stream size of the i-th node in the corresponding coding mode, (r x ,r y ,r z ) represents the prediction residual of the i-th node, λ1 represents the Lagrangian factor, k represents the corresponding coding mode index, and Mode represents the corresponding coding mode.

[0106] Furthermore, this embodiment uses point-to-point PSNR to calculate the cost factor, wherein the point-to-point PSNR calculation process is as follows:

[0107] Calculate the geometric distortion Dist between the reconstructed point of the current node and the original point, which is:

[0108] Dist=||r x ||+||r y ||+||r z ||;

[0109] Using the obtained geometric distortion measure sum DistSum, calculate the geometric mean distortion measure corresponding to the current node. The calculation formula is:

[0110] DistAve=(Dist+DistSum) / PointCount;

[0111] Among them, pointCount represents the number of points that have been reconstructed.

[0112] Then the point-to-point PSNR of the current node in this mode is:

[0113]

[0114] Among them, PeakValue represents the peak value of the noise.

[0115] The cost of each coding mode of the current node is obtained according to the geometric reconstruction quality of the current node in the mode and the geometric code stream R in the corresponding mode.

[0116] On the one hand, considering the balance between the quality of the reconstructed point cloud and the compressed bit stream of the geometric prediction coding, the coding mode corresponding to the maximum value of the first cost factor can be directly selected as the optimal coding mode for the current node.

[0117] On the other hand, since the geometric reconstruction value of the current node to be decoded cannot be obtained at the decoding end, the above-mentioned geometric distortion measurement algorithm cannot be directly used. Based on this, this embodiment introduces a first average cost factor to select the encoding method, which is as follows:

[0118] First, the first cost factor of each completed reconstructed point is calculated, and the first average cost factor of the completed reconstructed point is obtained, which is expressed as R-PSNR. The first preset threshold is set to Threshold1. If the first average cost factor is less than the threshold Threshold1, DirectMode is used for geometric reconstruction; otherwise, RegularMode is used for geometric reconstruction.

[0119] Then, the R-PSNR cost of the current node is calculated according to different modes, and the first average cost factor is updated to guide the adaptive mode selection of subsequent nodes.

[0120] In this embodiment, when the encoding switch is turned on, the best encoding mode for the current point can also be determined according to the number of reconstructed points.

[0121] Specifically, when the number of reconstruction points of the KD-Tree prediction tree is less than a preset first value, mode one is selected as the optimal encoding mode for the current node; when the number of reconstruction points is greater than or equal to the preset first value, mode two is selected as the optimal encoding mode for the current node.

[0122] Specifically, in the KD-Tree prediction tree structure, the DirectMode encoding mode is first used for encoding. As the KD-Tree structure continues to improve, after the number of reconstructed points reaches a certain threshold, the neighbor search for each point will be more accurate, and the predicted neighbors found for each point will also be more accurate. Based on this, for each prediction tree, a first quantity value ProcessedCnt1 is introduced to record that after each prediction tree completes the reconstruction of ProcessedCnt1 to a certain point, RegularMode is used for geometric reconstruction. This can reasonably reduce the encoding complexity and reduce the bit stream. At the same time, it can also reduce the complex structure of the KD-Tree established at the decoding end and improve the encoding efficiency.

[0123] In this embodiment, when the encoding switch is turned on, mode 2 can also be directly used as the optimal encoding mode at the current point.

[0124] Specifically, RegularMode may be used directly to replace DirectMode without performing mode selection, which can further improve the quality of geometric reconstruction and ensure geometric coding efficiency.

[0125] In another embodiment of the invention, before step S2, the method further comprises:

[0126] The original point cloud data is analyzed and divided into the first category of data and the second category of data.

[0127] Accordingly, determining the best encoding mode for each point also includes:

[0128] Mode 1 is used as the optimal encoding mode for each point in the first type of data, and mode 2 is used as the optimal encoding mode for each point in the second type of data.

[0129] Specifically, by analyzing and classifying the original point cloud at the encoding end, it can be divided into the first type of data suitable for DirectMode encoding and the second type of data suitable for RegularMode encoding, and then the prediction trees are established and encoded respectively. When the encoding switch is turned on, the DirectMode encoding mode is used for each point in the first type of data, and the RegularMode encoding mode is used for each point in the second type of data.

[0130] S4: Encode the prediction residual to obtain a geometric information code stream.

[0131] Specifically, when the DirectMode encoding mode is used for encoding, all information to be encoded including the prediction residual is encoded to obtain a geometric information code stream;

[0132] When the RegularMode coding mode is used for encoding, the prediction residual (r x ,r y ,r z ) is set to zero, all the information to be encoded is encoded to obtain the geometric information code stream.

[0133] In this embodiment, after the prediction residual is set to zero, only the coding mode may be encoded (for example, mode 1 may be encoded as 0 and mode 2 may be encoded as 1); the prediction residual after being set to zero may also be encoded, that is, zero may be directly encoded.

[0134] The adaptive point cloud geometry prediction coding method provided by the present invention selectively sets the geometry prediction residual to zero according to certain rules, and does not perform coding, thereby reducing the geometry code stream while ensuring a certain geometry distortion, thereby improving the geometry coding efficiency. In addition, the present invention further improves the geometry coding efficiency by setting different coding modes and adaptively selecting the optimal coding mode for each point.

[0135] Embodiment 2

[0136] For step S2 in the first embodiment, a prediction tree may be established based on the laser radar calibration information, and then prediction is performed to obtain a prediction residual.

[0137] Specifically, since the voxelized point cloud is reordered by sorting by azimuth when establishing the prediction tree structure according to the lidar calibration information, it is necessary to convert the Cartesian coordinates (x, y, z) of each point into polar coordinates (r, φ, tanθ) and sort each point in turn according to the azimuth φ, radius r and elevation tangent tanθ.

[0138] Please continue to see Figure 2 Before establishing the prediction tree structure based on the lidar calibration information, coordinate conversion is required. Specifically, the point is converted from Cartesian coordinates (x, y, z) to cylindrical coordinates (r, φ, i). The conversion formula is as follows:

[0139]

[0140]

[0141]

[0142] Among them, i is the laser scanner ID corresponding to the point (a typical lidar system may have 16, 32 or 64 laser scanners, and the prior information of each laser scanner is different, that is, the elevation angle θ and the height zLaser in the vertical direction are different).

[0143] For each laser scanner of the lidar, the collection points belonging to the same laser scanner should be regularly distributed in the cylindrical coordinate system. However, due to factors such as noise, measurement errors and equipment jitter, the actual data presents an uneven distribution, resulting in poor correlation between data, low prediction accuracy and coding efficiency.

[0144] In this embodiment, i is determined by looping the prior information of different laser scanners. In each loop, the z component of the point is calculated using the r, prior information and the above conversion formula of the point, and the deviation between the converted z component and the original z component of the point is calculated. Then, the point with the smallest deviation is selected from different laser scanner IDs as the i of the point. This process processes the non-uniform distribution of the point cloud in the vertical direction in space and regularizes it to improve the correlation between the data, thereby improving the coding efficiency.

[0145] Then, the cylindrical coordinates (r, φ, i) of the points determined in the previous step are used to establish a prediction tree. If a point with laser scanner ID i appears for the first time, this point is used as the root node of all points corresponding to laser scanner ID i; if not, this point is considered as a child node of the previously processed point with laser scanner ID i, and this process is repeated until all points are traversed. Finally, the root nodes of all laser scanner IDs are connected to complete the establishment of the prediction tree. See Figure 3 , Figure 3 It is a schematic diagram of the prediction tree structure of four laser scanners provided in an embodiment of the present invention.

[0146] Next, the prediction mode is selected. The prediction mode selection method in this embodiment is the same as the mode selection method based on the KD-Tree prediction tree in the above embodiment, and will not be described in detail here.

[0147] Finally, prediction is made for each node according to the corresponding prediction mode.

[0148] In this embodiment, the geometric prediction of the prediction tree node includes cylindrical coordinate prediction and Cartesian coordinate prediction, and the specific process is as follows:

[0149] If the current node is the root node of the prediction tree, the prediction mode is Mode0, that is, the cylindrical coordinates of the current node are not predicted, and the corresponding cylindrical coordinate prediction value is where r min The minimum value of the r component after the coordinate transformation of the entire point cloud. If the current node has no parent node, Set to 0, otherwise it is the cylindrical coordinate component of the parent node. The Cartesian coordinate prediction value is the original cylindrical coordinate of the point The inverse transformation

[0150] If the current node is not the root node and the prediction mode is Mode1, the cylindrical coordinates of the current point are obtained by multiplying the cylindrical coordinates of its parent node. The prediction can be made to get the predicted value of the cylindrical coordinates of the current point The Cartesian coordinates are predicted to be the original cylindrical coordinates. The inverse transformation

[0151] If the current node is not the root node, and the prediction mode is Mode2 or Mode3, the cylindrical coordinates of the current point are predicted using the corresponding prediction method, and the predicted value of the cylindrical coordinates of the current point is Likewise, the Cartesian coordinates predicted are the original cylindrical coordinates The inverse transformation

[0152] Using the best prediction mode of the current node, the cylindrical coordinates of the current point Make a prediction and get the corresponding prediction residual (r r ,r φ ,r i ), using the Cartesian coordinates (x, y, z) of the current point and the predicted Cartesian coordinates Perform differential prediction to obtain the prediction residual (r x ,r y ,r z ).

[0153] Among them, the predicted value of the azimuth of the current point The calculation is done according to the following formula:

[0154]

[0155] in, Indicates the predicted azimuth of the current point; n indicates the number of points that need to be skipped between the parent node and the current point according to the scanning speed, and this parameter also needs to be compressed into the bitstream; δ φ Indicates the scanning speed of the LiDAR.

[0156] At this point, the cylindrical coordinate prediction value of the current node is obtained and Cartesian coordinates predicted values and the corresponding prediction residual (r r ,r φ ,r i ) and (r x ,r y ,r z ).

[0157] The above geometric prediction values ​​and their corresponding prediction residuals, as well as the prediction mode adopted by the current point, the quantization parameter of the current point, the number of repeated points and the number of child nodes, the number of skipped points corresponding to each mode and other information are taken together as the information to be encoded.

[0158] Then, the optimal encoding mode for each point in the prediction tree based on the lidar calibration information is determined.

[0159] Specifically, similar to the first embodiment, this embodiment can construct a second cost factor according to the distortion measurement of the current laser scanner and the code stream size under different coding modes to determine the coding mode when the coding switch is turned on.

[0160] Then, the encoding mode corresponding to the minimum value of the second cost factor is selected as the optimal encoding mode for the current point; or the second average cost factor is calculated based on the second cost factor of the encoded point, and when the second average cost factor is less than the second preset threshold, mode one is selected as the optimal encoding mode for the current point, and when the second average cost factor is greater than or equal to the second preset threshold, mode two is selected as the optimal encoding mode for the current point.

[0161] Specifically, the second cost factor J2 can be constructed i , and its corresponding second rate-distortion optimization model is as follows:

[0162]

[0163] Among them, J2 i represents the RD cost of the i-th laser scanner, D i represents the distortion generated when the i-th laser scanner performs point cloud geometry encoding, R2 i It represents the code stream size generated when the i-th laser scanner performs point cloud geometry encoding, λ2 represents the Lagrangian factor, k represents the corresponding encoding mode index, and Mode represents the corresponding encoding mode.

[0164] In this embodiment, D1 (Point-to-Point) is used as the distortion measurement standard. Assuming that the original Cartesian coordinates of the current node are (x, y, z), and the reconstructed Cartesian coordinates of the corresponding mode of the current node are (x', y', z'), the point-to-point D1 calculation method is as follows:

[0165] D1=||x-x'||+||y-y'||+||z-z'||.

[0166] The cost of each coding mode of the current node is calculated by formula (2), and finally the coding mode corresponding to the minimum value of the second cost factor is selected as the optimal coding mode of the current node.

[0167] In addition, when performing coordinate conversion, you can also convert a point from Cartesian coordinates (x, y, z) to cylindrical coordinates (r, θ, φ) according to the following formula.

[0168]

[0169]

[0170]

[0171] When constructing a 2D regular structure, firstly, according to the number of laser scanners in the laser radar, the point cloud converted to the cylindrical coordinate system is clustered according to its pitch angle, so as to obtain the point cloud information belonging to each laser scanner, and then according to the clustering result, a "line" structure is constructed at the vertical sampling position corresponding to each laser scanner and the ideal points are set at equal intervals according to the sampling rate of the laser radar, so as to construct a regular structure. Since the cylindrical coordinate system can be expanded to two dimensions, it is equivalent to constructing a two-dimensional regular structure. Figure 4 shown.

[0172] In addition, after the prediction tree structure of the adjustment level is regularized, the r value of the adjusted node can be obtained by using the nearest neighbor interpolation calculation to further improve the geometric reconstruction quality and ensure the geometric coding efficiency.

[0173] When the input point cloud selects RegularMode, the point "○" is the original position (r, θ, φ) of the point in the cylindrical coordinate system. It can be seen that it exhibits the characteristics of non-uniform distribution. Therefore, when encoding, the ideal position (r0, θ0, φ0) in the 2D regular structure shown by the point "●" can be used instead. This process is equivalent to "moving" the input point cloud to its corresponding position in the two-dimensional regular structure. Finally, the regularized point cloud exhibits the characteristics of uniform distribution on the two-dimensional plane (θ, φ) of the cylindrical coordinate system. Therefore, there is no need to introduce auxiliary information in the horizontal direction during encoding, which can effectively improve the encoding efficiency, but the distortion caused by the movement of the point position will follow.

[0174] In this embodiment, the DirectMode mode predicts the cylindrical coordinates to obtain the prediction residual, encodes the prediction residual, and then inversely transforms the cylindrical coordinates to obtain the reconstructed Cartesian coordinates, and finally performs differential prediction on the reconstructed coordinates and encodes the differential prediction of the Cartesian coordinates. The RegularMode mode first regularizes the cylindrical coordinates of each node, that is, the (r r ,r φ ,r i ) in the horizontal prediction residual component r φNo encoding is performed, and the prediction residual of the Cartesian coordinates obtained after the inverse transformation of the cylindrical coordinates and the Cartesian coordinates of the current node is not encoded. A certain amount of distortion is introduced in exchange for improved coding efficiency. Specifically, by introducing a rate-distortion optimization model, the coding mode of each node is adaptively selected, and finally the optimal coding mode is used for geometric prediction. Among them, the introduction of the rate-distortion optimization model specifically includes the construction of the RD cost, the setting of the distortion measure D, and the selection of λ.

[0175] The construction of the RD cost is the same as the second cost function mentioned above, and the distortion measure based on D1 (Point-to-Point) is calculated using the following method:

[0176] Assume that the current node has multiple nearest neighbors in the original point cloud (x i ,y i ,z i ) i＝1...N , the reconstructed Cartesian coordinates of each mode of the current node are (x', y', z'), then

[0177]

[0178] By calculating the second cost factor of the corresponding mode selected by the current node, and obtaining the second average cost factor of the reconstructed point, the second preset threshold is set to Threshold2. If the second average cost factor is less than the threshold Threshold2, RegularMode is used for geometric reconstruction; otherwise, DirectMode is used for geometric reconstruction.

[0179] Then, the average cost factor is updated to guide the subsequent node mode selection.

[0180] In addition, for the prediction tree based on lidar calibration information, the optimal encoding mode of the current node can also be determined according to the number of reconstruction points.

[0181] When the number of reconstructed points of each laser scanner is less than a preset second value, mode one is selected as the optimal encoding mode for the current point; when the number of reconstructed points is greater than or equal to the preset second value, mode two is selected as the optimal encoding mode for the current point.

[0182] Specifically, in the prediction tree structure based on the laser radar calibration information, the DirectMode encoding mode is first used for encoding. With the establishment of the prediction tree structure, a better geometric prediction value can be found for each point later. Based on this, a second quantity value ProcessedCnt2 can be introduced for each laser scanner. After reconstructing ProcessedCnt2 points, RegularMode is uniformly used. This can reasonably control the node encoding method of each laser scanner, and at the same time, it can also save the complexity of establishing a prediction tree structure at the decoding end to calculate the quality of geometric reconstruction, thereby reducing the bit stream.

[0183] In this embodiment, when the encoding switch is turned on, mode 2 can also be directly used as the optimal encoding mode at the current point.

[0184] Finally, the determined encoding mode is used for encoding.

[0185] Specifically, when the DirectMode encoding mode is used for encoding, all information to be encoded including the prediction residual is encoded to obtain a geometric information code stream;

[0186] When the RegularMode coding mode is used for encoding, the prediction residual r φ and (r x ,r y ,r z ) is set to zero, the remaining information to be encoded is encoded to obtain a geometric information code stream.

[0187] Embodiment 3

[0188] Based on the above-mentioned embodiments 1 and 2, this embodiment provides an adaptive point cloud geometry encoding device. Figure 5 , Figure 5 : is a schematic diagram of the structure of an adaptive point cloud geometry encoding device provided by an embodiment of the present invention, comprising:

[0189] A first data acquisition module 11, used to acquire original point cloud data;

[0190] A first prediction module 12 is used to perform geometric prediction on each point in the original point cloud data to obtain a prediction residual;

[0191] A residual zeroing module 13, used for selectively setting the prediction residual to zero according to a preset rule;

[0192] The encoding module 14 is used to encode the prediction residual to obtain a geometric information code stream.

[0193] The device provided in this embodiment can implement the encoding method provided in the above-mentioned embodiment 1 or embodiment 2, and the specific implementation process will not be repeated here.

[0194] Embodiment 4

[0195] See also Figure 6 , Figure 6 Schematic diagram of an adaptive point cloud geometry decoding method provided by an embodiment of the present invention, comprising:

[0196] Step 1: Get the geometric information code stream and decode it to get the parsed data.

[0197] Specifically, the parsed data may include information such as the prediction mode of the current point, the encoding mode of the current point, and the prediction residual of the current point.

[0198] Step 2: Perform geometric prediction on the current point to obtain the geometric prediction value;

[0199] In this embodiment, the current point may be predicted according to the prediction mode obtained by analysis to obtain a predicted value.

[0200] Step 3: Determine whether the prediction residual of the current point is set to zero based on the analytical data, and obtain the prediction residual of the current point.

[0201] Specifically, this embodiment can obtain whether the current point is set to zero in the encoding stage according to the encoding mode of the current point. When the encoding mode is mode one, it means that the residual is not set to zero in the encoding stage, which means that the currently parsed prediction residual is the prediction residual of the current point. When the encoding mode is mode two, it means that the residual is set to zero in the encoding stage, then the prediction residual or part of the prediction residual of the current point is directly set to zero, and the remaining non-zero prediction residuals are parsed.

[0202] Step 4: Reconstruct the geometric point cloud according to the geometric prediction value and prediction residual of the current point to obtain the reconstructed point cloud data.

[0203] For details, see Figure 7 , Figure 7 is a block diagram of adaptive point cloud geometry decoding provided by an embodiment of the present invention;

[0204] In this embodiment, for establishing a prediction tree structure based on KD-tree, the decoding and reconstruction point cloud process is as follows:

[0205] According to the prediction mode adopted by the current point obtained by decoding and the reconstructed geometric coordinates of its ancestors, the geometric coordinates of the current point are predicted and the corresponding prediction values ​​are generated. The specific process is as follows:

[0206] If the current node is the root node of the prediction tree, the prediction mode selects Mode0, that is, there is no prediction for the cylindrical coordinates of the current node, and the corresponding Cartesian coordinate prediction value is (0,0,0).

[0207] If the current node is not the root node and the prediction mode is Mode1, the Cartesian coordinates of the current point are reconstructed through the geometry of its parent node. Make predictions;

[0208] If the current node is not the root node and the prediction mode is Mode2 or Mode3, the Cartesian coordinates of the current point are predicted using the corresponding prediction method. The predicted value of the Cartesian coordinates of the current node is

[0209] Then parse to get the encoding mode of the current node. If the encoding mode of the current node is DirectMode, it is necessary to parse the prediction residual (r x ,r y ,r z ); otherwise, there is no need to parse the prediction residual of the current node, let the geometric prediction residual of the current node (r x ,r y ,r z ) is zero.

[0210] Finally, the Cartesian coordinate residual (r x ,r y ,r z ) and the predicted Cartesian coordinates of the current point Computes the reconstructed Cartesian coordinates (x,y,z) of the current point.

[0211]

[0212] Finally, the reconstructed voxelized point cloud is obtained.

[0213] In another embodiment of the present invention, for establishing a prediction tree structure based on KD-tree, the decoding and reconstruction point cloud process may also be:

[0214] First, the prediction mode of the input geometric bitstream is decoded to obtain the prediction mode of each node; then, the geometric prediction value of the current node is obtained based on the prediction mode and the geometric prediction structure, and the number of child nodes of the current node is parsed to reconstruct the prediction tree structure; secondly, the average cost factor and threshold of the reconstructed point are used to determine the encoding mode of the current node. If the encoding mode of the current node is DirectMode, the prediction residual of the current node needs to be parsed; otherwise, the prediction residual of the current node does not need to be decoded. After completing the geometric reconstruction of the current node, the cost factor of the corresponding mode of the current node needs to be calculated to update the global average cost factor. Finally, the prediction value and prediction residual of the current node are used to reconstruct the geometric position information of the current node, and the geometric reconstruction of the decoding end is completed in sequence and iteratively.

[0215] For the prediction tree coding structure based on lidar information, the decoding and reconstruction point cloud process can be:

[0216] First, the cylindrical coordinates of the current point are predicted according to the prediction mode adopted by the decoded current point and the reconstructed cylindrical coordinates of its ancestors, and the corresponding prediction values ​​are generated. The specific process is as follows:

[0217] If the current node is the root node of the prediction tree, the prediction mode is Mode0, that is, the cylindrical coordinates of the current node are not predicted, and the corresponding cylindrical coordinate prediction value is where r min The minimum value of the r component after the coordinate transformation of the entire point cloud. If the current node has no parent node, then Set to 0, otherwise it is the cylindrical coordinate component of the parent node. The Cartesian coordinate prediction value is the original cylindrical coordinate of the point The inverse transformation

[0218] If the current node is not the root node and the prediction mode is Mode1, the cylindrical coordinates of the current point are obtained by multiplying the cylindrical coordinates of its parent node. The prediction can be made to get the predicted value of the cylindrical coordinates of the current point The Cartesian coordinates are predicted to be the original cylindrical coordinates. The inverse transformation

[0219] If the current node is not the root node, and the prediction mode is Mode2 or Mode3, the cylindrical coordinates of the current point are predicted using the corresponding prediction method, and the predicted value of the cylindrical coordinates of the current point is Likewise, the Cartesian coordinates predicted are the original cylindrical coordinates The inverse transformation

[0220] Using the best prediction mode of the current node, the cylindrical coordinates of the current point Make a prediction and get the corresponding prediction residual (r r ,r φ ,r i ), using the Cartesian coordinates (x, y, z) of the current point and the predicted Cartesian coordinates Perform differential prediction to obtain the prediction residual (r x ,r y ,r z ).

[0221] Note that the predicted value of the azimuth of the current point The calculation is done according to the following formula:

[0222]

[0223] in, Indicates the predicted azimuth of the current point; n indicates the number of points that need to be skipped between the parent node and the current point according to the scanning speed, and this parameter also needs to be compressed into the bitstream; δ φ Indicates the scanning speed of the LiDAR.

[0224] Then parse to get the encoding mode of the current node. If the encoding mode of the current node is DirectMode, it is necessary to parse the prediction residual (r x ,r y ,r z ) and (r r ,r φ ,r i ); otherwise, there is no need to parse the prediction residual r of the current node φ and (r x ,r y ,r z ), only the current node's (r r ,r i ), and let the geometric prediction residual of the current node (r x ,r y ,r z ) and r φ is zero.

[0225] Next, the predicted cylindrical coordinates of the current point can be obtained through the previous step Next, the cylindrical coordinate residual (r r ,r φ ,r i ) and the predicted cylindrical coordinates of the current point Compute the reconstructed cylindrical coordinates (r, φ, i) of the current point.

[0226]

[0227] After obtaining the reconstructed cylindrical coordinates of the current point, the position of the current point in the prediction tree can be further determined according to the reconstructed cylindrical coordinates (r, φ, i), thereby reconstructing the prediction tree.

[0228] Then convert the reconstructed cylindrical coordinates (r, φ, i) of the current point into Cartesian coordinates according to the following formula: This is the predicted Cartesian coordinate of the current point.

[0229]

[0230]

[0231]

[0232] Among them, i is the laser scanner ID corresponding to the point. The prior information of each laser scanner is different, that is, the elevation angle θ and the height zLaser in the vertical direction are different. Therefore, the elevation angle corresponding to the i-th laser scanner is θ(i), and the height in the vertical direction is zLaser(i).

[0233] Finally, the Cartesian coordinate residual (r x ,r y ,r z ) and the predicted Cartesian coordinates of the current point Computes the reconstructed Cartesian coordinates (x,y,z) of the current point.

[0234]

[0235] Finally, the reconstructed voxelized point cloud is obtained.

[0236] In another embodiment of the present invention, for the prediction tree coding structure based on laser radar information, the decoding and reconstruction point cloud process can also be:

[0237] First, the prediction mode of the input geometry bitstream is decoded to obtain the prediction mode of each node; then, the geometry prediction value (cylindrical coordinates) of the current node is obtained based on the prediction mode and the geometry prediction structure, and the number of child nodes of the current node is parsed to reconstruct the prediction tree structure; secondly, according to the average cost factor of the reconstructed points before the current node, when the average cost factor meets a certain threshold Threshold, the RegularMod mode is used for decoding, otherwise the DirectMode mode is used for decoding the current node. If the encoding mode of the current node is DirectMode, it is necessary to parse the prediction residual of the cylindrical coordinates of the current node and the prediction residual of the Cartesian coordinates; otherwise, it is not necessary to parse the horizontal component r of the cylindrical coordinates of the current node φ The prediction residual of the Cartesian coordinates does not need to be parsed. Finally, the cylindrical coordinates of the current node are used to perform an inverse transformation to obtain the predicted geometry of the Cartesian coordinates, and the Cartesian prediction residual is used to reconstruct the geometric position information of the current node, and the geometric reconstruction of the decoding end is completed in sequence and iteratively. Specifically, as follows:

[0238] If the encoding mode obtained by decoding is DirectMode, the above method is used to perform geometric reconstruction.

[0239] If the encoding mode obtained by decoding is RegularMode, geometric reconstruction is performed according to the following method. The specific process is as follows:

[0240] First, the geometric prediction residual (cylindrical coordinates) of the current node is obtained by using the prediction mode and geometric prediction structure corresponding to the current node. φ No decoding is performed;

[0241] The coordinate inverse transformation is to inversely transform the reconstructed cylindrical coordinates of the point into Cartesian coordinates. Specifically, the reconstructed cylindrical coordinates of the point (r0, θ0, φ0) are inversely transformed into Cartesian coordinates using the following formula: Get the reconstructed lossy voxelized point cloud.

[0242]

[0243]

[0244]

[0245] The reconstructed coordinates after coordinate reversal are the geometric reconstruction values ​​of the current node, and there is no need to decode the geometric prediction residuals of the current node in Cartesian coordinates.

[0246] Embodiment 5

[0247] Based on the above-mentioned fourth embodiment, this embodiment provides an adaptive point cloud geometry prediction decoding device. Figure 8 , Figure 8 : is a schematic diagram of the structure of an adaptive point cloud geometry decoding device provided by an embodiment of the present invention, comprising:

[0248] The second data acquisition module 21 is used to acquire the geometric information code stream and decode it to obtain parsed data;

[0249] The second prediction module 22 is used to perform geometric prediction on the current point to obtain a geometric prediction value;

[0250] A residual analysis module 23, used to determine whether the prediction residual of the current point is set to zero according to the analysis data, and obtain the prediction residual of the current point;

[0251] The reconstruction module 24 is used to reconstruct the geometric point cloud according to the geometric prediction value and the prediction residual of the current point to obtain the reconstructed point cloud data.

[0252] The device provided in this embodiment can implement the decoding method provided in the fifth embodiment above, and the specific implementation process will not be repeated here.

[0253] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

[0254] Appendix 1: Geometry parameter set syntax

[0255]

[0256]

[0257]

[0258] Appendix 2: Tile inventory syntax

[0259]

[0260]

[0261] Appendix 3: Geometry data unit header syntax

[0262]

[0263]

Claims

1. An adaptive point cloud geometry encoding method, characterized in that: include: Get the original point cloud data; Performing geometric prediction on each point in the original point cloud data to obtain a prediction residual; Selectively setting the prediction residual to zero according to a preset rule; Encoding the prediction residual to obtain a geometric information code stream; The step of selectively setting the prediction residual to zero according to a preset rule specifically includes: Setting a coding mode and a coding switch, wherein the coding mode includes mode 1 and mode 2; Determine the best encoding mode for each point according to the encoding switch, if the encoding switch is turned on, select the best encoding mode from the mode 1 and the mode 2 according to a preset rule, otherwise, use the mode 1 as the best encoding mode for the current point; If mode 1 is selected for encoding, the prediction residual is not set to zero; if mode 2 is selected for encoding, the prediction residual is set to zero or part of the prediction residual is set to zero.

2. The adaptive point cloud geometry encoding method according to claim 1, characterized in that: If the encoding switch is turned on, the best encoding mode is selected from the mode 1 and the mode 2, including: Determine, according to rate-distortion optimization, that the mode 1 or the mode 2 is the best encoding mode for the current point; or Determine, according to the number of reconstruction points, that the mode 1 or the mode 2 is the best encoding mode for the current point; or Mode 2 is directly used as the optimal encoding mode for the current point.

3. The adaptive point cloud geometry encoding method according to claim 2, characterized in that: Determine the best encoding mode at the current point based on rate-distortion optimization, including: Constructing a rate-distortion cost factor according to the prediction residual of the current point and the bitstream size under different coding modes; Selecting the coding mode corresponding to the optimal rate-distortion cost factor as the optimal coding mode for the current point; or An average cost factor is calculated according to the rate-distortion cost factor of the reconstructed point, and an optimal encoding mode for the current point is selected according to the average cost factor and a preset threshold.

4. The adaptive point cloud geometry encoding method according to claim 2, characterized in that: Determine the best encoding mode for the current point based on the number of reconstructed points, including: When the number of reconstructed points is less than a preset value, mode 1 is selected as the best encoding mode for the current point; when the number of reconstructed points is greater than or equal to the preset value, mode 2 is selected as the best encoding mode for the current point.

5. The adaptive point cloud geometry encoding method according to claim 1, characterized in that: Before performing geometric prediction on each point in the original point cloud data to obtain the prediction residual, the method further includes: The original point cloud data is analyzed and divided into a first type of data and a second type of data.

6. The adaptive point cloud geometry encoding method according to claim 5, characterized in that: Determine the best encoding mode for each point, including: Mode 1 is used as the optimal encoding mode for each point in the first category of data, and mode 2 is used as the optimal encoding mode for each point in the second category of data.

7. An adaptive point cloud geometry encoding device, used to implement the encoding method according to any one of claims 1 to 6, characterized in that: include: A first data acquisition module (11), used for acquiring original point cloud data; A first prediction module (12) is used to perform geometric prediction on each point in the original point cloud data to obtain a prediction residual; A residual zeroing module (13), used for selectively setting the prediction residual to zero according to a preset rule; The encoding module (14) is used to encode the prediction residual to obtain a geometric information code stream.

8. An adaptive point cloud geometry decoding method, characterized in that: include: Get the geometric information code stream and decode it to get the parsed data; Perform geometric prediction on the current point to obtain the geometric prediction value; Determine whether the prediction residual of the current point is set to zero according to the analytical data, and obtain the prediction residual of the current point; Reconstructing a geometric point cloud according to the geometric prediction value and the prediction residual of the current point to obtain reconstructed point cloud data; Wherein, determining whether the prediction residual of the current point is set to zero according to the analytical data to obtain the prediction residual of the current point comprises: Determine the encoding mode of the current point at the encoding end according to the state of the encoding switch. If the encoding switch is turned on, determine whether the encoding mode of the current point is mode 1 or mode 2 according to the encoding mode of the current point in the parsed data; otherwise, directly determine that the encoding mode of the current point is mode 1; If the encoding mode of the current point is mode one, the prediction residual of the current point is not set to zero, and the currently parsed prediction residual is directly determined as the prediction residual of the current point; if the encoding mode of the current point is mode two, the prediction residual of the current point is set to zero or partially set to zero, and the prediction residual or part of the prediction residual of the current point is directly set to zero, and the remaining non-zero prediction residuals are parsed to obtain the prediction residual of the current point.

9. An adaptive point cloud geometry decoding device, used to implement the decoding method of claim 8, characterized in that: include: A second data acquisition module (21) is used to acquire a geometric information code stream and decode it to obtain parsed data; A second prediction module (22) is used to perform geometric prediction on the current point to obtain a geometric prediction value; A residual analysis module (23), used for determining whether the prediction residual of the current point is set to zero according to the analysis data, and obtaining the prediction residual of the current point; A reconstruction module (24) is used to reconstruct a geometric point cloud according to the geometric prediction value and the prediction residual of the current point to obtain reconstructed point cloud data.

Images