A point cloud geometry inter-frame prediction method based on block motion estimation and motion compensation

Abstract

The present invention discloses a point cloud geometry inter-frame prediction method based on block motion estimation and motion compensation, comprising the following steps: 1) In order to ensure that the nodes of the frame to be encoded and the reference frame have a one-to-one correspondence, the same bounding box needs to be set for the frame to be encoded and the reference frame as the root node of the tree partition. After determining the bounding box, the frame to be encoded and the reference frame are synchronously divided into trees; 2) The size of the prediction starting unit PTU is set. If the current node size is equal to the set prediction block size, the current block is considered to be a prediction unit. A matching block is searched for the current node within a specific window range in the reference frame. The matching block with the smallest error is the obtained prediction block; 3) The occupancy information of the prediction block is used as context to help encode the occupancy information of the current block, thereby improving the efficiency of entropy coding.

Description

Technical Field

[0001] The present invention belongs to the technical field of point cloud data processing, and in particular relates to a point cloud geometry inter-frame prediction method based on block motion estimation and motion compensation. Background Art

[0002] Inter-frame prediction of point clouds has always been a technical challenge in point cloud encoding and decoding. Because point clouds lack a fixed spatial structure, it's difficult to intuitively obtain a regular block structure like with traditional video encoding and decoding. Furthermore, the sparsity and flexibility of point cloud spatial distribution also complicates block-based motion estimation and compensation.

[0003] The existing point cloud inter-frame prediction frameworks mainly include the following:

[0004] Prediction method based on occupancy code XOR operation: This method performs an XOR operation on the occupancy codes of the current frame and the predicted frame to obtain the geometric information residual of the two frames, and then encodes the residual. However, this method cannot capture the motion information of the point cloud object in time sequence. When the object moves with a large amplitude, the occupancy code residual obtained is also large.

[0005] Mapping-based prediction method: This method first maps the 3D point cloud onto a 2D plane at multiple angles to obtain multiple 2D images, which are then compressed using existing video encoding tools. This method is highly dependent on the quality of the mapping method, and this type of mapping-based compression method is mainly aimed at 3D human object surface scanning datasets. For sparse datasets such as lidar scanned maps, the inter-frame prediction effect is limited.

[0006] Inter-frame prediction method based on block partitioning: This method first directly divides the point cloud into blocks, then searches for the matching block with the smallest error from the reference frame as the prediction block, and uses the prediction block to improve the coding efficiency of the current block. There are currently two ways to use the prediction block. One is to directly replace the current block with the prediction block, calculate the prediction residual, and finally encode the residual coefficient. However, due to the sparsity of the spatial distribution of the point cloud, the residual coefficient of the prediction block is often still large. If the residual is not encoded to save codewords, lossless compression cannot be achieved. The second method is to use the geometric occupancy information of the prediction block as context to improve the entropy coding efficiency of the current block. The exploration platform EM13 proposed by the MPEG Point Cloud Task Force currently uses this prediction method, but it only supports octree partitioning of cubic bounding boxes. Therefore, the set bounding box range is too large, resulting in the introduction of too many empty blocks, and there are also defects such as high motion estimation complexity. Summary of the Invention

[0007] To overcome the shortcomings of the above-mentioned prior art and further improve the compression performance of point cloud attributes while taking into account computational complexity, the present invention provides a point cloud geometry inter-frame prediction method based on block motion estimation and motion compensation. This method, combined with point cloud quadtree and binary tree partitioning techniques, supports block partitioning of non-cubic bounding boxes, significantly reducing the codeword waste caused by empty blocks. It also combines the KD tree and iterative closest point algorithm for block matching calculations, significantly reducing the computational time complexity of motion estimation while ensuring compression rate. Furthermore, the present invention simplifies the inter-frame context model and, by integrating it with the intra-frame context model, significantly improves entropy coding efficiency without increasing the amount of context.

[0008] The present invention provides a point cloud geometry inter-frame prediction method based on block motion estimation and motion compensation, comprising the following steps: 1) To ensure a one-to-one correspondence between the nodes of the frame to be coded and the reference frame, the same bounding box is set for the frame to be coded and the reference frame as the root node of the tree partitioning. After determining the bounding box, the frame to be coded and the reference frame are synchronously partitioned into trees; 2) The size of the prediction starting unit PTU is set. If the current node size is equal to the set prediction block size, the current block is considered to be a prediction unit. A matching block is searched for the current node within a specific window range in the reference frame. The matching block with the smallest error is the obtained prediction block; 3) The occupancy information of the prediction block is used as context to help encode the occupancy information of the current block, thereby improving the efficiency of entropy coding.

[0009] Preferably, in the above-mentioned cloud geometry inter-frame prediction method, the calculation method of the reference frame and the current frame bounding box in step 1) includes but is not limited to the following two schemes: the first method is to count all frames in the entire sequence, and calculate the minimum bounding box that can contain all frames as the unified bounding box of the entire current sequence; the second method is to calculate the bounding box size only in the reference frame and the current frame, and take the maximum bounding box size as the root node of the tree division in the current frame encoding process.

[0010] Preferably, in the above-mentioned cloud geometry inter-frame prediction method, the tree partitioning operation in step 1) can be combined with octree, quadtree, binary tree and other partitioning methods according to different bounding box characteristics to adapt to non-cubic bounding boxes and reduce the proportion of empty blocks occupied.

[0011] In the cloud geometry inter-frame prediction method described above, in step 2), if the tree is divided into octree mode, the current node is a cube, and its side length is determined to be equal to the PTU size. If the tree is divided into quadtree or binary tree mode, the current node is a cuboid, and its shortest side length is determined to be equal to the PTU size, that is, the current block is considered to be the prediction starting unit.

[0012] Preferably, in the above-mentioned cloud geometry inter-frame prediction method, in step 2), for a prediction unit PU, we define a search window W, and the window size can be set according to different distribution characteristics of the data set and different bit rate points. If the point cloud distribution is relatively discrete, a larger window range is set; if the point cloud is relatively dense, a smaller window size can be set.

[0013] Preferably, in the above-mentioned cloud geometry inter-frame prediction method, in step 2), the matching block search process is to first find the most approximate matching block in the obtained window local reference point cloud using the ICP algorithm, that is, to minimize the Lagrangian cost. Here, the ICP algorithm is simplified accordingly, considering only translation transformation and ignoring rotation transformation. The translation vector obtained after ICP is the motion vector. The Lagrangian cost calculation formula is as follows:

[0014] Cost(MV)=Dist(Q(W,MV),B)+λEst(MV)

[0015] Where B is the current block to be encoded, W is the search window, Q is the set of nearest neighbor points found after B is translated by the motion vector MV, Est() estimates the codeword required to encode MV, which is obtained by calculating the codeword required for exponential Golomb coding, and Dist is the block matching loss function, which is as follows:

[0016]

[0017] The local point set corresponding to the MV with the minimum final coding cost is the optimal prediction block.

[0018] Preferably, in the above-mentioned cloud geometry inter-frame prediction method, in step 2), the above-mentioned search for matching blocks involves searching for the nearest neighbor points of each point in the block to be encoded within the window, and this search process is accelerated by establishing a KD tree.

[0019] Preferably, in the cloud geometry inter-frame prediction method described above, in step 3), the prediction block obtained in the above process is divided into sub-blocks synchronously with the current block to be encoded, and occupancy code information for the prediction block and the current block is obtained. For each sub-node, there are two types of inter-frame prediction context: 1) the sub-node at the corresponding position in the prediction block is occupied, and 2) the sub-node at the corresponding position in the prediction block is unoccupied. For each sub-node, an additional context bit is added to the original intra-frame context to indicate whether the corresponding position in the prediction block is occupied.

[0020] Preferably, in the above-mentioned cloud geometry inter-frame prediction method, in step 3), if the corresponding prediction sub-node is occupied, it is considered to be a strong prediction, and the prediction confidence is the same as that of the first 7 neighbor sub-nodes in the frame being occupied, and the occupancy codes of the first 7 sub-nodes in the Morton order in the corresponding intra-frame context mode [1] are all set to 1. If the corresponding prediction sub-node is not occupied, the original intra-frame context information is retained.

[0021] Preferably, in the above-mentioned cloud geometry inter-frame prediction method, in step 3), if the corresponding prediction sub-node is occupied, it is considered to be a strong prediction, and the prediction confidence is the same as that of the first three neighbor sub-nodes in the frame being occupied, and the occupancy codes of the first three sub-nodes in the Morton order in the corresponding intra-frame context mode [1] are all set to 1. If the corresponding prediction sub-node is not occupied, the original intra-frame context information is retained.

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] (1) For inter-frame prediction, non-cubic point cloud bounding boxes can be supported to further reduce the space occupied by empty blocks and reduce codeword waste.

[0024] (2) Using ICP for motion estimation makes MV not limited to a fixed search direction, and can more accurately find the optimal matching block while avoiding the complex motion vector search process.

[0025] (3) Inter-frame context methods 2 and 3 will not increase the total number of contexts and will not bring additional burden to hardware implementation. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the specific implementation or the description of the prior art.

[0027] Figure 1 It is an inter-frame prediction coding flow chart of the point cloud geometry inter-frame prediction method based on block motion estimation and motion compensation of the present invention.

[0028] Figure 2 This is a comparison result of the compression performance between the point cloud geometry inter-frame prediction method based on block motion estimation and motion compensation of the present invention and the existing method. DETAILED DESCRIPTION

[0029] The present invention will be further described below by way of examples in conjunction with the accompanying drawings, but the scope of the present invention is not limited in any way.

[0030] The block-partition-based point cloud geometry inter-frame prediction method of the present invention uses block motion estimation to capture point cloud temporal motion information and improve the compression performance of point cloud geometry information. For 3D point cloud geometry information compression, the spatial occupancy information of the previously encoded frame is used to predict the geometric occupancy information of the current block to be encoded, and the spatial occupancy information is used as context to improve the efficiency of occupancy code entropy coding, thereby improving the compression performance of point cloud geometry information. The method includes the following steps:

[0031] 1) Point cloud input:

[0032] Input the geometric information of the frame to be encoded and the frame to be encoded as a reference frame.

[0033] 2) Bounding box calculation:

[0034] In order to ensure that the nodes of the frame to be coded and the reference frame have a one-to-one correspondence, the same bounding box needs to be set for the frame to be coded and the reference frame as the root node of the tree division. After the bounding box is determined, the frame to be coded and the reference frame are synchronously divided into trees. Among them, the calculation method of the bounding box of the reference frame and the frame to be coded includes but is not limited to the following two schemes: the first method is to count all frames in the entire sequence and calculate the maximum bounding box size that can contain all frames as the unified bounding box of the entire sequence; the second method is to calculate the bounding box size only in the reference frame and the current frame, and take the maximum bounding box size as the root node of the tree division during the current frame coding process.

[0035] 3) Point cloud tree division:

[0036] The obtained point cloud bounding box space is divided into trees. According to the different shapes of the bounding box, the point cloud can be spatially decomposed using tree partitioning modes such as octree, quadtree and binary tree to obtain several coding blocks.

[0037] 4) Prediction unit division:

[0038] Similar to the concept of prediction blocks in video coding, we define a PTU (Prediction Tree Unit) as the starting prediction block. A PTU can be further divided into multiple PUs (Prediction Units) using an octree. Further division is determined based on coding cost, resulting in a PU tree structure. A PU is the basic prediction unit. Each PU calculates a motion vector (MV) to find the best matching prediction block in the reference frame.

[0039] In step 4), the PTU size is defined as PTU_size. The current point cloud is first divided into a tree. If the minimum edge length of the node is equal to PTU_size, the current node is regarded as a PTU. The PTU can be divided into a PU tree and multiple PUs through an octree. The PU tree contains two flag information. The split flags is used to indicate whether the nodes at each layer of the PU tree are further divided. If the division continues, the occupied flags is used to indicate the occupancy information of the divided sub-nodes. Each PU will calculate a motion vector MV, and the obtained PU tree and MV are encoded so that the decoder can also generate the prediction block and drive the arithmetic entropy encoder in the same way as the encoder.

[0040] 5) Motion Estimation and Motion Compensation:

[0041] For a PU, we define a search window W in the reference frame. Within this search window, we use the ICP algorithm to find the matching block that is closest to the current coded node, i.e., the one with the lowest Lagrangian cost, and obtain the corresponding MV. The local point set mapped to the MV with the lowest cost is the optimal prediction block. If there is no corresponding prediction block generated by motion estimation and motion compensation, the node at the corresponding position in the reference frame is directly used as the prediction block.

[0042] In step 5), the defined window size can be set according to the different distribution characteristics of the data set and different bit rates. If the point cloud distribution is relatively discrete, a larger window range is set; if the point cloud is relatively dense, a smaller window size can be set.

[0043] In step 5), the ICP algorithm is simplified accordingly, considering only translation transformation and ignoring rotation transformation. The translation vector obtained after ICP is the motion vector. The Lagrangian cost calculation formula is as follows:

[0044] Cost(MV)=Dist(Q(W,MV),B)+λEst(MV)(Formula 1)

[0045] Where B is the current block to be encoded, W is the search window, Q is the set of nearest neighbor points found after B is translated by MV, Est() estimates the codeword required to encode MV, which is obtained by calculating the codeword required for exponential Golomb coding, and Dist is the block matching loss function, which is as follows:

[0046] Dist(Q, W) = ∑ q∈Q log2(1+min w∈W ||qw||1)(Formula 2)

[0047] At the same time, searching for matching blocks involves searching for the nearest neighbor point of each point in the block to be encoded within the window. The present invention accelerates this search process by establishing a KD tree.

[0048] 6) Context entropy coding based on inter-frame prediction blocks

[0049] The occupancy information of the predicted block obtained in step 5) is used as context to help encode the occupancy information of the current node. [1] The point cloud intra-frame context entropy coding tool in , adds inter-frame context information to help further improve the entropy coding efficiency.

[0050] In step 6), for the prediction block obtained in step 5), we synchronize it with the current node to be coded and perform octree partitioning to obtain the occupancy code information of the prediction block and the current block respectively. For each child node, there are two inter-frame prediction context situations: 1) the corresponding position child node in the prediction block is occupied, 2) and the corresponding position child node in the prediction block is not occupied. Select one of the following three methods to implement inter-frame context: The first method is to add an additional context bit to each child node on the basis of the original intra-frame context to indicate whether the corresponding position in the prediction block is occupied. The second method is that if the corresponding prediction child node is occupied, it is considered to be a strong prediction, and the prediction confidence is the same as that of the first 7 neighboring child nodes in the frame are occupied, and the occupancy codes of the first 7 child nodes in the Morton order in the corresponding intra-frame context mode [1] are all set to 1. If the corresponding prediction child node is not occupied, the original intra-frame context information is retained. The third method is that if the corresponding predicted sub-node is occupied, it is considered as a strong prediction, and the prediction confidence is the same as if the first three neighboring sub-nodes in the frame are occupied, and the occupancy codes of the first three sub-nodes in the Morton order in the corresponding intra-frame context pattern [1] are all set to 1. If the corresponding predicted sub-node is not occupied, the original intra-frame context information is retained.

[0051] Figure 1 It is a flowchart of the method of the present invention, combined with Figure 1 The block-partition-based point cloud geometry inter-frame prediction method provided by the present invention is explained. The method of the present invention uses the spatial occupancy information of the previously encoded frame for point cloud data to predict the geometric occupancy information of the current block to be encoded, and uses it as context to improve the occupancy code entropy coding efficiency, thereby improving the compression performance of point cloud geometric information.

[0052] like Figure 2 As shown below, for the official point cloud dataset Ford_01_AVS_1mm sequence in the AVS point cloud compression working group, the method of the present invention is used to perform lossless compression of point cloud geometric information. The specific implementation steps are as follows:

[0053] 1. Calculate the bounding box as the root node of the spatial octree:

[0054] Here, we use the first bounding box calculation method as an example. We first traverse the entire sequence and calculate the minimum bounding box that can contain all frames. The bounding box is determined by the starting point's three-dimensional coordinates and its length, width, height, and side lengths. For this sequence, the corresponding bounding box information is: the starting point's xyz coordinates are (-115100, -115025, -44140), and the side lengths are (230239, 230316, 48208). For all frames in the sequence, the tree partitioning uses this bounding box as the root node.

[0055] 2. Taking the first frame as the reference frame and encoding the second frame as an example, the first frame uses full intra-frame coding because there is no forward coded frame. When encoding the second frame, the reconstructed first frame is input as the reference frame. The frame to be coded is used as the benchmark, and the point cloud space of the two frames is synchronously divided into two identical tree structures. Each tree node corresponds to a spatial block area. According to the octree, quadtree, or binary tree partitioning mode, a spatial block can be further divided into eight, four, or two sub-block spaces. Occupancy codes of different lengths are used to indicate whether there are points in the corresponding sub-block. Only nodes with points are further divided at each level.

[0056] 3. Set PTU_size to 4096. For child nodes generated by tree partitioning, if the minimum side length of the current node to be coded is greater than 4096, the node at the corresponding position in the reference frame is directly used as the prediction block. If the minimum side length of the previous node is equal to 4096, it is considered the prediction starting unit, and the prediction block with the minimum error is searched in the reference frame using the motion estimation.

[0057] 4. Set the window size to PTU_size+2*512, that is, for the range of the current PTU, expand the length of 512 in the three coordinate directions. The PTU can be further divided into a PU tree structure and corresponding multiple PUs through the octree. The PTU can also be regarded as a larger PU. The PU tree contains two flags: split flags is used to indicate whether the nodes at each layer of the PU tree are further divided. If the division continues, occupied flags is used to indicate the occupancy information of the divided child nodes. For each PU, first calculate the encoding cost when it is not further divided. The cost calculation formula is as follows:

[0058] Cost(MV)=Dist(Q(W,MV),B)+λEst(MV)

[0059] Where B is the current block to be encoded, W is the search window, MV is the motion vector corresponding to the closest matching block obtained by the iterative nearest point algorithm within the search window, Q is the set of nearest neighbor points found after B is translated by MV, Est() estimates the codeword required to encode MV, which is obtained by calculating the codeword required for exponential Golomb coding, and Dist is the block matching loss function, which is formulated as follows:

[0060]

[0061] Next, the coding cost of dividing the current PU into 8 sub-PUs is calculated. The total coding cost is obtained by summing the cost of each PU:

[0062]

[0063] The comparison of Cost (MV) and Cost (PU tree) determines whether the current PU should be further divided. The maximum depth of the PU tree is set to 2 layers. After motion estimation, each PU obtains an MV and a corresponding prediction block. Motion compensation is then performed, which replaces the node at the corresponding position in the reference frame with the predicted block. The corresponding motion vector and PU tree structure information are encoded so that the decoder can also generate the same prediction block.

[0064] 5. For a PU to be encoded, after obtaining the prediction block through step 3 or step 4, the prediction block and PU are further divided into trees to obtain the corresponding prediction occupancy code, which is used as context to help encode the occupancy information of the current PU. For the child nodes divided from the PU, there are two inter-frame prediction context situations: 1) the child node at the corresponding position in the prediction block is occupied, and 2) the child node at the corresponding position in the prediction block is not occupied. Select one of the following three methods to implement the inter-frame context: The first method is to add an additional context bit to each child node on the basis of the original intra-frame context to indicate whether the corresponding position in the prediction block is occupied. The second method is that if the corresponding prediction child node is occupied, it is considered to be a strong prediction, and the prediction confidence is the same as if the first 7 neighboring child nodes in the frame are all occupied, and the occupancy codes of the first 7 child nodes in the Morton order in the corresponding intra-frame context mode [1] are all set to 1. If the corresponding prediction child node is not occupied, the original intra-frame context information is retained. The third method is that if the corresponding predicted sub-node is occupied, it is considered as a strong prediction, and the prediction confidence is the same as if the first three neighboring sub-nodes in the frame are occupied, and the occupancy codes of the first three sub-nodes in the Morton order in the corresponding intra-frame context pattern [1] are all set to 1. If the corresponding predicted sub-node is not occupied, the original intra-frame context information is retained.

[0065] 6. Repeat steps 2-5 to obtain the current frame space occupancy code information stream, as well as the motion vector and PU tree code streams, to complete the encoding of the point cloud geometry information.

[0066] In order to verify the effect of the geometric inter-frame prediction compression method of the present invention, we perform lossy geometric coding with a quantization step size of 512 on the first 600 frames of the Ford_01_AVS_1mm sequence. Except for the first frame which uses full intra-frame prediction, the subsequent frames all use the inter-frame plus intra-frame prediction mode, and the inter-frame context uses the second method. The compression performance is compared with the existing method as shown below. Figure 2 shown.

[0067] from Figure 2 As can be seen, under the same test conditions, the combination of the block-based inter-frame prediction tool of this invention yielded stable performance gains across the tested multi-frame sequence of point clouds. By leveraging the temporal correlation of point clouds and reducing coding information redundancy, compression performance of point cloud geometry was improved at all bitrates.

[0068] The present invention's point cloud geometric inter-frame prediction method based on block motion estimation and motion compensation can be used to achieve geometric compression of point clouds. The method first calculates the bounding box information of the point cloud, then performs octree or quadtree and binary tree partitioning. When the size of the partitioned sub-nodes reaches the set prediction unit requirement, the corresponding prediction block with smaller error is searched for the current node in the reference frame. Finally, the geometric occupancy information of the prediction block is used to help improve the entropy coding efficiency of the occupancy information of the current node.

[0069] It should be noted that the purpose of disclosing the embodiments is to facilitate a further understanding of the present invention. However, those skilled in the art will appreciate that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments, and the scope of protection claimed by the present invention shall be determined by the scope defined in the claims.

[0070] References

[0071] [1] AVS Point Cloud Group Conference Proposal: EE11 Context-Based Point Cloud Geometry Compression, M5310, Jun. 2020

Claims

1. A point cloud geometry inter-frame prediction method based on block motion estimation and motion compensation, comprising the following steps: 1) To ensure a one-to-one correspondence between the nodes of the frame to be coded and the reference frame, the same bounding box must be set for the frame to be coded and the reference frame as the root node of the tree partitioning. After the bounding box is determined, the tree partitioning is performed synchronously for the frame to be coded and the reference frame. 2) Set the size of the prediction starting unit (PTU). If the current node size is equal to the set PTU size, the current block is considered to be a prediction unit. A matching block is searched for the current node within a specific window range in the reference frame. The matching block with the smallest error is the predicted block. For a prediction unit PU, a search window W is defined. The window size can be set according to the different distribution characteristics of the data set. If the point cloud distribution is relatively discrete, a larger window range is set; if the point cloud is relatively dense, a smaller window size can be set. 3) The occupancy information of the predicted block is used as context to help encode the occupancy information of the current block, thereby improving the efficiency of entropy coding. Specifically, the predicted block obtained in the above process is synchronously divided into sub-blocks with the current block to be encoded, and the occupancy code information of the predicted block and the current block are obtained respectively. For each sub-node, there are two types of inter-frame prediction context: ① The corresponding sub-node in the predicted block is occupied, and ② The corresponding sub-node in the predicted block is not occupied.

2. The cloud geometry inter-frame prediction method according to claim 1, characterized in that: There are two ways to calculate the bounding boxes of the reference frame and the current frame in step 1): the first method is to count all frames in the entire sequence and calculate the minimum bounding box that can contain all frames as the unified bounding box of the entire sequence; the second method is to calculate the bounding box size only in the reference frame and the current frame, and take the maximum bounding box size as the root node of the tree partition during the encoding process of the current frame.

3. The cloud geometry inter-frame prediction method according to claim 1, wherein: The tree partitioning operation in step 1) can be combined with octree, quadtree, and binary tree partitioning methods according to different bounding box characteristics to adapt to non-cubic bounding boxes and reduce the proportion of empty blocks.

4. The cloud geometry inter-frame prediction method according to claim 1, wherein: In step 2), if the tree is divided into octree mode, the current node is a cube, and its side length is determined to be equal to the PTU size. If the tree is divided into quadtree or binary tree mode, the current node is a cuboid, and its shortest side length is determined to be equal to the PTU size, which means that the current block is considered to be the prediction starting unit.

5. The cloud geometry inter-frame prediction method according to claim 1, wherein: In step 2), the matching block search process is to first find the most approximate matching block in the obtained window local reference point cloud using the ICP algorithm, that is, to minimize the Lagrangian cost. The ICP algorithm is simplified here and only considers translation transformation, not rotation transformation. The translation vector obtained after ICP is the motion vector. The Lagrangian cost calculation formula is as follows: in B is the current block to be encoded, W For the search window, Q for B After the motion vector MV The nearest neighbor point set found after translation, Est() estimates the codeword required to encode MV, which is obtained by calculating the codeword required for exponential Columbus encoding. Dist is the block matching loss function, and the formula is as follows: The final encoding cost is the smallest MV The corresponding local point set is the optimal prediction block.

6. The cloud geometry inter-frame prediction method according to claim 1, characterized in that: In step 2), the above search for matching blocks involves searching for the nearest neighbor point of each point in the block to be encoded within the window, and this search process is accelerated by establishing a KD tree.

7. The cloud geometry inter-frame prediction method according to claim 1, characterized in that: In step 3), for each child node, an additional context bit is added based on the original intra-frame context to indicate whether the corresponding position in the prediction block is occupied.

8. The cloud geometry inter-frame prediction method according to claim 1, wherein: In step 3), if the corresponding predicted sub-node is occupied, it is considered a strong prediction, and the prediction confidence is the same as if the first 7 neighbor sub-nodes in the frame are occupied. The occupancy codes of the first 7 sub-nodes in the Morton order in the corresponding intra-frame context pattern are all set to 1. If the corresponding predicted sub-node is not occupied, the original intra-frame context information is retained.

9. The cloud geometry inter-frame prediction method according to claim 1, wherein: In step 3), if the corresponding predicted sub-node is occupied, it is considered a strong prediction, and the prediction confidence is the same as if the first three neighbor sub-nodes in the frame are occupied. The occupancy codes of the first three sub-nodes in the Morton order in the corresponding intra-frame context pattern are all set to 1. If the corresponding predicted sub-node is not occupied, the original intra-frame context information is retained.

Images