A Point Cloud Hybrid Motion Prediction Method Based on Point Cloud Prediction Tree
By introducing hybrid motion prediction into the point cloud prediction tree, combining global and local motion, the problem of local motion complexity in point cloud compression is solved, achieving high-precision and efficient point cloud compression, which is suitable for complex scenarios such as autonomous driving.
Patent Information
- Application Number
- CN202610419166.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-01
- Publication Date
- 2026-05-26
AI Technical Summary
Existing point cloud compression techniques struggle to effectively characterize non-rigid deformations, local occlusions, or segmented motions when faced with complex local motions. This leads to unstable cross-frame correspondences, decreased prediction accuracy, and difficulty in achieving stable coding gains under different content and motion conditions.
A hybrid motion prediction method based on point cloud prediction trees is adopted. By expanding the candidate set and combining global and local motion, a hybrid motion prediction is generated. The local structure offset is used to compensate for the global motion model, thereby achieving high-precision prediction.
It significantly improves the accuracy and compression efficiency of point cloud prediction, and exhibits good robustness and coding gain, especially in complex scenarios such as autonomous driving, while reducing computational complexity.
Smart Images
Figure CN122093578A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of point cloud encoding and decoding technology, and particularly relates to a point cloud hybrid motion prediction method based on point cloud prediction trees. Background Technology
[0002] Point clouds are a type of data that represents a scene or object using a set of discrete points in three-dimensional space. Each point typically contains at least three-dimensional geometric coordinates (x, y, z) and can carry attribute information such as color (R, G, B or YUV), reflectance, and normal vectors. Point clouds can be used in immersive media, autonomous driving LiDAR (radar) perception, digital twins, 3D reconstruction, and AR / VR scenarios, but their data size is large and highly redundant, requiring efficient compression.
[0003] MPEG (Moving Picture Experts Group) has developed point cloud compression standards, among which Geometry-based Point Cloud Compression (G-PCC, often corresponding to ISO / IEC 23090-9) is based on a geometry-centric compression paradigm and is suitable for sparse to moderately dense point clouds (including LiDAR point clouds and dynamic point cloud sequences). G-PCC's approach is to separately encode the geometric information and attribute information of the point cloud: first encode the geometry, and then encode attributes (color / reflectivity, etc.) based on the geometric structure, thereby achieving a modular and scalable encoding and decoding process.
[0004] Predictive geometry coding (PredGeom), also often described as "predictive tree decoding," is a prediction focused on the "points themselves." The key to predictive tree geometry coding lies in its transformation of the geometric redundancy of point clouds into "local correlations within a tree structure," achieving efficient prediction through the geometric relationships between parent and child nodes. For point clouds with obvious scanning structures or local continuity (such as LiDAR point clouds), prediction trees have a significant advantage in certain scenarios.
[0005] For point cloud sequences (dynamic point clouds), there is usually a strong correlation between adjacent frames (e.g., the motion of an object and the motion of a sensor are relatively continuous in a short period of time). Therefore, the geometric coding of the prediction tree can further introduce inter-frame prediction: during the prediction tree traversal, the current frame node not only depends on the local correlation provided by the parent node of this frame, but can also generate candidate nodes from the reference frame (usually the previous reconstructed frame or multiple reference frames) as predictors, thereby effectively reducing the prediction residual and improving the overall compression efficiency.
[0006] However, inter-frame prediction still faces challenges arising from the complexity of local motion. Relying solely on global motion models (such as rigid body transformation) often fails to adequately characterize phenomena such as non-rigid deformation, local occlusion, or segmented motion, leading to unstable cross-frame correspondences and consequently, candidate matching mismatches and decreased prediction accuracy. In this context, while expanding and selecting the candidate set may improve prediction, it also introduces additional computational and bit overhead.
[0007] Therefore, the prediction tree inter-frame coding of G-PCC needs to be designed with a comprehensive trade-off between prediction accuracy, implementation complexity, flag overhead, and encoding / decoding reconstructability consistency in order to obtain stable gains under different content and motion conditions. Summary of the Invention
[0008] To address the above problems, this invention provides a point cloud hybrid motion prediction method based on point cloud prediction trees.
[0009] The technical solution of the present invention is as follows: A point cloud hybrid motion prediction method based on point cloud prediction trees, comprising the following steps: Step 1: Obtain geometric prediction parameters and initialize the sequence. The geometric configuration information of the current point cloud frame is obtained by parsing the bitstream, and each point to be decoded is obtained.
[0010] Step 2: Reference Frame Candidate Point Generation Step 2.1: Expanding Candidates For the current point to be decoded, each subsequent range is expanded to cover more complex local motions. This yields a set of candidate points for the reference frame. : in, The set of candidate points for the reference frame; For reference frames, including and , For a motionless reference frame, This serves as a reference frame for global motion compensation. For in set Reference frame candidate points inside, include and , Representing the Candidate points for global motion compensation reference frames Representing the Candidate points of a motionless reference frame There are a total of 2M candidate reference frames; and The radar coordinates are represented as (LaserID, azimuth, radial distance).
[0011] Step 2.2, Hybrid Global-Local Motion Prediction Using only global motion compensation reference frames For the One global motion compensation reference frame candidate point Make the necessary modifications.
[0012] Step 2.2.1: Local Structure Offset Extraction definition In order to be in Find the parent node of the current point to be decoded. A point with the same azimuth angle, for , define the first Local azimuth offset of relative same location : in, The azimuth angle of the point is denoted as .
[0013] Step 2.2.2: Generation of azimuth angle for hybrid global-local motion prediction Will Superimposed on the parent node of the current point to be decoded azimuth , obtained the Azimuth angle based on hybrid global-local motion prediction : Step 2.3: Sector Arbitration Based on Quantized Mesh To ensure alignment between the prediction and the angle quantization grid, the azimuth angle is quantized according to the angle quantization step size. The system is divided into discrete sectors, and the two prediction methods are analyzed. and Arbitration, among which, For the first An azimuth angle predicted based on global motion compensation.
[0014] Step 2.3.1, Sector Index Calculation by Calculate with the center as the center: in, For the first A sector index based on global motion compensation prediction. For the first A sector index based on hybrid global-local motion prediction, This is the floor function.
[0015] Step 2.3.2, Sector Consistency Judgment and Arbitration Trigger like They believe that global predictions are consistent with local structures. = , For the first A final hybrid global-local motion prediction sector index; like Trigger arbitration: compare the "centering degree" of the predictions within their respective sectors.
[0016] Step 2.3.3: Centering Degree Measurement Calculate the predicted distance to the center of each sector: in, For the first A sector distance based on global motion compensation prediction, No. A sector distance based on hybrid global-local motion prediction.
[0017] Step 2.3.4: Final sector selection and mesh snapping Select the sector with the smaller distance as the final sector: The final prediction is then snapped to the center of the sector to ensure strict alignment with the decoding grid. in, For the first The final predicted value of the azimuth angle based on hybrid global-local motion prediction.
[0018] Finally In Replace the azimuth of the point with The final output is a set of candidate reference points. Thus, candidate points for the reference frame are obtained.
[0019] Step 3: Prediction Point Selection Parse the refNodeIdx flag from the bitstream and select from the set. The predicted value is obtained by identifying the unique candidate reference frame point indicated by the flag bit.
[0020] Step 4: Point Cloud Reconstruction The quantization residual obtained from decoding the current point to be decoded is dequantized, and the dequantized residual is added to the predicted value to obtain the radar coordinates of the current point to be decoded. The radar coordinates (LaserID, azimuth, radial distance) of the current point to be decoded are converted into Cartesian coordinates (x, y, z) to complete the reconstruction of the current point; each point to be decoded is decoded to obtain the reconstructed point cloud, thus completing the point cloud hybrid motion prediction.
[0021] The beneficial effects of this invention are: (1) Break through the limitations of pure global motion model and achieve high-precision hybrid motion prediction.
[0022] The corresponding drawback is that the existing G-PCC only relies on the global motion model (such as rigid body transformation), which makes it difficult to characterize non-rigid deformation, local occlusion or segmented motion of independent objects, resulting in unstable cross-frame correspondence.
[0023] This invention no longer relies solely on global motion, but instead fuses coarse-grained global motion trajectories with fine-grained local geometric offsets. Specifically, the final predicted position is obtained by superimposing the local motion of a reference node relative to its parent node onto coordinates compensated for global motion. This mechanism effectively captures local object motion independent of the global background, fundamentally solving the prediction mismatch problem caused by the lack of reliable correspondences in dynamic point clouds (especially sparse LiDAR data), and significantly improving prediction accuracy.
[0024] (2) Scientifically expand the candidate set to fully cover complex and varied types of sports.
[0025] The corresponding drawback is that the local motion is extremely complex, and traditional candidate matching often results in a decrease in accuracy.
[0026] To address complex dynamic scenarios, this invention scientifically expands the candidate node list and introduces a wider range of predictors to cover various motion types (such as translation, rotation, and non-rigid deformation). This expansion breaks through the limitations of traditional single predictors, enabling the encoder to have richer matching options when facing complex local motions, thereby effectively reducing prediction residuals.
[0027] (3) Significantly improves the overall compression efficiency in specific scenarios (such as dynamic autonomous driving).
[0028] The corresponding drawback is that existing technologies struggle to achieve stable gains under different content and motion conditions.
[0029] For large-scale, sparse LiDAR point cloud sequences generated by autonomous driving and other applications, the hybrid prediction architecture of this invention exhibits strong robustness. Experimental results under MPEG general test conditions show that, compared with existing G-PCC anchors, the method of this invention can provide consistent and significant coding gains in various dynamic scenarios, especially in driving scenarios with dense independent moving objects. Attached Figure Description
[0030] Figure 1 This is a schematic diagram of the hybrid motion prediction method based on point cloud prediction tree of the present invention; Figure 2 This is a flowchart of the hybrid motion prediction method based on point cloud prediction tree of the present invention; Figure 3 This is a schematic diagram of a partially reconstructed point cloud sequence after compression using the hybrid motion prediction method based on point cloud prediction trees in the embodiment. Detailed Implementation
[0031] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.
[0032] like Figure 1 and Figure 2 As shown, the specific steps of the point cloud hybrid motion prediction method based on point cloud prediction tree of the present invention are as follows: Step 1: Obtain geometric prediction parameters and initialize the sequence. The geometric configuration information of the current point cloud frame is obtained by parsing the bitstream, and each point to be decoded is obtained.
[0033] Step 2: Reference Frame Candidate Point Generation Step 2.1: Expanding Candidates For the current point to be decoded, in order to cover more complex local motions, in this embodiment, the subsequent range is extended to... Thus, the reference frame candidate point set is obtained. : in, The set of candidate points for the reference frame; For reference frames, including and , For a motionless reference frame, This serves as a reference frame for global motion compensation. For in set Reference frame candidate points inside, include and , Representing the Candidate points for global motion compensation reference frames Representing the There are 8 motionless reference frame candidate points; and The radar coordinates are represented as (LaserID, azimuth, radial distance).
[0034] Step 2.2, Hybrid Global-Local Motion Prediction This step uses only the global motion compensation reference frame. For the One global motion compensation reference frame candidate point Make the necessary modifications.
[0035] Step 2.2.1: Local Structure Offset Extraction definition In order to be in Find the parent node of the current point to be decoded. A point with the same azimuth angle, for , define the first Local azimuth offset of relative same location : in, The azimuth angle of the point is denoted as .
[0036] Step 2.2.2: Generation of azimuth angle for hybrid global-local motion prediction Will Superimposed on the parent node of the current point to be decoded azimuth , obtained the Azimuth angle based on hybrid global-local motion prediction : This synthesis achieves "transferring the local angular topological offset of the reference frame to the current frame" to compensate for local motions that cannot be described by the global rigid body assumption.
[0037] Step 2.3: Sector Arbitration Based on Quantized Mesh To ensure alignment between the prediction and the angle quantization grid, the azimuth angle is quantized according to the angle quantization step size. The system is divided into discrete sectors, and the two prediction methods are analyzed. and Arbitration, among which, For the first An azimuth angle predicted based on global motion compensation.
[0038] Step 2.3.1, Sector Index Calculation by Calculate with the center as the center: in, For the first A sector index based on global motion compensation prediction. For the first A sector index based on hybrid global-local motion prediction, This is the floor function.
[0039] Step 2.3.2, Sector Consistency Judgment and Arbitration Trigger like They believe that global predictions are consistent with local structures. = , For the first A final hybrid global-local motion prediction sector index; like Trigger arbitration: compare the "centering degree" of the predictions within their respective sectors.
[0040] Step 2.3.3: Centering Degree Measurement Calculate the predicted distance to the center of each sector: in, For the first A sector distance based on global motion compensation prediction, No. A sector distance based on hybrid global-local motion prediction.
[0041] Step 2.3.4: Final sector selection and mesh snapping Select the sector with the smaller distance as the final sector: The final prediction is then snapped to the center of the sector to ensure strict alignment with the decoding grid. in, For the first The final predicted value of the azimuth angle based on hybrid global-local motion prediction.
[0042] Finally In Replace the azimuth of the point with The final output is a set of candidate reference points. Thus, candidate points for the reference frame are obtained.
[0043] Step 3: Prediction Point Selection Parse the refNodeIdx flag from the bitstream and select from the set. The predicted value is obtained by identifying the unique candidate reference frame point indicated by the flag bit.
[0044] Step 4: Point Cloud Reconstruction The quantization residual obtained from decoding the current point to be decoded is dequantized, and the dequantized residual is added to the predicted value to obtain the radar coordinates of the current point to be decoded. The radar coordinates (LaserID, azimuth, radial distance) of the current point to be decoded are converted into Cartesian coordinates (x, y, z) to complete the reconstruction of the current point; each point to be decoded is decoded to obtain the reconstructed point cloud, thus completing the point cloud hybrid motion prediction.
[0045] The method proposed in this embodiment is implemented in the MPEG G-PCC reference software, specifically version TMC13-v31. Performance testing strictly follows the MPEG Common Test Conditions (CTC), and Category 3 (Cat3) frame LiDAR sequences are selected as the test dataset. The Cat3 sequences are publicly available dynamic point cloud sequences, acquired by sensors mounted on a moving vehicle, and can characterize typical environmental features in autonomous driving scenarios. Specifically, the test sequences include multiple Ford and QNXADAS sequences, covering different point cloud densities, self-motion modes, and local target motion conditions, thus comprehensively verifying the applicability and effectiveness of the method of this invention.
[0046] This embodiment uses the predgeom--raht--inter anchor configuration as the comparison benchmark. Objective quality evaluation is performed using the MPEG standard tool pc_error to ensure the consistency and comparability of evaluation results across different test configurations. Encoding performance is evaluated using Bj... The ntegaard Delta Bitrate (BD-Rate) metric is used for quantitative characterization. When the BD-Rate is negative, it indicates that the method of this invention has achieved bitrate savings relative to the anchor point configuration, which means it has better compression performance.
[0047] Furthermore, to evaluate the computational complexity of the method of this invention, all encoding and decoding tests were performed on the same hardware and software platform. The test platform consisted of an Intel Core Ultra 7 265K processor, 64 GB of memory, and an Ubuntu 22.04 Linux operating system.
[0048] The experimental scope of this embodiment includes two standard coding conditions: geometrically lossy and attribute-lossy (C2) and geometrically lossless and attribute-lossless (CW). These two test conditions cover all geometric coding test scenarios defined in the common test conditions. Furthermore, computational complexity is evaluated using average relative processing time.
[0049] The experimental results of this embodiment are shown in Table 1.
[0050] Table 1 Experimental Results
[0051] Figure 3 The image shows the effect of lossless compression on frame 0109 of the dataset ford_01_q1mm. As can be seen from the image, the geometric reconstruction of the point cloud is well achieved using the method of this invention in lossless compression mode.
[0052] As shown in Table 1, the method of this invention achieves a stable geometric rate reduction effect on Cat3 sequences: under the C2 configuration, the average end-to-end rate distortion improvement of D1 and D2 indices is 0.55% and 0.94%, respectively. This indicates that the method can further reduce the geometric rate overhead without changing the existing coding framework.
[0053] In terms of computational complexity, this method maintains high engineering usability. As shown in Table 1, the average relative encoding time is approximately 102%, and the average relative decoding time is approximately 100%. This indicates that the method of this invention achieves a geometric rate improvement while introducing only negligible additional computational overhead, demonstrating good practicality and deployability.
[0054] In summary, this embodiment verifies that the method of the present invention can achieve stable geometric code rate reduction under standard test conditions and maintain low complexity overhead in most sequences.
Claims
1. A point cloud hybrid motion prediction method based on point cloud prediction trees, characterized in that, The steps are as follows: Step 1: Obtain geometric prediction parameters and initialize the sequence. Step 2: Reference Frame Candidate Point Generation Step 2.1: Expanding Candidates For the current point to be decoded, each subsequent range is expanded to cover more complex local motions. This yields a set of candidate points for the reference frame. ; Step 2.2, Hybrid Global-Local Motion Prediction Using only global motion compensation reference frames For the One global motion compensation reference frame candidate point Make modifications; Step 2.3: Sector Arbitration Based on Quantized Mesh To ensure alignment between the prediction and the angle quantization grid, the azimuth angle is quantized according to the angle quantization step size. The system is divided into discrete sectors, and the two prediction methods are analyzed. and Arbitration, among which, For the first An azimuth angle predicted based on global motion compensation; In Replace the azimuth of point with the first The final predicted value of the azimuth angle based on hybrid global-local motion prediction. The final output is a set of candidate reference points. Thus, candidate points for the reference frame are obtained; Step 3: Prediction Point Selection Parse the refNodeIdx flag from the bitstream and select from the set. The predicted value is obtained by identifying the unique candidate reference frame point indicated by this flag bit; Step 4: Point Cloud Reconstruction Perform inverse quantization on the quantization residual obtained from decoding the current point to be decoded, add the inverse quantization residual to the predicted value to obtain the radar coordinates of the current point to be decoded; convert the radar coordinates (LaserID, azimuth, radial distance) of the current point to be decoded into Cartesian coordinates (x, y, z) to complete the reconstruction of the current point; decode each point to be decoded to obtain the reconstructed point cloud, thus completing the point cloud hybrid motion prediction.
2. The point cloud hybrid motion prediction method based on point cloud prediction tree according to claim 1, characterized in that, Step 1 specifically involves parsing the geometric configuration information of the current point cloud frame from the bitstream to obtain each point to be decoded.
3. The point cloud hybrid motion prediction method based on point cloud prediction tree according to claim 1, characterized in that, In step 2.1: in, The set of candidate points for the reference frame; For reference frames, including and , For a motionless reference frame, This serves as a reference frame for global motion compensation. For in set Reference frame candidate points inside, include and , Representing the Candidate points for global motion compensation reference frames Representing the Candidate points of a motionless reference frame There are a total of 2M candidate reference frames; and The radar coordinates are represented as (LaserID, azimuth, radial distance).
4. The point cloud hybrid motion prediction method based on point cloud prediction tree according to claim 1, characterized in that, Step 2.2 is as follows: Step 2.2.1: Local Structure Offset Extraction definition In order to be in Find the parent node of the current point to be decoded. A point with the same azimuth angle, for , define the first Local azimuth offset of relative same location : in, The azimuth angle of the point; Step 2.2.2: Generation of azimuth angle for hybrid global-local motion prediction Will Superimposed on the parent node of the current point to be decoded azimuth , obtained the Azimuth angle based on hybrid global-local motion prediction : 。 5. The point cloud hybrid motion prediction method based on point cloud prediction tree according to claim 1, characterized in that, Step 2.3 is as follows: Step 2.3.1, Sector Index Calculation by Calculate with the center as the center: in, For the first A sector index based on global motion compensation prediction. For the first A sector index based on hybrid global-local motion prediction, It is a rounding function; Step 2.3.2, Sector Consistency Judgment and Arbitration Trigger like They believe that global predictions are consistent with local structures. = , For the first A final hybrid global-local motion prediction sector index; like Trigger arbitration: compare the "centering degree" of the predictions within their respective sectors; Step 2.3.3: Centering Degree Measurement Calculate the predicted distance to the center of each sector: in, For the first A sector distance based on global motion compensation prediction, No. Sector distances based on hybrid global-local motion prediction; Step 2.3.4: Final sector selection and mesh snapping Select the sector with the smaller distance as the final sector: The final prediction is then snapped to the center of the sector to ensure strict alignment with the decoding grid. in, For the first The final predicted value of the azimuth angle based on hybrid global-local motion prediction; Will In Replace the azimuth of the point with The final output is a set of candidate reference points. Thus, candidate points for the reference frame are obtained.