A lossless point cloud attribute compression method and system based on slice and intra prediction

Abstract

The application discloses a lossless point cloud attribute compression method and system based on slice and intra-block prediction, which firstly carries out slice processing on a three-dimensional point cloud along a main coordinate axis, divides the three-dimensional point cloud into smaller sub-blocks, and carries out adaptive blocking on the slices based on local attribute gradients; in the sub-blocks, representative key points are extracted, and attribute traversal orders are established with the key points as guides, and attribute residuals are obtained through intra-block prediction; subsequently, multi-source context features are constructed by fusing traversal history, spatial neighborhood, slice scanning history and key point statistical information, and a multi-layer perception (MLP) is used to dynamically model probability distribution parameters of the residuals; finally, arithmetic coding is performed on the residuals according to the probability distribution obtained by modeling to generate a bit stream. The application deeply mines the multi-dimensional correlation of point cloud attributes, significantly improves the bit rate efficiency of lossless compression, and is suitable for high-bandwidth demand scenes such as automatic driving and virtual reality.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional point cloud data processing and compression technology, specifically relating to a lossless point cloud attribute compression method and system based on slicing and intra-block prediction. Background Technology

[0002] With the rapid development of 3D acquisition equipment, point cloud data has become an important carrier for representing 3D information. Technologies such as LiDAR, depth cameras, and multi-view stereo reconstruction can capture the geometric structure and surface properties of objects and scenes, and are widely used in applications such as autonomous driving, virtual reality, augmented reality, smart cities, robot perception, and digital cultural heritage. Point clouds consist of millions or even hundreds of millions of 3D coordinate points, and usually also contain multi-dimensional attributes such as color, normal vectors, and reflectivity, resulting in an extremely large data volume.

[0003] In typical applications, such as autonomous driving, each vehicle can collect hundreds of megabytes of data per second; in film and VR / AR production, high frame rate point cloud sequences can generate terabytes of data. This massive volume not only poses a huge challenge to storage space but also places extremely high demands on real-time transmission and bandwidth. Therefore, how to efficiently compress point cloud data to achieve low bit rate, lossless or near-lossless transmission is a key issue of common concern to both academia and industry.

[0004] In recent years, deep learning has shown great potential in point cloud compression. Unlike traditional methods, learning-based methods can automatically model the spatial and contextual relationships of point clouds, bringing them closer to information theory optimality. However, most current research focuses on geometric compression, while attribute compression research lags behind. Existing attribute compression methods are often based on voxel or sequence unfolding structures, but these methods have limitations in areas such as insufficient block boundary context and difficulties in modeling non-uniform point clouds.

[0005] The international standard MPEG G-PCC represents the mainstream solution in the current industry. In attribute coding, it mainly adopts distance-based prediction / lifting transform and region adaptive hierarchical transform (RAHT). However, when processing point cloud attributes with complex textures or high dynamic range, G-PCC suffers from difficulties in fully eliminating cross-regional statistical redundancy in lossless compression scenarios because its prediction mechanism relies too heavily on fixed spatial neighborhood relationships.

[0006] Patent CN117579832A discloses a point cloud attribute compression method based on adaptive sampling and quantization. It calculates the sampling distance based on the texture complexity, generates a level of detail, and then generates a predictor based on the level of detail to achieve lossless point cloud compression. Although these techniques improve compression efficiency through localization, they often suffer from prediction discontinuities at block boundaries and lack organization of the compression order.

[0007] The academic community has also explored solutions based on graph transformation and deep learning. Examples include the attribute compression framework based on hierarchical segmentation proposed by Ke Zhang et al. (ICASSP 2018), and the fine-grained patch segmentation and rasterization method studied by Baoquan Zhao et al. (IEEE TCSVT 2021). While these methods improve accuracy through refined segmentation, they often face problems such as excessively high computational complexity (as seen in graph transformation methods) or limited modeling capabilities in lossless mode. In particular, existing learning-based modeling schemes (such as voxel-based methods) struggle to balance computational efficiency and lossless reconstruction when dealing with non-uniformly distributed point cloud attributes.

[0008] Therefore, there is an urgent need for a new compression framework that can significantly improve attribute compression performance while maintaining data integrity, and also take into account both lightweightness and efficiency. Summary of the Invention

[0009] This invention provides a lossless point cloud attribute compression method and system based on slicing and intra-block prediction. This invention fully utilizes the spatial continuity of point clouds through slicing and adaptive block division, reduces residual variance within blocks by using key points to guide the traversal order, and achieves accurate probabilistic modeling of the residuals through multi-dimensional contextual feature fusion, thereby realizing efficient entropy coding to solve the problem of insufficient context in existing point cloud attribute compression. This invention significantly improves compression bit rate efficiency while ensuring lossless decoding, and provides a lightweight, low-computational-complexity implementation suitable for large-scale practical applications.

[0010] The technical solution of the present invention is as follows: According to one aspect of the present invention, a lossless point cloud attribute compression method based on slicing and intra-block prediction is provided, comprising the following steps: S1. Multidimensional slicing and adaptive block division: The input three-dimensional point cloud is reduced to a two-dimensional slicing window along the principal coordinate axis, and an initial block of fixed size is divided. The local attribute gradient of the initial block is calculated, and the initial block is further divided into smaller sub-blocks along the attribute gradient edge according to the gradient distribution; S2. Keypoint-guided intra-block prediction: Feature salient points are selected as keypoints in each sub-block. The traversal order of each point in the sub-block is determined by using the keypoints as anchor points, and intra-block prediction is performed based on the encoded neighboring points to calculate the attribute prediction residual; S3. Multi-source context feature extraction and dynamic modeling: Multi-source context features, including traversal history features, spatial neighborhood features, slice scanning history features, and keypoint statistical features, are extracted from the current point to be encoded and input into a deep neural network to model the probability distribution of the attribute prediction residual; S4. Entropy coding and bitstream generation: Based on the probability distribution parameters obtained by dynamic modeling, entropy coding is performed on the attribute prediction residual to generate a compressed bitstream.

[0011] Optionally, in the above lossless point cloud attribute compression method based on slice and intra-block prediction, in the adaptive block segmentation step, if the attribute gradient within the initial block exceeds the adaptive threshold, the initial block is recursively subdivided along the gradient edge.

[0012] Optionally, in the above lossless point cloud attribute compression method based on slice and intra-block prediction, the keypoint-guided traversal order is an ordered prediction chain generated by calculating the spatial distance or geometric correlation between points within a sub-block and keypoints, and utilizing the shortest path algorithm.

[0013] Optionally, in the above lossless point cloud attribute compression method based on slice and block prediction, the deep neural network employs a multilayer perceptron (MLP) to predict the Laplace distribution parameters of the residual distribution, including location parameters and scale parameters.

[0014] Optionally, in the above lossless point cloud attribute compression method based on slice and intra-block prediction, the entropy coding employs an arithmetic encoder.

[0015] According to another aspect of the present invention, a lossless point cloud attribute compression system based on slicing and intra-block prediction is provided, comprising: a slicing module for reducing the dimensionality of a 3D point cloud along the principal coordinate axis to a 2D slicing window, dividing it into initial blocks of fixed size, and further subdividing the initial blocks into smaller sub-blocks based on local attribute gradients; a prediction module for extracting key points within the sub-blocks, determining the attribute traversal order, performing intra-block prediction based on encoded neighboring points, and calculating attribute prediction residuals; a modeling module for extracting multi-source contextual features of the current point to be encoded, including traversal history features, spatial neighborhood features, slicing scan history features, and key point statistical features, and inputting them into a deep neural network to model the probability distribution of the attribute prediction residuals; and an encoding module for entropy encoding the attribute prediction residuals based on the probability distribution parameters obtained from modeling, and outputting a compressed bitstream.

[0016] The beneficial effects of the technical solution of the present invention are as follows: This invention significantly improves attribute compression performance while maintaining data integrity, and balances lightweight design and high efficiency, making it particularly suitable for the efficient storage and transmission of large-scale point cloud data.

[0017] To better understand and illustrate the concept, working principle, and effects of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments: Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below.

[0019] Figure 1 This is a flowchart of the lossless point cloud attribute compression method based on slice and intra-block prediction of the present invention; Figure 2 This is the overall flowchart of the lossless point cloud attribute compression method based on slice and intra-block prediction of the present invention; Figure 3 This is a schematic diagram comparing the key point-guided traversal order of the present invention with the traditional center point prediction method; Figure 4 It predicts the correlation between various channels; Figure 5 This is a schematic diagram of the multidimensional context feature extraction and mapping module; Figure 6 It is an intrinsic self-attention module used for feature extraction; Figure 7 This is a schematic diagram of the scanning order of the sub-blocks. Detailed Implementation

[0020] To make the objectives, technical methods, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific examples. These examples are merely illustrative and not intended to limit the scope of the invention.

[0021] The lossless point cloud attribute compression method based on slicing and intra-block prediction of the present invention can significantly improve compression efficiency while ensuring lossless data recovery. Its overall process covers data preprocessing, slicing, key point-guided intra-block prediction, multi-dimensional context modeling and entropy coding, forming a complete encoding and decoding system.

[0022] This invention presents a lossless point cloud attribute compression method based on slicing and intra-block prediction. This method achieves efficient compression of point cloud attribute residuals through multi-dimensional space dimensionality reduction, adaptive geometric segmentation, and deep learning context modeling. The specific scheme includes multi-dimensional slicing and dimensionality reduction processing, adaptive block partitioning based on attribute gradients, keypoint-guided intra-block prediction, multi-source context fusion and dynamic modeling, and entropy coding and bitstream generation.

[0023] The multidimensional slicing and dimensionality reduction process divides the original input point cloud into a series of two-dimensional slice windows along the main axis of recognition. Adaptive block partitioning based on attribute gradients divides the slices into fixed-size square two-dimensional blocks, further subdividing them into sub-blocks along the areas of greatest attribute gradient change. Keypoint-guided intra-block prediction first selects representative points within the sub-blocks as keypoints based on the characteristics of the attribute thumbnails. Using these keypoints as anchors, and combining attribute similarity and spatial distance, an optimized attribute traversal sequence is established. Multi-source context fusion and dynamic modeling extract four types of core features: traversal history features, spatial neighborhood features, slice scanning history features, and keypoint features. These fused features are input into a pre-trained lightweight multilayer perceptron model, which dynamically outputs the Laplace distribution parameters of the residual distribution. Entropy coding and bitstream generation perform arithmetic coding on the attribute prediction residuals. Due to high modeling accuracy, the residual distribution is highly concentrated near zero, significantly compressing the number of bits required for coding and ultimately generating a lossless compressed bitstream.

[0024] Figure 2 This document demonstrates the complete process of the system, including multi-dimensional slicing and dimensionality reduction, adaptive block partitioning based on attribute gradients, keypoint-guided intra-block prediction, multi-source context fusion and dynamic modeling, entropy coding, and bitstream generation. The input point cloud is sliced ​​along the principal axis and divided into two-dimensional slicing windows. Each slicing window is further divided into sub-blocks, which are then adaptively segmented using local attribute gradients to maintain geometric and texture continuity. Within each sub-block, a keypoint-guided traversal order is generated to predict and encode attributes in the order with minimal residuals. Finally, multi-source contextual features, including traversal history, spatial neighborhood, slice scan history, and keypoint statistics, are fused to model the attribute residual distribution for entropy coding.

[0025] The lossless point cloud attribute compression method based on slice and intra-block prediction of the present invention includes the following steps: S1. Multidimensional Slicing and Adaptive Blocking: The input 3D point cloud is reduced to a 2D slice window along the principal coordinate axis and divided into initial blocks of fixed size. The local attribute gradient of the initial block is calculated, and the initial block is further subdivided into smaller sub-blocks along the attribute gradient edge according to the gradient distribution.

[0026] Multidimensional slicing and dimensionality reduction: First, the system acquires the 3D point cloud data to be compressed and identifies its principal coordinate axis (such as the Z-axis). The 3D spatial point cloud is then discretized and sliced ​​along this axis, mapping the 3D spatial positional relationships into a series of ordered 2D slice sequences. Through this partitioning operation, the complex 3D neighborhood correlations are transformed into cross-slice sequence correlations, laying the foundation for the subsequent construction of a cross-dimensional context reference structure.

[0027] Adaptive Blocking Based on Attribute Gradients For each 2D slice window, the system executes a hierarchical partitioning strategy: initial segmentation divides the slice into fixed-size initial square physical blocks, i.e., initial blocks; then adaptive subdivision is performed, calculating the local attribute gradient within each physical block (initial block). Along the boundaries where attribute changes drastically and textures are complex, the initial blocks are further decomposed into smaller-granularity sub-blocks. This step allows the block boundaries to conform to the geometric structure and attribute texture edges, greatly maintaining attribute continuity and effectively reducing prediction distortion at block boundaries. In the adaptive segmentation step, if the attribute gradient within the initial block exceeds an adaptive threshold, the block is recursively subdivided along the gradient edge.

[0028] S2. Keypoint-guided intra-block prediction: Select salient feature points as keypoints in each smaller-granularity sub-block, use keypoints as anchors to determine the traversal order of points in the smaller-granularity sub-block, and perform intra-block prediction based on encoded neighboring points to calculate attribute prediction residuals.

[0029] Within each subdivided sub-block (i.e., a smaller granularity sub-block), a feature-guided prediction process is implemented: first, representative points within the sub-block are selected as keypoints based on the characteristics in the attribute thumbnails. Using these keypoints as anchors, an optimized attribute traversal sequence is established by combining attribute similarity and spatial distance. Following the traversal order, the attribute information of the encoded points within the sub-block is used to predict the current point, and the residual between the original attribute value and the predicted value is calculated. The keypoint-guided traversal order is an ordered prediction chain generated using the shortest path algorithm by calculating the spatial distance or geometric correlation between points within the sub-block and keypoints.

[0030] Figure 3The invention demonstrates a comparison between the keypoint-guided intra-block prediction traversal order method and the traditional centroid-based method. The residuals between adjacent points are generally smaller, indicating that the invention can better maintain geometric and attribute continuity.

[0031] S3. Multi-source context fusion and dynamic modeling: Extract multi-source context features of the current point to be encoded, including traversal history features, spatial neighborhood features, slice scan history features, and key point statistical features, and input them into a deep neural network to model the probability distribution of the attribute prediction residual.

[0032] The deep neural network employs a multilayer perceptron (MLP) to predict the Laplace distribution parameters of the residual distribution, including location and scale parameters. The deep learning model accurately characterizes the residual distribution, extracting four core features: traversal history features: residual information of previous points in the current prediction chain; spatial neighborhood features: spatial correlation of the current point within the 2D slice and 3D neighborhood; slice scan history: attribute statistical distribution of corresponding positions in adjacent slices; and keypoint statistics: global attribute reference for keypoints in sub-blocks. Figure 4 The predicted correlation order between different channels is shown. The Y, Cg, and Co channels are compressed in that order, and the results of the previous compression become the context for the subsequent compression.

[0033] The fused features are input into a pre-trained lightweight multilayer perceptron model. The model dynamically outputs the Laplace distribution parameters of the residual distribution. Figure 5 The network structure and steps of the feature extraction and mapping module are shown. Figure 6 The structure of the internal self-attention module for feature extraction is shown. It combines the key residuals calculated from the preceding points and keypoints with the traversal residuals to perform feature extraction.

[0034] S4. Entropy Coding and Bitstream Generation: Based on the probability distribution parameters obtained from dynamic modeling, entropy coding is performed on the attribute prediction residuals to generate a compressed bitstream.

[0035] Entropy coding employs an arithmetic encoder, utilizing the Laplace probability distribution obtained in step S3 to perform arithmetic coding on the attribute prediction residuals. Due to the high modeling accuracy, the residual distribution is highly concentrated near zero, thereby significantly compressing the number of bits required for coding and ultimately generating a lossless compressed bitstream. Figure 7 The scanning sequence is shown, which is a diagonal raster scan.

[0036] The lossless point cloud attribute compression system based on slice and intra-block prediction of the present invention includes: Slicing and segmentation module: used to reduce the 3D point cloud to a 2D slice window along the principal coordinate axis, divide it into initial blocks of fixed size, and further subdivide the initial blocks into smaller sub-blocks based on local attribute gradients; Prediction module: used to extract key points within the sub-block, determine the attribute traversal order, perform intra-block prediction based on encoded neighboring points, and calculate attribute prediction residuals; Modeling module: Extracts multi-source context features of the current point to be encoded, including traversal history features, spatial neighborhood features, slice scan history features, and key point statistical features, and inputs them into a deep neural network to model the probability distribution of the attribute prediction residual; Encoding module: Used to entropy encode the attribute prediction residuals based on the probability distribution parameters obtained from modeling, and output a compressed bit stream.

[0037] In practical applications, point cloud data is typically massive in scale, with irregular point distribution and non-uniform attribute variations. Therefore, this invention first preprocesses the input point cloud, including coordinate normalization, scaling, and format conversion, to ensure the stability of subsequent processing. Then, Principal Component Analysis (PCA) is used to calculate the variance of the point cloud along each coordinate axis, and the direction with the largest variance is selected as the reference direction for slicing. This strategy maximizes the preservation of the point cloud's geometric and attribute flexibility. Subsequently, the point cloud is divided into multiple slices arranged along the reference direction. Each slice is further divided into two-dimensional grid blocks, thus transforming the three-dimensional spatial structure into a series of more manageable local regions. . To further adapt to the uneven distribution of attributes, this invention introduces an adaptive partitioning mechanism based on attribute gradients within the block: for regions with drastic attribute changes, the system automatically subdivides them into two sub-blocks based on a gradient threshold to capture local texture and edge information. Firstly, for each point within the block, we use neighboring points... Attributes Define the following gradient-based attribute metrics. To quantify the salience of local textures or edges. Gradient thresholding is then applied to identify regions with significant attribute changes, which may correspond to texture edges or geometric boundaries. and These are the mean and standard deviation of the attribute gradients within the block, respectively. Based on the results, the block was divided into two sub-blocks at the location where the attribute gradient changed most significantly. Optimal segmentation location. This is determined by maximizing the difference in average gradient between the two regions. This combination of slicing and block partitioning preserves the spatial continuity of the point cloud and provides a more stable foundation for subsequent prediction and contextual modeling. After obtaining the sub-blocks, this invention proposes a keypoint-guided intra-block prediction strategy. Traditional center point traversal methods often ignore the complex spatial and attribute relationships of point clouds, resulting in large residual variance. This invention selects several key points within the sub-blocks using the farthest point sampling (FPS) method; these key points can better cover the spatial structure and represent attribute features. Next, a residual-aware cost function is constructed, which simultaneously considers attribute value differences. Spatial geometric distance This ensures that the traversal order maintains spatial continuity while minimizing prediction residuals. During the traversal, the system starts from key points and selects the point with the lowest cost compared to the already visited set, until the entire sub-block is covered. When making predictions in this order, the attribute value of each point is inferred from previously encoded points, and the prediction error, i.e., the residual, is recorded. Compared to traditional methods, this key-point-guided prediction mechanism effectively reduces residual fluctuations and improves compression accuracy. After residual generation, accurate modeling becomes crucial for improving compression efficiency. This invention designs a multi-dimensional contextual feature extraction and fusion mechanism, comprehensively utilizing traversal history, cross-slice stacking, scan sequences, and keypoint statistical information to construct rich contextual information. First, the traversal context extracts features reflecting local statistical patterns by analyzing the residual sequence of encoded points. Second, the stacked context combines sub-blocks at the same coordinate position in different slices into a three-dimensional volume, using a sparse convolutional network to extract cross-slice spatial and attribute features. Third, the scan context statistically analyzes the residuals of previous sub-blocks according to the scan order within the slice, forming distribution information reflecting the overall trend of the local area. Finally, the keypoint context provides thumbnail-style statistical features for the current point using the mean, variance, and residual values ​​of keypoints. After the above multi-source contextual features are fused in the network, they can effectively capture the correlation between local and global factors, thus providing accurate priors for residual distribution modeling.

[0038] In the probabilistic modeling stage, this invention employs a Laplace distribution to characterize the statistical properties of the residuals. The fused contextual features are input into a multilayer perceptron (MLP) to generate parameters for the residual distribution. Unlike traditional methods that rely on a fixed model, this contextual conditional modeling dynamically adjusts the prediction results, making the residual distribution more consistent with reality. Finally, arithmetic coding is used to entropy-encode the residuals, generating a compressed bitstream. Since arithmetic coding is theoretically close to optimal, this invention can approximate the information theory limit to the greatest extent possible, achieving efficient and lossless compression of point cloud attributes.

[0039] On the decoding side, the system replicates the same slicing, segmentation, and traversal process as the encoding side. Using the same keypoint selection and context modeling strategies, attribute residuals are recovered point-by-point and superimposed onto the predicted values, thereby reconstructing point cloud attributes that are completely consistent with the original data. This process ensures complete symmetry and lossless operation between compression and decoding.

[0040] To verify the effectiveness of the method of this invention, experiments were conducted on multiple public datasets, including MVUB, 8iVFB, Owlii, and CAT1. The results show that compared with MPEG G-PCC v23, the proposed method improves bitrate efficiency by 38.31%, 14.40%, 17.08%, and 5.12%, respectively. Furthermore, ablation experiments also demonstrate the effectiveness of slice segmentation and keypoint-guided prediction: removing slices and adaptive segmentation significantly reduces compression efficiency; when using traditional centerpoint traversal instead of keypoint prediction, residual fluctuations increase, leading to an increase in compression bitrate. These results indicate that each module proposed in this invention plays a crucial role in improving overall performance.

[0041] Beyond experimental results, this invention has broad application value. In autonomous driving scenarios, vehicles need to collect and transmit point cloud data in real time for environmental perception and path planning. The method of this invention can achieve efficient compression while ensuring attribute integrity, reducing bandwidth consumption in vehicle-to-vehicle communication. In virtual reality and augmented reality, the transmission of high frame rate point cloud video places extremely high bandwidth demands. This method can maintain image quality under limited bandwidth conditions, enhancing the user's immersive experience. In the digitization of cultural heritage, high-precision 3D scans often contain rich texture information. This method can significantly reduce storage requirements, making it suitable for long-term archiving. In fields such as industrial inspection, robot perception, and telemedicine, this method can also provide an efficient and reliable point cloud attribute compression solution.

[0042] From a system implementation perspective, this invention can be deployed on high-performance servers or embedded in terminal devices. Both the encoding and decoding ends can run in GPU-accelerated environments. The training phase is suitable for completion on high-performance graphics cards such as the NVIDIA V100, while the inference phase can be processed in real time on consumer-grade GPUs or CPUs. In terms of software implementation, this method can implement sparse convolution operations based on the PyTorch framework and MinkowskiEngine. The system architecture can run independently or be embedded as a module into existing point cloud processing platforms, exhibiting good engineering adaptability and scalability.

[0043] In summary, this invention establishes a complete lossless point cloud attribute compression framework through slicing and block-based strategies, key-point guided prediction mechanisms, and multi-dimensional context modeling. While ensuring lossless data recovery, it significantly improves compression efficiency and reduces computational complexity, making it suitable for the storage and transmission of large-scale point clouds, and possessing broad application prospects and promotional value.

[0044] The above description represents the preferred embodiment based on the inventive concept and working principle. The above embodiments should not be construed as limiting the scope of protection of these claims; other embodiments and combinations of implementations based on the inventive concept are all within the scope of protection of this invention.

Claims

1. A lossless point cloud attribute compression method based on slice and intra-block prediction, characterized in that, Includes the following steps: S1. Multidimensional slicing and adaptive block division: The input 3D point cloud is reduced to a 2D slice window along the principal coordinate axis and divided into initial blocks of fixed size. The local attribute gradient of the initial block is calculated. According to the gradient distribution, the initial block is further divided into smaller sub-blocks along the attribute gradient edge. S2. Keypoint-guided intra-block prediction: Select salient feature points as keypoints in each sub-block, use keypoints as anchors to determine the traversal order of points in the sub-block, and perform intra-block prediction based on encoded neighboring points to calculate attribute prediction residuals. S3. Multi-source context feature extraction and dynamic modeling: Extract multi-source context features of the current point to be encoded, including traversal history features, spatial neighborhood features, slice scan history features and key point statistical features, and input them into a deep neural network to model the probability distribution of the attribute prediction residual; S4. Entropy Coding and Bitstream Generation: Based on the probability distribution parameters obtained from dynamic modeling, entropy coding is performed on the attribute prediction residuals to generate a compressed bitstream.

2. The lossless point cloud attribute compression method based on slice and intra-block prediction according to claim 1, characterized in that: In the adaptive segmentation step, if the attribute gradient within the initial segment exceeds the adaptive threshold, the initial segment is recursively subdivided along the gradient edge.

3. The lossless point cloud attribute compression method based on slice and intra-block prediction according to claim 1, characterized in that: The traversal order guided by the key points is an ordered prediction chain generated by calculating the spatial distance or geometric correlation between points within a sub-block and key points, and using the shortest path algorithm.

4. The lossless point cloud attribute compression method based on slice and intra-block prediction according to claim 1, characterized in that: The deep neural network employs a multilayer perceptron (MLP) to predict the Laplace distribution parameters of the residual distribution, which include location parameters and scale parameters.

5. The lossless point cloud attribute compression method based on slice and intra-block prediction according to claim 1, characterized in that: The entropy encoding uses an arithmetic encoder.

6. A lossless point cloud attribute compression system based on slicing and intra-block prediction, characterized in that: include: Slicing and segmentation module: used to reduce the 3D point cloud to a 2D slice window along the principal coordinate axis, divide it into initial blocks of fixed size, and further subdivide the initial blocks into smaller sub-blocks based on local attribute gradients; Prediction module: used to extract key points within the sub-block, determine the attribute traversal order, perform intra-block prediction based on encoded neighboring points, and calculate attribute prediction residuals; Modeling module: Extracts multi-source context features of the current point to be encoded, including traversal history features, spatial neighborhood features, slice scan history features, and key point statistical features, and inputs them into a deep neural network to model the probability distribution of the attribute prediction residual; Encoding module: Used to entropy encode the attribute prediction residuals based on the probability distribution parameters obtained from modeling, and output a compressed bit stream.

Images