Point cloud model adaptive rendering method based on cloud edge collaboration

Abstract

The invention relates to the technical field of point cloud model rendering, particularly provides a point cloud model adaptive rendering method based on cloud edge cooperation, and solves the technical problems of inaccurate point cloud data rendering partition, incapability of performing partition storage and edge device distribution on point cloud data and low rendering fineness of a generated three-dimensional point cloud model in the prior art. The method comprises the following steps: performing point cloud data acquisition on a target object by using a three-dimensional laser scanner to obtain point cloud data; obtaining a partition storage point cloud data set, and distributing the partition storage point cloud data set to an edge device set; sending the collected image to an edge device which is interactively associated; performing cross-modal local corresponding perception on the point cloud data set stored in the partition, and performing compensation by calculating the local correlation between the image semantic features and the point cloud features to obtain a compensated point cloud data set stored in the partition; and returning the compensation partition storage point cloud data set to the cloud end for integration rendering to obtain a rendered point cloud model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of point cloud model rendering, in particular to a point cloud model adaptive rendering method based on cloud edge collaboration. BACKGROUND

[0002] With the advancement of three-dimensional laser scanning technology and RGB-D camera technology, point cloud data has been widely used in many fields such as architecture, industry, autonomous driving, etc. Point cloud data is usually collected by three-dimensional laser scanners, and these data have the characteristics of high precision and rich details. However, due to the sparsity and high dimensionality of point cloud data, how to effectively process, store and render point cloud data becomes a challenge. Existing point cloud rendering methods mostly rely on a single computing platform for data processing, but when faced with large-scale point cloud data, traditional methods have limitations in computing efficiency, storage capacity and real-time rendering.

[0003] Prior art one, Chinese patent, application number CN118135080A discloses a laser point cloud model rendering method and system, which includes: acquiring point cloud data collected by a user, constructing a three-dimensional point cloud model based on MapGIS; importing the three-dimensional point cloud model, performing data analysis on the three-dimensional point cloud model by MapGISEarthforUnreal, and rendering the three-dimensional point cloud model based on texture mapping technology and dynamic lighting technology to parameterize the rendering scene of the three-dimensional point cloud model; adding the three-dimensional point cloud model to the rendering scene based on point cloud registration, and visualizing the point cloud model through human-computer interaction. Although the three-dimensional point cloud model is constructed and rendered by MapGIS, it does not mention how to store the point cloud data reasonably, and when processing large-scale point cloud data, the rendering partition may not be accurate, affecting the real-time performance and accuracy of rendering.

[0004] Prior art two, Chinese patent, application number CN114820902A discloses a method and system for rendering a three-dimensional model contour based on point cloud, wherein the method includes the following steps: acquiring medical images, constructing image data of sagittal and coronal planes; directly constructing a contour point cloud structure, and cutting the contour point cloud structure according to the coordinates given by the user; finding the boundary of the contour point cloud obtained by cutting to extract the contour coordinate data of the section; arranging, connecting and rendering the extracted contour to render the contour coordinate data on the image data of the sagittal and coronal planes generated by multi-plane reconstruction, so that doctors can intuitively view the contour under the selected coordinate section. Although this application constructs and renders a contour point cloud, it does not optimize the details and accuracy of the point cloud data, which may result in insufficient detail of the generated three-dimensional point cloud model

[0005] Prior art three, Chinese patent, application number CN120070700A discloses a mobile device digital twin model rendering optimization method, system, terminal and medium, the method comprises the following steps: based on the size of the rendering load, the image rendering task is distributed to the cloud rendering system or the edge node; when the image rendering task is distributed to the cloud rendering system, the cloud rendering system splits the image rendering task into multiple subtasks and distributes each subtask to different computing nodes, the cloud rendering engine on the computing node executes the subtask and outputs the first rendering result, the cloud rendering system collects the first rendering result and integrates it and then outputs it to the mobile device; when the image rendering task is distributed to the edge node, the rendering engine on the edge node executes the image rendering task and outputs the second rendering result, after completing the image rendering task, the edge node outputs the second rendering result to the mobile device. Although this method optimizes the distribution of image rendering tasks through the cloud or edge node, it mainly focuses on the distribution and processing of image rendering tasks, ignores the construction of a single image data source model, and is prone to cause high point cloud model deviation, ignores the construction of a point cloud model through multi-source data, and causes the point cloud model to lack necessary details.

[0006] At present, prior art one, prior art two and prior art three cannot solve the technical problems of inaccurate point cloud data rendering partitioning, inability to store and distribute point cloud data to edge devices, and low rendering precision of generated three-dimensional point cloud models in the prior art. Therefore, the present application provides a point cloud model adaptive rendering method based on cloud-edge collaboration. SUMMARY

[0007] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0008] In one aspect of the present application, a point cloud model adaptive rendering method based on cloud-edge collaboration is provided, which comprises:

[0009] The initial point cloud data of the target object is collected by a three-dimensional laser scanner, and then Gaussian filtering is used to smooth the data and remove noise to obtain the first-order point cloud data after processing. The first-order point cloud data is analyzed for outliers to remove abnormal points and generate point cloud data;

[0010] The point cloud data is uploaded to the cloud for partition storage to obtain a partition storage point cloud data set, and the partition storage point cloud data set is distributed to a set of edge devices, wherein each edge device is at least interactively associated with one RGB-D camera in an RGB-D camera array;

[0011] An RGB-D camera array is arranged around the target object, multi-point image acquisition of the target object is performed by using the RGB-D camera array, and the acquired images are sent to an interactive associated edge device, wherein each RGB-D camera in the RGB-D camera array captures images and depth data in a synchronous manner;

[0012] Through the edge device set, cross-modal local correspondence perception is performed on the partition storage point cloud data set by simultaneously or step by step receiving the acquired images, the partition storage point cloud data set is compensated by calculating the local correlation between image semantic features and point cloud features, and a compensated partition storage point cloud data set is obtained;

[0013] The compensated partition storage point cloud data set is returned to the cloud end through the network for integration processing, the cloud end merges, removes overlaps and gaps according to each part of the data, and generates a rendered three-dimensional point cloud model using a graphics rendering technique.

[0014] In an optional implementation, initial point cloud data of a target object is acquired by a three-dimensional laser scanner, then Gaussian filtering is used to smooth the data and remove noise, first-order point cloud data after processing is obtained, and outlier analysis is performed on the first-order point cloud data to remove abnormal points, and point cloud data is generated, including:

[0015] The three-dimensional laser scanner emits laser beams from different angles to the target object and calculates the time of reflection, measures the distance of the target object surface, and generates initial point cloud data, wherein the initial point cloud data contains spatial coordinate information of each point on the surface of the target object;

[0016] A preset Gaussian filtering window is obtained, a distance weighted average value of each point in the initial point cloud data and the neighborhood points in the preset Gaussian filtering window is calculated for high-frequency noise smoothing, and first-order processing point cloud data is obtained, wherein the preset Gaussian filtering window is set according to the density and noise characteristics of the initial point cloud data;

[0017] The distance average value of each point in the first-order processing point cloud and K nearest neighbor points is calculated, and it is judged whether the distance average value calculation result exceeds a set threshold value, if yes, it is determined as an outlier and removed, and point cloud data is obtained.

[0018] In an optional implementation, the point cloud data is uploaded to the cloud end for partition storage, a partition storage point cloud data set is obtained, and the partition storage point cloud data set is distributed to an edge device set, including:

[0019] The edge device set is traversed to collect data according to preset device performance indicators, the collected data is arranged in a standardized format, and an edge device performance indicator set is obtained, wherein the preset device performance indicators at least include computing power, storage capacity and network bandwidth;

[0020] A pre-constructed performance coefficient identifier is used to input the edge device performance indicator set into the performance coefficient identifier to obtain an edge device performance coefficient set;

[0021] According to the edge device performance coefficient set and the preset processable point cloud data volume, an edge device processable point cloud data volume set is determined, the point cloud data is structurally divided based on the edge device processable point cloud data volume set, and the division result is stored in the cloud in a partitioned manner to obtain a partitioned storage point cloud data set;

[0022] According to the similarity between the edge device processable point cloud data volume set and the partitioned storage point cloud data set, a one-to-one mapping relationship between the edge device and the partitioned storage point cloud data is constructed, and the partitioned storage point cloud data set is distributed to the edge device set according to the mapping relationship.

[0023] In an optional implementation, a pre-constructed performance coefficient identifier includes:

[0024] A performance coefficient identifier is constructed, wherein the performance coefficient identifier is trained based on a convolutional neural network, the input layer of the network includes edge device performance indicators of each edge device, collected sample edge device performance indicators and sample edge device performance coefficients are used as training data and are labeled to obtain labeled training data;

[0025] The convolutional neural network is trained, the labeled training data is input into the convolutional neural network, and the data is processed by each layer of the convolutional neural network to generate a prediction result, i.e., a predicted edge device performance coefficient of each edge device;

[0026] The difference between the predicted edge device performance coefficient and the labeled training data is calculated by a loss function, the weights of the convolutional neural network are updated by a back propagation algorithm, the error is transmitted back to each layer, and the parameters of each layer are adjusted until a preset number of times are met, and a trained performance coefficient identifier is obtained.

[0027] In an optional implementation, the partitioned storage point cloud data set is subjected to cross-modal local correspondence perception by the edge device set in combination with simultaneously or step-by-step received collected images, the partitioned storage point cloud data set is compensated by calculating the local correlation between image semantic features and point cloud features, a compensated partitioned storage point cloud data set is obtained, including:

[0028] The images received by the edge device set are resized, pixel value normalized, and rotated and flipped to obtain preprocessed images, and a pre-trained image semantic recognizer is used to extract features from the images received by the edge device set to obtain an image semantic feature set;

[0029] The point cloud data is divided into multiple local regions, and a matching window is set for each local region, wherein the window size is adjusted according to the density of the point cloud data and the resolution of the image features; the image semantic feature set and the partitioned stored point cloud data set are locally correlated by using cosine similarity, and a cloud map matching point set is determined;

[0030] Based on the cloud map matching point set, the images collected by the edge device set are matched and compensated with the corresponding partitioned stored point cloud data in the partitioned stored point cloud data set to obtain the compensated partitioned stored point cloud data set.

[0031] In an optional implementation, the cloud map matching point set is determined by locally correlating the image semantic feature set and the partitioned stored point cloud data set using cosine similarity, and includes:

[0032] For the image semantic feature set and the partitioned stored point cloud data set of each local region, a cosine similarity value is calculated, and if the cosine similarity value is greater than or equal to a preset similarity threshold, it is added to the initial cloud map matching point set;

[0033] The initial cloud map matching point set is taken as input, and an initial cloud map matching point is selected as a source point cloud, and another initial cloud map matching point is selected as a target point cloud, wherein the source point cloud is a point cloud that needs to be transformed, and the target point cloud is data that needs to be aligned with the source point cloud;

[0034] For each point in the source point cloud, the Euclidean distance is used to find the nearest point in the target point cloud, the best rigid transformation from the source point cloud to the target point cloud is calculated, including a rotation matrix and a translation vector, and the least squares method is used to minimize the squared error between the point clouds, and the calculated rotation matrix and translation vector are applied to the source point cloud to obtain the transformed point cloud. After multiple iterations, an optimized cloud map matching point set is obtained.

[0035] In an optional implementation, the compensated partitioned stored point cloud data set is obtained, including:

[0036] Based on the cloud map matching point set, the images collected by the edge device set are matched and compensated with the corresponding partitioned stored point cloud data in the partitioned stored point cloud data set to obtain the compensated partitioned stored point cloud data set.

[0037] Fuse the fusion partition storage point cloud data set with the partition storage point cloud data set to obtain the compensation partition storage point cloud data set.

[0038] In an optional implementation, the compensated partition storage point cloud data set is transmitted back to the cloud through a network for integrated processing, including:

[0039] The compensated partition storage point cloud data set is transmitted back to the cloud through a data transmission protocol, the cloud initializes each data partition by removing redundant data, merges the received data, integrates according to spatial information, and obtains an initial merged point cloud.

[0040] The initial merged point cloud is subjected to overlap region detection to obtain an overlap region set, and the ICP algorithm is used to remove the overlap between different data blocks in the overlap region set, and an interpolation method is used to fill the gap region in the point cloud data to complete the integrated processing.

[0041] In an optional implementation, the initial merged point cloud is subjected to overlap region detection to obtain an overlap region set, including:

[0042] The initial merged point cloud is subjected to voxel segmentation according to a preset voxel size, and each voxel after segmentation is subjected to overlap identification to determine whether there are points of other voxels, and if so, the region is identified as an overlap region to obtain a first overlap region set.

[0043] The initial merged point cloud is subjected to normal vector calculation, and the similarity of each pair of normal vectors obtained by calculation is identified, and if the identification result exceeds a preset threshold, the region where the normal vector is located is identified as a second overlap region to obtain a second overlap region set, and the first overlap region set and the second overlap region set are summarized to obtain an overlap region set.

[0044] In an optional implementation, a rendered three-dimensional point cloud model is generated using a graphics rendering technology, including:

[0045] A rendering API is selected, and a rendering context is initialized, GPU memory and a rendering buffer are allocated, a rendering target is configured as a window or a frame buffer, and a rendering viewport is set to complete rendering environment initialization.

[0046] In the initialized rendering environment, the integrated point cloud is converted into a triangular mesh format and stored as a vertex buffer object, the spatial coordinates of each point are converted into screen space coordinates through a vertex shader, a texture image is mapped to the surface of the point cloud based on UV mapping, and the three-dimensional point cloud model is obtained.

[0047] One or more technical solutions provided in the present application have at least the following technical effects or advantages

[0048] The application collects initial point cloud data of a target object by a three-dimensional laser scanner, then smoothes the data by Gaussian filtering to remove noise and obtain first-order point cloud data after processing, removes abnormal points from the first-order point cloud data by outlier analysis to generate point cloud data; uploads the point cloud data to the cloud for partition storage to obtain a partition storage point cloud data set, and distributes the partition storage point cloud data set to a set of edge devices, wherein each edge device is at least interactively associated with one RGB-D camera in an RGB-D camera array; arranges an RGB-D camera array around the target object, uses the RGB-D camera array to collect multi-point images of the target object, and sends the collected images to the interactively associated edge device, wherein each RGB-D camera in the RGB-D camera array captures images and depth data in a synchronous manner; through the set of edge devices, the collected images are received simultaneously or step by step to perform cross-modal local correspondence perception on the partition storage point cloud data set, compensate the partition storage point cloud data set by calculating the local correlation between image semantic features and point cloud features, and obtain a compensated partition storage point cloud data set; the compensated partition storage point cloud data set is returned to the cloud through a network for integration processing, the cloud merges, removes overlaps and gaps according to each part of the data, and uses a graphics rendering technology to generate a rendered three-dimensional point cloud model. The effective fusion of images and point cloud data is achieved, and the representation of point cloud data is optimized through local alignment and weighted fusion, so that the features between images and point cloud data can be better complementary, and the technical effect of improving the rendering quality of the point cloud model is achieved. BRIEF DESCRIPTION OF DRAWINGS

[0049] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, illustrate the application together with the embodiments thereof, and explain the application without limiting the application. In the drawings:

[0050] Figure 1 A cloud-edge collaborative point cloud model adaptive rendering method flowchart is provided in embodiment 1 of the application.

[0051] Figure 2 A process diagram flowchart for generating point cloud data is provided in embodiment 2 of the application.

[0052] Figure 3 A process diagram for obtaining a partition storage point cloud data set and distributing the partition storage point cloud data set to a set of edge devices is provided in embodiment 3 of the application.

[0053] Figure 4 A process diagram for obtaining a compensated partition storage point cloud data set is provided in embodiment 5 of the application.

[0054] Figure 5 A process diagram for determining a cloud map matching point set provided in Embodiment 6 of the present application;

[0055] Figure 6 A block diagram of an electronic device provided by the present application;

[0056] Figure 7 A block diagram of a computer readable storage medium provided by the present application. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments.

[0058] Hereinafter, the terms "first", "second", and the like are only used for description convenience, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second", and the like can explicitly or implicitly include one or more of the features. In the description of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more.

[0059] In the present application, unless otherwise explicitly specified and limited, the term "connection" should be understood broadly, for example, the "connection" can be a fixed mechanical connection, or a detachable mechanical connection, or integrated; or the "connection" can be direct connection, or indirect connection through an intermediate medium. In addition, unless otherwise explicitly specified and limited, the term "coupling" should be understood broadly, for example, the "coupling" can be direct electrical connection, for example, physical contact and electrical conduction between two components, or can be understood as electrical connection between different components through solid lines such as copper foil or wire of printed circuit board (PCB) in line structure, to transmit electrical signals; or the "coupling" can be indirect electrical connection between two components through an intermediate medium; or the "coupling" can be electrical connection between two components in a non-contact / empty space manner, for example, electrical connection between two components in a capacitive coupling manner to transmit electrical signals.

[0060] In the embodiments of the present application, the orientation terms such as "up", "down", "left", "right", and the like can include but not limited to the orientation defined by the relative position of the components in the drawings, and it should be understood that these directional terms can be relative concepts, which are used for relative description and clarification, and can be changed accordingly according to the change of the position of the components in the drawings.

[0061] Embodiment 1:

[0062] AsFigure 1 As shown, the embodiment of the present application provides a method comprising the following steps:

[0063] Step S100: Collecting initial point cloud data of a target object by a three-dimensional laser scanner, then performing smoothing processing on the data by using Gaussian filtering to remove noise and obtain first-order point cloud data after processing, performing outlier analysis on the first-order point cloud data to remove abnormal points, and generating point cloud data;

[0064] Step S200: uploading the point cloud data to the cloud for partition storage, obtaining a partition storage point cloud data set, and distributing the partition storage point cloud data set to a set of edge devices, wherein each edge device is at least interactively associated with one RGB-D camera in an RGB-D camera array;

[0065] Step S300: arranging an RGB-D camera array around the target object, using the RGB-D camera array to collect multi-point images of the target object, and sending the collected images to the interactively associated edge device, wherein each RGB-D camera in the RGB-D camera array captures images and depth data in a synchronous manner;

[0066] Step S400: using the set of edge devices to perform cross-modal local correspondence perception on the partition storage point cloud data set by simultaneously or step by step receiving the collected images, compensating the partition storage point cloud data set by calculating the local correlation between image semantic features and point cloud features, and obtaining a compensated partition storage point cloud data set;

[0067] Step S500: returning the compensated partition storage point cloud data set to the cloud for integration processing through a network, merging, removing overlaps and gaps according to each part of the data in the cloud, and generating a rendered three-dimensional point cloud model using a graphics rendering technique.

[0068] In the above embodiments, a three-dimensional laser scanner is used to scan the target object and collect initial point cloud data. The laser scanner calculates the distance of the object surface by emitting a laser beam and measuring its reflection time, generates spatial coordinate (x, y, z) data points, and forms the initial point cloud data. Gaussian filtering technology is applied to the initial point cloud data. Gaussian filter is a weighted average filter that smoothes the data by weighted average of the neighborhood points of each point in the point cloud, and removes high-frequency noise caused by measurement error or environmental interference. The result of filtering is to reduce the influence of noise points and improve the quality of point cloud data. After Gaussian filtering, the obtained point cloud data is called first-order processing point cloud data. This data has removed most of the high-frequency noise, but there may still be some abnormal points that need further processing. Outlier analysis is performed on the first-order processed point cloud data, and abnormal points that do not conform to the data distribution rule are identified and removed by outlier detection methods such as K-neighbor analysis, statistical methods, etc.

[0069] Further, the point cloud data is uploaded to the cloud through the network, and then the point cloud data is stored in partitions. The process of partition storage divides the point cloud data into multiple small subsets, each subset is stored independently, ensuring the efficiency of data management. After that, these partitioned and stored data are distributed to the edge device set. Each edge device is interactively associated with at least one RGB-D camera in the RGB-D camera array to provide support for subsequent image data processing and point cloud data compensation.

[0070] The RGB-D camera array simultaneously or continuously collects multiple images and depth data from different angles and positions of the target object, and sends the images collected by each RGB-D camera to the edge device associated with it. Each edge device is associated with at least one camera in the RGB-D camera array and is responsible for receiving, processing and transmitting image data. First, the RGB-D camera array is reasonably arranged around the target object to ensure that images and depth information of the target object are collected from multiple angles. Each RGB-D camera captures image and depth data in a synchronized manner, covering all aspects of the target object. Each RGB-D camera array captures image data (including RGB images and depth maps) of the target object in real time. The edge device can simultaneously or gradually receive data from multiple cameras for processing and storage according to the network bandwidth and computing power of the device. Through multi-point image acquisition, the edge device can obtain image data from different angles, achieving the technical effect of providing necessary input data for subsequent image and point cloud data fusion and cross-modal local correspondence perception.

[0071] The edge device receives and processes image data collected by the RGB-D camera array, and performs local feature matching in combination with the point cloud data set. By calculating the local correlation between image semantic features and point cloud features, the image data is used to compensate and refine the point cloud data, and the compensated partition storage point cloud data set is finally obtained. The compensated partition storage point cloud data set is returned to the cloud through the network, and the cloud merges, removes overlaps and gaps, and uses graphics rendering technology to generate the final three-dimensional point cloud model. Rendering techniques include ray tracing, surface reconstruction and other methods to ensure the accuracy and visual effect of the three-dimensional model.

[0072] In summary, through the cloud partition storage and edge device distribution strategy, point cloud data can be reasonably allocated according to device performance and network bandwidth, significantly improving data processing efficiency and reducing latency. Using the RGB-D camera array to synchronously collect image and depth data, and combining the cross-modal local correspondence perception method, the point cloud data can be more accurately aligned, improving the fineness and accuracy of the three-dimensional point cloud model. Through cross-modal feature matching and compensation, the point cloud data is not only spatially aligned, but also enhanced in detail, effectively reducing errors caused by data loss or noise, and obtaining more realistic and detailed point cloud data.

[0073] Embodiment 2:

[0074] As shown in Figure 2 On the basis of embodiment 1, in step S100 provided by the embodiment of the application, the initial point cloud data of the target object is collected by a three-dimensional laser scanner, and then Gaussian filtering is used to smooth the data and remove noise, obtaining the first-order point cloud data after processing. The first-order point cloud data is analyzed for outliers to remove abnormal points, and the point cloud data is generated, including the following steps:

[0075] Step S101: The three-dimensional laser scanner emits laser beams from different angles to the target object and calculates the time of reflection, measures the distance of the target object surface, and generates initial point cloud data, wherein the initial point cloud data contains the spatial coordinate information of each point on the surface of the target object;

[0076] Step S102: Obtain a preset Gaussian filter window, calculate the distance weighted average value of each point in the initial point cloud data and the neighborhood points in the preset Gaussian filter window for high-frequency noise smoothing, and obtain the first-order processed point cloud data, wherein the preset Gaussian filter window is set according to the density and noise characteristics of the initial point cloud data;

[0077] Step S103: traversing the distance average of each point in the first-order processing point cloud and K nearest neighbors, and judging whether the distance average calculation result exceeds the set threshold, if yes, it is determined as an outlier and removed, and the point cloud data is obtained.

[0078] In the above embodiment, the three-dimensional laser scanner measures the distance of the object surface by emitting a laser beam and calculating the time of the laser beam reflection, thereby generating point cloud data with high precision. By scanning the target object with the three-dimensional laser scanner, point cloud data reflecting the surface information of the target object is obtained. The three-dimensional laser scanner emits a laser beam to the target object and receives the returned laser beam to obtain initial point cloud data of the target object, wherein the initial point cloud data is the original data collected, usually containing the spatial coordinate information (x, y, z) of each point on the surface of the target object, but these data may contain noise or inaccurate points, which need to be denoised and analyzed for outliers, thereby obtaining processed point cloud data.

[0079] Preferably, the Gaussian filter window is pre-set by a person skilled in the art, and the window size usually depends on the density and noise characteristics of the point cloud data. The points within the window will be used for smoothing processing. According to the distance weighted average of each point in the point cloud data and the neighborhood points within the Gaussian filter window, high-frequency noise is smoothed out. According to the distance to each point, the weight gradually decreases from small to large, so that the noise points far from the target area are weakened. Thus, the first-order processing point cloud data after filtering is obtained. By smoothing out error points caused by device errors or environmental interference, the technical effect of preparing for subsequent outlier removal, feature extraction and other data processing steps is achieved. By calculating the K-neighbor distance of each point, it is judged whether the point far from other points is an outlier. When the K-neighbor distance is greater than or equal to the pre-set distance, it is defined as an outlier, which is removed from the first-order processing point cloud data, and the point cloud data is obtained.

[0080] For example, assuming that the density of the initial point cloud data is 0.1 points per square meter, i.e. about 0.1 points per square meter, and there is a certain noise in the measurement. In order to adapt to this data characteristic, we choose the radius of the Gaussian filter window to be 0.05 meters, i.e. the window will consider the neighborhood points within 0.05 meters around each point for weighted average. For each point in the point cloud, the distance between the point and other points in its neighborhood is calculated, and the weight of each neighborhood point is calculated according to the distance. Then, by weighted average, the smoothing value of the point is calculated, thereby eliminating noise.

[0081] In summary, by using a three-dimensional laser scanner to obtain initial point cloud data, combined with Gaussian filtering and smoothing processing, high-frequency noise caused by device errors, environmental interference, etc. is effectively removed, improving the accuracy and quality of the point cloud data. The filtered point cloud data retains the geometric features of the object surface and removes unnecessary noise, providing a clean data foundation for subsequent processing. Through outlier rejection, abnormal points that do not conform to the data distribution rule are removed, further optimizing the quality of the point cloud data and ensuring the reliability of the data. Finally, after noise removal and outlier rejection, the point cloud data is more accurate and stable, providing accurate support for subsequent three-dimensional modeling, point cloud registration, and surface reconstruction, improving the efficiency and accuracy of the entire data processing process. These technical effects ensure the high availability and high accuracy of point cloud data in applications, providing reliable technical support for high-precision three-dimensional analysis, virtual reality, and augmented reality fields.

[0082] Example 3:

[0083] As Figure 3 shown, based on example 2, the step S200 provided by the embodiment of the application uploads the point cloud data to the cloud for partition storage, obtains a partition storage point cloud data set, and distributes the partition storage point cloud data set to a set of edge devices. The process includes the following steps:

[0084] Step S201: Traverse the set of edge devices to collect according to the preset device performance indicators, arrange the collected data into a standardized format, and obtain a set of edge device performance indicators, wherein the preset device performance indicators at least include computing power, storage capacity and network bandwidth;

[0085] Step S202: Pre-construct a performance coefficient identifier, input the set of edge device performance indicators into the performance coefficient identifier, and obtain a set of edge device performance coefficients;

[0086] Step S203: According to the edge device performance coefficient in the set of edge device performance coefficients and the preset processable point cloud data volume, determine a set of edge device point cloud data processing volumes, divide the point cloud data based on the set of edge device point cloud data processing volumes, and store the division result in the cloud for partition storage, to obtain a set of partition storage point cloud data;

[0087] Step S204: According to the similarity between the set of edge device point cloud data processing volumes and the set of partition storage point cloud data, construct a one-to-one mapping relationship between the edge device and the partition storage point cloud data, and distribute the set of partition storage point cloud data to the set of edge devices according to the mapping relationship.

[0088] In the above embodiments, by analyzing the performance conditions of different edge devices, the scale for dividing the point cloud data is determined, and then the point cloud data is divided and uploaded to the cloud to obtain the partitioned storage point cloud data set. The performance coefficient identifier is a functional module for automatically identifying and calculating the performance coefficient according to the performance indicators of the edge device. It is supervised trained by a convolutional neural network (CNN), thereby effectively evaluating the performance of each device and determining the point cloud data partition that the device can handle. The size of the device performance coefficient reflects the actual ability of the edge device to process point cloud data, and the larger the device performance coefficient, the higher the ability of the corresponding edge device to process point cloud data. The performance indicators of each device are collected by traversing the edge device set to obtain the edge device performance coefficient set, wherein the preset device performance indicators include at least computing power, storage capacity and network bandwidth. The edge device performance indicator set is identified by the performance coefficient identifier to obtain the edge device performance coefficient set. Based on the size of the edge device performance coefficient set, the point cloud data is divided into data blocks suitable for processing by different edge devices. The cloud intelligently distributes data to appropriate edge devices for processing based on these divided data and the performance coefficients of the edge devices. This ensures efficient allocation of processing tasks and avoids the concentration of computing burden, thereby improving processing efficiency.

[0089] In one possible embodiment, the performance data of each edge device is collected, including multiple indicators such as computing power, storage capacity, network bandwidth, etc., to construct the sample edge device performance indicator set. According to the actual device capabilities and the requirements of the processing task, the corresponding sample edge device performance coefficient set is obtained. Based on the convolutional neural network, the framework is constructed, including the input layer, convolutional layer, pooling layer and fully connected layer. The sample edge device performance indicator set and sample edge device performance coefficient set are used to supervise the training of the framework until the training converges, and the trained performance coefficient identifier is obtained. By intelligently evaluating the processing capacity of each edge device and optimizing the distribution of point cloud data and task allocation based on its performance coefficient, the technical effect of improving the subsequent point cloud model rendering efficiency is achieved.

[0090] Through the pre-constructed performance coefficient identifier, the edge device performance indicator set is input into the identifier to obtain the edge device performance coefficient set, thereby evaluating the processing capacity of each device. Based on the performance coefficient, the system can determine the amount of processable point cloud data for each device, and then divide the point cloud data structure and store it in the cloud for partitioning, ensuring efficient management and storage of data. Next, according to the similarity between the edge device processing capacity and the partitioned storage point cloud data set, the system constructs the mapping relationship between the devices and the data, and intelligently distributes the data to appropriate edge devices for processing. This process optimizes the allocation of point cloud data, avoids device overload, improves data processing efficiency and reduces latency.

[0091] Embodiment 4:

[0092] Based on embodiment 3, the process of pre-constructing the performance coefficient identifier in step S202 provided by the embodiment of the application includes the following steps:

[0093] Step S2021: constructing a performance coefficient identifier, wherein the performance coefficient identifier is trained based on a convolutional neural network, the input layer of the network includes edge device performance indicators of each edge device; the collected sample edge device performance indicators and sample edge device performance coefficients are used as training data and are labeled to obtain labeled training data;

[0094] Step S2022: training the convolutional neural network, inputting the labeled training data into the convolutional neural network, and processing the data through each layer of the convolutional neural network to generate a prediction result, i.e., a predicted edge device performance coefficient of each edge device;

[0095] Step S2023: calculating the difference between the predicted edge device performance coefficient and the labeled training data through a loss function, updating the weights of the convolutional neural network through a backpropagation algorithm, transmitting the error back to each layer, and adjusting the parameters of each layer until a preset number of times is met, and obtaining a trained performance coefficient identifier.

[0096] In the above embodiment, by constructing a performance coefficient identifier based on a convolutional neural network (CNN), the system can accurately evaluate the processing capacity of each edge device. In step S2021, the collected edge device performance indicators and labeled performance coefficients are used as training data and are input into the convolutional neural network to generate a model capable of predicting device performance coefficients after training. This process ensures efficient evaluation of device performance and provides a reliable basis for subsequent data processing and distribution.

[0097] During the training process, the convolutional neural network continuously adjusts network parameters through learning of the labeled training data to improve prediction accuracy. In step S2022, the labeled data is input into the network to calculate the performance coefficient of each device, and the result is gradually optimized. In step S2023, the error between the predicted value and the actual value is calculated through a loss function, and the network weights are adjusted through backpropagation until the model converges. The finally obtained performance coefficient identifier can accurately evaluate the processing capacity of the edge device, and improve the accuracy and efficiency of resource allocation.

[0098] This process realizes automated evaluation of edge device processing capacity, optimizes data distribution strategies, and provides the system with flexible and efficient data processing and load balancing capabilities.

[0099] Embodiment 5:

[0100] As Figure 4 As shown in FIG. 1, based on the embodiment 1, the embodiment of the present application provides a process for acquiring a compensated partition storage point cloud data set by combining the received images through the edge device set in step S400, and performing cross-modal local correspondence perception on the partition storage point cloud data set, and compensating the partition storage point cloud data set by calculating the local correlation of the image semantic features and the point cloud features, and obtaining the compensated partition storage point cloud data set, including the following steps:

[0101] Step S401: Adjusting the size of the received image, normalizing the pixel value, and rotating and flipping the image to obtain a preprocessed image; using a pre-trained image semantic recognizer to extract features from the received image to obtain an image semantic feature set;

[0102] Step S402: Dividing the point cloud data into multiple local regions, and setting a matching window for each local region, wherein the window size is adjusted according to the density of the point cloud data and the resolution of the image features; using cosine similarity to match the local correlation of the image semantic feature set and the partition storage point cloud data set to determine a cloud map matching point set;

[0103] Step S403: Based on the cloud map matching point set, matching and compensating the image collected by the edge device set with the corresponding partition storage point cloud data in the partition storage point cloud data set to obtain the compensated partition storage point cloud data set.

[0104] In the above embodiment, first, the received image is adjusted in size, normalized in pixel value, and the image diversity is increased through rotation, flipping and other data enhancement methods, thereby improving the feature extraction capability of the image semantic recognizer. These extracted semantic feature sets provide accurate semantic information for subsequent point cloud matching. The point cloud data is divided into local regions according to the density and the resolution of the image features, and the matching window size is adjusted to ensure that the features of each local region are more fine. The local correlation of the image semantic features and the point cloud data is matched through cosine similarity to ensure accurate alignment in space. This step realizes the effective fusion of the image and the point cloud data, and improves the expression ability and accuracy of the data. Based on the cloud map matching point set, the image and the point cloud data are matched and compensated to further optimize the accuracy and details of the point cloud data. Finally, the compensated point cloud data set is more realistic and detailed, providing high-quality data support for subsequent three-dimensional modeling, reconstruction and visualization tasks. This process effectively improves the quality, processing efficiency and resource utilization of the point cloud data, and enhances the performance of the overall system.

[0105] Embodiment 6:

[0106] As Figure 5As shown, based on Embodiment 5, the process of utilizing cosine similarity to perform local correlation matching on the image semantic feature set and the partition-stored point cloud data set in step S402 to determine the cloud map matching point set provided by the present embodiment comprises the following steps:

[0107] Step S4021: For the image semantic feature set of each local region and the partition-stored point cloud data set, a cosine similarity value is calculated, and if the cosine similarity value is greater than or equal to a preset similarity threshold, the image semantic feature set is added to the initial cloud map matching point set;

[0108] Step S4022: The initial cloud map matching point set is taken as input, and one initial cloud map matching point is selected as a source point cloud and another initial cloud map matching point is selected as a target point cloud, wherein the source point cloud is a point cloud that needs to be transformed, and the target point cloud is data to be aligned with the source point cloud;

[0109] Step S4023: For each point in the source point cloud, the Euclidean distance is used to find the nearest point in the target point cloud, the best rigid transformation from the source point cloud to the target point cloud is calculated, including a rotation matrix and a translation vector, and the least squares method is used to minimize the squared error between the point clouds, the calculated rotation matrix and translation vector are applied to the source point cloud to obtain a transformed point cloud, and after multiple iterations, an optimized cloud map matching point set is obtained.

[0110] In the above embodiment, the cosine similarity value is used to calculate the correlation between the image semantic features and the point cloud data, and when the similarity reaches the preset threshold, it is added to the initial cloud map matching point set, ensuring the accuracy of the matching points. The source point cloud and the target point cloud are selected for point cloud registration. The Euclidean distance and the least squares method are used to calculate the best rigid transformation between the source point cloud and the target point cloud, including a rotation matrix and a translation vector, to ensure accurate alignment of the point cloud data. After multiple iterations, the optimized cloud map matching point set can effectively reduce errors and improve the accuracy and details of the point cloud data. This process significantly improves the matching accuracy of the image and the point cloud data and optimizes the geometric consistency of the point cloud.

[0111] Embodiment 7:

[0112] Based on Embodiment 5, the process of obtaining the compensation partition-stored point cloud data set in step S403 provided by the present embodiment comprises the following steps:

[0113] Step S4031: Based on the cloud map matching point set, the image collected by the edge device set is locally aligned with the corresponding partition-stored point cloud data in the partition-stored point cloud data set, and the image semantic features and the partition-stored point cloud data are respectively weighted and fused to obtain a fused partition-stored point cloud data set;

[0114] Step S4032: merging the fusion partition storage point cloud data set with the partition storage point cloud data set to obtain the compensation partition storage point cloud data set.

[0115] In the above embodiment, based on the cloud image matching point set, the image and the point cloud data are accurately matched through local alignment, ensuring that the semantic features of the image can be correctly mapped to the point cloud data. By weighting and fusing the image semantic features and the point cloud data, the texture, color and geometric information of the image data are effectively enhanced, and the expressiveness of the point cloud data is improved. Assuming that the weighting fusion process used gives a weight of 0.7 to the point cloud data and a weight of 0.3 to the image features, this ensures that the geometric features of the point cloud data are preferentially maintained, while the details of the image are supplemented. Finally, the obtained fusion partition storage point cloud data set has high accuracy, and the error is reduced by about 15%.

[0116] The merged partition storage point cloud data and the original point cloud data set are merged to remove redundant and inconsistent parts and optimize the spatial structure of the data. This processing further improves the quality of the data, and the final generated compensation partition storage point cloud data set has higher precision, ensuring the consistency of the final three-dimensional point cloud model in geometry and texture, and is suitable for high-precision three-dimensional modeling, virtual reality, augmented reality and other application fields, and improves the rendering effect by about 20%.

[0117] Embodiment 8:

[0118] On the basis of embodiment 1, the process of step S500 provided by the embodiment of the application for returning the compensated partition storage point cloud data set to the cloud for integration processing through the network includes the following steps:

[0119] Step S501: return the compensated partition storage point cloud data set to the cloud through a data transmission protocol, initialize each data partition by removing redundant data in the cloud, merge the received data, integrate according to the spatial information, and obtain an initial merged point cloud.

[0120] Step S502: perform overlap area detection on the initial merged point cloud to obtain an overlap area set, remove the overlap between different data blocks in the overlap area set using an ICP algorithm, and fill the gap area in the point cloud data using an interpolation method to complete the integration processing.

[0121] In the above embodiments, the compensated partitioned point cloud data is transmitted back to the cloud through the data transmission protocol, the cloud removes redundant data and initializes each data partition, and finally merges the data and generates an initial merged point cloud. For example, through spatial information merging, the error of the initial merged point cloud is reduced by about 10%. In step S502, the overlapping region detection method is used to accurately identify the overlapping regions between different data blocks, and the ICP algorithm is used to remove the overlapping parts to ensure the geometric consistency of the data. After ICP processing, the overlapping error is reduced by about 15%. For the generated gap region, an interpolation method is used to fill it, ensuring the integrity of the point cloud. After these processes, the final generated point cloud data is more accurate and complete, and the point cloud model accuracy is improved by about 20%.

[0122] Embodiment 9:

[0123] Based on embodiment 8, the process of detecting overlapping regions in the initial merged point cloud in step S502 to obtain the set of overlapping regions includes the following steps:

[0124] Step S5021: The initial merged point cloud is voxelized according to a preset voxel size, and each voxel after segmentation is identified for overlapping to determine whether there are points of other voxels. If yes, it is identified as an overlapping region to obtain a first set of overlapping regions.

[0125] Step S5022: The normal vector of the initial merged point cloud is calculated, and the similarity between each two normal vectors is identified. If the identification result exceeds a preset threshold, the region where it is located is identified as a second overlapping region to obtain a second set of overlapping regions. The first set of overlapping regions and the second set of overlapping regions are summarized to obtain a set of overlapping regions.

[0126] In the above embodiments, in step S5021, the initial merged point cloud is voxelized according to a preset voxel size, and the points in each voxel are identified according to the overlapping condition with other voxel points. If there are overlapping points, they are identified as overlapping regions. Assuming that the voxel size is 0.01 meters, this method can effectively identify the preliminary overlapping regions in the point cloud data. For example, about 12% of the point cloud data is identified as overlapping regions through voxel segmentation, ensuring efficient processing of the data. In step S5022, the normal vector of each point in the initial merged point cloud is calculated, and the similarity between each two normal vectors is calculated to identify the regions with similarity exceeding a preset threshold, and further identify the second overlapping regions. This method can eliminate the false matching caused by local noise or external interference, and optimize the identification process of the overlapping regions. Finally, by summarizing the first set of overlapping regions and the second set of overlapping regions, a complete set of overlapping regions is obtained, improving the accuracy and integration quality of the point cloud data and providing more accurate data support for subsequent processing.

[0127] Embodiment 10:

[0128] On the basis of Embodiment 1, the process of generating the rendered three-dimensional point cloud model in step S500 provided by the embodiment of the application comprises the following steps:

[0129] Step S503: Select a rendering API, and initialize a rendering context, allocate GPU memory and a rendering buffer, configure a rendering target as a window or a frame buffer, and set a rendering viewport, to complete rendering environment initialization;

[0130] Step S504: In the initialized rendering environment, convert the integrated point cloud into a triangular mesh format, and store it as a vertex buffer object, convert the spatial coordinates of each point into screen space coordinates through a vertex shader, map a texture image to the surface of the point cloud based on UV mapping, and obtain the three-dimensional point cloud model.

[0131] In the above embodiment, the rendering context is initialized, 4GB of GPU memory and a rendering buffer are allocated, the rendering target is configured as a frame buffer, the rendering viewport is set to a resolution of 1920x1080, the stability and efficiency of the rendering environment are ensured, and the graphics processing speed is improved. In step S504, the integrated point cloud data is converted into a triangular mesh format and stored as a vertex buffer object (VBO), ensuring efficient access to the data in the GPU. Through the vertex shader, the coordinates of each point are converted into screen space coordinates, and based on the UV mapping technology, a texture image is mapped to the surface of the point cloud, improving the visual effect of the point cloud.

[0132] Figure 6 A block diagram of an exemplary electronic device suitable for implementing an embodiment of the application is shown.

[0133] The electronic device can include a central processor / microprocessor / master control chip, etc. 4; a storage medium 5 coupled to the central processor / microprocessor / master control chip, etc. 4, and storing computer executable instructions therein for performing the steps of the various methods of the embodiments of the application when executed by the processor.

[0134] The central processor / microprocessor / master control chip, etc. 4 can include, but is not limited to, for example, one or more processors or microprocessors, etc.

[0135] The storage medium 5 can include, but is not limited to, for example, random access memory (RAM), read only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, computer storage media (such as hard disks, floppy disks, solid state disks, removable disks, CD-ROMs, DVD-ROMs, Blu-ray discs, etc.).

[0136] In addition, the electronic device can include, but is not limited to, a data bus 6, an input / output bus / external bus / device bus, etc. 7, a display 8, and an input / output device 9 (e.g., a keyboard, a mouse, a speaker, etc.), etc.

[0137] The central processing unit / microprocessor / master control chip, etc. 4 can communicate with external devices (8, 9, etc.) via a wired or wireless network (not shown) through the I / O bus 7.

[0138] The storage medium 5 can also store at least one computer-executable instruction for performing the steps of the various functions and / or methods in the embodiments described in the present technology when executed by the central processing unit / microprocessor / master control chip, etc. 4.

[0139] In one embodiment, the at least one computer-executable instruction can also be compiled or constitute a software product in which one or more computer-executable instructions are executed by a processor to perform the steps of the various functions and / or methods in the embodiments described in the present technology.

[0140] Figure 7 A schematic diagram of a computer-readable storage medium according to an embodiment of the present application is shown.

[0141] As Figure 7 shown, a non-transitory computer-readable storage medium 11 stores instructions, for example, computer-readable instructions 10. When the computer-readable instructions 10 are executed by a processor, the various methods described above can be performed. The non-transitory computer-readable storage medium includes, but is not limited to, for example, a volatile memory and / or a non-volatile memory. The volatile memory can include, for example, a random access memory (RAM) and / or a cache memory, etc. The non-transitory non-volatile memory can include, for example, a read-only memory (ROM), a hard disk, a flash memory, etc. For example, the non-transitory computer-readable storage medium 11 can be connected to a computing device such as a computer, and then, when the computing device executes the computer-readable instructions 10 stored on the computer-readable storage medium 11, the various methods described above can be performed.

[0142] In several embodiments provided by the present application, it should be understood that the disclosed apparatus and method can be implemented in other manners. For example, the apparatus embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There can be another division manner for the actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between the units can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or in other forms.

[0143] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., may be located in one place, or may be distributed to multiple network units. Part or all of the units may be selected according to actual needs to achieve the purpose of the embodiment.

[0144] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0145] If the integrated unit is realized in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the whole or part of the technical solutions that make essential contributions to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for executing all or part of the steps of the method of each embodiment of the present application by a computer device (which can be a personal computer, a server, or a network device, etc.). The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (English full name: Read-Only Memory, English abbreviation: ROM), a random access memory (English full name: Random Access Memory, English abbreviation: RAM), a magnetic disk or an optical disk, and various program code storage media.

[0146] The above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A point cloud model adaptive rendering method based on cloud-edge collaboration, characterized in that, The method includes: Initial point cloud data of the target object is acquired by a 3D laser scanner. Then, Gaussian filtering is used to smooth the data and remove noise to obtain processed first-order point cloud data. Outlier analysis is performed on the first-order point cloud data to remove outliers and generate point cloud data. The point cloud data is uploaded to the cloud for partitioned storage to obtain a partitioned storage point cloud data set, and the partitioned storage point cloud data set is distributed to an edge device set, wherein each edge device interacts with at least one RGB-D camera in the RGB-D camera array; An RGB-D camera array is deployed around the target object to capture multi-point images of the target object and send the captured images to the interactively associated edge device. Each RGB-D camera in the RGB-D camera array captures image and depth data synchronously. By combining images acquired simultaneously or gradually with a set of edge devices, cross-modal local correspondence perception is performed on the partitioned storage point cloud data set. By calculating the local correlation between image semantic features and point cloud features, the partitioned storage point cloud data set is compensated to obtain a compensated partitioned storage point cloud data set. The compensated partitioned storage point cloud data set is transmitted back to the cloud via the network for integration processing. The cloud merges the data according to its various parts, removes overlaps and gaps, and uses graphics rendering technology to generate a rendered 3D point cloud model.

2. The cloud-edge collaborative point cloud model adaptive rendering method as described in claim 1, characterized in that, Initial point cloud data of the target object is acquired using a 3D laser scanner. Gaussian filtering is then used to smooth the data and remove noise, yielding processed first-order point cloud data. Outlier analysis is performed on the first-order point cloud data to remove outliers, generating point cloud data including: A 3D laser scanner emits laser beams at a target object from different angles and calculates the time it takes for the beams to reflect back. It measures the distance to the surface of the target object and generates initial point cloud data, wherein the initial point cloud data contains the spatial coordinate information of each point on the surface of the target object. A preset Gaussian filtering window is obtained, and the distance weighted average between each point in the initial point cloud data and its neighboring points within the preset Gaussian filtering window is calculated to smooth high-frequency noise, thereby obtaining first-order processed point cloud data. The preset Gaussian filtering window is set according to the density and noise characteristics of the initial point cloud data. The average distance between each point in the first-order processed point cloud and its K nearest neighbors is calculated. If the average distance exceeds a set threshold, it is determined that the point is an outlier and removed to obtain the point cloud data.

3. The cloud-edge collaborative point cloud model adaptive rendering method as described in claim 1, characterized in that, The point cloud data is uploaded to the cloud for partitioned storage to obtain a partitioned point cloud data set, and the partitioned point cloud data set is distributed to a set of edge devices, including: The edge device set is traversed and collected according to preset device performance indicators. The collected data is organized into a standardized format to obtain a set of edge device performance indicators, wherein the preset device performance indicators include at least computing power, storage capacity and network bandwidth. A pre-built performance coefficient identifier is used to input the set of edge device performance indicators into the performance coefficient identifier to obtain the set of edge device performance coefficients. Based on the edge device performance coefficients in the edge device performance coefficient set and the preset processable point cloud data volume, determine the edge device point cloud data volume set, perform structural partitioning on the point cloud data based on the edge device point cloud data volume set, and store the partitioning results in the cloud partitioned storage to obtain a partitioned storage point cloud data set. Based on the similarity between the point cloud data set processed by edge devices and the point cloud data set stored in partitioned storage, a one-to-one mapping relationship is constructed between edge devices and point cloud data stored in partitioned storage. The point cloud data set stored in partitioned storage is then distributed to the edge device set according to the mapping relationship.

4. The cloud-edge collaborative point cloud model adaptive rendering method as described in claim 3, characterized in that, Pre-built performance coefficient identifier, including: A performance coefficient recognizer is constructed, wherein the performance coefficient recognizer is trained based on a convolutional neural network, and the input layer of the network includes the edge device performance index of each edge device; the collected sample edge device performance indexes and sample edge device performance coefficients are used as training data and labeled to obtain labeled training data; To train a convolutional neural network, labeled training data is input into the convolutional neural network. The data is processed through each layer of the convolutional neural network to generate prediction results, namely the predicted edge device performance coefficients for each edge device. The difference between the predicted performance coefficient of the edge device and the labeled training data is calculated by the loss function. The weights of the convolutional neural network are updated by the backpropagation algorithm. The error is propagated back to each layer, and the parameters of each layer are adjusted until the preset number of iterations is met, so as to obtain the trained performance coefficient recognizer.

5. The cloud-edge collaborative point cloud model adaptive rendering method as described in claim 1, characterized in that, By combining images acquired simultaneously or sequentially from an edge device set, cross-modal local correspondence sensing is performed on a partitioned storage point cloud data set. The partitioned storage point cloud data set is compensated by calculating the local correlation between image semantic features and point cloud features, resulting in a compensated partitioned storage point cloud data set. This process includes: The images received by the edge device set are resized, pixel value normalized, rotated, and flipped to obtain preprocessed images. A pre-trained image semantic recognizer is then used to extract features from the images received by the edge device set to obtain an image semantic feature set. The point cloud data is divided into multiple local regions, and a matching window is set for each local region. The window size is adjusted according to the density of the point cloud data and the resolution of the image features. Cosine similarity is used to perform local correlation matching between the image semantic feature set and the partitioned storage point cloud data set to determine the set of matching points in the cloud map. Based on the cloud map matching point set, the images collected by the edge device set are matched and compensated with the corresponding partition storage point cloud data in the partition storage point cloud data set to obtain the compensated partition storage point cloud data set.

6. The cloud-edge collaborative point cloud model adaptive rendering method as described in claim 5, characterized in that, Using cosine similarity to perform local correlation matching between the image semantic feature set and the partitioned storage point cloud data set, a set of matching points in the cloud map is determined, including: For each local region's image semantic feature set and the partitioned storage point cloud data set, a cosine similarity value is calculated. If the cosine similarity value is greater than or equal to a preset similarity threshold, it is added to the initial cloud map matching point set. The initial cloud map matching point set is used as input. One initial cloud map matching point is selected from it as the source point cloud, and another initial cloud map matching point is selected from it as the target point cloud. The source point cloud is the point cloud that needs to be transformed, and the target point cloud is the data that needs to be aligned with the source point cloud. For each point in the source point cloud, the Euclidean distance is used to find the nearest point in the target point cloud. The optimal rigid transformation from the source point cloud to the target point cloud is calculated, including the rotation matrix and translation vector. The squared error between the point clouds is minimized by the least squares method. The calculated rotation matrix and translation vector are applied to the source point cloud to obtain the transformed point cloud. After multiple iterations, the optimized set of matching points for the cloud map is obtained.

7. The cloud-edge collaborative point cloud model adaptive rendering method as described in claim 6, characterized in that, Obtaining the compensation partition storage point cloud data set includes: Based on the cloud map matching point set, the images collected by the edge device set are locally aligned with the corresponding partition storage point cloud data in the partition storage point cloud data set. By weighted fusion of image semantic features and partition storage point cloud data respectively, a fused partition storage point cloud data set is obtained. The fused partition storage point cloud data set is merged with the partition storage point cloud data set to obtain the compensated partition storage point cloud data set.

8. The cloud-edge collaborative point cloud model adaptive rendering method as described in claim 1, characterized in that, The compensated partitioned storage point cloud data set is transmitted back to the cloud via the network for integration processing, including: The compensated partitioned storage point cloud data set is transmitted back to the cloud via a data transmission protocol. The cloud initializes each data partition by removing redundant data, merges the received data, and integrates it according to spatial information to obtain an initial merged point cloud. Overlapping regions are detected in the initial merged point cloud to obtain an overlapping region set. The ICP algorithm is then used to remove the overlapping parts between different data blocks in the overlapping region set. Interpolation methods are then used to fill the gaps in the point cloud data to complete the integration process.

9. The cloud-edge collaborative point cloud model adaptive rendering method as described in claim 8, characterized in that, The initial merged point cloud is subjected to overlapping region detection to obtain a set of overlapping regions, including: According to the preset voxel size, the initial merged point cloud is segmented into voxels, and each segmented voxel is overlap identified to determine whether there are points of other voxels. If so, it is marked as an overlapping region to obtain the first overlapping region set. The initial merged point cloud is traversed to calculate normal vectors. The calculated normal vector set is then subjected to pairwise similarity identification. If the identification result exceeds a preset threshold, the region where it is located is marked as the second overlapping region, and the second overlapping region set is obtained. The first overlapping region set and the second overlapping region set are then summarized to obtain the overlapping region set.

10. The cloud-edge collaborative point cloud model adaptive rendering method as described in claim 1, characterized in that, The rendered 3D point cloud model is generated using graphics rendering technology, including: Select the rendering API and initialize the rendering context, allocate GPU memory and rendering buffer; configure the rendering target as a window or frame buffer, and set the rendering viewport to complete the rendering environment initialization; In the initialized rendering environment, the integrated point cloud is converted into a triangular mesh format and stored as a vertex buffer object. The spatial coordinates of each point are converted into screen space coordinates through the vertex shader. The texture image is mapped onto the point cloud surface based on UV mapping to obtain the three-dimensional point cloud model.

Images