Point cloud denoising method and system for cross-domain retrieval

Abstract

The invention discloses a point cloud denoising method for cross-domain retrieval, which comprises the following steps of: S1, jointly collecting point cloud data by using multiple sensors to obtain a camera point cloud and a radar point cloud; s2, point cloud registration: aligning two point clouds into the same coordinate system; s3, calculating an outlier weight coefficient through a deep learning network; s4, calculating a filtering weight coefficient of the original camera point cloud through a point cloud filtering network based on deep learning; and S5, denoising the camera point cloud through the following formula. According to the method, the camera point cloud is filtered through the environment information of the radar point cloud, and the improved machine learning network is combined to obtain the related denoising coefficient for denoising, so that the smoothness and integrity of the denoised point cloud can be remarkably improved, and a high-quality data basis is provided for subsequent three-dimensional reconstruction and semantic understanding.

Description

Technical Field

[0001] This invention relates to the field of point cloud denoising technology, and more specifically to a point cloud denoising method that uses cross-domain retrieval of point clouds with multi-view and multi-sensor fusion. Background Technology

[0002] With the rapid development of point cloud data acquisition technology, a large number of efficient and accurate point cloud reconstruction methods have emerged, such as Structure from Motion (SfM) reconstruction, real-time dense 3D reconstruction using monocular or depth cameras, and LiDAR-based mapping technology. Point clouds, as discrete sets of 3D points irregularly sampled from continuous curved surfaces, have become a mainstream 3D data representation widely used in fields such as autonomous driving, robotics, virtual reality, and immersive telepresence.

[0003] However, due to limitations in the physical performance of acquisition equipment and the uncertainty of matching in 3D reconstruction algorithms, point cloud data inevitably introduces noise. Noise not only affects the geometric accuracy of point clouds but also significantly interferes with downstream tasks such as rendering, reconstruction, and analysis, and may even lead to the loss and deformation of underlying structural information. Therefore, research on point cloud denoising (PCD) technology is of great significance for improving the robustness and reliability of 3D vision systems.

[0004] Traditional point cloud denoising techniques often rely on filtering. This filtering process typically considers the surrounding environment of the points in the point cloud. However, the original point cloud itself contains a significant amount of noise, and filtering based on a noisy environment can negatively impact the accuracy of the filtering results.

[0005] Among the many traditional filtering methods, each has its own advantages and suitable scenarios, but none can comprehensively outperform the others. This is because manually defined coefficients are difficult to characterize the diverse properties of data, limiting the generalization ability of traditional denoising methods. Therefore, the parameter selection of these traditional methods largely depends on the data structure and noise characteristics of the point cloud. In fact, the geometric structure of point cloud data is complex, and its noise characteristics are often unknown, making it difficult to obtain the optimal algorithm and parameters.

[0006] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0007] The purpose of this invention is to provide a denoising technology for cross-domain retrieval of multi-view, multi-sensor fusion point clouds. It combines the most widely used radar point cloud acquisition technology and camera point cloud reconstruction technology, making full use of the advantages of different point cloud acquisition technologies to achieve better point cloud denoising effect.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows: A point cloud denoising method for cross-domain retrieval includes the following steps: S1. Use multiple sensors to jointly collect point cloud data to obtain camera point cloud and radar point cloud; S2. Point cloud registration: Align two point clouds to the same coordinate system; S3. Calculate outlier weight coefficients using a deep learning network. ; Specifically, the steps include: S31. Using the registered point cloud, in the camera point cloud... Choose one Then select Radar point cloud within radius r k points in Calculate these k points The coordinates of the mean point are used to determine the mean point, and then the calculation is performed. The distance from the mean point is determined as the mean filtering distance. This can be expressed as the following formula:

[0009] S32. Concatenate the mean-filtered distance with the original camera point cloud data, input the result into the deep learning network, and obtain the outlier weight coefficients. ; S4. Calculate the filter weight coefficients from the camera point cloud using a deep learning-based point cloud filtering network. ; S5. Denoise the camera point cloud using the following formula:

[0010] in, Point The denoising result, i.e., the denoised coordinate values, Indicates the midpoint of the camera point cloud. The neighborhood of radius r, For points within this neighborhood, Represents the corresponding point in the neighborhood. The denoising weight coefficient values.

[0011] in ,in, For point The outlier weight coefficient value, For point The filter weight coefficient values.

[0012] Preferably, step S4 specifically includes the following steps: S41. Input the camera point cloud into the encoder to obtain multi-scale features. , S42. Multi-scale features Outlier weight coefficients Element-wise multiplication is performed, and then the results are input into a deep learning-based point cloud filtering network to obtain the filter weight coefficients. The formula is as follows:

[0013] in The asterisk (*) represents the decoder, and the asterisk (*) represents element-wise multiplication.

[0014] This invention also provides a cross-domain retrieval point cloud denoising system, the system comprising a point cloud acquisition module, a point cloud registration module, an outlier weight coefficient generation module, a filter weight coefficient generation module, and a denoised point cloud generation module; the point cloud acquisition module is used to generate camera point clouds and radar point clouds, the point cloud registration module is used to register the camera point clouds and radar point clouds, aligning the two point clouds to the same coordinate system, the outlier weight coefficient generation module is used to generate outlier weight coefficients, the filter weight coefficient generation module is used to generate filter weight coefficients, and the denoised point cloud generation module generates denoised point clouds using formula 4 or formula 5.

[0015] Preferably, the mean filtering distance of the points in the camera point cloud is calculated using the points in the aligned radar point cloud, as shown in Formula 2. Then, the mean filtering distance is concatenated with the camera point cloud data, and the concatenated data is used as the input of the outlier weight coefficient generation module.

[0016] Preferably, the outlier weight coefficient generation module adopts the PointNet network structure, removes T-Net, and uses softmax to replace the final normalization layer and activation layer to output the outlier weight coefficients.

[0017] Preferably, the filter weight coefficient generation module adopts an encoder-decoder architecture, using outlier weight coefficients before passing the features captured by the encoder module to the decoder module. Weight them accordingly.

[0018] In addition, the present invention also provides a cross-domain retrieval point cloud denoising device, including: a memory, a processor, and a cross-domain retrieval point cloud denoising program stored in the memory and executable on the processor, wherein the cross-domain retrieval point cloud denoising program is configured with a method for implementing cross-domain retrieval point cloud denoising.

[0019] In addition, the present invention also provides a storage medium storing a cross-domain retrieval point cloud denoising program, which, when executed, implements a cross-domain retrieval point cloud denoising method.

[0020] By adopting the above technical solution, the present invention has the following beneficial technical effects: 1. Traditional point cloud denoising techniques often rely on filtering. Filtering is typically performed based on the surrounding environment of the points in the point cloud. However, the original point cloud itself contains a significant amount of noise, and filtering based on a noisy environment can affect the accuracy of the filtering results. In multimodal point cloud denoising, radar point clouds are more accurate than camera-reconstructed point clouds. Therefore, this invention uses environmental information from radar point clouds to filter camera point clouds, which can improve the accuracy of denoising.

[0021] 2. Among numerous traditional filtering methods, each has its own advantages and suitable scenarios, but none can comprehensively outperform the others. This is because manually defined coefficients are difficult to characterize the diverse features of data, limiting the generalization ability of traditional denoising methods. Therefore, the parameter selection of these traditional methods largely depends on the data structure and noise characteristics of the point cloud. In fact, the geometric structure of point cloud data is complex, and noise characteristics are often unknown, making it difficult to obtain optimal algorithms and parameters. Therefore, this invention combines deep learning with filtering techniques to perform cross-source point cloud denoising, improving the generalization ability of the denoising method.

[0022] 3. This invention uses partial structures of PointNet and PointCleanNet networks to generate weight coefficients to calculate the final denoised displacement, instead of directly using the network to generate the denoised displacement, which can achieve better denoising results. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the system of the present invention.

[0024] Figure 2 This is a schematic diagram of the original PointNet network structure.

[0025] Figure 3 This is a schematic diagram of the original PointCleanNet network structure. Detailed Implementation

[0026] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0027] Example 1: There are many existing methods for collecting point cloud data. For example, radar-collected point clouds are accurate but sparse, while camera-reconstructed point clouds are dense and have color values ​​but are not accurate. The method proposed in this invention uses radar point clouds to guide camera reconstruction of point clouds for denoising. Since it involves joint denoising of point cloud data from different sources, it is called a cross-domain retrieval point cloud denoising method.

[0028] The point cloud denoising method for cross-domain retrieval of the present invention includes the following steps: S1. Use multiple sensors to jointly collect point cloud data.

[0029] Specifically, this includes using a real-time monocular camera dense reconstruction method to obtain camera point clouds, and using LiDAR to obtain high-precision radar point clouds.

[0030] The point cloud generation methods described above are all common knowledge, and those skilled in the art can implement them using existing methods. This embodiment is only briefly described here.

[0031] A real-time monocular camera dense reconstruction method consists of two parts: a pose estimation algorithm and dense reconstruction. Pose estimation: For each frame of the image, feature points and corresponding descriptors are extracted using Sift GPU; to obtain a georeferenced location, the camera's pose estimation integrates visual-based and GNSS-based positioning. Dense reconstruction: To ensure the accuracy of the point cloud, dense reconstruction uses images with a larger baseline from adjacent time frames as stereo image pairs; Bouguet correction aligns the stereo images to parallel optical axes, and efficient stereo matching is performed using the CUDA-accelerated SGM algorithm.

[0032] High-precision radar point clouds can be obtained using the FAST-LIO2 algorithm, which mainly consists of two modules: a motion model based on a tightly coupled Kalman filter and a mapping module. The motion model employs a Kalman filter-based motion model for efficient processing. This model performs forward propagation of IMU measurements and iterative updates for each LiDAR scan. Forward propagation estimates the current LiDAR position by integrating IMU measurements; specifically, when process noise is low, it transmits the current LiDAR state and covariance. Based on the IMU forward propagation estimation, the scanned point cloud is projected into the world coordinate system. Then, nonlinear optimization is used to align the point cloud with map points to minimize errors, resulting in more accurate attitude estimation. The mapping module uses a novel data structure, the IKD tree, to organize map points for large-scale mapping. To control map size, the system stores a region of a specific length around the current location. When the LiDAR scanning area touches the map boundary, the map moves along the boundary direction, and map points are deleted.

[0033] S2. Point cloud registration: Align two point clouds to the same coordinate system.

[0034] Point cloud registration is a conventional technique, and those skilled in the art can implement it using known methods.

[0035] This embodiment provides a simple example, and the specific steps are as follows: S21. Point Cloud Preprocessing: This includes downsampling and rough alignment. In this embodiment, a VoxelGrid filter is used to downsample the point cloud, thereby reducing its density without losing detail. Rough alignment uses GNSS constraints; however, the accuracy of GNSS is often limited, leading to inaccurate alignment; typically, only a rough positioning range can be determined.

[0036] S22. Point Cloud Segmentation: For cross-source point clouds, only a few sub-regions are similar due to significant differences in regional feature distribution. To address this issue, the point cloud is segmented into smaller blocks for individual registration. The overall registration pose of the point cloud can then be estimated based on the registered blocks. To segment the point cloud into blocks, the point cloud region is divided into a grid using a specified stride. Then, points are assigned to different grids based on their locations.

[0037] S23. Coarse Matching: The range of candidate blocks has a significant impact on the subsequent matching accuracy; therefore, it is necessary to select a good candidate block range. This embodiment calculates the centroid of the point cloud blocks reconstructed by the camera and combines it with GNSS information in the X... The Euclidean distance L on the Y-plane is used to delineate a ring around the candidate point cloud patches. Several point cloud patches are selected from this ring, and the Pearson correlation coefficient is calculated. Patches with high correlation are selected as candidate patches. The specific formula is as follows: Formula 1 In the formula, x and y represent points in two point cloud blocks, respectively. and These represent the average values ​​of points in the two point cloud blocks, respectively.

[0038] S24. Fine Matching: Extract points from each candidate block where d > h to filter out ground noise, where d is the distance from the point to the ground plane and h is the height threshold. Then, project the collected points onto X. In the Y-plane, boundary points are calculated using a boundary estimation algorithm. Next, a two-dimensional normal distribution transform (2DNDT) algorithm is used to register each point cloud patch. This allows finding the best-matching point cloud patch based on the minimum average distance cost of each point. Then, based on the 2D registration result, a three-dimensional normal distribution transform (3DNDT) is used for 3D registration, adjusting the pose along the z-axis. If the matching process converges, the registration is adopted; otherwise, it is discarded.

[0039] Then, the transformation matrix is ​​calculated: the transformation matrices of all blocks are selected and clustered. If the translation and angle between two different transformations are both less than a threshold, then the two transformations are considered to be of the same type. The cluster with the most transformations is selected, and its average value is calculated as the optimal transformation.

[0040] Then, the optimal transformation is used to accurately match the point cloud.

[0041] S3. Calculate the outlier weight coefficients using a deep learning network.

[0042] Outliers are points that are significantly far from the main body of the point cloud and can be considered as points severely affected by noise. We obtain the outlier weight coefficient for each point in the camera point cloud using a deep learning network. The higher the probability that a point is an outlier, the smaller its corresponding outlier weight coefficient.

[0043] Deep learning networks can use PointNet or its variants, and we can output the outlier weight coefficients for each point based on them, assuming the outlier probability of a point is... Then the outlier weight coefficient at that point .

[0044] This step specifically includes: S31. Using the registered point cloud, in the camera point cloud Choose one Then select Radar point cloud within radius r k points in Calculate these k points The coordinates of the mean point are used to determine the mean point, and then the calculation is performed. The distance from the mean point is determined as the mean filtering distance. This can be expressed as the following formula:

[0045] S32. Concatenate the mean-filtered distance with the camera point cloud data, input the result into the deep learning network, and obtain the outlier weight coefficients. .

[0046] To remove degrees of freedom that are not related to denoising, points within a local block can be normalized by dividing by the radius r before being input into the network, and the same operation can be performed in subsequent steps.

[0047] This embodiment uses the PointNet deep learning network, which expands point cloud coordinates into features using pointwise convolution and extracts global features using max pooling. Following PointNet's design in point cloud segmentation, and considering both model complexity and performance, this embodiment extracts four convolutional layers and one max pooling layer as the network structure for the outlier detector. A Softmax layer is then used after the final convolution to output the outlier weight coefficients.

[0048] S4. Calculate the filter weight coefficients using a point cloud filtering network based on deep learning.

[0049] Specifically, the steps include the following: S41. Input the camera point cloud into the encoder to obtain multi-scale features. , S42. Multi-scale features Outlier weight coefficients Element-wise multiplication is performed, and then the results are input into the decoder to obtain the filter weight coefficients. This can be expressed by the following formula:

[0050] in This represents a camera point cloud filtering network based on deep learning, and * indicates element-wise multiplication.

[0051] There are many types of point cloud filtering networks based on deep learning, such as the PointCleanNet network. This network outputs point displacements, but this embodiment outputs the filtering weight coefficients of the points within the neighborhood of each point in the point cloud. Therefore, certain adjustments need to be made to the network structure. Compared to the original PointCleanNet network, this invention removes the STN (Spatial Transformer Network) used to maintain network stability because its actual performance is generally poor. The original PointCleanNet network passes the high-dimensional features of the output neighborhood points through a multilayer perceptron, directly outputting the point displacement. This invention passes the high-dimensional features through a softmax layer to output the filter weight coefficients.

[0052] S5. Denoise the camera point cloud using the following formula:

[0053] in, Point The denoising result, i.e., the denoised coordinate values, Indicates the midpoint of the camera point cloud. The neighborhood of radius r, For points within this neighborhood, Represents the corresponding point in the neighborhood. The denoising weight coefficient values.

[0054] in ,in, For point The outlier weight coefficient value, For point The filter weight coefficient values.

[0055] Based on the formula above, first obtain the midpoint of the camera point cloud. Find the neighborhood of radius r, and then find each point within that neighborhood. coordinates and points Subtracting the coordinates of a point yields a point. Adjusting the associated coordinates ( The adjusted correlation coordinates of each point in the neighborhood are multiplied by the denoising weight coefficient of that point, and then summed to obtain the value of the point. Adjusting coordinates [ , will point Adding the original coordinates to the adjusted coordinates gives the point. The denoising result (denoised coordinates).

[0056] To normalize the coordinates, we can make some adjustments to the above formula, as follows:

[0057] Where r is the neighborhood radius, and is a hyperparameter.

[0058] Perform this operation on each point in the camera point cloud to obtain a denoised point cloud of the camera point cloud.

[0059] This embodiment also included comparative verification experiments. Verification experiments in a multi-source point cloud scenario demonstrated that the method described in this embodiment can effectively achieve cross-modal registration and fusion of point clouds from different sensors. The deep learning network can adaptively adjust denoising parameters based on the noise characteristics of the point cloud, significantly improving the smoothness and integrity of the point cloud, providing a high-quality data foundation for subsequent 3D reconstruction and semantic understanding. Specific data is shown in Table 1 below: Table 1

[0060] In comparison method 1, in step S31 of embodiment 1, the mean filtering distance is not calculated using radar point cloud coordinates, but the outlier weight coefficient is calculated directly using camera point cloud coordinates input into the deep learning network. All other steps are the same as in embodiment 1.

[0061] In comparison method 2, the coordinate displacement of the outlier point is calculated in step S3 of embodiment 1, and the point displacement of the denoised point is also calculated in step S4. Then, step S5 is skipped; the denoised point coordinates are obtained by directly subtracting the point displacement from the denoised point coordinates. All other steps are consistent with the method in embodiment 1.

[0062] The dataset used in the experiment was the PCN dataset. The comparison metric used was the chamfer distance (…). The smaller the value, the better the noise reduction effect.

[0063] Example 2: Additionally, this invention also provides a cross-domain retrieval point cloud denoising system, the system comprising a point cloud acquisition module, a point cloud registration module, an outlier weight coefficient generation module, a filter weight coefficient generation module, and a denoised point cloud generation module. For example... Figure 1 As shown.

[0064] The point cloud acquisition module is used to generate camera point clouds and radar point clouds.

[0065] The point cloud registration module is used to register the camera point cloud and the radar point cloud, aligning the two point clouds into the same coordinate system.

[0066] The outlier weight coefficient generation module is used to generate outlier weight coefficients.

[0067] like Figure 1 and 2 As shown, the outlier weight coefficient generation module adopts the PointNet network structure with certain improvements. Specifically, this invention removes T-Net because it performs poorly in subsequent optimization processes, and most PointNet-based works no longer use T-Net. This invention uses softmax instead of the final normalization and activation layers to output the outlier weight coefficients. To apply softmax for probability normalization, the network output classification probability size is set to n*2 instead of n*1. The first column of the output is the probability of non-outliers, i.e., the outlier weight coefficient, and the second column is the probability of outliers.

[0068] In this embodiment, the mean filtering distance of the points in the aligned radar point cloud is calculated using the points in the camera point cloud. Then, the mean filtering distance is concatenated with the camera point cloud data, and the concatenated data is used as the input of the outlier weight coefficient generation module. For details, please refer to step S31 of embodiment 1.

[0069] The filter weight coefficient generation module is used to generate filter weight coefficients.

[0070] The filter weight coefficient generation module adopts an encoder-decoder architecture.

[0071] like Figure 1 and 3 As shown, this embodiment refers to the PointCleanNet network structure, which itself includes an outlier identifier. This outlier identifier directly deletes outliers in the original point cloud, so the data input to the filtering module does not contain outliers. The outlier identification in this invention is processed by the outlier weight coefficient generation module, so this embodiment deletes the outlier identifier of the PointCleanNet network structure.

[0072] The original network structure had k channels in its output before entering the multilayer perceptron because the original method used to select the k nearest neighbors was different. In this invention, we use the more commonly used method of selecting points within a sampling radius r.

[0073] In addition, this embodiment removes the last multilayer perceptron in the original network and adds a softmax layer to allow the neural network to output filter weight coefficients.

[0074] The encoder module in this embodiment includes three convolutional layers and one max-pooling layer. Features corresponding to the center of a local block are denoted as center features. The features captured by the max-pooling layer are global features. Furthermore, the features at each point are subtracted from the center features to obtain relative features. Global features describe the features of the entire local block. Relative features reflect the relationship between each point and the center of the local block. Therefore, the captured features carry multi-scale information for subsequent calculation of filter weight coefficients. Before passing the features captured by the encoder module into the decoder module, we use outlier weight coefficients... Weight them accordingly.

[0075] The denoised point cloud generation module is used to generate denoised point clouds.

[0076] Furthermore, this embodiment is merely a basic description of the point cloud denoising system for cross-domain retrieval of the present invention. For technical details not described in detail in this embodiment, please refer to the methods provided in any embodiment of the present invention, which will not be repeated here.

[0077] Example 3: Those skilled in the art will clearly understand that the systems and methods of the above embodiments can be implemented using software plus necessary general-purpose hardware platforms. Of course, they can also be implemented using hardware, but in many cases, the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, node packaging device, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0078] Therefore, the present invention also provides a cross-domain retrieval point cloud denoising device, comprising: a memory, a processor, and a cross-domain retrieval point cloud denoising program stored in the memory and executable on the processor, wherein the cross-domain retrieval point cloud denoising program is configured with a method for implementing cross-domain retrieval point cloud denoising.

[0079] In addition, the present invention also provides a storage medium on which a point cloud denoising program for cross-domain retrieval is stored.

[0080] In reality, when deploying equipment or programs, a program may execute all steps, or it may execute only one step, with multiple programs working together to achieve the same result. Therefore, when the cross-domain retrieval point cloud denoising program is executed, it implements all or some of the processes in the cross-domain retrieval point cloud denoising method.

[0081] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for cross-domain retrieval denoising of point clouds, characterized in that, Includes the following steps: S1. Use multiple sensors to jointly collect point cloud data to obtain camera point cloud and radar point cloud; S2. Point cloud registration: Align two point clouds to the same coordinate system; S3. Compute outlier weight coefficients by a deep learning network ; Specifically comprising the following steps: S31. using the registered point cloud, selecting a point in the camera point cloud then selecting k points in the radar point cloud within a radius r neighborhood , calculating the coordinate mean of the k points , determining the mean point, and then calculating the distance from the mean point, determining the mean filtering distance , as follows:​​ S32. The mean filtering distance is spliced with the camera point cloud data, input into a deep learning network, and an outlier weight coefficient is obtained ; S4. Compute filtering weight coefficients for the camera point cloud by a deep learning based point cloud filtering network ; S5. Denoise the camera point cloud using the following formula: wherein, denotes the denoised result of the point, i.e. the denoised coordinate value, denotes a neighborhood of the point with radius r in the camera point cloud, is a point within the neighborhood, denotes a denoising weight coefficient value for the corresponding point within the neighborhood, wherein wherein, is a point outlier weight coefficient value, is a point filter weight coefficient value.

2. The point cloud denoising method for cross-domain retrieval as described in claim 1, characterized in that, Step S4 specifically comprises the following steps: S41. inputting the camera point cloud into an encoder to obtain multi-scale features , S42. The multi-scale features with outlier weight coefficients are element-wise multiplied and then input to the decoder to obtain the filter weight coefficients as follows: wherein denotes a decoder, * denotes element-wise multiplication.

3. The cross-domain retrieved point cloud denoising method of claim 1, wherein, Step S5 normalizes the coordinates using the following formula: Where r is the neighborhood radius, and is a hyperparameter.

4. A system for cross-domain retrieval of denoised point clouds, characterized in that, The system includes a point cloud acquisition module, a point cloud registration module, an outlier weight coefficient generation module, a filter weight coefficient generation module, and a denoised point cloud generation module. The point cloud acquisition module is used to generate camera point clouds and radar point clouds. The point cloud registration module is used to register camera point clouds and radar point clouds. The outlier weight coefficient generation module is used to generate outlier weight coefficients. The mean filtering distance of points in the camera point cloud is calculated using points in the aligned radar point cloud, as shown in the following formula: Then, the mean-filtered distance is concatenated with the camera point cloud data, and this concatenated data is used as input to the outlier weight coefficient generation module. The filter weight coefficient generation module is used to generate filter weight coefficients. , The denoised point cloud generation module generates a denoised point cloud using the following formula: 。 5. The point cloud denoising system for cross-domain retrieval as described in claim 4, characterized in that, The outlier weight coefficient generation module adopts the PointNet network structure, removes T-Net, and replaces the final normalization layer and activation layer with softmax to output the outlier weight coefficients.

6. The point cloud denoising system for cross-domain retrieval as described in claim 4, characterized in that, The filter weight coefficient generation module adopts an encoder-decoder architecture, using outlier weight coefficients before passing the features captured by the encoder module into the decoder module. Weight them accordingly.

7. Point cloud denoising devices for cross-domain retrieval, including: The device includes a memory, a processor, and a cross-domain retrieval point cloud denoising program stored in the memory and executable on the processor, wherein the cross-domain retrieval point cloud denoising program is configured with a point cloud denoising method for implementing cross-domain retrieval.

8. A storage medium storing a cross-domain retrieval point cloud denoising program, wherein the cross-domain retrieval point cloud denoising program, when executed, implements a cross-domain retrieval point cloud denoising method.

Images