A composite workpiece element automatic extraction method based on three-dimensional point cloud semantic segmentation

Abstract

The application belongs to the technical field of composite machining, and discloses a composite workpiece element automatic extraction method based on three-dimensional point cloud semantic segmentation, which comprises the following steps: obtaining a two-dimensional image and a three-dimensional point cloud of a composite workpiece element; respectively performing data preprocessing on the two-dimensional image and the three-dimensional point cloud of the composite workpiece element to obtain a co-configuration image set composed of a three-dimensional point set and a two-dimensional image; extracting the composite workpiece element in the two-dimensional image based on a convolutional neural network and obtaining a two-dimensional feature mapping set; establishing the corresponding relationship between the two-dimensional image and the three-dimensional point cloud of the composite workpiece element by using a multi-view aggregation model based on deep learning; and performing semantic segmentation on the three-dimensional point cloud to obtain corresponding point cloud sets of different composite workpiece elements. The application can realize the automatic denoising segmentation and extraction of a point cloud file scanned by a laser after the point cloud file is imported into an algorithm, and can segment the workpiece element according to the characteristics of the composite material, mark the belonging element into the belonging information of each point, and facilitate the extraction and calling.

Description

Technical Field

[0001] This invention belongs to the field of composite material processing technology, specifically relating to a method for semantic segmentation of scanned workpiece point clouds and extraction of point clouds belonging to the same element, based on the characteristics of composite workpieces. Background Technology

[0002] A mechanical workpiece refers to the object being machined in machining, and its elements refer to the points, lines, and surfaces that constitute the geometry of the mechanical workpiece. During machining, it is necessary to limit the geometric tolerances of the mechanical workpiece, such as flatness, cylindricity, perpendicularity, and symmetry. This requires the extraction of the corresponding feature surfaces of the mechanical workpiece. During assembly and inspection, it is also necessary to identify and extract the mating surfaces and the surfaces to be inspected.

[0003] Composite materials refer to materials with superior structure and performance, created by combining two or more different types, properties, forms, and phases through appropriate composite methods. For resin-based and ceramic-based composite materials, the reinforcing materials are often quartz fibers or carbon fibers and their woven fabrics. Taking quartz fiber-reinforced silica composites and quartz fiber-reinforced phenolic resin composites as examples, unlike traditional materials, workpieces made from these materials exhibit obvious woven textures on their surfaces, or noticeable surface defects due to phase structure defects, resulting in an uneven surface. These characteristics of composite materials make it difficult to extract workpiece features.

[0004] Current methods for extracting features from composite workpieces involve first scanning the workpiece with a laser scanner to obtain a 3D point cloud, and then manually segmenting the required feature point cloud regions. However, this approach has several drawbacks: firstly, it cannot achieve automated production line setup; secondly, it significantly impacts the efficiency and accuracy of feature extraction from composite workpieces, consequently affecting the precision of subsequent production and processing. Another approach is semantic segmentation of the 3D point cloud, but this technology is currently immature and inherently has low accuracy. Furthermore, unlike ordinary material workpieces, composite workpieces have uneven surfaces, which severely interferes with feature extraction. Summary of the Invention

[0005] This invention aims to solve the related technical problems of feature extraction of composite workpieces, and provides an automatic feature extraction method for composite workpieces based on three-dimensional point cloud semantic segmentation. Ultimately, after importing the laser-scanned point cloud file into the algorithm, the algorithm automatically performs noise reduction, segmentation and extraction of the point cloud file, and segments the workpiece features according to the characteristics of composite materials, storing the feature labels in the information of each point for easy extraction and retrieval.

[0006] To achieve the above-mentioned objectives, the present invention is implemented through the following technical solution:

[0007] This invention provides an automatic extraction method for composite workpiece features based on 3D point cloud semantic segmentation, the method comprising the following steps:

[0008] (1) Obtain two-dimensional images and three-dimensional point clouds of composite workpiece elements;

[0009] (2) Perform data preprocessing on the two-dimensional images and three-dimensional point clouds of the composite workpiece elements respectively, and set the set of three-dimensional points as P, and set the set of co-matched images composed of two-dimensional images as I.

[0010] (3) Based on a convolutional neural network, the composite workpiece elements in the two-dimensional image are extracted, and a two-dimensional feature mapping set is obtained, represented as follows:

[0011] (4) Using a deep learning-based multi-view aggregation model, the correspondence between the two-dimensional images and three-dimensional point clouds of composite workpiece elements is established.

[0012] (5) Perform semantic segmentation on the three-dimensional point cloud to obtain the corresponding point cloud set of different composite workpiece elements.

[0013] Furthermore, in step (2), the data preprocessing for the two-dimensional image includes image normalization and standardization.

[0014] Furthermore, in step (2), the data preprocessing for the three-dimensional point cloud includes data cleaning and data downsampling of the collected point cloud data.

[0015] Further, in step (3), the convolutional neural network structure includes convolutional layers, pooling layers, activation functions, and fully connected layers; the two-dimensional feature mapping set It is obtained from the convolutional layer.

[0016] Furthermore, in step (4), the deep learning-based multi-view aggregation model includes the following process:

[0017] Process 1. Detect whether a 3D point is in a 2D image using an occlusion detection model, and obtain the image mapping relationship of a single point pix(p, i);

[0018] Process 2. Composed of a single point image mapping relationship pix(p, i) and a two-dimensional feature mapping set. Obtain the element mapping set

[0019] Step 3. Calculate the view quality of feature block k based on visibility conditions and scale it to an attention score.

[0020] Step 4: Set a gating function to prevent the transmission of feature block k when the overall quality of image set v(p) is too low;

[0021] Process 5: Mapping from feature set Attention Score and gate functions The combined point image feature set P(f) is obtained. 2D Ultimately, the correspondence between the two-dimensional image and the three-dimensional point cloud of the composite workpiece elements is realized.

[0022] Further, in process 1, the occlusion detection model is used to indicate whether point p is visible in image i, where point p represents a point in P and i represents an image in I; if point p is visible in image i, and point p is in the view frustum of image i and is not occluded, it is represented as point-image pair (p, i) ∈ P × I, and pix(p, i) is defined as the mapping of a single pixel of image i to point p, and v(p) represents the visible atlas of point p.

[0023] Furthermore, in process 2, the element mapping set

[0024]

[0025] in, It is the association function obtained using a multilayer perceptron (MLP); the two-dimensional feature map set obtained in step three is represented as... It is associated with the co-matched image set I, and its width is C;

[0026] Obtain the element mapping set Then, it is divided into [C / K] consecutive modules, where K is the number of features, and each module is called feature block k:

[0027]

[0028] Furthermore, in process 3, the visibility conditions are solved for each feature block k separately, resulting in a vector group consisting of 8-dimensional vectors. This is used to quantify the visibility conditions for each feature block k; after solving the visibility conditions for each point image pair (p,i) in each feature block k, the view quality of each feature block k is further obtained. View quality It can be obtained from the following formula:

[0029]

[0030] Where φ1, φ2, and φ3 are the association functions obtained from three different multilayer perceptron (MLP) models; all images of i∈v(p) are used to learn and predict the view quality of each feature block k.

[0031] The visibility condition is a quantitative representation model of visibility, which uses an 8-dimensional vector to describe each pair of associated point images (p,i). The 8-dimensional vector includes normalized depth, local linearity description, local flatness description, local scattering description, viewpoint normal, pixel row, local density, and occlusion rate.

[0032] View quality Attention scores are obtained by scaling attention to [0,1]. Used to represent the relative correlation between feature block k (like i) and point p; attention score The formula is as follows:

[0033]

[0034] Where v(p) represents the visible atlas of point p.

[0035] Furthermore, in process 4, the gating parameters The formula is expressed as follows:

[0036]

[0037] Where, α k ,β k All of these are trainable parameters; View quality for each feature block k.

[0038] Furthermore, in process 5, for each point p seen in one or more images i, the feature map set is merged from each point image pair (p, i). For each feature block k, calculate the features in the view. The sum, and the attention score Weighted and multiplied by gating parameters Furthermore, we define the set of combined point image elements P(f) associated with point p. 2D ,p):

[0039]

[0040] Among them, f 2D Let p represent the set of features in a two-dimensional image, and p represent the corresponding point in a three-dimensional point cloud. The final result is achieved by combining the point image feature set P(f). 2D (p) is used to reflect the mapping of a two-dimensional image to a three-dimensional point cloud.

[0041] The beneficial effects of this invention are:

[0042] This invention utilizes the viewing conditions of 3D points in an image to select and merge the most relevant 2D features and 3D information. By combining convolutional neural networks and multi-view aggregation models, the 2D images and 3D point clouds of the scanned composite workpiece are processed to ultimately identify and extract composite workpiece elements containing shape and position information and texture details.

[0043] Compared to existing manual extraction methods for workpiece point cloud features, this invention enables automated operation, significantly improving accuracy and efficiency. It also allows for simultaneous spatiotemporal processing of large numbers of workpieces and their point cloud data. This is an essential step in the intelligent automation of a production line. Compared to traditional methods that directly perform semantic segmentation on 3D point clouds, this invention is better suited to the characteristics of composite workpieces. Specifically, for composite workpieces with uneven surfaces, prominent textures, and difficulties in direct point cloud clustering and segmentation, the method of first performing 2D recognition, then point image mapping, and finally point cloud semantic segmentation is more accurate, targeted, and robust. Attached Figure Description

[0044] Figure 1 This is a flowchart of the automatic extraction method for composite workpiece elements based on three-dimensional point cloud semantic segmentation provided by the present invention;

[0045] Figure 2 This is a schematic diagram of point-image relationship mapping provided by the present invention. Detailed Implementation

[0046] The present invention will be further described in detail below through specific embodiments. These embodiments are intended to enable those skilled in the art to gain a more comprehensive understanding of the present invention, but do not limit the invention in any way.

[0047] like Figure 1 As shown, this embodiment provides an automatic extraction method for composite workpiece features based on 3D point cloud semantic segmentation, which is performed according to the following steps:

[0048] Step 1: Obtain two-dimensional images and three-dimensional point clouds of the composite workpiece elements.

[0049] Using a laser scanner or depth camera, a surround scan of the composite workpiece is performed to obtain a 3D point cloud containing its position information, size, and texture details, as well as the corresponding 2D image. These images are then labeled and compiled into a dataset.

[0050] Step 2: Perform data preprocessing on the two-dimensional images and three-dimensional point clouds of the composite workpiece elements respectively, and denote the set of three-dimensional points as P and the set of co-matched images composed of two-dimensional images as I.

[0051] For two-dimensional images, image normalization and standardization are performed to avoid differences in pixel value range, image size, and color composition between different images, which could lead to differences in network input.

[0052] Preferably, image normalization and standardization refer to scaling the image pixel values ​​to a fixed range, such as [0,1]; and adjusting the size of the input image to a fixed size, such as 224x224.

[0053] For 3D point clouds, the collected point cloud data undergoes data cleaning and downsampling preprocessing.

[0054] Specifically, data cleaning refers to the process of removing noise and outliers from 3D point cloud data. For example, filtering algorithms such as Gaussian filtering and median filtering can be used to eliminate noise.

[0055] Specifically, data downsampling refers to the fact that 3D point cloud data is usually quite dense, while in practical applications only a portion of the data points need to be processed, thus requiring data downsampling. For example, voxel-based sampling algorithms such as voxel mesh filtering can be used to improve the effect of subsequent processing.

[0056] Step 3: Extract composite workpiece elements from the two-dimensional image based on a convolutional neural network.

[0057] First, a convolutional neural network (CNN) structure is constructed, consisting of convolutional layers, pooling layers, activation functions, and fully connected layers. Then, a 2D image is input into the CNN for forward propagation, and the CNN performs convolution, pooling, activation, and fully connected layers to obtain the output. Next, a loss function is calculated based on the output, and backpropagation is performed to update the CNN structure. This iterative training continues until the final model achieves the required accuracy and generalization ability for engineering applications, thus enabling the extraction of composite workpiece elements from 2D images.

[0058] The two-dimensional feature map set obtained from the convolutional layer is a two-dimensional data structure obtained through convolution operations, used to reflect the features of points, lines, and surfaces in a two-dimensional image, represented as follows:

[0059] Step Four: As Figure 2 As shown, a deep learning-based multi-view aggregation model is used to establish the correspondence between two-dimensional images and three-dimensional point clouds of composite workpiece elements, which is used to solve the semantic segmentation of composite workpieces in three-dimensional scenes.

[0060] Specifically, it includes the following five processes, which are described in detail below:

[0061] Process 1. Use an occlusion detection model to detect whether a 3D point is in a 2D image and obtain the mapping relationship of a single point image.

[0062] For the composite workpiece features to be extracted, different viewing angles and different workpiece poses may result in partial or complete occlusion of the desired features. Therefore, this invention provides an occlusion detection model, which is used to indicate whether point p is visible in image i, where p represents a point in P and i represents an image in I.

[0063] If point p is visible in image i, and point p is in the view frustum of image i and is not occluded, then it is represented as a point-image pair (p, i) ∈ P × I. pix(p, i) is defined as the mapping of a single pixel of image i to point p, and v(p) represents the visible set of point p.

[0064] Process 2. Composed of a single point image mapping relationship pix(p, i) and a two-dimensional feature mapping set. Obtain the element mapping set

[0065] This process employs a multilayer perceptron (MLP) to map two-dimensional feature sets based on the individual point image mapping relationship pix(p, i). The features in the image are transferred to the 3D point cloud, thus obtaining the feature mapping set.

[0066]

[0067] in, It is the correlation function obtained using a multilayer perceptron (MLP). The image features that a multilayer perceptron (MLP) can acquire include spatial relationships, size, and texture information.

[0068] The two-dimensional feature mapping set obtained in step three is represented as follows: It is associated with the co-matched image set I and has a width of C. The feature mapping set is obtained. Then, it is divided into [C / K] consecutive modules, where K is the number of features, and each module is called feature block k:

[0069]

[0070] Step 3. Calculate the view quality of feature block k based on visibility conditions and scale it to an attention score.

[0071] The visibility condition is a quantitative representation model for visibility, which uses a D-dimensional vector to describe each pair of associated point-image pairs (p,i), and represents this D-dimensional vector as o. (p,i) .

[0072] For the point cloud semantic segmentation problem of composite workpiece elements in this invention, the visibility condition is represented by a D=8 dimensional vector, which are (1) normalized depth, (2) local linearity description, (3) local flatness description, (4) local scattering description, (5) view normal, (6) pixel row, (7) local density, and (8) occlusion rate.

[0073] Specifically;

[0074] Normalized depth: The normalized depth is obtained by dividing the distance between point p and image i by the maximum observation distance of the laser scanner or depth camera, so as to obtain the relative relationship between each point-image pair (p,i).

[0075] Local geometric description of line dimension, flatness, and scattering: The local geometric description is calculated based on the eigenvalues ​​of the covariance matrix between point p and its 50 neighborhood points, thus obtaining the values ​​of line dimension, flatness, and scattering. This is because the geometry of a 3D point cloud affects the view quality in a 2D image. Planar surfaces are better captured by the camera, but highly irregular surfaces may exhibit numerous occlusions or glancing effects.

[0076] Viewpoint normal: The normal direction of each point p is obtained and compared with the fitted surface formed by 50 points in its neighborhood to obtain the value of the spatial line-surface angle.

[0077] Pixel row: The pixel row value is obtained by dividing the number of pixel rows of image i by the image height. It is used to compensate for the distortion caused by the camera lens.

[0078] Local density: Represented by the number of 3D points per unit area. Different densities will affect occlusion and serve as an indicator of the local accuracy of a 3D sensor.

[0079] Occlusion rate: Calculated by the occlusion ratio of a point p and its 50 neighboring points.

[0080] For each feature block k, the visibility conditions in eight dimensions are solved separately, resulting in a vector group consisting of eight-dimensional vectors. This is used to quantify the visibility conditions of each feature block k. That is, each image pair (p, i) corresponds to an 8-dimensional vector, and the 8-dimensional vectors of all image pairs (p, i) in a feature block k form a vector group.

[0081] After solving the visibility condition for each point image pair (p,i) in each feature block k, the view quality of each feature block k is further obtained. View quality It can be obtained from the following formula:

[0082]

[0083] Here, φ1, φ2, and φ3 are the association functions obtained from three different multilayer perceptron (MLPs). The construction of these three different MLPs is selected based on specific working conditions and belongs to existing technologies. All images i∈v(p) are used to learn and predict the view quality of each feature block k.

[0084] To facilitate subsequent processing, the view quality will be adjusted. Attention scores are obtained by scaling attention to [0,1]. This is used to represent the relative correlation between feature block k (like i) and point p.

[0085] Here, attention is achieved by assessing the view quality of all images i in the visible atlas v(p) of point p. The result is obtained by applying the softmax() function. Since each point p may have a different number of views, the softmax() function is adjusted according to the number of views that see point p, resulting in the final attention score. The formula is as follows:

[0086]

[0087] Step 4: Set the gating function

[0088] Since not all information in the visible atlas v(p) is effective for processing the results, the 2D image needs to be monitored before the final mapping result is established. Due to occlusion or limited viewpoints, for a given feature block k, some points p may not be seen by any image i, i.e., they are not seen in either near-view or far-view observations. In this case, discarding invalid information feature blocks from tilted viewpoint observations and relying purely on geometric relationships to infer the information is an effective approach. Therefore, a gating parameter is introduced. Its function is to prevent the transmission of feature block k when the overall quality of the image set v(p) is too low:

[0089]

[0090] Where, α k ,β k All of these are trainable parameters.

[0091] Process 5: Mapping from feature set Attention Score and gate functions The combined point image element set is obtained, and the correspondence between the two-dimensional image and the three-dimensional point cloud of the composite workpiece element is finally realized.

[0092] For each point p seen in one or more images i, merge the feature map set from each point image pair (p, i). For each feature block k, calculate the features in the view. The sum, and the attention score Weighted and multiplied by gating parameters Furthermore, we define the set of combined point image elements P(f) associated with point p. 2D ,p):

[0093]

[0094] f 2D Let p represent the set of features in a two-dimensional image, and p represent the corresponding point in a three-dimensional point cloud. The final result is achieved by combining the point image feature set P(f). 2D (p) is used to reflect the mapping of a two-dimensional image to a three-dimensional point cloud.

[0095] It can be seen that the mapping relationship between 3D point cloud and 2D image can be obtained through the multi-view aggregation model. Processes 1-5 consider the visibility condition and scale the final attention score according to the visible map set v(p). Therefore, for the observation of the same feature, the quality of the observed view is different depending on the viewing angle, and the corresponding mapping weight parameters are also different: (1) Far-view observation mostly obtains the shape and position information of composite workpiece; (2) Near-view observation can obtain the texture information of composite workpiece; (3) Tilt-view observation mostly obtains invalid information.

[0096] Step 5: Based on the correspondence between the two-dimensional images of composite workpiece elements and the three-dimensional point cloud, perform semantic segmentation on the three-dimensional point cloud to finally obtain the corresponding point cloud set of different composite workpiece elements.

[0097] As can be seen, the automatic extraction method for composite workpiece elements based on three-dimensional point cloud semantic segmentation provided by the present invention can extract the composite workpiece elements from the two-dimensional image and three-dimensional point cloud of the acquired composite workpiece elements through convolutional neural networks after data preprocessing. It can also establish the correspondence between the two-dimensional image and the three-dimensional point cloud of the composite workpiece elements using a deep learning-based multi-view aggregation model, and then perform semantic segmentation on the three-dimensional point cloud to achieve automatic extraction of composite workpiece elements.

[0098] Although the preferred embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many specific modifications under the guidance of the present invention without departing from the spirit of the invention and the scope of protection of the claims, and these modifications all fall within the scope of protection of the present invention.

Claims

1. A method for automatic extraction of composite workpiece elements based on 3D point cloud semantic segmentation, characterized in that, The method includes the following steps: (1) Obtain two-dimensional images and three-dimensional point clouds of composite workpiece elements; (2) Perform data preprocessing on the two-dimensional images and three-dimensional point clouds of the composite workpiece elements respectively, and set the set of three-dimensional points as P, and set the set of co-matched images composed of two-dimensional images as I. (3) Based on a convolutional neural network, the features of the composite workpiece in the two-dimensional image are extracted, and a two-dimensional feature mapping set is obtained, represented as follows: (4) Using a deep learning-based multi-view aggregation model, the correspondence between the two-dimensional images and three-dimensional point clouds of composite workpiece elements is established. The deep learning-based multi-view aggregation model includes the following process: Process 1. Detect whether a 3D point is in a 2D image using an occlusion detection model, and obtain the image mapping relationship of a single point pix(p,i); Process 2. Composed of a single point image mapping relationship pix(p,i) and a two-dimensional feature mapping set. Obtain the element mapping set Step 3. Calculate the view quality of feature block k based on visibility conditions and scale it to an attention score. Step 4: Set a gating function to prevent the transmission of feature block k when the overall quality of image set v(p) is too low; Process 5: Mapping from feature set Attention Score and gate functions The combined point image feature set P(f) is obtained. 2D ,p), ultimately realizing the correspondence between the two-dimensional image and the three-dimensional point cloud of the composite workpiece elements; (5) Semantic segmentation of the three-dimensional point cloud is performed to obtain the corresponding point cloud set of different composite workpiece elements.

2. The method for automatic extraction of composite workpiece elements based on three-dimensional point cloud semantic segmentation according to claim 1, characterized in that, In step (2), the data preprocessing for the two-dimensional image includes image normalization and standardization.

3. The method for automatic extraction of composite workpiece elements based on three-dimensional point cloud semantic segmentation according to claim 1, characterized in that, In step (2), the data preprocessing for the three-dimensional point cloud includes data cleaning and data downsampling of the collected point cloud data.

4. The method for automatic extraction of composite workpiece elements based on three-dimensional point cloud semantic segmentation according to claim 1, characterized in that, In step (3), the convolutional neural network includes convolutional layers, pooling layers, activation functions, and fully connected layers; the two-dimensional feature map set It is obtained from the convolutional layer.

5. The method for automatic extraction of composite workpiece elements based on three-dimensional point cloud semantic segmentation according to claim 1, characterized in that, In process 1, the occlusion detection model is used to indicate whether point p is visible in image i, where point p represents a point in P and i represents an image in I; If point p is visible in image i, and point p is in the view frustum of image i and is not occluded, then it is represented as a point-image pair (p,i)∈p×I. pix(p,i) is defined as the mapping of a single pixel of image i to point p, and v(p) represents the visible set of point p.

6. The method for automatic extraction of composite workpiece elements based on three-dimensional point cloud semantic segmentation according to claim 1, characterized in that, In process 2, the element mapping set in, It is the association function obtained using a multilayer perceptron (MLP); the two-dimensional feature map set obtained in step three is represented as... It is associated with the co-matched image set I, and its width is C; Obtain the element mapping set Then, it is divided into [C / K] consecutive modules, where K is the number of features, and each module is called feature block k:

7. The method for automatic extraction of composite workpiece elements based on three-dimensional point cloud semantic segmentation according to claim 1, characterized in that, In process 3, the visibility conditions are solved for each feature block k separately, resulting in a vector group consisting of 8-dimensional vectors. This is used to quantify the visibility conditions for each feature block k; after solving the visibility conditions for each point image pair (p,i) in each feature block k, the view quality of each feature block k is further obtained. View quality It can be obtained from the following formula: Where φ1, φ2, and φ3 are the association functions obtained from three different multilayer perceptron (MLP) models; all images of i∈v(p) are used to learn and predict the view quality of each feature block k. The visibility condition is a quantitative representation model of visibility, which uses an 8-dimensional vector to describe each pair of associated point images (p,i). The 8-dimensional vector includes normalized depth, local linearity description, local flatness description, local scattering description, viewpoint normal, pixel row, local density, and occlusion rate. View quality Attention scores are obtained by scaling attention to [0,1]. Used to represent the relative correlation between feature block k (like i) and point p; attention score The formula is as follows: Where v(p) represents the visible atlas of point p.

8. The method for automatic extraction of composite workpiece elements based on three-dimensional point cloud semantic segmentation according to claim 1, characterized in that, In process 4, the gating parameters The formula is expressed as follows: Where, α k ,β k All of these are trainable parameters; View quality for each feature block k.

9. The method for automatic extraction of composite workpiece elements based on three-dimensional point cloud semantic segmentation according to claim 1, characterized in that, In process 5, for each point p seen in one or more images i, the feature map set is merged from each point image pair (p, i). For each feature block k, calculate the features in the view. The sum, and the attention score Weighted and multiplied by gating parameters Furthermore, we define the set of combined point image elements P(f) associated with point p. 2 D,p): Among them, f 2D Let p represent the set of features in a two-dimensional image, and p represent the corresponding point in a three-dimensional point cloud. The final result is achieved by combining the point image feature set P(f). 2D The image is represented by ,p) to reflect the mapping of a two-dimensional image to a three-dimensional point cloud.

Images