A method for point cloud scene retrieval based on dynamic convolution to extract point cloud features

Abstract

The present invention provides a method for realizing point cloud scene retrieval based on point cloud feature extraction based on dynamic convolution, and relates to the technical field of scene point cloud intensity image retrieval. The present application extracts point cloud features of a target laser point cloud through a point cloud feature encoding network; the point cloud feature encoding network includes: a stacked feature rotation layer and a dynamic convolution layer for extracting local features of the point cloud, and the last dynamic convolution layer is connected to a maximum pooling layer for aggregating local features of the point cloud, and the point cloud features are obtained by feature fusion of the global features of the point cloud and the local features of the point cloud; the image features of the candidate scene point cloud intensity image are extracted through a scene point cloud intensity image feature encoding network; a lightweight multi-layer perceptron is used as a projection layer to map the point cloud features and the image features into point cloud comparison objects and image comparison objects respectively; the similarity of the point cloud comparison objects and the image comparison objects is used for comparative learning to match the laser point cloud with the corresponding scene point cloud intensity image.

Description

Technical Field

[0001] The present invention relates to the technical field of scene point cloud intensity image retrieval, and in particular to a method for realizing point cloud scene retrieval based on extracting point cloud features through dynamic convolution. Background Art

[0002] In recent years, the rise of 3D scanning technology has driven many applications that rely on 3D point cloud data, such as high-precision map creation, autonomous driving, and robot navigation. In addition to the point cloud, the point cloud intensity map is an important feature that can provide additional information about each point, such as reflectivity, signal strength, etc. Combining the point cloud with its intensity map can enhance the interpretability and usability of the data in many applications.

[0003] At present, many analysis methods of laser point clouds and point cloud intensity maps use neural network models. For the same scene, different point cloud intensity images can be obtained by sampling from different angles. The input of the neural network model for mapping needs to combine point cloud intensity images and point clouds from multiple angles. Therefore, during training, it is necessary to collect the scene point cloud and scene point cloud intensity images from different perspectives for the same scene, and construct a data set of scene point cloud intensity images corresponding to multiple perspectives for each point cloud to train the neural network for mapping. This requires retrieving the corresponding scene point cloud intensity image from a large number of scene point cloud intensity images based on the laser point cloud. Summary of the invention

[0004] In order to solve the above technical problem or at least partially solve the above technical problem, the present invention provides a method for realizing point cloud scene retrieval based on dynamic convolution to extract point cloud features.

[0005] In a first aspect, the present invention provides a method for implementing point cloud scene retrieval based on point cloud feature extraction based on dynamic convolution, comprising:

[0006] The point cloud features of the target laser point cloud are extracted by a point cloud feature encoding network based on dynamic convolution; wherein the constructed point cloud feature encoding network includes: a stacked feature rotation layer and a dynamic convolution layer for extracting local features of the point cloud, the dynamic convolution layer uses balls with different set radii to scan the laser point cloud to obtain local features of point clouds at different scales, and the last dynamic convolution layer is connected to the maximum pooling layer for aggregating local features of the point cloud, and the point cloud global features and the point cloud local features are obtained by feature fusion to obtain the point cloud features;

[0007] Extracting image features of candidate scene point cloud intensity images through a scene point cloud intensity image feature encoding network;

[0008] A lightweight multi-layer perceptron is used as the projection layer to map the orientation code and image features into image comparison objects, and to map the orientation code and point cloud features into point cloud comparison objects.

[0009] A lightweight multi-layer perceptron is used as the projection layer to map point cloud features and image features into point cloud comparison objects and image comparison objects respectively;

[0010] The similarity between point cloud comparison objects and image comparison objects is used for comparative learning to match the target laser point cloud with the scene point cloud intensity image.

[0011] Furthermore, the convolution kernel of the dynamic convolution layer is implemented as follows: the target laser point cloud is scanned by a ball with a set radius, and the ball is centered on the point in the target laser point cloud to obtain the center point All its neighbors The positional relationship between them is recorded as: , Center point The total number of neighbors of

[0012] Analyze the center point through the pre-trained weight coefficient network Each of its neighbors The positional relationship between them is used to obtain the weight coefficient that represents the correlation between the positional relationship and all the positional correlation weight matrices in the positional correlation weight matrix set:

[0013] ;

[0014] in, Represents the nonlinear mapping formed by the multi-layer perceptron, represents the Softmax layer, express The set of weight coefficients is expressed as , To represent the center point Each of its neighbors The position relationship between the position association weight matrix and the position association weight matrix set The weight coefficient of the correlation between them;

[0015] For each position relationship, the dynamic convolution kernel is constructed using the sum of the product of the weight coefficient and the weight matrix:

[0016] ;

[0017] For the center point All its neighbors The position relationship between them: ,have Group , expressed as , then the corresponding The group convolution kernel is expressed as: .

[0018] Furthermore, a set number of position-associated weight matrices are generated by random initialization , , regularize different position association weight matrices to obtain a set of position association weight matrices that represent different position relationships between points .

[0019] Furthermore, the weight coefficient network includes a multi-layer perceptron, and the last fully connected network layer of the multi-layer perceptron is connected to the Softmax layer.

[0020] Furthermore, the feature rotation layer is divided into two branches, one of which processes the input target laser point cloud data through the convolution layer, normalization layer and activation function, and then maps the data into 9 numbers that can form a regularized rotation matrix through the fully connected layer. After the 9 numbers are converted into the regularized rotation matrix, the rotation matrix is ​​multiplied by the processing result of the target laser point cloud data of the other branch.

[0021] Furthermore, the scene point cloud intensity image feature encoding network uses a series of multi-level resnet convolutional networks with downsampling layers to abstract multi-level image features of different scales from the candidate scene point cloud intensity image, and fuses the image features of different scales to obtain image features.

[0022] Furthermore, the contrastive learning uses the corresponding cross-modal point cloud features and image features of the same scene as positive pairs, and uses the point cloud features and image features of all different scenes as negative pairs, wherein, for any scene, the cosine similarity of the image features of the scene and all other scenes is calculated, and a set number of other scenes with the lowest cosine similarity are retained to form a negative pair with the scene;

[0023] In the data domain of the enhanced point cloud and scene point cloud intensity image, the model is optimized to maximize the cosine similarity between the point cloud comparison object and the image comparison object in the positive pair, and minimize the cosine similarity between the point cloud comparison object and the image comparison object in the negative pair, so that the model can model effective features that can distinguish between positive and negative pairs;

[0024] Among them, the loss function of the optimization model in the contrastive learning process is:

[0025] ;

[0026] Among them, A is the total number of targets, B is the total number of negative pairs, is the similarity function, is the exponential function of the natural logarithm, is the point cloud comparison object, is the image comparison object directly opposite to the point cloud comparison object, is the image contrast object in the negative pair of the point cloud contrast object, is the sensitivity coefficient.

[0027] Furthermore, in the contrastive learning process, data enhancement is used to process the point cloud and scene point cloud intensity images. The data enhancement process is as follows: the point cloud and scene point cloud intensity images are transformed by rotation, scaling and random flipping to obtain enhanced point cloud and scene point cloud intensity images.

[0028] In a second aspect, the present invention provides a device for realizing point cloud scene retrieval based on dynamic convolution to extract point cloud features, comprising: at least one processing unit, the processing unit and a storage unit are interconnected through a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the method for realizing point cloud scene retrieval based on dynamic convolution to extract point cloud features is implemented.

[0029] In a third aspect, the present invention provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the method for point cloud scene retrieval based on dynamic convolution to extract point cloud features.

[0030] The above technical solution provided by the embodiment of the present invention has the following advantages compared with the prior art:

[0031] The present application extracts the point cloud features of the target laser point cloud through a point cloud feature coding network; the point cloud feature coding network includes: a stacked feature rotation layer and a dynamic convolution layer for extracting local features of the point cloud, and the last dynamic convolution layer is connected to the maximum pooling layer for aggregating local features of the point cloud, and the point cloud features are obtained by feature fusion of the global features of the point cloud and the local features of the point cloud; the obtained convolution kernel is a dynamic convolution kernel that adapts to the positional relationship, and the dynamic convolution kernel is multiplied with the point cloud feature map, which can more flexibly model the irregular geometric structure of the three-dimensional point cloud. The image features of the candidate scene point cloud intensity image are extracted through the scene point cloud intensity image feature coding network; a lightweight multi-layer perceptron is used as a projection layer to map the orientation code and image features into image comparison objects, and to map the orientation code and point cloud features into point cloud comparison objects; the similarity between the point cloud comparison object and the image comparison object is used for comparative learning to match the target laser point cloud with the scene point cloud intensity image. In the data domain of the point cloud image, the model is optimized to maximize the cosine similarity of the point cloud comparison object and the image comparison object in the positive pair, and minimize the cosine similarity of the point cloud comparison object and the image comparison object in the negative pair, so that the model can model effective features that can distinguish between positive and negative pairs. In this application, the point cloud to be matched is mapped by the point cloud feature encoding network and the projection layer to model the point cloud comparison object that represents the transformation invariant properties related to the image in the point cloud, and the candidate scene point cloud intensity image is mapped by the scene point cloud intensity image feature encoding network and the projection layer to model the image comparison object that represents the transformation invariant properties related to the point cloud in the image. The similarity between the point cloud comparison object and the image comparison object is used for comparative learning to match the target laser point cloud with the scene point cloud intensity image, effectively achieving the matching of the point cloud and the image. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0034] Figure 1 A flowchart of a method for implementing point cloud scene retrieval based on point cloud feature extraction based on dynamic convolution provided by an embodiment of the present invention;

[0035] Figure 2 An architecture diagram of a model for implementing a method for extracting point cloud features based on dynamic convolution to realize point cloud scene retrieval provided by an embodiment of the present invention;

[0036] Figure 3 A schematic diagram of a point cloud feature encoding network provided by an embodiment of the present invention;

[0037] Figure 4 A schematic diagram of a dynamic convolutional layer in a point cloud feature encoding network provided by an embodiment of the present invention;

[0038] Figure 5 A schematic diagram of a scene point cloud intensity image feature encoding network provided by an embodiment of the present invention;

[0039] Figure 6 A schematic diagram of a point cloud scene retrieval device based on extracting point cloud features by dynamic convolution provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0042] Example 1

[0043] like Figure 1 and Figure 2 As shown, the technology of the present invention realizes a method for point cloud scene retrieval based on dynamic convolution to extract point cloud features. The present invention includes the following steps:

[0044] The point cloud features of the target laser point cloud are extracted through a point cloud feature encoding network based on dynamic convolution.

[0045] like Figure 3As shown, the constructed point cloud feature encoding network includes: a stacked feature rotation layer and a dynamic convolution layer for extracting local features of the point cloud. The dynamic convolution layer uses balls with different set radii to scan the laser point cloud to obtain local features of the point cloud at different scales. The last dynamic convolution layer is connected to the maximum pooling layer for aggregating local features of the point cloud, and the point cloud global features and local features are obtained by feature fusion to obtain the point cloud features.

[0046] like Figure 4 As shown in the figure, the convolution kernel of the dynamic convolution layer is implemented as follows:

[0047] Randomly initialize and generate a set number of position association weight matrices , , regularize different position association weight matrices to obtain a set of position association weight matrices that represent different position relationships between points .

[0048] Next, a weight coefficient network needs to be established. The weight coefficient network maps the discrete position association weight matrix into a dynamic convolution kernel in a continuous three-dimensional space based on the position relationship between the points in the point cloud. Specifically, a ball with a set radius scans the target laser point cloud. The ball takes the point in the target laser point cloud as the center and obtains the center point. All its neighbors The positional relationship between them is recorded as: , Center point The total number of neighbors of

[0049] Analyze the center point through the pre-trained weight coefficient network Each of its neighbors The weight coefficient is obtained by the positional relationship between them. The weight coefficient network includes a multi-layer perceptron, and the last layer of the multi-layer perceptron is connected to the Softmax layer.

[0050] The weight coefficient obtained by the weight coefficient network is expressed as:

[0051] ;

[0052] in, Represents the nonlinear mapping formed by the multi-layer perceptron, represents the Softmax layer, express The set of weight coefficients is expressed as , To represent the center point Each of its neighbors The position relationship between the position association weight matrix and the position association weight matrix set The weight coefficient of the association between the position relationship and the weight matrix; in this way, for each position relationship processing process, a weight coefficient is selected for each weight matrix, and a higher weight coefficient means a stronger association between the position relationship and the weight matrix.

[0053] For each position relationship, the dynamic convolution kernel is constructed using the sum of the product of the weight coefficient and the weight matrix:

[0054] ;

[0055] For the center point All its neighbors The position relationship between them: , then Group ,Right now , then the corresponding The group convolution kernel is expressed as: .

[0056] The obtained convolution kernel is a dynamic convolution kernel that adapts to the positional relationship. The dynamic convolution kernel is multiplied with the point cloud feature map, which can more flexibly model the irregular geometric structure of the three-dimensional point cloud.

[0057] Among them, the feature rotation layer is divided into two branches. One of them processes the input point cloud data through the convolution layer, normalization layer and activation function, and then maps the data into 9 numbers that can form a regularized rotation matrix through the fully connected layer. After the 9 numbers are converted into a regularized rotation matrix, they are multiplied with the point cloud data of the other branch.

[0058] The image features of the candidate scene point cloud intensity images are extracted through the scene point cloud intensity image feature encoding network.

[0059] like Figure 5 As shown, the scene point cloud intensity image feature encoding network uses a series of multi-level resnet convolutional networks with downsampling layers to abstract multi-level image features of different scales from the candidate scene point cloud intensity image, and fuses the image features of different scales to obtain image features.

[0060] A lightweight multi-layer perceptron is used as the projection layer to map point cloud features and image features into point cloud comparison objects and image comparison objects respectively;

[0061] The similarity between point cloud comparison objects and image comparison objects is used for comparative learning to match the target laser point cloud with the scene point cloud intensity image.

[0062] In the contrastive learning process, data enhancement is used to process point clouds and scene point cloud intensity images. Data enhancement can increase the difficulty of self-supervised learning, reduce overfitting, and encourage pre-trained weights to learn transformation-invariant features. The data enhancement process is as follows: point clouds and scene point cloud intensity images are transformed by rotation, scaling, and random flipping to obtain enhanced point clouds and scene point cloud intensity images.

[0063] After point cloud and scene point cloud intensity image enhancement, the corresponding cross-modal point cloud features and image features of the same scene are taken as positive pairs, and the cloud features and image features of all different scenes are taken as negative pairs, wherein the negative pairs of any scene are selected using the similarity of image features. Specifically, for any scene, the cosine similarity of the image features of the scene and all other scenes is calculated, and a set number of other scenes with the lowest cosine similarity are retained to form negative pairs with the scene.

[0064] During training, in the data domain of the point cloud image, the model is optimized to maximize the cosine similarity between the point cloud comparison object and the image comparison object in the positive pair, and minimize the cosine similarity between the point cloud comparison object and the image comparison object in the negative pair, so that the model can model effective features that can distinguish between positive and negative pairs. Among them, the loss function of the optimization model in the contrastive learning process is:

[0065] ;

[0066] Among them, A is the total number of targets, B is the total number of negative pairs, is the similarity function, is the exponential function of the natural logarithm, is the point cloud comparison object, is the image comparison object directly opposite to the point cloud comparison object, is the image contrast object in the negative pair of the point cloud contrast object, is the sensitivity coefficient.

[0067] The present application extracts the point cloud features of the target laser point cloud through a point cloud feature encoding network; the point cloud feature encoding network includes: a stacked feature rotation layer and a dynamic convolution layer for extracting local features of the point cloud, and the last dynamic convolution layer is connected to the maximum pooling layer for aggregating local features of the point cloud, and the point cloud features are obtained by feature fusion of the global features and local features of the point cloud; the obtained convolution kernel is a dynamic convolution kernel that adapts to the positional relationship, and the dynamic convolution kernel is multiplied with the point cloud feature map to more flexibly model the irregular geometric structure of the three-dimensional point cloud. The image features of the candidate scene point cloud intensity image are extracted through the scene point cloud intensity image feature encoding network; a lightweight multi-layer perceptron is used as the projection layer to map the point cloud features and image features into point cloud comparison objects and image comparison objects respectively; the similarity of the point cloud comparison object and the image comparison object is used for comparative learning to match the target laser point cloud with the scene point cloud intensity image. In the data domain of the point cloud image, the model is optimized to maximize the cosine similarity of the point cloud comparison object and the image comparison object in the positive pair, and minimize the cosine similarity of the point cloud comparison object and the image comparison object in the negative pair, so that the model can model effective features that can distinguish between positive and negative pairs. In this application, the point cloud to be matched is mapped by the point cloud feature encoding network and the projection layer to model the point cloud comparison object that represents the transformation invariant properties related to the image in the point cloud, and the candidate scene point cloud intensity image is mapped by the scene point cloud intensity image feature encoding network and the projection layer to model the image comparison object that represents the transformation invariant properties related to the point cloud in the image. The similarity between the point cloud comparison object and the image comparison object is used for comparative learning to match the target laser point cloud with the scene point cloud intensity image, effectively achieving the matching of the point cloud and the image.

[0068] Example 2

[0069] See also Figure 6 As shown, an embodiment of the present invention provides a device for realizing point cloud scene retrieval based on dynamic convolution point cloud feature extraction, comprising: at least one processing unit, the processing unit and the storage unit are interconnected through a bus unit, the storage unit is a computer-readable storage medium, and can be used to store software programs, computer executable programs and modules, such as the software programs, computer executable programs and modules corresponding to the method for realizing point cloud scene retrieval based on dynamic convolution point cloud feature extraction in the embodiment of the present invention. The processing unit implements the above-mentioned method for realizing point cloud scene retrieval based on dynamic convolution point cloud feature extraction by running the software programs, computer executable programs and modules stored in the storage unit, comprising:

[0070] The point cloud features of the target laser point cloud are extracted by a point cloud feature encoding network based on dynamic convolution; wherein the constructed point cloud feature encoding network includes: a stacked feature rotation layer and a dynamic convolution layer for extracting local features of the point cloud, wherein the feature rotation layer is divided into two branches, one of which processes the input point cloud data through a convolution layer, a normalization layer and an activation function, and then maps the data into 9 numbers that can form a regularized rotation matrix through a fully connected layer, and after the 9 numbers are converted into a regularized rotation matrix, they are multiplied with the point cloud data of the other branch; the dynamic convolution layer uses balls with different set radii to scan the laser point cloud to obtain local features of point clouds at different scales, and the last dynamic convolution layer is connected to the maximum pooling layer for aggregating local features of the point cloud, and the point cloud features are obtained after feature fusion of the global features of the point cloud and the local features of the point cloud;

[0071] Extracting image features of candidate scene point cloud intensity images through a scene point cloud intensity image feature encoding network;

[0072] A lightweight multi-layer perceptron is used as the projection layer to map point cloud features and image features into point cloud comparison objects and image comparison objects respectively;

[0073] The similarity between point cloud comparison objects and image comparison objects is used for comparative learning to match laser point clouds with images.

[0074] Of course, the computer program stored in the storage unit of the device for implementing point cloud scene retrieval based on dynamic convolution to extract point cloud features provided in an embodiment of the present invention is not limited to the method operations described above, but can also execute related operations in the method for implementing point cloud scene retrieval based on dynamic convolution to extract point cloud features provided in any embodiment of the present invention.

[0075] Example 3

[0076] An embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. When the computer program is executed, the method for extracting point cloud features based on dynamic convolution to realize point cloud scene retrieval is implemented, including:

[0077] The point cloud features of the target laser point cloud are extracted by a point cloud feature encoding network based on dynamic convolution; wherein the constructed point cloud feature encoding network includes: a stacked feature rotation layer and a dynamic convolution layer for extracting local features of the point cloud, wherein the feature rotation layer is divided into two branches, one of which processes the input point cloud data through a convolution layer, a normalization layer and an activation function, and then maps the data into 9 numbers that can form a regularized rotation matrix through a fully connected layer, and after the 9 numbers are converted into a regularized rotation matrix, they are multiplied with the point cloud data of the other branch; the dynamic convolution layer uses balls with different set radii to scan the laser point cloud to obtain local features of point clouds at different scales, and the last dynamic convolution layer is connected to the maximum pooling layer for aggregating local features of the point cloud, and the point cloud features are obtained after feature fusion of the global features of the point cloud and the local features of the point cloud;

[0078] Extracting image features of candidate scene point cloud intensity images through a scene point cloud intensity image feature encoding network;

[0079] A lightweight multi-layer perceptron is used as the projection layer to map point cloud features and image features into point cloud comparison objects and image comparison objects respectively;

[0080] The similarity between point cloud comparison objects and image comparison objects is used for comparative learning to match laser point clouds with images.

[0081] A computer-readable storage medium provided in an embodiment of the present invention stores a computer program which is not limited to the method operations described above, but can also execute related operations in a method for realizing point cloud scene retrieval based on dynamic convolution to extract point cloud features provided in any embodiment of the present invention.

[0082] In the embodiments provided by the present invention, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, structures or units, which can be electrical, mechanical or other forms.

[0083] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0084] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0085] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for realizing point cloud scene retrieval based on point cloud feature extraction using dynamic convolution, characterized in that: include: The point cloud features of the target laser point cloud are extracted by a point cloud feature encoding network based on dynamic convolution; wherein the constructed point cloud feature encoding network includes: a stacked feature rotation layer and a dynamic convolution layer for extracting local features of the point cloud, the dynamic convolution layer uses balls with different set radii to scan the laser point cloud to obtain local features of point clouds at different scales, and the last dynamic convolution layer is connected to the maximum pooling layer for aggregating local features of the point cloud, and the point cloud global features and the point cloud local features are obtained by feature fusion to obtain the point cloud features; Extracting image features of candidate scene point cloud intensity images through a scene point cloud intensity image feature encoding network; A lightweight multi-layer perceptron is used as the projection layer to map point cloud features and image features into point cloud comparison objects and image comparison objects respectively; The similarity between point cloud comparison objects and image comparison objects is used for comparative learning to match the target laser point cloud with the scene point cloud intensity image.

2. The method for realizing point cloud scene retrieval based on dynamic convolution point cloud feature extraction according to claim 1 is characterized in that: The convolution kernel of the dynamic convolution layer is implemented as follows: the target laser point cloud is scanned by a ball with a set radius, and the ball is centered on the point in the target laser point cloud to obtain the center point All its neighbors The positional relationship between them is recorded as: , Center point The total number of neighbors of Analyze the center point through the pre-trained weight coefficient network Each of its neighbors The positional relationship between them is used to obtain the weight coefficient that represents the correlation between the positional relationship and all the positional correlation weight matrices in the positional correlation weight matrix set: ； in, Represents the nonlinear mapping formed by the multi-layer perceptron, represents the Softmax layer, express The set of weight coefficients is expressed as , To represent the center point Each of its neighbors The position relationship between the position association weight matrix and the position association weight matrix set The weight coefficient of the correlation between them; For each position relationship, the dynamic convolution kernel is constructed using the sum of the product of the weight coefficient and the weight matrix: ； For the center point All its neighbors The position relationship between them: ,have Group , expressed as , then the corresponding The group convolution kernel is expressed as: .

3. The method for realizing point cloud scene retrieval based on dynamic convolution point cloud feature extraction according to claim 2 is characterized in that: Randomly initialize and generate a set number of position association weight matrices , , regularize different position association weight matrices to obtain a set of position association weight matrices that represent different position relationships between points .

4. The method for realizing point cloud scene retrieval based on dynamic convolution point cloud feature extraction according to claim 2 is characterized in that: The weight coefficient network includes a multi-layer perceptron, and the last fully connected network layer of the multi-layer perceptron is connected to the Softmax layer.

5. The method for realizing point cloud scene retrieval based on dynamic convolution point cloud feature extraction according to claim 1, characterized in that: The feature rotation layer is divided into two branches. One of them processes the input target laser point cloud data through the convolution layer, normalization layer and activation function, and then maps the data into 9 numbers that can form a regularized rotation matrix through the fully connected layer. After the 9 numbers are converted into a regularized rotation matrix, the rotation matrix is ​​multiplied by the processing result of the target laser point cloud data of the other branch.

6. The method for realizing point cloud scene retrieval based on dynamic convolution point cloud feature extraction according to claim 1, characterized in that: The scene point cloud intensity image feature encoding network uses a series of multi-level resnet convolutional networks with downsampling layers to abstract multi-level image features of different scales from the candidate scene point cloud intensity image, and fuses the image features of different scales to obtain image features.

7. The method for realizing point cloud scene retrieval based on dynamic convolution point cloud feature extraction according to claim 1 is characterized in that: The contrastive learning uses the corresponding cross-modal point cloud features and image features of the same scene as positive pairs, and uses the point cloud features and image features of all different scenes as negative pairs, where: For any scene, calculate the cosine similarity of the image features of the scene and all other scenes, and retain a set number of other scenes with the lowest cosine similarity to form a negative pair with the scene; In the data domain of the enhanced point cloud and scene point cloud intensity image, the model is optimized to maximize the cosine similarity between the point cloud comparison object and the image comparison object in the positive pair, and minimize the cosine similarity between the point cloud comparison object and the image comparison object in the negative pair, so that the model can model effective features that can distinguish between positive and negative pairs; Among them, the loss function of the optimization model in the contrastive learning process is: ； Among them, A is the total number of targets, B is the total number of negative pairs, is the similarity function, is the exponential function of the natural logarithm, is the point cloud comparison object, is the image comparison object directly opposite to the point cloud comparison object, is the image contrast object in the negative pair of the point cloud contrast object, is the sensitivity coefficient.

8. The method for realizing point cloud scene retrieval based on dynamic convolution point cloud feature extraction according to claim 7 is characterized in that: In the contrastive learning process, data enhancement is used to process the point cloud and scene point cloud intensity images. The data enhancement process is as follows: the point cloud and scene point cloud intensity images are transformed by rotation, scaling and random flipping to obtain enhanced point cloud and scene point cloud intensity images.

9. A point cloud scene retrieval device based on dynamic convolution to extract point cloud features, characterized in that: include: At least one processing unit, the processing unit and the storage unit are interconnected through a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the method for realizing point cloud scene retrieval based on dynamic convolution to extract point cloud features as described in any one of claims 1-8 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for point cloud scene retrieval based on dynamic convolution to extract point cloud features is implemented as described in any one of claims 1 to 8.

Images