Point cloud analysis method of gradient attention map convolution network

Abstract

The application discloses a point cloud analysis method of a gradient attention graph convolution network, sets an edge convolution layer and a gradient convolution layer as two branches, extracts multi-modal features of the point cloud by using different networks, so that the network can pay attention to the original information of the point cloud and the feature difference of the point cloud at the same time, and the application uses the original spatial position information of the point cloud and gradient information in the neighborhood to construct an attention matrix, increases the accuracy and robustness of the model, reduces the influence of some abnormal points, and directly and explicitly constructs the attention matrix when constructing a self-attention module, so that the attention module does not increase too much calculation amount and does not affect the performance of the original model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of point cloud analysis and computer vision, and particularly relates to a point cloud analysis method of gradient attention map convolution network. BACKGROUND

[0002] With the popularity of optical sensors such as lidar and millimeter wave radar and the development of related fields such as autonomous driving and indoor robots, point cloud analysis models have received widespread attention from academia and industry. Local aggregation symbol is one of the most important technologies in point cloud analysis models, which effectively aggregates the neighborhood features of point clouds and greatly improves the performance of models in processing point cloud data. At present, 3D point cloud analysis methods are divided into multi-view based methods, voxel based methods and point based methods. Among them, the multi-view based method projects the point cloud into each view; the voxel based method voxelizes the point cloud into regular voxels; the point based method directly extracts point features.

[0003] Chinese patent document CN114219009A discloses a point cloud downsampling method based on a biased attention mechanism and mutual supervision loss. It includes the following steps: using Simplify to simplify the original point cloud into a simplified point cloud; using Pointnet to extract the semantic features of the original point cloud and the simplified point cloud; finding the local neighbors of the simplified point cloud in the original point cloud through the KNN algorithm; performing feature fusion of the biased attention mechanism; using a similarity projection convolutional neural network to generate a similarity score based on the fused features, and using the distance between them and the score to correct the probability position distribution. The above technical solution uses a multi-layer perceptron to extract the point cloud feature difference and the Euclidean position distance between point clouds, uses a similarity projection convolutional neural network to generate a similarity score based on the fused features, and uses the distance between them and the score to correct the probability position distribution. The disadvantage is that it pays too much attention to feature differences or original position differences between point clouds, and ignores the original point cloud information, resulting in a decline in model performance. At the same time, this proposal only uses distance information to construct an attention matrix, which will cause fluctuations in model accuracy. SUMMARY

[0004] The application mainly solves the technical problems that the original technical scheme ignores the original point cloud information, leading to the decline of model performance, and only uses distance information to construct an attention matrix, leading to the fluctuation of model precision, and provides a point cloud analysis method of gradient attention map convolution network, by setting edge convolution layer and gradient convolution layer two branches, the multi-modal features of the point cloud are extracted by using different networks, so that the network of the application can pay attention to the original information of the point cloud and the feature difference of the point cloud at the same time, and the application uses the original space position information of the point cloud and the gradient information in the neighborhood to construct the attention matrix, increases the precision and robustness of the model, reduces the influence of some abnormal points, when constructing the self-attention module, the attention matrix is directly and explicitly constructed, so that the attention module does not increase too much calculation, and does not affect the performance of the original model.

[0005] The above technical problems of the application are mainly solved by the following technical scheme: the application comprises the following steps:

[0006] S1 constructing a gradient convolution layer;

[0007] S2 constructing an edge convolution layer;

[0008] S3 putting the original features of the point cloud into several edge convolution layers and gradient convolution layers respectively;

[0009] S4 obtaining multi-modal point cloud embedding and merging;

[0010] S5 putting the merging result into a gradient convolution layer and a maximum pooling layer to obtain the final output;

[0011] S6 outputting the global shape descriptor and inputting the classifier to perform three-dimensional tasks.

[0012] The gradient map convolution network proposed by us is a double-channel model composed of several parallel gradient convolution layers and edge convolution layers, which obtains the global shape descriptor by fusing spatial features of different dimensions. The input of the network is the original feature of the point cloud, after the input is respectively input into several edge convolution layers and gradient convolution layers with different sampling radii, different modal point cloud embedding is obtained, after merging, it is put into a gradient convolution layer with larger sampling radius, and after the maximum pooling layer, the global shape descriptor is obtained. The descriptor can be input into the classifier to perform three-dimensional shape recognition and three-dimensional segmentation tasks.

[0013] As preferred, the step S1 gradient convolution layer comprises two parallel gradient convolution modules with different sampling radii, the original feature of the point cloud and the absolute position contained in the original feature are taken as the input of the gradient convolution module, and the gradient convolution module construction process comprises the following steps:

[0014] S1.1 find the index of several center points P by using the farthest point sampling method;

[0015] S1.2 Find the index of M nearest neighbor points Q within the radius r range of the center point P using the ball query algorithm;

[0016] S1.3 Establish a directed graph between P and Q to obtain their relative positions and edge features;

[0017] S1.4 Calculate the zenith angle matrix and azimuth angle matrix between points through their relative positions, and construct the neighborhood gradient attention matrix using the two angles;

[0018] S1.5 Put the edge features into the multi-layer perceptron for extraction, multiply them with the neighborhood gradient attention matrix, and finally pass them through the max-pooling layer to obtain the output neighborhood features.

[0019] As a preferred, the gradient convolution module is the core of the gradient convolution layer, which extracts the global features of the point cloud and the absolute position as the input of the gradient convolution module, w-1 is the number of layers, F w-1 is the output channel number of w-1 layers, n is the number of input point clouds, the input global features X w-1 , the absolute position P, the sampling radius r, and the sampling point number k. The working process of the gradient convolution module using weighted aggregation edge features includes:

[0020] S3.1 Find the k nearest neighbor points q i within the radius r range of the center point p ij using the ball query algorithm;

[0021] S3.2 Establish a directed graph to obtain the relative position and the edge feature

[0022] S3.3 Calculate the relative distance dis ij , the zenith angle E ij and the azimuth angle A ij through the relative position ;

[0023] S3.4 Cross-multiply the zenith angle E ij and the azimuth angle A ij to obtain the edge feature attention matrix G;

[0024] S3.5 Multiply the edge feature attention matrix G after extracting high-dimensional features through the MLP, and finally use the sum function to aggregate the edge features to the center point.

[0025] As a preferred, the step S3.1 specifically includes, first finding a plurality of center points pi ∈ P, then use the ball query algorithm to select k nearest neighbors within a radius r of each center point, denoted as Q ij i = 1, …, n; j = 1, …, k.

[0026] As a preferred, the step S3.2 specifically includes,

[0027] A directed graph is established in the neighborhood to obtain the relative position vector and edge feature. By changing the size of the sampling radius r, we can flexibly adjust the receptive field of the gradient convolution module, so as to capture local semantics of different scales,

[0028] The relative position vector between the center point p i and its neighboring point q ij is denoted as:

[0029]

[0030] The edge feature between the center point p i and its neighboring point q ij is denoted as:

[0031]

[0032] As a preferred, the step S3.3 specifically includes, using the original geometric relationship of the point cloud itself to constrain the aggregation of the edge feature, giving each edge a different weight, the distance dis ij between the center point p i and its neighboring point q ij is obtained by the relative position vector :

[0033]

[0034] Subsequently, an orthogonal three-dimensional coordinate system XYZ is established, and then the relative position vector is projected onto the XY plane, the angle between the projection vector of the relative position vector and the XY plane in the Z direction is denoted as the zenith angle E ij , and the angle between the projection vector and the Y axis is denoted as the azimuth angle A ij , the zenith angle E ij and the azimuth angle A ij are calculated according to the distance between the position vector and the neighboring point:

[0035]

[0036] As preferred, the step S3.4 specifically comprises that the gradient convolution module extracts high-dimensional edge features by using a weight function H, and M(p, q) adds a distance constraint in the local point cloud, which means that the shorter the distance from a neighbor point to the center point, the greater the weight of the corresponding edge, wherein max(dis in ) represents the distance from the center point to the farthest neighbor point, dis ij represents the distance from the center point p i to the jth neighbor point q ij , and the purpose of the aggregation function G(p, q) is to aggregate the edge features to the center point p i , which is expressed as follows:

[0037]

[0038]

[0039] Wherein max represents the max-pooling operation, and sum represents the summation operation.

[0040] As preferred, the step S3.5 specifically comprises that the lower channel uses the relative distance between the center point and the neighbor point to constrain the input global feature. The first multi-layer perception is used to extract the attention weight of the global feature, and after cross multiplication with the global feature, the high-dimensional feature g w is extracted through the second multi-layer perception:

[0041]

[0042] The outputs of the two channels are added, and the features are aggregated to the center point through the max-pooling layer to obtain the final output X w :

[0043]

[0044] As preferred, the edge convolution layer input is the original point cloud feature, and the edge convolution layer workflow specifically comprises: S4.1 using the k-nearest neighbor search algorithm with the input of the point cloud feature to calculate the point cloud neighborhood, finding K neighborhood points around each point as the neighborhood of the point to obtain the neighborhood feature;

[0045] S4.2 expanding the dimension of the original point cloud feature to the same as the neighborhood feature, and calculating the difference between the original point cloud feature and the neighborhood feature after the dimension expansion, denoted as the relative feature;

[0046] S4.3 merging the relative feature and the expanded dimension of the original point cloud feature, and putting the output into the multi-layer perception and the max-pooling layer to obtain the final output.

[0047] The beneficial effects of the present application are:

[0048] 1. Two branches of edge convolution layer and gradient convolution layer are adopted to carry out point cloud feature extraction, so that the multi-modal features of the point cloud model can be effectively extracted, thereby improving the model precision;

[0049] 2. The attention module proposed does not increase the additional memory requirement, and the increased model inference time is very small, while the existing attention module needs to increase a large amount of memory and inference time;

[0050] 3. The current point cloud neighborhood attention module only uses the original position information of the point cloud, without considering the gradient information in the neighborhood, which leads to the low performance of the existing model;

[0051] 4. With the increase of the number of down-sampling layers, the receptive field of the model is also increasing, which is beneficial to the model to perceive different kinds of objects and help to improve the robustness of the model. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is a flowchart of the present application.

[0053] Figure 2 is a gradient graph convolution network flowchart of the present application.

[0054] Figure 3 is a gradient convolution layer flowchart of the present application.

[0055] Figure 4 is an edge convolution layer flowchart of the present application.

[0056] Figure 5 is a gradient convolution module schematic diagram of the present application. DETAILED DESCRIPTION

[0057] The technical solutions of the present application will be further specifically described below through examples and in combination with the drawings.

[0058] Example: the point cloud analysis method of the gradient attention graph convolution network of the present embodiment, as shown in Figure 2 , includes

[0059] Gradient graph convolution network:

[0060] The proposed gradient graph convolution network is a dual-channel model composed of several layers of gradient convolution layers and edge convolution layers in parallel, which obtains global shape descriptors by fusing spatial features of different dimensions. The input of the network is the original features of the point cloud. After the input is respectively passed through several edge convolution layers and gradient convolution layers with different sampling radii, point cloud embeddings of different modalities are obtained. After merging them, they are put into a gradient convolution layer with a larger sampling radius, and after a max-pooling layer, a global shape descriptor is obtained. The descriptor can be input into a classifier for three-dimensional shape recognition and three-dimensional segmentation tasks.

[0061] The gradient graph convolution network process is as follows:

[0062] The original features of the point cloud are respectively put into several edge convolution layers and gradient convolution layers to obtain multi-modal point cloud embeddings.

[0063] The multi-modal point cloud embeddings are merged, and the point cloud embeddings are put into a gradient convolution layer and a max-pooling layer to obtain the final output.

[0064] Gradient convolution layer:

[0065] As shown in Figure 3 , the gradient convolution layer includes two gradient convolution modules arranged in parallel and having different sampling radii. We use the original features of the point cloud and the absolute position contained in the original features as the input of the gradient convolution module. The gradient convolution module process is as follows:

[0066] 1. Find the indices of several center points P using the farthest point sampling method;

[0067] 2. Find the indices of M nearest neighbor points Q within the sampling radius r of the center point P using the ball query algorithm;

[0068] 3. Establish a directed graph between P and Q to obtain their relative positions and edge features;

[0069] 4. Calculate the zenith angle matrix and azimuth angle matrix between points by their relative positions, and construct a neighborhood gradient attention matrix using the two angles;

[0070] 5. Put the edge features into a multi-layer perceptron to extract them, multiply them with the neighborhood gradient attention matrix, and finally pass them through a max-pooling layer to obtain the output neighborhood features.

[0071] Gradient convolution module:

[0072] The gradient convolution module is the core of the gradient convolution layer. We use the global features and absolute position of the point cloud as the input of the gradient convolution module, w-1 is the number of layers, F w-1where w-1 is the output channel number of the w-1 layer, and n is the number of input point clouds. As shown in Equation 1, we summarize the key procedures of the weighted aggregation of edge features:

[0073] Equation 1 Gradient convolution module steps Input: global features X w-1 ; absolute positions P; sampling radius r; number of sampling points k

[0074] 1. Find the center point p using the ball query algorithm i k nearest neighbors within the sampling radius r of p, denoted as q ij ;

[0075] 2. Establish a directed graph to obtain the relative position and edge features

[0076] 3. Calculate the relative distance dis ij , zenith angle E ij , and azimuth angle A ij from the relative position ;

[0077] 4. Multiply the zenith angle E ij and the azimuth angle A ij to obtain the edge feature attention matrix G.

[0078] 5. After extracting high-dimensional features by MLP, multiply the edge feature attention matrix G, and finally use the sum function to aggregate the edge features to the center point.

[0079] Schematic diagram of the gradient convolution module. N represents the number of points, K represents the number of neighbor points of each center point, p represents the center point, and q represents a point in the neighborhood. The input (green box) is the absolute position (Nx3) and the global feature (NxC), C represents the number of input channels, and C1 represents the number of output channels. (a) Use the ball query algorithm on the input original point cloud to obtain the neighbor points of each point. (b) Establish a directed graph to obtain the edge features and the relative position vector. (c) Calculate the zenith angle and azimuth angle and construct the edge feature attention weight matrix, e pq is the relative distance vector between two points. (d) represents the aggregation process of the edge features and the global features. The relative coordinates and distances (Nx4) are the output of (b).

[0080] As shown in Figure 5 , first (a) uses the ball query algorithm to select k nearest neighbors within a radius r of each center point, denoted as q ij,i=1,...,n,j=1,...,k. Then (b) a directed graph is built in the neighborhood to obtain the relative position vector and edge features. By changing the size of r, we can flexibly adjust the receptive field of the gradient convolution module to capture local semantics at different scales. The gradient convolution module has two channels, which are used to constrain the aggregation process of edge features and global features, respectively. The upper channel calculates the two angles (E) between the relative position vector of the point cloud and the spatial coordinate system through (c). ij A ij And use these two included angles to explicitly construct the attention weight matrix G for edge features. Center point p i Its adjacent point q ij The relative position vectors between Represented as:

[0081]

[0082] Edge features Represented as:

[0083]

[0084] We use the original geometric relationships of the point cloud itself to constrain the aggregation of edge features, assigning different weights to each edge. Center point p i With adjacent point q ij Distance between ij It can be obtained through the relative position vector e ij get:

[0085]

[0086] like Figure 5 As shown in (c), we first establish an orthogonal three-dimensional spatial coordinate system XYZ, and then the relative position vector e ij Projected onto the XY plane. The angle between the relative position vector and the projection vector onto the XY plane in the Z direction is denoted as the zenith angle E. ij The angle between the projection vector and the Y-axis is called the azimuth angle A. ij They can be calculated based on the position vector and the distance between adjacent points:

[0087]

[0088] The gradient convolution module uses a weight function H to extract high-dimensional edge features. Distance constraints were added to the local point cloud: the shorter the distance from a neighboring point to its center point, the greater the weight of its corresponding edge. Where dis in Dis represents the set of distances from the center point to its k surrounding points. ij Representing the center point p idistance between the center point p ij and the neighboring points q The purpose is to aggregate the edge features to the center point p i . Its expression is as follows:

[0089]

[0090]

[0091] Where max represents the max-pooling operation, and sum represents the accumulation operation. The lower channel uses the relative distance between the center point and the neighboring points to constrain the global features of the input. The first multi-layer perceptron is used to extract the attention weight of the global feature, and after cross multiplication with the global feature, the high-dimensional feature is extracted through the second multi-layer perceptron:

[0092]

[0093] The outputs of the two channels are added, and the features are aggregated to the center point through the max-pooling layer, so as to realize the final output:

[0094]

[0095] The edge convolution layer is as shown in Figure 4 :

[0096] The input of the edge convolution layer is the original point cloud feature. The edge convolution layer process is as follows:

[0097] 1. Use the k-nearest neighbor search algorithm with the input of the point cloud feature to calculate the point cloud neighborhood, find K neighborhood points around each point as the neighborhood of the point, and obtain the neighborhood feature;

[0098] 2. Expand the dimension of the original point cloud feature to the same as the neighborhood feature, and calculate the difference between the expanded dimension of the original point cloud feature and the neighborhood feature, denoted as the relative feature;

[0099] 3. Merge the relative feature and the expanded dimension of the original point cloud feature, and put the output into the multi-layer perceptron and the max-pooling layer to obtain the final output.

[0100] The descriptor can be input into the classifier for 3D shape classification and 3D semantic segmentation tasks.

[0101] Specific embodiments of 3D shape classification are as follows:

[0102] The 3D shape classification network comprises several gradient convolution layers, edge convolution layers and a classification network. The gradient convolution branch comprises two gradient convolution layers and a separate MLP layer, and the MLP layer has a size of (64, 512). Each gradient convolution layer comprises an MLP layer and two gradient convolution modules in parallel and having different sampling radii. The first gradient convolution layer has MLP parameters of (3, 16), and the gradient convolution modules have sampling radii of 0.15 and 0.3 respectively; the second gradient convolution layer has MLP parameters of (64, 128), and the gradient convolution modules have sampling radii of 0.3 and 0.6 respectively. The edge convolution branch comprises three edge convolution layers, and the edge convolution layers have sizes of (3, 64), (64, 128) and (128, 256) respectively. Then, the two branches are fused into a gradient convolution layer, and the gradient convolution layer has no MLP and has gradient convolution modules with sampling radii of 0.6 and 0.8 respectively. Finally, a classification result is obtained through a classification network with a size of (1536, 10).

[0103] A specific embodiment of 3D semantic segmentation is performed.

[0104] The 3D semantic segmentation network comprises several gradient convolution layers, edge convolution layers and a segmentation network. The gradient convolution branch comprises two gradient convolution layers and a separate MLP layer, and the MLP layer has a size of (32, 64). Each gradient convolution layer comprises an MLP layer and two gradient convolution modules in parallel and having different sampling radii. The first gradient convolution layer has MLP parameters of (3, 16), and the gradient convolution modules have sampling radii of 0.15 and 0.3 respectively; the second gradient convolution layer has MLP parameters of (32, 32), and the gradient convolution modules have sampling radii of 0.3 and 0.6 respectively. The edge convolution branch comprises two edge convolution layers, and the edge convolution layers have sizes of (3, 64) and (64, 64) respectively. Then, the two branches are fused into a gradient convolution layer, and the gradient convolution layer has no MLP and has gradient convolution modules with sampling radii of 0.6 and 0.8 respectively. After passing through an MLP layer with a size of (384, 1024) again, the output features are put into a max-pooling layer, and then put into an MLP layer with a size of (1024, 10) to obtain a segmentation result.

[0105] Protection points of the present proposal

[0106] The gradient-based point cloud attention module: the present proposal explicitly calculates the zenith angle and the azimuth angle to construct a neighborhood attention matrix, and therefore, the present proposal first introduces gradient information to constrain the neighborhood aggregation process;

[0107] The present proposal is a dual-channel network, which extracts multi-modal information of point clouds through branch networks with different structures, which is conducive to improving the network performance;

[0108] The present proposal uses a ball query method for downsampling, however, in order to ensure the effect of neighborhood aggregation, the receptive field is constantly improved while downsampling, so that the network of the present application can better perceive the shape of the point cloud.

Claims

1. A point cloud analysis method of a gradient attention map convolutional network, characterized in that, The method comprises the following steps: S1: constructing a gradient convolution layer; the gradient convolution layer comprises two gradient convolution modules arranged in parallel and having different sampling radius sizes; the original feature of the point cloud and the absolute position contained in the original feature are taken as the input of the gradient convolution module, and the neighborhood feature is taken as the output; S2: constructing an edge convolution layer; S3: placing the original feature of the point cloud into a plurality of edge convolution layers and gradient convolution layers respectively; S4: obtaining a multi-modal point cloud embedding and merging the same; S5: placing the merged result into a gradient convolution layer and a maximum pooling layer to obtain a final output. 2.The point cloud analysis method of the gradient attention map convolutional network according to claim 1, characterized in that, The step S1: the gradient convolution layer comprises two gradient convolution modules arranged in parallel and having different sampling radius sizes; the original feature of the point cloud and the absolute position contained in the original feature are taken as the input of the gradient convolution module, and the gradient convolution module construction process comprises the following steps: S1.1: finding the indexes of a plurality of center points P by using a farthest point sampling method; S1.2: finding the indexes of M nearest neighbor points Q within a sampling radius r of the center point P by using a ball query algorithm; S1.3: establishing a directed graph between P and Q to obtain the relative position and edge feature therebetween; S1.4: calculating the zenith angle matrix and azimuth angle matrix between the points through the relative position between the points, and constructing a neighborhood gradient attention matrix by using the two angles; S1.5: placing the edge feature into a multilayer perceptron, multiplying the edge feature with the neighborhood gradient attention matrix, and finally obtaining the output neighborhood feature through a maximum pooling layer.

3. The point cloud analysis method of the gradient attention map convolutional network according to claim 1, characterized in that, The gradient convolution module is the core of the gradient convolution layer, which extracts the global features of the point cloud and absolute position As the input of the gradient convolution module, w-1 is the number of layers, F w-1 is the output channel number of w-1 layers, n is the number of input point clouds, and the input global feature X w-1 ; The absolute position P; the sampling radius r; The gradient convolution module workflow of the weighted aggregated edge feature specifically comprises: S3.1 Find the center point p using the ball query algorithm i ∈P Sample k nearest neighbors within radius r of p, denoted q ij ; S3.2 Build a directed graph to get relative positions and edge features S3.3 by relative position Calculate relative distance dis ij Zenith angle E ij and Azimuth angle A ij ; S3.4 obtain the zenith angle E ij cross-multiply the azimuth angle A ij to obtain the edge feature attention matrix G; S3.5 After extracting high-dimensional features by MLP, the edge feature attention matrix G is multiplied, and finally the sum function is used to aggregate the edge features to the center point.

4. The point cloud analysis method of a gradient attention map convolutional network according to claim 3, characterized in that, The step S3.1 specifically comprises: firstly, finding a plurality of center points p by a farthest point sampling method i ∈P, and then selecting k nearest neighbor points in a radius r range of each center point by using a ball query algorithm, denoted as Q ij ,i=1,…,n;j=1,…,k.

5. The point cloud analysis method of a gradient attention map convolutional network according to claim 3, characterized in that, The step S3.2 specifically comprises, A directed graph is established in the neighborhood to obtain a relative position vector and an edge feature; By changing the size of the sampling radius r, the receptive field of the gradient convolution module can be flexibly adjusted, so that local semantics of different scales can be captured, Central point p i Relative position vector between its neighboring point q ij is expressed as:​ Edge features is represented as: 6.The point cloud analysis method of a gradient attention map convolutional network according to claim 3, characterized in that, S3.3 specifically comprises using the original geometric relationship of the point cloud itself to constrain the aggregation of edge features, giving each edge a different weight, the center point p i The distance dis between the center point p and its adjacent point q ij ij Through the relative position vector Get:​ Subsequently, an orthogonal three-dimensional spatial coordinate system XYZ is established, and then the relative position vectors are... The angle between the relative position vector and the projection vector onto the XY plane in the Z direction is denoted as the zenith angle E. ij The angle between the projection vector and the Y-axis is called the azimuth angle A. ij The zenith angle E is calculated based on the position vector and the distance between adjacent points. ij and azimuth A ij :

7. The point cloud analysis method of a gradient attention map convolutional network according to claim 3, characterized in that, The step S3.4 specifically comprises that the gradient convolution module extracts high-dimensional edge features by using a weight function H, M(p, q) adds a distance constraint in a local point cloud, and represents that the shorter the distance from a certain neighbor point to the center point, the greater the weight of the corresponding edge, wherein max(dis in ) represents the distance from the center point to the farthest neighbor point, dis ij represents the distance from the center point p i to the jth neighbor point q ij , and the purpose of the aggregation function G(p, q) is to aggregate the edge features to the center point p i , which is expressed as follows: Where max represents a maximum pooling operation, and sum represents an accumulation operation. 8.The point cloud analysis method of a gradient attention map convolutional network according to claim 3, characterized in that, The step S3.5 specifically comprises that the lower passage uses the relative distance between the center point and the adjacent point to constrain the input global feature; the first multi-layer perception machine is used to extract the attention weight of the global feature, and after cross multiplication with the global feature, the high-dimensional feature g is extracted through the second multi-layer perception machine w : The outputs of the two channels are added, and the features are aggregated to the center point through a max-pooling layer to obtain the final output X w : 9.The point cloud analysis method of the gradient attention map convolutional network according to claim 1, characterized in that, The edge convolution layer input is an original point cloud feature, and the edge convolution layer workflow specifically comprises: S4.1: using a k-nearest neighbor search algorithm with the point cloud feature as the input to calculate the point cloud neighborhood, finding K neighborhood points around each point as the neighborhood of the point, and obtaining a neighborhood feature; S4.2: expanding the dimension of the original point cloud feature to the same dimension as the neighborhood feature, and calculating the difference between the expanded dimension original point cloud feature and the neighborhood feature, denoted as a relative feature; S4.3: merging the relative feature and the expanded dimension original point cloud feature, and placing the output into a multilayer perceptron and a maximum pooling layer to obtain a final output.

Images