Point cloud registration method of attention mechanism

Abstract

The application belongs to the field of three-dimensional digital measurement, and particularly relates to a point cloud registration method with an attention mechanism, which down-samples input point clouds and target point clouds to facilitate improvement of registration accuracy and efficiency. Spatial and channel attention mechanisms are used to focus on the overlapping area of the two point clouds. The method performs dot product operation on the weighted point cloud features to obtain a correlation matrix corresponding to the points, and then obtains a weighting matrix. The rigid transformation matrix is obtained by singular value decomposition algorithm using the weighting matrix and the two point cloud features, so that the fast and accurate point cloud registration task is realized. The application focuses on the similar areas between the point clouds to be registered by introducing the attention mechanism in deep learning, and extracts the feature information to perform correlation measurement, so as to solve the problem of point cloud registration.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of three-dimensional digital measurement, and particularly relates to a point cloud registration method based on attention mechanism. BACKGROUND

[0002] With the continuous development of digital technology and its positive application in various fields, the aviation manufacturing field also continuously proposes the demand for digital and automated detection. The difference between the actual manufacturing size and the theoretical design size of the aircraft parts directly reflects the quality of aircraft manufacturing. However, the existing detection methods are still conventional measuring tools such as vernier calipers, plug gauges and depth gauges, resulting in low measurement accuracy, low measurement efficiency, intensive manual work and low digitalization, etc. This seriously hinders the development of the aviation manufacturing industry and restricts the improvement of the quality of aviation products.

[0003] At present, three-dimensional measurement technology is becoming mature and has been successfully applied in various industrial fields. Among them, the structured light three-dimensional measurement technology is widely used because of its high measurement accuracy and efficiency. This technology mainly obtains point cloud data of the surface of the part through a three-dimensional scanning device, and compares it with the designed digital model to realize digital measurement. However, the parts in the aviation field are usually large, and it is difficult for the three-dimensional scanning device to obtain the complete point cloud data of the part at one time. Therefore, the point cloud data under different angles needs to be spliced and registered to obtain the point cloud data of the entire part, and the actual measurement point cloud and its design digital model also need to be registered to measure the deviation between them. Then, the point cloud registration technology is the most critical step in three-dimensional measurement technology. The point cloud registration accuracy will directly affect the measurement deviation between the measured value and the digital model value, resulting in the failure of the size measurement of the part. Therefore, many point cloud registration methods have been proposed to improve the registration accuracy between the measured point cloud and the measured point cloud, as well as the measured point cloud and the digital model. Patent No. CN 109472816 A discloses a point cloud registration method. This method uses a strategy of rough matching first and then fine matching to realize point cloud registration. First, it realizes adaptive field matching by judging whether the feature regions of the initial matching point pairs are the same and whether the GAN shapes of their fields are the same. Then, the initial positions of the point pairs are used for accurate registration. Patent No. CN 111311651 A discloses a point cloud registration method and device. Similarly, this method also uses the strategy of rough registration first and then fine registration. This method segments the point cloud of rough registration, then performs fine registration on the segmented point cloud, and repeatedly performs the above operation until the preset registration condition is met. However, the above methods are traditional point cloud registration algorithms, which are difficult to effectively utilize the text and semantic information of the point cloud, thereby making it difficult to further improve the point cloud registration accuracy. SUMMARY

[0004] To address the aforementioned technical problems, this invention discloses a point cloud registration method based on an attention mechanism. This method can effectively improve the registration accuracy and speed of point clouds, thereby enhancing the efficiency and accuracy of digital measurement of components.

[0005] To achieve the above-mentioned technical effects, the present invention is implemented through the following technical solution:

[0006] A point cloud registration method based on an attention mechanism includes the following steps:

[0007] Step 1. Obtain the input point cloud and the target point cloud;

[0008] Step 2. Downsample the input point cloud and the target point cloud;

[0009] Step 3. Use spatial attention mechanisms to focus on the overlapping area between the input point cloud and the target point cloud;

[0010] Step 4. Apply channel attention to the point cloud features of the overlapping region, focusing on the features corresponding to each channel point;

[0011] Step 5. Obtain the similarity matrix by aligning the input of the attention mechanism with the target point cloud features and performing a dot product operation.

[0012] Step 6. Obtain the first n maximum values ​​of the similarity matrix to construct a mask for the overlapping region of the two point clouds, and normalize the mask matrix to obtain a weighted matrix;

[0013] Step 7. Combine the weighted matrix with the two point cloud features mentioned above and use the singular value decomposition algorithm to obtain the rigid body transformation matrix;

[0014] Step 8: Point cloud registration complete.

[0015] Furthermore, the point cloud to be registered is the input point cloud. With target point cloud Its coordinate matrix can be represented as follows: The encoder is used to downsample the point cloud to be registered. The point cloud coordinate matrix after downsampling is: and Point cloud registration requires estimating the rigid body transformation matrix [R,T] such that y = Rx + T.

[0016] Furthermore, a feature extraction network is used to extract feature information from the input point cloud and the target point cloud respectively, and the feature parameters are shared, as shown below:

[0017] f' X =Net(P' X )

[0018] f' Y =Net(P' Y )

[0019] wherein, d denotes the number of feature channels; Net(·) denotes a feature extraction network.

[0020] Further, a spatial attention mechanism is used to pay more attention to the overlapping region between the input point cloud and the target point cloud, facilitating the calculation of the correlation matrix, as follows:

[0021]

[0022]

[0023] wherein, * denotes a convolution operation; ω * ,b * respectively denote the weights and bias of the one-dimensional convolution layer. After weighting by the spatial attention mechanism, the point cloud features pay more attention to the features of the overlapping region of the two point clouds. After weighting by the attention mechanism, the features of the two point clouds can be represented as:

[0024]

[0025]

[0026] In the formula, D(...,...) denotes a concatenation operation; and • denotes pixel-wise multiplication.

[0027] Further, a channel attention mechanism is used to pay attention to the feature information of the overlapping region in each channel of the input point cloud feature F' X and the target point cloud feature F' Y The channel attention mechanism can be represented as:

[0028]

[0029]

[0030] In the formula, AvgPool(·) denotes an average pooling operation. The features weighted by the channel attention mechanism are:

[0031]

[0032]

[0033] In the formula, ⊙ denotes feature-wise multiplication.

[0034] Further, after the input and target point cloud features pay attention to the features of the overlapping region by the attention mechanism, a dot product operation is used to obtain the similarity matrix of the two point cloud features, as follows:

[0035]

[0036] In the formula,

[0037] Further, the first n maximum values of the similarity measure matrix S are obtained, a mask matrix of the overlapping region of the two point cloud features is constructed, and the mask matrix is normalized to calculate a weight matrix C.

[0038] Further, the rigid body change matrix [R pre ,T pre ] is predicted by singular value decomposition algorithm (SVD) in combination with the weight matrix C, the input point cloud feature and the target point cloud feature, and then the point cloud registration task is realized.

[0039] The application has the advantages of:

[0040] The application focuses on the similar regions between the point clouds to be registered by introducing the attention mechanism in deep learning, extracts the feature information of the similar regions, and measures the correlation to solve the problem of point cloud registration. The application discloses a point cloud registration method based on attention mechanism to improve the accuracy and efficiency of point cloud registration. The method uses spatial and channel attention to focus on the overlapping regions between the input point cloud and the target point cloud, can fully utilize the text and semantic information of the point cloud, and further improves the registration accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a flowchart of the application.

[0042] Figure 2 is an input point cloud and a target point cloud. DETAILED DESCRIPTION

[0043] In order to more clearly illustrate the technical solutions provided by the application, the application will be further described below in combination with the drawings and embodiments. It should be pointed out that the provided embodiments are only a part of the embodiments of the application, not all the embodiments, and therefore should not be regarded as limiting the protection scope. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0044] Embodiment 1

[0045] As Figure 1 shown, a point cloud registration method based on attention mechanism includes the following steps:

[0046] Step 1. Obtain an input point cloud and a target point cloud;

[0047] Step 2. Downsample the input point cloud and the target point cloud;

[0048] Step 3. Use spatial attention mechanism to focus on the overlapping region of the input point cloud and the target point cloud;

[0049] Step 4. Channel attention is applied to the point cloud features in the overlapping area to focus on the corresponding features of each channel point;

[0050] Step 5. The similarity matrix is obtained by performing a dot product operation on the input of the attention mechanism and the target point cloud feature alignment;

[0051] Step 6. The top n maximum values of the similarity matrix are obtained to construct a two-point cloud overlapping area mask, and the mask matrix is normalized to obtain a weighted matrix;

[0052] Step 7. The weighted matrix is combined with the two point cloud features mentioned above to obtain the rigid transformation matrix by singular value decomposition algorithm;

[0053] Step 8. Point cloud registration is completed.

[0054] The point cloud to be registered is the input point cloud and the target point cloud The coordinate matrix can be represented as and The encoder is used to downsample the point cloud to be registered, and the coordinate matrix of the downsampled point cloud is and The rigid transformation matrix [R, T] needs to be estimated for point cloud registration, so that y = Rx + T.

[0055] The feature extraction network is used to extract the feature information of the input point cloud and the target point cloud respectively, and the feature parameters are shared, as follows:

[0056] f' X = Net(P' X )

[0057] f' Y = Net(P' Y )

[0058] wherein, d represents the number of feature channels; Net(·) represents the feature extraction network.

[0059] The spatial attention mechanism is used to pay more attention to the overlapping area between the input point cloud and the target point cloud, which facilitates the calculation of the correlation matrix, as follows:

[0060]

[0061]

[0062] wherein, * represents convolution operation; ω * ,b *respectively represent the weights and biases of the one-dimensional convolution layer, and the point cloud features weighted by the spatial attention mechanism pay more attention to the features of the overlapping area of the two point clouds. The two point cloud features weighted by the attention mechanism can be represented as:

[0063]

[0064]

[0065] In the formula, D(...,...) represents the concatenation operation; i represents pixel-by-pixel multiplication.

[0066] The channel attention mechanism is used to pay attention to the input point cloud feature F' X and the target point cloud feature F' Y The feature information of the overlapping area in each channel, and the channel attention mechanism can be represented as:

[0067]

[0068]

[0069] In the formula, AvgPool(·) represents the average pooling operation, and the feature weighted by the channel attention mechanism is:

[0070]

[0071]

[0072] In the formula, represents feature-by-feature multiplication.

[0073] The input and target point cloud features pay attention to the feature of the overlapping area by the attention mechanism, and then use the dot product operation to obtain the similarity matrix of the two point cloud features, as follows:

[0074]

[0075] In the formula,

[0076] The first n maximum values of the similarity matrix S are obtained, the mask matrix of the overlapping area of the two point cloud features is constructed, and the mask matrix is normalized to calculate the weight matrix C.

[0077] The weight matrix C, the input point cloud feature, and the target point cloud feature are combined to predict the rigid transformation matrix [R pre , T pre ] by the singular value decomposition algorithm (SVD), and then the point cloud registration task is realized.

[0078] Embodiment 2

[0079] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0080] This invention discloses a point cloud registration method based on an attention mechanism, as follows: Figure 1 As shown, it includes the following steps.

[0081] Step 1: Use CAD digital modeling to obtain the input point cloud, such as... Figure 2 As shown in the image on the right, Gaussian noise is added to the input point cloud to simulate the measured point cloud, as shown below. Figure 2 The airplane is shown on the left. The point clouds are represented as follows: and The input point cloud and the target point cloud have a relative positional relationship. A rigid body transformation matrix needs to be estimated to register the two point clouds for dimensional measurement of parts, complete machines, etc.

[0082] Step 2: Downsample the point cloud to be registered using the existing deep learning network KPCon. The point cloud coordinate matrix after downsampling is as follows: and

[0083] Step 3: Use a feature extraction network to extract feature information from the input point cloud and the target point cloud respectively, and share the feature parameters, as shown below:

[0084] f' X =Net(P' X )

[0085] f' Y =Net(P' Y )

[0086] in, d represents the number of feature channels; Net(·) represents the feature extraction network.

[0087] Step 4: Use spatial attention mechanism to focus more on the overlapping region between the input point cloud and the target point cloud, which facilitates the calculation of the correlation matrix, as shown below:

[0088]

[0089]

[0090] Where * represents the convolution operation; ω * ,b * Let represent the weights and biases of the one-dimensional convolutional layer, respectively. After weighting through the spatial attention mechanism, the point cloud features will focus more on the features of the overlapping region between the two point clouds. Therefore, the weighted features of the two point clouds after the attention mechanism can be represented as:

[0091]

[0092]

[0093] In the formula, D(...,...) represents the concatenation operation; • represents pixel-by-pixel multiplication.

[0094] Step 5: Focus on the input point cloud feature F' using the channel attention mechanism X and the target point cloud feature F Y The feature information of the overlapping region in each channel. The channel attention mechanism can be represented as:

[0095]

[0096]

[0097] In the formula, AvgPool(i) represents the average pooling operation. Therefore, the feature after the channel attention mechanism weighting is:

[0098]

[0099]

[0100] In the formula, ⊙ represents feature-by-feature multiplication.

[0101] Step 6: After the input and target point cloud features focus on their overlapping region features using the attention mechanism, use the dot product operation to obtain the similarity measure matrix of the two point cloud features, as follows:

[0102]

[0103] In the formula,

[0104] Step 7: Find the top n maximum values of the similarity measure matrix S to construct the mask matrix of the overlapping region of the two point cloud features. At the same time, normalize the mask matrix to calculate the weight matrix C.

[0105] Step 8: Combine the weight matrix C, the input point cloud feature, and the target point cloud feature to predict the rigid transformation matrix [R pre , T pre ] using the singular value decomposition algorithm (SVD), and then realize the point cloud registration task.

[0106] Step 9: Use the average of the rotation error and the translation error between the predicted value and the true value of the rigid transformation matrix as the loss function to update the network parameters, as follows:

[0107]

[0108]

[0109] In the formula, R, R pre respectively represent real rotation matrix and predicted rotation matrix; T, T pre respectively represent real translation matrix and predicted translation matrix.

[0110] Step 10: optimize the network model by using Adam optimization algorithm, and use Python language and Pytorch deep learning framework to build the whole point cloud registration model.

[0111] The part of the application that has not been described in detail belongs to the known technology of those skilled in the art.

[0112] The above is a further detailed description of the application in combination with specific preferred embodiments, and the specific embodiments of the application cannot be limited to these descriptions. Other embodiments obtained by those skilled in the art without departing from the technical solutions of the application should be included in the protection scope of the application.

Claims

1. A point cloud registration method based on an attention mechanism, characterized in that: Includes the following steps: Step 1. Obtain the input point cloud and the target point cloud; Step 2. Downsample the input point cloud and the target point cloud; Step 3. Use spatial attention mechanisms to focus on the overlapping area between the input point cloud and the target point cloud; Step 4. Apply channel attention to the point cloud features of the overlapping region, focusing on the features corresponding to each channel point; Step 5. Obtain the similarity matrix by aligning the input of the attention mechanism with the target point cloud features and performing a dot product operation. Step 6. Obtain the top N maximum values ​​of the similarity matrix to construct a mask for the overlapping region of the two point clouds, and normalize the mask matrix to obtain a weighted matrix; Step 7. Combine the weighted matrix with the two point cloud features mentioned above and use the singular value decomposition algorithm to obtain the rigid body transformation matrix; Step 8: Point cloud registration complete; In step 3, a spatial attention mechanism is used to focus more on the overlapping region between the input point cloud and the target point cloud, which facilitates the calculation of the correlation matrix, as shown below: in, and To extract features from the downsampled coordinate matrix. Indicates the convolution operation; Representing the weights and biases of a one-dimensional convolutional layer, respectively, after weighting through a spatial attention mechanism, the point cloud features will focus more on the features of the overlapping region between the two point clouds. The weighted features of the two point clouds after the attention mechanism can be represented as: In the formula, Indicates a cascading operation; This indicates pixel-by-pixel multiplication; Focusing on input point cloud features using channel attention mechanism Features of the target point cloud The feature information of the overlapping regions in each channel, the channel attention mechanism can be represented as: In the formula, The features after average pooling operation and weighting by channel attention mechanism are as follows: In the formula, This indicates multiplication by feature.

2. The point cloud registration method based on an attention mechanism according to claim 1, characterized in that: The point cloud to be registered is the input point cloud. With target point cloud Their coordinate matrices can be represented as follows: and ; The encoder is used to downsample the point cloud to be registered. The point cloud coordinate matrix after downsampling is: and Point cloud registration requires estimating the rigid body transformation matrix. , making .

3. The point cloud registration method based on an attention mechanism according to claim 2, characterized in that: Feature extraction networks are used to extract feature information from the input point cloud and the target point cloud respectively, and the feature parameters are shared, as shown below: in, , , Indicates the number of feature channels; This represents a feature extraction network.

4. The point cloud registration method based on an attention mechanism according to claim 1, characterized in that: After the input and target point cloud features are processed by an attention mechanism to focus on their overlapping region features, a dot product operation is used to obtain the similarity metric matrix of the two point cloud features, as shown below: In the formula, .

5. The point cloud registration method based on an attention mechanism according to claim 4, characterized in that: Find the similarity metric matrix The top N maximum values ​​are used to construct a mask matrix for the overlapping regions of the two point cloud features. Simultaneously, the mask matrix is ​​normalized to calculate the weight matrix. .

6. The point cloud registration method based on an attention mechanism according to claim 5, characterized in that: Combined with weight matrix The input point cloud features and the target point cloud features are used to predict the rigid body transformation matrix using the singular value decomposition algorithm. This enables the point cloud registration task.

Images