Point cloud registration model training method, device, equipment and storage medium

Abstract

The present disclosure relates to a point cloud registration model training method, device, equipment and storage medium, the method comprising: obtaining a first reference point cloud, a second reference point cloud and a real point cloud of a sampling object; transforming the real point cloud based on preset transformation parameters to obtain a transformed point cloud; determining the first reference point cloud, the second reference point cloud and the first registration relationship between the first reference point cloud and the second reference point cloud as a first training sample group, and determining the first reference point cloud, the real point cloud, the transformed point cloud and the transformation parameters as a second training sample group; training a pre-established point cloud registration model based on the first training sample group and the second training sample group, respectively, to obtain a target point cloud registration model. The present disclosure can realize the training of a point cloud registration model without manual annotation by establishing a training sample group containing an absolute transformation relationship and a relative transformation relationship, thereby saving labor costs while improving the accuracy of local point cloud registration.

Description

Technical Field

[0001] The present disclosure relates to the field of computer vision technology, and in particular to a point cloud registration model training method, device, equipment and storage medium. Background Art

[0002] Point cloud registration is a basic task in computer vision, which aims to align multiple point clouds collected from different perspectives or at different times in the same reference coordinate system by finding the geometric transformation relationship between the point clouds. With the continuous promotion of 3D acquisition equipment, point cloud registration has been applied in many fields such as 3D reconstruction, robot navigation, industrial robots, and augmented reality (AR).

[0003] Among the existing point cloud registration methods, the point cloud registration method based on feature matching can only obtain good registration results when targeting a complete point cloud, but a 3D camera can usually only capture a partial point cloud under a single viewing angle, resulting in limited registration accuracy; and the point cloud registration method based on deep learning performs well in the task of local point cloud registration, but this method is limited by the annotation of sample data, the workload of manual annotation is large, and the manually annotated point cloud conversion relationship may have systematic errors or random errors, thus affecting the final registration accuracy of the model. Therefore, how to improve the accuracy of the deep learning model in the task of local point cloud registration while saving the labor cost of manual annotation is a technical problem that needs to be solved. Summary of the invention

[0004] In order to solve the above technical problems, the present disclosure provides a point cloud registration model training method, device, equipment and storage medium.

[0005] A first aspect of an embodiment of the present disclosure provides a point cloud registration model training method, the method comprising:

[0006] Acquire a first reference point cloud, a second reference point cloud, and a real point cloud of the sampling object, wherein the first reference point cloud is the overall point cloud of the sampling object, the second reference point cloud is a local point cloud selected from the first reference point cloud based on a preset posture, and the real point cloud is the acquired local point cloud of the sampling object;

[0007] Transforming the real point cloud based on preset transformation parameters to obtain a transformed point cloud;

[0008] Determine the first reference point cloud, the second reference point cloud, and a first registration relationship between the first reference point cloud and the second reference point cloud as a first training sample group, and determine the first reference point cloud, the real point cloud, the transformed point cloud, and the transformation parameters as a second training sample group, wherein the first registration relationship is determined according to the preset posture;

[0009] The pre-established point cloud registration model is trained based on the first training sample group and the second training sample group respectively to obtain a target point cloud registration model.

[0010] A second aspect of an embodiment of the present disclosure provides a point cloud registration model training device, the device comprising:

[0011] A first acquisition module is used to acquire a first reference point cloud, a second reference point cloud, and a real point cloud of the sampling object, wherein the first reference point cloud is the overall point cloud of the sampling object, the second reference point cloud is a local point cloud selected from the first reference point cloud based on a preset posture, and the real point cloud is the local point cloud of the sampling object obtained by acquisition;

[0012] A transformation module, used for transforming the real point cloud based on preset transformation parameters to obtain a transformed point cloud;

[0013] a determination module, configured to determine the first reference point cloud, the second reference point cloud, and a first registration relationship between the first reference point cloud and the second reference point cloud as a first training sample group, and determine the first reference point cloud, the real point cloud, the transformed point cloud, and the transformation parameters as a second training sample group, wherein the first registration relationship is determined according to the preset posture;

[0014] The training module is used to train the pre-established point cloud registration model based on the first training sample group and the second training sample group respectively to obtain a target point cloud registration model.

[0015] A third aspect of an embodiment of the present disclosure provides a computer device, comprising a memory and a processor, and a computer program, wherein the computer program is stored in the memory, and when the computer program is executed by the processor, the point cloud registration model training method as described in the first aspect above is implemented.

[0016] A fourth aspect of an embodiment of the present disclosure provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the point cloud registration model training method as described in the first aspect above is implemented.

[0017] Compared with the prior art, the technical solution provided by the embodiments of the present disclosure has the following advantages:

[0018] In the point cloud registration model training method, apparatus, device and storage medium provided in the embodiments of the present disclosure, a first reference point cloud, a second reference point cloud and a real point cloud of a sampling object are obtained, wherein the first reference point cloud is the overall point cloud of the sampling object, the second reference point cloud is a local point cloud selected from the first reference point cloud based on a preset posture, and the real point cloud is the local point cloud of the sampling object obtained by collection, the real point cloud is transformed based on preset transformation parameters to obtain a transformed point cloud, the first reference point cloud, the second reference point cloud and a first registration relationship between the first reference point cloud and the second reference point cloud are determined as a first training sample group, and the first reference point cloud, the real point cloud and the real point cloud are transformed based on preset transformation parameters to obtain a transformed point cloud. The point cloud, transformed point cloud and transformation parameters are determined as the second training sample group, wherein the first registration relationship is determined according to a preset posture, and a pre-established point cloud registration model is trained based on the first training sample group and the second training sample group respectively to obtain a target point cloud registration model, which can obtain the absolute transformation relationship and relative transformation relationship between the point cloud data without manually annotating the point cloud data to construct the training sample data, and then train the point cloud registration model based on the first training sample group containing the absolute transformation relationship and the second training sample group containing the relative transformation relationship, thereby saving labor costs while improving the accuracy of local point cloud registration. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0020] 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.

[0021] Figure 1 is a flow chart of a point cloud registration model training method provided by an embodiment of the present disclosure;

[0022] Figure 2 is a flow chart of a method for screening a second reference point cloud provided by an embodiment of the present disclosure;

[0023] Figure 3 is a flow chart of a method for adjusting model parameters provided by an embodiment of the present disclosure;

[0024] Figure 4 is a flow chart of another method for adjusting model parameters provided by an embodiment of the present disclosure;

[0025] Figure 5 is a flow chart of a method for registering point clouds provided by an embodiment of the present disclosure;

[0026] Figure 6 is a flow chart of a method for calculating a registration result provided by an embodiment of the present disclosure;

[0027] Figure 7 is a flow chart of a method for fine point cloud registration provided by an embodiment of the present disclosure;

[0028] Figure 8 is a structural schematic diagram of a point cloud registration model training device provided by an embodiment of the present disclosure;

[0029] Fig. 9 It is a structural diagram of a computer device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0030] In order to more clearly understand the above-mentioned objectives, features and advantages of the present disclosure, the scheme of the present disclosure will be further described below. It should be noted that the embodiments of the present disclosure and the features in the embodiments can be combined with each other without conflict.

[0031] In the following description, many specific details are set forth to facilitate a full understanding of the present disclosure, but the present disclosure may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present disclosure, rather than all of the embodiments.

[0032] It should be understood that the various steps described in the method embodiments of the present disclosure may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this respect.

[0033] Figure 1 1 is a flow chart of a point cloud registration model training method provided by an embodiment of the present disclosure, and the method can be executed by a point cloud registration model training device. Figure 1 As shown, the point cloud registration model training method provided in this embodiment includes the following steps:

[0034] S101. Obtain a first reference point cloud, a second reference point cloud and a real point cloud of a sampling object, wherein the first reference point cloud is the overall point cloud of the sampling object, the second reference point cloud is a local point cloud selected from the first reference point cloud based on a preset posture, and the real point cloud is the collected local point cloud of the sampling object.

[0035] The first reference point cloud in the disclosed embodiment can be understood as the overall point cloud of the surface of the sampling object. For example, the first reference point cloud can be obtained by randomly sampling a pre-established three-dimensional model of the sampling object; the second reference point cloud is a partial point cloud in the first reference point cloud, and the second reference point cloud can be screened from the first reference point cloud based on a preset posture by specifying a preset posture of the three-dimensional camera viewpoint, that is, the first reference point cloud and the second reference point cloud are both simulated data synthesized by technical means, rather than real data obtained by observing and collecting the sampling object, wherein the preset posture can be obtained by randomly taking values ​​within a certain range; the real point cloud is data obtained by actually sampling the sampling object using a point cloud acquisition device, such as a three-dimensional camera. Since the three-dimensional camera can only capture a local point cloud under a single viewing angle, the real point cloud is a local point cloud of the sampling object.

[0036] In the disclosed embodiment, the three-dimensional coordinates of each point in the three groups of point clouds, namely, the first reference point cloud, the second reference point cloud and the real point cloud, correspond to three different coordinate systems respectively.

[0037] In the disclosed embodiment, the point cloud registration model training device can determine a sampling object when it is necessary to train the point cloud registration model, and obtain the first reference point cloud, the second reference point cloud and the real point cloud of the sampling object, wherein the point cloud registration model training device can, after obtaining the first reference point cloud, determine, based on a randomly determined preset pose, for each point in the first reference point cloud whether the point can be captured by the three-dimensional camera in a preset pose, and determine the point set that can be captured as the second reference point cloud.

[0038] In an exemplary implementation of the disclosed embodiment, the point cloud registration model training device can obtain the normal vector corresponding to each point in the point cloud after obtaining the first reference point cloud, the second reference point cloud and the real point cloud of the sampling object. For example, the normal vector can be calculated by the three-dimensional data processing library open3d.

[0039] S102: Perform transformation processing on the real point cloud based on preset transformation parameters to obtain a transformed point cloud.

[0040] The transformation parameters in the embodiments of the present disclosure may be randomly determined within a certain range of values.

[0041] In the disclosed embodiment, the point cloud registration model training device can, after obtaining the real point cloud of the sampling object, transform the real point cloud according to the pre-acquired transformation parameters. Specifically, each point in the real point cloud can be transformed, and the point cloud composed of the transformed points can be determined as the transformed point cloud.

[0042] Optionally, the transformation processing may include translation transformation processing and / or rotation transformation processing. The point cloud registration model training device may perform translation transformation processing on the real point cloud based on preset translation transformation parameters, or may perform rotation transformation processing on the real point cloud based on preset rotation transformation parameters. It may also first perform translation transformation processing on the real point cloud based on preset translation transformation parameters, and then perform rotation transformation processing on the real point cloud based on preset rotation transformation parameters, and determine the point cloud after transformation processing as the transformed point cloud.

[0043] S103. Determine the first reference point cloud, the second reference point cloud, and the first registration relationship between the first reference point cloud and the second reference point cloud as a first training sample group, and determine the first reference point cloud, the real point cloud, the transformed point cloud, and the transformation parameters as a second training sample group, wherein the first registration relationship is determined according to a preset posture.

[0044] The registration relationship in the embodiment of the present disclosure can be understood as a point-level pairing relationship and a relative transformation relationship between two groups of point clouds.

[0045] In the embodiment of the present disclosure, the point cloud registration model training device can determine the first reference point cloud and the second reference point cloud according to the preset posture after determining the preset posture corresponding to the first reference point cloud, and determine the first reference point cloud, the second reference point cloud and the first registration relationship as the first training sample group used when training the point cloud registration model, and determine the first reference point cloud, the real point cloud, the transformed point cloud and the transformation parameters as the second training sample group used when training the point cloud registration model.

[0046] In an exemplary implementation of the embodiments of the present disclosure, the preset pose may include the position coordinates and Euler angles of the three-dimensional camera viewpoint in the coordinate system where the first reference point cloud is located. The relative transformation relationship between the first reference point cloud and the second reference point cloud is the same as the relative transformation relationship between the coordinate origin of the coordinate system where the first reference point cloud is located and the three-dimensional camera viewpoint. Therefore, the translation parameters in the first registration relationship can be determined according to the position coordinates in the preset pose, and the rotation parameters in the first registration relationship can be determined according to the Euler angles in the preset pose. At the same time, since the second reference point cloud is obtained based on the first reference point cloud, the point-level pairing relationship between the first reference point cloud and the second reference point cloud can be obtained when determining the second reference point cloud.

[0047] S104 , training a pre-established point cloud registration model based on the first training sample group and the second training sample group respectively to obtain a target point cloud registration model.

[0048] The point cloud registration model in the embodiment of the present disclosure may be a deep closest point (DCP) model or other models based on deep learning algorithms, which are not limited here.

[0049] In the embodiment of the present disclosure, after determining the first training sample group and the second training sample group, the point cloud registration model training device can use the first registration relationship in the first training sample group and the transformation parameters in the second training sample group as training labels, and train the pre-established point cloud registration model based on the first training sample group and the second training sample group, respectively, to obtain a trained target point cloud registration model.

[0050] In an exemplary implementation of the presently disclosed embodiment, the point cloud registration model training device may first train the point cloud registration model based on a first training sample group to obtain a preliminarily trained point cloud registration model, and then optimize the preliminarily trained point cloud registration model based on a second training sample to obtain a trained target point cloud registration model.

[0051] In another exemplary implementation of the disclosed embodiments, the point cloud registration model training device may perform alternate training on the point cloud registration model based on the first training sample group and the second training sample group, that is, after performing n1 (n1≥1) rounds of training on the point cloud registration model using the first training sample group, the point cloud registration model is then trained m1 (m1≥1) rounds using the second training sample group, and then, after performing n2 (n2≥1) rounds of training on the point cloud registration model using the first training sample group, the point cloud registration model is then trained m2 (m2≥1) rounds using the second training sample group until a trained target point cloud registration model is obtained.

[0052] The disclosed embodiment obtains a first reference point cloud, a second reference point cloud, and a real point cloud of a sampling object, wherein the first reference point cloud is an overall point cloud of the sampling object, the second reference point cloud is a local point cloud selected from the first reference point cloud based on a preset posture, and the real point cloud is a local point cloud of the sampling object obtained by acquisition, the real point cloud is transformed based on preset transformation parameters to obtain a transformed point cloud, the first reference point cloud, the second reference point cloud, and a first registration relationship between the first reference point cloud and the second reference point cloud are determined as a first training sample group, and the first reference point cloud, the real point cloud, the transformed point cloud, and the transformation parameters are determined as The second training sample group is defined as the second training sample group, wherein the first registration relationship is determined according to a preset posture, and the pre-established point cloud registration model is trained based on the first training sample group and the second training sample group respectively to obtain a target point cloud registration model, which can obtain the absolute transformation relationship and the relative transformation relationship between the point cloud data without manually annotating the point cloud data to construct the training sample data, and then train the point cloud registration model based on the first training sample group containing the absolute transformation relationship and the second training sample group containing the relative transformation relationship, thereby saving labor costs and improving the accuracy of local point cloud registration.

[0053] Figure 2is a flow chart of a method for screening a second reference point cloud provided by an embodiment of the present disclosure, such as Figure 2 As shown, based on the above embodiment, the second reference point cloud can be screened by the following method.

[0054] S201. Determine a point cloud acquisition range of a point cloud acquisition device at a preset posture based on a preset posture and a pre-acquired acquisition viewing angle and imaging distance.

[0055] The acquisition viewing angle in the embodiment of the present disclosure can be understood as the maximum acquisition viewing angle of a point cloud acquisition device, such as a three-dimensional camera, when collecting point cloud data. The imaging distance can be understood as the imaging radial distance of the three-dimensional camera when collecting point cloud data. A truncated cone-shaped area can be determined by the acquisition viewing angle and the imaging distance. This area is the point cloud acquisition range of the point cloud acquisition device.

[0056] In the disclosed embodiment, the point cloud registration model training device can determine the point cloud acquisition range in which the three-dimensional camera in the preset posture can collect point cloud data according to the preset posture of the predetermined three-dimensional camera viewpoint and the pre-acquired acquisition viewing angle and imaging distance.

[0057] In an exemplary implementation of the present disclosure, the preset position and posture of the 3D camera viewpoint may be obtained randomly. Specifically, the point cloud registration model training device may randomly select a point p from the first reference point cloud. i , and in a given imaging distance range [d min ,d max ] and the angle range [θ min ,θ max ] is randomly obtained from the imaging distance d∈[d min ,d max ] and the field of view angle θ∈[θ min ,θ max ], combined with p i The normal vector n of the point i Calculate the position of the 3D camera viewpoint

[0058]

[0059] Center desired viewpoint heading vector In order to get the direction near the center, while the angle with the center vector does not exceed θ ′ The vector needs to increase the deviation vector off. First, randomly generate off∈R 3 , after normalization, we get the randomly perturbed orientation vector

[0060]

[0061] Then, the Euler angle of the three-dimensional camera viewpoint is determined according to the orientation vector after random perturbation.

[0062] S202: Filter point clouds within a point cloud acquisition range from the first reference point cloud as second reference point clouds.

[0063] In the disclosed embodiment, the point cloud registration model training device can, after determining the point cloud acquisition range, filter out a point cloud composed of points within the point cloud acquisition range from the first reference point cloud. Specifically, it can determine whether each point in the first reference point cloud is within the point cloud acquisition range, and determine the point cloud composed of points within the point cloud acquisition range as the second reference point cloud.

[0064] In an exemplary implementation of the disclosed embodiment, the point cloud registration model training device can determine the position coordinates and Euler angles of the three-dimensional camera viewpoint according to a preset posture, and then determine the position coordinates of the first reference point cloud in the three-dimensional viewpoint coordinate system according to the position coordinates and Euler angles of the three-dimensional viewpoint and the position coordinates of the first reference point cloud, and determine the normal vector of the first reference point cloud in the three-dimensional viewpoint coordinate system according to the Euler angles of the three-dimensional viewpoint and the normal vectors of each point in the first reference point cloud. The specific conversion method can be expressed as follows:

[0065] p i =R(e v ) -1 p i ′ -p v

[0066] n i =R(e v ) -1 n ′ i

[0067] Among them, p i represents the position coordinates of the i-th point in the first reference point cloud in the three-dimensional viewpoint coordinate system, R represents the rotation matrix, e v The Euler angle representing the three-dimensional viewpoint, p i ′ represents the position coordinates of the i-th point in the first reference point cloud, p v Represents the position coordinates of the three-dimensional viewpoint, n i Represents the normal vector of the i-th point in the first reference point cloud in the 3D viewpoint coordinate system, n ′ i Represents the normal vector of the i-th point in the first reference point cloud.

[0068] After completing the coordinate system conversion of the first reference point cloud, the screening condition can be determined according to the pre-acquired acquisition viewing angle and imaging distance, and it can be directly judged whether the position coordinates of the converted first reference point cloud meet the screening condition, which can be specifically expressed as follows:

[0069]

[0070]

[0071] n v n′ i <0

[0072] Among them, θ represents the maximum acquisition angle, H min and H max They represent the minimum imaging distance and the maximum imaging distance respectively. If the i-th point in the first reference point cloud satisfies the above screening conditions, it is determined that the point is within the point cloud acquisition range corresponding to the three-dimensional viewpoint.

[0073] The disclosed embodiment determines the point cloud acquisition range of the point cloud acquisition device in a preset posture based on the preset posture and the pre-acquired acquisition viewing angle and imaging distance, and selects the point cloud within the point cloud acquisition range from the first reference point cloud as the second reference point cloud. Without manually labeling the point cloud data, the training label can be determined by simulating the three-dimensional viewpoint posture, and the local point cloud that can be collected can be determined according to the three-dimensional viewpoint posture, so as to realize the training of the point cloud registration model as training data in the future.

[0074] Figure 3 is a flow chart of a method for adjusting model parameters provided by an embodiment of the present disclosure, such as Figure 3 As shown, based on the above embodiment, the model parameters can be adjusted by the following method.

[0075] S301. Input the first reference point cloud and the second reference point cloud into an encoder in a point cloud registration model respectively to obtain a first point cloud feature and a second point cloud feature.

[0076] In the disclosed embodiment, the point cloud registration model includes an encoder and a transformer. The encoder can be understood as a structure for extracting high-dimensional features from point cloud data, and the transformer can be understood as a structure based on a Transformer network for determining the relationship between two sets of point cloud features.

[0077] In the embodiment of the present disclosure, the point cloud registration model training device can input the first reference point cloud and the second reference point cloud into the encoder in the point cloud registration model respectively when training the point cloud registration model based on the first training sample group, and extract the first point cloud features and the second point cloud features.

[0078] S302: Input the first point cloud features and the second point cloud features into a transformer in a point cloud registration model to predict a second registration relationship between the first reference point cloud and the second reference point cloud.

[0079] In the embodiment of the present disclosure, the point cloud registration model training device can input the first point cloud features and the second point cloud features extracted by the encoder into the transformer in the point cloud registration model, obtain the translation vector and the rotation Euler angle output by the transformer, and determine them as the predicted second registration relationship between the first reference point cloud and the second reference point cloud.

[0080] S303: Calculate a first loss function based on the first registration relationship and the second registration relationship, and adjust the model parameters to make the first loss function converge.

[0081] In the disclosed embodiment, the point cloud registration model training device can construct and calculate a first loss function according to the deviation value between the first registration relationship and the second registration relationship after obtaining the second registration relationship output by the point cloud registration model, adjust the model parameters based on the calculated loss value, and determine that the point cloud registration model training is completed when the loss value corresponding to the first loss function converges.

[0082] In an exemplary implementation of the embodiments of the present disclosure, the point cloud registration model training device can update the model parameters through the gradient back propagation method.

[0083] Optionally, the first loss function may be a mean square error loss function.

[0084] The disclosed embodiment obtains first point cloud features and second point cloud features by respectively inputting the first reference point cloud and the second reference point cloud into an encoder in a point cloud registration model, inputs the first point cloud features and the second point cloud features into a transformer in the point cloud registration model, predicts a second registration relationship between the first reference point cloud and the second reference point cloud, calculates a first loss function based on the first registration relationship and the second registration relationship, and adjusts the model parameters to converge the first loss function, so that the point cloud registration model can learn the absolute transformation relationship between the two sets of point cloud data, so that it has preliminary registration capabilities.

[0085] Figure 4 FIG. 1 is a flow chart of another method for adjusting model parameters provided by an embodiment of the present disclosure. Figure 4 As shown, based on the above embodiment, the model parameters can be adjusted by the following method.

[0086] S401. Input the first reference point cloud, the real point cloud and the transformed point cloud into an encoder respectively to obtain first point cloud features, real point cloud features and transformed point cloud features.

[0087] In the embodiment of the present disclosure, the point cloud registration model training device can input the first reference point cloud, the real point cloud and the transformed point cloud into the encoder in the point cloud registration model respectively when training the point cloud registration model based on the second training sample group, and extract the first point cloud features, the real point cloud features and the transformed point cloud features.

[0088] In an exemplary implementation of the presently disclosed embodiment, the point cloud registration model training device may, after acquiring the first point cloud features of the first reference point cloud through S301, only input the real point cloud and the transformed point cloud into the encoder respectively to obtain the real point cloud features and the transformed point cloud features.

[0089] S402: Input the first point cloud features and the real point cloud features into a transformer to predict a third registration relationship between the first reference point cloud and the real point cloud.

[0090] In the embodiment of the present disclosure, the point cloud registration model training device can input the first point cloud features and the real point cloud features extracted by the encoder into the transformer in the point cloud registration model, obtain the translation vector and the rotation Euler angle output by the transformer, and determine them as the third registration relationship between the predicted first reference point cloud and the real point cloud features.

[0091] S403: Input the first point cloud features and the transformed point cloud features into a transformer, and predict a fourth registration relationship between the first reference point cloud and the transformed point cloud.

[0092] In the embodiment of the present disclosure, the point cloud registration model training device can input the first point cloud features and the transformed point cloud features extracted by the encoder into the transformer in the point cloud registration model, obtain the translation vector and the rotation Euler angle output by the transformer, and determine them as the fourth registration relationship between the predicted first reference point cloud and the transformed point cloud features.

[0093] S404: Calculate a second loss function based on the third registration relationship, the fourth registration relationship and the transformation parameters, and adjust the model parameters to make the second loss function converge.

[0094] In the disclosed embodiment, the point cloud registration model training device can determine the relative transformation relationship between the third registration relationship and the fourth registration relationship after obtaining the third registration relationship and the fourth registration relationship predicted by the point cloud registration model, and then construct a second loss function according to the deviation value between the relative transformation relationship and the transformation parameters, adjust the model parameters based on the calculated loss value, and determine that the point cloud registration model training is completed when the loss value corresponding to the second loss function converges.

[0095] In the embodiment of the present disclosure, if the point cloud registration model can output accurate results, the third registration relationship is the transformation relationship between the real point cloud and the first reference point cloud, the fourth registration relationship is the transformation relationship between the transformed point cloud and the first reference point cloud, and the transformation parameters represent the transformation relationship between the real point cloud and the transformed point cloud. Therefore, the relative transformation relationship between the third registration relationship and the fourth registration relationship should correspond to the transformation parameters.

[0096] In an exemplary implementation of the embodiment of the present disclosure, the second loss function may be a consistency loss function, which may be specifically expressed as follows:

[0097] L consistency =|t1-t-Δt|+|e1-e-Δe|

[0098] Among them, L consistency represents the consistency loss, t1 is the translation change in the fourth registration relationship, t is the translation change in the third registration relationship, Δt is the translation change corresponding to the transformation parameter, e1 is the rotation change in the fourth registration relationship, e is the rotation change in the third registration relationship, and Δe is the rotation change corresponding to the transformation parameter.

[0099] In the disclosed embodiment, the first reference point cloud, the real point cloud and the transformed point cloud are respectively input into an encoder to obtain a first point cloud feature, a real point cloud feature and a transformed point cloud feature, the first point cloud feature and the real point cloud feature are input into a transformer, a third registration relationship between the first reference point cloud and the real point cloud is predicted, the first point cloud feature and the transformed point cloud feature are input into a transformer, a fourth registration relationship between the first reference point cloud and the transformed point cloud is predicted, a second loss function is calculated based on the third registration relationship, the fourth registration relationship and the transformation parameters, and the model parameters are adjusted to make the second loss function converge. In this way, the real point cloud can be introduced into the training process without manually annotating the training labels corresponding to the real point cloud. Instead, the transformation parameters between the real point cloud and the transformed point cloud are used as training labels, so that the point cloud registration model learns the relative transformation relationship between the two groups of point cloud data, so that the model can achieve good registration effect when facing the real point cloud that may have noise and uneven distribution problems in actual collection and is significantly different from the point cloud simulated in the first training sample group, thereby further enhancing the processing capability of the enhanced point cloud registration model for the real collected point cloud.

[0100] In some embodiments of the present disclosure, when the point cloud registration model training device calculates a first loss function based on the first registration relationship and the second registration relationship, and adjusts the model parameters to make the first loss function converge, the first learning rate can be used to adjust the model parameters to make the first loss function converge; when the point cloud registration model training device calculates a second loss function based on the third registration relationship, the fourth registration relationship and the transformation parameters, and adjusts the model parameters to make the second loss function converge, the second learning rate can be used to adjust the model parameters to make the second loss function converge, wherein the first learning rate is greater than or equal to the second learning rate.

[0101] Specifically, the point cloud registration model training device can adopt different learning rates for different training sample groups when using the first training sample group and the second training sample group to jointly train the point cloud registration model. For example, in the initial stage of model training, a larger first learning rate can be set for the first training sample group, and a relatively smaller second learning rate can be set for the second training sample group. As the number of training rounds increases, the learning rate of the first training sample group and the learning rate of the second training sample group gradually decrease, and the two decrease at different rates. Finally, the first learning rate is equal to the second learning rate, which can improve the training process of the point cloud registration model for the real point cloud without falling into a local extreme value.

[0102] Figure 5 is a flow chart of a method for registering point clouds provided by an embodiment of the present disclosure, such as Figure 5 As shown, based on the above embodiment, the point cloud can be aligned by the following method.

[0103] S501 , obtaining a third reference point cloud and a point cloud to be registered of the target object, wherein the third reference point cloud is the overall point cloud of the target object, and the point cloud to be registered is the acquired local point cloud of the target object.

[0104] In the disclosed embodiment, the point cloud registration model training device can obtain a third reference point cloud of the entire target object corresponding to the point cloud to be registered, and a local point cloud to be registered obtained by collecting the target object through a point cloud acquisition device after the target point cloud registration model is trained.

[0105] S502: Input the third reference point cloud and the point cloud to be registered into the target point cloud registration model to obtain a first registration result output by the model.

[0106] In the disclosed embodiment, the point cloud registration model training device can input the third reference point cloud and the point cloud to be registered into the trained target point cloud registration model after obtaining the third reference point cloud and the point cloud to be registered, and the encoder in the target point cloud registration model can respectively extract the point cloud features of the third reference point cloud and the point cloud to be registered, and then the transformer in the target point cloud registration model can obtain the registration result between the third reference point cloud and the point cloud to be registered according to their point cloud features, that is, the first registration result output by the model.

[0107] The disclosed embodiment obtains a third reference point cloud and a point cloud to be registered of the target object, wherein the third reference point cloud is the overall point cloud of the target object, and the point cloud to be registered is the collected local point cloud of the target object, and inputs the third reference point cloud and the point cloud to be registered into a target point cloud registration model to obtain a first registration result output by the model. The first registration result of the third reference point cloud and the point cloud to be registered can be obtained through the trained target point cloud registration model, thereby improving the accuracy of point cloud registration.

[0108] Figure 6 is a flow chart of a method for calculating a registration result provided by an embodiment of the present disclosure, such as Figure 6 As shown, based on the above embodiment, the registration result can be calculated by the following method.

[0109] S601 , converting the point cloud to be registered based on the first registration result to obtain a target point cloud.

[0110] In the embodiment of the present disclosure, the point cloud registration model training device can, after obtaining the first registration result between the third reference point cloud and the point cloud to be registered, transform the point cloud to be registered according to the translation change and rotation change in the first registration result, and determine the transformed point cloud as the target point cloud.

[0111] Specifically, the i-th point in the roughly registered target point cloud The coordinates of can be expressed as follows:

[0112]

[0113] in, represents the coordinates of the i-th point in the third reference point cloud, t and e represent the translation change and rotation change in the first registration result, respectively, and R is the rotation matrix.

[0114] S602: Perform fine registration on the target point cloud and the third reference point cloud to obtain a second registration result.

[0115] In the disclosed embodiment, the point cloud registration model training device can convert the point cloud to be registered based on the first registration result. After obtaining the target point cloud, the target point cloud basically coincides with the third reference point cloud. At this time, the target point cloud and the third reference point cloud are finely registered again to determine which points in the third reference point cloud the points in the target point cloud correspond to, and then determine the translation and rotation changes between the corresponding points to obtain the second registration result.

[0116] In an exemplary implementation of the presently disclosed embodiment, the point cloud registration model training device can calculate the average translation and rotation changes of each point in the target point cloud and the corresponding point when the translation and rotation changes of each point in the target point cloud and the corresponding point in the third reference point cloud are different, and determine the average value as the second registration result.

[0117] S603: Calculate and obtain a target registration result based on the first registration result and the second registration result.

[0118] In the embodiment of the present disclosure, the point cloud registration model training device can calculate the target registration result based on the first registration result and the second registration result output by the target point cloud registration model after obtaining the second registration result after fine registration. Specifically, the translation and rotation changes in the first registration result can be summed with the translation and rotation changes in the second point cloud registration result to obtain the target registration result.

[0119] The disclosed embodiment obtains a target point cloud by converting the point cloud to be registered based on the first registration result, and obtains a second registration result by finely registering the target point cloud with a third reference point cloud. The target registration result is calculated based on the first registration result and the second registration result, and a secondary fine registration can be performed based on the registration result of the target point cloud registration model, thereby further improving the accuracy of the point cloud registration.

[0120] Figure 7 is a flow chart of a method for fine point cloud registration provided by an embodiment of the present disclosure. Figure 7 As shown, based on the above embodiment, the point cloud can be finely aligned by the following method.

[0121] S701. Match points in the target point cloud with points in the third reference point cloud based on the Hungarian algorithm.

[0122] In the disclosed embodiment, the point cloud registration model training device may use the Hungarian algorithm to determine the matching relationship between the points in the target point cloud and the points in the third reference point cloud before finely registering the points in the target point cloud with the points in the third reference point cloud.

[0123] S702: Based on the matching result, an iterative closest point algorithm is used to perform fine registration of the target point cloud and the third reference point cloud.

[0124] In the disclosed embodiment, the point cloud registration model training device can, after determining the matching relationship between the points in the target point cloud and the points in the third reference point cloud, determine the point pairs with the matching relationship as corresponding points in the iterative closest point algorithm, and then based on the iterative closest point algorithm, use the least squares method to iteratively calculate the translation transformation amount and the rotation transformation amount so that the sum of squared errors reaches a minimum value, and solve the optimal translation transformation amount and rotation transformation amount, and determine them as the second registration result.

[0125] The disclosed embodiment matches the points in the target point cloud with the points in the third reference point cloud based on the Hungarian algorithm, and based on the matching results, uses an iterative closest point algorithm to perform fine alignment on the target point cloud and the third reference point cloud, thereby obtaining a more accurate matching result, and then obtaining a more accurate alignment result, further improving the accuracy of point cloud alignment.

[0126] Figure 8 Schematic diagram of a point cloud registration model training device provided by an embodiment of the present disclosure. Figure 8 As shown, the point cloud registration model training device 800 includes: a first acquisition module 810, a transformation module 820, a determination module 830, and a training module 840, wherein the first acquisition module 810 is used to acquire a first reference point cloud, a second reference point cloud, and a real point cloud of a sampling object, wherein the first reference point cloud is the overall point cloud of the sampling object, the second reference point cloud is a local point cloud selected from the first reference point cloud based on a preset posture, and the real point cloud is the local point cloud of the sampling object obtained by collection; the transformation module 820 is used to transform the real point cloud based on preset transformation parameters. Perform transformation processing to obtain a transformed point cloud; determine module 830, which is used to determine the first reference point cloud, the second reference point cloud and the first registration relationship between the first reference point cloud and the second reference point cloud as a first training sample group, and determine the first reference point cloud, the real point cloud, the transformed point cloud and the transformation parameters as a second training sample group, wherein the first registration relationship is determined according to the preset posture; and train module 840, which is used to train a pre-established point cloud registration model based on the first training sample group and the second training sample group, respectively, to obtain a target point cloud registration model.

[0127] Optionally, the first acquisition module 810 includes: a range determination unit, used to determine the point cloud acquisition range of the point cloud acquisition device in the preset posture based on the preset posture and the pre-acquired acquisition viewing angle and imaging distance; a screening unit, used to screen the point cloud within the point cloud acquisition range from the first reference point cloud as the second reference point cloud.

[0128] Optionally, the training module 840 includes: a first encoding unit, used to input the first reference point cloud and the second reference point cloud into the encoder in the point cloud registration model respectively to obtain first point cloud features and second point cloud features; a first transformation unit, used to input the first point cloud features and the second point cloud features into the transformer in the point cloud registration model to predict the second registration relationship between the first reference point cloud and the second reference point cloud; a first adjustment unit, used to calculate a first loss function based on the first registration relationship and the second registration relationship, and adjust the model parameters to make the first loss function converge.

[0129] Optionally, the first loss function is a mean square error loss function.

[0130] Optionally, the transformation processing includes translation transformation processing and / or rotation transformation processing.

[0131] Optionally, the training module 840 includes: a second encoding unit, used to input the first reference point cloud, the real point cloud and the transformed point cloud into the encoder respectively to obtain first point cloud features, real point cloud features and transformed point cloud features; a second transformation unit, used to input the first point cloud features and the real point cloud features into the transformer, and predict a third registration relationship between the first reference point cloud and the real point cloud; a third transformation unit, used to input the first point cloud features and the transformed point cloud features into the transformer, and predict a fourth registration relationship between the first reference point cloud and the transformed point cloud; a second adjustment unit, used to calculate a second loss function based on the third registration relationship, the fourth registration relationship and the transformation parameters, and adjust the model parameters to make the second loss function converge.

[0132] Optionally, the first adjustment unit is specifically used to adjust the model parameters using a first learning rate so that the first loss function converges; the second adjustment unit is specifically used to adjust the model parameters using a second learning rate so that the second loss function converges, wherein the first learning rate is greater than or equal to the second learning rate.

[0133] Optionally, the point cloud registration model training device 800 also includes: a second acquisition module, used to acquire a third reference point cloud and a point cloud to be registered of the target object, wherein the third reference point cloud is the overall point cloud of the target object, and the point cloud to be registered is the collected local point cloud of the target object; a first registration module, used to input the third reference point cloud and the point cloud to be registered into the target point cloud registration model to obtain a first registration result output by the model.

[0134] Optionally, the point cloud registration model training device 800 also includes: a conversion module, used to convert the point cloud to be registered based on the first registration result to obtain a target point cloud; a second registration module, used to finely align the target point cloud with the third reference point cloud to obtain a second registration result; and a calculation module, used to calculate the target registration result based on the first registration result and the second registration result.

[0135] Optionally, the point cloud registration model training device 800 also includes: a matching module, used to match the points in the target point cloud with the points in the third reference point cloud based on the Hungarian algorithm; the second registration module is specifically used to perform fine registration of the target point cloud with the third reference point cloud based on the matching results, using an iterative closest point algorithm.

[0136] The point cloud registration model training device provided in this embodiment can execute the method described in any of the above embodiments. Its execution method and beneficial effects are similar and will not be repeated here.

[0137] Fig. 9 It is a structural diagram of a computer device provided in an embodiment of the present disclosure.

[0138] like Fig. 9 As shown, the computer device may include a processor 910 and a memory 920 storing computer program instructions.

[0139] Specifically, the processor 910 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0140] The memory 920 may include a large capacity memory for information or instructions. By way of example and not limitation, the memory 920 may include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a tape, or a universal serial bus (USB) drive or a combination of two or more of these. Where appropriate, the memory 920 may include a removable or non-removable (or fixed) medium. Where appropriate, the memory 920 may be inside or outside the integrated gateway device. In a particular embodiment, the memory 920 is a non-volatile solid-state memory. In a particular embodiment, the memory 920 includes a read-only memory (ROM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (Electrically Erasable Programmable ROM, EEPROM), an electrically rewritable ROM (EAROM) or a flash memory, or a combination of two or more of these.

[0141] The processor 910 reads and executes the computer program instructions stored in the memory 920 to perform the steps of the point cloud registration model training method provided in the embodiment of the present disclosure.

[0142] In one example, the computer device may further include a transceiver 930 and a bus 940. Fig. 9 As shown, the processor 910, the memory 920 and the transceiver 930 are connected via a bus 940 and communicate with each other.

[0143] The bus 940 includes hardware, software or both. For example, but not limitation, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industrial Standard Architecture (ISA) bus, an InfiniBand interconnect, a Low Pin Count (LPC) bus, a Memory bus, a MicroChannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses or a combination of two or more of these. Where appropriate, the bus 940 may include one or more buses. Although embodiments of the present application describe and illustrate a particular bus, the present application contemplates any suitable bus or interconnect.

[0144] The embodiments of the present disclosure also provide a computer-readable storage medium, which may store a computer program. When the computer program is executed by a processor, the processor implements the point cloud registration model training method provided by the embodiments of the present disclosure.

[0145] The above-mentioned storage medium may, for example, include a memory 920 of computer program instructions, and the above-mentioned instructions may be executed by the processor 910 of the point cloud registration model training device to complete the point cloud registration model training method provided by the embodiment of the present disclosure. Optionally, the storage medium may be a non-temporary computer-readable storage medium, for example, the non-temporary computer-readable storage medium may be a ROM, a random access memory (Random Access Memory, RAM), a compact disc read-only memory (Compact Disc ROM, CD-ROM), a magnetic tape, a floppy disk, and an optical data storage device, etc. The above-mentioned computer program may be written in any combination of one or more programming languages ​​to perform the program code of the operation of the embodiment of the present disclosure, and the programming language includes an object-oriented programming language, such as Java, C++, etc., and also includes a conventional procedural programming language, such as "C" language or a similar programming language. The program code may be executed entirely on the user computing device, partially on the user device, as an independent software package, partially on the user computing device and partially on the remote computing device, or entirely on the remote computing device or server.

[0146] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants 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, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0147] The above description is only a specific embodiment of the present disclosure, so that those skilled in the art can understand or implement the present disclosure. 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 disclosure. Therefore, the present disclosure will not be limited to the embodiments described herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A point cloud registration model training method, characterized in that: The method comprises: Acquire a first reference point cloud, a second reference point cloud, and a real point cloud of the sampling object, wherein the first reference point cloud is the overall point cloud of the sampling object, the second reference point cloud is a local point cloud selected from the first reference point cloud based on a preset posture, and the real point cloud is the acquired local point cloud of the sampling object; Transforming the real point cloud based on preset transformation parameters to obtain a transformed point cloud; Determine the first reference point cloud, the second reference point cloud, and a first registration relationship between the first reference point cloud and the second reference point cloud as a first training sample group, and determine the first reference point cloud, the real point cloud, the transformed point cloud, and the transformation parameters as a second training sample group, wherein the first registration relationship is determined according to the preset posture; Training a pre-established point cloud registration model based on the first training sample group and the second training sample group respectively to obtain a target point cloud registration model; The step of training a pre-established point cloud registration model based on the first training sample group includes: Inputting the first reference point cloud and the second reference point cloud into an encoder in the point cloud registration model respectively to obtain a first point cloud feature and a second point cloud feature; Inputting the first point cloud feature and the second point cloud feature into a transformer in the point cloud registration model to predict a second registration relationship between the first reference point cloud and the second reference point cloud; Calculating a first loss function based on the first registration relationship and the second registration relationship, and adjusting model parameters to make the first loss function converge; Training a pre-established point cloud registration model based on the second training sample group includes: Inputting the first reference point cloud, the real point cloud and the transformed point cloud into the encoder respectively to obtain first point cloud features, real point cloud features and transformed point cloud features; Inputting the first point cloud features and the real point cloud features into the transformer, and predicting a third registration relationship between the first reference point cloud and the real point cloud; Inputting the first point cloud features and the transformed point cloud features into the transformer, and predicting a fourth registration relationship between the first reference point cloud and the transformed point cloud; A second loss function is calculated based on the third registration relationship, the fourth registration relationship and the transformation parameters, and the model parameters are adjusted to make the second loss function converge.

2. The method according to claim 1, characterized in that Obtaining a second reference point cloud of the sampled object includes: Determining a point cloud acquisition range of a point cloud acquisition device in the preset posture based on the preset posture and the pre-acquired acquisition viewing angle and imaging distance; Point clouds within the point cloud acquisition range are selected from the first reference point clouds as second reference point clouds.

3. The method according to claim 1, characterized in that: The first loss function is a mean square error loss function.

4. The method according to claim 1, characterized in that: The transformation process includes a translation transformation process and / or a rotation transformation process.

5. The method according to claim 1, characterized in that: The adjusting the model parameters so as to make the first loss function converge includes: Using a first learning rate to adjust model parameters so that the first loss function converges; The adjusting the model parameters so that the second loss function converges includes: The model parameters are adjusted using a second learning rate to converge the second loss function, wherein the first learning rate is greater than or equal to the second learning rate.

6. The method according to any one of claims 1 to 5, characterized in that After the pre-established point cloud registration model is trained based on the first training sample group and the second training sample group respectively to obtain a target point cloud registration model, the method further includes: Acquire a third reference point cloud and a point cloud to be registered of the target object, wherein the third reference point cloud is the overall point cloud of the target object, and the point cloud to be registered is the acquired local point cloud of the target object; The third reference point cloud and the point cloud to be registered are input into the target point cloud registration model to obtain a first registration result output by the model.

7. The method according to claim 6, characterized in that After inputting the third reference point cloud and the point cloud to be registered into the target point cloud registration model to obtain a first registration result output by the model, the method further includes: Based on the first registration result, the point cloud to be registered is converted to obtain a target point cloud; Performing fine registration on the target point cloud and the third reference point cloud to obtain a second registration result; A target registration result is calculated based on the first registration result and the second registration result.

8. The method according to claim 7, characterized in that Before finely registering the target point cloud with the third reference point cloud, the method further includes: Matching points in the target point cloud with points in the third reference point cloud based on the Hungarian algorithm; The finely registering the target point cloud with the third reference point cloud comprises: Based on the matching result, an iterative closest point algorithm is used to finely align the target point cloud with the third reference point cloud.

9. A point cloud registration model training device, characterized in that: The device comprises: A first acquisition module is used to acquire a first reference point cloud, a second reference point cloud, and a real point cloud of the sampling object, wherein the first reference point cloud is the overall point cloud of the sampling object, the second reference point cloud is a local point cloud selected from the first reference point cloud based on a preset posture, and the real point cloud is the local point cloud of the sampling object obtained by acquisition; A transformation module, used for transforming the real point cloud based on preset transformation parameters to obtain a transformed point cloud; a determination module, configured to determine the first reference point cloud, the second reference point cloud, and a first registration relationship between the first reference point cloud and the second reference point cloud as a first training sample group, and determine the first reference point cloud, the real point cloud, the transformed point cloud, and the transformation parameters as a second training sample group, wherein the first registration relationship is determined according to the preset posture; A training module, used to train a pre-established point cloud registration model based on the first training sample group and the second training sample group to obtain a target point cloud registration model; The training module comprises: A first encoding unit, configured to input the first reference point cloud and the second reference point cloud into an encoder in the point cloud registration model respectively, to obtain a first point cloud feature and a second point cloud feature; A first transformation unit, configured to input the first point cloud feature and the second point cloud feature into a transformer in the point cloud registration model, and predict a second registration relationship between the first reference point cloud and the second reference point cloud; A first adjustment unit, configured to calculate a first loss function based on the first registration relationship and the second registration relationship, and adjust model parameters so that the first loss function converges; A second encoding unit, configured to input the first reference point cloud, the real point cloud and the transformed point cloud into the encoder respectively, to obtain a first point cloud feature, a real point cloud feature and a transformed point cloud feature; A second transformation unit, configured to input the first point cloud features and the real point cloud features into the transformer, and predict a third registration relationship between the first reference point cloud and the real point cloud; A third transformation unit, configured to input the first point cloud feature and the transformed point cloud feature into the transformer, and predict a fourth registration relationship between the first reference point cloud and the transformed point cloud; The second adjustment unit is used to calculate a second loss function based on the third registration relationship, the fourth registration relationship and the transformation parameters, and adjust the model parameters to make the second loss function converge.

10. The device according to claim 9, characterized in that The first acquisition module includes: A range determination unit, configured to determine a point cloud acquisition range of a point cloud acquisition device at the preset position and posture based on the preset position and posture and the pre-acquired acquisition viewing angle and imaging distance; A screening unit is used to screen point clouds within the point cloud acquisition range from the first reference point cloud as second reference point clouds.

11. The device according to claim 9, characterized in that The first loss function is a mean square error loss function.

12. The device according to claim 9, characterized in that The transformation process includes a translation transformation process and / or a rotation transformation process.

13. The device according to claim 9, characterized in that The first adjustment unit is specifically configured to adjust the model parameters using a first learning rate so as to make the first loss function converge; The second adjustment unit is specifically used to adjust the model parameters using a second learning rate so that the second loss function converges, wherein the first learning rate is greater than or equal to the second learning rate.

14. The device according to any one of claims 9 to 13, characterized in that The device also includes: A second acquisition module is used to acquire a third reference point cloud and a point cloud to be registered of the target object, wherein the third reference point cloud is the overall point cloud of the target object, and the point cloud to be registered is the acquired local point cloud of the target object; The first registration module is used to input the third reference point cloud and the point cloud to be registered into the target point cloud registration model to obtain a first registration result output by the model.

15. The device according to claim 14, characterized in that The device also includes: A conversion module, used for converting the to-be-registered point cloud based on the first registration result to obtain a target point cloud; A second registration module, used for finely registering the target point cloud with the third reference point cloud to obtain a second registration result; A calculation module is used to calculate a target registration result based on the first registration result and the second registration result.

16. The device according to claim 15, characterized in that The device also includes: A matching module, used for matching points in the target point cloud with points in the third reference point cloud based on the Hungarian algorithm; The second registration module is specifically used to perform fine registration of the target point cloud and the third reference point cloud using an iterative closest point algorithm based on the matching result.

17. A computer device, characterized in that: include: Memory; processor; and a computer program; wherein the computer program is stored in the memory and is configured to be executed by the processor to implement the method according to any one of claims 1 to 8.

18. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images