Target capturing method based on 3D confrontation point cloud, electronic equipment and medium

Abstract

The invention discloses a 3D confrontation point cloud-based target capturing method, electronic equipment and medium, and the method comprises the steps: obtaining an original first point cloud data set X0, inputting the first point cloud data set X0 into a diffusion model, and outputting to obtain a second point cloud data set XT; adding disturbance to each point cloud data in the second point cloud data set XT to generate disturbance point cloud, and obtaining a confrontation point cloud point cloud data set X 'T; constructing a 3D point cloud detection model, taking the confrontation point cloud data set X 'T as a training set of the 3D point cloud detection model, and training the 3D point cloud detection model through zero-order random coordinate descent and a multi-scale attack method; inputting to-be-processed point cloud data into the trained 3D point cloud detection model for target detection, and performing pose estimation on an identified target to obtain a target object coordinate; and the coordinates of the target object are converted into a robot coordinate system, the motion path of a mechanical arm is planned, and the mechanical arm is controlled to complete target grabbing.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning technology, and in particular relates to a target capture method, electronic equipment, and medium based on 3D adversarial point cloud. Background Art

[0002] Object detection is a task in the field of computer vision. Different from image recognition, object detection not only needs to identify the objects in the image and give the corresponding categories, but also needs to give the location of the object in the form of a minimum bounding box. With the emergence of neural network-based recognition algorithms such as Faster-CNN, SSD, and Yolo, the accuracy of 2D object detection has exceeded that of the human eye. For three-dimensional objects in complex real environments, using only two-dimensional information to characterize three-dimensional objects will inevitably cause information loss. 2D object detection can only provide the location and category of the target object in the two-dimensional image. At the same time, changes in light and dark may cause the features of the target area to become blurred or distorted, making it difficult for the model to accurately identify and locate the target.

[0003] 3D point cloud data retains the original, rich geometric structure information and fine-grained scale details of objects and scenes in three-dimensional space. Although 3D point cloud data solves the problem that 2D target detection is easily affected by light and dark and has low environmental adaptability, and the rich geometric structure information makes up for the information loss during the transition from three-dimensional to two-dimensional in the 2D target detection process. However, in many real-life scenarios, the irregularity of the detected objects may cause the point cloud data to contain more noise and outliers. The occlusion caused by the irregular shape of the detected object, the dynamic changes introduced by the movement of the device, and the difference in reflection intensity caused by uneven lighting conditions will all cause noise. Common noise is divided into two categories: one is the outliers caused by complex environments and sensor errors; the other is the outlier noise points generated by mobile devices. These noise points seriously affect the quality and accuracy of point cloud data, and then affect the performance of subsequent tasks such as three-dimensional modeling, object recognition, and visual positioning, resulting in poor robustness of engineering tasks based on 3D point clouds.

[0004] On the other hand, there is a potential conflict between point cloud sampling density and action execution efficiency. Point cloud sampling density includes the number and distribution of data points in the point cloud. A higher sampling density means that the point cloud contains more data points, which can more accurately describe the shape, structure, and surface details of the object. This is very important for applications that require high-precision results, such as 3D modeling, object recognition, and measurement. However, as the sampling density increases, the amount of computation required for data processing and analysis will also increase accordingly. The speed of processing and analyzing point cloud data will be affected, and a higher sampling density may cause the calculation process to become more complex and time-consuming, thereby reducing the efficiency of action execution. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a target grasping method, electronic device and medium based on 3D adversarial point cloud.

[0006] In a first aspect, an embodiment of the present invention provides a method for capturing a target based on a 3D adversarial point cloud, the method comprising the following steps:

[0007] Get the original first point cloud dataset X0, and input the first point cloud dataset X0 into the diffusion model, and output the second point cloud dataset X T ;

[0008] For the second point cloud dataset X T Add disturbance to each point cloud data in to generate a perturbed point cloud, and get the adversarial point cloud dataset X′ T ;

[0009] Build a 3D point cloud detection model and use the adversarial point cloud dataset X′ T As the training set of the 3D point cloud detection model, the 3D point cloud detection model is trained by zero-order random coordinate descent and multi-scale attack method;

[0010] The point cloud data to be processed is input into the trained 3D point cloud detection model for target detection, and the pose of the identified target is estimated to obtain the target object coordinates; the target object coordinates are transformed into the robot coordinate system, the robot arm motion path is planned, and the robot arm is controlled to complete the target grasping.

[0011] In a second aspect, an embodiment of the present invention provides an electronic device, comprising a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-mentioned target grasping method based on 3D adversarial point cloud.

[0012] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned target grasping method based on 3D adversarial point cloud.

[0013] In a fourth aspect, an embodiment of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the above-mentioned target grasping method based on 3D adversarial point cloud.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] The present invention provides a target grasping method based on 3D adversarial point cloud, which processes the original point cloud data set through a diffusion model to obtain a high-quality second point cloud data set; adds disturbance to each point cloud data in the second point cloud data set to generate a perturbed point cloud, and obtains an adversarial point cloud data set; uses the adversarial point cloud data set as a training set for a 3D point cloud detection model, and trains the 3D point cloud detection model through zero-order random coordinate descent and multi-scale attack method, thereby improving the robustness of the 3D point cloud detection model, thereby achieving a target grasping task with strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0017] Figure 1 A schematic diagram of a process of a target capture method based on 3D adversarial point cloud provided by an embodiment of the present invention;

[0018] Figure 2 A schematic diagram of a network structure of a diffusion model provided by an embodiment of the present invention;

[0019] Figure 3 A schematic diagram of a U-Net-like network structure in a diffusion model provided in an embodiment of the present invention;

[0020] Figure 4 A schematic diagram of a 3D point cloud generation process based on a diffusion model provided in an embodiment of the present invention;

[0021] Figure 5 A schematic diagram of the correspondence between the two-dimensional image texture coordinates and the three-dimensional spherical coordinates provided by an embodiment of the present invention;

[0022] Figure 6 A schematic diagram of a random fast search tree algorithm provided by an embodiment of the present invention;

[0023] Figure 7 A schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] It should be noted that, in the absence of conflict, the features in the following embodiments and implementations may be combined with each other.

[0026] like Figure 1 As shown, an embodiment of the present invention provides a target grasping method based on 3D adversarial point cloud, and the method comprises the following steps:

[0027] Step S1, obtaining an original first point cloud dataset X0, and inputting the first point cloud dataset X0 into a diffusion model, and outputting a second point cloud dataset X T .

[0028] Furthermore, if Figure 2 , Figure 3 and Figure 4 As shown, the present invention designs a U-Net network with a three-layer structure in a diffusion network, and each layer is composed of a convolution layer, a normalization layer, and an embedding layer. The three convolution layers are responsible for extracting features, capturing abstract information of the point cloud content, and eliminating some noise at the same time. The three deconvolution layers are used to restore the original details; the normalization layer is responsible for the distribution of the normalization layer input, so that the input data of each layer has a more stable distribution and improves the stability of the training process; the embedding layer is responsible for integrating the entire convolutional neural network. Considering that when processing realistic sparse point cloud scenes, the network may only be able to capture local features. When the network performs higher-level feature abstraction and combination, the gradient signal may be insufficient, resulting in the disappearance of the gradient; in addition, sparsity may also lead to numerical stability problems during normalization, thereby affecting the calculation of the gradient.

[0029] Furthermore, the diffusion model includes a forward process and a reverse process: the forward process applies noise to the point cloud step by step until the point cloud becomes a mass of random noise, and the reverse process reverses the forward process to predict the noise at each step.

[0030] The forward process of the diffusion model is as follows: for the initial data X0, noise interference is gradually applied to it until it becomes chaotic noise data X0 at the Tth step. T ; The formula for the process from step t-1 to step t is as follows:

[0031]

[0032] Among them, q(X t |X t-1 ) That is, when X is known t-1 Condition X t The probability distribution of t} represents the process of applying noise forward; I is a vector of all 1s, β t I represents the distribution of real noise; the noise addition strategy makes all conditional probabilities satisfy the same form of distribution. After the derivation of the formula, Xt It can be calculated directly from X0, the formula is as follows:

[0033]

[0034] in, ∈~N(0,I).

[0035] Furthermore, the reverse process of the diffusion model is as follows: starting from the noise distribution, using the neural network model to gradually remove the noise, and finally obtaining the expected shape. We use q(X t |X0) as a medium, through the conditional probability conversion formula, the posterior probability q(X t-1 |X t ,X0), the conversion formula is as follows:

[0036]

[0037] In the formula, q(X t |X t-1 ),q(X t-1 |X0) and q(X t |X0) are all Gaussian distributions, so q(X t-1 |X t ,X0) also satisfies the Gaussian distribution, and the mean u(X t ,X0) and variance ∑ satisfy the following relationship:

[0038]

[0039] By q(X t |X t-1 ) diffusion transforms the evaluation of the denoising process q(X t-1 |X t ,X0), X0 is unknown, so we choose to use model p θ (X t-1 |X t ,z) to fit. Using model p θ (X t-1 |X t ,z) The fitting reverse process is expressed as follows:

[0040] p θ (X t-1 |X t ,z)=N(X t-1 |μ θ (X t ,t,z),∑ θ (x t ,t))

[0041] where p θ(X te1 |X t ,z) indicates that X is known in the reverse process t , z condition X t-1 The probability distribution of μ θ is the predicted shape of the diffusion model, i.e., the mean term, ∑ θ (x t ,t) is the variance term.

[0042] The variance ∑ of the diffusion model is a linear schedule, which will make the 3D point cloud add noise too quickly. In fact, due to the sparsity of point clouds in 3D scenes, we do not want to add noise too quickly and increase a large number of noise points, but to ensure as many details as possible.

[0043] Therefore, the present invention improves the diffusion model, and for the model p θ (X t-1 |X t ,z) not only trains and learns μ θ (X t ,t,z), and also learned∑ θ (x t ,t) to improve the log-likelihood value, the formula is as follows:

[0044] min‖u t (X t ,X0)-u θ (X t ,t)‖ 2

[0045]

[0046] Where v is the model output, ∈ θ is the diffusion model's estimate of the noise.

[0047] After formula derivation and using parameterization techniques, the point cloud at step t-1 is expressed as:

[0048]

[0049] Calculate the difference between the noise predicted by the diffusion model and the actual noise and the variational lower bound vlb of the diffusion model. The calculation formula is as follows:

[0050] L simple (θ)=‖∈ t -∈ θ (X t ,t,z)‖ 2

[0051] L vlb (θ)=L0+L1+...+L t-1+L t

[0052] L0=-logp θ (X0|X1)

[0053] L t-1 =D KL (q(X t-1 |X t ,X0)||p θ (X t-1 |X t ))

[0054] L t =D KL (q(X t |X0)||p(X t ))

[0055] Among them, ∈ t is the noise corresponding to the time step in the diffusion process, ∈ θ (X t ,t,z) is the prediction noise of the diffusion model, D KL It is the operator for calculating KL divergence;

[0056] From this, we can write the desired objective function, the formula is as follows:

[0057] L hybrid =L simple +λL vlb

[0058] Where, λ=0.001, L simple Part of it is for u θ (X t ,t) to learn, L vlb Partially terminated θ (X t ,t) to avoid repeated calculations and guide ∑ θ (X t ,t) calculation.

[0059] Step S2: for the second point cloud data set X T Add disturbance to each point cloud data in to generate a perturbed point cloud, and get the adversarial point cloud dataset X′ T .

[0060] Specifically, the step S2 includes:

[0061] The generation of adversarial examples is optimized by maximizing the log-likelihood of the target class, expressed as:

[0062]

[0063] subject to‖x ′ -x‖ p <∈

[0064] In the formula, y t Represents the target class on the original sample, x is the real point cloud, x∈X T , x ′ is the generated disturbance point cloud;

[0065] Each disturbance is modeled as:

[0066] δ=E t～T [d(t(x ′ ),t(x))]

[0067] Where t(x) represents the transformation function, T represents the transformation distribution of the transformation function t, and d represents the distance between the real point cloud and the perturbed point cloud after being processed by the transformation function t(x);

[0068] Furthermore, in this example, the distribution T of the transformation function t(x) is selected for modeling; the transformation function t(x) maps the texture to the rendering of the object, simulating the object's texture mapping, rendering, lighting, rotation, lateral translation and perspective projection functions.

[0069] The specific steps of texture mapping transformation are as follows: a whole sphere is used as the intermediate transition surface between the two texture mappings, and the surface of the entire sphere is divided accordingly. Figure 5 , the horizontal line u of the texture coordinate on the left is mapped to the longitude θ corresponding to the entire sphere, and the vertical line v in the texture coordinate is mapped to the latitude corresponding to the entire sphere The texture coordinates (u, v) on the two-dimensional image and the points on the three-dimensional sphere are established The relationship between them can be expressed as follows using the mapping function:

[0070]

[0071] According to the mapping function, the texture coordinates (u, v) corresponding to the three-dimensional coordinates on the whole sphere are obtained, thereby obtaining a one-to-one correspondence between the three-dimensional coordinates (x, y, z) on the spherical surface and the texture coordinates (u, v) of the image. The formula is expressed as follows:

[0072]

[0073] Establish a hemisphere that can surround the entire point cloud model, establish a set of straight lines between the three-dimensional points of the point cloud and the projection center point, and then let the straight lines intersect the hemisphere to find out the relationship between the three-dimensional coordinate points of the point cloud and the spherical coordinates, and then map the two-dimensional plane texture image to the three-dimensional point cloud model.

[0074] Furthermore, the specific steps of translation and rotation transformation are as follows: obtain the position information (x, y, z) of the point cloud, and perform translation matrix transformation on the coordinates of the point cloud to achieve the translation effect. The formula is as follows:

[0075]

[0076] The rotation effect can be achieved by performing a rotation matrix transformation on the coordinates of the point cloud. The rotation matrix is ​​expressed as follows:

[0077]

[0078] The rotation here means rotating θ around the z-axis of the point cloud coordinate system.

[0079] The adversarial sample generation is expressed as:

[0080]

[0081] subject to E t～T [d(t(x ′ ),t(x))]<∈

[0082] The constrained Lagrangian relaxation form is used to solve the constrained optimization problem, combined with a quantitative evaluation of the probability of each adversarial point cloud being classified as the target class, to obtain the adversarial point cloud dataset X′ T , expressed as:

[0083]

[0084] Wherein, LAB represents LAB color space.

[0085] Set a quantitative evaluation standard function C(x,y) to indicate whether the input x is classified as class y, which can be expressed as:

[0086]

[0087] Then, by formula E t～T [C(t(x ′ ),y)], to evaluate the probability that the adversarial sample is classified as the target class.

[0088] Step S3: construct a 3D point cloud detection model and use the adversarial point cloud dataset X′ T As the training set of the 3D point cloud detection model, the 3D point cloud detection model is trained by zero-order random coordinate descent and multi-scale attack method.

[0089] Furthermore, in this example, the 3D point cloud detection model uses the bottom-up PointRCNN 3D point cloud detection model, and uses the complete box-based loss function BB-loss.

[0090] The implementation of the present invention selects a bottom-up PointRCNN 3D point cloud detection model. For the objective function of the model, the loss function based on the complete box BB-loss is selected here. A variable in a coordinate dimension is randomly selected, and the variable is continuously iterated and updated in the direction of approximately minimizing the objective function along the coordinate, which is expressed as:

[0091]

[0092] x i ←x i +δ *

[0093] Where f(x) is the objective function of the detection model; i is a standard basis vector used to describe the randomly selected coordinate dimension; δ * Denotes the variable update value δ obtained by minimizing the objective function, and δ * To update the corresponding variable x i ;

[0094] Calculate an estimate of the gradient and the estimate of the second-order gradient The formula is as follows:

[0095]

[0096] Here, h is a very small constant, and in this example, h=0.0001.

[0097] After estimating the gradient and the second-order gradient, the Newton iteration method is used to solve the minimization problem and obtain δ * ; The specific steps are as follows:

[0098] Choose an update step size η and update the randomly selected variable e i , calculated to obtain and like Then the negative gradient direction is used for updating, and the formula is expressed as:

[0099]

[0100] x i ←x i +δ *

[0101] Otherwise the update formula is expressed as:

[0102]

[0103] x i ←x i +δ*

[0104] Repeat the above steps to iterate and update the variable x i , until the variable x that minimizes the objective function is obtained i .

[0105] Furthermore, this example designs a series of dimensions m1, m2, ..., m n Transformations D1,D2,...,D n , first select a relatively small scale m1, whose corresponding transformation function is D1, and use these two parameters to update the objective function. As the number of iterations progresses, when the loss function no longer decreases, gradually increase m, increase the search space, use a larger m2, and its corresponding transformation function D2, and then update the objective function until a confrontation sample is found. The dimension increase is expressed by the following formula:

[0106] y j =D i -1 (D i-1 (y i-1 ))

[0107] Where y is the range of the attack point cloud, D -1 represents the inverse transform of D.

[0108] Furthermore, the zero-order random coordinate descent, importance sampling technique and multi-scale attack are combined to train the 3D point cloud detection model.

[0109] Step S4, input the point cloud data to be processed into the trained 3D point cloud detection model for target detection, estimate the pose of the identified target, and obtain the target object coordinates; transform the target object coordinates into the robot coordinate system, plan the robot arm motion path, and control the robot arm to complete the target grasping.

[0110] Specifically, step S4 includes the following sub-steps:

[0111] Step S401, collect point cloud data, and input the point cloud data into the trained 3D point cloud detection model for target detection; use the Euclidean clustering segmentation algorithm to segment the target instance; and use the template matching-based target recognition algorithm to perform posture estimation on the segmented target instance.

[0112] Step S402: Obtain the relationship between the workbench coordinate system {S} and the base coordinate system {B} according to the placement position of the robot arm and the workbench Relationship between the table coordinate system {S} and the tool coordinate system {T} And the relationship between the wrist coordinate system {W} and the tool coordinate system {T}

[0113] The relationship between the base coordinate system {B} and the tool coordinate system {T} is obtained, and the formula is expressed as follows:

[0114]

[0115] Step S403, using a random fast search tree algorithm to plan an optimal path, through the inverse solution of the robot arm, the movement of each node of the robot arm is obtained, thereby controlling the movement of each joint of the robot arm to achieve the grasping of the target object; Figure 6 As shown, including:

[0116] Set the starting point X of the random search tree start , set the end point X of the random search tree goal ;

[0117] Randomly generate a random point X in space that does not touch obstacles rand ; Find the distance from random point X rand The nearest random search tree node X near , connect random points X rand and X near , determine whether it touches the obstacle, if not, keep the random point X rand Node, and the random point X rand As a node for random search, repeat the above process until it is connected to the end point X of the random search tree. goal .

[0118] Accordingly, the present application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned target grasping method based on 3D adversarial point cloud. Figure 7 As shown, a hardware structure diagram of a device with data processing capability for the target capture method based on 3D adversarial point cloud provided by an embodiment of the present invention is shown, except Figure 7 In addition to the processor, memory and network interface shown, any device with data processing capability in which the apparatus in the embodiment is located may also include other hardware, generally based on the actual functions of the device with data processing capability, which will not be described in detail.

[0119] Accordingly, the present application also provides a computer-readable storage medium on which computer instructions are stored, and when the instructions are executed by the processor, the target capture method based on 3D adversarial point cloud as described above is implemented. The computer-readable storage medium can be an internal storage unit of any device with data processing capabilities described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), an SD card, a flash card (Flash Card), etc. equipped on the device. Furthermore, the computer-readable storage medium can also include both an internal storage unit and an external storage device of any device with data processing capabilities. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and can also be used to temporarily store data that has been output or is to be output.

[0120] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A target grasping method based on 3D adversarial point cloud, characterized in that: The method comprises the following steps: Get the original first point cloud dataset X0, and input the first point cloud dataset X0 into the diffusion model, and output the second point cloud dataset X T ; For the second point cloud dataset X T Add disturbance to each point cloud data in to generate a perturbed point cloud, and get the adversarial point cloud dataset X′ T ; Build a 3D point cloud detection model and use the adversarial point cloud dataset X′ T As the training set of the 3D point cloud detection model, the 3D point cloud detection model is trained by zero-order random coordinate descent and multi-scale attack method; The point cloud data to be processed is input into the trained 3D point cloud detection model for target detection, and the pose of the identified target is estimated to obtain the target object coordinates; the target object coordinates are transformed into the robot coordinate system, the robot arm motion path is planned, and the robot arm is controlled to complete the target grasping.

2. The target grasping method based on 3D adversarial point cloud according to claim 1, characterized in that: For the second point cloud dataset X T Add disturbance to each point cloud data in to generate a perturbed point cloud, and get the adversarial point cloud dataset X′ T The process includes: The generation of adversarial examples is optimized by maximizing the log-likelihood of the target class, expressed as: subject to||x′-x|| p <∈ In the formula, y t Represents the target class on the original sample, x is the real point cloud, x∈X T , x′ is the generated disturbance point cloud; Each disturbance is modeled as: δ=E t～T [d(t(x′),t(x))] Where t(x) represents the transformation function, T represents the transformation distribution of the transformation function t, and d represents the distance between the real point cloud and the perturbed point cloud after being processed by the transformation function t(x); The adversarial sample generation is expressed as: subject to E t～T [d(t(x′),t(x))]<∈ The constrained Lagrangian relaxation form is used to solve the constrained optimization problem in the adversarial point cloud dataset X′. T , expressed as: Wherein, LAB represents LAB color space.

3. The target grasping method based on 3D adversarial point cloud according to claim 2 is characterized in that: Modeling each disturbance involves: Select the distribution T of the transformation function t(x) for modeling; Among them, the transformation function t(x) maps the texture to the rendering of the object, simulating the rendering, lighting, rotation, translation and perspective projection of the object.

4. The target grasping method based on 3D adversarial point cloud according to claim 1, characterized in that: The process of training a 3D point cloud detection model through zero-order random coordinate descent and multi-scale attack method includes: Set the objective function of the 3D point cloud detection model, randomly select a variable, that is, a coordinate dimension, and iteratively update the variable in the direction of approximately minimizing the objective function along the coordinate. The expression is as follows: x i ←x i +d * Where f(x) is the objective function of the 3D point cloud detection model; i is a standard basis vector; δ * Denotes the variable update value δ obtained by minimizing the objective function, and δ * To update the corresponding variable x i ; Compute an estimate of the gradient and the estimate of the second-order gradient The formula is as follows: Where h is a constant; The minimization problem is solved by Newton iteration method, and δ is obtained * ,include: Choose an update step size η and update the randomly selected variable e i , calculated to obtain and like Then the negative gradient direction is used for updating, and the formula is expressed as: x i ←x i +d * Otherwise the update formula is expressed as: x i ←x i +d * Repeat the above steps to iteratively update the variable x i , until the variable x that minimizes the objective function is obtained i ; At the same time, a multi-scale attack method is used to perform layered attacks to find a suitable perturbation. In the optimization process, a series of dimensions m1, m2, ..., m are designed. n The transformations D1, D2, ..., D n , first select a relatively small scale m1, whose corresponding transformation function is D1, and use these two parameters to update the objective function. As the number of iterations proceeds, when the loss function no longer decreases, gradually increase m, increase the search space, use m2, and its corresponding transformation function D2, and then update the objective function until a confrontation sample is found; the dimension increase is expressed by the following formula: y j =D i -1 (D i-1 (y i-1 )) Where y is the range of the attack point cloud, D -1 represents the inverse transform of D.

5. The target grasping method based on 3D adversarial point cloud according to claim 1, characterized in that: The process of inputting the point cloud data to be processed into the trained 3D point cloud detection model for target detection and estimating the pose of the identified target to obtain the coordinates of the target object includes: Collect point cloud data and input the point cloud data into the trained 3D point cloud detection model for target detection; The target instance is obtained by segmentation using the Euclidean clustering segmentation algorithm; The pose of the segmented target instance is estimated through the target recognition algorithm based on template matching.

6. The target grasping method based on 3D adversarial point cloud according to claim 1, characterized in that: The process of transforming the target object coordinates to the robot coordinate system includes: According to the placement of the robot arm and the workbench, the relationship between the workbench coordinate system (S) and the base coordinate system {B} is obtained Relationship between the table coordinate system (S) and the tool coordinate system {T} And the relationship between the wrist coordinate system {W} and the tool coordinate system {T} The relationship between the base coordinate system {B} and the tool coordinate system {T} is obtained, and the formula is expressed as follows:

7. The target grasping method based on 3D adversarial point cloud according to claim 1, characterized in that: The process of planning the robot's motion path includes: The random fast search tree algorithm is used to plan an optimal path to control the movement of each joint of the robot arm; including: Set the starting point X of the random search tree start , set the end point X of the random search tree goal ; Randomly generate a random point X in space that does not touch obstacles rand ; Find the distance from a random point X rand The nearest random search tree node X near , connect random points X rand and X near , determine whether it touches the obstacle, if not, keep the random point X rand Node, and the random point X rand As a node for random search, repeat the above process until it is connected to the end point X of the random search tree. goal .

8. An electronic device comprising a memory and a processor, characterized in that: The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the target grasping method based on 3D adversarial point cloud as described in any one of claims 1-7 above.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the target grasping method based on 3D adversarial point cloud as described in any one of claims 1 to 7 is implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the target grasping method based on 3D adversarial point cloud described in any one of claims 1-7 is implemented.

Images