Complex workpiece 6D pose estimation method, device, electronic device and storage medium

Abstract

The present application relates to a method, apparatus, electronic device, and storage medium for 6D pose estimation of complex workpieces. The method introduces a farthest point sampling strategy into the point cloud feature extraction module, effectively reducing point cloud redundancy, improving feature representativeness, and effectively reducing computational complexity while retaining key geometric features. Furthermore, the method, combined with a Transformer encoder, further enhances the modeling capability of global features, significantly improving the robustness and accuracy of pose estimation in complex scenarios. A dual-branch regression structure is used to optimize translation and rotation parameters, achieving simultaneous improvements in pose regression accuracy and robustness.

Description

Technical Field

[0001] The present application relates to the technical field of workpiece pose estimation, and in particular to a method, device, electronic device and storage medium for estimating the 6D pose of a complex workpiece. Background Art

[0002] In intelligent manufacturing systems, autonomous robotic grasping technology is a key link in achieving flexible production and efficient logistics. This is especially true in complex industrial environments, where workpieces are diverse, stacked in disorder, and subject to significant occlusion and overlap, placing higher demands on robotic perception and manipulation capabilities. 6D pose estimation (i.e., position and posture in three-dimensional space) is fundamental to achieving high-precision grasping and assembly, and its accuracy directly impacts the success rate of robotic arm path planning and manipulation. Traditional RGB or RGB-D-based vision methods often struggle to maintain robustness and accuracy in complex scenarios such as occlusion, varying lighting conditions, and overlapping objects. In recent years, point cloud deep learning methods, with their powerful ability to model spatial geometric features, have become a research hotspot in the field of robotic perception.

[0003] Point cloud data has natural advantages in 6D pose estimation, but how to efficiently extract the multi-level features of point clouds, improve global information modeling capabilities, and balance computational efficiency and generalization capabilities remain key issues that need to be addressed.

[0004] With the development of depth sensors and point cloud learning technology, point cloud or depth-based methods have gradually attracted attention. CloudPose extracts features from point cloud data using point cloud processing networks (such as PointNet) and uses fully connected layers to directly regress 6D pose parameters from point cloud features. CloudAAE proposes a point cloud-based 6D pose estimation method that achieves high-precision pose estimation through enhanced autoencoders and lightweight data synthesis pipelines. Although the above methods have made some progress in point cloud feature extraction and global information modeling, in complex industrial scenarios, when faced with problems such as multi-object occlusion, stacking, and sparse point clouds, the robustness and accuracy of existing methods still have room for improvement. Summary of the Invention

[0005] Based on this, it is necessary to provide a 6D pose estimation method, device, electronic device and storage medium for complex workpieces to address the shortcomings of existing point cloud methods in feature expression, global modeling and computational efficiency.

[0006] A 6D pose estimation method for a complex workpiece, the method comprising:

[0007] Obtain the RGB image and corresponding depth map of the target workpiece captured by the 3D camera.

[0008] The Mask R-CNN instance segmentation network is used to process the RGB image to obtain the segmentation results.

[0009] Based on the segmentation processing results, the corresponding depth map and the camera intrinsic parameter matrix, a mapping relationship between the similarity coordinate system and the three-dimensional point cloud coordinate system is established, and the pixel points in the depth map are converted into three-dimensional point cloud data to obtain the point cloud of the target workpiece.

[0010] The point cloud feature extraction module is used to extract features from the point cloud of the target workpiece to obtain local features and position features; the point cloud feature extraction module is used to extract features from the point cloud using EdgeConv and the farthest point sampling algorithm.

[0011] After the local features and position features are positionally encoded, the global feature extraction module is used to extract features to obtain global features; the global feature extraction module is used to model the global features of the point cloud using the Transformer encoder.

[0012] The global features are processed using the pose regression module to obtain the 6D pose parameters of the target workpiece in three-dimensional space.

[0013] A 6D pose estimation device for a complex workpiece, comprising:

[0014] The point cloud construction unit of the target workpiece is used to obtain the RGB image and corresponding depth map of the target workpiece captured by the 3D camera; the Mask R-CNN instance segmentation network is used to process the RGB image to obtain the segmentation processing result; based on the segmentation processing result, the corresponding depth map, and the camera intrinsic parameter matrix, a mapping relationship between the similarity coordinate system and the 3D point cloud coordinate system is established, and the pixels in the depth map are converted into 3D point cloud data to obtain the point cloud of the target workpiece.

[0015] The point cloud feature extraction unit is used to extract features from the point cloud of the target workpiece using the point cloud feature extraction module to obtain local features and position features; the point cloud feature extraction module is used to extract features from the point cloud using EdgeConv and the farthest point sampling algorithm.

[0016] The global feature extraction unit is used to position-encode local features and position features and then use the global feature extraction module to extract features to obtain global features; the global feature extraction module is used to model the global features of the point cloud using the Transformer encoder.

[0017] The workpiece pose estimation unit is used to process the global features using the pose regression module to obtain the 6D pose parameters of the target workpiece in three-dimensional space.

[0018] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any of the above methods when executing the computer program.

[0019] A computer-readable storage medium stores a computer program, which implements the steps of any of the above methods when executed by a processor.

[0020] The above-mentioned 6D pose estimation method, device, electronic device and storage medium for complex workpieces effectively reduce point cloud redundancy and improve feature representativeness by introducing the farthest point sampling strategy in the point cloud feature extraction module, effectively reduce computational complexity and retain key geometric features, and further enhance the modeling capability of global features by combining with the Transformer encoder, significantly improving the robustness and accuracy of pose estimation in complex scenes; adopting a dual-branch regression structure to optimize the translation and rotation parameters respectively, achieving a simultaneous improvement in the accuracy and robustness of pose regression. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 1 is a flow chart of a method for estimating 6D pose of a complex workpiece in one embodiment;

[0022] Figure 2 This is a diagram of a network structure for estimating 6D pose of a complex workpiece in one embodiment;

[0023] Figure 3 A schematic diagram of the architecture of a point cloud feature extraction module in one embodiment;

[0024] Figure 4 A schematic diagram of a global feature extraction module framework in another embodiment;

[0025] Figure 5 A schematic diagram of a pose regression module framework in another embodiment;

[0026] Figure 6 This is an example diagram of an object model in another embodiment;

[0027] Figure 7 is a CAD model diagram of a workpiece in another embodiment;

[0028] Figure 8 A graph showing a translation loss function in another embodiment;

[0029] Figure 9 A rotation loss function curve diagram in another embodiment;

[0030] Figure 10 is a total loss function curve diagram in another embodiment;

[0031] Figure 11Graphs showing the success rate and ADD indicator results for the validation set in another embodiment, where (a) is a curve showing the success rate of this method at the ADD distance threshold (10% of the object diameter), and (b) is a diagram showing how the average ADD value changes with the number of iterations.

[0032] Figure 12 is a displacement loss curve diagram of a finished flange in another embodiment;

[0033] Figure 13 is a rotation loss curve diagram of a finished flange in another embodiment;

[0034] Figure 14 is a total loss curve diagram of a finished flange in another embodiment;

[0035] Figure 15 Graphs showing the success index of a finished flange in another embodiment, wherein (a) is a graph showing the success rate of ADD-S (0.1d), and (b) is a graph showing the ADD-S result.

[0036] Figure 16 A schematic diagram of a robotic arm grasping process in another embodiment;

[0037] Figure 17 FIG. 4 is a diagram showing the internal structure of an electronic device in another embodiment. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0039] In one embodiment, Figure 1 As shown, a method for estimating the 6D pose of a complex workpiece is provided, which includes the following steps:

[0040] Step 100: Obtain an RGB image and corresponding depth map of the target workpiece captured by a 3D camera; process the RGB image using the Mask R-CNN instance segmentation network to obtain a segmentation result; establish a mapping relationship from the similarity coordinate system to the 3D point cloud coordinate system based on the segmentation result, the corresponding depth map, and the camera intrinsic parameter matrix, convert the pixels in the depth map into 3D point cloud data, and obtain a point cloud of the target workpiece.

[0041] Specifically, step 100 is the first stage of this method, namely the preprocessing stage. First, using the RGB image acquired by the 3D camera as input, the Mask R-CNN instance segmentation network is used to obtain the bounding box, category label, and instance segmentation mask of a single workpiece. Then, combined with the depth map acquired by the 3D camera and the camera intrinsic parameter matrix, a mapping relationship is established from the pixel coordinate system to the 3D point cloud coordinate system. The pixels in the depth map are converted into 3D point cloud data to obtain the point cloud of the target workpiece.

[0042] This method is based on the Mask R-CNN instance segmentation network. It extracts point cloud data from the depth map and combines local and global feature extraction modules to achieve 6D pose estimation of the target workpiece.

[0043] Step 102: Use the point cloud feature extraction module to extract features from the point cloud of the target workpiece to obtain local features and position features; the point cloud feature extraction module is used to extract features from the point cloud using EdgeConv and the farthest point sampling algorithm.

[0044] Specifically, traditional CloudAAE relies solely on dynamic point cloud feature aggregation, which can easily lose spatial distribution information. Therefore, this method introduces a farthest point sampling (FPS) strategy, which significantly reduces feature redundancy and computational complexity while ensuring uniform spatial coverage of the point cloud, thereby improving sensitivity to object structural details.

[0045] Rich features are extracted from the point cloud through the feature extraction module composed of EdgeConv and FPS (Farthest Point Sampling).

[0046] Step 104: After position encoding the local features and position features, a global feature extraction module is used to extract features to obtain global features; the global feature extraction module is used to model the global features of the point cloud using a Transformer encoder.

[0047] Specifically, a multi-layer Transformer encoder is introduced after feature aggregation, and the self-attention mechanism is used to efficiently capture the global geometric relationship between distant point pairs in the point cloud, effectively solving the problem that traditional convolution / pooling methods are difficult to model complex spatial dependencies.

[0048] Before the features extracted by the point cloud feature extraction module are passed into the Transformer encoder, the feature expression ability is further enhanced using the learnable position encoding, and then the global features of the point cloud are modeled through the Transformer encoder.

[0049] Step 106: The global features are processed using a pose regression module to obtain 6D pose parameters of the target workpiece in three-dimensional space.

[0050] Specifically, the pose regression module is used to predict the translation and rotation parameters of the object in 3D space. Using the obtained global geometric features, the final pose parameters of the object can be predicted. The complex workpiece 6D pose estimation method uses a complex workpiece 6D pose estimation network such as Figure 2 As shown in the figure, the 6D pose estimation network for complex workpieces includes: Mask R-CNN instance segmentation network, point cloud feature extraction module, global feature extraction module and pose regression module.

[0051] In the above-mentioned 6D pose estimation method for complex workpieces, the method effectively reduces point cloud redundancy, improves feature representativeness, effectively reduces computational complexity and retains key geometric features by introducing the farthest point sampling strategy in the point cloud feature extraction module. It also further enhances the modeling capability of global features by combining with the Transformer encoder, significantly improving the robustness and accuracy of pose estimation in complex scenarios. A dual-branch regression structure is used to optimize the translation and rotation parameters respectively, achieving a simultaneous improvement in pose regression accuracy and robustness.

[0052] In one embodiment, the point cloud feature extraction module includes: 1 convolution layer, four EdgeConv layers and two downsampling operations using the FPS algorithm; step 102 includes: inputting the point cloud of the target workpiece into the convolution layer to obtain high-dimensional features; inputting the high-dimensional features into the first EdgeConv layer to obtain a first local feature and its position feature; downsampling the first local feature and its position feature using the FPS algorithm and inputting them into the second EdgeConv layer to obtain a second local feature and its position feature; inputting the second local feature into the third EdgeConv layer to obtain a third local feature and its position feature; downsampling the third local feature and its position feature using the FPS algorithm and inputting them into the fourth EdgeConv layer to obtain a local feature and its position feature.

[0053] In one embodiment, the EdgeConv layer includes: a feature extraction module, a first two-dimensional convolution layer, a first regularization layer, a maximum pooling layer, a splicing layer, a second two-dimensional convolution layer and a second regularization layer; wherein the feature extraction module is used to utilize the coordinate information of the point cloud itself and combine it with the KNN algorithm to construct a dynamic graph structure.

[0054] Specifically, in order for deep learning methods to be able to perform feature extraction on point cloud data, its unique characterization problem needs to be solved. Unlike image data, point cloud data has two key characteristics: permutation invariance and rotation invariance. Permutation invariance means that the order of points in the point cloud will not affect the shape and features of the object it describes, that is, the point cloud data can be arbitrarily arranged, and the same object will still be obtained; rotation invariance means that after the point cloud undergoes a rotation transformation, although the spatial coordinates of the points change, its local features should remain unchanged, and it is still the same object. These characteristics require that the feature extraction network must design a characterization method that is robust to geometric transformations. This embodiment explains the local feature extraction mechanism from two dimensions: dynamic graph convolution and sampling strategy. By using the FPS algorithm in the process of constructing the graph structure, point cloud graph structural features of different scales are obtained to ensure that important features of the point cloud can be captured at different scales. Optimize the fusion of point cloud feature expression.

[0055] (1) Dynamic Graph Convolutional Neural Network

[0056] Unlike PointNet and PointNet++, which directly use point clouds as network input, the Dynamic Graph Convolutional Neural Network (DGCNN) adopts a more sophisticated approach, capturing the geometric relationships between points by constructing a graph structure. In DGCNN, each point in the point cloud is treated as a node in the graph, and the relationships between points are represented by edges. This allows the network to capture the geometric relationships between points by performing convolution operations on the graph structure.

[0057] The calculation process of EdgeConv, the core module of the dynamic graph convolutional neural network, consists of three stages:

[0058] 1) Construct the graph structure. For input n points F dimensional point cloud , where pi represents the characteristics of each point (3D coordinates, color, normal, etc.), if F =3, then it can be expressed as the three-dimensional coordinates of each point . Use the distance metric to calculate its K nearest neighbors for each point ,in is a set of vertices, is the edge set.

[0059] 2) Calculate edge features, for each edge connecting the center point and neighbors , calculate the edge features. The edge features are calculated as follows:

[0060] ;

[0061] in, Typically a multi-layer perceptron (MLP), is a learnable parameter.

[0062] 3) Aggregate edge features, use maximum pooling to aggregate neighborhood information, and update center point features:

[0063] ;

[0064] in, is the updated central feature.

[0065] (2) Farthest point sampling algorithm

[0066] The Farthest Point Sampling (FPS) algorithm achieves point cloud downsampling by iteratively selecting the maximum distance points. Its core is to reduce the number of points while preserving the key geometric features of the point cloud as much as possible. The specific process of the algorithm includes: randomly selecting a starting point from the point cloud set P and adding it to the sampling point set S. This starting point can be selected arbitrarily because the iterative nature of the algorithm will ensure the consistency of the final result. Calculate the distance from all points in the point cloud to p 0 point, select the point with the largest distance as p 1, and add it to the sampling point set S. Repeat the above steps until the required N Up to one point.

[0067] (3) Point cloud feature extraction module

[0068] Point cloud feature extraction module such as Figure 3 As shown in the figure, the input of the point cloud feature extraction module is the original point cloud data. In order to map the point cloud data to a higher dimensional feature space, this method uses a one-dimensional convolution operation. Through this operation, the coordinate information of each point is converted into a vector with 8-dimensional features, thus laying the foundation for the subsequent feature extraction process. This preliminary feature mapping step enables the point cloud data to be effectively processed and analyzed in a higher dimensional space. After completing the initial feature mapping, the coordinate information of the point cloud itself is combined with the KNN (K-Nearest Neighbors) algorithm to construct a dynamic graph structure. Specifically, for each center point in the graph, the KNN algorithm is used to find its kneighboring points. To further enhance the expressiveness of the features, the feature difference between the center point and the adjacent edge points is used as the edge feature. In this way, a feature tensor containing both point features and edge features is obtained. Subsequently, a convolution operation is performed on the feature tensor, and the convolution kernel is used to slide the feature tensor to extract feature information within the local neighborhood. To prevent overfitting and ensure the stability and generalization ability of the features, the convolved features are regularized. Finally, a maximum pooling operation is performed on the features along the neighborhood dimension to obtain the aggregated features. This aggregated feature can effectively capture key information within the local neighborhood.

[0069] After completing the above steps, a complete point cloud feature extraction process is completed. The original CloudAEE network performed this operation four times, increasing the point cloud dimension from 3 to 128 while maintaining the number of points. This design effectively extracts multi-layered features from the point cloud, but it carries a high computational complexity, especially when processing large-scale point cloud data. To further reduce computational complexity and extract more representative features, this method uses the FPS algorithm to downsample the point cloud twice. The first downsampling operation downsamples the number of points from 2048 to 512. The second downsampling operation downsamples the number of points from 512 to 128. The above point cloud feature extraction process is repeated before and after downsampling. The post-downsampling extraction operation can be considered a secondary feature extraction of the point cloud data, further enhancing the hierarchical nature and representativeness of the features. After extracting point cloud features, the original network also concatenates the point cloud features of each dimension for subsequent input. However, due to the downsampling operation, the dimensions of each layer no longer align with the number of points in the cloud, making direct concatenation impossible. Therefore, this method uses the features and coordinates obtained from the last layer for subsequent input. The features of the last layer have undergone multiple downsampling and feature extraction processes, and have higher hierarchy and representativeness.

[0070] Through this improved point cloud feature extraction network architecture, the multi-level features of the point cloud can be effectively extracted, and the complexity of the point cloud can be gradually reduced through a reasonable sampling strategy, thereby providing high-quality input features for subsequent modules, thereby improving the performance and efficiency of the entire feature network.

[0071] In one embodiment, the global feature extraction module includes: a feature fusion module, a Transformer encoder, an MLP, a maximum pooling layer, and a fully connected layer; step 104 includes: position encoding the local features and position features and inputting them into the feature fusion module to obtain fused input features; inputting the fused input features into the Transformer encoder to obtain encoded features; and processing the encoded features successively through the MLP, the maximum pooling layer, and the fully connected layer to obtain global features.

[0072] Specifically, (1) Transformer structure

[0073] The Transformer was originally proposed by Vaswani et al. for natural language processing (NLP) tasks. It has since been expanded to computer vision (CV), point cloud, multimodal learning, and various other tasks. A standard Transformer encoder consists of five main components.

[0074] Input embedding: The main function of input embedding is to map the raw input data into a high-dimensional vector space so that the model can process and learn. Specifically, the role of input embedding varies depending on the input source. For text data, each word is mapped to a fixed-dimensional vector through a learnable embedding matrix. This can convert discrete data (such as words) into continuous data and also capture semantic information of the input data (such as semantic similarity between words). The formula is:

[0075] ;

[0076] in, E is the embedding matrix, i is the index of the word, E i is the corresponding word embedding vector. Positional Encoding: In the Transformer model, since the self-attention mechanism itself does not have the ability to explicitly model the order of the input sequence, it is necessary to introduce the position information of the elements in the sequence through positional encoding. Self-Attention Mechanism: The self-attention mechanism generates attention weights by calculating the similarity between each element in the sequence and other elements. These weights are used to weight the information of all elements in the aggregated sequence to capture the global contextual relationship. The calculation formula of the self-attention mechanism is as follows:

[0077] ;

[0078] Where Q, K, and V represent the query, key, and value matrices, respectively, and are the dimensions of the key vector. Feed-Forward Network (FFN): Following the self-attention mechanism, the Transformer uses a two-layer feed-forward neural network to perform nonlinear transformations on features. Residual Connections and Layer Normalization: To mitigate the vanishing gradient problem, the Transformer adds residual connections and layer normalization operations after each sublayer (self-attention and FFN).

[0079] (2) Global feature extraction module

[0080] In the original network, only one global pooling operation is used to extract global features, but the ability to capture global context information is limited. To make up for this shortcoming, this application introduces the Transformer architecture based on the CloudAEE network to enhance the global feature extraction capability. The global feature extraction module of this method is as follows Figure 4 As shown in the figure, the input to the Transformer encoder is the features processed by the point cloud feature extraction module. To further enhance the expressiveness of the features, the position information of the point cloud is fused with the local features. The position information is first processed through a fully connected layer for dimensionality increase. The dimensionality increase is then added to the local features to form the fused input features. After layer normalization, the fused input features are fed into the multi-head self-attention processing unit. The attention mechanism then connects the input with a residual and regularizes it. A feedforward neural network then applies a nonlinear transformation to the features. Residual connections and layer normalization are added. A dimensionality reduction mapping is then performed to restore the features to their original size. The features are then concatenated with the regularized attention mechanism features through the residual, ultimately yielding the final global features of the point cloud.

[0081] In one embodiment, the local features and position features are position-encoded and input into a feature fusion module to obtain fused input features, including: position-encoding the local features and position features, respectively processing them through a fully connected layer and adding them together, and then performing layer normalization processing to obtain fused input features.

[0082] In one embodiment, the pose regression module includes: two independent branch prediction networks, each branch prediction network includes three fully connected layers; the first branch prediction network is used to estimate the translation parameters, and the second branch prediction network is used to estimate the rotation parameters; step 106 includes: inputting the global feature into the first branch prediction network to obtain the three-dimensional coordinate offset of the target workpiece; inputting the global feature into the second branch prediction network to obtain the axis angle representation regression target of the target workpiece.

[0083] Specifically, the structure of the pose regression module is as follows Figure 5 As shown in Figure 2. This module uses two independent branch prediction networks to estimate translation and rotation parameters respectively. Each sub-branch consists of three layers of fully connected layers. The translation branch directly regresses the 3D coordinate offset. The global features are processed through the three layers of fully connected layers, and the final output is the translation offset. t pred The rotation branch uses axis-angle representation as the regression target. The axis-angle representation is represented by a three-dimensional vector Describes the geometric properties of rotation: its direction represents the unit vector of the rotation axisa , the module length corresponds to the rotation angle Based on the Rodrigue formula, the axis-angle parameters can be mapped to the rotation matrix R .

[0084] During the pose regression phase, a dual-branch architecture is employed to optimize translation and rotation parameters separately. The rotation branch employs quaternion regression and introduces a pose consistency loss, while the translation branch combines point cloud centers with geometric constraints to improve positioning accuracy. These improvements significantly enhance the robustness of pose estimation in complex scenarios like occlusion and stacking, while also maintaining inference efficiency.

[0085] In one embodiment, the overall loss function of the pose regression module is defined as the weighted sum of rotation loss and translation loss; the overall loss function expression of the pose regression module is:

[0086] ;

[0087] ;

[0088] ;

[0089] in, is the overall loss function of the pose regression module, 、 are the weights of translation loss and rotation loss respectively, 、 They are rotation loss and translation loss respectively, It means finding the trace of the relative transformation matrix of two rotations, is a function that outputs the minimum rotation angle between two rotations, R gt and R pred are the true rotation matrix and the predicted rotation matrix respectively, is the predicted translation value in the global coordinate system, is the actual translation value.

[0090] Specifically, in order to quantify the difference between the predicted rotation and true values, a rotation loss function and a translation loss function are designed respectively. R gt and R pred are the true rotation matrix and the predicted rotation matrix respectively. The rotation matrix loss function is defined as:

[0091] ;

[0092] in, It means finding the trace of the relative transformation matrix of two rotations, A function that outputs the minimum rotation angle between two rotations.

[0093] For the translation offset prediction task, since the predicted translation value is based on the local coordinate system after the input point cloud is centered, it is necessary to remap the predicted value back to the original coordinate system and then compare it with the true value. Let the centroid of the input point cloud be , the predicted translation offset is t pred , the complete predicted shift value should be calculated by the following formula:

[0094] ;

[0095] in, is the point cloud centroid, and the calculation formula is:

[0096] ;

[0097] in, N represents the number of points in the point cloud, Indicates the i The coordinates of the points. Get the predicted translation value in the global coordinate system Then, you can compare it with the real translation value Compare and calculate the translation loss function. The translation loss function usually uses the L2 norm to measure the difference between the two:

[0098] .

[0099] The overall loss function for the pose regression module can be defined as:

[0100] ;

[0101] in 、 They are the weights of translation loss and rotation loss respectively. They are used to balance the effects of rotation loss and translation loss. 、 You can control how much emphasis the model places on optimizing rotation and translation predictions.

[0102] Experimental results show that this method achieves excellent performance on both the standard LineMod dataset and a self-made industrial artifact dataset, verifying its application value in actual industrial scenarios.

[0103] It should be understood that although Figure 1The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0104] In a verification example, the experimental hardware platform uses an NVIDIA GeForce RTX 4090D graphics card (24G video memory), an Intel(R) Xeon(R) Platinum 8474C processor (32 cores and 64 threads, main frequency 3.60GHz), and 64GB of DDR4 memory; the system runs on Ubuntu 20.04.

[0105] At the software level, the algorithm is implemented in Python 3.8, using PyTorch 1.10.0 as the deep learning framework, and relying on the CUDA 11.3 and cuDNN 8.2 acceleration libraries. The specific configuration is shown in Table 1.

[0106] Table 1 Experimental environment

[0107]

[0108] To optimize the model training process, this experiment uses the following training strategy and parameter configuration: AdamW is used as the optimizer, the initial learning rate is set to 0.0003, and the learning rate is adjusted using the StepLR strategy; the weight decay parameter is 0.0001 to prevent overfitting; the batch size is set to 64; and the total number of training rounds is 200 to ensure sufficient model convergence. The specific parameter configurations are shown in Table 2.

[0109] Table 2 Training parameters

[0110]

[0111] (1) Dataset

[0112] 1) Benchmark Datasets

[0113] This example is primarily validated on two datasets: the LineMod dataset and a self-made artifact dataset. LineMod contains 13 low-texture everyday objects (such as beverage cans, toy ducks, and glue bottles), each corresponding to approximately 1,200 data units. Each data unit contains a strictly aligned RGB image, a depth map, camera intrinsics, and precise 6D pose annotations (rotation matrices and translation vectors). LineMod data is collected in real-world scenes, where the target object surfaces lack significant texture features and the background environment is complex with dynamic lighting changes. This example follows the consensus in the field, selecting 15% of each type of object for training and the remaining 85% for testing. Figure 6 Shows 13 object models selected from the LineMod dataset.

[0114] 2) Self-made dataset Although the LineMod dataset provides a standardized evaluation benchmark for 6D pose estimation algorithms, there are still significant differences in its application to specific industries. In order to effectively migrate the algorithm to the actual production environment, this embodiment constructs a self-made dataset for workpiece stacking scenarios. This dataset consists of two core components: real annotated data: 100 annotated samples of processed and unprocessed workpieces, collected through multi-view RGB-D image acquisition and manual calibration; CAD models of workpieces such as Figure 7 shown.

[0115] (2) Evaluation indicators

[0116] To evaluate the performance of the 6D pose estimation network, this example uses the ADD and ADD-S metrics. For asymmetric objects, the ADD metric calculates the average deviation of the distance between the predicted pose and the true pose of the 3D model points at each point, and determines whether the accuracy of the distance is less than a certain threshold of the object diameter (for example, ADD-0.1d). It is defined as follows:

[0117] ;

[0118] Where: R and t are the true rotation and translation, and are the estimated rotation and translation.

[0119] For symmetrical objects, the network prediction results and the true value may differ greatly on the rotation axis. However, this does not necessarily mean that the actual effect is poor. Therefore, the average distance between the closest points ADD-S is used to evaluate the network. It is defined as follows:

[0120] .

[0121] (3) Experimental comparison and analysis

[0122] 1) LineMod dataset

[0123] To verify the effectiveness of the improved model, this embodiment trains 13 classes on the LineMod dataset. Due to space limitations, this embodiment only analyzes the training process on the Cam class on the LineMod dataset. Figure 8 、 Figure 9 、 Figure 10 As shown in the figure, the translation loss (Translation Loss), rotation loss (Rotation Loss) and overall loss of the Cam class in the improved model are shown respectively.

[0124] Experimental results show that the rotation loss and translation loss for the training set (red curve) and validation set (blue curve) drop sharply in the first 50 epochs before gradually stabilizing and remaining at a low level. When the number of training iterations reaches 75 epochs, the loss functions converge. When the network iterations are complete, the translation loss and rotation loss of the final model converge to 1.2076×10-6 and 0.103408, respectively, on the training set. On the validation set, the translation loss and rotation loss converge to 2.4004×10-5 and 0.136103, respectively.

[0125] In terms of evaluation indicators, the success rate of this method under the ADD distance threshold (10% of the object diameter) and the change of the ADD average value with the number of iterations are shown as follows: Figure 11 As shown, Figure 11 (a) is the success rate curve of this method under the ADD distance threshold (10% of the object diameter). Figure 11 (b) is a schematic diagram of how the average value of ADD changes with the number of iterations.

[0126] After 200 epochs of training, the model achieved an accuracy of 95.4% and an average ADD value of 0.009989 in the category.

[0127] This method was experimentally validated on the LineMod dataset. Based on the differences in object symmetry, the ADD (for asymmetric objects) and ADD-S (for symmetric objects) metrics were used. For asymmetric objects, the symmetric objects in the LineMod dataset include Egg Box and Glue. The experimental results are shown in Table 3, demonstrating that this method achieves significant improvements across most object categories. A detailed analysis is provided below.

[0128] Table 3 Accuracy comparison on LineMod dataset

[0129]

[0130] The table above compares the performance of various methods on the LineMod dataset. After improving the CloudAEE network, the cam category saw the highest improvement compared to the original network across 13 categories, with the success rate increasing from 88.9% to 95.4% (+6.5%). Although there were slight fluctuations in the iron and lamp categories, with the iron category decreasing slightly by 1.3% and the lamp category decreasing by 0.8%, the improved method achieved positive improvements in 11 of the 13 categories, with an average success rate increase of 1.6% (95.5% → 97.1%), demonstrating the effectiveness of the improvements.

[0131] 2) Self-made dataset

[0132] Based on the validation results of the benchmark dataset, this example further applies the proposed pose estimation framework to a specific scenario. First, instance segmentation techniques are used to locate the target workpiece. A point cloud of the target object is then extracted and used as input to the pose estimation network. The improved pose estimation network is then used to calculate the workpiece's pose. The training parameters remain consistent with those in Table 2.

[0133] The experiment selected two types of industrial workpieces (blank flanges and finished flanges) for verification. This example focuses on the training process data of finished flanges, including the convergence curves of translation loss, rotation loss and total loss. Figure 12 、 Figure 13 、 Figure 14 shown.

[0134] Experimental results show that after 75 epochs, all loss functions converge, consistent with the performance on the benchmark dataset. This indicates that the model gradually stabilizes during training and the optimization effect is significant. Upon completion of the network iterations, the final model's translation loss and rotation loss converge to 3.130049×10-6 and 1.482548, respectively, on the training set. On the validation set, the translation loss and rotation loss converge to 1.010853×10-6 and 1.538814, respectively. These results demonstrate that the improved network exhibits good convergence and stability during both training and validation, and is capable of effectively estimating the pose of workpieces in real-world grasping scenarios.

[0135] The rough flange and finished flange in the self-made dataset are both symmetrical objects. This embodiment uses ADD-S to measure the accuracy. The success rate of the training process on the finished flange validation set is as follows: Figure 15 As shown, Figure 15 (a) is the success rate result of ADD-S (0.1d). Figure 15 (b) in the middle is the ADD-S result diagram.

[0136] The accuracy of ADD-S on the finished flange validation set showed an overall upward trend. When the training cycle reached 200 times, the improved model achieved 100% accuracy, with an ADD-S of 0.002906.

[0137] Table 4 Accuracy comparison on the self-made artifact dataset

[0138]

[0139] Table 4 shows that the improved network achieves significant improvements in accuracy on the self-made data. The accuracy on rough workpieces increases by 0.3%, reaching 100%, and the accuracy on finished workpieces increases by 0.5%, reaching 100%. The average ADD-S improves by 0.4%, from 99.6% to 100%. These experimental results demonstrate the effectiveness of the algorithmic improvements.

[0140] (4) Grasping experiment of workpiece grasping platform

[0141] In order to verify the effectiveness of the pose estimation algorithm, this method was deployed on an actual industrial workpiece grasping platform, and its grasping performance in typical industrial scenarios such as small amounts and dense stacking was systematically evaluated.

[0142] In order to verify the effectiveness of the algorithm, workpiece grasping experiments under different stacking degrees were designed, including a small number of workpiece stacking scenarios (8 workpieces of each type) and a large number of mixed workpiece stacking scenarios (14 workpieces of each type).

[0143] 1) Scenario of stacking a small number of workpieces

[0144] The experimental process of a small number of workpieces stacking scene is as follows Figure 16 As shown. The experimental goal is to place all the workpieces in the workpiece box into the material library. The grabbing process is (serial number corresponds to Figure 16 The robot arm is in its initial position and uses the vision system to identify the workpiece in the workpiece frame and obtain its position information. 2. Depending on the type of workpiece, the robot arm switches to the appropriate fixture. For example, if it is a finished flange, it switches to the internal support fixture; if it is a rough flange, it switches to the suction cup. 3. After the fixture switch is complete, the robot arm returns to its initial position. 4. The robot arm grasps the workpiece. 5. The robot arm places the grasped workpiece on the flipping table. 6. To flip the workpiece, the robot arm switches to the external support fixture. 7. The robot arm completes the flipping operation. 8. The robot arm switches back to the internal support fixture to grasp the workpiece. 9. The robot arm places the workpiece into the material storage. This process repeats until all workpieces in the workpiece frame have been grasped.

[0145] For the experiment involving a small number of workpieces, the workpiece frame contained eight blank flanges and eight finished flanges, and 80 grasping experiments were performed on each type of workpiece. The experimental results are shown in the following table.

[0146] Table 5 Experimental results of a small number of workpieces stacking scenario

[0147]

[0148] The blank flange and finished flange were grasped 80 times, with 76 success times (95% success rate) and 74 success times (92.5% success rate) respectively. The experimental results show that the pose estimation algorithm proposed in this application can meet the grasping requirements when facing a small amount of stacking.

[0149] 2) Scenario where a large number of workpieces are piled up

[0150] The workflow for the mixed stacking of a large number of workpieces was identical to that for the small number of workpieces. The workpiece frame contained 14 rough flanges and 14 finished flanges, and 140 grasping experiments were conducted for each type of workpiece. The experimental results are shown in Table 6.

[0151] Table 6 Experimental results of a small number of workpieces stacking scenario

[0152]

[0153] The robot grasped 140 blank flanges, achieving a success rate of 130 (92.9%) and 125 (89.3%), respectively. The data indicates that while the success rate for densely stacked workpieces is lower than for smaller stacks, it still achieved a relatively high grasping success rate. Overall, the success rate for blank flanges is higher than for finished flanges because the fixture used to grip the blank flanges is a suction cup, while the fixture used to grip the finished flanges is an internal support fixture. When the internal support fixture grips a flange close to the edge of the workpiece frame, the inherent characteristics of the internal support fixture apply an outward force to the workpiece, causing it to collide with the workpiece frame, triggering the robot arm's collision protection and resulting in a failed grasp. This is also the primary cause of failed grasping of finished flanges. The primary cause of failed grasping of blank flanges is often height deviation when acquiring the target workpiece's position, resulting in poor contact between the suction cup and the blank flange and a failed grasping of the workpiece.

[0154] This method was fully validated on the standard LineMod dataset and a self-made industrial artifact dataset. The results showed that the proposed method achieved an average ADD(-S) accuracy of 97.1% on the LineMod dataset and a 100% grasping success rate for both types of artifacts on the self-made dataset, significantly outperforming the original CloudAEE network and other mainstream methods, verifying the effectiveness of the method and its potential for engineering application. Furthermore, this embodiment deployed the algorithm on an actual industrial artifact grasping platform and conducted system experiments in a small amount of stacking and a dense stacking scenario, achieving grasping success rates of 93.8% and 91.1%, respectively. The experimental results show that the proposed method not only has good algorithmic performance, but can also effectively support actual industrial automated grasping tasks, and has strong potential for engineering application.

[0155] In one embodiment, a 6D pose estimation device for a complex workpiece is provided, comprising: a point cloud construction unit for a target workpiece, a point cloud feature extraction unit, a global feature extraction unit, and a workpiece pose estimation unit, wherein:

[0156] The point cloud construction unit of the target workpiece is used to obtain the RGB image and corresponding depth map of the target workpiece captured by the 3D camera; the Mask R-CNN instance segmentation network is used to process the RGB image to obtain the segmentation processing result; based on the segmentation processing result, the corresponding depth map, and the camera intrinsic parameter matrix, a mapping relationship between the similarity coordinate system and the 3D point cloud coordinate system is established, and the pixels in the depth map are converted into 3D point cloud data to obtain the point cloud of the target workpiece.

[0157] The point cloud feature extraction unit is used to extract features from the point cloud of the target workpiece using the point cloud feature extraction module to obtain local features and position features; the point cloud feature extraction module is used to extract features from the point cloud using EdgeConv and the farthest point sampling algorithm.

[0158] The global feature extraction unit is used to position-encode local features and position features and then use the global feature extraction module to extract features to obtain global features; the global feature extraction module is used to model the global features of the point cloud using the Transformer encoder.

[0159] The workpiece pose estimation unit is used to process the global features using the pose regression module to obtain the 6D pose parameters of the target workpiece in three-dimensional space.

[0160] In one embodiment, the point cloud feature extraction module includes: 1 convolution layer, four EdgeConv layers and two downsampling operations using the FPS algorithm; the point cloud feature extraction unit is also used to input the point cloud of the target workpiece into the convolution layer to obtain high-dimensional features; input the high-dimensional features into the first EdgeConv layer to obtain first local features and their position features; downsample the first local features and their position features using the FPS algorithm and input them into the second EdgeConv layer to obtain second local features and their position features; input the second local features into the third EdgeConv layer to obtain third local features and their position features; downsample the third local features and their position features using the FPS algorithm and input them into the fourth EdgeConv layer to obtain local features and their position features.

[0161] In one embodiment, the EdgeConv layer in the point cloud feature extraction unit includes: a feature extraction module, a first two-dimensional convolution layer, a first regularization layer, a maximum pooling layer, a splicing layer, a second two-dimensional convolution layer and a second regularization layer; wherein the feature extraction module is used to utilize the coordinate information of the point cloud itself and combine it with the KNN algorithm to construct a dynamic graph structure.

[0162] In one embodiment, the global feature extraction module includes: a feature fusion module, a Transformer encoder, an MLP, a maximum pooling layer and a fully connected layer; the global feature extraction unit is also used to position-encode local features and position features and input them into the feature fusion module to obtain fused input features; input the fused input features into the Transformer encoder to obtain encoded features; and the encoded features are processed by the MLP, the maximum pooling layer and the fully connected layer to obtain global features.

[0163] In one embodiment, the global feature extraction unit is further used to position-encode the local features and position features, respectively process them through a fully connected layer, and then add them together, and then perform layer normalization processing to obtain a fused input feature.

[0164] In one embodiment, the pose regression module includes: two independent branch prediction networks, each branch prediction network includes three fully connected layers; the first branch prediction network is used to estimate the translation parameters, and the second branch prediction network is used to estimate the rotation parameters; the workpiece pose estimation unit is also used to input the global features into the first branch prediction network to obtain the three-dimensional coordinate offset of the target workpiece; and input the global features into the second branch prediction network to obtain the axis angle representation regression target of the target workpiece.

[0165] In one embodiment, the overall loss function of the pose regression module in the workpiece pose estimation unit is defined as the weighted sum of rotation loss and translation loss; the overall loss function of the pose regression module is shown in the above-mentioned overall loss function expression of the pose regression module.

[0166] For the specific definition of the complex workpiece 6D pose estimation device, please refer to the definition of the complex workpiece 6D pose estimation method above, which will not be repeated here. The various modules in the above-mentioned complex workpiece 6D pose estimation device can be implemented in whole or in part by software, hardware, and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0167] In one embodiment, an electronic device is provided. The electronic device may be a terminal, and its internal structure diagram may be as follows: Figure 17 As shown. The electronic device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the electronic device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a method for estimating the 6D pose of a complex workpiece is implemented. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the electronic device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the electronic device, or an external keyboard, touchpad or mouse.

[0168] Those skilled in the art will understand that Figure 17 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the electronic device to which the solution of the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0169] In one embodiment, an electronic device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiment when executing the computer program.

[0170] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above method embodiment are implemented.

[0171] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0172] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0173] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the scope of the present application, and such modifications and improvements are all within the scope of protection of the present application.

Claims

1. A 6D pose estimation method for a complex workpiece, characterized in that: The method comprises: Obtain the RGB image and corresponding depth map of the target workpiece captured by the 3D camera; The RGB image is processed using a Mask R-CNN instance segmentation network to obtain a segmentation result; Based on the segmentation processing results, the corresponding depth map and the camera intrinsic parameter matrix, a mapping relationship between the similarity coordinate system and the three-dimensional point cloud coordinate system is established, and the pixels in the depth map are converted into three-dimensional point cloud data to obtain a point cloud of the target workpiece; A point cloud feature extraction module is used to extract features from the point cloud of the target workpiece to obtain local features and position features; the point cloud feature extraction module is used to extract features from the point cloud using EdgeConv and the farthest point sampling algorithm; After position encoding the local features and the position features, a global feature extraction module is used to extract features to obtain global features; the global feature extraction module is used to model the global features of the point cloud using a Transformer encoder; The global features are processed using a pose regression module to obtain 6D pose parameters of the target workpiece in three-dimensional space; wherein the pose regression module includes: two independent branch prediction networks, each branch prediction network includes three fully connected layers; the first branch prediction network is used to estimate translation parameters, and the second branch prediction network is used to estimate rotation parameters.

2. The complex workpiece 6D pose estimation method according to claim 1, characterized in that: The point cloud feature extraction module includes: 1 convolution layer, four EdgeConv layers and two downsampling operations using the FPS algorithm; The point cloud feature extraction module is used to extract features from the point cloud of the target workpiece to obtain local features and position features, including: Inputting the point cloud of the target workpiece into the convolution layer to obtain high-dimensional features; Input the high-dimensional features into the first EdgeConv layer to obtain the first local features and their position features; Downsampling the first local feature and its position feature using the FPS algorithm and inputting them into the second EdgeConv layer to obtain the second local feature and its position feature; Input the second local feature into the third EdgeConv layer to obtain the third local feature and its position feature; The third local feature and its position feature are downsampled using the FPS algorithm and then input into the fourth EdgeConv layer to obtain the local feature and its position feature.

3. The complex workpiece 6D pose estimation method according to claim 2, characterized in that: The EdgeConv layer includes: a feature extraction module, a first two-dimensional convolution layer, a first regularization layer, a maximum pooling layer, a splicing layer, a second two-dimensional convolution layer and a second regularization layer; The feature extraction module is used to utilize the coordinate information of the point cloud itself and combine it with the KNN algorithm to construct a dynamic graph structure.

4. The complex workpiece 6D pose estimation method according to claim 1, characterized in that: The global feature extraction module includes: feature fusion module, Transformer encoder, MLP, maximum pooling layer and fully connected layer; After position encoding the local features and the position features, a global feature extraction module is used to extract features to obtain global features, including: The local features and the position features are position-encoded and input into a feature fusion module to obtain fused input features; Inputting the fused input features into the Transformer encoder to obtain encoded features; The encoded features are processed successively by MLP, maximum pooling layer and fully connected layer to obtain global features.

5. The complex workpiece 6D pose estimation method according to claim 4, characterized in that: The local features and the position features are position-encoded and input into a feature fusion module to obtain fused input features, including: The local features and the position features are position-encoded and then processed by a fully connected layer and added together, and then layer normalization is performed to obtain a fused input feature.

6. The complex workpiece 6D pose estimation method according to claim 1, characterized in that: The global features are processed using a pose regression module to obtain the 6D pose parameters of the target workpiece in three-dimensional space, including: Inputting the global feature into a first branch prediction network to obtain a three-dimensional coordinate offset of the target workpiece; The global features are input into the second branch prediction network to obtain the axis angle representation regression target of the target workpiece.

7. The complex workpiece 6D pose estimation method according to claim 1, characterized in that: The overall loss function of the pose regression module is defined as the weighted sum of rotation loss and translation loss; the overall loss function of the pose regression module is: in, is the overall loss function of the pose regression module, 、 are the weights of translation loss and rotation loss respectively, 、 They are rotation loss and translation loss respectively, It means finding the trace of the relative transformation matrix of two rotations, is a function that outputs the minimum rotation angle between two rotations, R gt and R pred are the true rotation matrix and the predicted rotation matrix respectively, is the predicted translation value in the global coordinate system, is the actual translation value.

8. A 6D pose estimation device for a complex workpiece, characterized in that: The device comprises: The target workpiece point cloud construction unit is used to obtain an RGB image and corresponding depth map of the target workpiece captured by a 3D camera; process the RGB image using a Mask R-CNN instance segmentation network to obtain a segmentation result; establish a mapping relationship between a similar coordinate system and a 3D point cloud coordinate system based on the segmentation result, the corresponding depth map, and the camera intrinsic parameter matrix, and convert the pixels in the depth map into 3D point cloud data to obtain a point cloud of the target workpiece; A point cloud feature extraction unit is used to extract features from the point cloud of the target workpiece using a point cloud feature extraction module to obtain local features and position features; the point cloud feature extraction module is used to extract features from the point cloud using EdgeConv and the farthest point sampling algorithm; A global feature extraction unit is used to position-encode the local features and the position features and then use a global feature extraction module to extract features to obtain global features; the global feature extraction module is used to model the global features of the point cloud using a Transformer encoder; A workpiece pose estimation unit is used to process the global features using a pose regression module to obtain 6D pose parameters of the target workpiece in three-dimensional space; wherein the pose regression module includes: two independent branch prediction networks, each branch prediction network includes three fully connected layers; the first branch prediction network is used to estimate translation parameters, and the second branch prediction network is used to estimate rotation parameters.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for estimating the 6D pose of a complex workpiece according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for estimating the 6D pose of a complex workpiece according to any one of claims 1 to 7 are implemented.

Images