A training-free multi-target 6D pose estimation method, system, medium, and device

Abstract

This invention discloses a training-free multi-target 6D pose estimation method, system, medium, and device, relating to the field of computer technology. The method includes: acquiring an RGB texture map of the target to be tested; determining the texture map and corresponding segmented point cloud of the target to be tested; determining a scaled texture map of the target to be tested; acquiring texture maps from multiple offline reference frames; determining an initial similarity ranking between each texture map in the offline reference frames and the scaled texture map of the target to be tested; selecting point cloud data from multiple offline reference frames with the highest similarity to the target to be tested based on the initial similarity ranking, and determining the Chamfer distance between the point cloud data of the offline reference frames and the segmented point cloud to determine the most similar reference frame; extracting feature points and descriptors from the texture map of the most similar reference frame, matching them with feature points and descriptors in the texture map of the target to be tested to obtain feature matching point pairs; and combining RANSAC and SVD to determine the pose estimation result of the target to be tested.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a training-free multi-target 6D pose estimation method, system, medium, and device. Background Technology

[0002] Real-time pose estimation for multiple targets often requires a large amount of labeled data and computational resources. Therefore, low-cost, training-free real-time 6D estimation methods play a very important role in assembly and loading operations in industrial production environments such as wind power equipment, as well as in the rapidly evolving applications of embodied intelligence technology.

[0003] Training-free multi-object 6D pose estimation methods focus on extracting visual information through efficient pre-trained models to accurately estimate the position and pose of objects in 3D space. The main challenges they face are:

[0004] In practical applications, it is often necessary to acquire the 6D pose information of an object in real time, which places demands on the computational efficiency of the algorithm. At the same time, the accuracy of pose estimation is also crucial to ensuring the robot's operational precision.

[0005] Traditional training-free 6D pose estimation methods, such as PnP algorithm, ICP registration, and template matching, are not well adapted to complex environments and have difficulty handling situations where multiple targets are stacked or occluded, resulting in poor pose estimation accuracy. Summary of the Invention

[0006] This invention provides a training-free multi-target 6D pose estimation method, system, medium, and device to solve the aforementioned problems in the prior art, namely, how to improve the accuracy of pose estimation in the prior art. This invention provides a training-free multi-target 6D pose estimation method, which includes:

[0007] The RGB texture map of the target under test is obtained by RGBD stream, the RGB texture map is segmented, and the texture map of the target under test and the corresponding segmentation point cloud are determined.

[0008] The texture map of the target under test is scaled to obtain a scaled texture map of the target under test;

[0009] Acquire texture maps from multiple offline reference frames and determine the initial similarity ranking between each texture map in the offline reference frames and the scaled texture map of the target under test;

[0010] Acquire point cloud data from multiple offline reference frames of the target under test;

[0011] Based on the initial similarity ranking, the point cloud data of multiple offline reference frames with the highest similarity to the target to be tested are selected, and the Chamfer distance between the point cloud data of the offline reference frames and the segmented point cloud is determined. The reference frame with the smallest Chamfer distance is taken as the most similar reference frame.

[0012] Feature points and descriptors are extracted from the texture map of the most similar reference frame. These feature points and descriptors are then matched with the feature points and descriptors extracted from the texture map of the target to be tested to obtain feature matching point pairs. False feature matching point pairs are removed by RANSAC, and the remaining feature matching point pairs are decomposed by SVD to obtain the relative pose and thus the pose estimation result of the target to be tested.

[0013] Optionally, the step of selecting point cloud data of multiple offline reference frames with the highest similarity to the target under test based on the initial similarity sort, determining the Chamfer distance between the point cloud data of the offline reference frames and the segmented point cloud, and taking the reference frame with the smallest Chamfer distance as the most similar reference frame, specifically includes:

[0014] Based on the initial similarity ranking, offline reference point clouds of the target to be tested are selected. The Chamfer distance between the top 10 offline reference point clouds with the highest similarity to the target to be tested and the target point cloud is obtained using the following formula:

[0015]

[0016] Where |A| and |B| are the number of points in point cloud A and B respectively, and ||ab|| and ||ba|| are the Euclidean distances between points a and b;

[0017] Then, find the index of the minimum Chamfer distance for all Chamfer distances, and obtain the index of the reference frame that is most similar to the texture to be inspected.

[0018] Optionally, the step of segmenting the RGB texture map to determine the texture map of the target to be tested and the corresponding segmentation point cloud specifically includes:

[0019] A pre-trained basic visual FastSAM model for image segmentation is used to segment the RGB texture map, obtain the segmentation mask of multiple test targets, and obtain the texture map and corresponding segmentation point cloud of the test targets based on the segmentation mask.

[0020] Optionally, determining the initial similarity ranking between each texture map in the offline reference frame and the scaled texture map of the target under test specifically includes:

[0021] Obtain the VGG feature vector of the scaled texture map of the target under test, and obtain its cosine similarity with the VGG feature vector of each texture map in the offline reference frame.

[0022] The cosine similarity is aggregated, and the aggregated cosine similarity is compared with a preset value to determine the initial order of similarity between each texture map in the offline reference frame and the scaled texture map of the target to be tested.

[0023] Optionally, the feature points and descriptors can be extracted using a pre-trained OmniGlue model.

[0024] This invention provides a training-free multi-target 6D pose estimation system, comprising:

[0025] The acquisition module is used to acquire the RGB texture map of the target under test through the RGBD stream, segment the RGB texture map, and determine the texture map of the target under test and the corresponding segmentation point cloud;

[0026] The scaling module is used to scale the texture map of the target under test to obtain a scaled texture map of the target under test.

[0027] The sorting module is used to acquire texture maps from multiple offline reference frames and determine the initial sorting of similarity between each texture map in the offline reference frames and the scaled texture map of the target under test.

[0028] The point cloud data acquisition module is used to acquire point cloud data from multiple offline reference frames of the target under test;

[0029] The most similar reference frame determination module is used to select the point cloud data of multiple offline reference frames with the highest similarity to the target under test based on the initial similarity sorting, and to determine the Chamfer distance between the point cloud data of the offline reference frames and the segmented point cloud. The reference frame with the smallest Chamfer distance is taken as the most similar reference frame.

[0030] The pose estimation module is used to extract feature points and descriptors from the texture map of the most similar reference frame, match them with the feature points and descriptors extracted from the texture map of the target to be tested to obtain feature matching point pairs, remove false feature matching point pairs through RANSAC, and decompose the remaining feature matching point pairs through SVD to obtain the relative pose and obtain the pose estimation result of the target to be tested.

[0031] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described training-free multi-target 6D pose estimation method.

[0032] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-described training-free multi-target 6D pose estimation method.

[0033] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a training-free multi-target 6D pose estimation method. This method achieves training-free target detection by using a pre-trained visual base model for texture segmentation. By initially sorting the similarity between each texture map in the offline reference frame and the scaled texture map of the target to be tested, texture similarity and point cloud similarity are combined. This allows for the extraction of the reference frame that best matches the index image from the target to be tested reference dataset in one go, thereby enabling direct calculation of feature matching. The indexed RGBD target pose is calculated by using feature matching point pairs and RANSAC, avoiding the cumbersome feature point matching calculation caused by traversal polling, while ensuring image consistency and the accuracy of target pose estimation. Attached Figure Description

[0034] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0035] Figure 1 A flowchart illustrating a training-free multi-target 6D pose estimation method provided in an embodiment of the present invention;

[0036] Figure 2 A flowchart for offline creation of a reference RGB, depth map, and pose set using an RGBD stream or CAD model, provided in an embodiment of the present invention;

[0037] Figure 3 A schematic diagram illustrating the offline data creation process based on CAD provided in an embodiment of the present invention;

[0038] Figure 4 This is a flowchart of the real-time pose calculation process provided in an embodiment of the present invention;

[0039] Figure 5 This is a schematic diagram of the texture map feature point matching results provided in an embodiment of the present invention;

[0040] Figure 6 This is a schematic diagram of the pose calculation results provided in an embodiment of the present invention;

[0041] Figure 7 A schematic diagram of a computer device for a training-free multi-target 6D pose estimation method provided in an embodiment of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0043] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0044] Figure 1 This is a flowchart of a training-free multi-target 6D pose estimation method provided by an embodiment of the present invention, as shown below. Figure 1 As shown in the figure, this embodiment illustrates a training-free multi-target 6D pose estimation method, including:

[0045] S1: Obtain the RGB texture map of the target under test through the RGBD stream, segment the RGB texture map, and determine the texture map of the target under test and the corresponding segmentation point cloud.

[0046] Optionally, the RGB texture map is segmented to determine the texture map and corresponding segmentation point cloud of the target to be tested. Specifically, this includes: using a pre-trained basic vision FastSAM model for image segmentation to segment the RGB texture map, obtaining segmentation masks for multiple targets to be tested, and obtaining the texture map and corresponding segmentation point cloud of the target to be tested based on the segmentation masks.

[0047] S2: Scale the texture map of the target to be tested to obtain a scaled texture map of the target to be tested.

[0048] For example, spatial alignment of the segmented point cloud with the candidate target texture map can be achieved by projecting the point cloud onto the texture using a 2D transformation matrix. Figure 1 In the coordinate system, the scaled texture map of the target under test is obtained.

[0049] S3: Obtain texture maps from multiple offline reference frames and determine the initial similarity ranking between each texture map in the offline reference frames and the scaled texture map of the target under test.

[0050] For example, the texture image of the target to be inspected is scaled to 384*384 to obtain the scaled texture image and the instantaneous 2D transformation matrix. Then, a 128-dimensional VGG feature vector is calculated on the scaled texture image. Based on the obtained VGG feature vector and all offline calculated VGG feature vectors of the target to be inspected, the cosine similarity is calculated using the following formula:

[0051]

[0052] Where u·v is the dot product (inner product) of vectors u and v; u and v are the feature vectors output by the VGG model. In this application, the feature vector dimension can be 128, that is, 128 floating-point numbers, which are calculated by the deep neural network model VGG. The real-time acquired images need to calculate the VGG feature vectors, while the offline calculated VGG feature vectors of the target to be inspected are directly read from memory.

[0053] The cosine similarity is aggregated according to the target category. The highest category with an aggregated similarity value of more than 0.6 is taken as the detected target, and the scaled texture map and the corresponding original segmentation point cloud are saved. If the aggregated similarity value is lower than the threshold of 0.6, the candidate texture map is removed. The VGG feature cosine similarity of all reference frames of the detected target is sorted from high to low.

[0054] S4: Acquire point cloud data from multiple offline reference frames of the target under test.

[0055] S5: Based on the initial similarity ranking, select the point cloud data of multiple offline reference frames with the highest similarity to the target to be tested, and determine the Chamfer distance between the point cloud data of the offline reference frames and the segmented point cloud. The reference frame with the smallest Chamfer distance is taken as the most similar reference frame.

[0056] For example, the offline reference point cloud of the target is extracted based on the initial similarity ranking. Then, the Chamfer distance between the first 10 initially ranked offline reference point cloud clouds and the target point cloud is calculated using the following formula:

[0057]

[0058] Where |A| and |B| are the number of points in point cloud A and B respectively, and ||ab|| and ||ba|| are the Euclidean distances between points a and b;

[0059] Then, find the index of the minimum Chamfer distance among all the reference frames, which is the index of the reference frame that is most similar to the texture to be inspected.

[0060] S6: Extract feature points and descriptors from the texture map of the most similar reference frame, match them with the feature points and descriptors extracted from the texture map of the target to be tested to obtain feature matching point pairs, remove false feature matching point pairs by RANSAC, and decompose the remaining feature matching point pairs by SVD to obtain the relative pose and obtain the pose estimation result of the target to be tested.

[0061] For example, the OmniGlue model can be used to extract feature points and descriptors from the texture map of the most similar reference frame and the segmented point cloud (or corresponding texture map) of the current target. The feature points of the target are then matched with those of the most similar reference frame to obtain feature point matching pairs. The RANSAC (Random Sample Consensus) algorithm is used to remove incorrect matches from these pairs, and SVD (Singular Value Decomposition) is used to calculate the relative pose of the target with respect to the most similar reference frame. Combining this with the known pose of the most similar reference frame, the final 6D pose of the target in the global coordinate system can be determined.

[0062] For example, the multi-target 6D pose of the object under test can be derived from keyframes of CAD model rendering or RGBD streams, covering cases with and without CAD models, such as... Figure 2 As shown.

[0063] For example, when a CAD model is available, the reference dataset is constructed as follows:

[0064] 1. Uniformly sample 1000 points in a 3D spherical space with a radius of 2π, and use these points as rotation vectors for later use;

[0065] 2. In Pyrender, set a fixed camera pose and target background, import a textured target CAD model, and fix its center position in the camera's field of view;

[0066] 3. Use the rotation vector obtained in step 1 as the rotation orientation of the CAD model and perform a rotation operation around its center;

[0067] 4. Perform Pyrender rendering to generate the texture map, depth map, and pose of the target CAD relative to the camera in the current pose.

[0068] 5. Extract the minimum bounding box of the target object in the texture map, scale the effective foreground texture within the minimum bounding box to a fixed size of 384*384, and obtain the 2D transformation matrix of the image.

[0069] 6. Input the scaled texture map into the OmniGlue model and extract the feature points and descriptors corresponding to the texture map;

[0070] 7. Calculate the VGG feature vector of the scaled texture map. The feature vector length is 128.

[0071] 8. Repeat steps 3 to 7 until all poses corresponding to the 1000 rotation vectors are rendered and the above feature quantities are calculated;

[0072] 9. Save the scaled texture map, 2D transformation matrix, depth map, feature points and descriptors, VGG feature vectors and target pose obtained in the above steps as a reference dataset.

[0073] When there is no CAD model, a reference dataset can be constructed in the following way:

[0074] 1. Use a depth camera to acquire RGB video streams of real target objects and corresponding depth maps, and use the VOS video segmentation model Cutie to annotate the region masks of all target objects in the RGB video stream;

[0075] 2. Extract key RGBD frames and optimized pose information from the RGBD stream using the BundleTrack algorithm.

[0076] 3. For the keyframes obtained in step 2, extract the minimum bounding box of the target object in the keyframe texture map based on the target region mask, scale the effective foreground texture within the minimum bounding box to a fixed size of 384*384, and obtain the 2D transformation matrix of the image.

[0077] 4. Input the scaled texture map into the OmniGlue model and extract the feature points and descriptors corresponding to the texture map;

[0078] 5. Calculate the VGG feature vector of the scaled effective texture map. The feature vector length is 128.

[0079] 6. Repeat steps 3 to 6 until all keyframe texture map scaling and feature calculation operations are completed;

[0080] 7. Save the scaled texture map, 2D transformation matrix, depth map, feature points and descriptors, VGG feature vectors, and target pose as a reference dataset.

[0081] like Figure 3 As shown, in the process of establishing offline data based on CAD, random sampling points in the spherical space are converted into a rotation matrix. This matrix transforms the CAD model to change its pose, and the required offline reference frame is obtained through software rendering.

[0082] For example, target detection can be performed in the following manner:

[0083] 1. Drive the RGBD camera to acquire a real-time data stream containing RGB texture and depth;

[0084] 2. Obtain candidate target segmentation masks by segmenting RGB texture maps using the FastSAM model, and obtain candidate target texture maps and corresponding segmentation point clouds based on the segmentation masks;

[0085] 3. Scale the texture image of the target to be inspected to 384*384 to obtain the scaled texture image of the target to be inspected and the instantaneous 2D transformation matrix;

[0086] 4. Calculate the 128-dimensional VGG feature vector for the scaled texture map of the target object to be inspected;

[0087] 5. Calculate the cosine similarity between the VGG feature vector obtained in step 4 and all the offline calculated VGG feature vectors of the target to be detected;

[0088]

[0089] Here, u·v is the dot product (inner product) of vectors u and v.

[0090] 6. Aggregate the cosine similarity by target category. The highest category with an aggregation similarity value greater than 0.6 is taken as the detected target. Save the scaled texture map and the corresponding original segmentation point cloud. If the aggregation similarity value is lower than the threshold of 0.6, the candidate texture map is removed.

[0091] 7. Sort the VGG feature cosine similarity of all reference frames of the detected target from high to low;

[0092] 8. Repeat steps 4 to 7 until all candidate target texture maps have been detected, and obtain the final valid target detection result. This result includes the scaled texture map and original segmentation point cloud corresponding to the target, as well as the initial sorting of the similarity between the offline reference frame and the scaled texture map of the corresponding target.

[0093] like Figure 4 As shown in the flowchart, the real-time pose calculation process is divided into two parts: target detection, which determines unseen targets through similarity, and 3D-to-3D spatial transformation, which calculates matching point pairs. Figure 5 The image shows the results of texture map feature point matching. Figure 6 This is the result of the pose calculation.

[0094] The above are one or more embodiments of the training-free multi-target 6D pose estimation method provided in this specification. Based on the same idea, this specification also provides a corresponding training-free multi-target 6D pose estimation system, including:

[0095] The acquisition module is used to acquire the RGB texture map of the target under test through the RGBD stream, segment the RGB texture map, and determine the texture map of the target under test and the corresponding segmentation point cloud;

[0096] The scaling module is used to scale the texture map of the target under test to obtain a scaled texture map of the target under test.

[0097] The sorting module is used to acquire texture maps from multiple offline reference frames and determine the initial sorting of similarity between each texture map in the offline reference frames and the scaled texture map of the target under test.

[0098] The point cloud data acquisition module is used to acquire point cloud data from multiple offline reference frames of the target under test;

[0099] The most similar reference frame determination module is used to select the point cloud data of multiple offline reference frames with the highest similarity to the target under test based on the initial similarity sorting, and to determine the Chamfer distance between the point cloud data of the offline reference frames and the segmented point cloud. The reference frame with the smallest Chamfer distance is taken as the most similar reference frame.

[0100] The pose estimation module is used to extract feature points and descriptors from the texture map of the most similar reference frame, match them with the feature points and descriptors extracted from the texture map of the target to be tested to obtain feature matching point pairs, remove false feature matching point pairs through RANSAC, and decompose the remaining feature matching point pairs through SVD to obtain the relative pose and obtain the pose estimation result of the target to be tested.

[0101] Specific limitations regarding the training-free multi-target 6D pose estimation system can be found in the limitations of the training-free multi-target 6D pose estimation method described above, and will not be repeated here. Each module in the aforementioned training-free multi-target 6D pose estimation system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0102] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the training-free multi-target 6D pose estimation method provided above.

[0103] The present invention also provides Figure 7 The schematic diagram of the computer device shown is as follows: Figure 7 As shown, at the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the training-free multi-target 6D pose estimation method provided in the above embodiments.

[0104] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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 invention.

Claims

1. A training-free multi-target 6D pose estimation method, characterized in that, include: The RGB texture map of the target under test is obtained by RGBD stream, the RGB texture map is segmented, and the texture map of the target under test and the corresponding segmentation point cloud are determined. The texture map of the target under test is scaled to obtain a scaled texture map of the target under test; Acquire texture maps from multiple offline reference frames and determine the initial similarity ranking between each texture map in the offline reference frames and the scaled texture map of the target under test; Acquire point cloud data from multiple offline reference frames of the target under test; Based on the initial similarity ranking, the point cloud data of multiple offline reference frames with the highest similarity to the target to be tested are selected, and the Chamfer distance between the point cloud data of the offline reference frames and the segmented point cloud is determined. The reference frame with the smallest Chamfer distance is taken as the most similar reference frame. Feature points and descriptors are extracted from the texture map of the most similar reference frame. These feature points and descriptors are then matched with the feature points and descriptors extracted from the texture map of the target to be tested to obtain feature matching point pairs. False feature matching point pairs are removed by RANSAC, and the remaining feature matching point pairs are decomposed by SVD to obtain the relative pose and thus the pose estimation result of the target to be tested.

2. The training-free multi-target 6D pose estimation method as described in claim 1, characterized in that, The process involves selecting point cloud data from multiple offline reference frames that have the highest similarity to the target under test based on initial similarity ranking, determining the Chamfer distance between the point cloud data of the offline reference frames and the segmented point cloud, and selecting the reference frame with the smallest Chamfer distance as the most similar reference frame. Specifically, this includes: Based on the initial similarity ranking, offline reference point clouds of the target to be tested are selected. The Chamfer distance between the top 10 offline reference point clouds with the highest similarity to the target to be tested and the target point cloud is obtained using the following formula: Where |A| and |B| are the number of points in point cloud A and B respectively, and ||ab|| and ||ba|| are the Euclidean distances between points a and b; Then, find the index of the minimum Chamfer distance for all Chamfer distances, and obtain the index of the reference frame that is most similar to the texture to be inspected.

3. The training-free multi-target 6D pose estimation method as described in claim 1, characterized in that, The segmentation of the RGB texture map to determine the texture map of the target object and the corresponding segmentation point cloud specifically includes: A pre-trained basic visual FastSAM model for image segmentation is used to segment the RGB texture map, obtain the segmentation mask of multiple test targets, and obtain the texture map and corresponding segmentation point cloud of the test targets based on the segmentation mask.

4. The training-free multi-target 6D pose estimation method as described in claim 1, characterized in that, The initial similarity ranking between each texture map in the offline reference frame and the scaled texture map of the target under test specifically includes: Obtain the VGG feature vector of the scaled texture map of the target under test, and obtain its cosine similarity with the VGG feature vector of each texture map in the offline reference frame. The cosine similarity is aggregated, and the aggregated cosine similarity is compared with a preset value to determine the initial order of similarity between each texture map in the offline reference frame and the scaled texture map of the target to be tested.

5. The training-free multi-target 6D pose estimation method as described in claim 1, characterized in that, The feature points and descriptors are extracted using a pre-trained OmniGlue model.

6. A training-free multi-target 6D pose estimation system, characterized in that, include: The acquisition module is used to acquire the RGB texture map of the target under test through the RGBD stream, segment the RGB texture map, and determine the texture map of the target under test and the corresponding segmentation point cloud; The scaling module is used to scale the texture map of the target under test to obtain a scaled texture map of the target under test. The sorting module is used to acquire texture maps from multiple offline reference frames and determine the initial sorting of similarity between each texture map in the offline reference frames and the scaled texture map of the target under test. The point cloud data acquisition module is used to acquire point cloud data from multiple offline reference frames of the target under test; The most similar reference frame determination module is used to select the point cloud data of multiple offline reference frames with the highest similarity to the target under test based on the initial similarity sorting, and to determine the Chamfer distance between the point cloud data of the offline reference frames and the segmented point cloud. The reference frame with the smallest Chamfer distance is taken as the most similar reference frame. The pose estimation module is used to extract feature points and descriptors from the texture map of the most similar reference frame, match them with the feature points and descriptors extracted from the texture map of the target to be tested to obtain feature matching point pairs, remove false feature matching point pairs through RANSAC, and decompose the remaining feature matching point pairs through SVD to obtain the relative pose and obtain the pose estimation result of the target to be tested.

7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the training-free multi-target 6D pose estimation method according to any one of claims 1 to 5.

8. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the training-free multi-target 6D pose estimation method according to any one of claims 1 to 5.

Images