Industrial robot posture recognition method and device, and storage medium

Abstract

The application provides an industrial robot posture recognition method and device and a storage medium, the method is: importing an industrial robot video, extracting and adjusting an industrial robot 2D image from the industrial robot video; constructing a teacher model based on a human body posture estimation model DensePose, training the teacher model through the adjusted industrial robot 2D image, and outputting industrial robot 2D posture information through the trained teacher model; constructing a total FDPD distillation loss function based on a mean square error function MSE; constructing a student model based on the teacher model, training the student model through the total FDPD distillation loss function and the industrial robot 2D posture information, and recognizing the industrial robot posture in the industrial robot 2D image through the trained student model. The application realizes accurate and efficient industrial robot posture recognition, is suitable for abnormal detection of the industrial robot, does not need a large network, and improves the efficiency of recognizing the posture of the industrial robot.

Description

Technical Field

[0001] This invention relates to the field of robot posture recognition technology, specifically to an industrial robot posture recognition method, device, and storage medium. Background Technology

[0002] Because robots are highly flexible and have a large range of motion, and some models can move at high speeds, bear heavy loads, and have high rigidity under specific working conditions, they are very likely to cause serious safety accidents when sudden abnormal situations occur. For example, abnormal behavior of industrial robots may lead to robot damage, crushing, etc., which may threaten the personal safety of people in the work area to a certain extent. Therefore, the safe behavior of industrial robots is of utmost importance.

[0003] Research on industrial robot safety based on deep learning primarily focuses on pose estimation from single images, with most aiming to estimate human pose during human-robot interactions. Within the field of industrial robot pose estimation, 6D pose estimation of target objects in pick-and-place tasks is a major research focus. Traditional methods use vision-based approaches to detect collisions between industrial robots and obstacles, assuming the robot's pose is known. Since robot information is available from the controller and hand-eye calibration is performed, these studies focus on obstacle trajectory prediction. In recent research, many researchers have begun using deep learning methods to estimate the pose of industrial robots themselves. However, unlike human pose estimation tasks, deep learning-based industrial robot pose estimation requires the use of synthetic data or the creation of custom data corresponding to the task. Limited by insufficient datasets and the lack of standards, achievements in this area are still relatively few. At the 12th International Conference on Intelligent Robots and Applications (ICIRA) in 2019, F. Zhou et al. proposed a real-time deep learning-based method that directly estimates the 3D pose of a robotic arm using color images. This method aims to eliminate hand-eye calibration when the system needs reconfiguration and increases system flexibility by eliminating the requirement for the camera to be fixed relative to the robot, ultimately achieving good recognition accuracy. At the 14th IEEE International Conference on Automation Science and Engineering (CASE) in 2018, Thomas et al. proposed RoPose, a convolutional network-based method for industrial robot posture recognition. The core contribution of this method is the implementation of a RoPose annotation pipeline that automatically generates simulation datasets based on the robot system, utilizing the MoveIt! motion planning framework and the Gazebo integrated simulation suite. While this method addresses the current lack of open-source standard datasets for industrial robots, its simulation environment and specific motion planning requirements make its practical application challenging. Furthermore, both methods, in their pursuit of improved recognition accuracy, utilize excessive framework model suites, resulting in overly large networks and consequently, lower efficiency in posture recognition. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an industrial robot posture recognition method, device and storage medium to address the shortcomings of the prior art.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: An industrial robot posture recognition method, comprising the following steps:

[0006] Import industrial robot videos, extract multiple 2D images of industrial robots from the videos, organize the multiple 2D images of industrial robots, and obtain an industrial robot 2D image set from the organized multiple 2D images of industrial robots.

[0007] A teacher model is constructed based on the human pose estimation model DensePose. The teacher model is trained using a set of 2D images of industrial robots. The trained teacher model outputs the 2D pose information of the industrial robots.

[0008] Construct the overall FDPD distillation loss function based on the mean square error function (MSE);

[0009] A student model is constructed based on the trained teacher model. The student model is trained using the overall FDPD distillation loss function and the 2D pose information of the industrial robot. The trained student model is used to identify the pose of the industrial robot in each 2D image in the industrial robot 2D image set.

[0010] The beneficial effects of this invention are: by using a trained teacher model and constructing a student model based on the overall FDPD distillation loss function, a model for recognizing the posture of industrial robots is established by combining the posture distillation strategy with the existing human posture estimation model DensePose. This achieves accurate and efficient industrial robot posture recognition, lays the foundation for the safety monitoring of industrial robots, is suitable for anomaly detection of industrial robots, does not require an excessively large network, and improves the efficiency of industrial robot posture recognition.

[0011] Based on the above technical solution, the present invention can be further improved as follows.

[0012] Furthermore, the extraction of multiple 2D images of the industrial robot from the industrial robot video and the processing of these multiple 2D images specifically involves:

[0013] Multiple 2D images of the industrial robot are extracted from the video using a video segmentation keyframe method, and the 2D images of the industrial robot that do not meet the requirements are then removed from the multiple 2D images of the industrial robot.

[0014] Furthermore, the step of training the teacher model using the adjusted 2D image of the industrial robot specifically involves:

[0015] The process involves locating multiple joints of the industrial robot in a 2D image, determining the regions corresponding to each joint, obtaining the dense points and coordinates of each region based on the located joints, calculating the median of the coordinates of the dense points in each region, connecting the median points and outputting the results to obtain the trained teacher model.

[0016] Furthermore, the process of locating multiple joints of the industrial robot in the 2D image of the industrial robot, determining the regions corresponding to each joint, and obtaining the dense points and coordinates of the dense points in each region based on the located joints in each region, specifically involves:

[0017] The 2D image of the industrial robot is divided into n regions corresponding to each joint, based on the number n joints of the industrial robot. A two-dimensional coordinate system is established to parameterize each region, so as to obtain the coordinates of each joint of the industrial robot at any position in each region.

[0018] The region-based dense regression method obtains the Region of Interest (ROI) from the joints after all regions are located. The ROIs are concatenated, and the concatenated ROIs are used to generate features through ROI pooling to obtain the dense points and coordinates of the dense points of the ROIs.

[0019] Furthermore, the calculation of the median coordinates of dense points in each region, and the connection and output of the median points, specifically involves:

[0020] The coordinates of dense points within each Region of Interest (ROI) are traversed. The median of the leftmost and rightmost X-coordinates within each ROI is calculated to obtain the X-coordinate of the median point. Similarly, the median of the topmost and bottommost Y-coordinates within each ROI is calculated to obtain the Y-coordinate of the median point. This process yields n median points for each ROI. The median points are then connected and output to complete the training of the teacher model.

[0021] Furthermore, the construction of the overall FDPD distillation loss function based on the mean square error function (MSE) is specifically as follows:

[0022] For each joint, a joint label is assigned, and the Gaussian confidence score m corresponding to the k-th joint label is... k for:

[0023]

[0024] Where (x, y) represents the pixel position of the joint point, σ represents the pre-fixed spatial variance, and x k and y k These represent the coordinates of the median point;

[0025] Using Gaussian confidence m k The mean squared error function (MSE) is constructed by combining the prediction confidence plot of the k-th joint with the mean squared error function (MSE).

[0026]

[0027] in, This represents the prediction confidence plot for the k-th joint;

[0028] Based on the mean squared error function (MSE), a distillation loss function is constructed using the backpropagation learning method. The distillation loss function is as follows:

[0029]

[0030] in, and These represent the confidence maps of the k-th joint predicted by the teacher model and the student model being trained, respectively.

[0031] The mean squared error function (MSE) is constructed using the mean squared error function (MSE) and the distillation loss function. The overall FDPD loss function is as follows:

[0032] L fdpd =αL pd +(1-α)L mse ,

[0033] Where α represents the balance measure between the two losses estimated by cross-validation, and L fdpd L represents the overall FDPD loss function. mse Let MSE represent the mean square error function.

[0034] Furthermore, the training of the student model using the overall FDPD distillation loss function and the 2D pose information of the industrial robot specifically involves:

[0035] A backbone network is established between the trained teacher model and the student model;

[0036] The overall FDPD loss function and the 2D posture information of the industrial robot are input into the student model through the backbone network;

[0037] The student model is trained iteratively using the overall FDPD loss function and the 2D pose information of the industrial robot, thereby completing the training of the student model.

[0038] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: An industrial robot posture recognition device, comprising:

[0039] An image preprocessing module is used to import industrial robot videos and extract and adjust 2D images of the industrial robot from the videos.

[0040] The teacher model training module is used to construct a teacher model based on the human pose estimation model DensePose, train the teacher model using an adjusted 2D image of the industrial robot, and output the 2D pose information of the industrial robot through the trained teacher model.

[0041] The student model training module is used to construct the overall FDPD distillation loss function based on the mean squared error function (MSE).

[0042] A student model is constructed based on the teacher model. The student model is trained using the overall FDPD distillation loss function and the 2D pose information of the industrial robot. The trained student model is then used to identify the pose of the industrial robot in the 2D image of the industrial robot.

[0043] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: an industrial robot posture recognition device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the industrial robot posture recognition method described above is implemented.

[0044] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the industrial robot posture recognition method as described above. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the industrial robot posture recognition method provided in an embodiment of the present invention.

[0046] Figure 2 This is a functional block diagram of an industrial robot posture recognition device provided in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the DensePose model structure provided in an embodiment of the present invention. Detailed Implementation

[0048] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0049] Example 1:

[0050] like Figure 1-2 As shown, an industrial robot posture recognition method includes the following steps:

[0051] Import industrial robot videos, extract multiple 2D images of industrial robots from the videos, organize the multiple 2D images of industrial robots, and obtain an industrial robot 2D image set from the organized multiple 2D images of industrial robots.

[0052] A teacher model is constructed based on the human pose estimation model DensePose. The teacher model is trained using a set of 2D images of industrial robots. The trained teacher model outputs the 2D pose information of the industrial robots.

[0053] Construct the overall FDPD distillation loss function based on the mean square error function (MSE);

[0054] A student model is constructed based on the trained teacher model. The student model is trained using the overall FDPD distillation loss function and the 2D pose information of the industrial robot. The trained student model is used to identify the pose of the industrial robot in each 2D image in the industrial robot 2D image set.

[0055] Combination Figure 3 The DensePose model adopts an RCNN structure based on pyramid network features on the Caffe2 framework. It mainly uses a region feature aggregation method to obtain dense coordinates in each specific region. Through a multi-cascade structure and a region-based method, it decomposes the complex task into controllable modules.

[0056] In the above embodiments, a student model is constructed by training the teacher model and using the overall FDPD distillation loss function. That is, a model for recognizing the posture of industrial robots is established by combining the posture distillation strategy with the existing human posture estimation model DensePose. This achieves accurate and efficient industrial robot posture recognition, lays the foundation for the safety monitoring of industrial robots, is suitable for anomaly detection of industrial robots, does not require an excessively large network, and improves the efficiency of industrial robot posture recognition.

[0057] Based on Example 1, the extraction of multiple 2D images of the industrial robot from the industrial robot video and the processing of these multiple 2D images specifically involve:

[0058] Multiple 2D images of the industrial robot are extracted from the video using a video segmentation keyframe method, and the 2D images of the industrial robot that do not meet the requirements are then removed from the multiple 2D images of the industrial robot.

[0059] It should be understood that substandard 2D images of industrial robots refer to images where the robot is unclear or severely occluded. A standard resolution can be preset, and the clarity of the 2D image can be determined based on this preset resolution. Alternatively, existing image occlusion detection methods can be used to detect severely occluded images in the 2D image of the industrial robot.

[0060] In the above embodiments, unqualified 2D images of industrial robots can be screened out, thereby standardizing the 2D images of industrial robots and improving the accuracy of subsequent data processing.

[0061] Based on Example 1, the training of the teacher model using the adjusted 2D image of the industrial robot specifically involves:

[0062] The process involves locating multiple joints of the industrial robot in a 2D image, determining the regions corresponding to each joint, obtaining the dense points and coordinates of each region based on the located joints, calculating the median of the coordinates of the dense points in each region, connecting the median points and outputting the results to obtain the trained teacher model.

[0063] For example, if a six-axis industrial robot is used, we decompose it into six independent regions and parameterize each part using a local two-dimensional coordinate system that can be used to represent any position within the region around the joints of the industrial robot.

[0064] Based on Example 1, the process of locating multiple joints of the industrial robot in the 2D image of the industrial robot, determining the regions corresponding to each joint, and obtaining the dense points and coordinates of the dense points in each region based on the located joints in each region are as follows:

[0065] The 2D image of the industrial robot is divided into n regions corresponding to each joint, based on the number n joints of the industrial robot. A two-dimensional coordinate system is established to parameterize each region, so as to obtain the coordinates of each joint of the industrial robot at any position in each region.

[0066] The region-based dense regression method obtains the Region of Interest (ROI) from the joints after all regions are located. The ROIs are concatenated, and the concatenated ROIs are used to generate features through ROI pooling to obtain the dense points and coordinates of the dense points of the ROIs.

[0067] Based on Example 1, the calculation of the median points of the coordinates of dense points in each region, and the output of the connecting lines of each median point, specifically involves:

[0068] The coordinates of dense points within each Region of Interest (ROI) are traversed. The median of the leftmost and rightmost X-coordinates within each ROI is calculated to obtain the X-coordinate of the median point. Similarly, the median of the topmost and bottommost Y-coordinates within each ROI is calculated to obtain the Y-coordinate of the median point. This process yields n median points for each ROI. The median points are then connected and output to complete the training of the teacher model.

[0069] While using FCNs can simplify training, performing multiple tasks—part segmentation and pixel localization—within the same deep network while maintaining scale invariance presents a challenge for datasets. Therefore, referencing the Mask-R-CNN model, a region-based approach is employed, proposing a cascade of Regions of Interest (ROIs). ROI pooling inputs generated features into branches specific to these regions. This architecture decomposes the task's complexity into controllable modules and implements a region selection mechanism through ROI pooling.

[0070] To train the model in a supervised manner, the dataset was traversed during the training process. There are 6 images, each labeled with 6 joints, defined in image space as follows:

[0071] Where N represents the total amount of data in the dataset, and K represents 6.

[0072] After obtaining the coordinates of dense points within the region, first convert their UV coordinates into two-dimensional XY coordinates. Then, by traversing the coordinates of dense points within the irregular region, calculate the median of the leftmost and rightmost X coordinates as the X coordinate of the median point, and then take the median of the topmost and bottommost Y coordinates as the Y coordinate of the median point, thereby obtaining the six median points of the median of the six regions, and then connect them for output.

[0073] Based on Example 1, the output of the teacher model is used as the supervision information for training the student model. The supervision is mainly carried out through a modified loss function, where knowledge distillation from the classroom model to the target model is performed in each mini-batch iteration throughout the training process.

[0074] The key to knowledge distillation is designing a suitable simulation loss function that can effectively extract teacher knowledge and transfer it to the training of the student model. The distillation function in pose distillation, based on the softmax cross-entropy loss for object classification, is not suitable for transferring structured pose knowledge in a 2D image space.

[0075] To address the aforementioned issues, a loss function based on mean squared error (MSE) is used as the target loss function due to its greater compatibility. To represent the joint ground truth label, a Gaussian kernel is used around the labeled location Z. K =(x K y K The center is centered, thus generating a confidence graph for each individual joint label k (k∈1,...,6).

[0076] Based on Example 1, the construction of the overall FDPD distillation loss function based on the mean square error function (MSE) is specifically as follows:

[0077] For each joint, a joint label is assigned, and the Gaussian confidence score m corresponding to the k-th joint label is... k for:

[0078]

[0079] Where (x, y) represents the pixel position of the joint point, σ represents the pre-fixed spatial variance, and x k and y k These represent the coordinates of the median point;

[0080] Using Gaussian confidence m k The mean squared error function (MSE) is constructed by combining the prediction confidence plot of the k-th joint with the mean squared error function (MSE).

[0081]

[0082] in, This represents the prediction confidence plot for the k-th joint;

[0083] Based on the mean squared error function (MSE), a distillation loss function is constructed using the backpropagation learning method. The distillation loss function is as follows:

[0084]

[0085] in, and These represent the confidence maps of the k-th joint predicted by the teacher model and the student model being trained, respectively.

[0086] For example, if k is 6, there are dense point coordinates in 6 regions of interest. In order to calculate the error, it is necessary to calculate the Gaussian confidence of each point. The Gaussian confidence of each point is obtained by calculating the Gaussian kernel function of each point and its median point.

[0087] The MSE function was chosen as the distillation loss function to measure the difference between the student and teacher models, in order to maximize the comparability of the model with supervised learning loss.

[0088] The mean squared error function (MSE) is constructed using the mean squared error function (MSE) and the distillation loss function. The overall FDPD loss function is as follows:

[0089] L fdpd =αL pd +(1-α)L mse ,

[0090] Where α represents the balance measure between the two losses estimated by cross-validation, and L fdpd L represents the overall FDPD loss function. mse Let MSE represent the mean square error function.

[0091] Therefore, target network learning can both predict the labeled ground truth annotations of training samples and match the prediction structure of stronger teacher models, thereby enhancing the applicability of the method.

[0092] Based on Example 1, the training of the student model using the overall FDPD distillation loss function and the 2D posture information of the industrial robot specifically involves:

[0093] A backbone network is established between the trained teacher model and the student model;

[0094] The overall FDPD loss function and the 2D posture information of the industrial robot are input into the student model through the backbone network;

[0095] The student model is trained iteratively using the overall FDPD loss function and the 2D pose information of the industrial robot, thereby completing the training of the student model.

[0096] By inputting the 2D posture information of a single industrial robot into a high-precision posture estimation network model, namely the student model, and compressing the model by using a distillation method of fitting loss function, accurate and efficient industrial robot posture recognition is achieved, laying the foundation for the safety monitoring of industrial robots.

[0097] Example 2:

[0098] like Figure 2 As shown, an industrial robot posture recognition device includes:

[0099] The image preprocessing module is used to import industrial robot videos, extract multiple 2D images of industrial robots from the videos, and organize the multiple 2D images of industrial robots to obtain a set of 2D images of industrial robots.

[0100] The teacher model training module is used to construct a teacher model based on the human pose estimation model DensePose, train the teacher model using a set of 2D images of industrial robots, and output the 2D pose information of industrial robots through the trained teacher model.

[0101] The student model training module is used to construct the overall FDPD distillation loss function based on the mean squared error function (MSE).

[0102] A student model is constructed based on the trained teacher model. The student model is trained using the overall FDPD distillation loss function and the 2D pose information of the industrial robot. The trained student model is used to identify the pose of the industrial robot in each 2D image in the industrial robot 2D image set.

[0103] Example 3:

[0104] An industrial robot posture recognition device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The device is characterized in that when the processor executes the computer program, it implements the industrial robot posture recognition method as described above.

[0105] Example 4:

[0106] A computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the above-described industrial robot posture recognition method.

[0107] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0108] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for industrial robot posture recognition, characterized in that, Includes the following steps: Import industrial robot videos, extract multiple 2D images of industrial robots from the videos, organize the multiple 2D images of industrial robots, and obtain an industrial robot 2D image set from the organized multiple 2D images of industrial robots. A teacher model is constructed based on the human pose estimation model DensePose. The teacher model is trained using a set of 2D images of industrial robots. The trained teacher model outputs the 2D pose information of the industrial robots. Construct the overall FDPD distillation loss function based on the mean square error function (MSE); A student model is constructed based on the trained teacher model. The student model is trained using the overall FDPD distillation loss function and the 2D pose information of the industrial robot. The trained student model is used to identify the pose of the industrial robot in each 2D image in the industrial robot 2D image set. The process of training the teacher model using the adjusted 2D image of the industrial robot specifically involves: The system locates multiple joints of the industrial robot in the 2D image of the industrial robot, determines the regions corresponding to each joint based on the multiple joints, obtains the dense points and coordinates of the dense points in each region based on the located joints in each region, calculates the median of the coordinates of the dense points in each region, connects the median points and outputs them, thereby obtaining the trained teacher model. The process involves locating multiple joints of the industrial robot in the 2D image, determining the corresponding regions for each joint, and obtaining the dense points and coordinates of each region based on the located joints within each region. The 2D image of the industrial robot is divided into n regions corresponding to each joint, based on the number n joints of the industrial robot. A two-dimensional coordinate system is established to parameterize each region, so as to obtain the coordinates of each joint of the industrial robot at any position in each region. The region-based dense regression method obtains the Region of Interest (ROI) from the joints after all regions are located. The ROIs are concatenated, and the concatenated ROIs are used to generate features through ROI pooling to obtain the dense points and coordinates of the dense points of the ROIs.

2. The industrial robot posture recognition method according to claim 1, characterized in that, The process of extracting multiple 2D images of industrial robots from the industrial robot video and organizing these multiple 2D images specifically involves: Multiple 2D images of the industrial robot are extracted from the video using a video segmentation keyframe method, and the 2D images of the industrial robot that do not meet the requirements are then removed from the multiple 2D images of the industrial robot.

3. The industrial robot posture recognition method according to claim 1, characterized in that, The calculation of the median coordinates of dense points in each region, and the connection and output of the median points, specifically involves: The coordinates of dense points within each Region of Interest (ROI) are traversed. The median of the leftmost and rightmost X-coordinates within each ROI is calculated to obtain the X-coordinate of the median point. Similarly, the median of the topmost and bottommost Y-coordinates within each ROI is calculated to obtain the Y-coordinate of the median point. This process yields n median points for each ROI. The median points are then connected and output to complete the training of the teacher model.

4. The industrial robot posture recognition method according to claim 1, characterized in that, The overall FDPD distillation loss function constructed based on the mean square error function (MSE) is as follows: For each joint, assign a joint label, and the Gaussian confidence score corresponding to the k-th joint label is... for: ， Where (x, y) represents the pixel position of the joint. This represents a pre-fixed spatial variance. and These represent the coordinates of the median point; Gaussian confidence level The mean squared error function (MSE) is constructed by combining the prediction confidence plot of the k-th joint with the mean squared error function (MSE). ， in, This represents the prediction confidence plot for the k-th joint; Based on the mean squared error function (MSE), a distillation loss function is constructed using the backpropagation learning method. The distillation loss function is as follows: ， in, and These represent the confidence maps of the k-th joint predicted by the teacher model and the student model being trained, respectively. The overall FDPD distillation loss function is constructed using the mean square error function (MSE) and the distillation loss function. The overall FDPD distillation loss function is as follows: ， Where α represents the balance measure between the two losses estimated by cross-validation. This represents the overall FDPD distillation loss function. Let MSE represent the mean square error function.

5. The industrial robot posture recognition method according to claim 1, characterized in that, The process of training the student model using the overall FDPD distillation loss function and the 2D posture information of the industrial robot is as follows: A backbone network is established between the trained teacher model and the student model; The overall FDPD distillation loss function and the 2D posture information of the industrial robot are input into the student model through the backbone network; The student model is trained iteratively using the overall FDPD distillation loss function and the 2D posture information of the industrial robot, thereby completing the training of the student model.

6. An industrial robot posture recognition device, characterized in that, include: The image preprocessing module is used to import industrial robot videos, extract multiple 2D images of industrial robots from the videos, and organize the multiple 2D images of industrial robots to obtain a set of 2D images of industrial robots. The teacher model training module is used to construct a teacher model based on the human pose estimation model DensePose, train the teacher model using a set of 2D images of industrial robots, and output the 2D pose information of industrial robots through the trained teacher model. The student model training module is used to construct the overall FDPD distillation loss function based on the mean squared error function (MSE). A student model is constructed based on the trained teacher model. The student model is trained using the overall FDPD distillation loss function and the 2D pose information of the industrial robot. The trained student model is used to identify the pose of the industrial robot in each 2D image in the industrial robot 2D image set. The process of training the teacher model using the adjusted 2D image of the industrial robot specifically involves: The system locates multiple joints of the industrial robot in the 2D image of the industrial robot, determines the regions corresponding to each joint based on the multiple joints, obtains the dense points and coordinates of the dense points in each region based on the located joints in each region, calculates the median of the coordinates of the dense points in each region, connects the median points and outputs them, thereby obtaining the trained teacher model. The process involves locating multiple joints of the industrial robot in the 2D image, determining the corresponding regions for each joint, and obtaining the dense points and coordinates of each region based on the located joints within each region. The 2D image of the industrial robot is divided into n regions corresponding to each joint, based on the number n joints of the industrial robot. A two-dimensional coordinate system is established to parameterize each region, so as to obtain the coordinates of each joint of the industrial robot at any position in each region. The region-based dense regression method obtains the Region of Interest (ROI) from the joints after all regions are located. The ROIs are concatenated, and the concatenated ROIs are used to generate features through ROI pooling to obtain the dense points and coordinates of the dense points of the ROIs.

7. An industrial robot posture recognition device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the industrial robot posture recognition method as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the industrial robot posture recognition method as described in any one of claims 1 to 5.

Images