A human pose estimation method and system based on joint weight rotation
By using a joint weight rotation method to perform nonlinear mapping of the two-dimensional coordinates of human joints, the problems of accuracy and speed in 2D pose to 3D pose conversion are solved, and efficient 2D-3D human pose estimation is achieved.
Patent Information
- Application Number
- CN202210120713.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-02-09
- Publication Date
- 2025-11-11
- Estimated Expiration
- 2042-02-09
AI Technical Summary
Existing technologies struggle to efficiently and accurately map 2D human poses to 3D human poses, resulting in computers being unable to determine the corresponding 3D pose when understanding 2D poses.
A joint weight-based rotation method is adopted, which uses a rotation weight component and a rotation weight residual module to perform nonlinear mapping on the two-dimensional coordinates of human joints and calculates the three-dimensional coordinates of the joints.
It improves the accuracy and speed of 2D to 3D pose conversion, reduces regression error, and achieves simple and efficient 2D-3D human pose estimation.
Smart Images
Figure CN115223237B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of 3D human pose estimation technology, and in particular to a human pose estimation method and system based on joint weight rotation. Background Technology
[0002] 3D human pose estimation is a key technology for computers to understand human behavior. In recent years, 3D human pose estimation technology has been widely applied in fields such as animation, virtual reality, intelligent video analysis, and robotics. However, some problems still exist in computers' understanding of 2D poses. For example, a 2D pose may correspond to multiple 3D poses, and computers cannot understand from only the 2D pose which 3D pose corresponds to the real-world pose.
[0003] Therefore, how to efficiently handle the connections between 2D joints and simply and effectively map 2D human poses to 3D human poses has become an urgent problem to be solved. Summary of the Invention
[0004] The purpose of this invention is to provide a human posture estimation method and system based on joint weight rotation, which improves the accuracy and speed of two-dimensional posture to three-dimensional posture conversion.
[0005] To achieve the above objectives, the present invention provides the following solution:
[0006] A human pose estimation method based on joint weight rotation includes:
[0007] Obtain the two-dimensional coordinates of human joints;
[0008] The two-dimensional coordinates of the multiple joints are input into the first rotation weighting component to calculate the first joint features;
[0009] The first joint features are input into the rotation weight residual module to obtain the second joint features;
[0010] The second joint feature is input into the second rotation weight component to obtain the joint's three-dimensional coordinates;
[0011] The rotation weight residual module includes multiple rotation weight residual sub-modules, and each of the rotation weight residual sub-module, the first rotation weight component, and the second rotation weight component includes at least one rotation weight component.
[0012] Optionally, the process of determining the rotation weight component specifically includes:
[0013] Obtain multiple joint features to be updated from the dataset;
[0014] The multiple joint features to be updated are determined as a joint feature matrix;
[0015] Calculate the rotational joint feature matrix based on the joint feature matrix and the rotational identity matrix;
[0016] Weight sharing is performed on the rotational joint feature matrix to determine the calculation formula for the joint features after weight rotation;
[0017] Based on the calculation formula of the joint features after weight rotation, a rotation weight component is constructed.
[0018] Optionally, calculating the rotational joint feature matrix based on the joint feature matrix and the rotational identity matrix specifically includes:
[0019] According to formula H R =RH, calculate the feature matrix of the rotary joint;
[0020] Where R represents the rotational identity matrix and H represents the joint feature matrix. H R The feature matrix of the rotational joint. J represents the number of joint features to be updated. Let D represent the set of real number matrices, D represent the first feature dimension, h0 represent the joint feature to be updated for the 0th joint, h1 represent the joint feature to be updated for the 1st joint, and h J-1 This represents the joint features to be updated for the (J-1)th joint.
[0021] Optionally, the step of weight sharing on the rotational joint feature matrix to determine the calculation formula for the joint features after weight rotation specifically includes:
[0022] According to the formula Determine the joint features after weight rotation;
[0023] Where H′ represents the joint feature after weight rotation, HR represents the rotational joint feature matrix. Let D' represent the set of real matrix values, D′ represent the dimension of the second feature, and W represent the rotation-learned weights. J represents the number of joint features to be updated.
[0024] Optionally, the weighted rotating residual module includes a first rotating weighted residual submodule, a second rotating weighted residual submodule, and a third rotating weighted residual submodule;
[0025] The input terminal of the first rotation weight residual submodule is connected to the output terminal of the first rotation weight component, and the output terminal of the first rotation weight residual submodule is connected to the input terminal of the second rotation weight residual submodule.
[0026] The output of the second rotation weight residual submodule is connected to the input of the third rotation weight residual submodule, and the output of the third rotation weight residual submodule is connected to the input of the second rotation weight component.
[0027] Optionally, the first rotation weight residual submodule includes a third rotation weight component and a fourth rotation weight component;
[0028] The output of the first rotation weight component is connected to the input of the third rotation weight component, and the output of the third rotation weight component is connected to the input of the fourth rotation weight component.
[0029] The output of the first rotation weight component is also connected to the first output of the fourth rotation weight component in a jump connection and perform feature addition.
[0030] To achieve the above objectives, the present invention also provides the following technical solutions:
[0031] A human pose estimation system based on joint weight rotation, the human pose estimation system comprising:
[0032] The coordinate acquisition unit is used to acquire the two-dimensional coordinates of human joints;
[0033] The first feature calculation unit is used to input the two-dimensional coordinates of the multiple joints into the first rotation weight component to calculate the first joint features;
[0034] The second feature calculation unit is used to input the first joint feature into the rotation weight residual module to obtain the second joint feature;
[0035] The third feature calculation unit is used to input the second joint features into the second rotation weight component to obtain the three-dimensional coordinates of the joint;
[0036] The rotation weight residual module includes multiple rotation weight residual sub-modules, and each of the rotation weight residual sub-module, the first rotation weight component, and the second rotation weight component includes at least one rotation weight component.
[0037] Optionally, in determining the rotation weight component, the human pose estimation system further includes:
[0038] The dataset acquisition sub-unit is used to retrieve multiple joint features to be updated from the dataset.
[0039] The joint feature matrix determination subunit is used to determine multiple joint features to be updated as a joint feature matrix.
[0040] A rotary joint feature determination subunit is used to calculate the rotary joint feature matrix based on the joint feature matrix and the rotational unit matrix;
[0041] The first feature calculation subunit is used to perform weight sharing on the rotational joint feature matrix and determine the calculation formula of the joint features after weight rotation.
[0042] The component construction subunit is used to construct the rotation weight component according to the calculation formula of the joint features after the weight rotation.
[0043] Optionally, the weighted rotating residual module includes a first rotating weighted residual submodule, a second rotating weighted residual submodule, and a third rotating weighted residual submodule;
[0044] The input terminal of the first rotation weight residual submodule is connected to the output terminal of the first rotation weight component, and the output terminal of the first rotation weight residual submodule is connected to the input terminal of the second rotation weight residual submodule.
[0045] The output of the second rotation weight residual submodule is connected to the input of the third rotation weight residual submodule, and the output of the third rotation weight residual submodule is connected to the input of the second rotation weight component.
[0046] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0047] This invention performs nonlinear mapping processing on the two-dimensional coordinates of multiple joints in a human skeleton image through the rotation weight component in the first rotation weight component, the rotation weight component in the rotation weight residual module, and the rotation weight component in the second rotation weight component, and finally obtains the three-dimensional coordinates of the joints, thereby greatly reducing the error of 2D pose to 3D pose conversion, which is simple and efficient. Attached Figure Description
[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0049] Figure 1 This is a flowchart illustrating the human pose estimation method based on joint weight rotation of the present invention.
[0050] Figure 2 This is a schematic diagram illustrating the composition of the joint features of the present invention under the weighted rotation update method;
[0051] Figure 3This is a schematic diagram illustrating the correspondence between weights and joint features in the weighted rotation component of this invention. Figure 1 ;
[0052] Figure 4 This is a schematic diagram illustrating the correspondence between weights and joint features in the weighted rotation component of this invention. Figure 2 ;
[0053] Figure 5 This is a schematic diagram illustrating the correspondence between weights and joint features in the weighted rotation component of this invention. Figure 3 ;
[0054] Figure 6 This is a 2D-3D human pose estimation network diagram using weighted rotation in this embodiment of the invention;
[0055] Figure 7 This is a schematic diagram of the human posture estimation system based on joint weight rotation according to the present invention. Detailed Implementation
[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0057] The purpose of this invention is to provide a human pose estimation method and system based on joint weight rotation, which can simply and effectively map 2D human pose to 3D human pose more accurately, thereby improving the speed of 2D-3D human pose estimation and reducing regression error.
[0058] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0059] Example 1
[0060] like Figure 1 As shown, this invention provides a human pose estimation method based on joint weight rotation, comprising:
[0061] Step 100: Obtain the two-dimensional coordinates of the human joints.
[0062] Step 200: Input the two-dimensional coordinates of the multiple joints into the first rotation weighting component to calculate the first joint features.
[0063] Step 300: Input the first joint feature into the rotation weight residual module to obtain the second joint feature.
[0064] Step 400: Input the second joint features into the second rotation weight component to obtain the joint's three-dimensional coordinates. The rotation weight residual module includes multiple rotation weight residual sub-modules, and each of the rotation weight residual sub-modules, the first rotation weight component, and the second rotation weight component includes at least one rotation weight component.
[0065] Preferably, such as Figure 6 As shown, the weighted rotation residual module includes a first rotation weight residual submodule, a second rotation weight residual submodule, and a third rotation weight residual submodule. The input terminal of the first rotation weight residual submodule is connected to the output terminal of the first rotation weight component, and the output terminal of the first rotation weight residual submodule is connected to the input terminal of the second rotation weight residual submodule. Specifically, the output feature data of the first rotation weight component is used as the input feature data of the first rotation weight residual submodule, and the output feature data of the first rotation weight residual submodule is used as the input feature data of the second rotation weight residual submodule.
[0066] The output of the second rotation weight residual submodule is connected to the input of the third rotation weight residual submodule, and the output of the third rotation weight residual submodule is connected to the input of the second rotation weight component. Specifically, the output feature data of the second rotation weight residual submodule is used as the input feature data of the third rotation weight residual submodule, and the output feature data of the third rotation weight residual submodule is used as the input feature data of the second rotation weight component.
[0067] Furthermore, the first rotation weight residual submodule includes a third rotation weight component and a fourth rotation weight component. The output of the first rotation weight component is connected to the input of the third rotation weight component, and the output of the third rotation weight component is connected to the input of the fourth rotation weight component; the output feature data of the first rotation weight component is used as the input feature data of the third rotation weight component, and the output feature data of the third rotation weight component is used as the input feature data of the fourth rotation weight component.
[0068] The output of the first rotation weight component is also connected to the first output of the fourth rotation weight component and feature addition is performed. The feature obtained after feature addition is the output feature of the first rotation weight residual submodule, and then the output feature of the first rotation weight residual submodule is used as the input feature of the second rotation weight residual submodule.
[0069] The structures of the second and third rotation weight residual submodules are the same as those of the first rotation weight residual submodule, and will not be described again here.
[0070] Preferably, the process of determining the rotation weight component specifically includes:
[0071] (1) Obtain multiple joint features to be updated from the dataset; specifically, the joint features to be updated are two-dimensional coordinates of the joints to be updated. In a specific embodiment, firstly, the two-dimensional coordinates of J joints in the dataset are obtained, and then the multiple two-dimensional coordinates are normalized and mapped to [0,1] to serve as the input of the 2D-3D human pose estimation network.
[0072] (2) The multiple joint features to be updated are determined as a joint feature matrix.
[0073] (3) Calculate the rotational joint feature matrix based on the joint feature matrix and the rotational unit matrix.
[0074] (4) Perform weight sharing on the rotational joint feature matrix to determine the calculation formula of the joint features after weight rotation; the updated joint coordinates can be determined according to the calculation formula of the joint features after weight rotation.
[0075] (5) Construct a rotation weight component based on the calculation formula of the joint features after the weight rotation.
[0076] The rotation process in the rotation weight component is as follows:
[0077] 1) Arrange multiple joint features to be updated in a closed loop. In one specific embodiment, such as... Figure 2 As shown, six human joint features are extracted from a human skeleton image, and the coordinates of the joint to be updated corresponding to each feature are obtained. These six coordinates are then evenly arranged clockwise on a circle. In practical applications, the number of human joints can be changed as needed.
[0078] 2) Arrange multiple trainable rotation learning weights in a closed loop, with each rotation learning weight corresponding to a joint coordinate to be updated. Specifically, for example... Figure 3 As shown, the rotation learning weights W are divided into 6 parts and evenly distributed on another circle. The circle containing the rotation learning weights is concentric with the circle containing the joint coordinates to be updated. Each weight corresponds to a joint coordinate, that is, the number of rotation learning weights is equal to the number of joint features.
[0079] 3) Rotate the multiple joint features to be updated that are evenly distributed in the closed loop, so that each joint feature still corresponds to a rotation learning weight after rotation. Specifically, rotate the joint features to be updated located on the outer circle counterclockwise, rotating one position at a time, as shown below: Figure 3 The diagram shows the correspondence between weights and joint features when the number of rotations is 0. In this case, the 0th joint feature is h′0. Where J represents the number of human joints to be updated, and J = 6; for example Figure 4 The diagram shows the correspondence between weights and joint features when the number of rotations is 1. In this case, the first joint feature is h′1. like Figure 5 The diagram shows the correspondence between weights and joint features when the number of rotations is 2. In this case, the second joint feature is h′2. In summary, we can obtain the following: the features of each joint to be updated after each rotation and the rotation learning weights corresponding to each feature can be obtained using the formula... Iterative calculations are performed to obtain the shared updated joint features for each rotation.
[0080] Where, h′ t Indicates the updated joint features shared after the t-th joint rotation; h i h represents the unupdated features of the i-th joint. i+1 h represents the features of the (i+1)th joint before the update. J-1 hj-1 represents the pre-update features of the (J-1)th joint, h0 represents the pre-update features of the 0th joint, h1 represents the pre-update features of the 1st joint, and hj represents the pre-update features of the 1st joint. i-1 Let Wi represent the unupdated features of the (i-1)th joint, where i, t ∈ 0, 1, 2, ..., J-1; W0 represents the learning weight of the 0th rotation of the joint; W1 represents the learning weight of the 1st rotation of the joint; Wi ... J-i+1 W represents the (J-i+1)th rotational learning weight of the joint; J-i W represents the learned weight of the i-th rotation of the joint; J-i-1 W represents the (i-1)th rotational learning weight of the joint; J-1 This represents the (J-1)th rotational learning weight of the joint.
[0081] Further, the step of calculating the rotational joint feature matrix based on the joint feature matrix and the rotational identity matrix specifically includes: calculating the rotational joint feature matrix according to formula H. R =RH, calculate the characteristic matrix of the rotary joint.
[0082] Where R represents the rotation identity matrix, and H represents the joint feature matrix. Specifically, H is a matrix composed of multiple joint features before rotation sharing update. D represents the dimension of the first feature, H R Let J be the feature matrices of the rotational joints. J represents the number of joint features to be updated, where the number of joint features to be updated, the number of human joints to be updated, and the number of rotation learning weights are equal. Let h represent a set of real-valued matrices of dimension J×D, h0 represent the joint features to be updated for the 0th joint, h1 represent the joint features to be updated for the 1st joint, and h... J-1 This represents the joint features to be updated for the (J-1)th joint.
[0083] Preferably, the step of weight sharing on the rotational joint feature matrix to determine the calculation formula for the joint features after weight rotation specifically includes:
[0084] According to the formula Determine the joint features after weight rotation.
[0085] Where H' represents the joint feature after weight rotation. H R Let D' represent the rotational joint feature matrix, D' represent the second feature dimension, and W represent the rotational learning weights. J represents the number of joint features to be updated.
[0086] Furthermore, the rotation weight component also includes a BatchNorm layer, a ReLU layer, and a Dropout layer; the joint features determined after weight rotation sharing are input to the BatchNorm layer, the output of the BatchNorm layer is connected to the input of the ReLU layer, and the output of the ReLU layer is connected to the input of the Dropout layer; a rotation weight component is constructed through the above connections.
[0087] In a specific embodiment, as shown in Figure 6, a 2D-3D human pose estimation network consisting of multiple rotation weight components is constructed. Its specific working process is as follows:
[0088] Step 1: Input the 2D coordinates of J joints.
[0089] Step 2: Use the rotation weight component to map the 2D coordinates of the joint to 128 dimensions (number of channels), which is to convert the 2D features into 128-dimensional features.
[0090] Step 3: Connect the two rotation weight components and use residual connections to perform joint feature transformation, so that the input and output dimensions of the features remain at 128.
[0091] Step 4: Repeat step 3 three times.
[0092] Step 5: Use the weight rotation method in the rotation weight component to convert the output features of step 4 into 3D coordinates.
[0093] Step 6: Calculate the loss for regressing 3D coordinates using the coordinate mean square error loss function; specifically, according to the formula... Calculate the training loss for the three-dimensional coordinates; where L CThe training loss represents the three-dimensional coordinates. C represents the calculated three-dimensional coordinates of the i-th joint. i This represents the 3D coordinates of the i-th joint in the dataset annotation.
[0094] In one specific embodiment of the present invention, after obtaining the two-dimensional coordinates of the human joints, these are used as input features of the 2D-3D human pose estimation system. Simultaneously, the three-dimensional coordinates of the human joints are obtained and used as the true values (target values) of the 2D-3D human pose estimation system to supervise the estimation network for efficient fitting and loss calculation. The 2D-3D human pose estimation system performs nonlinear mapping processing on the input two-dimensional coordinates and outputs the 3D coordinates of the human joints, which are used as the predicted values output by the pose estimation system. The error between the true value and the predicted value (training loss L) is calculated. C By continuously learning and updating the rotation learning weights in the posture learning system, the error is minimized, thus determining the final calculation parameters of the posture estimation network, resulting in the final 2D-3D human posture estimation system. In practical applications, only the two-dimensional coordinates of J joints need to be input, and the network can output high-precision, low-error three-dimensional coordinates.
[0095] Example 2
[0096] like Figure 7 As shown, this embodiment provides a human pose estimation system based on joint weight rotation, the human pose estimation system comprising:
[0097] The coordinate acquisition unit 101 is used to acquire the two-dimensional coordinates of human joints.
[0098] The first feature calculation unit 201 is used to input the two-dimensional coordinates of the multiple joints into the first rotation weight component to calculate the first joint features.
[0099] The second feature calculation unit 301 is used to input the first joint feature into the rotation weight residual module to obtain the second joint feature.
[0100] The third feature calculation unit 401 is used to input the second joint feature into the second rotation weight component to obtain the joint three-dimensional coordinates; wherein, the rotation weight residual module includes multiple rotation weight residual sub-modules, and the rotation weight residual sub-module, the first rotation weight component and the second rotation weight component each include at least one rotation weight component.
[0101] Specifically, regarding the determination of the rotation weight component, the human pose estimation system further includes:
[0102] The dataset acquisition sub-unit is used to retrieve multiple joint features from the dataset that need to be updated.
[0103] The joint feature matrix determination sub-unit is used to determine multiple joint features to be updated as a joint feature matrix.
[0104] The rotary joint feature determination subunit is used to calculate the rotary joint feature matrix based on the joint feature matrix and the rotational identity matrix.
[0105] The first feature calculation subunit is used to perform weight sharing on the rotational joint feature matrix and determine the calculation formula for the joint features after weight rotation.
[0106] The component construction subunit is used to construct the rotation weight component according to the calculation formula of the joint features after the weight rotation.
[0107] Preferably, the weighted rotation residual module includes a first rotation weight residual submodule, a second rotation weight residual submodule, and a third rotation weight residual submodule; the input terminal of the first rotation weight residual submodule is connected to the output terminal of the first rotation weight component, and the output terminal of the first rotation weight residual submodule is connected to the input terminal of the second rotation weight residual submodule; specifically, the output feature data of the first rotation weight component is used as the input feature data of the first rotation weight residual submodule, and the output feature data of the first rotation weight residual submodule is used as the input feature data of the second rotation weight residual submodule.
[0108] The output of the second rotation weight residual submodule is connected to the input of the third rotation weight residual submodule, and the output of the third rotation weight residual submodule is connected to the input of the second rotation weight component. Specifically, the output feature data of the second rotation weight residual submodule is used as the input feature data of the third rotation weight residual submodule, and the output feature data of the third rotation weight residual submodule is used as the input feature data of the second rotation weight component.
[0109] Compared with the prior art, the present invention also has the following advantages:
[0110] (1) The weighted rotation component used in this invention realizes the nonlinear mapping from 2D coordinates to 3D coordinates, which can greatly reduce the error of 2D to 3D pose conversion. It has a small number of parameters, high speed, and is simple and efficient.
[0111] (2) This invention achieves stable performance and efficient reduction of 3D pose prediction error.
[0112] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0113] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A human pose estimation method based on joint weight rotation, characterized in that, The human pose estimation method includes: Obtain the two-dimensional coordinates of human joints; The two-dimensional coordinates of the multiple joints are input into the first rotation weighting component to calculate the first joint features; The first joint features are input into the rotation weight residual module to obtain the second joint features; The second joint feature is input into the second rotation weight component to obtain the joint's three-dimensional coordinates; The rotation weight residual module includes multiple rotation weight residual sub-modules, and each of the rotation weight residual sub-module, the first rotation weight component, and the second rotation weight component includes at least one rotation weight component. The process of determining the rotation weight component specifically includes: Obtain multiple joint features to be updated from the dataset; determine the multiple joint features to be updated as a joint feature matrix; calculate the rotation joint feature matrix based on the joint feature matrix and the rotation unit matrix; perform weight sharing on the rotation joint feature matrix to determine the calculation formula of the joint features after weight rotation; construct a rotation weight component based on the calculation formula of the joint features after weight rotation. The step of calculating the rotational joint feature matrix based on the joint feature matrix and the rotational identity matrix specifically includes: According to formula H R =RH, calculate the feature matrix of the rotary joint; Where R represents the rotational identity matrix and H represents the joint feature matrix. H R The feature matrix of the rotational joint. J represents the number of joint features to be updated. Let D represent the set of real number matrices, D represent the first feature dimension, h0 represent the joint feature to be updated for the 0th joint, h1 represent the joint feature to be updated for the 1st joint, and h J-1 This represents the joint features to be updated for the (J-1)th joint.
2. The human pose estimation method based on joint weight rotation according to claim 1, characterized in that, The step of weight sharing on the rotational joint feature matrix to determine the calculation formula for the joint features after weight rotation specifically includes: According to the formula Determine the joint features after weight rotation; Where H' represents the joint feature after weight rotation. H R Represents the feature matrix of a rotational joint. Let D' represent the set of real matrix values, D′ represent the dimension of the second feature, and W represent the rotation-learned weights. J represents the number of joint features to be updated.
3. The human pose estimation method based on joint weight rotation according to claim 1, characterized in that, The rotated weight residual module includes a first rotated weight residual submodule, a second rotated weight residual submodule, and a third rotated weight residual submodule; The input terminal of the first rotation weight residual submodule is connected to the output terminal of the first rotation weight component, and the output terminal of the first rotation weight residual submodule is connected to the input terminal of the second rotation weight residual submodule. The output of the second rotation weight residual submodule is connected to the input of the third rotation weight residual submodule, and the output of the third rotation weight residual submodule is connected to the input of the second rotation weight component.
4. The human pose estimation method based on joint weight rotation according to claim 3, characterized in that, The first rotation weight residual submodule includes a third rotation weight component and a fourth rotation weight component; The output of the first rotation weight component is connected to the input of the third rotation weight component, and the output of the third rotation weight component is connected to the input of the fourth rotation weight component. The output of the first rotation weight component is also connected to the first output of the fourth rotation weight component in a jump connection and perform feature addition.
5. A human pose estimation system based on joint weighted rotation, characterized in that, The human pose estimation system includes: The coordinate acquisition unit is used to acquire the two-dimensional coordinates of human joints; The first feature calculation unit is used to input the two-dimensional coordinates of the multiple joints into the first rotation weight component to calculate the first joint features; The second feature calculation unit is used to input the first joint feature into the rotation weight residual module to obtain the second joint feature; The third feature calculation unit is used to input the second joint features into the second rotation weight component to obtain the three-dimensional coordinates of the joint; The rotation weight residual module includes multiple rotation weight residual sub-modules, and each of the rotation weight residual sub-module, the first rotation weight component, and the second rotation weight component includes at least one rotation weight component. Regarding the determination of the rotation weight component, the human pose estimation system further includes: The dataset acquisition sub-unit is used to retrieve multiple joint features to be updated from the dataset. The joint feature matrix determination subunit is used to determine multiple joint features to be updated as a joint feature matrix. A rotary joint feature determination subunit is used to calculate the rotary joint feature matrix based on the joint feature matrix and the rotational unit matrix; The first feature calculation subunit is used to perform weight sharing on the rotational joint feature matrix and determine the calculation formula of the joint features after weight rotation. The component construction subunit is used to construct the rotation weight component according to the calculation formula of the joint features after the weight rotation; The step of calculating the rotational joint feature matrix based on the joint feature matrix and the rotational identity matrix specifically includes: According to formula H R =RH, calculate the feature matrix of the rotary joint; Where R represents the rotational identity matrix and H represents the joint feature matrix. H R The feature matrix of the rotational joint. J represents the number of joint features to be updated. Let D represent the set of real number matrices, D represent the first feature dimension, h0 represent the joint feature to be updated for the 0th joint, h1 represent the joint feature to be updated for the 1st joint, and h J-1 This represents the joint features to be updated for the (J-1)th joint.
6. The human pose estimation system based on joint weight rotation according to claim 5, characterized in that, The rotated weight residual module includes a first rotated weight residual submodule, a second rotated weight residual submodule, and a third rotated weight residual submodule; The input terminal of the first rotation weight residual submodule is connected to the output terminal of the first rotation weight component, and the output terminal of the first rotation weight residual submodule is connected to the input terminal of the second rotation weight residual submodule. The output of the second rotation weight residual submodule is connected to the input of the third rotation weight residual submodule, and the output of the third rotation weight residual submodule is connected to the input of the second rotation weight component.