Adaptive pose estimation method based on global feature cooperation and graph structure driving

Abstract

The present application relates to a kind of self-adapting pose estimation methods based on global feature cooperation and graph structure driving, comprising the following steps: step S1: preliminary feature extraction is carried out using ResNet main network, and the multiscale feature map containing global shape of human body and joint position detail information is obtained;Step S2: multiscale feature map is input into global adaptive feature fusion module to carry out adaptive feature fusion, integrate global and local information, capture the overall structure and details of human posture;Step S3: the adaptive fusion feature obtained in step S2 is input into dynamic visual feature encoder, and through deformable multi-head attention mechanism, key point recognition and positioning are optimized;Step S4: the interdependence between key points is analyzed using graph reasoning technology, and the key point position prediction is further refined.The method is conducive to improving the accuracy of pose estimation in complex scenes.

Description

Technical Field

[0001] The present invention relates to the field of computer vision, and in particular to an adaptive posture estimation method based on global feature collaboration and graph structure driving. Background Art

[0002] In recent years, with the rapid development of computer vision technology, deep learning-based pose estimation methods have become a research hotspot in this field. These methods have achieved significant breakthroughs on public datasets and continue to set new performance records. As a key branch of computer vision, pose recognition technology has a wide range of applications, from human-computer interaction to virtual reality, from security monitoring to medical rehabilitation. However, while pose estimation technology has achieved promising results in many scenarios, its performance is often greatly challenged in complex environments. Complex environments can contain numerous uncertainties, such as varying lighting, occlusions, background interference, and varying human postures, all of which can significantly impact the accuracy of pose estimation. Therefore, improving the accuracy of pose estimation in complex scenarios has become a pressing issue in the field. Summary of the Invention

[0003] The purpose of the present invention is to provide an adaptive pose estimation method based on global feature collaboration and graph structure driving, which is conducive to improving the accuracy of pose estimation in complex scenes.

[0004] To achieve the above objectives, the present invention adopts a technical solution: an adaptive pose estimation method based on global feature collaboration and graph structure driving, comprising the following steps:

[0005] Step S1: Use the ResNet backbone network to perform preliminary feature extraction to obtain a multi-scale feature map containing the global morphology of the human body and detailed information of joint positions;

[0006] Step S2: Input the multi-scale feature map into the global adaptive feature fusion module for adaptive feature fusion, integrating global and local information to capture the overall structure and details of the human posture;

[0007] Step S3: Input the adaptive fusion features obtained in step S2 into the dynamic visual feature encoder, and optimize the key point recognition and positioning through the deformable multi-head attention mechanism;

[0008] Step S4: Use graph reasoning technology to analyze the interdependencies between key points and further refine the key point position prediction.

[0009] Furthermore, the step S1 specifically includes the following steps:

[0010] Step S11: Obtain a public human pose estimation dataset from the Internet and obtain relevant annotations of the data in the dataset;

[0011] Step S12: ResNet-50 is used as the feature extraction backbone network to preliminarily extract multi-scale feature maps from the input data;

[0012] Step S13: In the absence of differentiable matching, the minimum BatchSize is first set to 14, the initial learning rate α0 is set to 0.001, and the cosine decay strategy is used for adjustment. The network optimization is performed by the Adam optimizer.

[0013] Furthermore, the step S2 specifically includes the following steps:

[0014] Step S21: For the multi-scale feature map extracted in step S1, each layer of feature X i Calculating global attention weights and local attention weights Where i represents the level, taking C3, C4, C5; the global attention weight By X i Perform global average pooling (GAP) and pass it through a fully connected layer (FC); the local attention weight By i Apply a small 3×3 convolution kernel to obtain;

[0015]

[0016] Step S22: Get the global attention weight and local attention weights Weight each layer of the multi-scale feature map and fuse the global and local weighted features to form enhanced features

[0017]

[0018] Among them, σ(·) is the activation function used to normalize the weights;

[0019] Step S23: In order to integrate information at different scales, the enhanced features of the C3, C4, and C5 layers are concatenated in the channel dimension using the Concat(·) function. The concatenated features are then further optimized through convolution to meet the requirements of subsequent tasks.

[0020]

[0021] Among them, X adaptiverepresents the final adaptive fusion feature, which integrates the comprehensive information of global context and local details.

[0022] Furthermore, the step S3 specifically includes the following steps:

[0023] Step S31: Adaptively fusion feature X obtained in step S2 adaptive Input is processed into the encoder; in the encoder, the deformable attention mechanism first processes X through a set of queries Q, keys K and values ​​V adaptive , where Q, K, and V are obtained by the following formula:

[0024] Q=X adaptive W Q

[0025] K=X adaptive W K

[0026] V=X adaptive W V

[0027] Among them, W Q , W K , W V are the weight matrices for query, key, and value respectively;

[0028] Step S32: After processing in step S31, the deformable attention mechanism dynamically adjusts the focus position of the key K and value V by learning the offset, so as to focus on the features more flexibly; this process is achieved through the offset function f offsets To do this, it adjusts K and V based on the query Q:

[0029] K'=K+f offset (Q)

[0030] V'=V+f offset (Q)

[0031] Among them, K' and V' are the adjusted key and value, which can dynamically adjust the focus area according to the background noise;

[0032] Step S33: The encoder uses the adjusted K' and V' and the original query Q to calculate the deformable multi-head attention output to obtain the transformed feature X trans :

[0033]

[0034] Among them, the softmax function ensures the regularization of attention weights, and d k is the dimension of the key vector, used to scale the dot product for training stability.

[0035] Furthermore, the step S4 specifically includes the following steps:

[0036] Step S41: Define the human body structure as an undirected graph G = (Ver, E), where Ver is a vertex set representing the joints of the human body, i.e., key points, and E is an edge set representing the physical connections between key points; use the adjacency matrix to formally describe this human body structure map, where each element of the matrix identifies whether a pair of key points are directly connected; assuming that the human body has key points, then the dimension of the adjacency matrix A is Each element A mn Indicates whether the mth and nth key points are directly connected; if there is a direct physical connection between key points m and n, then A mn =1, otherwise A mn =0;

[0037] Step S42: Graph reasoning technology is combined with the decoder to further enhance the ability to detect human joints. In the early stage of the decoder, the features of each node, i.e., key points, are enhanced using predefined human structure map information. This enhancement operation is achieved by aggregating the features of neighboring nodes, which can be expressed as:

[0038]

[0039] Among them, x' id is the enhanced feature, N(id) is the neighbor set of node id, W0 and W1 correspond to the weight matrices of node features and neighbor features respectively, and σ is a nonlinear activation function;

[0040] Step S43: Enhanced node feature x' id Processed through a fully connected layer (FC) to predict the location of key points; expressed as:

[0041] y id =FC(x' id )=σ(W fc x' id +b fc )

[0042] Among them, W fc and b fc are the weights and biases of the fully connected layer, y id is the predicted key point position;

[0043] Step S44: A multi-head attention mechanism is introduced after the fully connected layer; the attention of each head is expressed as:

[0044]

[0045] in, are query, key and value respectively, through the enhanced feature x' id Calculated, d k is the dimension of the key vector; the output of the multi-head attention mechanism is passed through the fully connected layer to refine the predicted key point position to further improve the accuracy of the prediction;

[0046] Finally, the model outputs a prediction of the location of each keypoint;

[0047] Step S45: Use classification loss L c The classification head is applied to the pose decoder; in the pose and joint decoders, L1 loss L is used respectively. r and OKS loss L oks To regress pose and joints, we use heatmap loss L h To assist training to achieve rapid convergence; OKS loss L oks The specific formula is as follows:

[0048]

[0049] in, represents the Euclidean distance between the true value and the predicted value; s represents the scale of the target key point, v o A sign indicating whether the true value is visible, δ is the Dirac function, k o represents the key point constant;

[0050] The overall loss function formula is:

[0051] L=L c +αL r +βL h +γL oks

[0052] Among them, α, β, and γ are loss weights.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. The present invention can extract and process feature maps containing key information such as the global shape and joint positions of the human body, effectively identify human posture in complex environments, improve the accuracy of posture estimation, and effectively reduce the impact of environmental noise.

[0055] 2. The dynamic visual feature encoder and its deformable multi-head attention mechanism enable this method to adapt to complex and changeable human postures and improve the accuracy of key point recognition and positioning.

[0056] 3. The present invention uses graph reasoning technology to enable the present invention to have a deeper understanding of the interdependence between key points, further improve the accuracy and stability of posture estimation, and solve the problem caused by multi-person occlusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flowchart of a method implementation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0059] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0060] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0061] like Figure 1 As shown, this embodiment provides an adaptive pose estimation method based on global feature collaboration and graph structure driving, including the following steps:

[0062] Step S1: Use the ResNet backbone network to perform preliminary feature extraction to obtain a multi-scale feature map containing the global morphology of the human body and detailed information of joint positions;

[0063] Step S2: Input the multi-scale feature map into the global adaptive feature fusion module for adaptive feature fusion, integrating global and local information to capture the overall structure and details of the human posture;

[0064] Step S3: Input the adaptive fusion features obtained in step S2 into the dynamic visual feature encoder, and optimize the key point recognition and positioning through the deformable multi-head attention mechanism;

[0065] Step S4: Use graph reasoning technology to analyze the interdependencies between key points and further refine the key point position prediction.

[0066] In this embodiment, the specific implementation steps of step S1 are as follows.

[0067] Step S11: Obtain a public human pose estimation dataset from the Internet and obtain relevant annotations of the data in the dataset.

[0068] Step S12: Using the data obtained from the dataset as input, ResNet-50 is used as the feature extraction backbone network to preliminarily extract multi-scale feature maps from the input data.

[0069] Step S13: In the absence of differentiable matching, the minimum BatchSize is first set to 14, the initial learning rate α0 is set to 0.001, and the cosine decay strategy is used for adjustment. The network optimization is performed by the Adam optimizer. In this embodiment, the image input size of the CrowdPose dataset is 320×256, and the learning rate adjustment of CrowdPose is decayed to one tenth of the original at the 40th and 70th iterations, respectively.

[0070] In this embodiment, the specific implementation steps of step S2 are as follows.

[0071] Step S21: For the multi-scale feature map extracted in step S1, each layer of feature X i (where i represents the level, take C3, C4, C5) to calculate the global attention weight and local attention weights The global attention weight By X i Perform global average pooling (GAP) and pass it through a fully connected layer (FC); the local attention weight By i Apply a small 3×3 convolution kernel to obtain;

[0072]

[0073] Step S22: Get the global attention weight and local attention weights Weight each layer of the multi-scale feature map and fuse the global and local weighted features to form enhanced features

[0074]

[0075] Here, σ(·) is an activation function, such as Sigmoid, used to normalize the weights, and · represents element-wise multiplication.

[0076] Step S23: In order to integrate information at different scales, the enhanced features of the C3, C4, and C5 layers are concatenated in the channel dimension using the Concat(·) function. The concatenated features are then further optimized through convolution to meet the requirements of subsequent tasks.

[0077]

[0078] Among them, X adaptive represents the final adaptive fusion feature, which integrates the comprehensive information of global context and local details.

[0079] In this embodiment, the specific implementation steps of step S3 are as follows.

[0080] Step S31: Adaptively fusion feature X obtained in step S2 adaptive Input to the encoder for processing. Specifically, suppose X adaptive It is the feature matrix obtained from the input data after preprocessing and adaptive feature extraction. In the encoder, the deformable attention mechanism first processes X through a set of queries Q, keys K and values ​​V. adaptive , where Q, K, and V are obtained by the following formula:

[0081] Q=X adaptive W Q

[0082] K=X adaptive W K

[0083] V=X adaptive W V

[0084] Among them, W Q , W K , W V are the weight matrices for query, key, and value respectively.

[0085] Step S32: After step S31, the deformable attention mechanism dynamically adjusts the focus position of the key K and value V by learning the offset, so as to focus on the features more flexibly. This process can be achieved by the offset function f offsets To do this, it adjusts K and V based on the query Q:

[0086] K'=K+f offset (Q)

[0087] V'=V+f offset (Q)

[0088] Here, K' and V' are the adjusted key and value, which can dynamically adjust the focus area according to the background noise.

[0089] Step S33: The encoder uses the adjusted K' and V' and the original query Q to calculate the deformable multi-head attention output to obtain the transformed feature X trans :

[0090]

[0091] Among them, the softmax function ensures the regularization of attention weights, and d k is the dimension of the key vector, used to scale the dot product for training stability.

[0092] In this embodiment, the specific implementation steps of step S4 are as follows.

[0093] Step S41: Define the human body structure as an undirected graph G = (Ver, E), where Ver is a vertex set representing the joints of the human body, also known as key points, and E is an edge set representing the physical connections between key points. Furthermore, the adjacency matrix is ​​used to formally describe this human body structure map, where each element of the matrix identifies whether a pair of key points is directly connected. Such a mathematical expression not only facilitates the calculation and analysis of the map, but also provides an intuitive data input form for subsequent machine learning models. Assume that the human body has key points, then the dimension of the adjacency matrix A is Each element A mn Indicates whether the mth and nth key points are directly connected; if there is a direct physical connection between key points m and n, then A mn =1, otherwise A mn =0.

[0094] Step S42: Combine graph reasoning technology with the decoder to further enhance the ability to detect human joints; in the early stage of the decoder, use predefined human body structure atlas information to enhance the features of each node (key point). This can be achieved by combining the node features with the aggregated features of its neighbors, thereby introducing graph structure information without changing the overall network architecture. This process not only includes the mathematical expression of the human body structure atlas, but also involves the specific steps of how to optimize the posture estimation process through this atlas. Specifically, the atlas information is first used to enhance the features of each node (i.e., key point); this enhancement operation is achieved by aggregating the features of neighboring nodes, expressed as:

[0095]

[0096] Among them, x' id is the enhanced feature, N(id) is the neighbor set of node id, W0 and W1 correspond to the weight matrices of node features and neighbor features respectively, and σ is a nonlinear activation function.

[0097] Step S43: Enhanced node feature x' id The image is processed through a fully connected layer (FC) to predict the location of the key points. The role of the fully connected layer is to transform the input features so as to learn the nonlinear relationship from the feature to the key point location mapping. This can be expressed as:

[0098] y id =FC(x′ id )=σ(W fc x id +b fc )

[0099] Among them, W fc and b fc are the weights and biases of the fully connected layer, y id is the predicted key point location.

[0100] Step S44: A multi-head attention mechanism is introduced after the fully connected layer. Here, the multi-head attention mechanism can help the model capture the dependencies between different key points, especially in complex human postures. At this stage, the attention of each head can be expressed as:

[0101]

[0102] in, are query, key and value respectively, through the enhanced feature x' id Calculated, d k is the dimension of the key vector; the output of the multi-head attention mechanism is passed through the fully connected layer to refine the predicted key point position, which can further improve the accuracy of the prediction.

[0103] Finally, the model outputs a prediction of the location of each keypoint;

[0104] Step S45: Use classification loss (denoted as L c ), applied to the classification head of the pose decoder. In the pose and joint decoders, L1 loss (denoted as L r ) and OKS loss (denoted as L oks ) to regress the pose and joints. At the same time, the heat map loss (denoted as L h ) to assist training to achieve fast convergence, but it is only used in training and not in inference. When dealing with postures with large scale differences, L1 loss shows limitations due to the similarity of relative errors; therefore, this method supplements the use of object keypoint similarity (OKS) loss to improve this problem. OKS loss L oks The specific formula is as follows:

[0105]

[0106] in, represents the Euclidean distance between the true value and the predicted value; s represents the scale of the target key point, v o A sign indicating whether the true value is visible, δ is the Dirac delta function, k o represents the key point constant;

[0107] The overall loss function formula is:

[0108] L=L c +αL r +βL h +γL oks

[0109] Among them, α, β, and γ are loss weights.

[0110] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0111] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0112] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0113] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0114] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. An adaptive pose estimation method based on global feature collaboration and graph structure driving, characterized in that: The following steps are involved: Step S1: Use the ResNet backbone network to perform preliminary feature extraction to obtain a multi-scale feature map containing the global morphology of the human body and detailed information of joint positions; Step S2: Input the multi-scale feature map into the global adaptive feature fusion module for adaptive feature fusion, integrating global and local information to capture the overall structure and details of the human posture; Step S3: Input the adaptive fusion features obtained in step S2 into the dynamic visual feature encoder, and optimize the key point recognition and positioning through the deformable multi-head attention mechanism; Step S4: Graph reasoning technology is used to analyze the interdependencies between key points and further refine the key point position prediction; The step S2 specifically includes the following steps: Step S21: For the multi-scale feature map extracted in step S1, each layer of feature X i Calculating global attention weights and local attention weights Where i represents the level, taking C3, C4, C5; the global attention weight By X i Perform global average pooling (GAP) and pass it through a fully connected layer (FC); the local attention weight By i Apply a small 3×3 convolution kernel to obtain; Step S22: Get the global attention weight and local attention weights Weight each layer of the multi-scale feature map and fuse the global and local weighted features to form enhanced features Among them, σ(·) is the activation function used to normalize the weights; Step S23: In order to integrate information at different scales, the enhanced features of the C3, C4, and C5 layers are concatenated in the channel dimension using the Concat(·) function. The concatenated features are then further optimized through convolution to meet the requirements of subsequent tasks. Among them, X adaptive represents the final adaptive fusion feature, which integrates the comprehensive information of global context and local details.

2. The adaptive pose estimation method based on global feature collaboration and graph structure driving according to claim 1 is characterized in that: The step S1 specifically includes the following steps: Step S11: Obtain a public human pose estimation dataset from the Internet and obtain relevant annotations of the data in the dataset; Step S12: ResNet-50 is used as the feature extraction backbone network to preliminarily extract multi-scale feature maps from the input data; Step S13: In the absence of differentiable matching, the minimum BatchSize is first set to 14, the initial learning rate α0 is set to 0.001, and the cosine decay strategy is used for adjustment. The network optimization is performed by the Adam optimizer.

3. The adaptive pose estimation method based on global feature collaboration and graph structure driving according to claim 1 is characterized in that: The step S3 specifically includes the following steps: Step S31: Adaptively fusion feature X obtained in step S2 adaptive Input is processed into the encoder; in the encoder, the deformable attention mechanism first processes X through a set of queries Q, keys K and values ​​V adaptive , where Q, K, and V are obtained by the following formula: Q=X adaptive W Q K=X adaptive W K V=X adaptive W V Among them, W Q , W K , W V are the weight matrices for query, key, and value respectively; Step S32: After processing in step S31, the deformable attention mechanism dynamically adjusts the focus position of the key K and value V by learning the offset, so as to focus on the features more flexibly; this process is achieved through the offset function f offsets To do this, it adjusts K and V based on the query Q: K′=K+f offset (Q) V′=V+f offset (Q) Among them, K′ and V′ are the adjusted key and value, which can dynamically adjust the focus area according to the background noise; Step S33: The encoder uses the adjusted K′ and V′ and the original query Q to calculate the deformable multi-head attention output to obtain the transformed feature X trans : Among them, the softmax function ensures the regularization of attention weights, and d k is the dimension of the key vector, used to scale the dot product for training stability.

4. The adaptive pose estimation method based on global feature collaboration and graph structure driving according to claim 3 is characterized in that: The step S4 specifically includes the following steps: Step S41: Define the human body structure as an undirected graph G = (Ver, E), where Ver is a vertex set representing the joints of the human body, i.e., key points, and E is an edge set representing the physical connections between key points; use the adjacency matrix to formally describe this human body structure map, where each element of the matrix identifies whether a pair of key points are directly connected; assuming that the human body has key points, then the dimension of the adjacency matrix A is Each element A mn Indicates whether the mth and nth key points are directly connected; if there is a direct physical connection between key points m and n, then A mn =1, otherwise A mn =0; Step S42: Graph reasoning technology is combined with the decoder to further enhance the ability to detect human joints. In the early stage of the decoder, the features of each node, i.e., key points, are enhanced using predefined human structure map information. This enhancement operation is achieved by aggregating the features of neighboring nodes, which can be expressed as: Where x′ id is the enhanced feature, N(id) is the neighbor set of node id, W0 and W1 correspond to the weight matrices of node features and neighbor features respectively, and σ is a nonlinear activation function; Step S43: Enhanced node feature x′ id Processed through a fully connected layer (FC) to predict the location of key points; expressed as: y id =FC(x′ id )=σ(W fc x′ id +b fc ) Among them, W fc and b fc are the weights and biases of the fully connected layer, y id is the predicted key point position; Step S44: A multi-head attention mechanism is introduced after the fully connected layer; the attention of each head is expressed as: in, are query, key and value respectively, through the enhanced feature x′ id Calculated, d k is the dimension of the key vector; the output of the multi-head attention mechanism is passed through the fully connected layer to refine the predicted key point position to further improve the accuracy of the prediction; Finally, the model outputs a prediction of the location of each keypoint; Step S45: Use classification loss L c The classification head is applied to the pose decoder; in the pose and joint decoders, L1 loss L is used respectively. r and OKS loss L oks To regress pose and joints, we use heatmap loss L h To assist training to achieve rapid convergence; OKS loss L oks The specific formula is as follows: in, represents the Euclidean distance between the true value and the predicted value; s represents the scale of the target key point, v o A sign indicating whether the true value is visible, δ is the Dirac function, k o represents the key point constant; The overall loss function formula is: L=L c +αL r +βL h +γL oks Among them, α, β, and γ are loss weights.

Images