A high-resolution lightweight human pose estimation method and system combined with an attention mechanism

Abstract

This invention discloses a high-resolution, lightweight human pose estimation method incorporating an attention mechanism. The invention designs a high-resolution, lightweight human pose estimation network. The core of the network consists of specially designed GEBasicblock and GEBottleneck modules, which utilize inverse residual structures and Ghost modules to extract features and employ an ECA channel attention mechanism to enhance the capture of key information. These modules construct a multi-stage GEHRNet network, which, through multi-scale feature fusion technology, increases the number of channels while reducing resolution, effectively balancing feature resolution and capacity, thereby helping the model focus on regions near keypoints. This invention, under limited parameter and computational constraints, can better extract features from the input image, ensuring high resolution while achieving higher inference speed, making it suitable for deployment on resource-constrained embedded devices.

Description

Technical Field

[0001] This invention relates to a human pose estimation technique, and more particularly to a high-resolution, lightweight 2D human pose estimation method and system. Background Technology

[0002] Human pose estimation (HPE) is a fundamental task in computer vision, aiming to detect the positional information of various joints such as the head, shoulders, wrists, and knees in a 2D image. It serves as an auxiliary and preparatory step for many other vision tasks and currently has important and widespread applications in fields such as intelligent video surveillance, human-computer interaction, autonomous driving, and intelligent healthcare. The development of deep learning has provided new opportunities for the improvement and optimization of human pose estimation algorithms, especially the application of convolutional neural networks (CNNs) and recurrent neural networks (RNNs). However, this technology typically requires large-scale models and high computational power. This poses a significant challenge for edge devices, especially mobile devices. These devices often have limited computing power and storage space, making it impractical to deploy complex human pose estimation models on them. Therefore, to apply high-precision human pose estimation technology to real-world scenarios, it is necessary to develop more lightweight models to adapt to the hardware limitations of these devices.

[0003] The high-resolution representation learning network HRNet (Sun, Ke, et al. "Deep high-resolution representation learning for human pose estimation." Proceedings of the IEEE / CVF conference on compute vision and pattern recognition. 2019.) has demonstrated outstanding performance in human pose estimation tasks. It employs a multi-branch structure with different resolutions, enabling feature extraction at various scales. Feature fusion between different branches allows the high-resolution branch to obtain more effective features, and the output of the high-resolution branch is finally used as a heatmap to calculate keypoint coordinates. While high-resolution networks offer high accuracy, their large number of model parameters and computational cost limit their deployment on computationally constrained embedded devices. Lightweight design is necessary to improve model inference efficiency.

[0004] In the field of deep learning, channel attention mechanisms have become one of the key techniques for improving model performance. However, while early channel attention mechanisms such as SENet were effective, they increased computational complexity and potentially overlooked some important inter-channel relationships due to the introduction of dimensionality reduction operations. To address these issues, ECA-Net proposed a lightweight channel attention module that uses 1D convolutions to achieve inter-channel interactions, avoiding unnecessary dimensionality reduction, and adaptively selecting the convolution kernel size to determine the scope of local interactions. This design makes the ECA module both efficient and significantly improves model performance, becoming an important optimization technique in computer vision tasks. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a high-resolution, lightweight 2D human pose estimation method to solve the problem of achieving efficient and accurate human pose estimation on resource-constrained devices. This method aims to improve model performance by optimizing the network structure and utilizing attention mechanisms, while reducing the number of model parameters and computational complexity, enabling it to run in real-time on mobile devices or edge computing devices.

[0006] The objective of this invention is achieved through the following technical solution: a lightweight, high-resolution 2D human pose estimation method incorporating an attention mechanism, comprising the following steps:

[0007] Step 1: Construct a human pose estimation dataset and preprocess the dataset;

[0008] Step 2: Construct high-resolution, lightweight basic network modules GEBasicblock and GEBottleneck that combine attention mechanisms. Both GEBasicblock and GEBottleneck modules are inverted residual structures, including the lightweight module Ghost, ECA channel attention, and residual connections.

[0009] Step 3: Combining the GEBasicblock module and GEBottleneck module built in step S2, construct the high-resolution lightweight feature extraction network GEHRNet based on the HRNet high-resolution network and the attention mechanism, which is used to extract high-resolution features.

[0010] Step 4: Construct a human pose estimator through convolution operations, encode the high-resolution features extracted in S3, and obtain K heat maps of human joints, where K represents the number of predefined human joint categories in the dataset.

[0011] Step 5: Analyze the heat maps of the K human joints predicted in step S4. Compared with the Gaussian distribution true value heatmap constructed based on the true values Calculate the loss using the mean squared error loss function;

[0012] Step 6: Combine the loss from Step 5 to train the entire high-resolution lightweight pose estimation network with attention mechanism, which consists of Steps 2-4.

[0013] Step 7: Use the trained network model to estimate the pose of the human body image.

[0014] Furthermore, the GEBasiicblock module consists of two 3x3 Ghost modules connected sequentially, followed by an ECA channel attention module. Finally, the outputs are summed through a residual connection. In the GEBasiicblock module, a 1x1 Ghost module is first used to increase the dimensionality of the input, and then a 3x3 depthwise separable convolution is used for feature extraction. Next, a 1x1 Ghost module is used to reduce the dimensionality of the features. After these operations, the results are input into an ECA channel attention module, and finally, the outputs are summed through a residual connection.

[0015] Furthermore, the Ghost module first uses traditional convolution operations to generate m intrinsic feature maps Y', the process of which is expressed as: Y = X' * f, where X is the input feature map and f is a convolution kernel of size k x k; the generated m intrinsic feature maps Y' are then processed using the formula... A series of inexpensive linear operations are applied to each intrinsic feature in Y' to generate s Ghost feature maps, φ i,j It generates the j-th Ghost feature map y. ij The i-th linear operation; through the above operations, n = m·s output feature maps Y = {y 11 ,y 12 ...y ms This is the final output of the Ghost module.

[0016] Furthermore, the ECA module first adaptively calculates the kernel size k of the one-dimensional convolution based on the number of channels. Where C is the number of channels in the input feature, and γ and b are hyperparameters; after obtaining the kernel size k, the ECA module applies a one-dimensional convolution to the input feature to learn the importance of each channel relative to other channels; then a sigmoid function is used to obtain the channel weights; finally, the obtained channel weights are applied to the original feature map to obtain a new feature map X′.

[0017] Furthermore, the lightweight feature extraction network GEHRNet is divided into four stages: stage 1, stage 2, stage 3, and stage 4. First, the input image is processed through a stem stage consisting of two 3x3 standard convolutions with a stride of 2, generating a feature map with a size of 1 / 4 of the original image, which serves as the input to stage 1. In stage 1, multiple GEBottleneck modules are stacked. The resulting feature map is then directly fed into stage 2 without any further processing. Simultaneously, the feature map generated in stage 1 is reduced in resolution by half and its channel count is doubled, before being used as a new branch in stage 2. Starting from stage 2, each stage contains multiple parallel subnetworks, each processing feature maps at different resolutions.

[0018] In stage 2, information is exchanged and fused between feature maps of different resolutions between the two branches. A multi-resolution feature fusion unit converts the feature map of the previous stage into a feature map of different resolutions required in the current stage.

[0019] Similarly, the three branches in stage 3 and the four branches in stage 4 are processed in the same way as in stage 2. Finally, the results of the four branches in stage 4 are merged as the final output.

[0020] In the three stages of stage 2, stage 3, and stage 4, each branch of the stage consists of multiple GEBasicblock modules, and the resolution of the feature map is adjusted by upsampling and downsampling within each stage to ensure the fusion of multi-resolution feature maps.

[0021] Furthermore, the specific formula for the multi-resolution feature fusion unit to fuse from the high-resolution branch to the low-resolution branch is as follows: In the formula, n is the total number of branches; A i H represents the output of the i-th branch before feature fusion. i This represents the output of the i-th branch during high-resolution to low-resolution fusion; F(i-1,i,H) i-1 The function represents the output result H of the i-1 branch. i-1 Perform channel and resolution conversion to enable fusion with branch i, where branch i-1 is the preceding branch adjacent to branch i.

[0022] The specific formula for the multi-resolution feature fusion unit to fuse from the low-resolution branch to the high-resolution branch is as follows: In the formula, n is the total number of branches; A i L represents the output of the i-th branch before feature fusion. iThis represents the fusion result of the i-th branch when fusing from low resolution to high resolution; F(i+1,i,L) i+1 The function represents the output result L of the i+1 branch. i+1 Perform channel and resolution conversion to enable fusion with branch i, where branch i+1 is the next branch adjacent to branch i;

[0023] The final output of the multi-resolution feature fusion unit is obtained by summing the two fusion results: one from the low-resolution branch to the high-resolution branch and the other from the high-resolution branch to the low-resolution branch. The specific formula is as follows: Q i =H i +L i -A i In the formula (1 <= i <= n), n is the total number of branches, A i L represents the output of the i-th branch before feature fusion. i The fusion result of the i-th branch is used for low-resolution to high-resolution fusion; H i The fusion result of the i-th branch is used for high-resolution to low-resolution fusion; Q i This is the output of the i-th branch of the multi-resolution feature fusion unit.

[0024] Furthermore, both stage 3 and stage 4 are processed several times.

[0025] Furthermore, the formula for the mean squared error loss function is as follows:

[0026] This invention provides a high-resolution, lightweight human pose estimation system that incorporates an attention mechanism, comprising:

[0027] The processor and memory are used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a high-resolution, lightweight human pose estimation method that combines an attention mechanism as described in the above technical solution.

[0028] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0029] This invention replaces the standard 3x3 convolution with a 3x3 Ghost module using an improved Basicblock module. Compared to the traditional 3x3 convolution, the Ghost module offers stronger spatial feature extraction capabilities while reducing the number of convolution parameters. In the feature extraction branch, the improved Bottleneck module first performs dimensionality increase using a 1x1 Ghost module, then extracts features through a 3x3 depthwise separable convolution, and finally performs dimensionality reduction using another 1x1 Ghost module. Furthermore, the introduction of the ECA channel attention mechanism, using a simple 1x1 convolution operation, avoids the complex dimensionality increase and reduction processes, adding only a small number of parameters and computational cost, but significantly improving network performance.

[0030] In the feature extraction branch, adding a batch normalization (BN) layer after each convolutional layer can accelerate training and has a certain anti-overfitting effect. Furthermore, using the ReLU activation function after the BN layer can enhance the non-linearity of the network and improve its performance.

[0031] This invention combines two lightweight techniques—inverted residual structures and Ghost modules—to reconstruct modules that meet lightweight standards. It also introduces an ECA attention mechanism to ensure network accuracy. With limited parameters and computational resources, it can better extract features from the input image, achieving higher inference speed while maintaining high resolution, making it suitable for deployment on resource-constrained embedded devices. Attached Figure Description

[0032] Figure 1 This is a flowchart of the invention;

[0033] Figure 2 This is a network structure diagram of the GE Basicblock and GE Bottleneck modules in this invention;

[0034] Figure 3 This is a network structure diagram of the GEHRNet module in this invention;

[0035] Figure 4 This is a heat map display of the detection results of this invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments.

[0037] This invention employs the COCO image dataset to train a high-resolution, lightweight human pose estimation network. The core of the network consists of specially designed GEBasicblock and GEBottleneck modules, which utilize inverse residual structures and Ghost modules to extract features and employ ECA channel attention mechanisms to enhance the capture of key information. These modules construct a multi-stage GEHRNet network, using multi-scale feature fusion techniques to increase the number of channels while reducing resolution, effectively balancing feature resolution and capacity, thereby helping the model focus on regions near keypoints.

[0038] The following is a reference to Table 1 and... Figures 1 to 3 The present invention will be described in detail below.

[0039] like Figure 1 As shown, a high-resolution, lightweight human pose estimation method incorporating attention mechanisms includes the following steps:

[0040] Step S1: In this embodiment of the invention, the COCO image dataset is used, which contains over 200,000 images and annotates approximately 250,000 people with 17 keypoints. Specifically, the train2017 training subset of the COCO dataset consists of 57,000 images and 150,000 annotated person instances, while the val2017 validation subset consists of 5,000 images. The images from these training subsets are used to train the network model, while the images from the validation subset are used to evaluate the model's performance.

[0041] Step S2: Construct high-resolution, lightweight basic network modules GEBasicblock and GEBottleneck, which incorporate an attention mechanism. Both modules employ an inverse residual structure, consisting of the lightweight Ghost module, an ECA channel attention mechanism, and residual connections. Figure 2 As shown, GEBasicblock consists of two sequentially connected 3x3 Ghost modules, followed by an ECA channel attention module. The outputs of these modules are then merged via residual connections. In the design of the GEBasicblock module, a 1x1 Ghost module is first used to increase the dimensionality of the input, followed by feature extraction using a 3x3 depthwise separable convolution. Next, another 1x1 Ghost module is used to reduce the dimensionality of the features. After these operations, the results are fed into an ECA channel attention module to enhance the model's focus on important features. Finally, residual connections are used to sum the outputs of these processing steps to improve model performance.

[0042] Adding a batch normalization (BN) layer after each convolutional layer can accelerate training and has a certain anti-overfitting effect. Using the ReLU activation function after the BN layer can enhance the non-linearity of the network and make the network perform better.

[0043] Furthermore, the Ghost module first uses traditional convolution operations to generate m intrinsic feature maps Y', the process of which is expressed as: Y' = X * f, where X is the input feature map and f is a convolution kernel of size k x k. The generated m intrinsic feature maps Y' are then processed using the formula... A series of inexpensive linear operations are applied to each intrinsic feature in Y' to generate s Ghost feature maps, φ i,j It generates the j-th Gh ost feature map y ij The i-th linear operation; through the above operations, n = m·s output feature maps Y = {y 11 ,y 12 ...y ms This is the final output of the Ghost module.

[0044] Furthermore, the ECA module first adaptively calculates the kernel size k of the one-dimensional convolution based on the number of channels. Where C is the number of channels in the input feature, and γ and b are hyperparameters; after obtaining the kernel size k, the ECA module applies a one-dimensional convolution to the input feature to learn the importance of each channel relative to other channels; then a sigmoid function is used to obtain the channel weights; finally, the obtained channel weights are applied to the original feature map to obtain a new feature map X′.

[0045] Step S3: Based on the lightweight basic network modules GEBasicblock and GEBottleneck constructed in Step S2, a high-resolution lightweight feature extraction network, GEHRNet, combining an attention mechanism, is constructed based on the high-resolution HRNet network. The lightweight network GEHRNet is mainly divided into four parts: stage 1, stage 2, stage 3, and stage 4. The network structure of GEHRNet is shown in Table 1. Figure 3As shown, the input image first undergoes a STEM stage consisting of two 3x3 standard convolutions with a stride of 2, generating a feature map with a size 1 / 4 of the original image, which serves as the input for the next stage. In stage 1, four GEBottleneck modules are stacked, and the processed feature map is input to stage 2. The feature map from stage 1 is then halved in resolution and doubled in channel count, before being used as a new branch in stage 2. Starting from stage 2, a multi-scale feature fusion approach is introduced. Each stage contains multiple parallel sub-networks, each processing feature maps at different resolutions. As the number of stages increases, the network adds branches with halved resolutions while doubling the channel count. Within stage 2, feature maps of different resolutions exchange and fuse between the two branches. A multi-resolution feature fusion unit converts the feature map from the previous stage into a feature map of the required resolution for the current stage. This maintains high resolution while improving the ability to perceive details of human pose. Similarly, the processing method for the three branches in stage 3 and the four branches in stage 4 is the same as that for stage 2; stage 3 is processed 4 times and stage 4 is processed 3 times. Finally, the results of the four branches in stage 4 are merged as the final output.

[0046] In the subsequent three stages, each stage's sub-network consists of four GEBasicblock modules. Within each stage, upsampling and downsampling are used to adjust the feature map resolution to ensure the fusion of multi-resolution feature maps. The number of channels in the four different resolution branches increases exponentially, namely C, 2C, 4C, and 8C, with C being set to 32 channels. Then, the input image cropped in step S1 is used to generate high-resolution visual features through the constructed high-resolution lightweight network.

[0047] The specific formula for the multi-resolution feature fusion unit to fuse from the high-resolution branch to the low-resolution branch is as follows: In the formula, n is the total number of branches; A i H represents the output of the i-th branch before feature fusion. i This represents the output of the i-th branch during high-resolution to low-resolution fusion; F(i-1,i,H) i-1 The function represents the output result H of the i-1 branch. i-1 Perform channel and resolution conversion to enable fusion with branch i, where branch i-1 is the preceding branch adjacent to branch i.

[0048] The specific formula for the multi-resolution feature fusion unit to fuse from the low-resolution branch to the high-resolution branch is as follows: In the formula, n is the total number of branches; A iL represents the output of the i-th branch before feature fusion. i The fusion result of the i-th branch is used to convert low-resolution to high-resolution data; F(i+1,i,L) i+1 The function represents the output result L of the i+1 branch. i+1 Perform channel and resolution conversion to enable merging with branch i, where branch i+1 is the next branch adjacent to branch i.

[0049] The final output of the multi-resolution feature fusion unit is obtained by summing the two fusion results: one from the low-resolution branch to the high-resolution branch and the other from the high-resolution branch to the low-resolution branch. The specific formula is as follows: Q i =H i +L i -A i In the formula (1 <= i <= n), n is the total number of branches, A i L represents the output of the i-th branch before feature fusion. i H represents the fusion result of the i-th branch when fusing from low resolution to high resolution; i The fusion result of the i-th branch is used for high-resolution to low-resolution fusion; Q i This is the output of the i-th branch of the feature fusion unit.

[0050] Table 1. GEHRNet Network Structure of the Invention

[0051]

[0052]

[0053] Step S4: Construct a human pose estimator through convolution operations, encode the high-resolution features extracted in S3, and obtain K heat maps of human joints, where K represents the number of predefined human joint categories in the dataset. For the COCO 2017 dataset, K = 17.

[0054] Step S5: Heat maps of the K human joints predicted in step S4 Compared with the Gaussian distribution true value heatmap constructed based on the true values The loss is calculated using the mean squared error loss function, the formula for which is:

[0055] Step S6: Using the Adam optimizer, train the entire high-resolution lightweight pose estimation network incorporating the attention mechanism based on the error obtained in Step S4. Specifically, we set the hyperparameters of Adam to: learning rate 2e-3, and coefficients β1 = 0.9 and β2 = 0.999 for calculating the gradient and the running average of the squared gradient.

[0056] Furthermore, in step 6, regarding the optimizer, the lightweight HRNet network built during training uses the Adaptive Moment Estimation (Adam) optimizer, which can adaptively adjust the learning rate to better control the convergence speed and performance of the model.

[0057] Furthermore, regarding the learning rate, for the MS-COCO dataset, the initial learning rate was set to 10⁻³, and then reduced to 10⁻⁴ and 10⁻⁵ in the 90th and 120th epochs respectively, for a total of 140 epochs of training.

[0058] Furthermore, the network training loss function consists of two parts: heatmap prediction loss and offset regression loss. The heatmap prediction loss uses L2 loss to constrain the prediction of the heatmap, and the offset regression loss uses normalized Smooth L1 Loss to constrain the regression of the offset.

[0059] Step S7: Use the trained high-resolution lightweight human pose estimation network with attention mechanism to estimate the pose of the human images to verify the detection accuracy of the trained high-resolution lightweight human pose estimation network with attention mechanism. Specifically, we obtain the index position of the corresponding peak of the heat map of the K human joints predicted in step S4, which is the final prediction result of the joints.

[0060] like Figure 4 As shown, based on the final detection results, the improved network in this invention can accurately identify and locate key points of the skier's body, such as the head, shoulders, elbows, wrists, hips, knees, and ankles. Through the connection of these key points, the network can analyze and understand the skier's posture, demonstrating its effectiveness in 2D human posture estimation. This network structure enables efficient and accurate 2D human posture estimation on resource-constrained devices, providing strong technical support for various applications.

[0061] On the other hand, embodiments of the present invention provide a high-resolution, lightweight human pose estimation system incorporating an attention mechanism, comprising:

[0062] The processor and memory are used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a high-resolution, lightweight human pose estimation method that combines an attention mechanism as described in the above technical solution.

[0063] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A high-resolution, lightweight human pose estimation method incorporating an attention mechanism, characterized in that, Includes the following steps: Step 1: Construct a human pose estimation dataset and preprocess the dataset; Step 2: Construct high-resolution, lightweight basic network modules GEBasicblock and GEBottleneck that combine attention mechanisms. Both GEBasicblock and GEBottleneck modules are inverted residual structures, including the lightweight module Ghost, ECA channel attention, and residual connections. Step 3: Combining the GEBasicblock module and GEBottleneck module built in step S2, construct the high-resolution lightweight feature extraction network GEHRNet based on the HRNet high-resolution network and the attention mechanism, which is used to extract high-resolution features. The lightweight feature extraction network GEHRNet is divided into four stages: stage 1, stage 2, stage 3, and stage 4. First, the input image is processed by a stem stage consisting of two 3x3 standard convolutions with a stride of 2, generating a feature map with a size of 1 / 4 of the original image, which serves as the input to stage 1. In stage 1, multiple GEBottleneck modules are stacked. The resulting feature map is then directly fed into stage 2 without any further processing. Simultaneously, the feature map generated in stage 1 is reduced in resolution by half and its number of channels is doubled, before being used as a new branch in stage 2. Starting from stage 2, each stage contains multiple parallel subnetworks, each processing feature maps at different resolutions. In stage 2, information is exchanged and fused between feature maps of different resolutions between the two branches. A multi-resolution feature fusion unit converts the feature map of the previous stage into a feature map of different resolutions required in the current stage. Similarly, the three branches in stage 3 and the four branches in stage 4 are processed in the same way as in stage 2. Finally, the results of the four branches in stage 4 are merged as the final output. In the three stages of stage 2, stage 3, and stage 4, each branch of the stage consists of multiple GEBasicblock modules, and the resolution of the feature map is adjusted by upsampling and downsampling within each stage to ensure the fusion of multi-resolution feature maps. The specific formula for the multi-resolution feature fusion unit to fuse from the high-resolution branch to the low-resolution branch is as follows: In the formula, n is the total number of branches; A i H represents the output of the i-th branch before feature fusion. i This represents the output of the i-th branch during high-resolution to low-resolution fusion; F(i-1,i,H) i-1 The function represents the output result H of the i-1 branch. i-1 Perform channel and resolution conversion to enable fusion with branch i, where branch i-1 is the preceding branch adjacent to branch i. The specific formula for the multi-resolution feature fusion unit to fuse from the low-resolution branch to the high-resolution branch is as follows: In the formula, n is the total number of branches; A i L represents the output of the i-th branch before feature fusion. i This represents the fusion result of the i-th branch when fusing from low resolution to high resolution. F(i+1,i,L i+1 The function represents the output result L of the i+1 branch. i+1 Perform channel and resolution conversion to enable fusion with branch i, where branch i+1 is the next branch adjacent to branch i; The final output of the multi-resolution feature fusion unit is obtained by summing the two fusion results: one from the low-resolution branch to the high-resolution branch and the other from the high-resolution branch to the low-resolution branch. The specific formula is as follows: Q i =H i +L i -A i In the formula (1 <= i <= n), n is the total number of branches, A i L represents the output of the i-th branch before feature fusion. i H represents the fusion result of the i-th branch when fusing from low resolution to high resolution; i The fusion result of the i-th branch is used for high-resolution to low-resolution fusion; Q i This is the output of the i-th branch of the multi-resolution feature fusion unit; Step 4: Construct a human pose estimator through convolution operations, encode the high-resolution features extracted in S3, and obtain K heat maps of human joints, where K represents the number of predefined human joint categories in the dataset. Step 5: Analyze the heat maps of the K human joints predicted in step S4. Compared with the Gaussian distribution true value heatmap constructed based on the true values Calculate the loss using the mean squared error loss function; Step 6: Combine the loss from Step 5 to train the entire high-resolution lightweight pose estimation network with attention mechanism, which consists of Steps 2-4. Step 7: Use the trained network model to estimate the pose of the human body image.

2. The high-resolution, lightweight human pose estimation method incorporating an attention mechanism as described in claim 1, characterized in that: The GEBasicblock module consists of two 3x3 Ghost modules connected sequentially, followed by an ECA channel attention module. Finally, the outputs are summed through a residual connection. In the GEBasicblock module, a 1x1 Ghost module is first used to increase the dimensionality of the input, and then a 3x3 depthwise separable convolution is used for feature extraction. Next, a 1x1 Ghost module is used to reduce the dimensionality of the features. After these operations, the results are input into an ECA channel attention module, and finally, the outputs are summed through a residual connection.

3. The high-resolution, lightweight human pose estimation method incorporating an attention mechanism as described in claim 2, characterized in that: The Ghost module first uses traditional convolution operations to generate m intrinsic feature maps Y', the process of which is represented as: Y′=X*f, where X is the input feature map and f is a convolution kernel of size kxk; the generated m intrinsic feature maps Y' are then processed by the formula... A series of inexpensive linear operations are applied to each intrinsic feature in Y' to generate s Ghost feature maps, φ i,j It generates the j-th Ghost feature map y. ij The i-th linear operation; through the above operations, n = m·s output feature maps Y = {y 11 ,y 12 ...y ms This is the final output of the Ghost module.

4. The high-resolution, lightweight human pose estimation method incorporating an attention mechanism as described in claim 2, characterized in that: The ECA module first adaptively calculates the kernel size k of the one-dimensional convolution based on the number of channels. Where C is the number of channels in the input feature, and γ and b are hyperparameters; after obtaining the kernel size k, the ECA module applies a one-dimensional convolution to the input feature to learn the importance of each channel relative to other channels; then a sigmoid function is used to obtain the channel weights; finally, the obtained channel weights are applied to the original feature map to obtain a new feature map X′.

5. The high-resolution, lightweight human pose estimation method incorporating an attention mechanism as described in claim 1, characterized in that: Both stage 3 and stage 4 are processed repeatedly several times.

6. The high-resolution, lightweight human pose estimation method incorporating an attention mechanism as described in claim 1, characterized in that: The formula for the mean squared error loss function is:

7. A high-resolution, lightweight human pose estimation system incorporating an attention mechanism, characterized in that, include: The processor and memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a high-resolution, lightweight human pose estimation method incorporating an attention mechanism as described in any one of claims 1-6.

Images