Human body posture accurate estimation method based on two-stage lightweight framework

Abstract

The invention discloses a human body posture accurate estimation method based on a two-stage lightweight framework. The method comprises the following steps: collecting a plurality of human body posture images and labeling key points, and further constructing a training set for human body posture estimation; establishing a human body posture detection network, inputting the training set into the human body posture detection network for training, and obtaining a trained human body posture detection network; and obtaining a to-be-detected human body posture image, and inputting the to-be-detected human body posture image into the trained human body posture detection network for posture detection to obtain a human body posture estimation result. According to the method, an attention mechanism is introduced, a key channel is focused, attention on important key points is enhanced, meanwhile, a lightweight basic module GCBlock is constructed, compared with a current mainstream method, under the condition that the number of parameters and calculation complexity are remarkably reduced, the performance is equivalent or better, and balance of high precision and lightweight is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and particularly relates to a human pose accurate estimation method based on a two-stage lightweight framework. BACKGROUND

[0002] Two-dimensional human pose estimation is a basic task in the field of computer vision, which aims to estimate the spatial positions of human key points (such as shoulders, hips, knees, ankles, etc.) in images or videos to better understand human motion and behavior. As a fundamental and important component in many computer vision applications, with the advent of advanced human pose estimation methods, it has been applied to many fields such as behavior recognition, human-computer interaction, etc.

[0003] In recent years, deep learning and related technologies have been widely used in human pose estimation and have become the mainstream research method in this field. Human pose estimation methods based on deep learning mainly fall into two categories: coordinate regression-based methods and key point heat map regression-based methods. The coordinate regression-based method takes the position coordinates of human key points as the target value, and the model directly infers the specific coordinates of these key points from the input image. This method relies on accurate prediction of key point positions and utilizes the spatial perception ability learned by the model from data. The key point heat map regression-based method first generates a heat map and then predicts the key points based on the generated heat map, which is the more mainstream method in human pose estimation at present.

[0004] Although deep learning methods effectively improve the accuracy of human pose estimation, these methods have the problems of large number of model parameters and large number of floating point operations, which are not suitable for resource-constrained devices and are contrary to the trend of lightweight modeling. Existing lightweight networks are mainly designed from two angles: adopting different network structures and using model compression techniques. In the exploration of network structure, since the cost of network convolution operation dominates, some works aim to reduce the computational load by reducing the redundancy in matrix-vector multiplication, such as MobileNet, ShuffleNet, etc. In the use of model compression techniques, the existing network architecture is mainly compressed and optimized, and some new lightweight modules are proposed to replace the original modules for processing, such as Lite-HRNet, Dite-HRNet, etc.

[0005] However, the current lightweight human pose estimation network still has many deficiencies, such as the inability to achieve a good balance between lightweight and model accuracy, and the difficulty in meeting the dual demands of high precision and high efficiency in practical applications. SUMMARY

[0006] In order to solve the problems of too many parameters and large calculation amount of the human posture estimation model in the prior art, and the insufficient balance between accuracy and efficiency of the lightweight model, the present application provides a human posture accurate estimation method based on a two-stage lightweight framework, so as to reduce the model parameters and the calculation complexity, ensure or even improve the accuracy of human posture estimation, and realize the balance between accuracy and efficiency.

[0007] The technical scheme of the present application is: The present application comprises the following steps: S1, a plurality of human posture images are collected and key points are labeled, and then a training set for human posture estimation is constructed; S2, a human posture detection network is established, the training set is input into the human posture detection network for training, and a trained human posture detection network is obtained; S3, a human posture image to be detected is obtained and input into the trained human posture detection network for posture detection to obtain a human posture estimation result.

[0008] The human posture detection network comprises a Process Net module and a Reprocess Net module connected in sequence, the human posture image is input into the Process Net module for feature extraction processing to obtain a first-stage feature map, and then the first-stage feature map is input into the Reprocess Net module for feature enhancement processing to obtain the human posture estimation result.

[0009] The Process Net module comprises a first stage, a second stage, a third stage and a fourth stage connected in sequence; The first stage comprises a Stem layer, the human posture image is input into the Stem layer for feature extraction processing to output an initial feature map; The second stage comprises an exchange block and a high-resolution branch and a medium-resolution branch arranged in parallel, the initial feature map output by the first stage is input into the high-resolution branch and the medium-resolution branch of the second stage as the input of the second stage; The third stage comprises an exchange block and a high-resolution branch, a medium-resolution branch and a lower-resolution branch arranged in parallel, the output of the high-resolution branch of the second stage is input into the high-resolution branch of the third stage as the input of the high-resolution branch of the third stage, and the output of the medium-resolution branch of the second stage is input into the medium-resolution branch and the lower-resolution branch of the third stage as the input of the medium-resolution branch and the lower-resolution branch of the third stage; The fourth stage includes an exchange block and high-resolution branches, medium-resolution branches, lower-resolution branches and low-resolution branches arranged in parallel, the output of the high-resolution branch of the third stage is taken as the input of the high-resolution branch of the fourth stage, the output of the medium-resolution branch of the third stage is taken as the input of the medium-resolution branch of the fourth stage, the output of the lower-resolution branch of the third stage is taken as the input of the lower-resolution branch and the low-resolution branch of the fourth stage respectively, the four branches of the fourth stage correspond to feature maps with output resolutions from high to low respectively, and then multi-scale fusion is performed on the feature maps output by the four branches of the fourth stage to obtain a first stage feature map. The features with different resolutions are aligned through the exchange block in each stage between each two different resolution branches, and the aligned features are fused.

[0010] The Stem layer includes two sequentially connected feature extraction units, each of which includes a first convolutional layer, a first batch normalization layer and a first ReLU activation function connected in sequence, a human pose image is taken as the input of the Stem layer, and after being sequentially processed by the first convolutional layer, the first batch normalization layer and the first ReLU activation function of the first feature extraction unit, the human pose image is input to the second feature extraction unit, and after being sequentially processed by the first convolutional layer, the first batch normalization layer and the first ReLU activation function of the second feature extraction unit, an initial feature map is finally output.

[0011] The high-resolution branch, the medium-resolution branch, the lower-resolution branch and the low-resolution branch in each stage include a plurality of sequentially connected GCBlock modules, the GCBlock modules in each stage align and fuse the features with different resolutions of different resolution branches through the exchange block in the respective stage, the GCBlock module includes a Ghost module, a second batch normalization layer, a second ReLU activation function, a second convolutional layer, a third batch normalization layer, an adaptive fusion operation and a third ReLU activation function connected in sequence, the first GCBlock module of each branch in each stage takes the input of the branch in the stage as the input, the non-first GCBlock module of each branch in each stage takes the feature map fused by the exchange block after the previous GCBlock module as the input, the input of each GCBlock module is sequentially input to the Ghost module, the second batch normalization layer, the second ReLU activation function, the second convolutional layer and the third batch normalization layer to output an intermediate feature, then the intermediate feature and the input of each GCBlock module are subjected to adaptive fusion operation and input to the third ReLU activation function, and the output of the third ReLU activation function in each GCBlock module is taken as the output of each GCBlock module.

[0012] The Ghost module includes convolutional layers, linear operations, and concatenation operations. The input of each Ghost module to its respective GCBlock module is used as the input of its respective Ghost module. The input of the Ghost module is processed by the convolutional layer to obtain an intrinsic feature map. Then, the intrinsic feature map is subjected to linear operations to obtain NNN. Finally, the NNN and the intrinsic feature map are concatenated to obtain the output of the Ghost module.

[0013] The Reprocess Net module includes four sequentially connected ECA attention mechanism modules. The first-stage feature map serves as the input to the first ECA attention mechanism module, the output of the previous ECA attention mechanism module serves as the input to the next ECA attention mechanism module, and the output of the last ECA attention mechanism module is the human pose estimation result. Each ECA attention mechanism module includes a globally average pooling layer, a convolutional layer, a sigmoid activation function, and an element-wise multiplication operation connected in sequence. The input of each ECA attention mechanism module is first processed by the globally average pooling layer to obtain a channel-compressed feature map. Then, the channel-compressed feature map is input into the convolutional layer for convolution, and then mapped by the sigmoid activation function to obtain a weight vector. Finally, the weight vector and the input of each ECA attention mechanism module are multiplied element-wise to obtain the output of each ECA attention mechanism module.

[0014] The key points in S1 include 17 key points of the human body, namely the nose, left and right eyes, left and right ears, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles. The human pose estimation result in S3 includes the coordinates (X, Y) of 17 skeletal key points of the human body in a two-dimensional coordinate system.

[0015] The beneficial effects of this invention are: This invention achieves a balance between high accuracy and lightweight by using Process Net to maintain high-resolution global detection of key points and Reprocess Net to further process key points by emphasizing important features and suppressing irrelevant features.

[0016] The Ghost module and the regular convolution module are integrated to build a lightweight base module GCBlock, which significantly compresses the model size while maintaining human pose estimation performance; the model structure is optimized and adjusted to achieve performance comparable to the original model with a simpler model structure.

[0017] By introducing an attention mechanism to construct a Reprocess Net, the model focuses on key channels, enhances attention to important key points, and improves model accuracy.

[0018] Experimental results show that, compared with current mainstream methods, the present invention achieves comparable or better performance with a significant reduction in the number of parameters and computational complexity. On the MSCOCO dataset, compared with HRNet-W32, the accuracy is improved by 0.1%, the number of parameters is reduced by 32%, and the computational complexity is reduced by 29%. Compared with other small networks, the mAP score is significantly improved, for example, by 10.7% compared with YOLOv5s6-pose and by 9.7% compared with Lite-HRNet-18. Attached Figure Description

[0019] Figure 1 This is a diagram illustrating the overall architecture of a two-dimensional human pose estimation method.

[0020] Figure 2 This is a diagram of the specific architecture of GCBlock. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] This embodiment includes the following steps: like Figure 1 As shown, this embodiment includes: S1. Collect several human pose images and label key points to construct a training set for human pose estimation. S2. Establish a human pose detection network, input the training set into the human pose detection network for training, and obtain the trained human pose detection network. S3. Obtain the human pose image to be detected and input it into the trained human pose detection network to obtain the human pose estimation result, that is, the key point heatmap corresponding to the human pose image to be detected.

[0023] The human pose detection network consists of a Process Net module and a Reprocess Net module connected in sequence. The human pose image is input into the Process Net module for feature extraction to obtain the first-stage feature map. The first-stage feature map is then input into the Reprocess Net module for feature enhancement to obtain the human pose estimation result.

[0024] The Process Net module uses HRNet as its backbone network and integrates the Ghost module and regular convolutional modules to build a new basic module GCBlock. The network structure is optimized and adjusted to reduce the number of parameters and computational cost while ensuring the model's prediction accuracy. In other words, the original residual units in the traditional HRNet network are replaced by building the new basic module GCBlock.

[0025] The Process Net module consists of four stages connected in sequence: the first stage, the second stage, the third stage, and the fourth stage. The first stage includes the Stem layer. After the human pose image is input into the Stem layer, feature extraction processing is performed to output the initial feature map. The second stage includes swap blocks and parallel high-resolution and medium-resolution branches. The initial feature map output from the first stage is used as the input to the second stage, and is input to the high-resolution and medium-resolution branches of the second stage respectively. The third stage includes a swap block and parallel settings for high-resolution, medium-resolution, and low-resolution branches. The output of the high-resolution branch in the second stage is used as the input of the high-resolution branch in the third stage, and the output of the medium-resolution branch in the second stage is used as the input of the medium-resolution and low-resolution branches in the third stage, respectively. The fourth stage includes a swap block and parallel settings for a high-resolution branch, a medium-resolution branch, a lower-resolution branch, and a low-resolution branch. The output of the high-resolution branch in the third stage is used as the input of the high-resolution branch in the fourth stage, the output of the medium-resolution branch in the third stage is used as the input of the medium-resolution branch in the fourth stage, and the output of the lower-resolution branch in the third stage is used as the input of the lower-resolution branch and the low-resolution branch in the fourth stage, respectively. The four branches in the fourth stage correspond to four feature maps with output resolutions from high to low. Then, the feature maps output by the four branches in the fourth stage are fused at multiple scales to obtain the feature map of the first stage. In each stage, features of different resolutions are aligned between two branches with different resolutions through the exchange blocks in their respective stages, and the aligned features are then fused.

[0026] Multi-scale fusion specifically involves first upsampling the feature maps of the three lower-resolution branches output from the fourth stage to match the resolution of the highest-resolution branch feature map. Then, the four branches are fused into a feature map using a weighted average. Finally, the feature map of this stage is output after convolutional smoothing and denoising, balancing information at each scale and avoiding the dominance of a single scale.

[0027] The exchange block is used for feature interaction between branches in each stage. Specifically, the GCBlocks in different branches connected in parallel in each stage first generate feature maps. Then, the generated feature maps are upsampled and downsampled to generate feature maps of branches with a higher and lower resolution than the current branch. The branch with the highest resolution is downsampled only, and the branch with the lowest resolution is upsampled only.

[0028] The Stem layer consists of two sequentially connected feature extraction units. Each feature extraction unit includes a first convolutional layer, a first batch of normalized layers, and a first ReLU activation function, all connected in sequence. The human pose image is used as the input to the Stem layer. After passing through the first convolutional layer, the first batch of normalized layers, and the first ReLU activation function of the first feature extraction unit, it is input to the second feature extraction unit. The second feature extraction unit then processes the image through the first convolutional layer, the first batch of normalized layers, and the first ReLU activation function, finally outputting the initial feature map.

[0029] Each stage's high-resolution branch, medium-resolution branch, lower-resolution branch, and low-resolution branch each include multiple sequentially connected GCBlock modules. The GCBlock modules in each stage align and fuse features of different resolutions across the different resolution branches through their respective exchange blocks. Figure 2 As shown, the GCBlock module includes a Ghost module, a second batch normalization layer, a second ReLU activation function, a second convolutional layer, a third batch normalization layer, an adaptive fusion operation, and a third ReLU activation function, connected in sequence. The first GCBlock module in each branch of each stage takes the input of the branch of its current stage as its input. The non-first GCBlock modules in each branch of each stage take the feature map fused by the previous GCBlock module through the exchange block as their input. The input of each GCBlock module passes through the Ghost module, the second batch normalization layer, the second ReLU activation function, the second convolutional layer, and the third batch normalization layer in sequence to obtain intermediate features. Then, the intermediate features and the inputs of each GCBlock module are adaptively fused and input into the third ReLU activation function. The output of the third ReLU activation function in each GCBlock module is the output of each GCBlock module.

[0030] Specifically, GCBlock consists of two different convolutional modules (a regular convolutional module and a Ghost module), BN layers, and ReLU functions stacked together. It first obtains the intrinsic feature map through a small number of regular convolutions, then performs a series of linear operations on the intrinsic feature map to generate the "Ghost" feature map, and finally concatenates the intrinsic feature map and the "Ghost" feature map to obtain the final result.

[0031] Regular convolution: for input data Where c is the number of input channels, h and w are the height and width of the input data, and the convolution operation that generates n feature maps is represented as: ,in For convolution operations, To output the feature map, The convolutional filter for this layer has a parameter size of c×k×k×n, and the required FLOPs are calculated as n×h'×w'×c×k×k; Ghost module: First, it obtains the intrinsic feature map Y1 through a small number of regular convolutions. ,in , m≤n, parameter size is c×k×k×m, required FLOPs are m×h'×w'×c×k×k; then a series of linear operations are performed on the intrinsic feature maps to generate "Ghost" feature maps Y2. Assume that each intrinsic feature map uses a 3×3 linear kernel to obtain s "Ghost" feature maps Y2. ,in Let i be the i-th intrinsic feature map in Y1. The j-th linear kernel of the i-th intrinsic feature map has a parameter size of s×3×3×m, and the required FLOPs are m×h'×w'×s×3×3; finally, Y1 and Y2 are concatenated to obtain .

[0032] The adaptive fusion operation specifically involves extracting the channel mean and variance of the intermediate features and the inputs of each GCBlock module, then concatenating them into a statistical vector, which is then input into a single-layer fully connected network and processed by the Sigmoid function to obtain the weights. Finally, according to weight The intermediate features are fused with the inputs of each GCBlock module, whereby... It is dynamically adjusted according to the feature quality.

[0033] The Ghost module includes convolutional layers, linear operations, and concatenation operations. The input of each Ghost module to its respective GCBlock module is used as the input of its Ghost module. The input of the Ghost module is processed by the convolutional layer to obtain the intrinsic feature map. Then, the intrinsic feature map is processed by linear operations to obtain NNN. Finally, the NNN and the intrinsic feature map are concatenated to obtain the output of the Ghost module.

[0034] The Reprocess Net module further processes the feature maps output by Process Net by introducing a lightweight and efficient ECA attention mechanism, enhancing the ability to capture cross-channel information and improving the accuracy of key point prediction.

[0035] The Reprocess Net module consists of four sequentially connected ECA attention mechanism modules. The first-stage feature map serves as the input to the first ECA attention mechanism module, and the output of the previous ECA attention mechanism module serves as the input to the next ECA attention mechanism module. That is, the output of the second ECA attention mechanism module serves as the input to the third ECA attention mechanism module, the output of the third ECA attention mechanism module serves as the input to the last ECA attention mechanism module, and the output of the last ECA attention mechanism module is the human pose estimation result. Each ECA attention mechanism module includes a globally average pooling layer, a convolutional layer, a sigmoid activation function, and an element-wise multiplication operation connected in sequence. The input of each ECA attention mechanism module is first processed by the globally average pooling layer to obtain a channel-compressed feature map. Then, the channel-compressed feature map is input into the convolutional layer for convolution, and then mapped by the sigmoid activation function to obtain a weight vector. Finally, the weight vector and the input of each ECA attention mechanism module are multiplied element-wise to obtain the output of each ECA attention mechanism module.

[0036] Specifically, the implementation process of the ECA attention mechanism includes: 1) Apply global average pooling to compress the input features into a channel-dimensional statistic 1×1×C, where C represents the number of channels; 2) The input feature vector is processed by a 1D convolutional layer with a kernel size of k to capture cross-channel correlation within the local receptive field. The value of k is adaptively adjusted according to the number of input channels. 3) The Sigmoid activation function maps the values ​​in the feature vector to the range of 0-1, resulting in a weight vector of size 1×1×C. , This represents the attention weight for each channel; 4) Perform element-wise multiplication to obtain the final output feature map, using the following formula: ,in This represents a one-dimensional convolution, where k represents the number of parameters in the one-dimensional convolution. Let y represent the Sigmoid activation function and y represent the output feature.

[0037] In this embodiment, the Process Net module uses HRNet as the backbone network and includes four stages: The first stage: The image is passed through two convolutional layers with a kernel size of 3×3 and a stride of 2. Each convolutional layer is followed by a BN layer and a ReLU function. The image is downsampled four times, and the channel size is increased from 3 to 64. Subsequently, the number of channels is adjusted and features are extracted by repeatedly stacking Bottleneck, without changing the size of the feature layer.

[0038] Stages two through four: Each stage adds a new branch based on the previous stage. The new branch downsamples the lowest resolution layer of the previous stage by a factor of 2, and the number of channels is twice that of the corresponding low-resolution branch in the previous stage. For example, the second stage uses two 3×3 convolutional layers in parallel to obtain two different scale branches with downsampling scales of 4x and 8x based on the output of the previous stage. Each stage uses a feature fusion module, where GCBlock first extracts image feature information, and then interactively fuses the feature information of each branch. The output of each scale branch is obtained by fusing the outputs of all branches in the previous stage. Finally, only the highest resolution feature map is output for the next processing step.

[0039] In this embodiment, the four swap blocks in the third stage are adjusted to two swap blocks, the original residual units are replaced by GCBlock, and the expansion method of different scale branches is changed to direct downsampling from the lowest scale branch.

[0040] The key points in S1 include 17 key points of the human body in the acquired image, namely the nose, left and right eyes, left and right ears, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles. The human pose estimation result in S3 includes the coordinates (X, Y) of 17 skeletal key points of the human body in the image in a two-dimensional coordinate system. At the same time, the visualization output of the two-dimensional human skeletal key point prediction result generated by connecting the above 17 human key points can be selected.

[0041] The experimental data for this embodiment are as follows: Dataset: The model training and testing used the MSCOCO dataset, with 118,287 human pose images as the training set, 5,000 as the validation set, and 20,000 as the test set. Experimental environment: PyTorch 1.10.1 framework, GPU GeForce RTX 4090, system memory 24GB, programming language Python 3.7, not pre-trained on ImageNet dataset; Training parameters: AdamW optimizer was used, initial learning rate was 0.0001, total training epochs were 210, and batch size was 32; Data preprocessing: The images are uniformly cropped to 256×192 pixels and then randomly rotated [-45°, 45°], randomly scaled [0.65, 1.35], and randomly flipped.

[0042] Experimental results show that, compared with current mainstream methods, the present invention achieves comparable or better performance with a significant reduction in the number of parameters and computational complexity. On the MSCOCO dataset, compared with HRNet-W32, the accuracy is improved by 0.1%, the number of parameters is reduced by 32%, and the computational complexity is reduced by 29%. Compared with other small networks, the mAP score is significantly improved, for example, by 10.7% compared with YOLOv5s6-pose and by 9.7% compared with Lite-HRNet-18.

[0043] The above embodiments and accompanying drawings are for illustrative purposes only and do not constitute any limitation. The actual scope of protection of this invention is set forth in the claims. It should be understood that any modifications and changes can be made without departing from the spirit of this invention.

Claims

1. A method for accurate human pose estimation based on a two-stage lightweight framework, characterized in that, include: S1. Collect several human pose images and label key points to construct a training set for human pose estimation. S2. Establish a human pose detection network, input the training set into the human pose detection network for training, and obtain the trained human pose detection network. S3. Obtain the human pose image to be detected and input it into the trained human pose detection network to perform pose detection and obtain the human pose estimation result.

2. The method for accurate human pose estimation based on a two-stage lightweight frame according to claim 1, characterized in that: The human pose detection network includes a Process Net module and a Reprocess Net module connected in sequence. The human pose image is input into the Process Net module for feature extraction to obtain a first-stage feature map. The first-stage feature map is then input into the Reprocess Net module for feature enhancement to obtain the human pose estimation result.

3. The method for accurate human pose estimation based on a two-stage lightweight frame according to claim 2, characterized in that: The Process Net module includes a first stage, a second stage, a third stage, and a fourth stage connected in sequence; The first stage includes the Stem layer. After the human pose image is input into the Stem layer, feature extraction processing is performed to output the initial feature map. The second stage includes swap blocks and parallel high-resolution and medium-resolution branches. The initial feature map output from the first stage is used as the input to the second stage, and is input to the high-resolution and medium-resolution branches of the second stage respectively. The third stage includes a swap block and parallel settings for high-resolution, medium-resolution, and low-resolution branches. The output of the high-resolution branch in the second stage is used as the input of the high-resolution branch in the third stage, and the output of the medium-resolution branch in the second stage is used as the input of the medium-resolution and low-resolution branches in the third stage, respectively. The fourth stage includes a swap block and parallel high-resolution, medium-resolution, low-resolution, and low-resolution branches. The output of the high-resolution branch in the third stage is used as the input of the high-resolution branch in the fourth stage, the output of the medium-resolution branch in the third stage is used as the input of the medium-resolution branch in the fourth stage, and the output of the low-resolution branch in the third stage is used as the input of the low-resolution and low-resolution branches in the fourth stage, respectively. The four branches in the fourth stage correspond to feature maps with output resolutions from high to low. Then, multi-scale fusion is performed on the feature maps output by the four branches in the fourth stage to obtain the feature map of the first stage. In each stage, features of different resolutions are aligned between two branches with different resolutions through the exchange blocks in their respective stages, and the aligned features are then fused.

4. The method for accurate human pose estimation based on a two-stage lightweight frame according to claim 3, characterized in that: The Stem layer includes two sequentially connected feature extraction units. Each feature extraction unit includes a first convolutional layer, a first batch of normalized layers, and a first ReLU activation function, all connected in sequence. The human pose image is used as the input to the Stem layer. After passing through the first convolutional layer, the first batch of normalized layers, and the first ReLU activation function of the first feature extraction unit, it is input to the second feature extraction unit. The second feature extraction unit then processes the image through the first convolutional layer, the first batch of normalized layers, and the first ReLU activation function, and finally outputs the initial feature map.

5. The method for accurate human pose estimation based on a two-stage lightweight frame according to claim 3, characterized in that: Each stage's high-resolution branch, medium-resolution branch, lower-resolution branch, and low-resolution branch includes multiple sequentially connected GCBlock modules. Each GCBlock module aligns and fuses features of different resolutions across branches using a swap block within its respective stage. A GCBlock module comprises a Ghost module, a second batch normalization layer, a second ReLU activation function, a second convolutional layer, a third batch normalization layer, an adaptive fusion operation, and a third ReLU activation function, all connected sequentially. The first GCBlock module in each branch of each stage takes the input of its own branch as input. Subsequent GCBlock modules in each branch take the feature map fused by the previous GCBlock module through the swap block as input. The inputs of each GCBlock module sequentially pass through the Ghost module, the second batch normalization layer, the second ReLU activation function, the second convolutional layer, and the third batch normalization layer to obtain intermediate features. These intermediate features and the inputs of each GCBlock module are then adaptively fused and input into the third ReLU activation function. The output of the third ReLU activation function in each GCBlock module serves as the output of that module.

6. The method for accurate human pose estimation based on a two-stage lightweight frame according to claim 3, characterized in that: The Ghost module includes convolutional layers, linear operations, and concatenation operations. The input of each Ghost module to its respective GCBlock module is used as the input of its respective Ghost module. The input of the Ghost module is processed by the convolutional layer to obtain an intrinsic feature map. Then, the intrinsic feature map is subjected to linear operations to obtain NNN. Finally, the NNN and the intrinsic feature map are concatenated to obtain the output of the Ghost module.

7. The method for accurate human pose estimation based on a two-stage lightweight frame according to claim 2, characterized in that: The Reprocess Net module includes four sequentially connected ECA attention mechanism modules. The first-stage feature map serves as the input to the first ECA attention mechanism module, the output of the previous ECA attention mechanism module serves as the input to the next ECA attention mechanism module, and the output of the last ECA attention mechanism module is the human pose estimation result. Each ECA attention mechanism module includes a globally average pooling layer, a convolutional layer, a sigmoid activation function, and an element-wise multiplication operation connected in sequence. The input of each ECA attention mechanism module is first processed by the globally average pooling layer to obtain a channel-compressed feature map. Then, the channel-compressed feature map is input into the convolutional layer for convolution, and then mapped by the sigmoid activation function to obtain a weight vector. Finally, the weight vector and the input of each ECA attention mechanism module are multiplied element-wise to obtain the output of each ECA attention mechanism module.

8. The method for accurate human pose estimation based on a two-stage lightweight frame according to claim 1, characterized in that: The key points in step S1 include 17 key points of the human body, namely the nose, left and right eyes, left and right ears, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles. The human pose estimation result in step S3 includes the coordinate values ​​of 17 skeletal key points of the human body in a two-dimensional coordinate system.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images