A method and system for estimating human body posture

Abstract

The present invention discloses a method and system for human posture estimation. The method comprises: obtaining a target image to be processed, inputting the target image into a backbone network for human posture estimation for processing to obtain predicted key point coordinates; extracting and fusing the multi-scale feature maps in the feature extraction process of the backbone network to obtain a high-resolution feature map; dividing the high-resolution feature map into two parts according to the channel, deforming the two-dimensional features into one-dimensional features respectively, and aggregating the features using a multi-layer perceptron to obtain the offset feature map H in the x-direction and y-direction. x and H y ; According to the offset feature map H x and H y Calculate the predicted offset O x and O y , and performs offset correction and error compensation on the predicted key point coordinates based on the predicted offset. An improved offset guidance network is added to the backbone network to reduce the error caused by quantization during the offset calculation process, improve the accuracy of offset correction, and enhance prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology, and more specifically, relates to a method and system for estimating human body posture. Background Art

[0002] Human pose estimation is an important and challenging research area in computer vision. Its goal is to use images to locate anatomical key points or key parts of the human body. Human pose estimation has many practical applications, such as human action recognition, human-computer interaction, animation, and film and television production. Research in this area focuses on improving the accuracy of estimating key points or key parts of the human body and increasing the speed of human pose estimation.

[0003] The existing human pose estimation methods with good performance are all based on deep learning methods. The typical method is to input the image into a network. The network is mostly composed of a series of sub-networks from high resolution to low resolution, and finally use deconvolution to improve the output resolution.

[0004] During feature extraction in the backbone network, as resolution decreases from high to low and then to high, quantization error inevitably occurs. This pixel-level error significantly impacts the final result. Common data augmentation methods used in human pose estimation, such as cropping, scaling, rotating, and folding, can introduce sub-pixel errors due to the difference between image coordinates and pixel coordinates. These errors also exhibit certain patterns. The discrepancy between predicted and true keypoints significantly impacts prediction accuracy. Summary of the Invention

[0005] In response to the defects of the related art, the purpose of the present invention is to provide a method and system for human posture estimation, which aims to solve the problem of prediction accuracy error caused by the deviation between the predicted key points and the actual key points in the existing human posture estimation methods.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for estimating a human body posture, comprising:

[0007] Obtaining a target image to be processed, and inputting the target image into a backbone network for human posture estimation for processing to obtain predicted key point coordinates;

[0008] Extract and fuse the multi-scale feature maps in the feature extraction process of the backbone network to obtain a high-resolution feature map;

[0009] The high-resolution feature map is divided into two parts according to the channel, and the two-dimensional features are transformed into one-dimensional features respectively. The multi-layer perceptron is used to aggregate the features to obtain the offset feature maps H in the x-direction and y-direction. x and H y ;

[0010] According to the offset feature map H x and H y Calculate the predicted offset O x and O y , and perform offset correction and error compensation on the predicted key point coordinates according to the predicted offset to obtain corrected key point coordinates.

[0011] Optionally, the high-resolution feature map is divided into two parts according to the channel, the two-dimensional features are deformed into one-dimensional features, and the multi-layer perceptron is used to aggregate the features to obtain the offset feature maps H in the x-direction and y-direction x and H y ,include:

[0012] The high-resolution feature map S0 is divided into two parts according to the number of channels, namely S x and S y ,where K is the number of key points on the human body;

[0013] The convolution layer is used to change the resolution size, and the two-dimensional features of each channel are split and deformed into one dimension. The multi-layer perceptron is used to aggregate the one-dimensional features and transform S x and S y Transformed into H x and H y ,in, W and H are the width and height of the image input to the network.

[0014] Optionally, the offset feature map H x and H y Calculate the predicted offset O x and O y ,include:

[0015] To predict key point coordinates As the center, r is the interval radius to select the offset feature map H x and H y The offset feature h in x and h y ;

[0016] The offset feature h x and h y Input linear layer, the resulting offset prediction is and The offset labels are:

[0017]

[0018] Optionally, performing offset correction and error compensation on the predicted key point coordinates according to the predicted offset to obtain corrected key point coordinates includes:

[0019] The predicted key point coordinates are corrected for offset and compensated for errors according to the predicted offset:

[0020]

[0021] Among them, (x f ,y f ) is the corrected key point coordinate, To predict the coordinates of key points;

[0022] r is the calculation range of the offset, and R is the calculation range of the offset feature, satisfying:

[0023] r x =γR x , 0<γ<1;

[0024] δ is the offset correction error, and its value is:

[0025]

[0026] Optionally, before acquiring the target image to be processed and inputting the target image into a backbone network for human posture estimation for processing to obtain predicted key point coordinates, the method further includes:

[0027] The total loss function is used to supervise the predicted heat map, predicted offset feature map and predicted offset output by the backbone network; wherein the total loss function is the loss function of the backbone network Loss function L for predicting offset feature maps H And the loss function L of the predicted offset o sum.

[0028] Optionally, the loss function of the backbone network is According to the predicted heat map output by the backbone network The heat map S generated by the real key point coordinates is calculated, and the loss function in the backbone network training process is expressed as:

[0029]

[0030] Among them, the real key point coordinates are expressed as z k =(x k ,y k ), K is the number of key points of the human body.

[0031] Optionally, the loss function L for predicting the offset feature map H for:

[0032]

[0033] Among them, the offset feature map in the x direction Offset feature map in the y direction The number of channels is K. and are the predicted offset feature maps in the x and y directions, respectively, H kx and H ky is the label of the predicted offset feature map, α1 and β1 are the parameters for adjusting the loss function of the predicted offset feature;

[0034] The predicted offset feature map label H kx and H ky They are:

[0035]

[0036]

[0037] Among them, the predicted position of each key point is The true position of the key point is (x k ,y k ), R is the calculation range of the predicted offset feature.

[0038] Optionally, the loss function L of the predicted offset o According to the predicted offset O x and O y The loss function of the predicted offset is calculated as follows:

[0039]

[0040] Among them, α2 and β2 are parameters for adjusting the loss function, and are the predicted offsets in the x and y directions respectively.

[0041] Optionally, extracting and fusing the multi-scale feature maps in the feature extraction process of the backbone network to obtain a high-resolution feature map includes:

[0042] Select multiple feature maps of different resolutions containing high-level semantic information in the last stage of the backbone network;

[0043] According to the resolution size, the lowest resolution feature map is upsampled to the resolution size of the higher-level feature through deconvolution, connected in series with the higher-level feature, and then the number of channels is reduced to the number of channels of the next-level feature through convolution. This process is repeated until features of all scales are fused to obtain a high-resolution feature map.

[0044] In a second aspect, the present invention further provides a system for estimating human body posture, comprising:

[0045] The key point prediction module is used to input the target image, extract the features of the target image, and obtain the predicted key point coordinates;

[0046] The multi-scale feature fusion module is used to extract and fuse the multi-scale feature maps in the feature extraction process of the backbone network to obtain high-resolution feature maps;

[0047] The offset feature map prediction module is used to divide the high-resolution feature map into two parts according to the channel, transform the two-dimensional features into one-dimensional features, and use the multi-layer perceptron to aggregate the features to obtain the offset feature map H in the x-direction and y-direction x and H y ;

[0048] The offset correction module is used to convert the offset feature map H x and H y The offset guidance network is input and the predicted offset is output. The predicted key point coordinates are offset corrected and error compensated according to the predicted offset to obtain the corrected key point coordinates.

[0049] In general, the above technical solutions conceived by the present invention have the following advantages compared with the prior art:

[0050] (1) The method for human posture estimation proposed in the present invention adds an independent improved offset guidance network to the existing human posture estimation method based on heat map, aggregates the multi-scale features in the posture estimation network to perform offset correction on the prediction results, and effectively reduces the quantization error of the human posture estimation method based on heat map.

[0051] (2) The improved offset-guided network structure provided by the present invention integrates multi-scale features on the basis of the original one. In the fusion stage, the intermediate feature maps are gradually fused from bottom to top while upsampling is performed using deconvolution. In the resolution improvement stage, the feature map resolution is no longer directly increased to the original image resolution, but is converted into a one-dimensional feature map. This reduces the computational complexity without affecting accuracy, thus improving efficiency.

[0052] (3) The offset calculation branch network added to the improved offset guidance network provided by the present invention avoids the accuracy loss caused by coordinate quantization, and integrates the distribution of offset features into the offset calculation, which can improve the prediction accuracy of the offset.

[0053] (4) The method of human posture estimation based on the improved offset guidance network proposed in the present invention improves the correction method and introduces a calculation radius parameter. Even in the case of large deviations, the predicted value can be effectively corrected to the true value, and the theoretical accuracy of the correction is higher; and the correction error parameter is introduced, which can effectively reduce the error caused by quantization in the calculation offset process. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A schematic flow chart of a method for estimating human body posture provided by the present invention;

[0055] Figure 2 Schematic diagram of the backbone network structure used in the present invention;

[0056] Figure 3 Schematic diagram of the improved offset guidance network structure for fusing multi-scale features of the present invention;

[0057] Figure 4 Schematic diagram of the overall structure combining the backbone network with the offset guidance network;

[0058] Figure 5 (a) and (b) are the heat map labels generated by the real coordinates of the image and the predicted heat map generated by the backbone network according to the input target image;

[0059] Figure 6 (a), (b), (c) and (d) are the change curves of loss and accuracy in the training set and validation set respectively. DETAILED DESCRIPTION

[0060] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is 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 intended to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0061] The contents involved in the above embodiment are described below in conjunction with a preferred embodiment.

[0062] like Figure 1 As shown, the present invention provides a method for estimating human body posture, comprising:

[0063] S1. Obtain the target image to be processed, and input the target image into the backbone network for human posture estimation for processing to obtain the predicted key point coordinates;

[0064] S2, extract and fuse the multi-scale feature maps in the feature extraction process of the backbone network to obtain a high-resolution feature map;

[0065] S3, divide the high-resolution feature map into two parts according to the channel, transform the two-dimensional features into one-dimensional features, and use the multi-layer perceptron to aggregate the features to obtain the offset feature map H in the x-direction and y-direction x and H y ;

[0066] S4, according to the offset feature map H x and H y Calculate the predicted offset O x and O y , and perform offset correction and error compensation on the predicted key point coordinates according to the predicted offset to obtain the corrected key point coordinates.

[0067] The existing arbitrary human posture detection network is used as the prediction backbone network, and an independent improved offset guidance network is added on the basis of the backbone network. Taking the more popular heat map-based human posture estimation network as an example, the real coordinates of the key points are expressed as Among them, z k =(x k ,y k ), K is the number of key points, and the heat map S generated according to the real coordinates of the key points is used as the label. The heat map S is as follows Figure 5 As shown in (a), the generation method is:

[0068] S k (p)~N(z k ,Σ)

[0069] in, represents the position coordinates, and Σ represents the variance matrix.

[0070] The target image to be processed is input into the backbone network, and the predicted heat map output by the backbone network is The predicted key point coordinates calculated by the output prediction heat map are The calculation method is:

[0071]

[0072] During the heatmap prediction process in the backbone network, feature maps of varying sizes and resolutions are generated. The basic offset-guided network approach simply adds an offset network to the output of the backbone network, lacking scale information. This embodiment provides an independent, improved offset-guided network that integrates feature maps of multiple resolutions from the backbone network, extracting multi-scale information from the network.

[0073] For example, a high-resolution network structure such as Figure 2 As shown, in this example, a 4-stage high-resolution network (HRNet) is used, and the image input of the network is The output heat map is In order to reduce the amount of calculation, the output heat map resolution is generally smaller than the original image.

[0074] The independent improved offset-guided network structure integrates the multi-scale features in the backbone network, selects intermediate feature maps of different resolutions in the backbone network, and increases the resolution of different feature maps to the resolution of the final output heat map with K channels. The deconvolution layer is then used to increase the feature map to the resolution of the network input image to obtain a high-resolution feature map with the number of channels adjusted to 2K, where K is the number of key points.

[0075] Among them, S2 specifically includes:

[0076] S21. Select multiple feature maps of different resolutions containing high-level semantic information in the last stage of the backbone network.

[0077] The feature map in the backbone network is denoted as s stands for stage and r stands for resolution index. The resolution is represented by w, the number of channels is represented by c, The resolution of the first sub-network in this stage of The number of channels is 2 r -1 ,In this implementation, the four feature maps of the final stage are selected,

[0078] S22. According to the resolution size, the feature map with the lowest resolution is upsampled to the resolution size of the higher-level feature through deconvolution, and then connected in series with the higher-level feature. Then, the number of channels is reduced to the number of channels of the next-level feature through convolution, and this process is repeated until the features of all scales are fused to obtain a high-resolution feature map.

[0079] Specifically: Use the improved offset guidance network to fuse the multi-resolution feature maps, and improve the offset guidance network structure as follows: Figure 3 As shown. Upsample to a resolution and number of channels comparable to Same as The stacked feature map is represented as S3, and the resolution and number of channels are S3 upsamples to the same resolution and number of channels as Same as The stacked feature map is represented as S2, and the resolution and number of channels are S2 upsamples to the same resolution and number of channels as Same as The stacked feature map is denoted as S1, with a resolution and number of channels of (w, 2c). S1 incorporates the network's multi-scale intermediate feature information. S1 is upsampled to increase the resolution and reduce the number of channels, resulting in a high-resolution feature map S0. The resolution and number of channels are adjusted to (2w, 2K), where K is the number of key points to be detected.

[0080] After obtaining the high-resolution feature map, in order to avoid the inefficiency caused by redundant information in the deviation calculation, the high-resolution feature map is deformed and divided into two parts according to the channel. The number of channels in each part is K, and the offset feature map H in the x-direction and y-direction is obtained. x and H y S3 specifically includes:

[0081] S31, the high-resolution feature map S0 is divided into two parts according to the number of channels, namely S x and S y ,K is the number of key points on the human body;

[0082] S32, use the convolution layer to change the resolution size, and split the two-dimensional features of each channel into one dimension, use the multi-layer perceptron to aggregate the length of the one-dimensional features, and convert S x and S y Transformed into H x and H y ,in, W and H are the width and height of the image input to the network.

[0083] In the process of calculating the offset correction using the offset feature map, the quantization error of the predicted key point coordinates will cause a loss of accuracy. In order to solve the error caused by the quantization problem in the offset calculation, an offset calculation prediction branch is introduced in the improved offset guidance network to predict the offset.

[0084] Output prediction heatmap of the backbone network The key point coordinates are like Figure 5 As shown in (b), the offset feature map H x and H y Input the predicted offset calculation branch of the improved offset guidance network, and perform offset correction on the key point coordinates according to the output predicted offset. The corrected key point coordinates are (x f ,y f ).

[0085] Wherein, step S4 specifically includes:

[0086] S41, to predict key point coordinates As the center, r is the interval radius to select the offset feature map H x and H yThe offset feature h in x and h y ;

[0087] S42, the offset feature h x and h y Input linear layer, the resulting offset prediction is and The offset labels are:

[0088]

[0089] S43, performing offset correction and error compensation on the predicted key point coordinates according to the predicted offset:

[0090]

[0091] Among them, (x f ,y f ) is the corrected key point coordinate, To predict the coordinates of key points;

[0092] r is the calculation range of the offset, and R is the calculation range of the offset feature, satisfying:

[0093] r x =γR x , 0<γ<1

[0094] δ is the offset correction error, and its value is:

[0095]

[0096] In the technical solution of this embodiment, an independent improved offset guidance network is added to the existing human posture estimation network, and the multi-scale feature information in the backbone network is integrated to improve the accuracy of offset correction. In the process of offset calculation, repeated features are removed, and the two-dimensional offset features are reduced to one dimension, thereby improving efficiency. An offset prediction branch is also added to the improved offset guidance network, and the offset features are calculated using a linear network to obtain an accurate offset for correcting the prediction of the backbone network, which can effectively reduce the influence of quantization error. The original calculation method only calculates the offset by summing the values ​​in the offset feature map. The introduced linear layer has learnable parameters, can learn high-level feature information, and can integrate the distribution characteristics of the offset feature map into the calculation process, thereby improving the accuracy of offset correction. Compared with the existing human posture estimation method based on heat maps, the method of the present invention can use offset correction to compensate for the quantization error caused by downsampling, and the introduction of multi-scale feature information improves the adaptability of offset correction and improves prediction accuracy. The improved correction method introduces a calculation radius parameter, which can effectively correct the predicted value to the true value even in the case of large deviation, and the theoretical accuracy of the correction is higher. The introduction of the correction error parameter can effectively reduce the error caused by quantization in the calculation offset process.

[0097] Furthermore, the improved offset-guided network adopts an independent network structure, which can be integrated with other heat map-based two-stage human pose estimation methods, and can also be easily transferred to other frameworks to solve the problem of quantization error caused by resolution transformation in other deep learning machine vision.

[0098] Based on the above embodiment, optionally, before acquiring the target image to be processed and inputting the target image into the backbone network for human posture estimation for processing to obtain the predicted key point coordinates, the method further includes:

[0099] The total loss function is used to supervise the predicted heat map, predicted offset feature map and predicted offset output by the backbone network.

[0100] Among them, the total loss function and the loss function of the backbone network Loss function L for predicting offset feature maps H And the loss function L of the predicted offset o The relationship is expressed as:

[0101] Improve the overall structure of the offset guidance network as follows Figure 4 As shown in the figure, during the training process, the label of the backbone network is generated by the real coordinates of the key points of the dataset into a heat map S, which is composed of S and The calculated loss function is The labels of the improved offset guidance network are generated by the real coordinates of the key points of the dataset, denoted as Hx and H y , by H and The calculated loss function is L H The label of the offset feature map offset calculation is generated by the predicted coordinates and the real coordinates of the key points. The loss function of the offset prediction is L o .

[0102] Optionally, the loss function of the backbone network is According to the predicted heat map output by the backbone network The heat map S generated by the real key point coordinates is calculated, and the loss function in the backbone network training process is expressed as:

[0103]

[0104] Among them, the prediction heat map The key point coordinates are expressed as The real key point coordinates are expressed as z k =(x k ,y k ), K is the number of key points of the human body.

[0105] For each key point position predicted by the backbone network and its corresponding real position (x k ,y k ), predict the label H of the offset feature map kx and H ky They are:

[0106]

[0107]

[0108] Where R is the calculation range of the predicted offset feature.

[0109] Loss function L for predicting feature maps H for:

[0110]

[0111] Among them, the offset feature map in the x direction Offset feature map in the y direction The number of channels is K. and They are the predicted offset feature maps in the x and y directions, H kx and H ky is the label of the predicted offset feature map, and α1 and β1 are the parameters for adjusting the loss function of the predicted offset feature.

[0112] Optionally, a loss function L for predicting the offset o According to the predicted offset O x and O y The loss function of the offset prediction is calculated as follows:

[0113]

[0114] Among them, α2 and β2 are the parameters for adjusting the loss function, and are the predicted offsets in the x and y directions respectively.

[0115] The loss function is used to perform back propagation of gradients during training, making the supervised features closer to the actual features.

[0116] To verify the effectiveness of the present invention, the COCO dataset is selected for training and testing. The COCO dataset is a large-scale object detection dataset, which is mainly captured from complex daily scenes.

[0117] The evaluation index is based on the Object Keypoint Similarity (OKS)

[0118]

[0119] where d i To predict key points The actual coordinate z of the corresponding key point k The Euclidean distance, δ(v i ) is the key point visual weight function, s is the target scale parameter, k i Is a constant used to control attenuation. Reporting indicators use precision AP and recall AR, AP 50 is the accuracy when OKS=0.50, AP M For medium-sized targets, AP L is the average precision of large-sized objects, and AR is the average recall.

[0120] The model was trained on Ubuntu 18.04 using the pytorch 1.7 framework for 210 epochs. The loss and accuracy curves of the model on the training set and validation set during training are shown in the following figure. Figure 6 As shown in (a), (b), (c) and (d).

[0121] The comparison results of the models on the test set are shown in Table 1. The improved offset guidance network and offset calculation correction method proposed in the present invention have significantly improved the accuracy compared with the heat map method with only the backbone network. The increase in the number of parameters and computing power is not high. The method provided by the present invention has certain practical value.

[0122] Table 1

[0123]

[0124] Based on the above embodiments, the present invention further provides a system for estimating human body posture, comprising:

[0125] The key point prediction module is used to input the target image, extract the features of the target image, and obtain the predicted key point coordinates;

[0126] The multi-scale feature fusion module is used to extract and fuse the multi-scale feature maps in the feature extraction process of the backbone network to obtain high-resolution feature maps;

[0127] The offset feature map prediction module is used to divide the high-resolution feature map into two parts according to the channel, transform the two-dimensional features into one-dimensional features, and use the multi-layer perceptron to aggregate the features to obtain the offset feature map H in the x-direction and y-direction x and H y ;

[0128] The offset correction module is used to convert the offset feature map H x and H y The offset guidance network is input and the predicted offset is output. The predicted key point coordinates are offset corrected and error compensated according to the predicted offset to obtain the corrected key point coordinates.

[0129] The human body posture estimation system provided by the embodiment of the present invention can execute the human body posture estimation method provided by any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0130] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for estimating human body posture, characterized in that: include: Obtain a target image to be processed, and input the target image into the backbone network for processing to obtain predicted key point coordinates; Extract and fuse the multi-scale feature maps in the feature extraction process of the backbone network to obtain a high-resolution feature map; The high-resolution feature map is divided into two parts according to the channel, the two-dimensional features are transformed into one-dimensional features, and the multi-layer perceptron is used to aggregate the features to obtain Direction and Directional offset feature map and ; To predict key point coordinates As the center, Select the offset feature map for the interval radius and Offset feature in and ; The offset feature and Input linear layer, the resulting offset prediction is and The offset labels are: ； The predicted key point coordinates are corrected for offset and compensated for errors according to the predicted offset: ； Among them, (x f ,y f ) is the corrected key point coordinate, To predict the coordinates of key points; is the calculation range of the offset, is the calculation range of the offset feature, satisfying: ； is the offset correction error, and its value is: 。 2. The method according to claim 1, wherein The high-resolution feature map is divided into two parts according to the channel, and the two-dimensional features are transformed into one-dimensional features respectively. The multi-layer perceptron is used to aggregate the features to obtain the offset feature maps H in the x-direction and y-direction. x and H y ,include: The high-resolution feature map S0 is divided into two parts according to the number of channels, namely S x and S y ,where K is the number of key points on the human body; The convolution layer is used to change the resolution size, and the two-dimensional features of each channel are split and deformed into one dimension. The multi-layer perceptron is used to aggregate the one-dimensional features and transform S x and S y Transformed into H x and H y ,in, W and H are the width and height of the image input to the network.

3. The method according to claim 1, wherein Before obtaining the target image to be processed and inputting the target image into the backbone network for processing to obtain the predicted key point coordinates, the method further includes: The total loss function is used to supervise the predicted heat map, predicted offset feature map and predicted offset output by the backbone network; wherein the total loss function is the loss function of the backbone network , loss function for predicting offset feature maps And the loss function of the predicted offset sum.

4. The method according to claim 3, wherein The loss function of the backbone network According to the predicted heat map output by the backbone network Heatmap generated from the true keypoint coordinates It is calculated that the loss function during the backbone network training process is expressed as: Among them, the real key point coordinates are expressed as , , is the number of key points of the human body.

5. The method according to claim 3, wherein Loss function for predicting offset feature maps for: in, Directional offset feature map , Directional offset feature map , the number of channels is , and Respectively in Direction and The predicted offset feature map of the direction, H kx and H ky To predict the label of the offset feature map, and To adjust the parameters of the prediction offset feature loss function; The predicted offset feature map label H kx and H ky They are: Among them, the predicted position of each key point is , the true position of the key point is , The calculation range for the prediction offset feature.

6. The method according to claim 3, wherein The loss function L of the prediction offset o According to the predicted offset O x and O y The loss function of the predicted offset is calculated as follows: Among them, α2 and β2 are parameters for adjusting the loss function, and are the predicted offsets in the x and y directions respectively.

7. The method according to claim 1, wherein The multi-scale feature maps in the feature extraction process of the backbone network are subjected to feature extraction and fusion to obtain a high-resolution feature map, including: Select multiple feature maps of different resolutions containing high-level semantic information in the last stage of the backbone network; According to the resolution size, the lowest resolution feature map is upsampled to the resolution size of the higher-level feature through deconvolution, connected in series with the higher-level feature, and then the number of channels is reduced to the number of channels of the next-level feature through convolution. This process is repeated until features of all scales are fused to obtain a high-resolution feature map.

8. A system for estimating human body posture, characterized in that: include: The key point prediction module is used to input the target image, extract the features of the target image, and obtain the predicted key point coordinates; The multi-scale feature fusion module is used to extract and fuse the multi-scale feature maps in the feature extraction process of the backbone network to obtain high-resolution feature maps; The offset feature map prediction module is used to divide the high-resolution feature map into two parts according to the channel, transform the two-dimensional features into one-dimensional features, and use the multi-layer perceptron to aggregate the features to obtain Direction and Directional offset feature map and ; Offset correction module, used to predict key point coordinates As the center, Select the offset feature map for the interval radius and Offset feature in and ; The offset feature and Input linear layer, the resulting offset prediction is and The offset labels are: ； The predicted key point coordinates are corrected for offset and compensated for errors according to the predicted offset: Among them, (x f ,y f ) is the corrected key point coordinate, To predict the coordinates of key points; r is the calculation range of the offset, and R is the calculation range of the offset feature, satisfying: r x =γR x ,0<γ<1; δ is the offset correction error, and its value is:

Images