An Efficient Human Pose Estimation Method Based on Transformer Token Masking

Abstract

The present invention discloses an efficient human posture estimation method based on Transformer token masking. The network model TMPose adopted has excellent performance and can enable the network to complete high-quality human posture estimation tasks very efficiently. The network model TMPose of the present invention includes a backbone network, a token generator, a Transformer module and a heat map regression module. It has great scientific research value and practical application value. Since the token masking method has a certain generalization, it can be easily combined with other network models to inspire more work. In addition, since TMPose greatly reduces the computational complexity while maintaining a high level of accuracy, it can be efficiently applied to more scenarios in actual deployment.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision and human posture estimation, and in particular relates to an efficient human posture estimation method based on Transformer token masking. Background Art

[0002] In recent years, visual Transformers have made great progress, with representative works including image classification, object detection, and semantic segmentation. Although these models have guaranteed accuracy and surpass CNN-based models, they all come at a huge computational cost.Therefore, researchers have begun to propose many algorithms to improve the efficiency of the model, such as pruning, distillation and quantization in model compression; in addition to these traditional methods, some researchers have also proposed pruning tokens. In the field of model pre-training, many efficient self-supervised learning methods have achieved good results. For example, the literature [He, Kaiming, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. "Masked autoencoders are scalable vision learners." In Proceedings of the IEEE / CVFConference on Computer Vision and Pattern Recognition, pp. 16000-16009.2022] proposed a self-supervised model of the encoder-decoder structure, which supervises model training by randomly masking patches in the image as labels; the literature [Li, Yanghao, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. "Scaling language-image pre-training via masking." In Proceedingsof the IEEE / CVF Conference on Computer Vision and Pattern Recognition, pp. 23390-23400.2023], the image patches are masked and only the visible patches are encoded, and then the model is trained by comparative learning with text samples; in the literature [Bao, Hangbo, Li Dong, Songhao Piao, and FuruWei. "Beit: Bert pre-training of image transformers." arXiv preprint arXiv: 2106.08254 (2021)], the authors proposed a masked image modeling method to pre-train the ViT model. By masking the image patch at the pixel level and the visual token at the semantic level, the ViT model can be efficiently trained. After the model is pre-trained, it returns to the downstream task.

[0003] The idea of ​​token pruning has also been reflected in many works. Specifically: in Token-to-tokens, the number of tokens is reduced by concatenating adjacent tokens into one token; in DynamicViT, tokens are pruned through a learnable token selector; in PPT [Ma, Haoyu, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. "PPT: token-Pruned Pose Transformer for monocular and multi-view human pose estimation." In Computer Vision–ECCV 2022:17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part V, pp. 424-442. Cham: Springer Nature Switzerland, 2022], the authors added human token recognition (HTI) and pruned the number of tokens through the attention operation. Although these methods have achieved impressive results, they essentially add additional learnable parameters to the Transformer to reduce the number of tokens and thus reduce the complexity of the model.

[0004] From the above, we can see that Transformer's work in the field of vision is developing rapidly. The same is true for the task of human pose estimation, and many excellent works have emerged: TransPose introduced Transformer into human pose estimation for the first time. The model first extracts features through the backbone network, then converts the feature map into Token form and sends it to Transformer, and finally predicts the key points through the heat map method. The author achieved the SOTA effect at the time through such a simple framework. Then TFPose removed the heat map method on the basis of TransPose and adopted the method of directly regressing the key points. TokenPose added keypoint token on the basis of TransPose, in order to enable the network to learn the constraint information between joint points. In such a framework, DPIT integrates top-down and bottom-up methods through a parallel structure. A new backbone network is proposed in HRFormer to achieve the effect of HRNet through pure Transformer.

[0005] In recent years, Transformer has achieved remarkable results in the field of vision. Its built-in attention layer can well model the global dependencies of images and convert image features into token form as input of Transformer. This paradigm has been deeply rooted in people's minds, but it comes with high computational costs. Summary of the Invention

[0006] In view of the above, the present invention provides an efficient human posture estimation method based on Transformer token mask. The network model TMPose adopted has excellent performance, which enables the network to complete high-quality human posture estimation tasks very efficiently.

[0007] An efficient human pose estimation method based on Transformer token masking includes the following steps:

[0008] (1) Obtain a large number of pictures containing people and mark the joint positions of the people in the pictures;

[0009] (2) Construct the TMPose network model framework, which includes the backbone network, token generator, Transformer module and heat map regression module, where:

[0010] The backbone network is used to extract features from the input image;

[0011] The token generator is used to convert the feature map extracted by the backbone network into a one-dimensional sequence form and generate tokens through partial masking;

[0012] The Transformer module is used to encode the input Token, thereby outputting M one-dimensional sequences, where M is the number of labeled joint points;

[0013] The heatmap regression module is used to reshape the M one-dimensional sequences to generate M heatmaps, from which the prediction results of the position coordinates of each joint point are decoded;

[0014] (3) Using the images and their annotation information obtained in step (1) to train the above model framework;

[0015] (4) Input the image containing the person to be estimated into the trained model, and the position coordinates of each joint point of the person in the image can be predicted.

[0016] Furthermore, to balance the Transformer's overhead, the backbone network only retains the first three stages of HRNet, with only 25% of the original parameters. The specific operation is: first, the input image is cropped to a uniform size, and then feature extraction is performed through a deep convolutional neural network to generate a feature map with a size of 1 / 4 of the original image.

[0017] Furthermore, the specific operation process of the token generator is as follows:

[0018] First, the feature map extracted by the backbone network is divided into grids, flatten each grid into a size of P h ×P w ×C one-dimensional sequence, and then the flattened sequence is mapped into visual tokens through a linear layer. At the same time, M learnable key point tokens are randomly initialized and generated. The sequence length is consistent with the visual token, where H, W, and C are the height, width, and channel dimensions of the feature map respectively;

[0019] Then, the visual tokens are masked into a size of l×(M·R) by a mapping function. The mapping functions used in the masking process include random sampling, uniform sampling, and patch sampling. The ratio R is used is 0.4, 0.6, or 0.8, where L and M are the sequence length and number of visual tokens, respectively.

[0020] Finally, the visual token is positionally encoded, that is, the position of each sequence in the visual token is marked using two-dimensional position embedding, and then the key point token and the position-encoded visual token are concatenated together to obtain a token, which is then sent to the Transformer module.

[0021] Furthermore, the Transformer module is composed of a cascade of multiple layers of encoding structures, and each layer of the encoding structure is composed of a regularization layer L1, a multi-head self-attention mechanism layer, a regularization layer L2, and a multi-layer perceptron MLP connected in sequence, wherein the input of the regularization layer L2 is the output of the multi-head self-attention mechanism layer superimposed with the input of the regularization layer L1, and the output of the multi-layer perceptron MLP is superimposed with the input of the regularization layer L2 as the output of the encoding structure.

[0022] Furthermore, the multi-head self-attention mechanism layer is composed of multiple Self-Attentions, each of which combines the input of the multi-head self-attention mechanism layer with the linear transformation matrix W Q 、W K and W VMultiplying them gives the query vector matrix Q, key vector matrix K, and value vector matrix V. Then, the output of Self-Attention is calculated using the following formula. Finally, the outputs of multiple Self-Attention layers are concatenated and passed through a linear layer to obtain the output of the multi-head self-attention mechanism layer.

[0023]

[0024] Among them: Attention(Q,K,V) is the output result of Self-Attention, s is the number of columns of the query vector matrix Q, T Indicates transpose.

[0025] Furthermore, the heatmap regression module only takes M keypoint tokens as output, and then remaps them back to a two-dimensional heatmap form through a multi-layer perceptron. Specifically, the output X of the Transformer module is fed into the multi-layer perceptron, where X is first mapped back to a two-dimensional form P, whose size is M×H * ×W * , H * and W * The height H and width W of the original image are 1 / 4 respectively. Finally, P is reshaped into M heat maps, and the heat map size is the same as the original Figure 1 On this basis, the coordinates of each joint of the human body are located by finding the maximum response position on the heat map.

[0026] Furthermore, the specific implementation of step (3) is as follows:

[0027] 3.1 Initialize model parameters, including the bias vector and weight matrix of each layer, learning rate, and optimizer;

[0028] 3.2 Input the image containing the person into the model, and the model forward propagates the output to obtain the corresponding prediction result, that is, the position coordinates of the joint points, and calculate the loss function between the prediction result and the annotation information;

[0029] 3.3 Based on the loss function, the optimizer is used to iteratively update the model parameters through the gradient descent method until the loss function converges and the training is completed.

[0030] Furthermore, the loss function is expressed as follows:

[0031]

[0032] Where: L MSE is the loss function, P k The heat map generated for the real coordinates of the k-th joint point in the annotation information, Predict the heat map of the k-th joint point for the model.

[0033] The proposed network model, TMPose, has significant scientific research and practical application value. Due to the generalizability of the token masking method, it can be easily combined with other network models, inspiring further research. Furthermore, TMPose significantly reduces computational complexity while maintaining a high level of accuracy, making it highly applicable to a wider range of scenarios in practical deployments. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a structural diagram of the TMPose network model framework of the present invention.

[0035] Figure 2 This is a schematic diagram comparing the computational complexity of the TMPose of the present invention and the mainstream Pose Transformer, with TMPose on the left and Pose Transformer on the right.

[0036] Figure 3 Schematic diagram comparing the computational overhead of the TMPose framework of the present invention and mainstream frameworks.

[0037] Figure 4 This is the joint point heat map predicted by the TMPose network of the present invention.

[0038] Figure 5 This is a visualization of the prediction results of TMPose on the COCO Validation dataset.

[0039] Figure 6 This is a comparison chart of the visual prediction results of TMPose in the present invention and other lightweight human pose estimation networks on the COCO Validation dataset. DETAILED DESCRIPTION

[0040] In order to describe the present invention more specifically, the technical solution of the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] The present invention provides an efficient human posture estimation method based on Transformer token masking, comprising the following steps:

[0042] (1) Obtain a large number of pictures containing people and mark the joint positions of the people in the pictures.

[0043] In this implementation, we use images from the COCO dataset for model training. COCO is a large dataset proposed by Microsoft that can be used for image classification, object detection, semantic segmentation, and pose estimation tasks. It is one of the most influential and widely used datasets in the field of vision, and the images it contains cover the vast majority of daily scenes. For each person in the image, 17 joints are marked. The dataset contains a total of 200K images and 250K human instances, of which 118K images are used for training and 5K images are used for verification. The evaluation indicators used are average precision AP and average recall AR, and AP and AR are calculated using OKS (Object Keypoint Similarity). OKS is the similarity between the true value and the predicted value, and the calculation formula is as follows:

[0044]

[0045] Where: i represents the i-th joint point, d i Represents the Euclidean distance between the true value and the predicted value, s is the scale of the target, v i Indicates whether the target is visible, and δ represents the regularization parameter of the key point.

[0046] The calculation formulas for AP and AR are as follows:

[0047]

[0048] Where: p is the number of the human detection frame, when t is 0.5 and 0.75, we record it as AP 50 and AP 75 , when 32 2 2 <96 2 or 96 2 2 When we denote it as AP m and AP l ; When t is 0.5, 0.55,…, 0.9, 0.95, we record the average value as AP. Similarly, AR can also be divided.

[0049] (2) Build the TMPose model architecture, such as Figure 1 As shown in Figure 3, the architecture consists of a backbone network, a token generator, a Transformer module, and a heatmap regression module.

[0050] ​​To balance the Transformer's overhead, we only retain the first three stages of HRNet as the backbone network, named HRNet-s, with only 25% of the original parameters. Specifically, the network first crops the input single-person image to a uniform size, then extracts features through a deep convolutional neural network to generate a feature map that is 1 / 4 the size of the original image. In TMPose, our visual tokens are obtained from these feature maps rather than from image patches. We use CNN to effectively extract low-level features of the image.

[0051] Since Transformer is a sequence-to-sequence network, it is necessary to first map the two-dimensional feature map into a one-dimensional sequence. This paper follows the idea of ​​ViT and assumes that the feature map output by the backbone network is x∈R H×W×C , we divide it into grids, and then flatten each grid into a size of P h ×P w ×C one-dimensional sequence, and then the flattened sequence is mapped to the visual token we need through a linear layer; at the same time, the network randomly initializes to generate M learnable key point tokens, whose sequence length is consistent with the visual token.

[0052] Then assume that the size of the visual token is v∈R L×M , where L is the sequence length and M is the number of sequences. We mask it into v′∈R through the mapping function L×(M·R) ,f:v→v′∈R L×(M·R) In this invention, we adopt three different mapping functions f, namely random sampling, uniform sampling, and patch sampling, and also adopt three different R (ratio) values ​​of 0.4, 0.6, and 0.8.

[0053] Since the self-attention mechanism in Transformer has no position awareness, it is necessary to position encode the newly generated visual tokens, and the two-dimensional position embedding pe i Each sequence of input can be marked with a position {visual token} = {v′1+pe1,v′2+pe2,…,v′ H +pe H Where H is the number of sequences H = M·R. Next, the key point tokens are concatenated with the position-encoded visual tokens and sent to the transformer for training. Finally, the sequence output by the transformer network is used to predict the key points using the heat map method.

[0054] The Transformer module consists of a cascade of N layers of encoding structures, each of which is composed of a regularization layer, a multi-head self-attention mechanism, and a multi-layer perceptron connection. Specifically, the input sequence is R3=R c +R s +R k , L = 192, first through the matrix Project X3 to query key value The calculation is as follows:

[0055] Q=X3·W Q , K=X·W K , V=X3·W V

[0056] Then the self-attention score matrix SA (self-attention) can be expressed as:

[0057]

[0058] The multi-head self-attention mechanism (MSA) is an extension of the self-attention mechanism with h self-attention operations, where h is called head. In MSA, the sequence length L h It is often set to L / h, and the specific calculation formula is as follows:

[0059] MSA(X3)=[SA1(X3);SA2(X3);…;SA h (X3)]W P

[0060] Among them: X3 is the input, The obtained attention score determines the degree of attention of the current query to the embedding; after processing through N layers of Transformer, the network outputs M embeddings, and the size of M depends on the number of joint points.

[0061] The heatmap regression module only takes M key point tokens as output at the output stage of the Transformer, and then remaps them back to a two-dimensional heatmap through a multi-layer perceptron. Specifically, assuming that the output of the Transformer is X∈R M×H Where M is the number of sequences and H is the length of the sequence. Then X is fed into a multi-layer perceptron, where it is first mapped back to a two-dimensional generator P∈R M×H*×W* , where H * ,W * are 1 / 4 of the original image size, and then reshaped into P∈R M×H×W; The heat map size at this time is the same as the original Figure 1 Finally, based on this, we locate the coordinates of the joints of the human body by finding the maximum response position on the heat map.

[0062] like Figure 2 As shown in Figure 2, the TMPose of the present invention has lower computational complexity than the mainstream Pose Transformer. Figure 3 As shown in the figure, the token masked method proposed in the present invention has a smaller computational overhead than the Pure Transformer type and CNN+Transformer type human posture estimation networks.

[0063] In this implementation, we use the heat map method to predict key points. The heat map prediction results are as follows: Figure 4 As shown in the figure, the network predicts the heat map of the character. Each heat map in each row represents the response to a different joint point. It can be seen that although the character is severely occluded, the network can still distinguish different joint points well.

[0064] (3) Use the images and labels in step (1) to train the TMPose model framework.

[0065] During the training process, we use the MSE loss function to train the network. The calculation formula of MSE loss is as follows:

[0066]

[0067] Where: M is the number of joint points (COCO: M = 17, MPII: M = 16), P k The keypoint heatmap generated for the k-th ground-truth coordinate, is the heat map of the predicted k-th joint point.

[0068] (4) By feeding the character picture into the trained model, the coordinates of the joint points can be predicted.

[0069] Verification Example

[0070] For fair comparison, we adopt a model variant similar to that in TokenPose, as shown in Table 1. In the following, we will use B to replace Base and L to replace Large, and HRNet-s indicates that only the first three stages of HRNet are used.

[0071] Table 1

[0072]

[0073] During the training process, we followed a top-down paradigm, uniformly cropping all single-person images in the COCO dataset to 256×192 size, and all single-person images in the MPII dataset to 256×256 size. We used the Adam optimizer with an initial learning rate of 1e-3, which was reduced to 1e-4 and 1e-5 in the 200th and 260th epochs, respectively. The total number of epochs was set to 300. All models were implemented in Pytorch and performed on a server with 8 NVIDIA RTX 3090 24G graphics cards.

[0074] COCO validation dataset: Table 2 shows the test results of TMPose on COCO Validation. It can be seen that TMPose has greatly reduced the computational complexity without significant loss of accuracy. Among them, TMPose-B has only lost 0.3% of AP while the computational complexity of the Transformer part has decreased by 60.5%. TMPose-L / D24 has a significant improvement in GFLOPs. T With a 61.1% decrease, AP only dropped by 0.2%. Compared with TransPose-H-A6, TMPose-L / D24 has achieved a significant lead in both speed and accuracy, and is also significantly better than EfficientPose and Lite-HRNet.

[0075] Table 2

[0076]

[0077] Note: Test results on the COCO Validation dataset, where the input image size is 256×192, * indicates the use of GTBox, GFLOPs T Indicates the computational complexity of the Transformer part. Since TMPose focuses on accelerating the Transformer, we use GFLOPs T As the main evaluation indicator.

[0078] MPII validation dataset: Table 3 shows the test results of TMPose on the MPII dataset. It can be seen that the computational complexity of TMPose-L / D6 is reduced by 59.4%, while Mean only loses 0.3%. Specifically, it is on par with TokenPose-L / D6 on Sho and even surpasses it on Kne, achieving an accuracy of 86.2%. This shows that the token masking concept does not affect overall accuracy. This concept allows the network to perform pose estimation more efficiently, achieving a balance between speed and accuracy.

[0079] Table 3

[0080]

[0081] Note: Test results on the MPII dataset, where the input image size is 256×256, the mean is when PCK@0.5, and * indicates the use of GTBox.

[0082] Qualitative experimental results analysis: Figure 5 It can be seen that TMPose can also predict the final joint points well in the case of Mask, and can show excellent performance no matter in single or multi-person scenes, crowded or occluded scenes. Figure 6 The following figure compares the experimental results of TMPose and mainstream lightweight network models. The first two figures show that in occlusion scenarios, TMPose can infer more reasonable joint point positions. The fourth figure shows that in scenarios with more complex image features, TMPose can more accurately predict the final result. In short, TMPose is better than EfficientPose and Lite-HRNet in both speed and accuracy.

[0083] Ablation experiment results analysis: In order to verify the impact of mask type and sampling rate on the experiment, we conducted an ablation experiment as shown in Table 4. It can be seen that when the sampling rate is the same, random sampling achieves the best accuracy, followed by uniform sampling, and patch sampling loses the most performance. When the sampling rate gradually decreases, the computational complexity also gradually decreases. When the sampling rate is 0.4, GFLOPs T A 79% decrease in GFLOPs at a sampling rate of 0.6 T A decrease of 60%; in summary, we believe that Random 0.6 strikes a balance between speed and accuracy and is the best combination.

[0084] Table 4

[0085]

[0086]

[0087] Summary: For the token mask-based human pose estimation network TMPose of the present invention, the effectiveness of the network is verified on the COCO and MPII datasets. Compared with TokenPose, TMPose has a significantly reduced computational complexity with almost no loss of accuracy. We also verified through qualitative experimental analysis that in practical applications, compared with the mainstream lightweight human pose estimation network, TMPose can predict the coordinates of human joints more accurately and reasonably. Finally, we verified the influence of different Mask Ratios and Mask Types on the results through ablation experiments.

[0088] The above description of the embodiments is intended to facilitate understanding and application of the present invention by those skilled in the art. It is apparent that those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without requiring creative effort. Therefore, the present invention is not limited to the above embodiments. Any improvements or modifications made by those skilled in the art based on the disclosure of the present invention should fall within the scope of protection of the present invention.

Claims

1. An efficient human pose estimation method based on Transformer token masking, comprising the following steps: (1) Obtain a large number of pictures containing people and mark the joint positions of the people in the pictures; (2) Construct the TMPose network model framework, which includes the backbone network, token generator, Transformer module and heat map regression module, where: The backbone network is used to extract features from the input image; The token generator is used to convert the feature map extracted by the backbone network into a one-dimensional sequence form and generate tokens through partial masking. The specific operation process is as follows: First, the feature map extracted by the backbone network is divided into grids, flatten each grid into a size of P h ×P w ×C one-dimensional sequence, and then the flattened sequence is mapped into visual tokens through a linear layer. At the same time, M learnable key point tokens are randomly initialized and generated. The sequence length is consistent with the visual token, where H, W, and C are the height, width, and channel dimensions of the feature map respectively; Then, the visual tokens are masked into a size of L×(M·R) by a mapping function. The mapping functions used in the masking process include random sampling, uniform sampling, and patch sampling. The ratio R is 0.4, 0.6, or 0.8, where L and M are the sequence length and number of visual tokens, respectively. Finally, the visual token is positionally encoded, that is, the position of each sequence in the visual token is marked using a two-dimensional position embedding. Then, the key point token and the position-encoded visual token are concatenated together to obtain a token, which is then fed into the Transformer module. The Transformer module is used to encode the input Token, thereby outputting M one-dimensional sequences, where M is the number of labeled joint points; The heatmap regression module is used to reshape the M one-dimensional sequences to generate M heatmaps, from which the prediction results of the position coordinates of each joint point are decoded; the heatmap regression module only takes M key point tokens as output, and then remaps them back to a two-dimensional heatmap form through a multi-layer perceptron. Specifically: the output X of the Transformer module is fed into the multi-layer perceptron, where X is first mapped back to a two-dimensional form P, whose size is M×H * ×W * , H * and W * The height H and width W of the original image are 1 / 4 respectively. Finally, P is reshaped into M heat maps with the same size as the original image. On this basis, the coordinates of each joint of the human body are located by finding the maximum response position on the heat map; (3) Using the images and their annotation information obtained in step (1) to train the above model framework; (4) Input the image containing the person to be estimated into the trained model, and the position coordinates of each joint point of the person in the image can be predicted.

2. The method for efficient human posture estimation according to claim 1, wherein: To balance the Transformer overhead, the backbone network only retains the first three stages of HRNet, with only 25% of the original parameters. The specific operation is: first, the input image is cropped to a uniform size, and then feature extraction is performed through a deep convolutional neural network to generate a feature map with a size of 1 / 4 of the original image.

3. The efficient human posture estimation method according to claim 1, wherein: The Transformer module is composed of a cascade of multiple layers of encoding structures. Each layer of the encoding structure is composed of a regularization layer L1, a multi-head self-attention mechanism layer, a regularization layer L2, and a multi-layer perceptron MLP connected in sequence. The input of the regularization layer L2 is the output of the multi-head self-attention mechanism layer superimposed with the input of the regularization layer L1, and the output of the multi-layer perceptron MLP is superimposed with the input of the regularization layer L2 as the output of the encoding structure.

4. The method for efficient human posture estimation according to claim 3, wherein: The multi-head self-attention mechanism layer is composed of multiple Self-Attention layers, each of which combines the input of the multi-head self-attention mechanism layer with the linear transformation matrix W Q 、W K and W V Multiplying them gives the query vector matrix Q, key vector matrix K, and value vector matrix V. Then, the output of Self-Attention is calculated using the following formula. Finally, the outputs of multiple Self-Attention layers are concatenated and passed through a linear layer to obtain the output of the multi-head self-attention mechanism layer. Where: Attention(Q,K,V) is the output of Self-Attention, s is the number of columns in the query vector matrix Q, and T represents transpose.

5. The efficient human posture estimation method according to claim 1, wherein: The specific implementation of step (3) is as follows: 3.1 Initialize model parameters, including the bias vector and weight matrix of each layer, learning rate, and optimizer; 3.2 Input the image containing the person into the model, and the model forward propagates the output to obtain the corresponding prediction result, that is, the position coordinates of the joint points, and calculate the loss function between the prediction result and the annotation information; 3.3 Based on the loss function, the optimizer is used to iteratively update the model parameters through the gradient descent method until the loss function converges and the training is completed.

6. The efficient human posture estimation method according to claim 5, wherein: The expression of the loss function is as follows: Where: L MSE is the loss function, P k The heat map generated for the real coordinates of the k-th joint point in the annotation information, Predict the heat map of the k-th joint point for the model.

Images