Human body posture estimation method and device based on multi-stage knowledge distillation and storage medium
Through multi-stage knowledge distillation method and layered and progressive distillation strategy, the deployment problems and semantic mismatch of the human pose estimation model on resource-limited devices are solved, and efficient and accurate human pose estimation is achieved, which is suitable for mobile devices.
Patent Information
- Application Number
- CN202411547958.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-01
- Publication Date
- 2025-02-14
AI Technical Summary
The existing high-precision human pose estimation model is difficult to deploy on mobile devices with limited resources due to the requirements of computing resources and storage space, and the knowledge distillation method has semantic mismatch problems when transmitting information on the spatial structure of the teacher model.
Using a multi-stage knowledge distillation method, light student models are trained through pre-trained teacher models, and the student model parameters are optimized by using a layered progressive distillation strategy and a multi-receptive field fusion module. , solve the problem of semantic mismatch, and reduce inference memory usage by key point representation method based on classification coordinates.
Significantly improves the deployment efficiency and accuracy of lightweight networks on resource-constrained devices, solves the problem of semantic mismatch, and reduces inference time and memory usage without sacrificing performance.
Smart Images
Figure CN119445667A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer vision technology, and specifically relates to a human posture estimation method, device and storage medium based on multi-stage knowledge distillation. Background Art
[0002] Human pose estimation is an important technology in the field of computer vision and is widely used in many fields such as behavior analysis, interactive design, and augmented reality. Existing high-precision human pose estimation models often rely on complex network structures and a large amount of computing resources, which makes them difficult to deploy on mobile devices with limited resources. Traditional high-precision human pose estimation models are usually based on deep learning, especially convolutional neural networks (CNN). These models, such as HRNet and OpenPose, achieve excellent accuracy through a large number of network layers and complex architectures. However, these high-precision models usually require a lot of computing resources and storage space, which limits their application on devices with limited computing power, such as smartphones and embedded systems. In order to deploy human pose estimation models on such devices, researchers have explored a variety of model lightweight strategies, including network pruning, low-rank decomposition, quantization, and knowledge distillation. Although these methods can reduce the model burden, they often come at the expense of certain performance. Knowledge distillation is an effective model compression technique that improves the performance of the student model by transferring knowledge from a complex "teacher" model to a simple "student" model. During the knowledge distillation process, the student model learns to imitate the output or feature behavior of the teacher model, thereby improving accuracy without significantly increasing the computational burden. Although knowledge distillation has been successful in tasks such as image classification, it remains a challenge to effectively transfer the spatial structure information in the teacher model in the task of human pose estimation, which has a highly structured output. In addition, existing distillation methods rarely consider the semantic mismatch problem caused by the difference in depth and width between the teacher and student networks, which is particularly important in the task of human pose estimation. Summary of the invention
[0003] In view of the deficiencies in the prior art, the present invention provides a human posture estimation method, device and storage medium based on multi-stage knowledge distillation, which effectively solves the semantic mismatch problem caused by the difference in the network structure between the teacher and the student in the prior art, thereby significantly improving the deployment efficiency and accuracy of lightweight networks on resource-constrained devices without sacrificing model performance.
[0004] The present invention provides the following technical solutions:
[0005] In a first aspect, a method for estimating human posture based on multi-stage knowledge distillation is provided, comprising:
[0006] A lightweight student model is trained using a pre-trained teacher model; during the training process, the teacher model transfers knowledge to the student model through multi-stage knowledge distillation; the multi-stage knowledge distillation includes: intermediate layer distillation and output layer distillation; the intermediate layer distillation is: during the training process, the teacher model uses a hierarchical and progressive distillation strategy to guide the feature extraction of the student model, and the hierarchical and progressive distillation strategy is: the feature extraction of each stage of the teacher model not only guides the feature extraction of the same stage of the student model, but also guides the feature extraction of all subsequent stages of the student model; the output layer distillation is: at the output stage, the vanilla knowledge distillation method is used to adjust the output of the student model;
[0007] The multi-stage knowledge distillation loss and the original human pose estimation loss are used to optimize the parameters of the student model until a student model that meets the preset conditions is generated, and the trained student model is used for human pose estimation.
[0008] Optionally, the teacher model adopts a hierarchical and progressive distillation strategy to guide the feature extraction of the student model, and introduces a multi-receptive field fusion module; the multi-receptive field fusion module combines multi-size convolution and pooling operations.
[0009] Optionally, during the intermediate layer distillation process, 1×1 convolution is used to align feature channels, and when the size of the feature map output by the student model during the feature extraction process at this stage is inconsistent with the size of the feature map output by the teacher model at all stages, transposed convolution is used to adjust the size of the feature map output by the student model at this stage.
[0010] Optionally, the teacher model adopts a hierarchical and progressive distillation strategy to guide the feature extraction of the student model, and also introduces a cross attention weight allocation module to obtain the intermediate layer distillation loss;
[0011] Specifically, based on the extracted student feature map of the mth stage and the teacher characteristics of the first stage Obtain the autocorrelation matrix of the teacher model at stage l and the autocorrelation matrix of the student model at stage m;
[0012] Perform dimension reduction and normalization operations on the autocorrelation matrix of the teacher model in the lth stage and the autocorrelation matrix of the student model in the mth stage to obtain the student eigenvector S of the mth stage m [i] and the teacher feature vector T of the first stage l [i];
[0013] Based on the student feature vector S of the mth stage m [i] and the teacher feature vector T of the first stage l[i], get the attention weight γ of the m-th stage student model and the l-th stage teacher model i (t l ,s m );
[0014]
[0015] Where L is the number of stages of the teacher model, M is the number of stages of the student model, and i represents the i-th instance in the sample batch;
[0016] According to the attention weights of the student model in the mth stage and the teacher model in the lth stage, the intermediate layer distillation loss l is determined fkd ;
[0017]
[0018] Where B is the number of samples in each batch, D(·) is the distance function, and f align (·) is the feature alignment function; is the i-th instance extracted by the l-th layer of the teacher model; is the i-th instance extracted by the m-th layer of the student model.
[0019] Optionally, the representation method of the student model output adopts a key point representation method based on classification coordinates;
[0020] The key point representation method based on classified coordinates is: converting the key point positioning task into the categories of predicting the horizontal coordinates and vertical coordinates of the key points.
[0021] Optionally, based on the representation method of the student model output, the original loss L of the student model for human pose estimation is designed task for:
[0022]
[0023] Among them, N is the total number of people in a batch, K is the number of key points of human posture estimation output by the model, and M n,k is the weight mask of the unobserved key points, Q is the number of categories of vertical and horizontal coordinates, V q is the label value, S q The coordinate category probability value predicted by the student model.
[0024] Optionally, in the case where the vanilla knowledge distillation method is used to adjust the output of the student model, when the representation method of the output of the student model adopts a key point representation method based on classification coordinates, the loss L of the output layer distillation is vkd for:
[0025]
[0026] Among them, T q is the coordinate category probability value predicted by the teacher model, and KL(·) is the KL loss function.
[0027] Optionally, in the process of training the model, the overall loss function L of the student model is constructed using the loss of the intermediate layer distillation, the loss of the output layer distillation and the original loss of the human posture estimation;
[0028] L=L task +α·L fkd +β·L vkd
[0029] Among them, l task is the original loss of human pose estimation, l fkd is the intermediate distillation loss, l vkd is the output layer distillation loss, α is the hyperparameter of the intermediate layer distillation loss, and β is the hyperparameter of the output layer distillation loss.
[0030] In a second aspect, a computer device is provided, comprising a processor and a memory; wherein, when the processor executes a computer program stored in the memory, the steps of the human body posture estimation method based on multi-stage knowledge distillation as described in any one of the first aspects are implemented.
[0031] In a third aspect, a computer-readable storage medium is provided for storing a computer program; when the computer program is executed by a processor, the steps of the method for human posture estimation based on multi-stage knowledge distillation as described in any one of the first aspects are implemented.
[0032] Compared with the prior art, the present invention has the following beneficial effects:
[0033] (1) The present invention uses a multi-layer teacher network to guide the student network at different stages, thereby solving the semantic mismatch problem between the teacher and student networks caused by different complexities in traditional knowledge distillation, thereby significantly improving the deployment efficiency and accuracy of lightweight networks on resource-constrained devices without sacrificing model performance.
[0034] (2) The present invention also achieves effective cross-layer distillation interaction and optimizes the learning process by expanding the receptive field of the student network output feature map, calculating the similarity matrix between the student and teacher features, and assigning distillation weights at the corresponding stage.
[0035] (3) The present invention adopts a key point representation method based on classified coordinates to replace the traditional key point representation method based on heat map, which effectively reduces the memory usage of reasoning and improves the reasoning speed, and further performs knowledge distillation based on the generated results; while reducing the reasoning time and memory usage, the present invention can maintain or even improve the accuracy of human posture estimation. The present invention is particularly suitable for resource-constrained mobile devices and has important practical application value. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 It is the overall network structure diagram of the multi-stage knowledge distillation of the present invention.
[0037] Figure 2 It is a structural schematic diagram of the multi-receptive field fusion module of the present invention.
[0038] Figure 3 It is a structural schematic diagram of the cross-attention weight allocation module of the present invention.
[0039] Figure 4 It is a classification decoupling coordinate representation diagram of the present invention. DETAILED DESCRIPTION
[0040] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.
[0041] It should be noted that the term "comprises" and any variations of the term in the specification and claims of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or inherent to these processes, methods, products or apparatus.
[0042] Example 1
[0043] A human posture estimation method based on multi-stage knowledge distillation is provided, comprising: using a pre-trained teacher model to train a lightweight student model; during the training process, the teacher model transfers knowledge to the student model through multi-stage knowledge distillation, and uses the multi-stage knowledge distillation loss and the original human posture estimation loss to optimize the parameters of the student model until a student model that meets preset conditions is generated, so as to use the trained student model to perform human posture estimation.
[0044] The teacher model is usually a complex and high-precision complex model. As an option, the network structure of the teacher model can adopt HRNet. The student model is usually a lightweight network. As an option, the network structure of the student model is LiteHRNet.
[0045] like Figure 1 As shown in the figure, the network of the teacher model and the student model mainly consists of Stage 1, Stage 2, Stage 3, and Stage 4.
[0046] The multi-stage knowledge distillation includes: intermediate layer distillation and output layer distillation.
[0047] 1. Intermediate distillation
[0048] The intermediate layer distillation is: during the training process, the teacher model adopts a hierarchical and progressive distillation strategy to guide the feature extraction of the student model. The hierarchical and progressive distillation strategy is: the feature extraction of each stage of the teacher model not only guides the feature extraction of the student model at the same stage, but also guides the feature extraction of all subsequent stages of the student model.
[0049] Specifically, due to the difference in complexity between the teacher model and the student model, in the shallow stage of the network structure, the teacher features often have deeper semantic features. In the distillation process, in order to effectively transfer the deep semantic features from the teacher model to the student model, a hierarchical and progressive distillation strategy is adopted. Specifically, each stage of the teacher model not only guides the same stage corresponding to the student model, but also guides its subsequent multiple stages; for example, the first stage (Stage1) of the teacher model will guide the first stage (Stage1) to the fourth stage (Stage4) of the student model, and the second stage (Stage2) of the teacher model will guide the student model from the second stage (Stage2) to the fourth stage (Stage4). This hierarchical and progressive guidance method helps the student model to more comprehensively capture and imitate the complex features of the teacher model, thereby improving its performance at all levels.
[0050] As a further step, the teacher model adopts a hierarchical and progressive distillation strategy to guide the feature extraction of the student model, and introduces a multi-field fusion module (Multi-Field Fusion Module Transform, MFT); this module is to enhance the ability of the student network in processing complex features inherited from the teacher network, such as Figure 2As shown. The module MFT increases the receptive field of the student network by combining multi-size convolution and pooling operations, which enables the student network to perceive contextual information in a wider area during the multi-stage knowledge distillation process, thereby better capturing the mutual relationship and spatial layout of each joint in complex postures. When the feature size output by the student network at different stages is inconsistent with a specific stage of the teacher model, a transposed convolution is used to adjust the size of the student feature map. In addition, 1×1 convolution is used to align feature channels so that the features of the student network can be better aligned with those of the teacher network. Distillation is performed through branches with multiple receptive fields, which make the knowledge transfer from the teacher's features to the student's features more effective. This design allows the student network to not only replicate the features of the teacher network, but also gradually improve its understanding and processing capabilities of complex data during the learning process.
[0051] In this embodiment, in the middle layer distillation process of multi-stage distillation, a cross-attention weight allocation model (CWA) is also introduced to dynamically adjust the association weights between teacher and student features, such as Figure 3 shown.
[0052] Specifically, based on the extracted student feature map of the mth stage and the teacher characteristics of the first stage Obtain the autocorrelation matrix of the teacher model at stage l and the autocorrelation matrix of the student model at stage m;
[0053] The autocorrelation matrix T of the teacher model at stage l l :
[0054]
[0055] The autocorrelation matrix S of the student model at stage m m :
[0056]
[0057] Among them, Transpose means transpose.
[0058] Perform dimension reduction and normalization operations on the autocorrelation matrix of the teacher model in the lth stage and the autocorrelation matrix of the student model in the mth stage to obtain the student eigenvector S of the mth stage m [i] and the teacher feature vector T of the first stage l [i].
[0059] Specifically, because T l and S mBoth are B×B matrices. The two autocorrelation matrices obtained are projected through two linear layers. Each sample vector is processed by a linear layer followed by an activation function from the obtained autocorrelation matrix, and then passed through a normalization layer; the output vector is standardized using the L2 norm, which effectively refines the sample features; finally, we get the corresponding sample vector S m [i] and T l [i].
[0060] Based on the student feature vector S of the mth stage m [i] and the teacher feature vector T of the first stage l [i], get the attention weight γ of the m-th stage student model and the l-th stage teacher model i (t l ,s m );
[0061]
[0062] Among them, the attention weight γ i (t l ,s m ) must meet L is the number of stages of the teacher model, M is the number of stages of the student model, and i represents the i-th instance in the sample batch.
[0063] According to the attention weights of the student model in the mth stage and the teacher model in the lth stage, the intermediate layer distillation loss K is determined fkd .
[0064]
[0065] Where B is the number of samples in each batch, D(·) is the distance function, and the mean square error MSE is usually used as the distance function during training. align (·) is the feature alignment function, which is used to adjust the dimension and size of student features; is the i-th instance extracted by the l-th layer of the teacher model; is the i-th instance extracted by the m-th layer of the student model.
[0066] 2. Output Layer Distillation
[0067] The output layer distillation is as follows: In the output stage, the vanilla knowledge distillation method is used to adjust the output of the student model.
[0068] The representation method of the student model output adopts the key point representation method based on classification coordinates.
[0069] The key point representation method based on classified coordinates is: converting the key point positioning task into the categories of predicting the horizontal and vertical coordinates of the key points.
[0070] Specifically, the key point localization task is transformed into a classification problem, that is, predicting the categories of the horizontal and vertical coordinates of the key points, replacing the traditional 2D Gaussian heat map coordinate representation method, such as Figure 4 As shown in Figure 2, for the features output from the backbone network stage, horizontal and vertical coordinate classifiers are designed, and the coordinate vectors of the corresponding key points are output through multiple fully connected layers, with a dimension of (H×W, 1).
[0071] The original loss L of the student model for human pose estimation task for:
[0072]
[0073] Among them, N is the total number of people in a batch, K is the number of key points of human posture estimation output by the model, and M n,k is the weight mask of the unobserved key points, Q is the number of categories of vertical and horizontal coordinates, V q is the label value, S q The coordinate category probability value predicted by the student model.
[0074] That is, a one-dimensional Gaussian distribution is used to generate the supervision signal, and the cross entropy is used as the loss function.
[0075] Furthermore, when the representation method of the student model output adopts the key point representation method based on classification coordinates, the vanilla knowledge distillation method is used to further adjust the output of the student model. The core of this process is to use the KL loss function to ensure that the output of the student model imitates the prediction results of the teacher model as much as possible.
[0076] The loss L of the output layer distillation vkd for:
[0077]
[0078] Among them, T q is the coordinate category probability value predicted by the teacher model, and KL(·) is the KL loss function.
[0079] 3. Overall loss function
[0080] In the process of model training, the overall loss function L of the student model is constructed by using the loss of intermediate layer distillation, the loss of output layer distillation and the original loss of human posture estimation;
[0081] L=L task +α·Lfkd +β·L vkd
[0082] Among them, L task is the original loss of human pose estimation, L fkd is the intermediate distillation loss, L vkd is the output layer distillation loss, α is the hyperparameter of the intermediate layer distillation loss, and β is the hyperparameter of the output layer distillation loss.
[0083] Example 2
[0084] This embodiment uses NVIDIA GPU graphics card as the computing platform and adopts Pytorch and MMPose as the network model framework.
[0085] Dataset acquisition
[0086] Obtain the annotated dataset from the official website of the MPII dataset, and configure the dataset locally according to the specific requirements of the model code file.
[0087] Model Training
[0088] In the training phase, we first pre-trained the teacher network HRNet32, and then set up the distillation network, set the teacher model HRNet32 and the student model LiteHRNet18, and import the corresponding weights of the teacher model. We set the input batch size to 32, the optimizer to use the Adam optimizer, the initial learning rate to 0.001, the final learning rate to 0.0001, the training rounds to 210, and the hyperparameters α and β in the loss function to 0.005 and 0.04, respectively.
[0089] Model Testing
[0090] The accuracy of the model is tested using a validation set divided from the data set. The testing method is basically the same as that of training, that is, after the model performs forward inference on the image, it directly outputs the result without backpropagation.
[0091] Example 3
[0092] The present invention provides a computer device, comprising a processor and a memory; wherein, when the processor executes a computer program stored in the memory, the steps of the above-mentioned human body posture estimation method based on multi-stage knowledge distillation are implemented.
[0093] For more specific processes of the above method, please refer to the corresponding contents disclosed in the aforementioned embodiments, which will not be repeated here.
[0094] Example 4
[0095] The present invention provides a computer-readable storage medium for storing a computer program; when the computer program is executed by a processor, the steps of the human body posture estimation method based on multi-stage knowledge distillation are implemented.
[0096] For more specific processes of the above method, please refer to the corresponding contents disclosed in the aforementioned embodiments, which will not be repeated here.
[0097] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device and storage medium disclosed in the embodiment, since they correspond to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0098] Those skilled in the art can clearly understand that the technology in the embodiments of the present invention can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the technical solution in the embodiments of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention or some parts of the embodiments.
[0099] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.
Claims
1. A human posture estimation method based on multi-stage knowledge distillation, characterized in that: include: Use the pre-trained teacher model to train the lightweight student model; During the training process, the teacher model transfers knowledge to the student model through multi-stage knowledge distillation; the multi-stage knowledge distillation includes: intermediate layer distillation and output layer distillation; the intermediate layer distillation is: during the training process, the teacher model adopts a hierarchical and progressive distillation strategy to guide the feature extraction of the student model, and the hierarchical and progressive distillation strategy is: the feature extraction of each stage of the teacher model not only guides the feature extraction of the same stage of the student model, but also guides the feature extraction of all subsequent stages of the student model; the output layer distillation is: at the output stage, the vanilla knowledge distillation method is used to adjust the output of the student model; The multi-stage knowledge distillation loss and the original human pose estimation loss are used to optimize the parameters of the student model until a student model that meets the preset conditions is generated, and the trained student model is used for human pose estimation.
2. The method for human posture estimation based on multi-stage knowledge distillation according to claim 1, characterized in that: The teacher model adopts a hierarchical and progressive distillation strategy to guide the feature extraction of the student model, and introduces a multi-receptive field fusion module; the multi-receptive field fusion module combines multi-size convolution and pooling operations.
3. The method for human posture estimation based on multi-stage knowledge distillation according to claim 1, characterized in that: During the distillation of the intermediate layers, 1×1 convolution is used to align the feature channels. When the size of the feature map output by the student model during the feature extraction process at this stage is inconsistent with the size of the feature map output by the teacher model at all stages, transposed convolution is used to adjust the size of the feature map output by the student model at this stage.
4. The method for human posture estimation based on multi-stage knowledge distillation according to claim 1, characterized in that: The teacher model adopts a hierarchical and progressive distillation strategy to guide the feature extraction of the student model, and also introduces a cross-attention weight allocation module to obtain the intermediate layer distillation loss; Specifically, based on the extracted student feature map of the mth stage and the teacher characteristics of the first stage Obtain the autocorrelation matrix of the teacher model at stage l and the autocorrelation matrix of the student model at stage m; Perform dimension reduction and normalization operations on the autocorrelation matrix of the teacher model in the lth stage and the autocorrelation matrix of the student model in the mth stage to obtain the student eigenvector S of the mth stage m [i] and the teacher feature vector T of the first stage l [i]; Based on the student feature vector S of the mth stage m [i] and the teacher feature vector T of the first stage l [i], get the attention weight γ of the m-th stage student model and the l-th stage teacher model i (t l ,s m ); Where L is the number of stages of the teacher model, M is the number of stages of the student model, and i represents the i-th instance in the sample batch; According to the attention weights of the student model in the mth stage and the teacher model in the lth stage, the intermediate layer distillation loss L is determined fkd ; Where B is the number of samples in each batch, D(·) is the distance function, and f align (·) is the feature alignment function; is the i-th instance extracted by the l-th layer of the teacher model; is the i-th instance extracted by the m-th layer of the student model.
5. The method for human posture estimation based on multi-stage knowledge distillation according to claim 1, characterized in that: The representation method of the student model output adopts the key point representation method based on classification coordinates; The key point representation method based on classified coordinates is: converting the key point positioning task into the categories of predicting the horizontal coordinates and vertical coordinates of the key points.
6. The method for human posture estimation based on multi-stage knowledge distillation according to claim 5, characterized in that: Based on the representation method of the student model output, the original loss L of the student model for human posture estimation is designed task for: Among them, N is the total number of people in a batch, K is the number of key points of human posture estimation output by the model, and M n,k is the weight mask of the unobserved key points, Q is the number of categories of vertical and horizontal coordinates, V q is the label value, S q The coordinate category probability value predicted by the student model.
7. The method for human posture estimation based on multi-stage knowledge distillation according to claim 5, characterized in that: In the method of using the vanilla knowledge distillation to adjust the output of the student model, when the representation method of the student model output adopts the key point representation method based on the classification coordinates, the loss L of the output layer distillation vkd for: Among them, T q is the coordinate category probability value predicted by the teacher model, and KL(·) is the KL loss function.
8. The method for human posture estimation based on multi-stage knowledge distillation according to claim 1, characterized in that: In the process of model training, the overall loss function L of the student model is constructed by using the loss of intermediate layer distillation, the loss of output layer distillation and the original loss of human posture estimation; L=L task +α·L fkd +β·L vkd Among them, L task is the original loss of human pose estimation, L fkd is the intermediate distillation loss, L vkd is the output layer distillation loss, α is the hyperparameter of the intermediate layer distillation loss, and β is the hyperparameter of the output layer distillation loss.
9. A computer device, characterized in that: It comprises a processor and a memory; wherein, when the processor executes the computer program stored in the memory, it implements the steps of the human body posture estimation method based on multi-stage knowledge distillation as described in any one of claims 1-8.
10. A computer-readable storage medium, characterized in that: Used to store computer programs; when the computer programs are executed by the processor, the steps of the human body posture estimation method based on multi-stage knowledge distillation according to any one of claims 1 to 8 are implemented.