Machine learning-based complex gesture behavior analysis method
By using machine learning methods to label, calculate similarity, and optimize models for pose data, the problems of low efficiency and insufficient analysis in the processing of complex pose data are solved, and efficient and accurate pose behavior parsing is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SUZHOU AUTOMOBILE RES INST OF TSINGHUA UNIV (WUJIANG)
- Filing Date
- 2025-07-24
- Publication Date
- 2026-06-23
AI Technical Summary
Existing technologies are inefficient when processing complex posture data and lack unified and effective integration and analysis methods, making it difficult to achieve accurate and comprehensive understanding and interpretation of posture behavior.
Machine learning methods are used to label, calculate similarity, and filter pose data. Similar feature data are combined to build and optimize an analytical model. K-fold cross-validation is used to train and validate the model, generating accurate analytical results.
It improves data utilization efficiency and analytical value, enhances model training effectiveness and generalization ability, improves the accuracy and stability of pose resolution, and meets the requirements of real-time performance and high efficiency.
Smart Images
Figure CN121167252B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of posture recognition technology, specifically a method for parsing complex posture behaviors based on machine learning. Background Technology
[0002] With the development of technology, complex posture behavior analysis has important application needs in many fields, such as human-computer interaction, intelligent security, sports training analysis, and rehabilitation medicine.
[0003] Patent application CN116343341B discloses a motion prediction and analysis method and system based on posture recognition. By acquiring the attribute information of the posture to be identified, and based on the attribute information of the posture to be identified and a preset strategy, the method determines the action that the posture to be identified will take place, records the motion state of the posture to be identified, and selects to update multiple posture feature vector sets of the posture to be identified to the state of the action that the action to be identified has taken, or updates multiple posture feature vector sets of the posture to be identified to the state of the first action and the action that the action to be identified has taken, respectively.
[0004] However, traditional attitude analysis methods often have many limitations when faced with complex and diverse attitude data. They are inefficient when processing large-scale attitude data and cannot meet the requirements of real-time performance and high efficiency. At the same time, they lack unified and effective integration and analysis methods for different types of attitude data, such as images, videos and sensor data, making it difficult to achieve accurate and comprehensive attitude behavior understanding and interpretation in practical applications. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a complex posture behavior parsing method based on machine learning. This method solves the problem of lacking unified and effective integration and analysis methods, resulting in low efficiency when processing large-scale posture data, which makes it difficult to achieve accurate and comprehensive posture behavior understanding and interpretation in practical applications.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a complex pose behavior parsing method based on machine learning, which specifically includes the following steps:
[0007] Step S1: Acquire images, videos, or sensor data of complex pose behaviors, and label the acquired pose data;
[0008] Step S2: Analyze the obtained posture data, filter similar posture data by calculating the similarity between posture data, and combine the obtained similar posture data to obtain posture data of the same type;
[0009] Step S3: Analyze the obtained feature data of the same type, analyze the correctness of the posture behavior corresponding to the feature data of the same type, and correct the incorrect posture behavior to generate correct data information;
[0010] Step S4: Divide the obtained correct data information into different datasets, and at the same time, establish corresponding analytical models based on the different datasets, and optimize the analytical models;
[0011] Step S5: Obtain the corresponding pose to be analyzed, substitute the pose into the analysis model for analysis to generate the corresponding analysis results, and display the analysis results to the corresponding operator.
[0012] As a further aspect of the present invention, the specific method for analyzing the attitude data in step S2 is as follows:
[0013] All pose data are acquired and labeled as i, where i = 1, 2, ..., j, and j represents the number of pose data. Then, the pose data labeled i = 1 is taken as the target object. At the same time, the similarity value between the remaining pose data and the target object is calculated. Based on the similarity value, pose data that are similar to the target object are filtered and recorded as similar pose data.
[0014] All similar pose data are acquired, and the corresponding behavioral features are obtained. Then, similar pose data with the same behavioral features are combined to obtain the same type of feature data. The obtained same type of feature data are labeled as n, and n=1, 2, ..., m.
[0015] As a further aspect of the present invention, the specific method for analyzing similar feature data in step S3 is as follows:
[0016] Acquire all similar feature data, compare the posture behavior of similar feature data with the standard behavior. If the posture behavior of similar feature data is different from the standard behavior, it indicates that the posture behavior corresponding to the current similar feature data is abnormal, and an abnormal posture signal is generated. Conversely, if the posture behavior of similar feature data is the same as the standard behavior, it indicates that the posture behavior corresponding to the current similar feature data is normal, and a normal posture signal is generated.
[0017] The characteristic data of the same type corresponding to the attitude anomaly signal is recorded as the abnormal same type data, and the attitude behavior of the abnormal same type data is corrected according to the standard behavior, while generating correct data information.
[0018] As a further aspect of the present invention, the specific method for dividing the correct data information in step S4 is as follows:
[0019] Obtain all the correct data information, analyze the correct data information to select the standard model, and divide the correct data information into training set and validation set according to K-fold cross-validation method;
[0020] The labeled correct data information is divided into K subsets of equal size, which are called "folds". For example, when K=5, the dataset is divided into 5 subsets and K rounds of training and validation are performed. In each round, one subset is selected as the validation set and the remaining K-1 subsets are selected as the training set. This process is repeated to obtain the training set and validation set.
[0021] As a further aspect of the present invention, the specific method for establishing corresponding analytical models based on the different datasets obtained in step S4, and optimizing the analytical models, is as follows:
[0022] The training set is input into the standard model. In each iteration, the model performs forward propagation based on the input data, calculates the output result, and then calculates the loss value between the predicted result and the true label using the loss function. Next, the selected optimizer is used to backpropagate and update the model parameters based on the loss value. After each training cycle, the validation set is used to evaluate the model's performance.
[0023] As a further aspect of the present invention, the specific method for using a validation set to evaluate the model's performance in step S4 is as follows:
[0024] Obtain a validation set and acquire different types of pose data from the validation set. Substitute different types of pose data into the model and calculate the accuracy corresponding to different types of pose data. Then compare the calculated accuracy with the preset accuracy and filter out pose data with an accuracy lower than the preset accuracy as data to be analyzed.
[0025] Obtain the data quantity and attitude characteristics corresponding to the data to be analyzed, and determine the specific reasons based on the two. Optimize the analytical model based on the specific reasons obtained.
[0026] As a further aspect of the present invention, the specific method for generating the parsing result in step S5 is as follows:
[0027] Obtain the pose to be parsed and the corresponding data format. Match the data format with the standard format of the parsing model. If the data format is different from the standard format, it needs to be converted. Otherwise, if the data format is the same as the standard format, no processing is required. Input the pose to be parsed into the parsing model to generate the parsing result.
[0028] As a further aspect of the present invention, step S5, which involves inputting the attitude to be analyzed into the analytical model to generate the analytical result, is performed as follows:
[0029] By performing convolution operations between the convolution kernel and the input pose data, local features of the pose are extracted. These local features include the shape, angle, and edge information of body parts. Based on the obtained local features, comprehensive reasoning is performed to generate corresponding analytical results, which are then displayed to the corresponding operators.
[0030] This invention provides a method for parsing complex pose behaviors based on machine learning. Compared with existing technologies, it has the following advantages:
[0031] This invention uses similarity calculation and screening methods to accurately extract similar pose data from a large amount of complex pose data, and further combines them to obtain the same type of pose data and the same type of feature data. This effectively integrates data resources, improves the efficiency of data utilization and analytical value, and can better cope with the diversity and complexity of pose data. It also reduces the subjectivity of manual intervention and feature design, thus providing a more representative and accurate data foundation for subsequent model training.
[0032] By selecting a targeted model architecture based on the specific type of pose data and combining it with K-fold cross-validation for dataset partitioning and model training, the training effect and generalization ability of the model can be improved. During the model optimization process, by strictly monitoring the accuracy of the validation set and conducting in-depth analysis of the data to be analyzed, the shortcomings of the model can be accurately located, and corresponding data augmentation or structural optimization measures can be taken to solve the problem of insufficient learning of pose features by the model, thereby improving the accuracy and stability of the model, which is superior to traditional general model building and optimization methods. Attached Figure Description
[0033] Figure 1 This is a diagram illustrating the steps and methods of the present invention. Detailed Implementation
[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0035] Example 1, please refer to Figure 1 This application provides a method for parsing complex pose behaviors based on machine learning, which specifically includes the following steps:
[0036] Step S1: Acquire images, videos, or sensor data of complex pose behaviors, and label the acquired pose data.
[0037] Step S2: Analyze the obtained posture data, filter similar posture data by calculating the similarity between posture data, and combine the obtained similar posture data to obtain posture data of the same type.
[0038] All pose data are acquired and labeled as i, where i = 1, 2, ..., j, and j represents the number of pose data. Then, the pose data labeled i = 1 is taken as the target object. At the same time, the similarity value between the remaining pose data and the target object is calculated. For example, the Euclidean distance between the feature vectors of the pose data is calculated. The smaller the distance, the higher the similarity. For ease of illustration, we convert it into a similarity value between 0 and 1 (1 means completely the same and 0 means completely different). The similarity value is calculated by taking the pose data labeled i = 2 as the standard. Based on the similarity value, pose data that are similar to the target object are selected and recorded as similar pose data. The selection criterion here is to compare the similarity value with a preset value. If the similarity value is greater than the preset value, it means that the two are similar data. Otherwise, if it is less than the preset value, it means that the two are dissimilar data.
[0039] For example, the similarity value between the pose data of i=2 and the target object of i=1 is calculated to be 0.7; the similarity value between the pose data of i=3 and the target object is 0.4; the similarity value between the pose data of i=4 and the target object is 0.8, and so on. The similarity value between all the remaining pose data and the target object of i=1 is calculated in turn. Assuming that this preset value is set to 0.6, for example, the similarity values corresponding to i=2 and i=4 are greater than the preset value. Therefore, the pose data corresponding to i=2 and i=4 are marked as similar pose data to the target object.
[0040] All similar posture data are acquired, and the corresponding behavioral features are obtained. These behavioral features represent all similar posture data of the same type. Then, similar posture data with the same behavioral features are grouped together. By default, only one behavioral feature exists in each set of similar posture data. Behavioral features can be represented by different postures to obtain similar feature data. These similar feature data are then collectively labeled with n, where n = 1, 2, ..., m. For example, in a dataset about human movement postures, a set of similar posture data might show postures such as arms extended upwards, body slightly leaning forward, and legs bent. By analyzing the common behavior reflected by these postures, their behavioral feature is determined to be "preparing to jump." Similarly, if multiple sets of similar posture data exist, and although they differ slightly in details such as the degree of arm extension and the angle of body bending, their overall behavioral feature all points to "preparing to jump," then these similar posture data are grouped into a set of similar feature data.
[0041] Step S3: Analyze the obtained feature data of the same type, analyze the correctness of the posture behavior corresponding to the feature data of the same type, and correct the incorrect posture behavior to generate correct data information.
[0042] All similar feature data are acquired, and then the corresponding posture behaviors are analyzed. The posture behaviors are compared with standard behaviors. The standard behaviors here are specifically a set of standardized templates built on domain knowledge, experience, and precise definitions and detailed explanations of various behavioral features. It clearly defines the typical posture behavior patterns that each behavioral feature should present and its internal logical relationships. The posture behaviors represent the definition and explanation of the feature data corresponding to the similar feature data. If the posture behaviors of the similar feature data are different from the standard behaviors, it means that the posture behaviors corresponding to the current similar feature data are abnormal, and a posture abnormality signal is generated. Conversely, if the posture behaviors of the similar feature data are the same as the standard behaviors, it means that the posture behaviors corresponding to the current similar feature data are normal, and a posture normal signal is generated.
[0043] The characteristic data of the same type corresponding to the attitude anomaly signal is recorded as the abnormal same type data, and the attitude behavior of the abnormal same type data is corrected according to the standard behavior, while generating correct data information.
[0044] Step S4: Divide the obtained correct data information into different datasets, establish corresponding analytical models based on the different datasets, and optimize the analytical models.
[0045] First, we comprehensively collect and organize all accurate posture data. This data should cover a variety of different posture scenarios and individual performances to ensure its comprehensiveness and representativeness. For example, in a human posture analysis project, we collected human posture data from different age groups, genders, and various daily activities (such as walking, running, jumping, sitting, standing, etc.) and professional sports (such as basketball shooting, dance movements, yoga postures, etc.). This data was acquired through high-precision sensors or professional image acquisition equipment and underwent rigorous data cleaning and preprocessing to ensure its accuracy and completeness, forming the basis of our subsequent analysis—a correct data information set.
[0046] Based on the specific type and characteristics of the data, a suitable standard model architecture is carefully selected. For posture data presented in image form, given the superior performance of CNNs (Convolutional Neural Networks) in image feature extraction, classic CNN architectures such as VGG and ResNet, and their optimized variants, are given priority. For example, in the image analysis of yoga postures, the multi-layer convolutional structure of the VGG network can effectively capture key features such as human contours, limb positions, and joint angles, thereby accurately identifying various yoga postures. For data with obvious time-series characteristics, such as continuous human movement trajectories, RNNs (Recurrent Neural Networks) and their advanced improved versions (such as LSTM and GRU) are more advantageous. They can properly handle the time dependencies in the data and accurately capture the dynamic changes of posture over time. Taking the analysis of an athlete's running posture sequence as an example, the LSTM network can remember previous posture information through memory units, thereby better understanding the time-series features such as stride rhythm, limb swing patterns, and speed change trends during the running process, providing strong support for accurate analysis of running posture.
[0047] Next, K-fold cross-validation is used to scientifically and reasonably divide the correct data set to generate training and validation sets, ensuring the reliability and effectiveness of model training and evaluation. Taking K=5 as an example, the complete dataset is evenly divided into 5 equal-sized subsets, each subset preserving as much of the original dataset's feature distribution as possible. In each round of cross-validation, one subset is selected alternately as the validation set, and the remaining 4 subsets are combined as the training set. Specifically, in the first round, the first subset is set as the validation set, and the remaining 2 to 5 subsets form the training set for model training; in the second round, the second subset is selected as the validation set, and subsets 1, 3, 4, and 5 are used as the training set, and so on, for a total of 5 rounds of training and validation.
[0048] During the model training phase, the pre-defined training set data is input batch by batch into the selected standard model. The model performs forward propagation calculations based on the input pose data and outputs the corresponding prediction results. For example, when training a human pose classification model based on the training set, each time a pre-processed pose image or pose data vector is input, the model outputs the predicted probability distribution of the pose's category through its internal complex network structure and parameter calculations. Taking a training set containing 1000 samples as an example, a complete training epoch means that all 1000 samples are sequentially input into the model for one forward propagation and one backpropagation to update the parameters. In each iteration, the degree of difference between the model's prediction result and the true label is calculated using a pre-defined loss function (such as the cross-entropy loss function used for classification tasks), i.e., the loss value. Subsequently, the selected optimizer (such as the Adam optimizer) performs backpropagation based on the loss value, adjusting the model's parameters according to the rules of the optimization algorithm, so that the model gradually learns the inherent characteristics and patterns of the pose data during training, continuously optimizing its prediction performance.
[0049] After each training cycle, a validation set is introduced to rigorously evaluate the model's performance. First, different types of posture data are acquired from the validation set. These posture types should be meticulously categorized according to actual behavioral characteristics. For example, a posture category might be described as arms extended upwards, body slightly leaning forward, and legs bent, which could represent a specific preparatory posture for movement or a daily activity posture. These different types of posture data are then input into the trained model one by one, and the prediction accuracy for each posture category is calculated. Accuracy is calculated as the ratio of the number of correctly predicted samples to the total number of samples for that posture category. For example, if the model outputs 100 predictions for a specific posture category, and 95 of them are correct, then the accuracy for that posture category is 95%. Next, the calculated accuracy for each posture category is compared with a pre-set accuracy threshold. This pre-set accuracy threshold is set by experienced operators based on the specific needs and actual conditions of the project. For example, in some professional sports posture analysis projects with extremely high accuracy requirements, the pre-set accuracy might be set above 98%, while in some general posture recognition applications, the pre-set accuracy can be appropriately reduced to around 90%. Pose data with an accuracy rate lower than the preset accuracy rate are selected and marked as data to be analyzed in order to further explore the reasons for the poor model performance.
[0050] For the data to be analyzed, obtain detailed information on the quantity of data and the pose features it presents. By comparing and analyzing with other well-performing pose data, determine the specific reasons for the low accuracy. If the quantity of data to be analyzed is significantly less than other normal pose data, this may indicate insufficient data, preventing the model from fully learning the feature patterns of that pose category. In this case, data augmentation strategies are needed, such as increasing the number of similar pose images through image synthesis techniques, or collecting more pose data from real-world scenarios to enrich the data sample. If analysis determines that the problem stems from the model's insufficient ability to learn and recognize pose features—for example, failing to fully extract deep features of complex poses in a CNN model, or failing to accurately capture key changes in pose time series in an RNN model—then appropriate optimization of the model structure is necessary. For CNN models, the number and structural complexity of convolutional and pooling layers can be reasonably increased to enhance their ability to extract pose image features. For RNN models, the number of neurons in the hidden layers can be appropriately increased, or more advanced gating mechanisms (such as the optimization of update and reset gates in GRU) can be adopted to improve their ability to learn and memorize pose time series features. Thus, the parsing model can be optimized in a targeted manner according to specific reasons, improving the model's parsing accuracy and generalization performance for various pose data, and ultimately achieving accurate and efficient parsing of pose data to meet the actual needs of different application scenarios.
[0051] The optimized process described above enables a more systematic and comprehensive approach to building and optimizing attitude analysis models using accurate data, effectively improving the accuracy and reliability of the models in processing complex attitude data, and providing solid technical support for practical applications related to attitude analysis.
[0052] Step S5: Obtain the corresponding pose to be analyzed, substitute the pose into the analysis model for analysis to generate the corresponding analysis results, and display the analysis results to the corresponding operator.
[0053] The process involves acquiring the pose to be parsed, along with the corresponding data format. This data format can be either image or text. The data format is then matched against the standard format of the parsing model. If the data format differs from the standard format, it needs to be converted; otherwise, no processing is required. For example, if the parsing model is a convolutional neural network (CNN) and requires fixed-size image data as input, the pose data (such as joint coordinates obtained from sensors) needs to be converted into an image format. The pose contour can be drawn on a blank image based on the joint coordinates, or a thermal image can be used to represent the position and intensity of the joints.
[0054] Next, the pose to be analyzed is input into the analytical model, and the data format of the pose to be analyzed is the same as the analytical format of the analytical model. The input pose data is convolved with the convolution kernel to extract local features of the pose. Local features include the shape, angle, edge and other information of body parts. Based on the obtained local features, comprehensive reasoning is performed to generate the corresponding analytical result, and the analytical result is displayed to the corresponding operator.
[0055] For example, the task of an analytical model is to classify postures, and the output will be the category to which the posture belongs. For a complex analytical model of industrial robot operation posture, the output may be different operation categories such as "grasping operation", "assembly operation", and "welding operation". When analyzing the jumping posture of an athlete, the output may be the estimated value of the jump height, the take-off speed, and other values.
[0056] Example 2 is implemented based on Example 1, and the difference from Example 1 is as follows:
[0057] The similarity value between the remaining pose data and the target object is calculated. In the first embodiment, the Euclidean distance between the feature vectors of the pose data is calculated. However, in this second embodiment, the analysis is performed by a simple matching coefficient. Specifically, the target object and the remaining pose data are vectorized, and the proportion of the number of the same category in the two vectors is calculated to the total number of categories. The similarity is judged by the calculated proportion.
[0058] Example 3, as Example 3 of the present invention, focuses on combining the implementation processes of Example 1 and Example 2.
[0059] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.
[0060] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.
Claims
1. A method for parsing complex pose behaviors based on machine learning, characterized in that, The method specifically includes the following steps: Step S1: Acquire images, videos, or sensor data of complex pose behaviors, and label the acquired pose data; Step S2: Analyze the obtained posture data, filter similar posture data by calculating the similarity between posture data, and combine the obtained similar posture data to obtain feature data of the same type; Step S3: Analyze the obtained feature data of the same type. Analyze the correctness of the posture behavior corresponding to the feature data of the same type, and correct the incorrect posture behavior to generate correct data information. The specific processing method is as follows: Acquire all similar feature data, compare the posture behavior of similar feature data with the standard behavior. If the posture behavior of similar feature data is different from the standard behavior, it indicates that the posture behavior corresponding to the current similar feature data is abnormal, and an abnormal posture signal is generated. Conversely, if the posture behavior of similar feature data is the same as the standard behavior, it indicates that the posture behavior corresponding to the current similar feature data is normal, and a normal posture signal is generated. The same type of feature data corresponding to the attitude anomaly signal is recorded as the abnormal same type data, and the attitude behavior of the abnormal same type data is corrected according to the standard behavior, while generating correct data information. Step S4: Divide the obtained correct data information into different datasets, and at the same time, establish corresponding analytical models based on the different datasets, and optimize the analytical models; Step S5: Obtain the corresponding pose to be analyzed, substitute the pose into the analysis model for analysis to generate the corresponding analysis results, and display the analysis results to the corresponding operator.
2. The complex pose behavior parsing method based on machine learning according to claim 1, characterized in that, The specific method for analyzing the attitude data in step S2 is as follows: All pose data are acquired and labeled as i, where i = 1, 2, ..., j, and j represents the number of pose data. Then, the pose data labeled i = 1 is taken as the target object. At the same time, the similarity value between the remaining pose data and the target object is calculated. Based on the similarity value, pose data that are similar to the target object are filtered and recorded as similar pose data. All similar pose data are acquired, and the corresponding behavioral features are obtained. Then, similar pose data with the same behavioral features are combined to obtain the same type of feature data. The obtained same type of feature data are labeled as n, and n=1, 2, ..., m.
3. The complex pose behavior parsing method based on machine learning according to claim 1, characterized in that, The specific method for dividing the correct data information in step S4 is as follows: Obtain all the correct data information, analyze the correct data information to select the standard model, and divide the correct data information into training set and validation set according to K-fold cross-validation method; The labeled correct data information is divided into K equal-sized subsets, called "folds", and K rounds of training and validation are performed. In each round, one subset is selected as the validation set, and the remaining K-1 subsets are selected as the training set. This process is repeated to obtain the training set and validation set.
4. The complex pose behavior parsing method based on machine learning according to claim 1, characterized in that, The specific method for establishing corresponding analytical models based on the different datasets obtained in step S4, and optimizing the analytical models, is as follows: The obtained training set is input into the standard model. In each iteration, the model performs forward propagation based on the input data, calculates the output result, and then calculates the loss value between the predicted result and the true label through the loss function. The selected optimizer is used to backpropagate based on the loss value to update the model parameters. After each training cycle, the validation set is used to evaluate the performance of the model.
5. The complex pose behavior parsing method based on machine learning according to claim 4, characterized in that, The specific method for using the validation set to evaluate the model's performance in step S4 is as follows: Obtain a validation set and acquire different types of pose data from the validation set. Substitute different types of pose data into the model and calculate the accuracy corresponding to different types of pose data. Then compare the calculated accuracy with the preset accuracy and filter out pose data with an accuracy lower than the preset accuracy as data to be analyzed. Obtain the data quantity and attitude characteristics corresponding to the data to be analyzed, and determine the specific reasons based on the two. Optimize the analytical model based on the specific reasons obtained.
6. The complex pose behavior parsing method based on machine learning according to claim 1, characterized in that, The specific method for generating the parsing result in step S5 is as follows: Obtain the pose to be parsed and the corresponding data format. Match the data format with the standard format of the parsing model. If the data format is different from the standard format, it needs to be converted. Otherwise, if the data format is the same as the standard format, no processing is required. Input the pose to be parsed into the parsing model to generate the parsing result.
7. The complex pose behavior parsing method based on machine learning according to claim 6, characterized in that, The method for generating the analytical result by inputting the attitude to be analyzed into the analytical model in step S5 is as follows: By performing convolution operations between the convolution kernel and the input pose data, local features of the pose are extracted. These local features include the shape, angle, and edge information of body parts. Based on the obtained local features, comprehensive reasoning is performed to generate the corresponding analytical results.