A motion form recognition method and system based on fusion vision and clothing AI

Abstract

The present invention proposes a motion form recognition method and system based on fusion vision and clothing AI. It belongs to the field of artificial intelligence and computer vision technology. The method includes: real-time acquisition of video or image data of human body motion by wearable visual sensors, and based on clothing AI technology, clothing recognition of the acquired video or image data, extraction of clothing features, recognition of human body posture based on deep learning algorithms, extraction of key posture features, analysis of human body motion trajectory in combination with time information, and extraction of dynamic features. By integrating the image data collected by visual sensors and clothing AI technology, combined with deep learning algorithms for the recognition of human body posture, key posture features and dynamic features can be extracted more accurately, thereby improving the accuracy of motion form recognition.

Description

Technical Field

[0001] The present invention proposes a motion form recognition method and system based on fusion vision and clothing AI, belonging to the technical field of artificial intelligence and computer vision. Background Art

[0002] Traditional motion form recognition methods mainly rely on sensor data, such as accelerometers and gyroscopes. Although these methods can provide relatively accurate motion data, they have problems such as high equipment cost, inconvenience in wearing, and complex data processing. In recent years, with the development of computer vision technology, vision-based motion form recognition methods have gradually become a research hotspot. However, most existing visual recognition methods are limited to the recognition of simple motion postures. For complex motion forms, especially when combined with clothing changes, their recognition accuracy and robustness still need to be improved. Summary of the invention

[0003] The present invention provides a motion form recognition method and system based on fusion vision and clothing AI to solve the problems mentioned in the above background technology:

[0004] The present invention proposes a motion form recognition method based on fusion vision and clothing AI, the method comprising:

[0005] S1. Use wearable visual sensors to collect video or image data of human movement in real time, and use clothing AI technology to perform clothing recognition on the collected video or image data and extract clothing features;

[0006] S2. Based on deep learning algorithms, human posture is recognized, key posture features are extracted, and the human motion trajectory is analyzed and dynamic features are extracted by combining time information.

[0007] S3, fusing the extracted clothing features with the motion features through a multimodal feature fusion algorithm, and correcting the fused features based on a machine learning algorithm; wherein the motion features include key posture features and dynamic features;

[0008] S4. Based on the fused features, a motion form recognition model is constructed, and the input motion data is classified through the trained model to identify the specific motion form;

[0009] S5. Feedback the recognition results to the user or system in real time, and iteratively optimize the model based on user feedback and actual application effects.

[0010] The present invention proposes a motion form recognition system based on fusion vision and clothing AI, the system comprising:

[0011] Data acquisition module: Use wearable visual sensors to collect video or image data of human movement in real time, and use clothing AI technology to perform clothing recognition and extract clothing features from the collected video or image data;

[0012] Feature extraction module: Based on deep learning algorithms, it recognizes human postures, extracts key posture features, combines time information, analyzes human motion trajectories, and extracts dynamic features;

[0013] Feature fusion module: The extracted clothing features are fused with motion features through a multimodal feature fusion algorithm, and the fused features are corrected based on a machine learning algorithm;

[0014] Model building module: Based on the fused features, a motion form recognition model is built. The input motion data is classified through the trained model to identify the specific motion form.

[0015] Real-time feedback module: Provides real-time feedback of recognition results to users or systems, and iteratively optimizes the model based on user feedback and actual application effects.

[0016] The invention has the following beneficial effects: by fusing image data collected by visual sensors and clothing AI technology, combined with deep learning algorithms for the recognition of human postures, key posture features and dynamic features can be extracted more accurately, thereby improving the accuracy of motion form recognition; fusing clothing features with motion features can provide more comprehensive data analysis, enhance the recognition ability of the model, and make the recognition result more reliable; the real-time feedback mechanism can provide users with sports guidance or health monitoring in a timely manner, improve user experience, and provide instant feedback and suggestions in practical applications, thereby enhancing the practicality of the system; the model is iteratively optimized through user feedback and actual application effects, so that the system can continuously learn and improve and adapt to different user needs and application scenarios; the multi-channel parallel transmission protocol and data compression algorithm are adopted to improve the efficiency and reliability of data transmission, and ensure the real-time and accuracy of data; the fault detection mechanism and load balancing strategy are used to improve the stability and reliability of the system, and ensure the continuity of data transmission and processing; the method is not only applicable to specific motion form recognition, but also can adapt to different types of sports and user needs by adjusting and optimizing model parameters; by collecting user feedback and behavior data, users can directly participate in the optimization and improvement process of the system, thereby improving user participation and satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a step diagram of the method of the present invention;

[0018] Figure 2 This is a system module diagram of the present invention. DETAILED DESCRIPTION

[0019] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0020] One embodiment of the present invention, as Figure 1 As shown, a motion form recognition method based on fusion vision and clothing AI, the method includes:

[0021] S1. Use wearable visual sensors to collect video or image data of human movement in real time, and use clothing AI technology to perform clothing recognition on the collected video or image data and extract clothing features;

[0022] S2. Based on the deep learning algorithm, human posture is recognized, key posture features are extracted, and the human motion trajectory is analyzed in combination with time information to extract dynamic features, including motion speed, acceleration and direction;

[0023] S3, fusing the extracted clothing features with the motion features through a multimodal feature fusion algorithm, and correcting the fused features based on a machine learning algorithm; wherein the motion features include key posture features and dynamic features;

[0024] S4. Based on the fused features, a motion form recognition model is constructed, and the input motion data is classified through the trained model to identify specific motion forms; the motion forms include running, jumping and swimming;

[0025] S5. Feedback the recognition results to the user or system in real time, and iteratively optimize the model based on user feedback and actual application effects.

[0026] The working principle of the above technical solution is as follows: through wearable visual sensors (such as smart glasses, sports cameras, etc.), video or image data of the human body in motion is captured in real time; using clothing AI technology, the collected video or image is intelligently analyzed to identify the clothing worn and extract clothing features, including color, texture and style, which together constitute a comprehensive description of the clothing; based on the deep learning algorithm, the human posture in the video or image is identified, and the key posture features are extracted, mainly including joint positions (such as shoulders, elbows, knees, etc.) and posture angles. The key posture features reflect the specific posture of the human body in motion; combined with time information, the human body motion trajectory is further analyzed to extract dynamic features, including motion speed, acceleration and direction. The dynamic features jointly describe the dynamic changes of the human body in motion; through the multimodal feature fusion algorithm, the extracted clothing features are fused with the motion features (including key posture features and dynamic features) to form a more comprehensive and rich feature set; based on the machine learning algorithm, the fused features are corrected and optimized; based on the fused features, a motion morphology recognition model is constructed. Through training and learning, the model can identify specific motion forms, such as running, jumping, swimming, etc. The input motion data is classified through the trained model to identify the corresponding motion form. This process realizes the extraction of useful information from video or image data and converts it into specific motion form recognition results; the recognition results are fed back to the user or system in real time. Users can perform intelligent sports training or health monitoring based on the recognition results, and the system can make corresponding decisions or adjustments based on the recognition results; according to user feedback and actual application effects, the motion form recognition model is iteratively optimized, including adjusting model parameters and improving algorithms.

[0027] The effects of the above technical solutions are as follows: through wearable visual sensors, video or image data of human body movements are collected in real time, ensuring the real-time nature of the data, so that the system can respond to the user's movement status immediately; based on clothing AI technology and deep learning algorithms, clothing features and human postures are accurately identified, improving the accuracy of motion form recognition; through multimodal feature fusion algorithms, clothing features and motion features (including key posture features and dynamic features) are fused to form a more comprehensive and rich feature set, providing more information support for motion form recognition, which not only improves the accuracy of recognition, but also enhances the robustness of the system, enabling it to work stably in different environments and under different clothing conditions; recognition The results can be fed back to users or systems in real time, providing strong support for application scenarios such as intelligent sports training and health monitoring. Users can adjust their training plans and monitor their health status based on the recognition results, and the system can make intelligent recommendations, early warnings and other decisions based on the recognition results. The motion morphology recognition model is iteratively optimized based on user feedback and actual application effects to continuously improve the recognition accuracy and generalization ability of the model. The continuous optimization mechanism enables the system to adapt to the needs of different users and different sports scenarios, improving the practicality and flexibility of the system. At the same time, it also provides important reference information for health management, helping users to understand their exercise status and health level in a timely manner, so as to take appropriate measures to intervene and adjust.

[0028] In one embodiment of the present invention, the S1 includes:

[0029] S11. Real-time collection of video or image data of human body motion by wearable visual sensors deployed at key parts of the human body;

[0030] S12, performing clothing recognition on the collected video or image data based on the deep learning model, and extracting clothing features, wherein the clothing features include color, texture, and style; wherein the extraction of clothing features is implemented by the following pseudo code:

[0031] #Extract clothing features

[0032] # Initialize a deep learning model (e.g., a pre-trained convolutional neural network CNN)

[0033] model=load_pretrained_cnn_model()

[0034] # Load image data (video or image frames acquired from a wearable vision sensor)

[0035] image=load_image_from_sensor()

[0036] #Preprocess the image (e.g. resize, normalize, etc.)

[0037] preprocessed_image=preprocess_image(image)

[0038] #Use deep learning model to extract features

[0039] #Note: The model has multiple output layers, corresponding to color, texture and style features

[0040] #In actual implementation, train or use different models to extract these features

[0041] #Extract color features (through additional color histogram calculation module)

[0042] #Assume that the color feature is obtained through an auxiliary output layer of the model (just for example)

[0043] color_features=model.predict_color_features(preprocessed_image)

[0044] #Or, if the color features are obtained by post-processing:

[0045] #color_histogram=calculate_color_histogram(image)

[0046] #color_features=process_color_histogram(color_histogram)

[0047] #Extract texture features (possibly through Gabor filter or LBP algorithm)

[0048] #Assume that the texture feature is obtained through an intermediate layer or auxiliary output layer of the model (just an example)

[0049] texture_features=model.predict_texture_features(preprocessed_image)

[0050] #Or, if the texture features are obtained by traditional image processing algorithms:

[0051] #gabor_filters=apply_gabor_filters(image)

[0052] #texture_features=extract_texture_from_gabor_filters(gabor_filters)

[0053] #or

[0054] #lbp_image=apply_lbp(image)

[0055] #texture_features=extract_texture_from_lbp(lbp_image)

[0056] #Extract style features (through the main classification output layer of the model)

[0057] style_features=model.predict_style_features(preprocessed_image)

[0058] #Style features are usually classification labels or classification probabilities, and may need to be post-processed to obtain more useful representations

[0059] #For example, convert the classification probability into a one-hot encoding or select the category with the highest probability as the style label

[0060] #Preprocess the extracted features (e.g., normalization, dimensionality reduction, etc.)

[0061] normalized_color_features=normalize_features(color_features)

[0062] normalized_texture_features=normalize_features(texture_features)

[0063] #Style features may already be probability distributions or one-hot encodings, so no additional normalization is needed

[0064] # Output the extracted features

[0065] output_features={

[0066] 'color_features':normalized_color_features,

[0067] 'texture_features':normalized_texture_features,

[0068] 'style_features':style_features

[0069] }

[0070] S13, and preprocessing the extracted clothing features.

[0071] The working principle of the above technical solution is as follows: wearable visual sensors are deployed at key parts of the human body (such as the head, wrists, ankles, etc.). The sensors have built-in high-definition cameras or image sensors, which can capture video or image data of the human body in real time during movement. The data contains rich movement information and background environment information; the sensors send the collected data to the data processing center or cloud server in real time through wireless transmission methods (such as Wi-Fi, Bluetooth, etc.) to ensure the timeliness and effectiveness of the data; the pre-trained deep learning model (such as convolutional neural network CNN) is used to perform clothing recognition on the collected video or image data. The model has been trained with a large number of clothing images and can accurately identify the type, color, texture and other characteristics of clothing; the color characteristics of clothing are quantified through color histograms. Color histograms are a method of statistically analyzing the distribution of colors in an image, which can intuitively reflect the main colors of clothing and their proportions; the texture characteristics of clothing are extracted using methods such as Gabor filters or local binary patterns (LBP). Gabor filter can simulate the frequency and direction selectivity of human visual system and has strong ability to capture texture information; LBP extracts texture features by comparing the gray value difference between pixels and pixels in its neighborhood; the style features of clothing are obtained through the classification results of deep learning models. The model will output the style classification results of clothing (such as T-shirts, shirts, skirts, etc.) based on the input image data as part of the style features; the extracted clothing features are cleaned to remove noise and redundant information to ensure the validity and accuracy of the features; the features such as color and texture are normalized to make them at the same order of magnitude to facilitate subsequent feature fusion and model training; if the feature dimension is too high, it may lead to increased computational complexity and storage cost. Therefore, feature dimensionality reduction techniques (such as principal component analysis PCA) can be used to reduce the number of features while retaining the information of the original features as much as possible.

[0072] The effect of the above technical solution is: by deploying wearable visual sensors at key parts of the human body, video or image data of human body movement can be collected in real time. Using the deep learning model to perform clothing recognition on the collected data, key features such as color, texture and style of clothing can be efficiently extracted. These features are not only helpful for motion form analysis, but also provide valuable information for applications such as personalized recommendation and fashion matching; color features are quantified through color histograms, making color recognition more accurate and stable; texture features are extracted through Gabor filters or local binary patterns, which can capture subtle texture changes on the surface of clothing; style features are obtained through the classification results of the deep learning model, with high recognition accuracy; preprocessing of the extracted clothing features, including data cleaning, normalization and feature dimensionality reduction, can significantly improve the quality and availability of data, reduce the impact of noise and redundant information on subsequent analysis, and improve the accuracy and efficiency of the algorithm; by accurately extracting clothing features and using them as part of motion form recognition, the robustness of the recognition algorithm can be enhanced. Clothing features not only provide an additional dimension of information, but also help the algorithm better understand and distinguish different forms of movement, especially when similar movements are difficult to distinguish; this technical solution provides strong support for application scenarios such as personalized recommendations, intelligent sports training, and health monitoring. By collecting and analyzing users' sports data and clothing features in real time, the system can provide users with more accurate and personalized suggestions and services, improving user experience and satisfaction.

[0073] In one embodiment of the present invention, the S2 includes:

[0074] S21. Perform preliminary posture recognition on human images or video frames based on a pre-trained deep learning model; based on the preliminary recognition, perform key posture feature extraction;

[0075] S22, judging the stability of the initially recognized posture, calculating the position change of the same joint point in several consecutive frames of images, if the change is within the preset threshold range, the joint point is judged to be stable, and its coordinates can be used for subsequent analysis; otherwise, the joint point is judged to be unstable, and re-identification or error correction is performed; the position change is calculated by the following formula:

[0076]

[0077] in, represents the coordinates of the joint point at time t, Represents the coordinates of the joint point at time t+1;

[0078] S23, constructing a human motion trajectory based on the initially identified joint point coordinates, wherein the human motion trajectory includes a path of each joint point changing over time, and a relative motion trajectory between the joint points;

[0079] S24, according to the motion trajectory, by calculating the change amount of the joint point coordinates between adjacent frames, the dynamic features are preliminarily extracted, and the dynamic features include motion speed, acceleration and direction; in particular, the speed feature is obtained by calculating the change rate of the joint point position over time; the acquisition method is as follows:

[0080]

[0081] in, Represents the time interval, that is, the time difference between two adjacent frames;

[0082] For the acceleration characteristics, it is obtained by calculating the rate of change of the velocity characteristics over time; the acquisition method is as follows:

[0083]

[0084] The directional feature is obtained by calculating the tangent direction of the joint point motion trajectory; the acquisition method is as follows:

[0085]

[0086] S25, judging the continuity of the motion trajectory, calculating the smoothness of the motion trajectory of the same joint point in several consecutive frames of images, if the smoothness is within a certain threshold range, then the motion trajectory of the joint point is judged to be continuous, and its dynamic characteristics can be used for subsequent analysis; otherwise, the motion trajectory of the joint point is judged to be discontinuous, and re-analysis or error correction is performed; the smoothness is evaluated by calculating the trajectory curvature, and the calculation formula of the curvature is:

[0087]

[0088] in, Represents the first-order derivative of the joint trajectory, that is, the velocity, It represents the second-order arrival time of the joint point motion trajectory, that is, the component of acceleration in the trajectory direction. In practical applications, if discrete data is processed, the curvature is approximately calculated by the following formula:

[0089]

[0090] in, Indicates the distance between two adjacent frames; , , Respectively represent the positions of the joint points of three adjacent frames;

[0091] S26. Further optimize the initially extracted key posture features, and optimize the initially extracted dynamic features.

[0092] The working principle of the above technical solution is: use a pre-trained deep learning model (such as a convolutional neural network CNN or a posture estimation network) to perform preliminary posture recognition on human images or video frames. The model can efficiently and accurately detect and locate human joints, such as the head, shoulders, elbows, wrists, hips, knees, ankles, etc.; based on the preliminary recognition, extract key posture features from the image or video frame, including the two-dimensional or three-dimensional coordinates of the joints (depending on the sensor type and data acquisition method), and the relative position relationship between the joints (such as distance, angle, etc.). For key postures in motion (such as starting, jumping, landing, etc.), special attention should be paid to and the coordinates and relative position relationship of the joints under these postures should be extracted for subsequent in-depth analysis; the stability of the initially recognized posture is judged. By calculating the position change of the same joint point in several consecutive frames of images, if the change is within the preset threshold range, the joint point is judged to be stable and its coordinates can be used for subsequent analysis; if the position change of the joint point exceeds the preset threshold, the joint point is judged to be unstable, which may be caused by noise, occlusion or recognition error. At this time, it is necessary to re-identify the joint point or perform error correction to improve the accuracy of posture recognition; based on the coordinates of the initially identified joint points, construct the human body motion trajectory. This includes the path of each joint point changing over time, and the relative motion trajectory between the joint points. According to the motion trajectory, the dynamic features are preliminarily extracted by calculating the change in the joint point coordinates between adjacent frames. The velocity feature is obtained by calculating the rate of change of the joint point position over time; the acceleration feature is obtained by calculating the rate of change of the velocity feature over time; the direction feature is obtained by calculating the tangent direction of the joint point motion trajectory. The continuity of the motion trajectory is judged. By calculating the smoothness of the motion trajectory of the same joint point in several consecutive frames of images (such as by calculating the curvature or second-order derivative of the trajectory), if the smoothness is within a certain threshold range, the motion trajectory of the joint point is judged to be continuous, and its dynamic features can be used for subsequent analysis; if the smoothness of the joint point motion trajectory exceeds the preset threshold, the motion trajectory of the joint point is judged to be discontinuous, which may be caused by noise, recognition error or motion mutation. At this time, it is necessary to re-analyze the motion trajectory of the joint point or perform error correction to ensure the accuracy of the subsequent analysis; further optimize the key posture features initially extracted. This includes steps such as removing redundant features (such as repeated or highly correlated features), feature dimensionality reduction (such as through principal component analysis PCA or linear discriminant analysis LDA) and feature normalization (such as scaling the eigenvalues ​​to the same scale). These optimization operations help reduce computational complexity, improve recognition accuracy and enhance the generalization ability of the model; optimize the initially extracted dynamic features. This includes steps such as removing noise (such as through filtering methods), feature smoothing (such as through moving average or weighted average methods) and feature selection (such as through correlation analysis or importance evaluation methods).

[0093] The effect of the above technical solution is: through the pre-trained deep learning model, it is possible to perform high-precision posture recognition on human images or video frames, accurately detect and locate key joints of the human body, and provide a solid foundation for subsequent key posture feature extraction and dynamic feature analysis. On the basis of preliminary recognition, the two-dimensional or three-dimensional coordinates of the joints and the relative position relationship between the joints are extracted in detail, with special attention paid to key postures in motion, such as starting, jumping, landing, etc. These detailed feature information helps to deeply analyze the laws and characteristics of human motion; by calculating the position change and motion trajectory smoothness of the same joint in several consecutive frames of images, the stability of the joints and the continuity of the motion trajectory can be effectively judged, which can reduce misjudgment caused by noise, occlusion or recognition error, and improve the accuracy of subsequent analysis; according to the motion trajectory, dynamic features such as motion speed, acceleration and direction are fully extracted. The key posture features and dynamic features extracted initially are optimized, including the steps of removing redundant features, feature dimensionality reduction, feature normalization, noise removal, feature smoothing and feature selection, which helps to reduce computational complexity and improve analysis efficiency, while retaining key information and enhancing the generalization ability of the model.

[0094] In one embodiment of the present invention, the S21 includes:

[0095] Preprocess the input human body image or video frame, load the pre-trained deep learning model, and convert the format and resize the preprocessed image or video frame according to the input requirements of the model;

[0096] Use the deep learning model to perform preliminary human joint detection on the preprocessed images or video frames, and automatically identify and mark the positions of the human joints;

[0097] Perform preliminary screening on the detected joint points, remove erroneous detection results, further fine-tune the initial detected joint points, and optimize the positions of the joint points based on the regression capability of the model;

[0098] According to the sensor type and data acquisition method, extract the two-dimensional or three-dimensional coordinates of the joint points. For a two-dimensional image, extract the coordinates of the joint points on the image plane; for a three-dimensional video or depth sensor data, extract the coordinates of the joint points in the three-dimensional space; and calculate the relative position relationship between the joint points, the relative position relationship includes distance and angle;

[0099] Based on the classification ability of the deep learning model, the key postures in the image or video frame are identified, and for the identified key postures, the coordinates and relative position relationships of the joint points under the key postures are extracted.

[0100] The working principle of the above technical solution is as follows: receiving human body images or video frames as input data; denoising the input data to reduce noise interference in the image or video and improve the accuracy of subsequent processing; improving the visual effect of the image or video through contrast enhancement technology to make the joints more visible; loading a pre-trained deep learning model, which should have the ability to accurately detect and locate human body joints. The model has been trained with a large number of human posture data sets to ensure its generalization performance and accuracy; according to the input requirements of the model, the pre-processed image or video frame is converted into a format (such as from RGB to the specific format required by the model) and resized (such as scaling to the resolution required by the model); using the deep learning model to perform preliminary human body joint detection on the pre-processed image or video frame. The model will automatically identify and mark the positions of the joints of the human body, such as the head, shoulders, elbows, wrists, hips, knees, ankles, etc.; perform preliminary screening of the detected joints to remove erroneous detection results, which may include abnormal positions (such as joints appearing outside the human body) or points with unreasonable relationships with other joints (such as the connection between joints does not conform to the human anatomical structure); further fine-tune the initial detected joints. Involves fine-tuning the positions of the joints to improve the accuracy of positioning; based on the regression capability of the model, the positions of the joints are optimized. The regression capability allows the model to predict the precise position of the joints based on the input data (such as images or video frames), thereby further improving the accuracy of positioning; extract the two-dimensional or three-dimensional coordinates of the joints according to the sensor type and data acquisition method. For two-dimensional images, extract the coordinates of the joints on the image plane; for three-dimensional videos or depth sensor data, extract the coordinates of the joints in three-dimensional space; calculate the relative position relationship between the joints, including distance and angle. Based on the classification ability of the deep learning model, the key postures in the image or video frame are identified, including starting, jumping, landing and other actions with specific motion characteristics; for the identified key postures, the coordinates and relative position relationship of the joint points under the key postures are extracted. These feature information is of great significance for subsequent tasks such as human motion analysis, posture recognition and behavior understanding.

[0101] The effects of the above technical solution are: through preprocessing steps such as denoising, contrast enhancement and grayscale (if applicable), the quality of the input image or video frame is improved, the impact of noise and illumination changes on posture recognition is reduced, and thus the accuracy of subsequent posture recognition is improved; the deep learning model used has been trained with a large number of human posture data sets, so that the model can accurately detect and locate human joints, and even when faced with human images or video frames of different body shapes, postures and backgrounds, it can maintain high recognition accuracy and generalization ability; after the preliminary detection of the joint points through preliminary screening and fine positioning, the position of the joint points is optimized based on the regression ability of the model, which further improves the accuracy of joint point positioning and provides a more reliable data basis for subsequent feature extraction and posture analysis; depending on the type of sensor and the data acquisition method, the technical solution can extract the two-dimensional or three-dimensional coordinates of the joint points. This enables the technical solution to be widely used in different scenarios and devices, such as two-dimensional image analysis, three-dimensional video processing, and the use of depth sensor data; based on the classification capability of the deep learning model, the technical solution can automatically identify key postures in images or video frames, improve the efficiency of posture recognition, and provide strong support for subsequent behavior analysis and pattern recognition; extract the coordinates and relative position relationships of joint points under key postures, and provide rich feature information for subsequent tasks such as human motion analysis, posture recognition, and behavior understanding.

[0102] In one embodiment of the present invention, S3 includes:

[0103] S31. Before feature fusion, clothing features and motion features are standardized, and dimension matching is performed through feature selection; after feature preprocessing and dimension matching, the quality of each feature is evaluated; the quality of the feature is evaluated by the following formula:

[0104]

[0105] in, represents the quality score of feature F; represents the variance of feature F; Represents the mutual information between feature F and target variable Y; Represents the average correlation coefficient of feature F and other features; , , They represent the maximum possible values ​​of variance, mutual information and correlation among all features, respectively, and are used for normalization; , , represents the weights of variance, mutual information and correlation coefficient, and ; , , represents the exponential parameter, which is used to introduce nonlinear relationships and can be adjusted to emphasize or weaken the impact of specific indicators; n represents the total number of features;

[0106] S32, according to the characteristics of clothing features and motion features and the needs of subsequent model training, feature fusion strategy selection is performed; after the selection is completed, the feature fusion strategy is implemented; wherein the technical solution of S32 can be implemented by the following pseudo code:

[0107] # Input clothing features and motion features

[0108] F_clothing,F_motion=get_features()

[0109] #Feature Standardization

[0110] F_clothing_std=standardize(F_clothing)

[0111] F_motion_std=standardize(F_motion)

[0112] #Feature Embedding

[0113] E_clothing=embedding_layer(F_clothing_std,Theta_clothing)

[0114] E_motion=embedding_layer(F_motion_std,Theta_motion)

[0115] #Feature Fusion

[0116] F_concat=concatenate(E_clothing,E_motion)

[0117] #Deep Learning Network

[0118] H_1=relu(dense_layer(F_concat,Theta_1))

[0119] H_2=relu(dense_layer(H_1,Theta_2))

[0120] #...More layers

[0121] H_L=dense_layer(H_{L-1},Theta_L)

[0122] #Output layer

[0123] Y_pred=output_layer(H_L,Theta_output)

[0124] #Loss calculation and parameter update

[0125] loss=loss_function(Y_pred,Y_true)

[0126] update_parameters(Theta,loss)

[0127] S33. After feature fusion, the fusion results are evaluated and the fused features are corrected using a machine learning algorithm.

[0128] The working principle of the above technical solution is as follows: before feature fusion, clothing features and motion features are first standardized. Standardization usually involves scaling the data to a common scale, such as zero-mean unit variance standardization, to ensure that different features are numerically comparable; followed by feature selection, which simplifies the model by reducing the number of features while retaining as much information as possible. Feature selection is based on statistical tests, domain knowledge, or machine learning algorithms to identify the most important features; ensuring that the dimensions of clothing features and motion features match, that is, their feature vectors have the same length, for subsequent feature fusion operations. If there is no match, the dimension is adjusted by adding zero vectors, dimension reduction or feature selection. The quality of each feature is evaluated, usually by calculating the variance of the feature (measurement of the discreteness of the data), mutual information (measurement of the correlation between the feature and the target variable), correlation coefficient (measurement of the linear relationship between features) and other indicators. For low-quality features, such as features with small variance and low correlation with other features, you can choose to discard or perform further processing, such as feature transformation or feature dimension reduction. According to the characteristics of clothing features and motion features and the needs of subsequent model training, select appropriate feature fusion strategies, including feature splicing, feature weighting and deep feature fusion. For example, if the feature dimension is low and the linear relationship is obvious, direct splicing is performed, which is mainly used for simple classification tasks, such as clothing style classification based on a small number of statistical features. If the feature dimension is high and dimension reduction is required, principal component analysis (PCA) or autoencoder is used for dimension reduction and then splicing is mainly used for the fusion of image features and text features. If there is a nonlinear relationship between features, they are fused through a deep learning network, which is mainly used for the fusion of video and audio features for emotion recognition. Clothing features and motion features are directly spliced ​​together to form a new feature vector. The features are weighted according to their importance. The weights can be learned by machine learning algorithms or set manually based on domain knowledge. The weighted feature values ​​are calculated by multiplying the weights by the original feature values. The clothing features and motion features are used as inputs to the deep learning model. The deep learning model learns the complex relationships between features through multiple layers of nonlinear transformations and outputs fused features. The fused features are evaluated to verify the effectiveness of the fusion strategy. Evaluation methods may include calculating the correlation between the fused features and the target variable, training the model using the fused features and evaluating the model performance. The fused features are further corrected using machine learning algorithms. This includes using regression models to fine-tune the feature values ​​or using classification models to optimize the relationship between the features and the target variable. The purpose of the correction is to further improve the quality of the fused features to make them more suitable for subsequent model training and application.

[0129] The effects of the above technical solution are as follows: before feature fusion, clothing features and motion features are standardized and feature selected to ensure that the features are comparable in value and reduce redundant and noise features, thereby improving the efficiency and accuracy of subsequent model training; through feature selection and dimension matching, the dimensions of clothing features and motion features are ensured to be consistent, which is convenient for subsequent feature fusion operations. It avoids the difficulty of feature fusion or information loss caused by dimension mismatch; according to the characteristics of clothing features and motion features and the needs of subsequent model training, a suitable feature fusion strategy is selected. Strategies such as feature splicing, feature weighting and deep feature fusion can make full use of the complementary information between different features and improve the quality of fused features; the fused features contain richer information, which can more comprehensively describe human posture and clothing features, and improve the generalization ability of subsequent models for tasks such as human posture recognition and clothing classification; evaluating and correcting the fused features can further optimize the feature quality and reduce the impact of noise and redundant information on model performance. Using machine learning algorithms to correct the fused features can further improve the accuracy and robustness of the model; through feature selection and feature fusion, the number of model input features can be reduced without losing too much information, thereby reducing the complexity and computational cost of the model; this technical solution is not only applicable to the fusion of clothing features and motion features, but can also be extended to the fusion of other multi-source information, such as audio, video, text, etc., providing new ideas and methods for cross-modal learning and fusion. The above evaluation formula is normalized to ensure that each indicator is compared on the same scale, which improves the fairness and consistency of the evaluation results; the correlation term takes into account the correlation of all other features, avoids the influence of a single extreme value, and better reflects the overall correlation between features through the average value; by adjusting the weight and index parameters, the importance of each indicator can be flexibly adjusted according to the specific application scenario; it helps to more accurately evaluate the quality of features, thereby optimizing the feature selection and model building process and improving model performance.

[0130] In one embodiment of the present invention, the S4 includes:

[0131] S41. Collect data containing different motion forms from multiple data sources (such as videos, sensors, user input, etc.), and clean the collected data;

[0132] S42, labeling the cleaned data, determining the motion form corresponding to each data sample, and extracting features from the cleaned and labeled data. After the feature extraction, perform feature selection;

[0133] S43. According to the characteristics of the data and the needs of subsequent applications, the motion form recognition model is selected, and the selected model is trained using the labeled data set; after the model is trained, the model is verified through the validation set, and the performance of the model is evaluated.

[0134] S44. Select a classifier based on the output of the model and the needs of subsequent applications. The classifiers include hard classifiers (such as threshold classifiers) and soft classifiers (such as probability classifiers). The hard classifier directly gives the final classification result, while the soft classifier gives the probability or score of each category and optimizes the selected classifier.

[0135] S45, using the test set to evaluate the model and calculate various evaluation indicators of the model. If the performance of the model meets the preset requirements, the model is deployed in actual applications; if the performance of the model does not meet the preset requirements, return to the previous steps for further optimization and adjustment;

[0136] The working principle of the above technical solution is: collect data from multiple data sources (such as videos, sensors, user input, etc.), including various motion forms, to ensure the generalization ability of the model; pre-process the collected data, including removing noise, filling missing values, correcting erroneous data, etc., to improve data quality. Label the cleaned data to determine the motion form corresponding to each data sample. The labeling process needs to ensure accuracy and consistency for subsequent model training; extract features from the labeled data. Feature extraction methods include time domain analysis (such as mean, variance, etc.), frequency domain analysis (such as Fourier transform), time-frequency analysis (such as short-time Fourier transform, wavelet transform, etc.) and deep learning feature extraction (such as convolutional neural network, recurrent neural network, etc.); after feature extraction, perform feature selection to remove redundant and irrelevant features and retain features that have an important impact on motion form recognition. Select a suitable motion form recognition model based on the characteristics of the data and the needs of subsequent applications. The model includes traditional machine learning models and deep learning models; the selected model is trained using the labeled data set. During the training process, adjust the parameters of the model (such as learning rate, number of iterations, regularization parameter, etc.). At the same time, monitor the training process of the model, such as the change of loss function and accuracy, to ensure that the model can correctly learn the characteristics of the data; after the model is trained, verify the model through the validation set and evaluate the performance of the model. If the performance of the model is not ideal, it can be optimized by adjusting the model parameters, increasing the training data, improving the feature extraction method, etc. At the same time, use methods such as cross-validation to evaluate the stability and generalization ability of the model; select the appropriate classifier according to the output of the model and the needs of subsequent applications. Classifiers include hard classifiers (such as threshold classifiers) and soft classifiers (such as probabilistic classifiers). Hard classifiers directly give the final classification results, while soft classifiers give the probability or score of each category; optimize the selected classifier to improve the classification accuracy and robustness. Optimization methods include adjusting the parameters of the classifier, using ensemble learning methods, etc.; use the test set to evaluate the model and calculate the various evaluation indicators of the model (such as accuracy, recall, F1 score, etc.). If the performance of the model meets the preset requirements, the model is deployed in actual applications. In actual applications, the model can process input data in real time or in batches and output motion form recognition results; if the performance of the model does not meet the preset requirements, it returns to the previous steps for further optimization and adjustment, including adjusting model parameters, adding training data, improving feature extraction methods, selecting more appropriate classifiers, etc. Through iterative optimization, the performance of the model is continuously improved.

[0137] The effect of the above technical solution is: collecting data from multiple data sources (such as videos, sensors, user input, etc.) can cover more diverse motion forms, thereby enhancing the generalization ability of the model. This diversity is not only reflected in the richness of motion types, but also includes data collected in different scenes, different lighting conditions, and different devices, making the model more robust in practical applications; cleaning and labeling the collected data are key steps in building a high-quality data set. The cleaning process can remove noise and redundant information and improve the accuracy and consistency of the data; the labeling process ensures that each data sample has a clear motion form label, providing a reliable foundation for subsequent model training; the use of a variety of feature extraction methods (such as time domain analysis, frequency domain analysis, time-frequency analysis, deep learning feature extraction, etc.) can comprehensively capture the key information of motion forms. The feature selection process further removes redundant and irrelevant features, retains features that have an important impact on motion form recognition, and thus optimizes the performance of the model; according to the characteristics of the data and the needs of subsequent applications, select a suitable motion form recognition model for training, which can adapt to different application scenarios and needs. At the same time, by adjusting the parameters of the model and optimizing the training process, the performance of the model can be further improved; after the model is trained, the model is verified by the validation set and the performance of the model is evaluated, which can ensure the reliability of the model in practical applications. At the same time, by adjusting the model parameters, increasing training data, improving the feature extraction method, etc., the performance of the model can be continuously improved; according to the output of the model and the needs of subsequent applications, select a suitable classifier and optimize the classifier. The choice of hard classifier and soft classifier depends on the specific application scenario and needs, and the optimization process can further improve the accuracy of classification; the model is evaluated using the test set and the various evaluation indicators of the model are calculated, which can fully reflect the performance of the model and ensure the stability and accuracy of the model in practical applications. If the performance of the model does not meet the preset requirements, return to the previous steps for further optimization and adjustment.

[0138] In one embodiment of the present invention, S5 includes:

[0139] S51, determining the goals and functions of the real-time feedback system, and selecting a feedback mechanism based on the goals;

[0140] S52, capturing the user's motion data (such as position, speed, acceleration, posture, etc.) in real time through the data acquisition module, and transmitting the collected data to the processing center through a multi-channel transmission protocol;

[0141] S53, setting a feedback trigger condition according to the result of motion form recognition and a preset rule, and executing a corresponding feedback operation in real time according to the set trigger condition;

[0142] S54, through the user feedback collection module, collect users' satisfaction, suggestions and improvement opinions on the real-time feedback system through user surveys, online comments, social media feedback, etc.;

[0143] S55, integrating and analyzing the user data, user feedback, and output data of the motion form recognition model collected by the real-time feedback system;

[0144] S56. Evaluate the performance of the motion form recognition model using new data sets (including data collected by the real-time feedback system), and develop a model optimization strategy based on the performance evaluation and user feedback;

[0145] S57. Retrain the model using new training data and verify the model using the validation set. Deploy the optimized model to the real-time feedback system and perform a new round of iterative optimization. During the iteration process, continuously collect user data, user feedback, and model performance data.

[0146] S58. Regularly analyze user behavior data to identify common problems and needs of users when using the real-time feedback system, and adjust the feedback mechanism based on the analysis results.

[0147] The working principle of the above technical solution is as follows: first, the core goal of the real-time feedback system is clarified, such as providing instant exercise guidance, correcting incorrect postures or monitoring abnormal behaviors; based on the goal, the specific functions that the planning system should have, such as data collection, processing, feedback execution, etc.; according to the target user group and scenario, select the appropriate feedback mechanism, such as visual feedback (screen prompts, animations), auditory feedback (voice prompts, warning sounds) or tactile feedback (vibration, pressure sensing) to ensure that the information can be accurately and timely conveyed to the user; through data acquisition modules, such as sensor networks, cameras, etc., the user's motion data, including position, speed, acceleration, posture, etc., is captured in real time; multi-channel transmission protocols, such as Wi-Fi, Bluetooth, etc., are used to efficiently and stably transmit the collected data to the processing center; feedback trigger conditions are set according to the results of motion pattern recognition and preset rules. For example, when it is recognized that the user's posture is incorrect, visual and auditory feedback is triggered; when the exercise intensity is too high, tactile feedback is triggered; when the trigger conditions are met, the corresponding feedback operations are performed in real time, such as displaying prompt information, playing voice prompts or generating vibrations, etc.; through user surveys, online comments, social media feedback and other methods, users' satisfaction, suggestions and improvement opinions on the real-time feedback system are widely collected; the collected user feedback is sorted and analyzed to identify the user's main concerns, satisfaction level and potential problems; the user data collected by the real-time feedback system, user feedback and the output data of the motion form recognition model are integrated; through data analysis, the scenarios in which the model performs poorly and the common problems mentioned in user feedback are identified to provide a basis for model optimization; using new data sets (including real The system uses the data collected by the real-time feedback system to evaluate the performance of the motion form recognition model and identify the model's weaknesses and improvement directions. Based on the performance evaluation and user feedback, a model optimization strategy is formulated, including adjusting model parameters, improving feature extraction methods, adding new training data, introducing new model architectures, etc. The model is retrained using the new training data to improve model performance. The model is verified using the validation set to ensure that the model still maintains good performance after optimization. The optimized model is deployed to the real-time feedback system for a new round of iterative optimization. User behavior data is regularly analyzed to identify common problems and needs of users when using the real-time feedback system. According to the analysis results, the feedback mechanism is adjusted, such as optimizing the feedback content, adjusting the feedback timing, or adding new feedback methods, to better meet user needs.

[0148] The effect of the above technical solution is: through the real-time feedback system, it can provide users with instant exercise guidance and posture correction, help users better master exercise skills and improve exercise effects. At the same time, it monitors abnormal behavior during exercise, promptly reminds users to avoid sports injuries, and ensures user safety; the real-time feedback system can continuously collect user data and feedback, which are used to evaluate and optimize the performance of the motion form recognition model. By continuously iterating and optimizing the model, its adaptability and accuracy in different scenarios can be improved to better meet user needs; this technical solution allows personalized adjustment of the feedback mechanism based on user feedback and behavior data. The system can be closer to the actual needs of users, provide more intimate and effective feedback, and enhance user satisfaction and loyalty; by integrating and analyzing the data collected by the real-time feedback system, enterprises can have a deep understanding of user behavior and needs, so as to make more accurate and scientific decisions, which will help improve the market competitiveness of products and promote business growth; this technical solution adopts a modular design concept, which makes the system highly flexible and scalable. As technology develops and user needs change, new functional modules can be easily added or existing functions can be upgraded to meet the ever-changing market needs; through the real-time feedback system, enterprises can understand user needs and behaviors more accurately, thereby optimizing the allocation and utilization of resources, helping to reduce unnecessary cost waste and improve the overall operational efficiency and cost-effectiveness of the enterprise; collecting user opinions through user surveys, online comments and social media feedback can not only help optimize products, but also enhance users' sense of participation and belonging, help build a closer user community, and enhance the brand's reputation and influence.

[0149] In one embodiment of the present invention, the S52 includes:

[0150] S521, performing preliminary data cleaning on the collected user data through the data collection module, and compressing the data through a data compression algorithm;

[0151] S522, transmitting the compressed data through a multi-channel parallel transmission protocol, and dynamically allocating transmission channels according to real-time requirements; for example, data with high real-time requirements are transmitted first;

[0152] S523, detecting errors in data transmission through the CRC integrated in the protocol, and retransmitting through the automatic retransmission request mechanism if an error is detected;

[0153] S524, opening multiple transmission streams, and dynamically adjusting the data transmission rate according to the network status; the network status indicator is obtained in the following manner:

[0154]

[0155] in, Indicates the average bandwidth over a period of time; Indicates the average delay over a period of time; Indicates the average packet loss rate over a period of time; Indicates the average throughput over a period of time; and Indicates the start and end time of the time window; , , , They represent the bandwidth, delay, packet loss rate and throughput changing with time t respectively; , , , Represents the weight coefficient, which represents bandwidth fluctuation, delay change, packet loss rate and throughput. The relative importance of

[0156] S525. Through the transmission performance monitoring system, key indicators such as data transmission delay, packet loss rate, and throughput are tracked in real time. Based on the monitoring data, the transmission strategy is predicted and optimized through machine learning algorithms;

[0157] S526. Based on the fault detection mechanism, once a channel failure is detected, it will immediately switch to the backup channel and implement redundant deployment at key nodes, such as multi-server load balancing and data backup.

[0158] The working principle of the above technical solution is as follows: the data acquisition module first performs preliminary processing on the user data, removes invalid or redundant data, and ensures the accuracy and consistency of the data; then, the cleaned data is compressed using a data compression algorithm (such as Huffman coding, LZW algorithm, etc.). According to the real-time requirements and data characteristics, select a multi-channel parallel transmission protocol, such as MQTT, WebSocket, etc. According to the real-time requirements, dynamically allocate transmission channels. For data with high real-time requirements, high-bandwidth, low-latency channels are allocated first to ensure that the data can reach the processing center in time; during the data transmission process, the CRC (cyclic redundancy check) mechanism integrated in the protocol is used to verify the data. CRC can detect errors in data transmission, such as bit flips, data packet damage, etc.; if an error is detected, the sender is requested to retransmit the erroneous data packet through the automatic retransmission request (ARQ) mechanism. In order to further improve the transmission efficiency, multiple transmission streams are opened at the same time to realize parallel transmission of data; the data transmission rate is dynamically adjusted according to the network conditions (such as bandwidth fluctuations, delay changes). When the network is in good condition, increase the transmission rate; when the network is congested, reduce the transmission rate to avoid network congestion and ensure smooth transmission; through the transmission performance monitoring system, track key indicators such as data transmission delay, packet loss rate, and throughput in real time; based on the monitoring data, use machine learning algorithms to predict and optimize the transmission strategy. For example, according to the historical data transmission situation, dynamically adjust channel allocation, transmission rate and other strategies to improve transmission efficiency and reliability; through the fault detection mechanism, monitor the status of the transmission channel in real time. Once a channel failure is detected, the fault handling process is immediately triggered; when a channel failure is detected, immediately switch to the backup channel to ensure the continuity and reliability of data transmission; implement redundant deployment at key nodes, such as multi-server load balancing, data backup, etc.

[0159] The effect of the above technical solution is: through the data cleaning and compression steps, redundant data can be effectively removed, the amount of data transmission can be reduced, and the efficiency of data transmission can be improved. Data compression algorithms such as Huffman coding and LZW algorithm can identify and compress repeated patterns in data, further reducing the amount of data; the use of multi-channel parallel transmission protocols such as MQTT and WebSocket can dynamically allocate transmission channels according to real-time requirements, ensure that high real-time data is transmitted first, and improve the real-time performance of data transmission; CRC checksum and ARQ mechanism can detect and correct errors in data transmission, ensure the integrity and accuracy of data, and enhance the reliability of data transmission; by opening multiple transmission streams and dynamically adjusting the data transmission rate according to network conditions, network congestion can be effectively avoided and the smoothness of transmission can be ensured. It helps to optimize the utilization of network resources and improve the overall performance of the network; through the transmission performance monitoring system, the key indicators of data transmission can be tracked in real time, and the transmission strategy can be predicted and optimized by using machine learning algorithms, which can improve the adaptive ability of the system. The system can dynamically adjust the transmission strategy according to the network conditions and data transmission requirements to ensure the efficiency and stability of data transmission; the fault detection mechanism and redundant deployment strategy can enhance the fault tolerance and stability of the system. Once a channel failure is detected, the system can immediately switch to the backup channel to ensure the continuity of data transmission. At the same time, implementing redundant deployment at key nodes, such as multi-server load balancing and data backup, can further improve the reliability and stability of the system. The above formula provides a quantitative network status assessment, which helps to realize the intelligent and automated management of data transmission, thereby improving the overall performance and efficiency of the network. In addition, the above formula can provide a comprehensive network status assessment by comprehensively considering the four key network performance indicators of bandwidth, delay, packet loss rate and throughput. This comprehensive assessment helps to more accurately reflect the actual performance of the network, so as to make more appropriate transmission strategy adjustments; by calculating the network status indicators in real time, the system can dynamically adjust the data transmission rate and routing according to the current network status, increase the transmission rate when the network conditions are good, and reduce the transmission rate or take other measures when the network conditions are poor to ensure the reliability of data transmission; by optimizing the transmission strategy, the efficiency of data transmission can be improved. For example, when the bandwidth is sufficient and the packet loss rate is low, the transmission rate can be increased to reduce the transmission delay; and when the network is congested or the packet loss rate is high, flow control measures can be taken to reduce data loss; because the network status is dynamically changing, the formula enables the system to quickly adapt to network changes by real-time monitoring and evaluation of the network status, thereby enhancing the adaptability and robustness of the system. This is particularly important for ensuring the quality of data transmission for critical applications; , , , , the system can pay different attention to different network indicators according to the actual needs and the importance of network conditions, which helps to optimize the allocation of network resources and improve the overall network performance; through real-time monitoring and evaluation of network conditions, the system can promptly discover potential network problems and take preventive measures or quickly recover to reduce the impact of network failures on data transmission; in the case of multi-channel parallel transmission, the formula can help the system dynamically allocate transmission channels and priorities according to network conditions and real-time requirements of data, ensuring the timely transmission of high-priority data; by optimizing data transmission strategies, it can reduce unnecessary network resource consumption, reduce operating costs, and improve service quality.

[0160] In one embodiment of the present invention, the S526 includes:

[0161] The status of each transmission channel is continuously monitored through the heartbeat packet mechanism. When abnormal parameters of a channel are detected (such as a significant increase in latency or a sharp increase in error rate), an early warning is triggered and the fault response process is automatically started;

[0162] Through machine learning algorithms, historical fault data is analyzed to establish a fault mode library, identify the current fault type, and locate the specific location of the fault in combination with the network topology map;

[0163] If a channel failure is detected, it will automatically switch to the backup channel according to the preset priority, and start the fault recovery program to repair the faulty channel, such as restarting the network equipment and reconfiguring the route; and after the faulty channel is restored, perform data transmission test;

[0164] Implement load balancing strategies at key nodes of data transmission, such as server clusters, to disperse the pressure of data transmission, and regularly back up transmitted data to redundant storage systems.

[0165] The working principle of the above technical solution is as follows: the system continuously monitors the status of each transmission channel through the heartbeat packet mechanism, including key parameters such as connection status, data transmission rate, error rate, etc.; the heartbeat packet is a short data packet sent periodically to confirm the connection status of the channel and the existence of data transmission activities; when the system detects abnormal parameters such as a significant increase in the heartbeat packet response delay of a certain channel and a sharp increase in the error rate, the early warning mechanism will be triggered immediately; once the early warning is triggered, the system will send an alarm to the administrator or monitoring system to indicate the existence of a potential channel failure; at the same time, the system will automatically start the fault response process to prepare for subsequent fault handling; the system uses machine learning algorithms to analyze historical fault data and establish a fault mode library; by comparing and matching the current fault data, the machine learning algorithm can identify the type of fault; combined with the network topology map, the system can further locate the specific location where the fault occurs; the network topology map shows the connection relationship between various devices, transmission lines and servers in the network; by analyzing the fault data and the network topology map, the system can determine whether the fault occurs in the network device, transmission line or server If a channel failure is detected, the system will automatically switch to the backup channel according to the preset priority. The backup channel is pre-configured to take over the data transmission task when the main channel fails. During the switching process, the system will ensure the continuity and integrity of data transmission to avoid data loss or interruption. After switching to the backup channel, the system will start the fault recovery program to try to repair the faulty channel. The fault recovery program may include restarting network equipment, reconfiguring routing and other operations. After the repair is completed, the system will perform a data transmission test to ensure that the faulty channel has resumed normal function. At key nodes of data transmission, such as server clusters, the system implements a load balancing strategy. The load balancing strategy can disperse the pressure of data transmission and avoid single-point overload and bottleneck problems. By evenly distributing data transmission tasks to multiple servers, the system can improve the overall data transmission efficiency and stability. The system regularly backs up transmission data to a redundant storage system. A redundant storage system is an additional storage device or system used to store data backups. When data is lost or damaged, the system can quickly recover data from the redundant storage system to ensure data integrity and availability.

[0166] The effect of the above technical solution is: through the heartbeat packet mechanism, the system can monitor the status of each transmission channel in real time, including key parameters such as connection status, data transmission rate and error rate; once the channel parameter abnormality is detected, such as a significant increase in delay or a sharp increase in error rate, the system will immediately trigger an early warning and automatically start the fault response process. This real-time monitoring and early warning mechanism can quickly discover potential problems and gain valuable time for subsequent fault handling; by using machine learning algorithms to analyze historical fault data and establish a fault mode library, the system can more accurately identify the current fault type; combined with the network topology map, the system can quickly locate the specific location of the fault, whether it is a network device failure, a transmission line problem or a server abnormality, it can be handled in a timely manner. This rapid identification and positioning capability helps to reduce troubleshooting time and improve fault handling efficiency; if a channel fault is detected, the system will automatically switch to the backup channel according to the preset priority to ensure the continuity and stability of data transmission; at the same time, the system will start the fault recovery program and try to repair the faulty channel, such as restarting the network device, reconfiguring the route, etc.; after the faulty channel is restored, the system will also perform data transmission tests to ensure that the channel has resumed normal function. This automatic switching and fault recovery mechanism can minimize data loss and interruption and ensure the reliability of data transmission. At key nodes of data transmission, such as server clusters, the system implements a load balancing strategy to disperse the pressure of data transmission and avoid single-point overload and bottleneck problems. At the same time, the system regularly backs up the transmitted data to a redundant storage system to ensure rapid recovery when data is lost or damaged. This load balancing and data backup strategy helps to improve the overall performance and stability of the system and ensure data integrity and availability.

[0167] One embodiment of the present invention, as Figure 2 As shown, a motion form recognition system based on fusion vision and clothing AI, the system includes:

[0168] Data acquisition module: Use wearable visual sensors to collect video or image data of human movement in real time, and use clothing AI technology to perform clothing recognition and extract clothing features from the collected video or image data;

[0169] Feature extraction module: Based on deep learning algorithms, it recognizes human postures, extracts key posture features, combines time information, analyzes human motion trajectories, and extracts dynamic features;

[0170] Feature fusion module: The extracted clothing features are fused with motion features through a multimodal feature fusion algorithm, and the fused features are corrected based on a machine learning algorithm; the motion features include key posture features and dynamic features;

[0171] Model building module: Based on the fused features, a motion form recognition model is built, and the input motion data is classified through the trained model to identify specific motion forms; the motion forms include running, jumping and swimming;

[0172] Real-time feedback module: Provides real-time feedback of recognition results to users or systems, and iteratively optimizes the model based on user feedback and actual application effects.

[0173] The working principle of the above technical solution is as follows: through wearable visual sensors (such as smart glasses, sports cameras, etc.), video or image data of the human body in motion is captured in real time; using clothing AI technology, the collected video or image is intelligently analyzed to identify the clothing worn and extract clothing features, including color, texture and style, which together constitute a comprehensive description of the clothing; based on the deep learning algorithm, the human posture in the video or image is identified, and the key posture features are extracted, mainly including joint positions (such as shoulders, elbows, knees, etc.) and posture angles. The key posture features reflect the specific posture of the human body in motion; combined with time information, the human body motion trajectory is further analyzed to extract dynamic features, including motion speed, acceleration and direction. The dynamic features jointly describe the dynamic changes of the human body in motion; through the multimodal feature fusion algorithm, the extracted clothing features are fused with the motion features (including key posture features and dynamic features) to form a more comprehensive and rich feature set; based on the machine learning algorithm, the fused features are corrected and optimized; based on the fused features, a motion morphology recognition model is constructed. Through training and learning, the model can identify specific motion forms, such as running, jumping, swimming, etc. The input motion data is classified through the trained model to identify the corresponding motion form. This process realizes the extraction of useful information from video or image data and converts it into specific motion form recognition results; the recognition results are fed back to the user or system in real time. Users can perform intelligent sports training or health monitoring based on the recognition results, and the system can make corresponding decisions or adjustments based on the recognition results; according to user feedback and actual application effects, the motion form recognition model is iteratively optimized, including adjusting model parameters and improving algorithms.

[0174] The effects of the above technical solutions are as follows: wearable visual sensors are used to collect video or image data of human motion in real time, ensuring the real-time nature of the data, so that the system can respond to the user's motion status instantly; based on clothing AI technology and deep learning algorithms, clothing features and human postures are accurately identified, improving the accuracy of motion form recognition; through multimodal feature fusion algorithms, clothing features and motion features (including key posture features and dynamic features) are fused to form a more comprehensive and rich feature set, providing more information support for motion form recognition, which not only improves the accuracy of recognition, but also enhances the robustness of the system, enabling it to work stably in different environments and under different clothing conditions; the recognition results can be fed back to the user or system in real time, providing intelligent It provides strong support for application scenarios such as sports training and health monitoring. Users can adjust their training plans and monitor their health status based on the recognition results, and the system can make intelligent recommendations, early warnings and other decisions based on the recognition results. The motion morphology recognition model is iteratively optimized based on user feedback and actual application effects, and the recognition accuracy and generalization ability of the model are continuously improved. Through the continuous optimization mechanism, the system can adapt to the needs of different users and different sports scenarios, improving the practicality and flexibility of the system. This technical solution provides scientific data support for sports training, making the training process more accurate and efficient. At the same time, it also provides important reference information for health management, which helps users to understand their exercise status and health level in a timely manner, so as to take corresponding measures to intervene and adjust.

[0175] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A motion form recognition method based on fusion vision and clothing AI, characterized in that: The method comprises: S1. Use wearable visual sensors to collect video or image data of human movement in real time, and use clothing AI technology to perform clothing recognition on the collected video or image data and extract clothing features; S2. Based on deep learning algorithms, human posture is recognized, key posture features are extracted, and the human motion trajectory is analyzed and dynamic features are extracted by combining time information. S3, fusing the extracted clothing features with the motion features through a multimodal feature fusion algorithm, and correcting the fused features based on a machine learning algorithm; wherein the motion features include key posture features and dynamic features; S4. Based on the fused features, a motion form recognition model is constructed, and the input motion data is classified through the trained model to identify the specific motion form; S5. Feedback the recognition results to the user or system in real time, and iteratively optimize the model based on user feedback and actual application effects.

2. According to claim 1, a motion form recognition method based on fusion vision and clothing AI is characterized in that: Said S1 comprises: S11. Real-time collection of video or image data of human body motion by wearable visual sensors deployed at key parts of the human body; S12, performing clothing recognition on the collected video or image data based on a deep learning model, and extracting clothing features, wherein the clothing features include color, texture, and style; S13, and preprocessing the extracted clothing features.

3. According to claim 1, a motion form recognition method based on fusion vision and clothing AI is characterized in that: The S2 comprises: S21. Perform preliminary posture recognition on human images or video frames based on a pre-trained deep learning model; based on the preliminary recognition, perform key posture feature extraction; S22, judging the stability of the initially recognized posture, calculating the position change of the same joint point in several consecutive frames of images, and if the change is within a preset threshold range, the joint point is judged to be stable, and its coordinates can be used for subsequent analysis; otherwise, the joint point is judged to be unstable, and re-recognition or error correction is performed; S23, constructing a human motion trajectory based on the initially identified joint point coordinates, wherein the human motion trajectory includes a path of each joint point changing over time, and a relative motion trajectory between the joint points; S24, preliminarily extracting dynamic features according to the motion trajectory by calculating the change in the coordinates of the joint points between adjacent frames, wherein the dynamic features include motion speed, acceleration and direction; S25, judging the continuity of the motion trajectory, calculating the smoothness of the motion trajectory of the same joint point in several consecutive frames of images, if the smoothness is within a certain threshold range, then the motion trajectory of the joint point is judged to be continuous, and its dynamic characteristics can be used for subsequent analysis; otherwise, the motion trajectory of the joint point is judged to be discontinuous, and re-analysis or error correction is performed; S26. Further optimize the initially extracted key posture features, and optimize the initially extracted dynamic features.

4. According to claim 3, a motion form recognition method based on fusion vision and clothing AI is characterized in that: The S21 includes: Preprocess the input human body image or video frame, load the pre-trained deep learning model, and convert the format and resize the preprocessed image or video frame according to the input requirements of the model; Use the deep learning model to perform preliminary human joint detection on the preprocessed images or video frames, and automatically identify and mark the positions of the human joints; Perform preliminary screening on the detected joint points, remove erroneous detection results, further locate the initially detected joint points, and optimize the positions of the joint points based on the regression capability of the model; According to the sensor type and data acquisition method, extract the two-dimensional or three-dimensional coordinates of the joint points. For a two-dimensional image, extract the coordinates of the joint points on the image plane; for a three-dimensional video or depth sensor data, extract the coordinates of the joint points in the three-dimensional space; and calculate the relative position relationship between the joint points, the relative position relationship includes distance and angle; Based on the classification ability of the deep learning model, the key postures in the image or video frame are identified, and for the identified key postures, the coordinates and relative position relationships of the joint points under the key postures are extracted.

5. According to claim 1, a motion form recognition method based on fusion vision and clothing AI is characterized in that: The S3 includes: S31. Before feature fusion, clothing features and motion features are standardized and dimension matching is performed through feature selection; after feature preprocessing and dimension matching, the quality of each feature is evaluated; S32, selecting a feature fusion strategy according to the characteristics of clothing features and motion features and the needs of subsequent model training; and implementing the feature fusion strategy after the selection is completed; S33. After feature fusion, the fusion results are evaluated and the fused features are corrected using a machine learning algorithm.

6. According to claim 1, a motion form recognition method based on fusion vision and clothing AI is characterized in that: The S4 comprises: S41, collecting data containing different motion forms from multiple data sources, and cleaning the collected data; S42, labeling the cleaned data, determining the motion form corresponding to each data sample, and extracting features from the cleaned and labeled data; after feature extraction, performing feature selection; S43. According to the characteristics of the data and the requirements of subsequent applications, a motion form recognition model is selected, and the selected model is trained using the labeled data set; after the model is trained, the model is verified using the validation set, and the performance of the model is evaluated; S44, selecting a classifier according to the output of the model and the requirements of subsequent applications, and optimizing the selected classifier; S45. Use the test set to evaluate the model and calculate the various evaluation indicators of the model. If the performance of the model meets the preset requirements, the model is deployed in actual applications; if the performance of the model does not meet the preset requirements, return to the previous steps for further optimization and adjustment.

7. According to claim 1, a motion form recognition method based on fusion vision and clothing AI is characterized in that: The S5 comprises: S51, determining the goals and functions of the real-time feedback system, and selecting a feedback mechanism based on the goals; S52, capturing the user's motion data in real time through a data acquisition module, and transmitting the collected data to a processing center through a multi-channel transmission protocol; S53, setting a feedback trigger condition according to the result of motion form recognition and a preset rule, and executing a corresponding feedback operation in real time according to the set trigger condition; S54, collecting user satisfaction, suggestions and improvement opinions on the real-time feedback system through the user feedback collection module; S55, integrating and analyzing the user data, user feedback, and output data of the motion form recognition model collected by the real-time feedback system; S56. Evaluate the performance of the motion form recognition model using a new data set, and develop a model optimization strategy based on the performance evaluation and user feedback; S57. Retrain the model using new training data and verify the model using the validation set. Deploy the optimized model to the real-time feedback system and perform a new round of iterative optimization. During the iteration process, continuously collect user data, user feedback, and model performance data. S58. Regularly analyze user behavior data to identify common problems and needs of users when using the real-time feedback system, and adjust the feedback mechanism based on the analysis results.

8. According to claim 7, a motion form recognition method based on fusion vision and clothing AI is characterized in that: The S52 includes: S521, performing preliminary data cleaning on the collected user data through the data collection module, and compressing the data through a data compression algorithm; S522, transmitting the compressed data through a multi-channel parallel transmission protocol, and dynamically allocating transmission channels according to real-time requirements; S523, detecting errors in data transmission through the CRC integrated in the protocol, and retransmitting through the automatic retransmission request mechanism if an error is detected; S524, opening multiple transmission streams, and dynamically adjusting the data transmission rate according to the network conditions; S525. Through the transmission performance monitoring system, the key indicators of data transmission are tracked in real time. Based on the monitoring data, the transmission strategy is predicted and optimized through machine learning algorithms; S526. Based on the fault detection mechanism, once a channel failure is detected, the system immediately switches to the backup channel and implements redundant deployment at key nodes.

9. According to claim 8, a motion form recognition method based on fusion vision and clothing AI is characterized in that: The S526 includes: The status of each transmission channel is continuously monitored through the heartbeat packet mechanism. When a parameter of a channel is detected to be abnormal, an early warning is triggered and the fault response process is automatically started; Through machine learning algorithms, historical fault data is analyzed to establish a fault mode library, identify the current fault type, and locate the specific location of the fault in combination with the network topology map; If a channel failure is detected, it will automatically switch to the backup channel according to the preset priority, start the fault recovery program, repair the faulty channel, and perform data transmission test after the faulty channel is restored; At key nodes of data transmission, implement load balancing strategies to disperse data transmission pressure, and regularly back up transmission data to redundant storage systems.

10. A motion form recognition system based on fusion vision and clothing AI, characterized in that: The system comprises: Data acquisition module: Use wearable visual sensors to collect video or image data of human movement in real time, and use clothing AI technology to perform clothing recognition and extract clothing features from the collected video or image data; Feature extraction module: Based on deep learning algorithms, it recognizes human postures, extracts key posture features, combines time information, analyzes human motion trajectories, and extracts dynamic features; Feature fusion module: The extracted clothing features and motion features are fused through a multimodal feature fusion algorithm, and the fused features are corrected based on a machine learning algorithm; Model building module: Based on the fused features, a motion form recognition model is built. The input motion data is classified through the trained model to identify the specific motion form. Real-time feedback module: Provides real-time feedback of recognition results to users or systems, and iteratively optimizes the model based on user feedback and actual application effects.

Images