Search

CN-122020432-A - Track and field training data management method and system based on visual analysis

CN122020432ACN 122020432 ACN122020432 ACN 122020432ACN-122020432-A

Abstract

The application provides an athletic training data management method and system based on visual analysis, which are characterized by extracting text feature vectors and human body posture data in a multi-mode data set, carrying out time sequence decomposition and alignment on action stages in a training process according to the text feature vectors and the human body posture data to obtain joint action features of each action stage, carrying out variable self-coding on all joint feature vectors, restraining mutual information capacity of discrete potential variables and continuous potential variables, decoupling all joint features into a cross-sample common technical code and an individual state code of individual state deviation, and carrying out weight fitting adjustment on the common technical code and the individual state code to obtain a visual analysis report of common technical scores and individual state indexes. By adopting the scheme of the application, the effective migration of personalized and common knowledge can be realized according to the common characteristics and individual characteristics of the track and field training data so as to improve the accuracy of training analysis.

Inventors

  • CHEN XIAOCHUN
  • ZHANG XUEMIN

Assignees

  • 贵州理工学院

Dates

Publication Date
20260512
Application Date
20260413

Claims (10)

  1. 1. A track and field training data management method based on visual analysis is characterized by comprising the following steps: collecting a multi-modal data set in the training process of track and field athletes, and extracting text feature vectors and human body posture data in the multi-modal data set; Performing time sequence decomposition and alignment on action stages in the training process according to the text feature vector and the human body posture data to obtain joint action features of all the action stages in the track and field training process; Performing variation self-coding on all the combined feature vectors, constructing discrete potential variables and continuous potential variables in the coding process, and restraining mutual information capacity of the discrete potential variables and the continuous potential variables so as to decouple all the combined features into a cross-sample commonality technical code and an individual state code of individual state deviation of the track and field athletes; performing emphasis fitting adjustment of sharing technology and personalized knowledge on the commonality technical code and the individual state code to further obtain a visualized analysis report of the commonality technical score and the individual state index of the track and field athlete; And superposing the visual analysis report on the training video of the track and field athlete, and loading the visual analysis report on a display terminal.
  2. 2. The method of claim 1, wherein collecting the multimodal dataset during the athletic training session specifically comprises: disposing a plurality of high-definition cameras and wearable sensors on an athletic training field, and synchronously collecting video stream data and inertial measurement unit data of athletic training actions; collecting technical comment documents and injury record documents written in the current training; and performing time alignment on the video stream data, the inertia measurement unit data and the document data to obtain a multi-mode data set.
  3. 3. The method of claim 1, wherein extracting text feature vectors and human body pose data in the multimodal dataset comprises: Detecting human body key points of the video stream data in the multi-mode data set, and extracting two-dimensional coordinate sequences of the joint points of the track and field athletes in each video frame to obtain human body posture data; Performing word segmentation and entity identification on the technical comment document and the injury record document in the multi-mode data set, and marking; And carrying out semantic coding on the marked technical comment document and the marked injury record document to obtain a text feature vector.
  4. 4. The method of claim 1, wherein the performing time sequence decomposition and alignment on the motion phases in the training process according to the text feature vector and the human body posture data to obtain the joint motion characteristics of each motion phase in the track and field training process specifically comprises: Performing speed curve analysis and angle change rate detection on a skeleton joint point coordinate sequence in the human body posture data, and identifying starting and ending time points of each action stage to obtain a plurality of time sequence boundaries; Extracting technical description fragments which are semantically matched with each action stage from the text feature vector according to all time sequence boundaries; And dividing the human body posture data into a plurality of subsequences according to a time sequence boundary, and carrying out feature cascading on each subsequence and a corresponding technical description fragment to obtain the combined action features of each action stage in the track and field training process.
  5. 5. The method of claim 1, wherein performing variable self-encoding on all joint feature vectors and constructing discrete latent variables and continuous latent variables during encoding specifically comprises: inputting the joint feature vectors of each action stage into an encoder network to perform iterative training of variable-division self-coding; Mapping a plurality of full connection layers of the encoder network into a first statistical parameter and a second statistical parameter of a hidden space; sampling from the first statistical parameter to generate discrete potential variables with dimensions equal to the number of preset standard technical primitive categories; Sampling from the second statistical parameter to generate a continuous latent variable.
  6. 6. The method of claim 1, wherein constraining the mutual information capacity of the discrete latent variable and the continuous latent variable comprises: determining the current coded mutual information capacity of the discrete potential variables, and comparing the current coded mutual information capacity with a preset discrete capacity upper limit; determining the current coded mutual information capacity of the continuous potential variable, and comparing the current coded mutual information capacity with a preset continuous capacity upper limit; When the mutual information capacity of the discrete potential variables exceeds the discrete capacity upper limit, enhancing constraint to enable the discrete potential variables to only retain core information related to action phase categories; When the mutual information capacity of the continuous latent variable exceeds the continuous capacity upper limit, the constraint is enhanced so that the continuous latent variable only retains the change information related to individual state fluctuation.
  7. 7. The method of claim 1, wherein decoupling all joint features into a commonality technique code across samples and an individual status code of individual status deviations of the track and field athlete comprises: selecting an index position corresponding to the maximum probability value from the discrete potential variables; Mapping the index position into an action primitive number in a preset standard technical library, and generating a cross-sample commonality technical code; And taking the numerical value of each dimension in the continuous latent variable as the individual state code of the individual state deviation of the track and field athlete.
  8. 8. The method of claim 1, wherein performing a weighted fit adjustment of shared technology and personalized knowledge on the commonality technical code and the individual status code to obtain a visual analysis report of commonality technical scores and individual status indicators of the track and field athlete specifically comprises: Determining the fitting degree of the individual state codes and the multi-mode data set in the initial training stage; taking the fitting degree as an optimization target, and carrying out first round of emphasis fitting updating on the individual state codes; gradually increasing similarity constraint between the individual state codes and the commonality technical codes at a later stage of training; introducing the commonality technical code as a regular term into an iterative updating process of the individual state code, and performing a second round of emphasis fitting updating on the individual state code; Calculating similarity scores between training actions of the track and field athletes and standard technical primitives based on the updated individual state codes, and obtaining common technical scores of all action stages; And extracting the numerical value of each dimension in the updated individual state code to obtain an individual state index curve of the track and field athlete, and obtaining a visual analysis report.
  9. 9. The method of claim 1, wherein overlaying the visual analysis report on the athletic athlete's training video and loading at a display terminal specifically comprises: Suspending and displaying the commonality technical scores in the visual analysis report in a digital label form at the side of the body part of the track and field athlete in a video frame; The individual state indexes in the visual analysis report are presented at the bottom of a video playing interface in a time axis form, and are synchronously updated in a rolling way along with the video progress; and generating an interactive training report page, and supporting a coach to view the comparison analysis results of a plurality of groups of training videos on the same interface.
  10. 10. An athletic training data management system based on visual analysis for performing an athletic training data management method based on visual analysis as claimed in any one of claims 1 to 9, comprising: the acquisition module is used for acquiring a multi-modal data set in the training process of the track and field athlete and extracting text feature vectors and human body posture data in the multi-modal data set; the processing module is used for carrying out time sequence decomposition and alignment on the action stages in the training process according to the text feature vector and the human body posture data to obtain the joint action characteristics of each action stage in the track and field training process; the processing module is further used for performing variation self-coding on all the combined feature vectors, constructing discrete potential variables and continuous potential variables in the coding process, restraining mutual information capacity of the discrete potential variables and the continuous potential variables, and further decoupling all the combined features into a cross-sample commonality technical code and an individual state code of individual state deviation of the track and field athlete; The processing module is also used for carrying out emphasis fitting adjustment on shared technology and personalized knowledge on the common technology code and the individual state code so as to obtain a visual analysis report of the common technology score and the individual state index of the track and field athlete; and the execution module is used for superposing the visual analysis report on the training video of the track and field athlete and loading the visual analysis report on a display terminal.

Description

Track and field training data management method and system based on visual analysis Technical Field The application relates to the technical field of visual analysis, in particular to an athletic training data management method and system based on visual analysis. Background With the rapid development of computer vision and wearable sensing technology, track and field training is gradually changed from empirical guidance to data-driven analysis, and technical actions of athletes are quantitatively evaluated and visually presented, so that a coach is helped to accurately find problems and optimize a training scheme. The conventional track and field training data management method generally extracts a human body key point sequence from a video through an attitude estimation technology, then combines sensor data to perform action phase division and kinematic parameter calculation, and finally presents the action phase division and the kinematic parameter calculation to a coach in a curve, however, the conventional track and field training data management method cannot effectively decouple common technical characteristics of athlete actions and individual state characteristics, so that technical estimation and fatigue monitoring are interfered with each other, secondly, the conventional track and field training data management method lacks self-adaptive adjustment capability, fusion weights cannot be dynamically adjusted when the quality of a certain data source is reduced, analysis distortion is easily caused by noise introduction, and further, a model is difficult to simultaneously consider individual suitability and generalization capability. Therefore, how to realize the effective migration of individuation and common knowledge according to the common characteristics and individual characteristics of track and field training data so as to improve the accuracy of training analysis becomes a difficult problem facing the industry. Disclosure of Invention The application provides an athletic training data management method and system based on visual analysis, which can realize effective migration of individuation and common knowledge according to common characteristics and individual characteristics of athletic training data so as to improve the accuracy of training analysis. In a first aspect, the present application provides an athletic training data management method based on visual analysis, including the steps of: collecting a multi-modal data set in the training process of track and field athletes, and extracting text feature vectors and human body posture data in the multi-modal data set; Performing time sequence decomposition and alignment on action stages in the training process according to the text feature vector and the human body posture data to obtain joint action features of all the action stages in the track and field training process; Performing variation self-coding on all the combined feature vectors, constructing discrete potential variables and continuous potential variables in the coding process, and restraining mutual information capacity of the discrete potential variables and the continuous potential variables so as to decouple all the combined features into a cross-sample commonality technical code and an individual state code of individual state deviation of the track and field athletes; performing emphasis fitting adjustment of sharing technology and personalized knowledge on the commonality technical code and the individual state code to further obtain a visualized analysis report of the commonality technical score and the individual state index of the track and field athlete; And superposing the visual analysis report on the training video of the track and field athlete, and loading the visual analysis report on a display terminal. In some embodiments, collecting a multimodal dataset during an athletic training process specifically includes: disposing a plurality of high-definition cameras and wearable sensors on an athletic training field, and synchronously collecting video stream data and inertial measurement unit data of athletic training actions; collecting technical comment documents and injury record documents written in the current training; and performing time alignment on the video stream data, the inertia measurement unit data and the document data to obtain a multi-mode data set. In some embodiments, extracting the text feature vector and the human body pose data in the multimodal dataset specifically includes: Detecting human body key points of the video stream data in the multi-mode data set, and extracting two-dimensional coordinate sequences of the joint points of the track and field athletes in each video frame to obtain human body posture data; Performing word segmentation and entity identification on the technical comment document and the injury record document in the multi-mode data set, and marking; And carrying out semantic coding on the marked technical commen