CN-121256393-B - Sprint starting training method based on action decomposition

CN121256393BCN 121256393 BCN121256393 BCN 121256393BCN-121256393-B

Abstract

The application relates to the technical field of motion recognition, in particular to a sprint starting training method based on motion decomposition, which comprises the following steps: extracting starting motion characteristics, inputting the starting motion characteristics into an individuation model for executing stage judgment, calculating the fitness score of the starting motion characteristics and preset motion characteristics in the individuation model according to stage judgment results, and carrying out fitness judgment according to the fitness score. According to the application, the sprint starting process is divided into three stages, staged identification and targeted analysis are realized, the matching degree of actions and a preset model is quantized through a fitness grading and fuzzy interval judging mechanism, and samples in a fuzzy interval are automatically marked and manually rechecked, so that misjudgment is reduced and self-adaptability is enhanced. Meanwhile, in the gesture similarity judgment, a source joint positioning mechanism is introduced, key joints causing gesture deviation are identified, and individual correction suggestions in the aspects of joint angle, force-generating time sequence and gravity center are output, so that targeted training guidance of sprint starting is realized.

Inventors

ZHOU YONGPING
ZHOU XIAOXIANG
LIU DADUO

Assignees

吉林师范大学

Dates

Publication Date: 20260512
Application Date: 20251016

Claims (7)

1. The sprint starting training method based on action decomposition is characterized by comprising the following steps of: Acquiring multisource training data, extracting a human body boundary frame of an athlete, and generating an individualized motion sequence by adopting an improved key point tracking algorithm; Extracting a first parameter and a second parameter of a starting action according to the personalized motion sequence, and performing self-supervision learning by using a time sequence convolution network to construct a personalized model; Extracting starting motion characteristics, inputting the starting motion characteristics into an individuation model execution stage judgment, calculating the fitness score of the starting motion characteristics and preset motion characteristics in the individuation model according to a stage judgment result, carrying out fitness judgment according to the fitness score, entering gesture similarity judgment when the fitness score is more than or equal to a first threshold value, judging that deviation exists in starting motion and outputting a stage adjustment suggestion when the fitness score is less than the first threshold value, marking the starting motion as a sample to be rechecked when the fitness score is in a preset fuzzy interval, and feeding back to the individuation model execution parameter update after manual confirmation; the method comprises the steps of executing gesture similarity judgment, calculating a similarity index of starting motion characteristics and preset gestures in an individuation model, judging that starting motion is normal when the similarity index is more than or equal to a second threshold value, positioning a source joint causing abnormal starting motion when the similarity index is less than the second threshold value, and outputting correction advice; the first parameter is a time sequence parameter of joint angle and angular velocity, and is used for representing the flexion and extension amplitude and the change rate of main joints of lower limbs, and specifically comprises time sequence data of angle changes and angular velocities of hip joints, knee joints and ankle joints on a sagittal plane; the second parameter is a body posture and gravity center space displacement parameter, and is used for representing the posture state and gravity center balance characteristic of the whole body, and specifically comprises time sequence data of the trunk forward inclination angle, the horizontal displacement and vertical fluctuation of a gravity center projection point relative to a starting line; based on the extracted first parameter and second parameter, constructing a time sequence convolution network model, inputting different time sequence segments into a network for self-supervision learning, capturing the space-time dependency relationship among parameters of each stage in the starting action according to the self-supervision learning, optimizing through staged reconstruction errors, enabling the network to automatically learn typical motion characteristics of each stage, and generating motion hidden vectors of each stage in the starting action; Constructing an individuation model by using the trained time sequence convolution network, storing the statistical mean value of the motion hidden vectors of each stage according to the stage by the individuation model as the stage standard characteristic of the current athlete, and setting a first threshold value of the fitness score and a second threshold value of the gesture similarity index according to the statistical distribution of the historical optimal starting sample of the athlete; The preset characteristics of the individuation model comprise standard flexion and extension angle ranges and angular speed ranges of hip joints, knee joints and ankle joints in each stage; The preset gesture of the individuation model comprises a trunk front inclination angle standard interval, a gravity center projection point standard displacement interval and a joint gesture configuration reference of each stage; The individuation model also comprises a parameter updating interface, and after the new effective sample is confirmed in the manual rechecking link, the individuation model corrects and updates the preset characteristics and the preset gesture through a staged increment learning mechanism.
2. The sprint start training method based on motion decomposition as defined in claim 1, wherein the steps of obtaining multi-source training data, extracting athlete body bounding boxes, and generating an individualized motion sequence using an improved key point tracking algorithm comprise: the multisource training data comprise starting video data of athletes, triaxial acceleration data, pressure distribution data acquired by plantar pressure sensors and muscle activation data acquired by electromyographic signal acquisition equipment; Extracting a boundary frame of a sportsman body from multi-source training data through a target detection algorithm, screening out non-starting action frames by taking the change rate of the outline of the human body and the space pose as constraint conditions, reserving the starting action frames, extracting the human boundary frame in the starting action frames, and keeping the starting action frames aligned with the acquired time; Extracting key points including head, shoulder, elbow, hip, knee and ankle in the human body boundary frame according to the extracted human body boundary frame and an improved key point tracking algorithm; determining stress time sequences of the supporting feet and the force-exerting feet in the starting process according to the pressure distribution data, and keeping the key point track consistent with the actual force-exerting direction according to the angle change direction of the stress time sequences; Identifying the stress sequence of each main muscle group according to the muscle activation data, verifying whether the angle change of the key points accords with the physiological stress rule, and carrying out smooth correction on the motion trail of the key points when the key point change which does not accord with the physiological stress rule is detected; and generating an individuation motion sequence according to the key points after the dynamic calibration, the stress time sequence analysis and the motion track smooth correction.
3. The sprint training method based on motion decomposition as claimed in claim 1, wherein the extraction of the feature of the starting motion, input to the execution stage decision of the individualization model, comprises: Extracting starting motion characteristics based on multi-source training data, wherein the starting motion characteristics comprise plantar pressure distribution data, gravity center projection point displacement data and lower limb joint angle data of the current athlete, and synchronously aligning the starting motion characteristics with uniform time stamps; Inputting the aligned starting movement characteristics as input data into an execution stage of the individuation model for judgment, wherein the stages comprise three stages, namely a preparation stage, a stress stage and a transition stage; The judgment standard of the preparation stage is that when the fact that the two feet of the athlete bear supporting force at the same time is detected, the pressure ratio of the front foot to the rear foot is kept in a preset balance range, and the gravity center projection point is positioned behind the starting line, the current action is judged to be in the preparation stage; when the supporting pressure of the forefoot is detected to rise and exceed a preset stress threshold, and the gravity center projection point is changed from a static state to continuously move forwards, judging that the current action is in the stress stage; And the judgment standard of the transition stage is that when the rear foot supporting force is detected to be reduced to nearly zero and not less than 3 sampling frames are continuously detected, and meanwhile, the trunk forward inclination angle is kept in a gesture stable interval preset by the individuation model, the current action is judged to be in the transition stage.
4. The sprint training method based on motion decomposition as claimed in claim 1, wherein the step of performing fitness judgment on the starting motion feature according to the stage judgment result, and calculating a fitness score of the starting motion feature and a preset feature in the individuation model, specifically comprises: If the starting movement characteristic is in the preliminary stage, selecting a preset characteristic corresponding to the preliminary stage in the personalized model; if the starting movement characteristics are in the stress stage, selecting preset characteristics corresponding to the stress stage in the personalized model; if the starting movement characteristics are in the transition stage, selecting preset characteristics corresponding to the transition stage in the personalized model; The center of gravity projection point displacement, plantar pressure distribution and lower limb joint angle parameters in starting movement characteristics form a characteristic vector, unified fitness calculation is carried out on the characteristic vector and a preset characteristic set in a corresponding stage, and a fitness score is obtained through the following formula: ; In the formula, Represents a fitness score that is indicative of the fitness, Representing the first of the features of the starting motion The current value of the individual characteristic parameter is, Representing the preset feature set of the corresponding stage in the personalized model The reference value of the individual characteristic parameter is, Represent the first The weight coefficient of each characteristic parameter is used for reflecting the importance degree of the parameter in the overall action coordination; representing the total number of characteristic parameters.
5. The method for training the starting of the sprint based on the motion decomposition according to claim 1, wherein when the fitness score is in a preset fuzzy interval, the starting motion is marked as a sample to be checked, and the sample is manually confirmed and fed back to the individuation model to update parameters, comprising the following steps: Setting a preset fuzzy interval according to the fitness grading distribution of the historical training samples in the individuation model, wherein the preset fuzzy interval is used for distinguishing a special section of the motion characteristic between normal fluctuation and abnormal deviation and is determined by an upper floating range and a lower floating range of a first threshold value, wherein the upper floating limit of the fuzzy interval is a grading interval of which the first threshold value is increased by 0.05 times, and the lower floating limit is a grading interval of which the first threshold value is reduced by 0.05 times; When the calculated fitness score is in a preset fuzzy interval, marking the feature of the running motion to be a sample to be rechecked, and generating a unique rechecked sample number, wherein the rechecked sample number is stored in an individuation model together with a corresponding running stage label, a timestamp, plantar pressure distribution data, gravity center projection point displacement data and lower limb joint angle data; in the manual rechecking stage, the gravity center track curve, the pressure change curve and the joint angle curve of the sample to be rechecked are visually displayed, training and guiding personnel compare the data result of visual display with the preset standard gesture of the individuation model, the training and guiding personnel confirm that the sample to be rechecked belongs to a normal fluctuation sample or an abnormal sample, and the confirmation result is recorded in a rechecking label; When the review tag is marked as a normal fluctuation sample, the marked sample features are brought into a parameter updating interface of the individuation model, preset features and preset postures are adjusted in stages in an incremental learning mode, when the review tag is marked as an abnormal sample, the marked sample features are used for error reverse correction calculation, and a first threshold and a second threshold of the individuation model are optimized according to correction calculation results.
6. The sprint training method based on motion decomposition as claimed in claim 1, wherein performing gesture similarity determination, calculating similarity index of the starting motion feature and preset gesture in the individualization model, specifically comprises: respectively selecting preset postures matched with corresponding stages in the personalized model according to posture data of different stages in the starting motion characteristics; The preset gesture comprises a trunk forward inclination angle, a hip joint angle, a knee joint angle, an ankle joint angle and a gravity center projection point displacement parameter, the gesture parameter of the starting motion characteristic at the same stage is subjected to space mapping calculation with the preset gesture, and a similarity index is obtained through cosine similarity of a gesture vector included angle, wherein a calculation formula is as follows: ; In the formula, The index of similarity of the poses is represented, Representing the corresponding stage preset gesture set in the individuation model Normalized vector components of the individual pose parameters; representing the first of the features of the starting motion Normalized vector components of the individual pose parameters; representing the total number of gesture parameters.
7. A sprint starting training method based on motion decomposition as defined in claim 1, wherein locating the source joint causing the abnormality of the starting motion and outputting corrective advice comprises: The method comprises the steps of determining the similarity of the gestures, starting a source joint positioning process when the similarity index is lower than a second threshold value, respectively carrying out difference analysis on projection points of a hip joint, a knee joint, an ankle joint, a trunk and a gravity center according to gesture data in starting motion characteristics, and calculating gesture deviation amounts of all joints by comparing the deviation degree of the current starting motion characteristics and preset gestures in an individuation model; When a plurality of joints are simultaneously judged as abnormal joint candidates, determining a source joint according to an upstream priority principle of a moving chain, and preferentially selecting an upstream joint with a larger influence range as the source joint causing abnormal actions; After the source joint is determined, combining pressure distribution, angle change rate and muscle activation time sequence of the corresponding joint in the multi-source training data to further verify whether the abnormality of the source joint has persistence or not; And outputting correction advice according to the positioning result of the source joint, wherein the correction advice comprises a joint angle adjustment range, a force-generating time sequence correction prompt, a gravity center displacement adjustment target and action coordination feedback information and is used for guiding an athlete to conduct targeted posture correction training.

Description

Sprint starting training method based on action decomposition Technical Field The application relates to the technical field of motion recognition, in particular to a sprint starting training method based on motion decomposition. Background The existing human behavior action recognition method is generally based on key point extraction and time sequence analysis in a video sequence, and is used for integrally recognizing and classifying human actions. For example, a common method extracts key point information of a human body by acquiring continuous video frames and utilizing a human body target detection model and a gesture estimation model, then generates a gesture track image sequence based on a matching result between adjacent frames, and performs behavior recognition by combining a preset human body action gesture index so as to determine a specific action type occurring in the video sequence. However, such methods have significant drawbacks in situations such as sprint start training where high precision time series decomposition and local pose determination are required. Firstly, the traditional human behavior recognition only carries out global recognition on a complete action sequence, and lacks of fine division of different stages of starting actions, and cannot realize stage-level action evaluation and targeted guidance, so that the judgment on fine stress difference or action continuity is not accurate enough. Secondly, in the existing scheme, a fixed threshold is mostly adopted for performing action matching judgment in the identification process, random fluctuation caused by sensor noise or individual difference of athletes is not considered, and boundary sample misjudgment and unstable evaluation are easy to occur. Therefore, a sprint start training method based on action decomposition is proposed in an effort to solve the above-described problems. Disclosure of Invention In order to achieve the above object, the present application provides a sprint start training method based on motion decomposition, comprising: And acquiring multi-source training data, extracting a human body boundary box of the athlete, and generating an individualized motion sequence by adopting an improved key point tracking algorithm. And extracting a first parameter and a second parameter of the starting action according to the individuation motion sequence, and utilizing a time sequence convolution network to perform self-supervision learning to construct an individuation model. Extracting starting motion characteristics, inputting the starting motion characteristics into an execution stage judgment of an individuation model, calculating a fitness score of the starting motion characteristics and preset motion characteristics in the individuation model according to a stage judgment result, carrying out fitness judgment according to the fitness score, entering gesture similarity judgment when the fitness score is more than or equal to a first threshold value, judging that deviation exists in starting motion and outputting a stage adjustment suggestion when the fitness score is less than the first threshold value, marking the starting motion as a sample to be rechecked when the fitness score is in a preset fuzzy interval, and feeding back the sample to the individuation model after manual confirmation to execute parameter updating. The method comprises the steps of executing gesture similarity judgment, calculating similarity indexes of starting motion features and preset gestures in an individuation model, judging that starting motions are normal when the similarity indexes are larger than or equal to a second threshold value, positioning source joints causing abnormal starting motions when the similarity indexes are smaller than the second threshold value, and outputting correction suggestions. Preferably, the multi-source training data is obtained, the athlete human body boundary box is extracted, and an improved key point tracking algorithm is adopted to generate an individualized motion sequence, which specifically comprises the following steps: the multisource training data comprise starting video data of athletes, triaxial acceleration data, pressure distribution data acquired by plantar pressure sensors and muscle activation data acquired by electromyographic signal acquisition equipment. And extracting the boundary frame of the athlete body from the multi-source training data through a target detection algorithm, screening out non-starting action frames by taking the change rate of the outline of the body and the space pose as constraint conditions, reserving the starting action frames, extracting the body boundary frame in the starting action frames, and keeping the starting action frames aligned with the acquired time. Extracting key points including head, shoulder, elbow, hip, knee and ankle according to improved key point tracking algorithm in the human body boundary frame, correcting the movement amplitude