CN-121999529-A - Sports training evaluation method and system based on action recognition

CN121999529ACN 121999529 ACN121999529 ACN 121999529ACN-121999529-A

Abstract

The invention relates to the technical field of physical training evaluation, in particular to a physical training evaluation method and a physical training evaluation system based on motion recognition, which are used for constructing a shoulder-knee connection path and extracting pixel lengths to form a time sequence table according to key moment and image frames of reversing positioning motion of knee-shoulder angular velocity, and monitoring the non-smooth mutation, comparing the standard trend to generate an action offset list, identifying a rhythm fluctuation time period based on key frame interval change, analyzing offset and fluctuation information in a crossing manner, and obtaining a training action identification evaluation result according to the distribution characteristics of the overlapping region parts. According to the invention, through angular velocity transformation and identification action conversion, shoulder and knee track extraction extension length is constructed, deviation is judged based on mutation and rhythm trend is combined, central characteristics of overlapping region parts are cross-analyzed, synchronous deviation and rhythm interference are accurately extracted, action continuity and stability discrimination are improved, key state identification resolution is enhanced, and fine tracking and effective early warning of complex action flow are realized.

Inventors

LIU WEI
LIU XIXIAO
LING HONG
HU JING
DING HAITAO
XIE MINGJUN

Assignees

安徽水利水电职业技术学院

Dates

Publication Date: 20260508
Application Date: 20260115

Claims (9)

1. A method for action recognition-based physical training assessment, comprising the steps of: S1, acquiring angular velocity time series data of knee and shoulder motion tracking equipment, grouping continuous time points according to the angular velocity change direction, identifying a direction reversal section, extracting a motion conversion starting time point, and matching corresponding image frame numbers to obtain a motion reversing time set; S2, positioning shoulder joint and knee joint pixel points according to the image frames corresponding to the image frame numbers in the motion reversing moment set, determining shoulder joint and knee joint connection paths, constructing a cross-frame continuous image displacement track, extracting pixel extension lengths of the connection paths of all frames, and forming a time sequence motion line segment length table; s3, comparing extension changes of the shoulder joint and knee joint connection paths in continuous frames by using connection line segment change information of an image frame sequence in the time sequence action line segment length table, identifying non-smooth mutation, extracting overall extension trend, and comparing with a preset standard action to obtain an action offset mark list; And S4, extracting image frame time points of key moments in the action reversing moment set, representing training rhythms by using adjacent image frame time intervals, comparing the change trend of the initial training rhythms with the change trend of the subsequent rhythms, and identifying image frame time periods with prolonged, shortened or frequently alternated intervals to form a rhythmic fluctuation time period list.
2. The method for evaluating physical training based on motion recognition according to claim 1, wherein the motion reversing time set comprises reversing starting time, image frame number and direction reversing identification, the time sequence motion line segment length table comprises a connection path length value, an image frame time index and a shoulder and knee position pixel mapping relation, the motion deviation mark list comprises an abnormal image frame index, a deviation part number and an extension trend abnormality type, and the rhythm fluctuation time period list comprises a time interval change type, a fluctuation duration and a rhythm stability label.
3. The method for evaluating physical training based on action recognition according to claim 1, wherein the step of S1 is: S101, acquiring time sequence angular velocity data recorded by knee and shoulder action tracking equipment of an athlete during training, extracting an angular velocity direction value and a corresponding position label of each time point according to time sequence, comparing the angular velocity directions between adjacent time points, dividing continuous time periods with the same direction value into the same classification number, numbering according to time sequence, and obtaining an angular velocity direction continuous numbering sequence; S102, screening whether positive and negative conversion occurs between adjacent numbering directions according to the change condition of direction values between adjacent numbering in the continuous numbering sequence of the angular velocity direction, extracting time sections between a front numbering end point and a rear numbering start point for the position where conversion exists, and orderly arranging all the time sections with direction conversion according to the numbering sequence to generate a direction conversion time section sequence; And S103, extracting corresponding image frame numbers in the original data according to the starting time points of each time period in the direction conversion time period sequence, integrating the corresponding frame numbers of all starting points in time sequence, constructing an image frame key time point set, and generating an action reversing time set by taking the image frame key time point set as a positioning basis for action conversion.
4. The method for motion recognition based physical training assessment according to claim 1, wherein the step of S2 is: S201, acquiring corresponding image frame image information according to image frame numbers in the action reversing moment set, sequentially extracting coordinate positions of shoulder joints and knee joints in an image in a two-dimensional pixel region, calling an image position difference value between each pair of joint coordinate points and an adjacent state of the image boundary to screen, eliminating joint coordinate pairs with incomplete marks, and then establishing corresponding relations between two joint pixel points in all effective frames to generate a joint pixel coordinate sequence; S202, extracting relative position direction values among pixel points in each frame according to coordinate points between shoulder joint and knee joint positions of each frame in the joint pixel coordinate sequence, constructing a frame-by-frame pixel connection path according to connection positions among the pixel points according to the arrangement sequence of the pixel points in a two-dimensional image area, calling sequence numbers of the pixel connection paths in all image frames as indexes, and generating a pixel path number list; s203, extracting image frame time positions as transverse indexes according to the number of pixel points contained in each frame path in the pixel path number list, correspondingly matching each frame path length information with the image frame time points, completing the record of the extension number of the cross-frame paths in the image sequence, and generating a time sequence action line segment length table.
5. The method for motion recognition based physical training assessment according to claim 1, wherein the step of S3 is: S301, calling extension length values of shoulder to knee paths in adjacent image frames based on connection path information of each frame image in the action line extension length sequence, extracting length difference value change conditions between adjacent frames according to frame number sequences, and performing section division on areas with continuously increasing, continuously decreasing or suddenly changing change values to generate connection path sudden change frame intervals; S302, screening the extending trend directions of the first frame and the last frame of each section according to the image frame numbers in the connecting path abrupt change frame sections, matching the starting and stopping relation of path direction change and trend fluctuation amplitude values according to the path fluctuation starting positions indicated by the frame numbers, and extracting the frame numbers which do not meet the direction consistency standard to obtain a path trend departure frame number group; S303, according to the corresponding time points of the image frames in the path trend deviation frame number group, combining the shoulder and knee pixel coordinate positions in the time point image frames, extracting offset position labels in the connecting path, identifying the positions of deviation of extension trend in each image frame, establishing a corresponding matching table of the time points and the positions, and generating an action offset mark list.
6. The method for motion recognition based physical training assessment according to claim 1, wherein the step of S4 is: S401, extracting image frame time points corresponding to each key time in the action reversing time set, sequentially arranging the time points, sequentially calling time difference values between two adjacent time points, taking each time difference value as a rhythm representation value, and establishing a rhythm change reference sequence in a continuous time period after eliminating repeated time points and abnormal frame numbers to generate an image frame rhythm interval sequence; s402, extracting the first five groups of rhythm intervals according to the numbering sequence of each rhythm value in the image frame rhythm interval sequence, calculating an average time interval as a reference rhythm in the initial stage of training, calling the difference relation between the reference rhythm and each subsequent group of rhythm values, marking the trend of the difference according to the change direction of the time interval, establishing a trend comparison identification sequence related to the reference rhythm, and generating a rhythm difference trend group; S403, detecting whether the state of rhythm interval lengthening, shortening or alternate fluctuation occurs in the continuous time period according to trend identifiers between adjacent numbers in the rhythm difference trend group, dividing the time period of the numbered paragraphs with trend change characteristics, extracting an image frame time point interval corresponding to each change period, and generating a rhythm fluctuation time period list.
7. The method for action recognition based physical training assessment according to claim 1, wherein said method further comprises: S5, cross analysis is carried out on the abnormal time points and the connection path parts in the action deviation mark list and the start and stop time in the rhythm fluctuation time period list, the concentrated distribution and continuous characteristics of the abnormal parts in the superposition section are analyzed, key monitoring content in the training process is determined, and a sports training action recognition evaluation result is obtained; the physical training action recognition evaluation result comprises an abnormal time and rhythm fluctuation coincidence interval, a concentrated offset part distribution sequence and a training performance abnormal type.
8. The method for athletic training assessment based on action recognition according to claim 7, wherein the step of S5 is: S501, acquiring an abnormal occurrence time point and a connection path position mark recorded in the action deviation mark list, acquiring start and stop time information of each time period in a rhythm fluctuation time period list, comparing the abnormal time point with the start and stop range of the time period item by item according to a time axis sequence, judging whether the abnormal time point falls into a corresponding time period range, and marking a time point combination meeting a time coverage relation to generate an abnormal rhythm superposition time interval; S502, based on the associated connection path position identifiers in the abnormal rhythm superposition time interval, summarizing position identifier sequences appearing in the same interval according to time sequence, judging whether position identifiers of adjacent time points keep continuous consistency or repeatedly appearing, grouping and recording the position identifiers meeting continuous appearance conditions, establishing interval and position corresponding relation, and generating an abnormal position concentrated distribution sequence; S503, screening interval records with time coincidence features and position concentrated distribution features according to the position distribution state corresponding to each time interval in the abnormal position concentrated distribution sequence, combining the duration length of the interval and the position appearance sequence, summarizing related information according to time sequence, establishing a training performance evaluation data set, and generating a sports training action recognition evaluation result.
9. A motion recognition based physical training assessment system for use in a motion recognition based physical training assessment method as claimed in any one of claims 1 to 8, the system comprising: The motion reversing identification module is used for acquiring angular velocity time series data of the knee and shoulder motion tracking equipment, grouping continuous time points according to the angular velocity change direction, identifying a direction reversing section, extracting a motion conversion starting time point, and matching corresponding image frame numbers to obtain a motion reversing time set; the joint path construction module is used for positioning the pixel points of the shoulder joint and the knee joint according to the image frames corresponding to the image frame numbers in the motion reversing moment set, determining the connection paths of the shoulder joint and the knee joint, constructing a frame-crossing continuous image displacement track, extracting the pixel extension length of each frame of connection path, and forming a time sequence motion line segment length table; The motion deviation judging module is used for comparing extension changes of the shoulder joint and knee joint connection paths in continuous frames by using the connection line segment change information of the image frame sequence in the time sequence motion line segment length table, identifying non-smooth mutation, extracting overall extension trend, and comparing with preset standard motion alignment to obtain a motion deviation mark list; The training rhythm analysis module extracts image frame time points of key moments in the action reversing moment set, characterizes training rhythms by adjacent image frame time intervals, compares the change trend of the rhythms at the initial stage and the follow-up stage of training, and identifies image frame time periods with prolonged, shortened or frequently alternated intervals to form a rhythmic fluctuation time period list; And the comprehensive evaluation monitoring module is used for carrying out cross analysis on the abnormal time points and the connection path parts in the action deviation mark list and the start and stop time in the rhythm fluctuation time period list, analyzing the concentrated distribution and continuous characteristics of the abnormal parts in the superposition section, determining key monitoring content in the training process and obtaining a sports training action recognition evaluation result.

Description

Sports training evaluation method and system based on action recognition Technical Field The invention relates to the technical field of physical training evaluation, in particular to a physical training evaluation method and system based on action recognition. Background The technical field of physical training evaluation comprises a systematic method for carrying out multidimensional analysis and quantitative judgment on key elements such as action execution conditions, body function performances, training progress and the like of a motion participant in a training process, the core content of the technical field mainly surrounds how to accurately acquire motion behavior data of the training participant through scientific means, and comprehensively evaluates training quality and action standard by means of image processing, biomechanical modeling, data analysis and the like, in the process, action capture is used as one of key data acquisition means, dynamic recording of spatial positions, motion tracks and time sequences of all parts of a human body is usually realized through optical sensing, inertial measurement, visual recognition and the like, and in the whole, the technical field of physical training evaluation aims at constructing a training feedback mechanism based on demonstration data, so that the establishment of a personalized, structured and continuously optimized motion training system is promoted. The physical training evaluation method based on motion recognition is characterized in that motion capture is used as a basis, key body node motion data of a trainer in a designated training scene is collected, motion characteristics are extracted and matched through a specific recognition flow, different motion types and execution modes are distinguished, technical matters such as gesture recognition, motion compliance analysis, rhythm consistency judgment and the like in the training process are mainly developed, motion parameters of each joint of the body in space are calculated through a continuous motion sequence of the trainer, feature vector comparison is carried out with a preset standard motion template, the training motion types are recognized, evaluation basis is output, the state change of a key frame after the training item is combined in the recognition process, and the fluency and stability of the motion transition process are analyzed, so that a comprehensive evaluation system of the training motion is constructed. In the prior art, global gestures or whole motion forms are used as main lines for unfolding analysis in the training evaluation process, an extension trend judging mechanism based on joint path mutation points is lacked, instantaneous changes of motion fluctuation are difficult to capture from a microscopic part level, an abnormal recognition channel with multiple coupling between time and space is difficult to build under the condition of unstable training frequency or frequent local motion deviation, the lack of sensitivity and local pertinence of an evaluation result are easy to cause, and particularly in a scene with severe rhythm change or concentrated position deviation, the problem that details of training state change cannot be accurately revealed exists, so that the capture efficiency of key training abnormal states is influenced. Disclosure of Invention In order to solve the technical problems in the prior art, the embodiment of the invention provides a physical training evaluation method and a physical training evaluation system based on action recognition. The technical scheme is as follows: A method of action recognition based physical training assessment comprising the steps of: S1, acquiring angular velocity time series data of knee and shoulder motion tracking equipment, grouping continuous time points according to the angular velocity change direction, identifying a direction reversal section, extracting a motion conversion starting time point, and matching corresponding image frame numbers to obtain a motion reversing time set; S2, positioning shoulder joint and knee joint pixel points according to the image frames corresponding to the image frame numbers in the motion reversing moment set, determining shoulder joint and knee joint connection paths, constructing a cross-frame continuous image displacement track, extracting pixel extension lengths of the connection paths of all frames, and forming a time sequence motion line segment length table; s3, comparing extension changes of the shoulder joint and knee joint connection paths in continuous frames by using connection line segment change information of an image frame sequence in the time sequence action line segment length table, identifying non-smooth mutation, extracting overall extension trend, and comparing with a preset standard action to obtain an action offset mark list; And S4, extracting image frame time points of key moments in the action reversing moment set, representing training rhythms