CN-121982778-A - Physical training posture real-time correction and feedback system based on machine vision

CN121982778ACN 121982778 ACN121982778 ACN 121982778ACN-121982778-A

Abstract

The invention relates to the technical field of vision analysis, in particular to a machine vision-based real-time correction and feedback system for physical training gestures, which comprises a data acquisition module, a gesture deviation calculation module, an evaluation decision module, a multi-mode feedback generation module and a real-time correction guide module, wherein the real-time video of a sportsman in the target training process is acquired, the gesture of the sportsman is estimated to obtain joint coordinates of the sportsman, the joint coordinates are mapped to a historical personalized gesture reference library of the sportsman to obtain a dimensional gesture deviation quantification result, the dimensional gesture deviation quantification result is input to a preset dynamic evaluation rule engine to obtain a structural evaluation result, the structural evaluation result is subjected to strategic analysis to obtain a composite feedback instruction, the real-time gesture is guided and regulated based on the composite feedback instruction to correct the physical training gesture, and the personalized deviation judgment and the real-time correction feedback efficiency of the physical training gesture can be improved.

Inventors

HE YANG
WANG JUNWEI
XIE JUN
Du Diejun
DU YANLI

Assignees

绍兴理工学院

Dates

Publication Date: 20260505
Application Date: 20260130

Claims (10)

1. The physical training posture real-time correction and feedback system based on machine vision is characterized by comprising a data acquisition module, a posture deviation calculation module, an evaluation decision module, a multi-mode feedback generation module and a real-time correction guide module, wherein: The data acquisition module is used for acquiring real-time videos of athletes in the target training process, and estimating human body postures of the real-time videos to obtain joint coordinates of the athletes; The attitude deviation calculation module is used for mapping the joint coordinates to a historical personalized attitude reference library of the athlete to obtain a dimension attitude deviation quantification result of the athlete; The evaluation decision module is used for inputting the dimension attitude deviation quantification result to a preset dynamic evaluation rule engine so as to obtain a structural evaluation result of the athlete; The multi-mode feedback generation module is used for strategically analyzing the structural evaluation result to obtain a composite feedback instruction of the athlete; and the real-time correction guiding module is used for guiding and controlling the real-time posture of the athlete based on the composite feedback instruction so as to correct the physical training posture of the athlete.
2. The machine vision-based real-time correction and feedback system for physical training poses of a player according to claim 1, wherein the data acquisition module is configured to, when performing acquisition of real-time video of a target training process, perform human body pose estimation on the real-time video to obtain joint coordinates of the player: collecting multi-angle real-time videos of athletes; Denoising and enhancing the multi-angle real-time video to obtain an optimized video of the athlete, and carrying out dynamic foreground separation on the optimized video to obtain a purified foreground image of the athlete; the method comprises the steps of performing human body region image segmentation on a purified foreground image to obtain a complete posture outline of an athlete; and carrying out gesture feature interpretation on the complete gesture outline to obtain the joint position of the athlete, and carrying out coordinate mapping on the joint position to obtain the joint coordinates of the athlete.
3. The machine vision based physical training posture real-time correction and feedback system of claim 1, wherein said posture deviation calculation module, when performing mapping of said joint coordinates to said athlete's historical personalized posture reference library, is specifically configured to: establishing a historical personalized attitude benchmark library of the athlete according to the historical training video of the athlete; screening personalized gesture reference actions of athletes in a historical personalized gesture reference library based on joint coordinates; Performing space offset quantization on the joint coordinates and the joint coordinates of the personalized gesture reference action to obtain a coordinate difference value of the athlete; and carrying out parameter synthesis on the coordinate difference values to obtain a dimension attitude deviation quantization result of the athlete.
4. A machine vision based physical training posture real time correction and feedback system according to claim 3, wherein said posture deviation calculation module, when executing the establishment of said athlete's historical personalized posture reference library from said athlete's historical training video, is specifically adapted to: screening key action time periods in the historical training video to obtain an initial video segment of the athlete; detecting an action optimal frame of the initial video segment to obtain a standard training action gesture frame of the athlete; Carrying out time sequence recombination on the standard training action gesture frame to obtain a standard training action instance video clip of the athlete; carrying out three-dimensional attitude estimation on the video clips of the standard training action examples to obtain the standard joint coordinates of the athlete; and (3) calibrating the gesture reference to the standard joint coordinates to obtain a historical personalized gesture reference library of the athlete.
5. The machine vision based physical training posture real-time correction and feedback system of claim 3, wherein said posture deviation calculation module is specifically configured to, when performing parametric synthesis on said coordinate difference values to obtain a dimension posture deviation quantization result of said athlete: Performing joint angle conversion on the coordinate difference value to obtain a multi-joint angle deviation set of the athlete; multidimensional vector synthesis is carried out on the multi-joint angle deviation set to obtain the integral deviation vector of the limb segment of the athlete; performing time domain filtering of a moving average window on the whole deviation vector of the limb segment to obtain a smooth limb segment deviation vector of the athlete; and carrying out dimensionless standardized mapping on the smooth limb segment deviation vector based on a preset limb movement range extremum to obtain a dimensionality attitude deviation quantification result of the athlete.
6. The machine vision based physical training posture real-time correction and feedback system of claim 1, wherein the assessment decision module is specifically configured to, when executing the input of the posture deviation quantization result to a preset dynamic assessment rules engine to obtain the structural assessment result of the athlete: removing redundant data in the dimension attitude deviation quantization result to obtain adaptability deviation data of the athlete; The adaptive deviation data are subjected to individual decoupling to obtain independent deviation data of athletes; performing threshold matching on the independent deviation data based on a preset deviation threshold gradient standard to obtain a deviation severity level of the athlete; according to the deviation severity level, carrying out hierarchical dynamic matching on independent deviation data to obtain a level association deviation list of the athlete; Performing identification coding on the grade association deviation list to obtain a grade identification evaluation sub-result of the athlete; and integrating the grade identification evaluation sub-result into a structural evaluation result of the athlete.
7. The machine vision based physical training posture real time correction and feedback system of claim 6, wherein said evaluation decision module, when executing a hierarchical dynamic matching of said independent deviation data according to said deviation severity level, is specifically configured to: Performing level mapping on the independent deviation data to obtain a level association characteristic matrix of the independent deviation data; performing similarity matching on the grade association characteristic matrix and the characteristic set of the deviation severity grade to generate an initial pairing set of a grade association deviation list; Confidence screening is carried out on the initial pairing set, and an effective pairing set of the level association deviation list is obtained; And carrying out structured packaging on the effective pairing set to obtain a rank association deviation list of the athlete.
8. The machine vision-based physical training posture real-time correction and feedback system of claim 1, wherein the multi-modal feedback generation module performs strategic analysis on the structured assessment result to obtain a composite feedback instruction of the athlete, specifically for: carrying out dimension analysis on the structural evaluation result to obtain deviation information of athletes; Performing structural mapping on the deviation information to obtain corrected text description of the deviation information; performing three-dimensional posture correction simulation on the deviation information to obtain dynamic correction animation of the deviation information; and carrying out collaborative coding on the correction text description and the dynamic correction animation to obtain a composite feedback instruction of the athlete.
9. The machine vision-based physical training posture real-time correction and feedback system of claim 8, wherein the multi-modal feedback generation module is specifically configured to, when performing a three-dimensional posture correction simulation on the deviation information to obtain a dynamic correction animation of the deviation information: According to the joint deviation information, carrying out reverse reconstruction on the real-time posture of the athlete to obtain an error posture of the athlete; Calculating the posture deviation of the athlete based on the standard actions and the wrong postures in the historical personalized posture reference library; And based on the gesture deviation, performing deviation driving animation synthesis on the motion trail of the athlete to obtain the dynamic correcting animation of the athlete.
10. The machine vision based physical training posture real-time correction and feedback system of claim 1, wherein said real-time correction guidance module, when executing said composite feedback instruction based on said real-time correction guidance module, is specifically configured to: Carrying out multidimensional information fusion on the correction text description and the dynamic correction animation to obtain multi-mode feedback content of the athlete; dynamically guiding interface synthesis is carried out on the multi-mode feedback content to obtain a correction guiding interface of the athlete; Based on the correction guiding interface, the real-time training gesture of the athlete is subjected to visual guiding adjustment so as to correct the physical training gesture of the athlete.

Description

Physical training posture real-time correction and feedback system based on machine vision Technical Field The invention relates to the technical field of vision analysis, in particular to a physical training posture real-time correction and feedback system based on machine vision. Background In the field of physical training posture correction, the prior art generally lacks a personalized posture reference system based on individual characteristics of athletes, relies on generalized posture standard to develop deviation judgment, is difficult to adapt to physical structures, exercise habits and training level differences of different athletes, and results in disjoint of quantized results of posture deviation and actual training requirements, and is low in accuracy and pertinence. Meanwhile, in the attitude data acquisition link, the technical means of frame synchronous processing, denoising enhancement and multi-view image fusion of the multi-angle real-time video are not mature enough, the incomplete segmentation of human body areas and low extraction precision of joint coordinates are easily caused, so that larger errors exist in basic data of subsequent deviation calculation, and the integral effect of attitude correction is directly influenced. Most of the posture evaluation rules in the prior art are fixedly set, and lack of the ability of dynamically adapting different training action types and real-time training states of athletes, and cannot accurately classify and structurally disassemble the posture deviation, so that the evaluation result is difficult to provide clear correction direction. In addition, the feedback form is single, the multi-mode collaborative feedback of text description and dynamic visualization cannot be realized, and the real-time performance of feedback transmission is insufficient, so that a sportsman cannot quickly obtain clear and visual correction guidance, and is difficult to adjust the wrong gesture in time, the training efficiency is reduced, and the exercise injury is possibly caused by accumulation of the long-term wrong gesture, so that the real-time correction feedback efficiency of the physical training gesture is improved, and the problem to be solved urgently is solved. Disclosure of Invention In order to achieve the above purpose, the machine vision-based real-time physical training posture correction and feedback system provided by the invention is characterized by comprising a data acquisition module, a posture deviation calculation module, an evaluation decision module, a multi-mode feedback generation module and a real-time correction guide module, wherein: The data acquisition module is used for acquiring real-time videos of athletes in the target training process, and estimating human body postures of the real-time videos to obtain joint coordinates of the athletes; The attitude deviation calculation module is used for mapping the joint coordinates to a historical personalized attitude reference library of the athlete to obtain a dimension attitude deviation quantification result of the athlete; The evaluation decision module is used for inputting the dimension attitude deviation quantification result to a preset dynamic evaluation rule engine so as to obtain a structural evaluation result of the athlete; the multi-mode feedback generation module is used for strategically analyzing the structured evaluation result to obtain a composite feedback instruction of the athlete; And the real-time correction guiding module is used for guiding and controlling the real-time posture of the athlete based on the composite feedback instruction so as to correct the physical training posture of the athlete. In a preferred embodiment, the data acquisition module is specifically configured to, when performing acquisition of real-time video of an athlete during target training and performing human body posture estimation on the real-time video to obtain joint coordinates of the athlete: collecting multi-angle real-time videos of athletes; denoising and enhancing the multi-angle real-time video to obtain an optimized video of the athlete, and carrying out dynamic foreground separation on the optimized video to obtain a purified foreground image of the athlete; performing human body region image segmentation on the purified foreground image to obtain a complete posture outline of the athlete; and carrying out gesture feature interpretation on the complete gesture outline to obtain the joint position of the athlete, and carrying out coordinate mapping on the joint position to obtain the joint coordinate of the athlete. In a preferred embodiment, the attitude deviation calculation module is specifically configured to, when performing mapping the joint coordinates to a historical personalized attitude reference library of the athlete to obtain a dimension attitude deviation quantification result of the athlete: establishing a historical personalized attitude benchmark library o