Real-time posture correction feedback system based on computer vision

Abstract

The invention discloses a real-time posture correction feedback system based on computer vision, and relates to the technical field of computer vision and health monitoring, the system comprises the following components: a computer vision acquisition module, a core index extraction module, a risk grading evaluation module, a prospective intervention scheme generation module and a multi-mode feedback module; a user posture image is captured in real time through the computer vision acquisition module, core posture indexes such as the spine curvature, the pelvic inclination angle and the cervical vertebra anteflexion angle are dynamically calculated in combination with the core index extraction module, and more importantly, the system uses the risk grading evaluation module, based on a time sequence risk evaluation algorithm and a user portrait, and the user posture index is evaluated. According to the method and the system, personalized threshold adjustment is realized, posture abnormity is accurately divided into three levels, namely a mild level, a moderate level and a severe level, and the personalized and precise evaluation ensures that each user can obtain a posture correction scheme which is most suitable for the user, so that the correction effect is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and health monitoring, in particular to a real-time body posture correction feedback system based on computer vision. BACKGROUND

[0002] With the rapid development of computer vision technology and health monitoring technology, how to use these advanced technologies to improve people's body posture health has become an important research topic. Body posture problems not only affect personal appearance, but also may cause a series of skeletal muscle system diseases, such as cervical spondylosis, lumbar spondylosis, etc.

[0003] Traditional body posture correction technology mainly has the following shortcomings: first, the evaluation method is highly subjective, depending on the experience and observation of the evaluator, and there may be large differences between different evaluators, leading to inconsistency of the evaluation results; second, traditional methods are difficult to realize real-time monitoring and dynamic evaluation, and cannot timely discover and correct the user's body posture problems, which is easy to let the user form the wrong body posture habits; third, traditional technology lacks personalized correction schemes, and cannot accurately adjust according to the user's age, gender, occupation and other characteristics, resulting in limited correction effect; finally, traditional body posture correction methods often lack effective feedback mechanisms, and users cannot intuitively understand their body posture status and correction progress, affecting the enthusiasm and effect of correction.

[0004] In view of the many shortcomings of traditional body posture correction technology, the present application proposes a real-time body posture correction feedback system based on computer vision, which is particularly important. SUMMARY

[0005] The purpose of the present application is to make up for the shortcomings of the prior art, and to provide a real-time body posture correction feedback system based on computer vision, which can realize real-time monitoring, accurate evaluation, personalized correction scheme generation and multi-modal feedback of user body posture by integrating computer vision acquisition module, core index extraction module, risk grading evaluation module, forward-looking intervention scheme generation module and multi-modal feedback module. This system not only overcomes the shortcomings of traditional methods such as strong subjectivity and poor real-time performance, but also ensures that each user can obtain the most suitable body posture correction scheme through personalized and accurate evaluation, greatly improving the correction effect. At the same time, the introduction of multi-modal feedback and forward-looking intervention mechanism further improves the user experience and the safety of the system, providing strong technical support for modern body posture health management.

[0006] To solve the above technical problems, the present application provides the following technical scheme: a real-time body posture correction feedback system based on computer vision, which comprises the following components: computer vision acquisition module, core index extraction module, risk grading evaluation module, forward-looking intervention scheme generation module and multi-modal feedback module. The computer vision acquisition module supports multi-scene adaptive acquisition parameter adjustment and image preprocessing, and captures user body image data in real time. The core index extraction module calculates core body posture indexes such as spinal curvature, pelvic tilt angle, and cervical spine flexion angle using a dynamic feature fusion algorithm, and includes a noise removal and dynamic update mechanism. The risk grading evaluation module, based on a time series risk evaluation algorithm, combines user portraits to achieve personalized threshold adjustment, and divides body posture abnormalities into three levels of mild, moderate, and severe. The prospective intervention scheme generation module generates a personalized relaxation training scheme for mild risk, initiates real-time correction guidance for moderate risk, and generates a professional intervention scheme for severe risk in conjunction with a medical platform. The multi-modal feedback module supports adaptive switching of visual, audio, and tactile multi-modal feedback, accurately feeds back intervention schemes to users, and realizes real-time body posture correction and risk warning.

[0007] Further, the specific steps of multi-scene adaptive acquisition in the computer vision acquisition module are as follows: First, real-time acquisition of environmental light intensity and scene texture feature data is performed through an image sensor, and a scene classification model trained based on a convolutional neural network is used to automatically identify the current scene type, including office, home, outdoor, and sports scenes. Second, acquisition parameters are dynamically adjusted for different scenes. In the office scene, the acquisition frame rate is adjusted to 30 fps and the exposure time is set to 1 / 60 s to reduce screen reflection interference. In the outdoor scene, automatic white balance adjustment and backlight compensation functions are started. In the sports scene, high-speed acquisition mode is switched to and a motion blur suppression algorithm is turned on. Third, the original images collected are preprocessed. Environmental noise is removed through Gaussian filtering, and the contrast of the skeletal key point region is enhanced using histogram equalization. Finally, the preprocessed image data is transmitted to the core index extraction module in real time through edge computing technology, and image frames within 10 seconds are cached to avoid data loss caused by sudden network fluctuations or device lag, ensuring the stability and clarity of body image acquisition in different scenes.

[0008] Further, the calculation of spinal curvature in the core index extraction module uses a dynamic feature fusion algorithm, and the formula is wherein is the maximum perpendicular distance of the spinal curve and the fitting straight line, is the change amount of the maximum distance of the spinal curve within 5 seconds, is the straight line length of the spinal key point connection line, is the time weight coefficient, The parameter source and determination method are as follows: by collecting 50,000+ dynamic image data of different ages, body shapes and users, covering 12 scenes such as static sitting posture and dynamic walking, using gradient descent algorithm to iteratively train the sample data, and finally calibrating The weight distribution can effectively consider the influence of static bending degree and dynamic change trend on health risk; And The key point coordinates of the spine are fitted by a Bezier curve, the fitting error is controlled within ±0.2mm, and the L is calculated by the Euclidean distance between the cervical C7 point and the lumbar L5 point, to ensure the accuracy and dynamic adaptability of the spine curvature index, and to provide core data support for risk classification.

[0009] Further, the specific steps of noise elimination and dynamic updating in the core index extraction module are as follows: first, the 3σ criterion is used to detect outliers of the core index data of single sampling, and the index data exceeding the mean value ±3 times the standard deviation is determined as noise data, and the linear interpolation method is used to supplement the missing data; secondly, the effective sampling data is summarized every 5 seconds, the average value, maximum value and change slope of the index within 5 seconds are calculated, and the index dynamic change curve is constructed; then, the dynamic change curve is smoothed based on the sliding window algorithm, the high frequency fluctuation interference is eliminated, and the trend characteristics of the body shape change are retained; finally, the processed index data is pushed to the risk classification evaluation module in real time, and the historical data in the user local database is updated, the historical data is compressed and archived once every 24 hours, and the statistical characteristics of the daily core index are extracted, to provide high-quality data support for risk assessment model optimization and user body shape trend analysis.

[0010] Further, the risk classification evaluation module adopts a time sequence risk assessment algorithm, and the formula is: Wherein R is the comprehensive risk value, S is the spine curvature, is the pelvic tilt angle, is the cervical lordosis angle, t is the duration of body shape abnormality, is the index weight coefficient, is the time sequence attenuation factor coefficient, is the attenuation rate constant, and the parameters and weight determination method: based on the body shape monitoring data of 30,000+ users and 10,000+ orthopedic clinical diagnosis cases, the feature importance analysis is carried out by using random forest algorithm, combined with medical expert score, and finally ; the association data between abnormal duration and health risk is fitted by survival analysis model, and is calibrated, the time sequence factor can quantify the cumulative risk of long-term abnormal body shape; the risk grade classification standard is: is the mild risk, is the moderate risk, For severe risk, the model test set accuracy reaches 96.3%, realizing the scientific quantification and accurate division of risk levels.

[0011] Further, the system realizes the individual adjustment of risk threshold through the following specific technical steps: first, the computer vision acquisition module guides the user to complete the static posture shooting from three angles of front, side and back, extracts biological characteristic data such as user height, weight, skeletal ratio and joint activity degree through video frame analysis, and obtains the user's age, gender, occupation, and past medical history information; second, the K-means clustering algorithm is used to classify user characteristics, and six core groups of teenagers, adults, the elderly, office workers, sports groups and special health status groups are divided; then, based on the 20,000+ sample data of each group, the threshold adjustment model is trained, and for the teenager group, the mild risk threshold of spinal curvature is lowered to S<4.0%, for the elderly group, the moderate risk threshold of pelvic tilt angle is raised to , and for the office worker group, the mild risk duration threshold of cervical flexion angle is shortened to <20 minutes; finally, the group threshold is optimized again combined with the individual characteristics of the user. Further, the forward-looking intervention scheme generation module generates an individualized relaxation training scheme for mild risk through the following technical steps: first, determine the abnormal part based on the core index data; second, filter the training content according to the age, occupation and use scenario information in the user profile, and prioritize the spine development protection actions for teenagers, select the fragmented office scenario adaptive actions for office workers, and push the complete training sequence for home scenarios; then, match the training actions with corresponding difficulty for the user through the action difficulty grading algorithm, with 60% of the initial difficulty, 30% of the intermediate difficulty, and 10% of the advanced difficulty, to avoid muscle damage caused by excessive training intensity; finally, generate a structured training plan, including action name, duration, breathing rhythm guidance, and set training reminder time according to user daily use habits, and the training scheme supports offline caching and progress synchronization, automatically records training effect after user completes training, and dynamically adjusts the difficulty and frequency of subsequent training content.

[0012] Further, the prospective intervention program generation module corrects and guides in real time for moderate risk by the following technical steps: first, call up the preset standard body model library, adaptively adjust the standard model parameters according to the user's biological characteristics, and generate a user-specific standard body model; second, compare the current body shape with the standard model in real time through computer vision, calculate the body shape similarity using the IoU algorithm, and accurately locate the abnormal joints and parts; third, start the multi-dimensional guidance coordination mechanism, superimpose three-dimensional correction guide lines and dynamic target boxes on the real-time picture in the visual layer, and mark the abnormal parts with gradually changing red color, and the brightness is enhanced when the similarity is less than 80%; based on natural language processing technology, generate personalized instructions in the voice layer, send short instructions every 3 seconds in the early stage, and send detailed guidance every 10 seconds when the similarity is improved to more than 85%; finally, calculate the correction compliance rate in real time, switch to intermittent guidance when the compliance rate is greater than or equal to 90% for 3 minutes in a row, and start intensive guidance when the compliance rate is less than 70%, to ensure the balance between correction effect and user comfort.

[0013] Further, the prospective intervention program generation module performs medical platform linkage for severe risk by the following technical steps: first, automatically integrate user core body shape index data, abnormal continuous trajectory, risk assessment report and user portrait information, generate structured medical archives according to the DICOM standard format of the medical industry, and ensure that the data format meets the compatibility requirements of the hospital information system; second, establish secure communication with the orthopedic platform of the cooperative Internet hospital through an encrypted API interface, use the HTTPS protocol to transmit medical archive data, and perform AES encryption processing on sensitive information during transmission; third, the platform matches orthopedic doctors based on the risk characteristics in the medical archives, the screening conditions include the professional direction of the doctor, the on-call state, and the geographical location distance of the user, generates a list of 3 candidate doctors and pushes it to the user; at the same time, the user end APP provides one-key reservation function, supports online video consultation or offline outpatient registration, sends appointment reminders and precautions after successful reservation; finally, generate a temporary safe body shape guidance program during the user's waiting period, limit high-risk actions through real-time monitoring, push posture adjustment reminders every 15 minutes, and send emergency warnings if the risk value continues to rise, and notify the bound guardians or emergency contacts in synchronization, to ensure that severe risk users receive timely and professional medical intervention.

[0014] Further, the multi-modal feedback module realizes adaptive switching through the following specific technical steps: first, the computer vision acquisition module analyzes the current state of the user in real time, judges the user's concentration through facial feature recognition, and determines the current scene attribute in combination with environmental sound intensity and scene type; second, based on the scene attribute and the user's historical feedback preference, the feedback mode priority is set, the priority order in the concentration scene is: tactile feedback > visual weak prompt > voice feedback, the priority order in the relaxation scene is: voice feedback > visual highlight prompt > tactile feedback, and the priority order in the motion scene is: tactile feedback > voice concise prompt > visual guidance; then, the feedback parameters are dynamically adjusted, the vibration intensity of tactile feedback is adjusted according to the risk level, the mild risk is set to level 2, the moderate risk is set to level 3-4, and the severe risk is set to level 5; the volume of voice feedback is automatically adjusted according to the environmental noise intensity to ensure that the user can hear clearly; finally, the response effect of the user to the feedback is recorded in real time, if there is no adjustment action for 3 times of tactile feedback, the voice + visual joint feedback is automatically switched, the feedback mode combination is continuously optimized through the reinforcement learning algorithm, and the user's compliance and response efficiency to the correction guidance are improved.

[0015] Compared with the prior art, the real-time body posture correction feedback system based on computer vision has the following beneficial effects: First, the system captures user body posture images in real time through the computer vision acquisition module, and dynamically calculates core posture indicators such as spinal curvature, pelvic tilt angle and cervical lordosis angle in combination with the core indicator extraction module, more importantly, the system uses the risk grading evaluation module to realize individualized threshold adjustment based on the time sequence risk evaluation algorithm and the user portrait, accurately divides the body posture abnormalities into three levels of mild, moderate and severe, and ensures that each user can obtain the most suitable body posture correction scheme, thereby improving the correction effect.

[0016] Second, the system is equipped with a multi-modal feedback module, supports adaptive switching of visual, voice and tactile feedback modes, ensures that the intervention scheme can be fed back to the user in the optimal way in different scenes, and at the same time, the forward-looking intervention scheme generation module provides individualized relaxation training scheme, real-time correction guidance and medical platform linkage and other professional intervention measures for different risk levels. This all-round and multi-level feedback and intervention mechanism not only improves the user experience, but also ensures that the user can obtain professional medical help in time when facing severe body posture risks, thereby greatly enhancing the safety and practicality of the system.

[0017] Other advantages, objects, and features of the present application will be apparent to those skilled in the art from the following specification, in some respects, will be apparent to those skilled in the art from the following specification, and in some respects, will be apparent to those skilled in the art from the following specification, or can be learned from the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0019] Figure 1 A running flowchart of a real-time body posture correction feedback system based on computer vision; Figure 2 A risk threshold personalized adjustment flowchart of a real-time body posture correction feedback system based on computer vision; Figure 3 A multi-modal feedback adaptive switching flowchart of a real-time body posture correction feedback system based on computer vision. DETAILED DESCRIPTION

[0020] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined application purpose, the following will combine the drawings and preferred embodiments to specifically describe the specific embodiments, structures, features and effects of the present application. EMBODIMENT

[0021] After the system is started, the computer vision acquisition module first acquires the light intensity and scene texture feature data of the office environment in real time through the image sensor, automatically identifies the current office scene based on the scene classification model trained by the convolutional neural network, and then dynamically adjusts the acquisition parameters, sets the acquisition frame rate to 30fps and the exposure time to 1 / 60s, so as to reduce the interference caused by the reflection of the computer screen. After that, the original user body posture image collected is preprocessed, the environmental noise is removed through Gaussian filtering, the contrast of the skeleton key point region is enhanced by using histogram equalization, and finally the preprocessed image data is transmitted to the core index extraction module in real time through edge computing technology.

[0022] After receiving the image data, the core index extraction module calculates the core body posture indexes of the user such as the spine curvature, the pelvic tilt angle and the cervical spine flexion angle by using a dynamic feature fusion algorithm, and the formula is as follows: , wherein is the maximum vertical distance of the spine curve and the fitting straight line, is the change amount of the maximum distance of the spine curve within 5 seconds, is the straight line length of the spine key point connecting line, is the time weight coefficient, For the rate of change weight coefficient, in the data processing process, the 3σ criterion is used to detect outliers of the core index data of single sampling, the index data exceeding ± 3 times the standard deviation of the mean is judged as noise data, and the missing data is supplemented by linear interpolation method; every 5 seconds, the effective sampling data is summarized, the average value, the maximum value and the change slope of the index in the period are calculated, the dynamic change curve of the index is constructed, and the curve is smoothed based on the sliding window algorithm, then the processed index data is pushed to the risk grading evaluation module in real time, and the historical data in the user local database is updated, the historical data will be compressed and archived every 24 hours in the future, and the statistical characteristics of the daily core index are extracted, such as Figure 1

[0023] After the risk grading evaluation module receives the core index data, the time sequence risk assessment algorithm is used to calculate the comprehensive risk value, the formula is: wherein R is the comprehensive risk value, S is the curvature of the spine, is the pelvic tilt angle, is the cervical flexion angle, t is the duration of abnormal posture, is the index weight coefficient, is the time sequence attenuation factor coefficient, is the attenuation rate constant, combined with user portrait for personalized threshold adjustment, as shown in Figure 2 Before that, the system has guided the user to complete the static posture shooting of three angles of front, side and back, extracted the biological characteristic data such as height, weight, skeletal proportion and joint activity, and obtained the information of age, gender, occupation and past medical history filled by the user, classified the user into office crowd through K-means clustering algorithm, and based on the threshold adjustment model trained by 20,000+ sample data of the group, the threshold value of the duration of mild risk of cervical flexion angle is shortened to <20 minutes, and the group threshold value is optimized again combined with the individual characteristics of the user, finally the current abnormal posture of the user is determined as mild risk.

[0024] ​The proactive intervention module initiates a personalized relaxation training program generation process for mild risks: First, based on core indicator data, the abnormal areas of the user are identified as the cervical and lumbar spine; second, considering the occupational characteristics and work scenarios of office workers, fragmented and suitable movements are selected to avoid occupying too much work time; then, a movement difficulty grading algorithm is used to match training movements of corresponding difficulty to the user, with 60% being beginner level, 30% intermediate level, and 10% advanced level; finally, a structured training plan is generated, including movement names, durations, breathing rhythm guidance, and training reminders set at 10:00 AM, 3:00 PM, and one hour before the end of the workday, based on the user's daily commuting and rest habits. The training plan supports offline caching and progress synchronization. After the user completes the training, the system automatically records the training effect to dynamically adjust the difficulty and frequency of subsequent training content.

[0025] The multimodal feedback module first analyzes the user's current state through a computer vision acquisition module. It determines the user is focused on work by using facial feature recognition. Based on the characteristics of the office environment, it prioritizes feedback methods as follows: tactile feedback > subtle visual cues > voice feedback. Figure 3 As shown, based on a mild risk level, the vibration intensity of the tactile feedback is set to level 2, while the volume of the voice feedback is automatically adjusted according to the sound intensity of the office environment. The system records the user's response to the feedback in real time. If the user does not make any posture adjustment after three consecutive tactile feedbacks, it automatically switches to a combination of voice and visual feedback. The system continuously optimizes the combination of feedback methods through reinforcement learning algorithms to ensure that the intervention plan is accurately delivered to the user. Example

[0026] After the system starts, the computer vision acquisition module first guides the user to complete static body posture photography from three angles: front, side, and back. It extracts biometric data such as the user's height, weight, bone ratio, and joint mobility, while also acquiring information such as age, gender, and past health history filled in by the user's guardian. Subsequently, the image sensor collects light intensity and scene texture feature data of the home environment. The scene classification model trained based on a convolutional neural network automatically identifies the current scene as a home environment. The acquired raw images are preprocessed by removing environmental noise through Gaussian filtering and enhancing the contrast of key skeletal areas through histogram equalization. Finally, the preprocessed image data is transmitted in real time to the core indicator extraction module through edge computing technology.

[0027] The core indicator extraction module uses a dynamic feature fusion algorithm to calculate core postural indicators such as spinal curvature, pelvic tilt angle, and cervical flexion angle. In data processing, outliers are detected using the 3σ criterion, and missing data is supplemented using linear interpolation. Valid sampled data is summarized every 5 seconds, and the average, maximum, and slope of the indicators are calculated. Dynamic curves are constructed and smoothed using a sliding window algorithm. The processed indicator data is then pushed to the risk grading assessment module, while simultaneously updating historical data in the user's local database. Data is compressed and archived every 24 hours, and daily core indicator statistical features are extracted, such as... Figure 1 As shown.

[0028] The risk grading and assessment module uses a time-series risk assessment algorithm to calculate a comprehensive risk value, and then adjusts the thresholds individually based on user profiles. Figure 2 As shown, the user was classified into the adolescent group by K-means clustering algorithm. Based on the threshold adjustment model trained with 20,000+ sample data of this group, the threshold for mild risk of spinal curvature was lowered to S<4.0%. Then, the group threshold was optimized by combining individual user characteristics, and finally the user's current abnormal posture was determined to be of moderate risk.

[0029] The prospective intervention program generation module initiates a real-time correction guidance process for moderate risk: First, it retrieves a preset standard posture model library, adaptively adjusts the standard model parameters based on the user's biometrics, and generates a user-specific standard posture template; second, it uses computer vision to compare the user's current posture with the standard template in real time, calculates posture similarity using the IoU algorithm, and accurately locates abnormal joints and areas; then, it activates a multi-dimensional guidance collaboration mechanism. Visually, it overlays three-dimensional correction guidance lines and dynamic target boxes on the real-time screen, with abnormal areas marked in gradient red. When the similarity is below 80%, the marking brightness increases. In terms of voice, it generates personalized instructions based on natural language processing technology, initially sending a short instruction every 3 seconds, and then sending detailed guidance every 10 seconds when the similarity increases to above 85%; finally, it calculates the correction achievement rate in real time. If the achievement rate is ≥90% for 3 consecutive minutes, it switches to intermittent guidance; if the achievement rate is below 70%, it activates reinforced guidance.

[0030] The multimodal feedback module analyzes the user's current state through a computer vision acquisition module, determines it to be a relaxation scenario based on the attributes of the home environment, and sets the priority of feedback methods as follows: voice feedback > visual highlighting > tactile feedback, such as... Figure 3 As shown, based on a moderate risk level, the vibration intensity of the haptic feedback is set to level 3-4, while the volume of the voice feedback is automatically adjusted according to the noise level of the home environment. The system records the user's response to the feedback in real time. If the user does not adjust their posture after three consecutive haptic feedbacks, it automatically switches to a combination of voice and visual feedback. The system continuously optimizes the combination of feedback methods through reinforcement learning algorithms to ensure the effectiveness of real-time corrective guidance.

[0031] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A real-time posture correction feedback system based on computer vision, characterized in that, The system comprises the following components: a computer vision acquisition module, a core indicator extraction module, a risk grading and assessment module, a prospective intervention plan generation module, and a multimodal feedback module; The computer vision acquisition module supports multi-scene adaptive acquisition parameter adjustment and image preprocessing, and captures user body image data in real time. The core indicator extraction module uses a dynamic feature fusion algorithm to calculate core postural indicators such as spinal curvature, pelvic tilt angle, and cervical flexion angle, and includes noise removal and dynamic update mechanisms. The risk grading and assessment module: based on the time-series risk assessment algorithm and combined with user profiles, it realizes personalized threshold adjustment and classifies abnormal body posture into three levels: mild, moderate and severe. The prospective intervention program generation module generates personalized relaxation training programs for mild risks, initiates real-time corrective guidance for moderate risks, and links with the medical platform to generate professional intervention programs for severe risks. The multimodal feedback module supports adaptive switching between visual, voice, and tactile multimodal feedback, accurately conveying the intervention plan to the user.

2. The real-time posture correction feedback system based on computer vision according to claim 1, characterized in that, The specific steps of the multi-scene adaptive acquisition in the computer vision acquisition module are as follows: First, ambient light intensity and scene texture feature data are acquired in real time through an image sensor. The scene classification model trained based on a convolutional neural network automatically identifies the current scene type, including office scenes, home scenes, outdoor scenes, and sports scenes. Second, the acquisition parameters are dynamically adjusted for different scenes. In office scenes, the acquisition frame rate is adjusted to 30fps and the exposure time is set to 1 / 60s to reduce screen reflection interference. In outdoor scenes, automatic white balance adjustment and backlight compensation functions are activated. In sports scenes, the high-speed acquisition mode is switched and the motion blur suppression algorithm is enabled. Then, the acquired raw images are preprocessed. Gaussian filtering is used to remove environmental noise, and histogram equalization is used to enhance the contrast of the skeletal key point area. Finally, the preprocessed image data is transmitted to the core indicator extraction module in real time through edge computing technology.

3. The real-time posture correction feedback system based on computer vision according to claim 1, characterized in that, The calculation of spinal curvature in the core indicator extraction module adopts a dynamic feature fusion algorithm, and the formula is as follows: ,in This represents the maximum perpendicular distance between the spinal curve and the fitted straight line. This represents the maximum change in the spinal curve within 5 seconds. The length of the straight line connecting the key points of the spine. For time weighting coefficients, This is the rate of change weighting coefficient.

4. The real-time posture correction feedback system based on computer vision according to claim 1, characterized in that, The specific steps for noise removal and dynamic updating in the core indicator extraction module are as follows: First, outlier detection is performed on the core indicator data of a single sample using the 3σ criterion. Indicator data exceeding the mean ± 3 times the standard deviation are identified as noise data, and missing data are supplemented using linear interpolation. Second, valid sampled data are summarized every 5 seconds, and the average, maximum, and slope of the indicators within 5 seconds are calculated to construct a dynamic change curve for the indicators. Then, the dynamic change curve is smoothed based on the sliding window algorithm. Finally, the processed indicator data is pushed to the risk classification assessment module in real time, while the historical data in the user's local database is updated. The historical data is compressed and archived every 24 hours, and the statistical characteristics of the daily core indicators are extracted to provide data support for risk assessment model optimization and user body posture trend analysis.

5. The real-time posture correction feedback system based on computer vision according to claim 1, characterized in that, The risk grading and assessment module uses a time-series risk assessment algorithm, the formula of which is: Where R is the overall risk value and S is the spinal curvature. The pelvic tilt angle, t represents the cervical flexion angle, and t represents the duration of the postural abnormality. The index weighting coefficient, For time-series decay factor coefficients, is the decay rate constant.

6. The real-time posture correction feedback system based on computer vision according to claim 1, characterized in that, The system achieves personalized adjustment of risk thresholds through the following specific technical steps: First, the computer vision acquisition module guides users to complete static body posture photography from three angles: front, side, and back. Combined with video frame analysis, it extracts biometric data such as user height, weight, bone structure, and joint mobility, while also acquiring user-provided information on age, gender, occupation, and past health history. Second, the K-means clustering algorithm is used to classify user characteristics, dividing them into six core groups: adolescents, adults, the elderly, office workers, athletes, and people with special health conditions. Then, based on over 20,000 sample data points for each group, a threshold adjustment model is trained. For the adolescent group, the risk threshold for mild spinal curvature is lowered to S < 4.0%, and for the elderly group, the risk threshold for moderate pelvic tilt angle is raised to... For office workers, the threshold for the duration of mild risk of cervical flexion angle was shortened to <20 minutes; finally, the threshold for the group was further optimized based on individual user characteristics.

7. A real-time posture correction feedback system based on computer vision according to claim 1, characterized in that, The prospective intervention program generation module generates personalized relaxation training programs for mild risks through the following technical steps: First, it identifies abnormal areas based on core indicator data; second, it selects suitable training content by combining age, occupation, and usage scenario information from the user profile, prioritizing spinal development protection exercises for adolescents, selecting exercises suitable for fragmented office scenarios for office workers, and pushing complete training sequences for home scenarios; third, it matches training exercises of corresponding difficulty levels to users through an exercise difficulty grading algorithm, with 60% being beginner level, 30% intermediate level, and 10% advanced level; finally, it generates a structured training plan, including exercise names, durations, breathing rhythm guidance, and sets training reminder times based on the user's daily usage habits. The training plan supports offline caching and progress synchronization, automatically recording training effects after the user completes the training, and dynamically adjusting the difficulty and frequency of subsequent training content.

8. A real-time posture correction feedback system based on computer vision according to claim 1, characterized in that, The prospective intervention program generation module provides real-time correction and guidance for moderate risk through the following technical steps: First, it retrieves a preset standard posture model library, adaptively adjusts the standard model parameters based on the user's biometrics, and generates a user-specific standard posture template. Second, it uses computer vision to compare the current posture with the standard template in real time, calculates posture similarity using the IoU algorithm, and accurately locates abnormal joints and areas. Then, it activates a multi-dimensional guidance and collaboration mechanism. Visually, it overlays three-dimensional correction guidance lines and dynamic target boxes on the real-time screen, with abnormal areas marked in gradient red, and the marking brightness increases when the similarity is below 80%. In terms of voice, it generates personalized instructions based on natural language processing technology, initially sending a short instruction every 3 seconds, and then sending a detailed instruction every 10 seconds after the similarity increases to above 85%. Finally, it calculates the correction achievement rate in real time. When the achievement rate is ≥90% for 3 consecutive minutes, it switches to intermittent guidance; when the achievement rate is below 70%, it activates enhanced guidance.

9. A real-time posture correction feedback system based on computer vision according to claim 1, characterized in that, The prospective intervention plan generation module, targeting severe risks, integrates with the medical platform through the following technical steps: First, it automatically integrates the user's core postural indicators, abnormal trajectory, risk assessment report, and user profile information to generate a structured medical record according to the medical industry's DICOM standard format. Second, it establishes secure communication with the partner internet hospital's orthopedic platform via an encrypted API interface, transmitting medical record data using the HTTPS protocol, and encrypting sensitive information with AES during transmission. Then, the platform matches orthopedic doctors based on the risk characteristics in the medical record, using criteria including doctor's specialty, outpatient status, and user's geographical distance, generating a list of three candidate doctors and pushing it to the user. Simultaneously, the user's app provides a one-click appointment function, supporting online video consultations or offline outpatient registration, sending appointment reminders and precautions after successful appointment. Finally, while the user waits for their appointment, a temporary safe posture guidance plan is generated, restricting high-risk movements through real-time monitoring, sending posture adjustment reminders every 15 minutes, and sending an emergency warning if the risk value continues to rise, simultaneously notifying the linked guardian or emergency contact.

10. A real-time posture correction feedback system based on computer vision according to claim 1, characterized in that, The multimodal feedback module achieves adaptive switching through the following specific technical steps: First, the computer vision acquisition module analyzes the user's current state in real time, judging the user's focus through facial feature recognition and determining the current scene attributes by combining ambient sound intensity and scene type; Second, based on scene attributes and the user's historical feedback preferences, the priority of feedback methods is set. In focused scenes, the priority order is: tactile feedback > weak visual cues > voice feedback; in relaxed scenes, the priority order is: voice feedback > visual highlight cues > tactile feedback; in motion scenes, the priority order is: tactile feedback > concise voice cues > visual guidance; Then, the feedback parameters are dynamically adjusted. The vibration intensity of tactile feedback is adjusted according to the risk level: level 2 for mild risk, level 3-4 for moderate risk, and level 5 for severe risk; the volume of voice feedback is automatically adjusted according to the ambient noise intensity; Finally, the user's response to the feedback is recorded in real time. If there is no adjustment action after three consecutive tactile feedback responses, the system automatically switches to a combined voice and visual feedback, continuously optimizing the combination of feedback methods through reinforcement learning algorithms.

Images