Method for recognizing sports posture based on big data and image recognition

Abstract

The application discloses a motion posture recognition method based on big data and image recognition, and comprises the following steps: acquiring a motion video stream of a target user in real time, and extracting a user posture image sequence in the motion video stream; performing key point detection on the user posture image sequence, generating a user real-time posture skeleton graph, and calculating a real-time multi-dimensional posture feature vector based on the skeleton graph; extracting a benchmark posture feature vector set of a current sports item based on a preset motion posture feature big data cloud library; matching the real-time multi-dimensional posture feature vector with a benchmark posture feature vector in the benchmark posture feature vector set to obtain a matching degree; and identifying the category and quality level of the current motion posture according to the matching degree. The application compares and analyzes the current motion posture by collecting motion images in real time and combining the motion posture feature big data cloud library, thereby improving the accuracy of motion posture recognition and providing a basis for subsequent motion posture adjustment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of motion posture recognition, and more particularly to a motion posture recognition method based on big data and image recognition. BACKGROUND

[0002] With the rapid development of computer vision and artificial intelligence technology, motion posture analysis based on image recognition has become a research hotspot in the fields of sports science, rehabilitation training and public fitness. The existing technology usually captures human motion images through a single camera or a depth sensor (such as Kinect), extracts two-dimensional or three-dimensional human key points by using open source libraries such as OpenPose and AlphaPose, and then calculates simple parameters such as joint angles and motion trajectories, and compares them with pre-set limited standard action models to realize action recognition or normative evaluation.

[0003] However, the existing technology has significant limitations when applied to professional, complex and even extreme sports scenarios; first, data silos and insufficient model generalization ability: most systems rely on standardized action data sets collected in small-scale laboratory environments, and their action libraries are sparse and insufficient in covering individual differences, resulting in a sharp decline in recognition accuracy when faced with the diverse personal styles, body types and non-standard actions in the real world, and failing to provide evaluations with broad reference value; second, poor environmental adaptability: most existing algorithms are designed for indoor environments with good lighting and simple backgrounds, and in outdoor complex environments (such as snow, water, mountains), strong light, reflection, complex background, dynamic occlusion (such as water splashes, snow mist) and special clothing (such as heavy ski suits) will seriously interfere with image quality and key point detection accuracy, causing the system to fail; third, single analysis dimension and delayed feedback: technical solutions focus on static posture or simple dynamic recognition, lack of deep integration analysis of motion timing logic, phase importance weight and multi-dimensional motion biomechanical features (such as acceleration, momentum change), and feedback is often in the form of post-video playback and data reports, which cannot provide real-time, intuitive and operable guidance during the movement, especially in high-speed and high-risk extreme sports, where delayed feedback has little training value; fourth, lack of personalized evolution and safety warning mechanism: the system usually uses a "one-size-fits-all" universal standard, cannot track the long-term progress curve of individual users, and is difficult to predict and warn potential high-risk error postures based on historical data, lacking functionality in preventing sports injuries. SUMMARY

[0004] To solve at least one of the above technical problems, the purpose of the present application is to provide a motion posture recognition method based on big data and image recognition, which can improve the accuracy of motion posture recognition.

[0005] The present application provides a motion posture recognition method based on big data and image recognition, comprising:

[0006] acquiring a motion video stream of a target user in real time, and extracting a user posture image sequence in the motion video stream;

[0007] detecting key points of the user posture image sequence, generating a user real-time posture skeleton graph, and calculating a real-time multi-dimensional posture feature vector based on the skeleton graph;

[0008] extracting a set of reference posture feature vectors of a current motion item based on a preset motion posture feature big data cloud library;

[0009] matching the real-time multi-dimensional posture feature vector with reference posture feature vectors in the set of reference posture feature vectors to obtain a matching degree;

[0010] identifying a category and a quality level of the current motion posture according to the matching degree.

[0011] In the scheme, the step of calculating the real-time multi-dimensional posture feature vector based on the skeleton graph specifically comprises:

[0012] calculating a skeleton vector between the joint nodes based on the skeleton graph;

[0013] calculating a set of angles between the key bones based on the skeleton vector , wherein m is the number of angles of interest;

[0014] calculating a motion speed and acceleration feature set of the joint nodes based on changes of the skeleton graph between consecutive frames of the user posture image sequence , wherein n is the number of joint nodes;

[0015] normalizing and splicing the set of angles A and the motion feature set M to form the real-time multi-dimensional posture feature vector .

[0016] In the scheme, the formula for matching the real-time multi-dimensional posture feature vector with reference posture feature vectors in the set of reference posture feature vectors to obtain a matching degree is specifically: , wherein S represents the matching degree; K represents the total number of divided motion stages; i belongs to k; represents a weight coefficient of the i-th stage; represents the real-time multi-dimensional posture feature vector in the i-th stage; represents a standard posture feature vector in the set of reference posture feature vectors that is closest to the Euclidean distance.

[0017] The scheme further comprises:

[0018] When the matching degree is higher than a preset matching degree threshold, a real-time multi-dimensional posture feature vector corresponding to the current matching degree is extracted;

[0019] The real-time multi-dimensional posture feature vector corresponding to the matching degree is set as a high-quality user posture feature vector, and is supplemented to the reference posture feature vector set.

[0020] In the scheme, the following are further included:

[0021] The matching degree of the current frame is set as The matching degree of the previous frame is set as The matching degree difference between adjacent frames is ;

[0022] If the matching degree difference between adjacent frames is greater than a preset continuity factor, an alert information is triggered;

[0023] Based on the alert information, if the matching degree of the current frame is less than a preset second matching degree threshold, a preset safety capsule is triggered;

[0024] Based on the alert information, if the matching degree of the current frame is greater than or equal to the preset second matching degree threshold, the posture feature vector of the current frame is optimized according to the posture feature vectors in a preset sliding time window to recalculate the matching degree of the current frame.

[0025] In the scheme, the following are further included:

[0026] When the matching degree is higher than a preset matching degree threshold, a real-time multi-dimensional posture feature vector corresponding to the current matching degree is extracted;

[0027] After traversing all the motion video streams, all the marked positions are obtained;

[0028] Taking any one marked position as a center and setting a distance as a radius, an influence range is constructed, and the number of markings in the influence range is extracted;

[0029] If the number of markings in the influence range is greater than a preset number threshold, a current influence range maintenance processing alert information is generated.

[0030] In the scheme, the following are further included:

[0031] Based on a preset sliding time window, a matching degree set of different time periods is obtained;

[0032] The matching degrees in the matching degree set are averaged to obtain average matching degrees of the corresponding time periods;

[0033] The average matching degree of the current time period is subtracted from the average matching degree of the previous time period to obtain an average matching degree difference;

[0034] If the absolute value of the average matching degree difference value is less than a preset first average matching degree threshold, motion state stability prompt information is generated;

[0035] If the average matching degree difference value is greater than a preset second average matching degree threshold, motion state improvement prompt information is generated;

[0036] If the average matching degree difference value is less than a preset third average matching degree threshold, motion state decline prompt information is generated.

[0037] In the scheme, after the motion state decline prompt information is generated, the scheme further includes:

[0038] The matching degree set of the current time period is extracted, and the minimum matching degree in the corresponding matching degree set is extracted;

[0039] According to the minimum matching degree in the corresponding matching degree set, a real-time multi-dimensional posture feature vector at the current motion time is determined;

[0040] According to the real-time multi-dimensional posture feature vector at the current motion time and the corresponding standard posture feature vector, a difference value is calculated to obtain a motion posture feature vector adjustment value;

[0041] The motion posture feature vector adjustment value is sent to a preset screen end for display.

[0042] In the scheme, after the user posture image sequence is detected for key points to generate a user real-time posture skeleton map, the scheme further includes:

[0043] If the current motion scene is a skiing scene, a user body surface temperature distribution map is obtained in real time based on a preset infrared thermal imaging sensor;

[0044] The temperature point cloud in the user body surface temperature distribution map is analyzed for clustering to obtain a user joint part point cloud map, and a user joint node range is determined;

[0045] If the joint node in the user real-time posture skeleton map falls within the corresponding joint node range, the current user real-time posture skeleton map is normal;

[0046] If the joint node in the user real-time posture skeleton map does not fall within the corresponding joint node range, the position of the joint node is revised based on the corresponding joint node range.

[0047] The one or more technical solutions proposed in the application have at least the following technical effects:

[0048] By constructing and continuously updating the motion posture feature big data cloud library, the benchmark posture feature vector set is continuously enriched and evolved, thereby improving the accuracy and professionalism of the recognition and evaluation structure;

[0049] For the special scene of snow limit skiing, an infrared thermal imaging sensor is added to obtain the user body temperature distribution map, and the user joint range is recognized through the higher temperature of the joint point; in combination with the user real-time posture skeleton map, the error of user real-time posture skeleton recognition caused by thick clothes worn by the user when skiing in the snow is reduced.

[0050] In summary, the application realizes the all-round breakthrough of the motion posture recognition technology in precision, reliability, real-time performance, safety and individualization by constructing the motion posture feature big data cloud library and combining the depth image processing and intelligent analysis algorithm for special scenes, and provides a complete, efficient and intelligent solution for the fields of high-performance sports training, public scientific fitness and rehabilitation evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 The flowchart of the motion posture recognition method based on big data and image recognition of the application is shown. DETAILED DESCRIPTION

[0052] In order to more clearly understand the above-mentioned purposes, features and advantages of the application, the application will be further described in detail below in combination with the drawings and specific embodiments. It should be noted that the embodiments of the application and the features in the embodiments can be combined with each other without conflict.

[0053] In the following description, many specific details are set forth in order to provide a thorough understanding of the application, but the application can also be implemented in other ways different from those described herein, therefore, the protection scope of the application is not limited by the specific embodiments disclosed below.

[0054] Figure 1 The flowchart of the motion posture recognition method based on big data and image recognition of the application is shown.

[0055] As Figure 1 shown, the application discloses a motion posture recognition method based on big data and image recognition, which comprises:

[0056] S101, real-time acquisition of the motion video stream of the target user, and extraction of the user posture image sequence in the motion video stream;

[0057] S102, key point detection on the user posture image sequence, generation of the user real-time posture skeleton map, and calculation of the real-time multi-dimensional posture feature vector based on the skeleton map;

[0058] S103, extraction of the reference posture feature vector set of the current sports project based on the pre-set motion posture feature big data cloud library;

[0059] S104, matching the real-time multi-dimensional posture feature vector and the reference posture feature vector in the reference posture feature vector set to obtain a matching degree;

[0060] S105, identifying the category and quality level of the current motion posture according to the matching degree.

[0061] According to the embodiment of the present application, the original data is obtained through image acquisition, and the digital feature vector representing the essence of the user posture is extracted through key point detection and skeleton graph generation The preset motion posture feature big data cloud library contains a reference posture feature vector set of various sports, such as table tennis; the reference posture feature vector corresponding to the current multi-dimensional posture feature vector is determined according to the matching degree, and the current motion posture category is determined according to the reference posture feature vector; each reference posture feature vector in the reference posture feature vector set corresponds to a motion posture category, such as "snow skiing big turn" in skiing; the quality of the corresponding motion posture category is determined through the matching degree, the matching degree is divided into multiple matching degree ranges, each matching degree range corresponds to a quality level, and the higher the matching degree in the matching degree range, the higher the corresponding quality level.

[0062] According to the embodiment of the present application, the step of calculating the real-time multi-dimensional posture feature vector based on the skeleton graph specifically includes:

[0063] Based on the skeleton graph, the bone vector between the joint nodes is calculated;

[0064] Based on the bone vector, the angle set between the key bones is calculated , wherein m is the number of angles of interest;

[0065] Based on the change of the skeleton graph between the continuous frames of the user posture image sequence, the motion speed and acceleration feature set of the joint nodes is calculated ; wherein n is the number of joint nodes;

[0066] The angle set A and the motion feature set M are normalized and spliced to form the real-time multi-dimensional posture feature vector .

[0067] It should be noted that the relative spatial position relationship of each link of the human body at a certain frame moment is accurately described by the set A of angles between the skeleton vectors and the key bones (for example, knee joint angles, torso inclination angles), which is the core basis for evaluating whether the posture is "in place"; the set M of angular velocities and accelerations of the joints is calculated to capture the rate and rhythm of the posture change, which is crucial for evaluating the fluency of the motion, the explosive force, and identifying specific dynamic errors (such as rapid deceleration); for example, in a sports turn, the angular acceleration of the hip joint can effectively reflect the agility of the turning force; the static A and the dynamic M are normalized (to eliminate the influence of individual body size differences) and spliced to form a unified mathematical expression with high information density .

[0068] According to the embodiment of the present application, the formula for matching the real-time multi-dimensional posture feature vector and the reference posture feature vector in the reference posture feature vector set to obtain the matching degree is as follows: , wherein S represents the matching degree; K represents the total number of divided motion stages; i belongs to k; represents the weight coefficient of the i-th stage; represents the real-time multi-dimensional posture feature vector in the i-th stage; represents the reference posture feature vector in the reference posture feature vector set that is closest to the Euclidean distance.

[0069] It should be noted that the action cycle is divided into K stages for segmented matching, and different weight coefficients are set for different stages, such as higher weight coefficients for key stages; the normalized distance between the feature vector of each stage of the user and the nearest standard vector in the cloud library is calculated, and the total score S is obtained by weighted summation, which directly reflects the overall similarity between the user's action and the ideal action, such as S = 0.85, which can be set as the action completion quality is good, and the D represents the normalized distance function for calculating the difference between the corresponding two feature vectors, and the smaller the D value is, the more similar it is.

[0070] According to the embodiment of the present application, it further comprises:

[0071] When the matching degree is higher than the preset matching degree threshold, the real-time multi-dimensional posture feature vector corresponding to the current matching degree is extracted;

[0072] The real-time multi-dimensional posture feature vector corresponding to the matching degree is set as a high-quality user posture feature vector, and is supplemented to the reference posture feature vector set.

[0073] It should be noted that with the use of a large number of users, the big data cloud library can collect different body types, different styles but equally efficient motion patterns, so that the standard set is more diverse and inclusive, and even better motion paradigms can be discovered and summarized, for example, if the preset matching threshold is 0.85, the real-time multi-dimensional posture feature vector with a matching degree higher than 0.85 is fed back to the corresponding reference posture feature vector set for storage.

[0074] Further, the reference posture feature vectors in the reference posture feature vector set are sorted in chronological order, and when the number of reference posture feature vectors exceeds the preset number threshold, the earliest reference posture feature vector is deleted until the number of reference posture feature vectors in the reference posture feature vector set does not exceed the preset number threshold.

[0075] According to an embodiment of the present application, further comprising:

[0076] The matching degree of the current frame is set to , and the matching degree of the previous frame is set to , and the matching degree difference between adjacent frames is ;

[0077] If the matching degree difference between adjacent frames is greater than the preset continuity factor, an alarm information is triggered;

[0078] Based on the alarm information, if the matching degree of the current frame is less than the preset second matching degree threshold, a preset safety capsule is triggered;

[0079] Based on the alarm information, if the matching degree of the current frame is greater than or equal to the preset second matching degree threshold, the posture feature vector of the current frame is optimized according to the posture feature vectors in the preset sliding time window to recalculate the matching degree of the current frame.

[0080] It should be noted that in normal continuous motion, the posture matching degree between adjacent frames should change smoothly, and when the matching degree difference between adjacent frames is greater than the preset continuity factor, it is determined that the current frame has a large difference caused by external interference; if the matching degree of the current frame is less than the preset second matching degree threshold, it means that the posture feature vector of the current motion and the standard posture feature vector differ greatly, which can easily cause damage to the athlete, so a preset safety capsule is triggered; if the matching degree of the current frame is greater than or equal to the preset second matching degree threshold, it means that the posture feature vector of the current motion and the standard posture feature vector differ less, and the motion posture is more standardized and will not cause damage; further, the historical feature vectors of the current frame and its previous N frames are extracted, and weighted average or median filtering is performed to obtain the posture feature vector after optimization of the current frame; the preset second matching degree threshold is less than the preset matching degree threshold.

[0081] According to the embodiment of the present application, further comprising:

[0082] When the alarm information is triggered, the location of the current video frame is marked;

[0083] After traversing all the motion video streams, all the marked locations are obtained;

[0084] Taking any one of the marked locations as the center and setting the distance as the radius, an influence range is constructed, and the number of markings in the influence range is extracted;

[0085] If the number of markings in the influence range is greater than a preset number threshold, a current influence range maintenance processing alarm information is generated.

[0086] It should be noted that if multiple video streams are marked at the same location and the number of markings is greater than a preset number threshold, such as 3, it indicates that the motion field where the current location is located may have unreasonable conditions, such as too small curvature, unevenness, etc. Therefore, an alarm information is generated to prompt timely processing.

[0087] According to the embodiment of the present application, further comprising:

[0088] Based on a preset sliding time window, a matching degree set of different time periods is obtained;

[0089] The matching degrees in the matching degree set are averaged to obtain an average matching degree corresponding to the time period;

[0090] The average matching degree of the current time period is subtracted from the average matching degree of the previous time period to obtain an average matching degree difference;

[0091] If the absolute value of the average matching degree difference is less than a preset first average matching degree threshold, a motion state stable prompt information is generated;

[0092] If the average matching degree difference is greater than a preset second average matching degree threshold, a motion state improvement prompt information is generated;

[0093] If the average matching degree difference is less than a preset third average matching degree threshold, a motion state decline prompt information is generated.

[0094] It should be noted that the preset third average matching degree threshold is less than zero, and the absolute value of the preset third average matching degree threshold is greater than the preset first average matching degree threshold, the preset first average matching degree threshold is greater than zero and less than the preset second average matching degree threshold, the matching degree set of the athlete in a period of time is intercepted through the preset sliding time window, and the matching degree set in the adjacent time period is compared, the average matching degree of the time period is quantified to evaluate whether the athlete in the current state reaches or approaches the personal peak level, and the state fluctuation is monitored; for example, if the average matching degree difference is greater than the preset second average matching degree threshold, a sports state improvement prompt information such as "your rotation angle stability today has increased by 5% than the best record in the previous stage" is generated; such personalized feedback can greatly improve the stickiness and training motivation of the user, and make the system change from a cold action judge to a personalized intelligent coach who understands and accompanies the user's growth.

[0095] According to the embodiment of the application, after the sports state decline prompt information is generated, the method further comprises:

[0096] extracting the matching degree set of the current time period and extracting the minimum matching degree in the corresponding matching degree set;

[0097] determining the real-time multi-dimensional posture feature vector at the current sports time according to the minimum matching degree in the corresponding matching degree set;

[0098] performing difference calculation according to the real-time multi-dimensional posture feature vector at the current sports time and the corresponding standard posture feature vector to obtain a sports posture feature vector adjustment value;

[0099] sending the sports posture feature vector adjustment value to a preset screen end for display.

[0100] It should be noted that after the sports state decline prompt information is generated, the real-time multi-dimensional posture vector corresponding to the minimum matching degree of the current time period is found, and the current sports posture feature vector is adjusted according to the feature difference between the real-time multi-dimensional posture vector and the corresponding standard posture feature vector, thereby effectively improving the rationality of the sports.

[0101] According to the embodiment of the application, after the user posture image sequence is detected for key points to generate a user real-time posture skeleton graph, the method further comprises:

[0102] if the current sports scene is a skiing scene, a user body surface temperature distribution graph is acquired in real time based on a preset infrared thermal imaging sensor;

[0103] performing clustering analysis on the temperature point cloud in the user body surface temperature distribution graph to obtain a user joint part point cloud graph and determine a user joint range;

[0104] If the joint in the real-time posture skeleton graph of the user falls in the corresponding joint range, the current real-time posture skeleton graph of the user is normal.

[0105] If the joint in the real-time posture skeleton graph of the user does not fall in the corresponding joint range, the position of the joint is revised based on the corresponding joint range.

[0106] It should be noted that if the current sports scene is a skiing scene, it indicates that the athlete wears heavy clothes, and there may be errors in identifying the real-time posture skeleton graph of the user through the video stream. Therefore, by combining the comprehensive analysis of the user's body surface temperature distribution graph, the accuracy of the position of the joint in the real-time posture skeleton graph of the user is improved.

[0107] Further, if the joint in the real-time posture skeleton graph of the user does not fall in the corresponding joint range, when the joint range is after the joint in time, the point closest to the corresponding joint in the joint range is extracted and set as the position after the joint is revised; when the joint range is before the joint in time, the point closest to the corresponding joint in the joint range is extracted and set as the nearest point, the position of the nearest point and the position of the joint in the real-time posture skeleton graph of the user are connected, and based on a preset weight, the position on the corresponding connecting line is extracted and set as the position after the joint is revised. For example, the preset weight is set to 0.5, and the middle point position of the corresponding connecting line is taken as the position of the joint after revision.

[0108] According to the embodiment of the application, further comprising:

[0109] If the current sports scene is a skiing scene, a subset of reference posture feature vectors of the current ski slope gradient and snow hardness is extracted;

[0110] The standard posture of the reference posture feature vector set in the preset sports posture feature big data cloud library has been annotated with the applicable slope range and snow hardness range;

[0111] According to the real-time slope information fed back by the slope sensor and the snow hardness selected by the user, the reference posture feature vector set of the current environment is dynamically selected, so that the reference posture feature vector is adaptively adjusted according to the environmental conditions.

[0112] The application discloses a sports posture recognition method based on big data and image recognition, and comprises the following steps: acquiring a sports video stream of a target user in real time, and extracting a user posture image sequence in the sports video stream; performing key point detection on the user posture image sequence, generating a user real-time posture skeleton graph, and calculating a real-time multi-dimensional posture feature vector based on the skeleton graph; extracting a set of reference posture feature vectors of a current sports item based on a preset sports posture feature big data cloud library; matching the real-time multi-dimensional posture feature vector with the reference posture feature vectors in the set of reference posture feature vectors, obtaining a matching degree; and identifying the category and quality level of the current sports posture according to the matching degree.

[0113] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other manners. The above described device embodiments are merely schematic. For example, the division of the units is only a logical function division. There can be another division manner for the actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling, or direct coupling or communication connection between the components can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or in other forms.

[0114] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units; they can be located in one place, or distributed on multiple network units; and part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0115] In addition, each functional unit in each embodiment of the present application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be realized in the form of hardware, or in the form of hardware plus software functional unit.

[0116] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by relevant hardware of program instructions, and the foregoing program can be stored in a computer readable storage medium, and the program performs the steps of the above-mentioned method embodiments when executed; and the foregoing storage medium includes: a mobile storage device, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0117] Alternatively, the integrated unit of the present application can also be stored in a computer readable storage medium if it is realized in the form of a software function module and sold or used as an independent product. Based on such understanding, the technical solutions of the embodiments of the present application can be embodied in the form of a software product, and the computer software product is stored in a storage medium, includes several instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the methods described in the embodiments of the present application. The foregoing storage medium includes: a mobile storage device, a ROM, a RAM, a magnetic disk or an optical disk, and various media that can store program codes.

Claims

1. A motion posture recognition method based on big data and image recognition, characterized in that, include: The motion video stream of the target user is acquired in real time, and the user posture image sequence in the motion video stream is extracted. Key point detection is performed on the user pose image sequence to generate a real-time user pose skeleton map, and a real-time multi-dimensional pose feature vector is calculated based on the skeleton map. Based on a pre-set big data cloud database of motion posture features, extract the benchmark posture feature vector set for the current sport. The matching degree is obtained by matching the real-time multi-dimensional attitude feature vector with the reference attitude feature vector set. The category and quality level of the current motion posture are identified based on the matching degree; Also includes: When the matching degree is higher than the preset matching degree threshold, extract the real-time multi-dimensional pose feature vector corresponding to the current matching degree; Set the real-time multi-dimensional posture feature vector corresponding to the matching degree as the high-quality user posture feature vector and supplement it to the baseline posture feature vector set. After performing key point detection on the user pose image sequence to generate a real-time user pose skeleton map, the method further includes: If the current activity is skiing, the user's body surface temperature distribution map will be obtained in real time based on the preset infrared thermal imaging sensor. Cluster analysis is performed on the temperature point cloud in the user's body surface temperature distribution map to obtain the point cloud map of the user's joints and determine the range of the user's joints. If the joints in the user's real-time pose skeleton graph fall within the corresponding joint range, then the current user's real-time pose skeleton graph is normal. If a joint in the user's real-time pose skeleton graph does not fall within the corresponding joint range, the position of that joint will be revised based on the corresponding joint range.

2. The motion posture recognition method based on big data and image recognition according to claim 1, characterized in that, The steps for calculating real-time multi-dimensional pose feature vectors based on the skeleton graph specifically include: Based on the skeleton diagram, calculate the bone vectors between the joints; Based on the bone vectors, calculate the set of angles between key bones. , where m is the number of angles of interest; Based on the changes of the skeleton map across consecutive frames in the user pose image sequence, the motion velocity and acceleration feature set of the joints is calculated. Where n is the number of joints; The angle set A and the motion feature set M are normalized and concatenated to form the real-time multi-dimensional pose feature vector. .

3. The motion posture recognition method based on big data and image recognition according to claim 1, characterized in that, The formula for matching the real-time multi-dimensional attitude feature vector with the reference attitude feature vector set to obtain the matching degree is as follows: Where S represents the matching degree; K represents the total number of motion stages; and i belongs to k. This represents the weight coefficient for the i-th stage; This represents the real-time multi-dimensional pose feature vector at the i-th stage; This indicates that in the set of reference attitude feature vectors, and The standard pose feature vector with the closest Euclidean distance.

4. The motion posture recognition method based on big data and image recognition according to claim 1, characterized in that, Also includes: Set the matching degree of the current frame to The matching degree of the previous frame is set to The difference in matching degree between adjacent frames is then... ; If the difference in matching degree between adjacent frames is greater than the preset coherence factor, an alert message will be triggered. Based on the warning information, if the matching degree of the current frame is less than the preset second matching degree threshold, the preset safety liner will be triggered. Based on the warning information, if the matching degree of the current frame is greater than or equal to the preset second matching degree threshold, the attitude feature vector of the current frame is optimized according to the attitude feature vector within the preset sliding time window to recalculate the matching degree of the current frame.

5. The motion posture recognition method based on big data and image recognition according to claim 4, characterized in that, Also includes: Once the warning message is triggered, the current video frame location is marked. After traversing all motion video streams, obtain all marked positions; Using any marked location as the center and a distance as the radius, construct an influence range and extract the number of times the mark is made within the influence range; If the number of markings within the affected area exceeds a preset threshold, a maintenance warning message for the current affected area will be generated.

6. The motion posture recognition method based on big data and image recognition according to claim 4, characterized in that, Also includes: Based on a preset sliding time window, obtain the matching degree set for different time periods; The average matching score for the corresponding time period is obtained by averaging the matching scores in the matching score set. Subtract the average matching score of the previous time period from the average matching score of the current time period to obtain the average matching score difference. If the absolute value of the average matching degree difference is less than the preset first average matching degree threshold, a smooth motion state prompt message is generated. If the average matching degree difference is greater than the preset second average matching degree threshold, a motion state improvement prompt message will be generated. If the average matching degree difference is less than the preset third average matching degree threshold, a motion state decline prompt message will be generated.

7. The motion posture recognition method based on big data and image recognition according to claim 6, characterized in that, After generating the motion state decline warning message, the method further includes: Extract the matching score set for the current time period, and extract the minimum matching score in the corresponding matching score set; The real-time multi-dimensional pose feature vector for the current motion is determined based on the minimum matching degree in the corresponding matching degree set. The difference between the real-time multi-dimensional posture feature vector during the current motion and the corresponding standard posture feature vector is calculated to obtain the motion posture feature vector adjustment value. The adjusted values ​​of the motion posture feature vector are sent to a preset screen for display.

Images