Human posture detection method and equipment based on machine vision

Abstract

The present invention provides a machine vision-based human posture detection method and device, which is applied to posture recognition in physical fitness testing. The method acquires a multi-frame continuous image sequence of a physical fitness training scene, performs human body region positioning processing on the multi-frame continuous image sequence, generates a dynamic human body region set of a target user, performs posture feature extraction processing on the dynamic human body region set, obtains a spatiotemporal posture feature set of the target user, and then performs posture state recognition processing on the spatiotemporal posture feature set based on a preset physical fitness assessment model to generate a posture state recognition result for the target user. Based on the posture state recognition result, physical fitness assessment parameters are generated, and the physical fitness assessment parameters are fed back to a physical fitness training guidance system. The present invention can improve the ability to identify subtle posture changes in complex physical fitness movements, achieve simultaneous assessment of movement standardization and physiological load status, and overcome the limitations of traditional single-dimensional movement evaluation.

Description

Technical Field

[0001] The present invention relates to the fields of machine vision and posture detection, and in particular to a method and device for detecting human posture based on machine vision. Background Art

[0002] In physical fitness testing, posture recognition technology provides a key basis for exercise effect evaluation and training program optimization by analyzing changes in body posture during the execution of training movements. Existing technologies usually perform static joint detection based on single-frame images or use time-series averaging to extract basic motion features, and judge the degree of movement standardization through preset threshold comparison; however, the rapid movement of the human body in dynamic training scenarios can easily lead to inaccurate regional positioning between consecutive frames and incomplete capture of limb extension range. The separation of three-dimensional spatial relationships and motion differential features in the feature extraction process weakens the ability to identify subtle differences in complex movements. The single-dimensional movement deviation judgment of the evaluation model cannot synchronously reflect abnormal physiological load distribution, resulting in physical fitness assessment parameters that are difficult to fully characterize the correlation between movement execution quality and body function status, limiting the accuracy and real-time nature of training guidance feedback. Summary of the Invention

[0003] The present invention provides a human body posture detection method and equipment based on machine vision.

[0004] In the first aspect, an embodiment of the present invention provides a human posture detection method based on machine vision, which is applied to posture recognition in physical fitness testing, including: obtaining a multi-frame continuous image sequence of a physical fitness training scene; performing human body region positioning processing on the multi-frame continuous image sequence to generate a dynamic human body region set of a target user; performing posture feature extraction processing on the dynamic human body region set to obtain a spatiotemporal posture feature set of the target user; performing posture state recognition processing on the spatiotemporal posture feature set based on a preset physical fitness assessment model to generate a posture state recognition result of the target user; generating physical fitness assessment parameters based on the posture state recognition result, and feeding the physical fitness assessment parameters back to a physical fitness training guidance system. In the second aspect, an embodiment of the present invention provides a posture detection device, including: a memory, in which a computer program is stored; and a processor, which is used to load the computer program to implement the human posture detection method based on machine vision as described above.

[0005] The human posture detection method based on machine vision provided by the present invention effectively solves the problem of inaccurate human body area tracking during movement through dynamic human body area positioning processing of multi-frame continuous image sequences, combined with temporal continuity verification and spatial calibration mechanism, ensures the complete capture of limb extension range, and provides high-precision dynamic human body area data for subsequent feature extraction; by temporally and spatially fusing the joint coordinates mapped in three-dimensional space with the differential features of the motion trajectory, and integrating the spatial topological coding of the joint distance and angle, a composite posture feature that simultaneously characterizes the kinematic state and dynamic characteristics is constructed, significantly improving the ability to identify subtle posture changes in complex physical fitness movements; based on the deviation analysis of feature space mapping and standard movement clustering centers, combined with the identification of abnormal areas of joint movement energy distribution, the synchronous evaluation of movement standardization and physiological load status is realized, breaking through the limitations of traditional single-dimensional movement evaluation, so that the physical fitness evaluation parameters can simultaneously reflect the quality of movement execution and the body function status, providing the training guidance system with both real-time and comprehensive feedback data, and achieving the coordinated optimization of accurate movement recognition and scientific load monitoring in dynamic training scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 This is a flow chart of a human posture detection method based on machine vision provided by an embodiment of the present invention.

[0007] Figure 2 The figure is a schematic diagram of the composition of a posture detection device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0008] See also Figure 1 , Figure 1 A flowchart of a human posture detection method based on machine vision is provided in an embodiment of the present invention. The method can be executed by a posture detection device and may include the following steps: Step S100: Acquire a multi-frame continuous image sequence of a physical fitness training scene. A multi-frame continuous image sequence refers to a series of images continuously acquired at time intervals in a physical fitness training scene, and these images can record the posture changes of the target user during the training process. It can be understood that in the embodiment of the present invention, the acquired user image information is a non-privacy-intrusive image acquired after the user's authorization. Acquiring a multi-frame continuous image sequence is the basic data source for subsequent human posture detection and analysis. In actual operation, the image acquisition device can be used to acquire images at a preset frame rate to ensure that the acquired image sequence has temporal continuity and integrity.

[0009] Optionally, before step S100, the following steps S101 to S107 may also be included: Step S101: Configure a multi-perspective visual acquisition device array, where the multi-perspective visual acquisition device array includes at least three spatially distributed image acquisition devices. The multi-perspective visual acquisition device array is a collection of multiple spatially distributed image acquisition devices, which capture images of the physical fitness training scene from different angles. By configuring the multi-perspective visual acquisition device array, image information of the target user at different perspectives can be obtained, thereby more comprehensively capturing the posture characteristics of the target user. The image acquisition device can be a camera, a camcorder, or other device with an image acquisition function. For example, three high-definition cameras are installed at different locations of the physical fitness training venue, such as in front, on the left, and on the right, to form a simple multi-perspective visual acquisition device array.

[0010] Step S102: Perform spatiotemporal synchronization processing on the multi-view visual acquisition device array and establish a unified timestamp system. Spatiotemporal synchronization processing refers to performing temporal and spatial synchronization operations on each image acquisition device in the multi-view visual acquisition device array to ensure that the images acquired by each device are consistent in time and space. The purpose of the unified timestamp system is to give each acquired image a unique time identifier so that images acquired by different devices can be accurately aligned in time. In actual operation, spatiotemporal synchronization processing can be achieved by hardware synchronization or software synchronization. For example, a high-precision clock source is used to provide a unified time signal for each image acquisition device, or the clocks of each device are calibrated through a network synchronization protocol to establish a unified timestamp system.

[0011] Step S103: Frame rate calibration is performed on each image acquisition device using a unified timestamp system. Frame rate calibration involves adjusting and calibrating the frame rates of each image acquisition device in the multi-view visual acquisition device array based on the unified timestamp system to ensure consistent frame rates across all devices. Frame rate refers to the number of frames captured per second by an image acquisition device. Inconsistent frame rates can lead to image timing mismatches in subsequent processing. Using a unified timestamp system, the frame rates of each device can be precisely controlled and adjusted.

[0012] Step S104: Calculate and process the overlapping area of ​​the field of view angles of each calibrated image acquisition device to determine the effective acquisition area. Calculating and processing the overlapping area of ​​the field of view angles refers to analyzing and calculating the field of view angles of each calibrated image acquisition device to find the overlapping parts between the field of view angles of each device. The effective acquisition area refers to the area where the multi-view visual acquisition device array can comprehensively and accurately capture the target user's posture information. By calculating the overlapping area of ​​the field of view angles, it is possible to avoid the occurrence of acquisition blind spots and ensure that all postures of the target user during physical fitness training can be captured. In actual calculations, the overlapping area of ​​the field of view angles can be determined using geometric calculation methods based on information such as the installation position and field of view angle parameters of the image acquisition device. For example, the field of view angle range of each camera is calculated using trigonometric functions, and then the overlapping parts between them are found to determine the effective acquisition area.

[0013] Step S105: Setting a physical fitness training motion capture space within the effective acquisition area. The physical fitness training motion capture space refers to a set space demarcated within the effective acquisition area, in which the target user performs physical fitness training so that the multi-view visual acquisition device array can better capture their posture information. When setting the physical fitness training motion capture space, it is necessary to consider the target user's training motion range and the acquisition capabilities of the multi-view visual acquisition device array. For example, a rectangular space can be demarcated within the effective acquisition area using lines or markers as the physical fitness training motion capture space to ensure that any movement of the target user within the space can be clearly captured by the image acquisition device.

[0014] Step S106: Perform adaptive lighting adjustment processing on the physical fitness training motion capture space to eliminate ambient light interference factors. Adaptive lighting adjustment processing refers to adjusting and optimizing the lighting conditions in the physical fitness training motion capture space to eliminate the interference of ambient light on image acquisition. Ambient light interference factors may cause problems such as uneven brightness and shadows in the captured images, affecting the subsequent human body area positioning and posture feature extraction. In actual operation, a variety of lighting adjustment methods can be used, such as using professional lighting equipment, adjusting lighting angles and brightness, etc. For example, a dimmable LED light is installed in the physical fitness training motion capture space to automatically adjust the brightness and angle of the light according to the intensity and direction of the ambient light, so that the lighting in the space is uniform and stable, eliminating ambient light interference.

[0015] Step S107: When the target user enters the physical fitness training motion capture space, the multi-view visual acquisition device array is triggered to synchronously capture images. The target user is a person who performs physical fitness training and needs to perform posture detection. When the target user enters the physical fitness training motion capture space, the entry of the target user is detected by a sensor or other detection means, and then the multi-view visual acquisition device array is triggered to start synchronous image capture. Synchronous image capture can ensure that the images captured by each device are consistent in time, providing an accurate data basis for subsequent processing. For example, an infrared sensor is installed at the entrance of the physical fitness training motion capture space. When the target user is detected to enter, the sensor sends a signal to the multi-view visual acquisition device array, triggering each device to start capturing images at the same time.

[0016] Step S200: Perform human body region positioning processing on a multi-frame continuous image sequence to generate a dynamic human body region set of the target user. Human body region positioning processing refers to accurately identifying and locating the human body region of the target user from a multi-frame continuous image sequence. The dynamic human body region set refers to a set of human body regions of the target user that dynamically changes over time in a multi-frame continuous image sequence. By performing human body region positioning processing on a multi-frame continuous image sequence, the human body region information of the target user at different times can be obtained, providing a basis for subsequent posture feature extraction and posture state recognition. In actual operation, a variety of image processing and computer vision technologies can be used to implement human body region positioning processing. For example, a background subtraction algorithm is used to separate the background and foreground in the image, and then the human target in the foreground is further identified and located.

[0017] Optionally, step S200 may specifically include the following steps S210 to S280: Step S210: Perform background separation processing on the multi-frame continuous image sequence to obtain a set of foreground moving objects. Background separation processing refers to the operation of separating the background and foreground portions of the multi-frame continuous image sequence. The set of foreground moving objects refers to the set of objects that move relative to the background in the image. In fitness training scenarios, this primarily refers to the human body of the target user. Through background separation processing, the target user can be separated from the background, facilitating subsequent processing of the human body region. In actual operation, a variety of background separation algorithms can be used, such as the Gaussian Mixture Model (GMM) algorithm and codebook algorithm. For example, a Gaussian Mixture Model algorithm can be used to process the multi-frame continuous image sequence. This algorithm models the pixel values ​​in the image and classifies the pixels into background pixels and foreground pixels, thereby achieving background and foreground separation and obtaining a set of foreground moving objects.

[0018] Step S220: Biometric matching is performed on the foreground moving object set to determine the target moving object outline of the target user. Biometric matching refers to the operation of matching objects in the foreground moving object set with the biometric features of the target user. The target moving object outline refers to the human body outline of the target user in the image. Through biometric matching, the human body outline of the target user can be accurately determined, eliminating interference from other non-target objects. In actual operation, biometric features such as facial features and body proportions of the target user can be used for matching. For example, a facial image of the target user is pre-captured, and then a face recognition algorithm is used to match the foreground moving object set to determine the target user's position and outline.

[0019] Step S230: Perform bounding box fitting on the target moving object's outline to generate an initial human body region bounding box set. Bounding box fitting refers to the operation of using a rectangular box to fit the target moving object's outline. The initial human body region bounding box set refers to a set of rectangular boxes obtained through the bounding box fitting process, which roughly surround the target user's human body area. Through the bounding box fitting process, the target moving object's outline can be simplified into a rectangular box, which facilitates subsequent processing and calculation. In actual operation, the minimum bounding rectangle algorithm can be used to perform the bounding box fitting process. For example, for an irregular human body outline, the minimum bounding rectangle algorithm is used to find a minimum rectangular box that can completely surround the outline, thereby generating an initial human body region bounding box set.

[0020] Step S240: Perform temporal continuity check processing on the initial human body region bounding box set to detect the regional displacement change between adjacent frames. Temporal continuity check processing refers to the operation of checking and verifying the continuity of the initial human body region bounding box set in the time series. The regional displacement change refers to the change in the position and size of the initial human body region bounding box in two adjacent frames. Through the temporal continuity check processing, the movement of the target user between adjacent frames can be detected to determine whether the positioning of the human body region is accurate and continuous. In actual operation, the difference in the coordinates of the center point of the initial human body region bounding box between adjacent frames, the rate of change of the size of the bounding box and other parameters can be calculated to detect the regional displacement change. For example, for the initial human body region bounding boxes in two adjacent frames, the difference in the x-coordinate and y-coordinate of their center points and the rate of change of the bounding box area are calculated as a measure of the regional displacement change.

[0021] Step S250: If the regional displacement change is less than the preset displacement threshold, spatial calibration is performed on the initial human body region bounding box set to generate a calibrated human body region bounding box set. The preset displacement threshold is a pre-set threshold used to determine whether the regional displacement change is normal. Spatial calibration refers to adjusting and optimizing the initial human body region bounding box set so that it more accurately reflects the human body region of the target user. The calibrated human body region bounding box set refers to the human body region bounding box set obtained after spatial calibration. When the regional displacement change is less than the preset displacement threshold, it indicates that the positioning of the human body region between adjacent frames has changed little, and errors may exist, requiring spatial calibration. In actual operation, an image registration algorithm can be used to adjust the initial human body region bounding box set. For example, an image registration algorithm based on feature point matching is used to find the feature point correspondence between adjacent frames. Then, based on these correspondences, the initial human body region bounding box is subjected to transformations such as translation, rotation, and scaling to generate a calibrated human body region bounding box set.

[0022] Step S260: Keyframe sampling is performed on the calibrated human body region bounding box set, and a sampling frame set that meets the posture change amplitude condition is selected. Keyframe sampling refers to the operation of selecting a representative set of frames from the calibrated human body region bounding box set as keyframes. The posture change amplitude condition refers to a pre-set condition used to determine whether the target user's posture change in a frame is significant. The sampling frame set refers to the set of frames selected through the keyframe sampling process that meet the posture change amplitude condition. Keyframe sampling can reduce the amount of data required for subsequent processing while retaining key information about the target user's posture change. In actual operation, the change in the position and angle of the calibrated human body region bounding box between adjacent frames can be calculated. When the change exceeds a preset posture change amplitude threshold, the frame is selected as a keyframe. For example, for each two adjacent frames in the calibrated human body region bounding box set, the change in the displacement and rotation angle of the bounding box center points is calculated. If the change exceeds the preset posture change amplitude threshold, the frame is added to the sampling frame set.

[0023] Step S270: Perform multi-scale region expansion processing on each sampling frame in the sampling frame set to generate a set of expanded human body regions that include the complete limb extension range. Multi-scale region expansion processing refers to the operation of expanding the human body region at different scales for each sampling frame in the sampling frame set. The expanded human body region set refers to the set of human body regions that include the complete limb extension range obtained after the multi-scale region expansion processing. Through the multi-scale region expansion processing, it is possible to ensure that the collected human body region contains the complete limb information of the target user, avoiding information loss caused by limb extension. In actual operation, a pyramid image scaling algorithm can be used to perform multi-scale processing on the sampling frames, and then the human body region is expanded at each scale. For example, the sampling frames are scaled by different multiples to obtain multiple images of different scales. The human body region bounding box is then expanded at each scale to include the possible limb extension range. Finally, the human body regions expanded at different scales are merged to generate the set of expanded human body regions.

[0024] Step S280: Perform spatiotemporal alignment processing on the extended human body region set and the calibrated human body region bounding box set of the unsampled frame to generate a dynamic human body region set. Spatiotemporal alignment processing refers to the operation of aligning the extended human body region set and the calibrated human body region bounding box set of the unsampled frame in time and space. The dynamic human body region set refers to the set of human body region information of the target user in the entire multi-frame continuous image sequence obtained after the spatiotemporal alignment processing. Through the spatiotemporal alignment processing, the extended human body region information of the key frame can be accurately mapped to the unsampled frame to obtain a continuous and complete dynamic human body region set. In actual operation, the spatiotemporal alignment processing can be implemented using a motion estimation algorithm based on the optical flow method. For example, for the calibrated human body region bounding box in the unsampled frame, the optical flow method is used to calculate its motion vector with the extended human body region in the adjacent sampled frame, and then the extended human body region information is mapped to the unsampled frame according to the motion vector to generate a dynamic human body region set.

[0025] Optionally, step S200, when performing human body region positioning processing on a multi-frame continuous image sequence, may further include the following steps S201 to S207: Step S201: Acquire a synchronized multi-perspective image set through a multi-perspective visual acquisition device array. A synchronized multi-perspective image set refers to a set of images acquired from different angles by a multi-perspective visual acquisition device array at the same time. By acquiring a synchronized multi-perspective image set, image information of the target user at different perspectives can be obtained, providing richer data for subsequent three-dimensional human body surface model reconstruction and human body region positioning. In actual operation, since the multi-perspective visual acquisition device array has been subjected to spatiotemporal synchronization processing, it can be ensured that each device acquires images at the same time. For example, in a physical fitness training scenario, the three cameras in the front, left, and right sides respectively acquire images of the front, left side, and right side of the target user at the same time, and these images constitute a synchronized multi-perspective image set.

[0026] Step S202: Perform disparity calculation processing on the synchronized multi-view image set to generate a depth information map. Disparity calculation processing refers to the operation of analyzing and calculating images from different perspectives in the synchronized multi-view image set to find the disparity between corresponding points. The depth information map refers to an image obtained through disparity calculation processing that reflects the depth information of objects in the image. In stereo vision, the position of the same object in the image will be different from different perspectives. This position difference is the disparity. By calculating the disparity, the depth information of the object can be obtained. In actual operation, a stereo matching algorithm can be used to perform disparity calculation processing on the synchronized multi-view image set. For example, a semi-global matching (SGM) algorithm is used to find corresponding points in the synchronized multi-view image set, calculate the disparity between them, and then generate a depth information map based on the disparity and information such as the camera's intrinsic and extrinsic parameters.

[0027] Step S203: Perform spatial point cloud reconstruction processing on the depth information map to generate a three-dimensional human body surface model. Spatial point cloud reconstruction processing refers to the operation of converting the depth information in the depth information map into point cloud data in three-dimensional space, and reconstructing a three-dimensional human body surface model based on these point cloud data. A three-dimensional human body surface model refers to a model that uses geometric elements such as points, lines, and surfaces in three-dimensional space to represent the surface shape of the target user's body. Through spatial point cloud reconstruction processing, the three-dimensional human body surface information of the target user can be obtained, providing a more accurate basis for subsequent posture analysis. In actual operation, a point cloud generation algorithm based on depth information can be used to convert the depth information map into point cloud data, and then the three-dimensional human body surface model can be generated using point cloud processing and surface reconstruction algorithms. For example, a Poisson surface reconstruction algorithm is used to construct a smooth three-dimensional human body surface model based on point cloud data.

[0028] Step S204: Perform motion artifact detection processing on the three-dimensional human body surface model to identify model distortion areas caused by rapid motion. Motion artifact detection processing refers to the operation of inspecting and analyzing the three-dimensional human body surface model to find out the model distortion areas caused by the rapid motion of the target user. The model distortion area refers to the area in the three-dimensional human body surface model where the information is inaccurate or missing due to rapid motion. When the target user moves rapidly, the multi-view visual acquisition device array may not be able to capture the posture information of the human body in a timely and accurate manner, resulting in distortion of the three-dimensional human body surface model. In actual operation, the model distortion area can be detected by analyzing the geometric features, texture information, etc. of the three-dimensional human body surface model. For example, the curvature of the three-dimensional human body surface model is calculated using a curvature analysis algorithm. When the curvature changes abnormally large, it is determined that the area may be a model distortion area.

[0029] Step S205: Perform multi-frame data compensation processing on the distorted area of ​​the model, and use the three-dimensional human body surface model of the adjacent frames to perform data repair. Multi-frame data compensation processing refers to the operation of supplementing and repairing the distorted area of ​​the model using the information in the three-dimensional human body surface model of the adjacent frames. The accuracy and integrity of the three-dimensional human body surface model can be improved by multi-frame data compensation processing. In actual operation, the three-dimensional human body surface model of the adjacent frames can be aligned with the model of the current frame using image registration and data fusion algorithms, and then the information corresponding to the distorted area of ​​the model can be extracted from the adjacent frames to perform data repair on the distorted area of ​​the model of the current frame. For example, an image registration algorithm based on feature point matching is used to find the correspondence between adjacent frames, and then the point cloud data of the corresponding area in the adjacent frames is copied to the model distorted area of ​​the current frame to achieve data repair.

[0030] Step S206: Project the repaired three-dimensional human body surface model onto a two-dimensional image plane to generate an optimized human body contour map. The projection operation refers to the operation of projecting an object in three-dimensional space onto a two-dimensional plane. The optimized human body contour map refers to a more accurate and clearer human body contour image obtained after projecting the repaired three-dimensional human body surface model onto a two-dimensional image plane. By projecting the three-dimensional human body surface model onto a two-dimensional image plane, the three-dimensional information can be converted into two-dimensional information, which is convenient for subsequent human body area positioning processing. In actual operation, the camera's projection model can be used to project the points in the three-dimensional human body surface model onto the two-dimensional image plane, and then generate an optimized human body contour map based on the projection results. For example, based on the camera's internal and external parameters, each point in the three-dimensional human body surface model is projected onto the two-dimensional image plane, and the projected points are connected to form a human body contour to generate an optimized human body contour map.

[0031] Step S207: Input the optimized human body contour image into the human body region positioning process to assist in generating a dynamic human body region set. Inputting the optimized human body contour image into the human body region positioning process can provide more accurate human body contour information for human body region positioning, thereby assisting in generating a more accurate dynamic human body region set. In actual operation, the optimized human body contour image can be integrated and verified with the human body region positioning results previously obtained by other methods. For example, the optimized human body contour image is compared with the human body region obtained through background subtraction and target recognition, and inconsistent areas are adjusted and corrected to ultimately generate a more accurate dynamic human body region set.

[0032] Step S300: Perform posture feature extraction processing on the dynamic human body region set to obtain the spatiotemporal posture feature set of the target user. Posture feature extraction processing refers to the operation of extracting features that can reflect the posture information of the target user from the dynamic human body region set. The spatiotemporal posture feature set refers to a set containing the posture features of the target user in time and space. By performing posture feature extraction processing on the dynamic human body region set, the posture features of the target user at different times and different spatial positions can be obtained, providing key information for subsequent posture state recognition and physical fitness assessment. In actual operation, a variety of computer vision and machine learning technologies can be used to implement posture feature extraction processing. For example, a deep learning model is used to analyze the dynamic human body region set to extract features such as the coordinates of key joint points and posture angles.

[0033] Optionally, step S300 may specifically include the following steps S310 to S380: Step S310: Perform joint key point detection on each dynamic human region in the dynamic human region set to generate a basic key point coordinate set. Joint key point detection refers to the operation of accurately identifying and locating key points of human joints within each dynamic human region in the dynamic human region set. The basic key point coordinate set refers to the set of coordinates of human joint key points on the image plane obtained through joint key point detection. Human joint key points refer to characteristic points at human skeletal joints, such as the center points of joints such as the shoulder, elbow, wrist, hip, knee, and ankle. By detecting joint key points, the target user's skeletal structure information can be obtained, providing a basis for subsequent posture analysis. In practice, a deep learning-based joint key point detection algorithm can be used to process the dynamic human region set. For example, the OpenPose algorithm can be used. This algorithm analyzes dynamic human regions using a convolutional neural network (CNN), outputs the coordinates of human joint key points, and generates a basic key point coordinate set.

[0034] Step S320: Perform three-dimensional space mapping processing on the basic key point coordinate set, convert the two-dimensional image coordinates into three-dimensional space coordinates, and generate a spatial key point coordinate set. Three-dimensional space mapping processing refers to the operation of converting the two-dimensional image coordinates in the basic key point coordinate set into three-dimensional space coordinates. The spatial key point coordinate set refers to the set of coordinates of the human joint key points in three-dimensional space obtained after the three-dimensional space mapping processing. In a two-dimensional image, only the planar position information of the joint key points can be obtained. Through the three-dimensional space mapping processing, this two-dimensional information can be converted into position information in three-dimensional space, which more accurately reflects the posture of the target user. In actual operation, it is necessary to combine the camera's intrinsic parameters, extrinsic parameters and other information to perform three-dimensional space mapping processing. For example, based on the camera's calibration parameters, the principle of triangulation is used to convert the two-dimensional image coordinates in the basic key point coordinate set into three-dimensional space coordinates to generate a spatial key point coordinate set.

[0035] Step S330: Perform motion trajectory modeling on the spatial key point coordinate set to construct a key point motion trajectory curve. Motion trajectory modeling refers to the operation of analyzing and modeling the changes in the spatial key point coordinate set in the time series to construct a motion trajectory curve of the key points of the human joints. The key point motion trajectory curve refers to the use of mathematical functions or curves to represent the trajectory of the key points of the human joints that changes over time in three-dimensional space. Through motion trajectory modeling, the movement of the joints of the target user during physical fitness training can be obtained, providing a basis for subsequent motion feature analysis. In actual operation, the spatial key point coordinate set can be processed using methods such as polynomial fitting and spline interpolation to construct a key point motion trajectory curve. For example, the cubic spline interpolation method is used to construct a smooth key point motion trajectory curve based on the coordinate values ​​in the spatial key point coordinate set.

[0036] Step S340: Perform differential feature calculation on the key point motion trajectory curve to extract the key point motion velocity feature and motion acceleration feature. Differential feature calculation refers to the operation of performing a derivative operation on the key point motion trajectory curve to calculate the motion velocity and motion acceleration of the key point. The key point motion velocity feature is the motion velocity information of the key points of the human joint in three-dimensional space, and the motion acceleration feature is the motion acceleration information of the key points of the human joint in three-dimensional space. By extracting the key point motion velocity feature and motion acceleration feature, the target user's motion dynamic information during fitness training can be obtained, providing an important basis for subsequent posture analysis and fitness assessment. In actual operation, the key point motion trajectory curve can be processed using a numerical differentiation method. For example, using the finite difference method, the coordinate values ​​of the key point motion trajectory curve at adjacent time points are differentially calculated to obtain the motion velocity and motion acceleration of the key point, and the key point motion velocity feature and motion acceleration feature are extracted.

[0037] Step S350: Performing time-dimensional fusion processing on the key point motion velocity features and motion acceleration features to generate a short-term motion feature vector. Time-dimensional fusion processing refers to the operation of integrating and processing the key point motion velocity features and motion acceleration features in the time dimension to generate a feature vector that comprehensively reflects the target user's short-term motion situation. The short-term motion feature vector refers to a feature vector containing key point motion velocity and motion acceleration information obtained through time-dimensional fusion processing. Through time-dimensional fusion processing, the motion velocity and motion acceleration information at different time points can be integrated to obtain a more representative short-term motion feature. In actual operation, the key point motion velocity features and motion acceleration features can be arranged in chronological order to form a vector, and then the vector can be normalized and processed to generate the short-term motion feature vector. For example, the key point motion velocity and motion acceleration at each time point can be used as vector elements to form a high-dimensional vector, and then the vector is processed using a normalization algorithm to ensure that the values ​​of its elements are within a preset range to generate the short-term motion feature vector.

[0038] Step S360: Perform relative position relationship calculation processing on the spatial key point coordinate set to generate adjacent joint distance features and joint angle features. Relative position relationship calculation processing refers to the operation of analyzing and calculating the relative position relationship between the human joint key points in the spatial key point coordinate set to obtain features such as adjacent joint distances and joint angles. The adjacent joint distance feature refers to the distance information between two adjacent joints, and the joint angle feature refers to the angle information between adjacent joints. By generating adjacent joint distance features and joint angle features, static structural information of the target user's human posture can be obtained, providing a basis for subsequent posture analysis and physical fitness assessment. In actual operation, the Euclidean distance formula can be used to calculate the distance between adjacent joints, and the vector angle formula can be used to calculate the angle between adjacent joints. For example, for two adjacent joint key points, the square root of the sum of the squares of their coordinate differences in three-dimensional space is calculated to obtain the adjacent joint distance; based on the vector formed by the two joint key points and the joint key point they are connected to, the vector dot product formula is used to calculate the angle between them to generate the joint angle feature.

[0039] Step S370: Perform spatial topological coding on the adjacent joint distance features and joint angle features to generate a static posture feature vector. Spatial topological coding refers to the operation of encoding and processing the adjacent joint distance features and joint angle features to generate a feature vector that can reflect the static posture structure information of the target user. The static posture feature vector refers to the feature vector containing adjacent joint distance and joint angle information obtained through spatial topological coding. Through spatial topological coding, the adjacent joint distance features and joint angle features can be integrated and compressed to facilitate subsequent posture state recognition and analysis. In actual operation, a dimensionality reduction algorithm such as principal component analysis (PCA) can be used to process the adjacent joint distance features and joint angle features to generate a static posture feature vector. For example, using the PCA algorithm, the covariance matrix of the adjacent joint distance features and joint angle features is calculated to find their principal components. The original features are then projected onto the principal components to obtain a reduced-dimensional static posture feature vector.

[0040] Step S380: Perform spatiotemporal feature fusion processing on the short-term motion feature vector and the static posture feature vector to generate a spatiotemporal posture feature set. Spatiotemporal feature fusion processing refers to the operation of integrating and processing the short-term motion feature vector and the static posture feature vector to generate a feature set that comprehensively reflects the target user's posture information in time and space. The spatiotemporal posture feature set refers to the set of spatiotemporal posture features of the target user obtained after the spatiotemporal feature fusion processing. Through the spatiotemporal feature fusion processing, the dynamic motion information and static posture structure information of the target user can be combined to more comprehensively reflect the target user's posture. In actual operation, the short-term motion feature vector and the static posture feature vector can be spliced ​​into a longer vector, or they can be fused using methods such as weighted summation to generate a spatiotemporal posture feature set. For example, the short-term motion feature vector and the static posture feature vector can be weighted summed according to set weights to obtain a new feature vector as an element of the spatiotemporal posture feature set.

[0041] Step S400: Based on a preset physical fitness assessment model, posture state recognition processing is performed on the spatiotemporal posture feature set to generate a posture state recognition result for the target user. The physical fitness assessment model is a pre-trained model used to assess the physical fitness and posture state of the target user. Posture state recognition processing refers to the use of the physical fitness assessment model to analyze and determine the spatiotemporal posture feature set to identify the target user's posture state. The posture state recognition result refers to information about the target user's posture state obtained after posture state recognition processing, such as whether the posture is standard and whether there are any movement distortions. By performing posture state recognition processing on the spatiotemporal posture feature set, the target user's physical fitness training status can be evaluated and feedback can be provided. In practice, the physical fitness assessment model can adopt a variety of machine learning and deep learning models, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs). For example, using a CNN-based physical fitness assessment model, the spatiotemporal posture feature set is input into the model. After a series of convolution, pooling, and fully connected layers, the model outputs the posture state recognition result.

[0042] Optionally, step S400 may specifically include the following steps S410 to S480: Step S410: Input the spatiotemporal posture feature set into the posture feature encoding module of the physical fitness assessment model for high-dimensional feature mapping to generate an abstract posture feature space. The posture feature encoding module is a key component of the physical fitness assessment model, responsible for encoding and processing the input spatiotemporal posture feature set. High-dimensional feature mapping involves mapping the features in the spatiotemporal posture feature set into a higher-dimensional space to better represent and distinguish different posture features. The abstract posture feature space is an abstract feature space obtained after high-dimensional feature mapping, where the features are more representative and discriminative. In practice, the posture feature encoding module may employ a neural network structure such as a multi-layer perceptron (MLP). For example, an MLP comprising multiple hidden layers is used to process the spatiotemporal posture feature set. The feature set is input to the input layer of the MLP, and after nonlinear transformation by the hidden layers, the feature set is output to the output layer to generate the abstract posture feature space.

[0043] Step S420: Perform classification boundary division processing in the abstract posture feature space to determine the standard action feature cluster center. Classification boundary division processing refers to the operation of dividing different classification boundaries according to different posture categories in the abstract posture feature space. The standard action feature cluster center refers to the cluster center representing the standard action feature in the abstract posture feature space. By performing classification boundary division processing and determining the standard action feature cluster center, a reference and basis can be provided for subsequent posture state recognition. In actual operation, a clustering algorithm can be used to perform cluster analysis on the feature points in the abstract posture feature space to determine the standard action feature cluster center. For example, the K-means clustering algorithm is used to divide the feature points in the abstract posture feature space into different clusters, and the center point of each cluster is the standard action feature cluster center.

[0044] Step S430: Calculate the feature deviation between the abstract posture feature space and the standard action feature cluster center. Feature deviation refers to the degree of difference between the feature points in the abstract posture feature space and the standard action feature cluster center. By calculating the feature deviation, the similarity between the target user's posture and the standard action can be determined. In actual operation, measurement methods such as Euclidean distance and cosine similarity can be used to calculate the feature deviation. For example, the Euclidean distance formula is used to calculate the distance between each feature point in the abstract posture feature space and the standard action feature cluster center as a measure of feature deviation.

[0045] Step S440: Dynamic threshold comparison processing is performed on the feature deviation, and a motion deformation indicator is generated when the feature deviation exceeds the first deviation threshold. Dynamic threshold comparison processing refers to the operation of comparing the feature deviation with a pre-set first deviation threshold. The first deviation threshold is a threshold used to determine whether the target user's posture is deformed. The motion deformation indicator refers to an indicator generated when the feature deviation exceeds the first deviation threshold, which is used to indicate that the target user's posture may be deformed. Through dynamic threshold comparison processing and the generation of motion deformation indicators, posture problems of the target user can be discovered in a timely manner. In actual operation, the size of the first deviation threshold can be adjusted according to different training movements and user groups. For example, for some more complex training movements, the first deviation threshold can be appropriately increased; for beginners, the first deviation threshold can be appropriately reduced. When the feature deviation exceeds the first deviation threshold, the system generates a motion deformation indicator, recording that there may be a problem with the posture of the frame.

[0046] Step S450: Perform energy consumption pattern recognition processing in the abstract posture feature space and extract joint motion energy distribution characteristics. Energy consumption pattern recognition processing refers to the operation of analyzing and identifying the energy consumption pattern of the target user during the physical fitness training process in the abstract posture feature space. Joint motion energy distribution characteristics refer to the characteristics of the motion energy distribution of each joint of the target user during the physical fitness training process. Through energy consumption pattern recognition processing and extraction of joint motion energy distribution characteristics, the energy consumption of the target user during the training process can be understood, providing a reference for subsequent physical fitness evaluation. In actual operation, the energy consumption model can be used to analyze the features in the abstract posture feature space and extract the joint motion energy distribution characteristics. For example, based on information such as the joint's movement speed, acceleration and joint mass, the kinetic energy and potential energy of the joint are calculated to obtain the joint motion energy distribution characteristics.

[0047] Step S460: Perform pattern matching processing on the energy distribution characteristics of joint motion and identify abnormal energy consumption areas. Pattern matching processing refers to the operation of matching and comparing the energy distribution characteristics of joint motion with a pre-set energy consumption pattern. The abnormal energy consumption area refers to an area in the energy distribution characteristics of joint motion that is significantly different from the standard energy consumption pattern. Through pattern matching processing and identifying abnormal energy consumption areas, it is possible to discover unreasonable energy consumption that may exist in the target user during the training process. In actual operation, a template matching algorithm can be used to process the energy distribution characteristics of joint motion. For example, a correlation-based template matching algorithm is used to compare the energy distribution characteristics of joint motion with the standard energy consumption pattern. When the correlation is lower than a threshold, the area is judged to be an abnormal energy consumption area.

[0048] Step S470: Perform correlation analysis on the motion deformation identifier and the abnormal energy consumption area to generate a posture stability evaluation index. Correlation analysis refers to the operation of analyzing and judging the correlation relationship between the motion deformation identifier and the abnormal energy consumption area. The posture stability evaluation index refers to an index for evaluating the posture stability of the target user obtained through correlation analysis. Motion deformation and abnormal energy consumption may both be related to the posture stability of the target user. By performing correlation analysis on them, the posture stability of the target user can be more comprehensively evaluated. In actual operation, a statistical analysis method can be used to perform correlation analysis on the motion deformation identifier and the abnormal energy consumption area. For example, the correlation coefficient between the motion deformation identifier and the abnormal energy consumption area is calculated, and the posture stability evaluation index is generated based on the correlation coefficient.

[0049] Step S480: Compare the posture stability evaluation index with the preset physical fitness standard interval to generate a posture state recognition result. The preset physical fitness standard interval is a pre-set interval for evaluating the physical fitness and posture state of the target user. Comparison processing refers to the operation of comparing and judging the posture stability evaluation index with the preset physical fitness standard interval. Through the comparison processing, it can be determined whether the posture state of the target user meets the standard. In actual operation, if the posture stability evaluation index is within the preset physical fitness standard interval, the posture state of the target user is judged to be normal; if it exceeds the preset physical fitness standard interval, it is judged that there is a problem with the posture state of the target user, and the corresponding posture state recognition result is generated.

[0050] Optionally, the physical fitness assessment model includes a motion standardization analysis module and a fatigue state prediction module. Based on this, after step S480, steps S490 to S4150 may also be included: Step S490: The motion standardization analysis module performs motion cycle decomposition processing on the spatiotemporal posture feature set to identify repeated motion units. The motion standardization analysis module is a module in the physical fitness assessment model used to analyze the target user's motion standardization. Motion cycle decomposition processing refers to the operation of decomposing the motion sequence in the spatiotemporal posture feature set into repeated motion units. Repeated motion units refer to basic motion units that appear repeatedly during physical fitness training. Through motion cycle decomposition processing and identification of repeated motion units, a more detailed analysis of the target user's motion standardization can be performed. In actual operation, the spatiotemporal posture feature set can be processed using a time domain analysis method. For example, repeated motion units can be identified by detecting periodic changes in the spatiotemporal posture feature set.

[0051] Step S4100: Phase alignment is performed on the repetitive action units to generate a standardized action cycle sequence. Phase alignment refers to the operation of aligning the repetitive action units in time so that their starting points and ending points are consistent in time. A standardized action cycle sequence refers to a standardized action cycle sequence obtained after phase alignment, in which each action cycle has the same time length and characteristics. By phase alignment and generating a standardized action cycle sequence, the degree of action standardization of the target user can be more conveniently compared and analyzed. In actual operation, template matching and other methods can be used to perform phase alignment on the repetitive action units. For example, a standard action cycle is selected as a template, and other repetitive action units are matched and aligned with the template to generate a standardized action cycle sequence.

[0052] Step S4110: Perform amplitude fluctuation analysis on the standardized action cycle sequence and calculate the action execution consistency parameter. Amplitude fluctuation analysis refers to the operation of analyzing and calculating the action amplitude in the standardized action cycle sequence to find out its fluctuation. The action execution consistency parameter refers to a parameter obtained through amplitude fluctuation analysis for measuring the consistency of the target user's action execution. The action execution consistency reflects the stability of the target user when repeatedly performing the same action. In actual operation, the amplitude mean and standard deviation of each action cycle in the standardized action cycle sequence can be calculated, and the standard deviation can be used as a measurement indicator of the action execution consistency parameter. For example, for each action cycle in the standardized action cycle sequence, the amplitude mean and standard deviation are calculated. The smaller the standard deviation, the higher the action execution consistency.

[0053] Step S4120: The fatigue state prediction module performs muscle activation pattern extraction processing on the spatiotemporal posture feature set to generate muscle synergy features. The fatigue state prediction module is a module in the physical fitness assessment model used to predict the fatigue state of the target user. Muscle activation pattern extraction processing refers to the operation of extracting the target user's muscle activation pattern from the spatiotemporal posture feature set. Muscle synergy features refer to a feature obtained through muscle activation pattern extraction processing that can reflect the target user's muscle synergy. During physical fitness training, different muscles work together to complete various movements. By extracting muscle synergy features, the target user's muscle fatigue status can be understood. In actual operation, the spatiotemporal posture feature set can be processed using methods such as surface electromyography (sEMG) analysis.

[0054] Step S4130: Perform time-domain attenuation analysis on the muscle synergy features to identify muscle fatigue trend parameters. Time-domain attenuation analysis refers to the operation of analyzing and calculating the changes in the muscle synergy features in the time series to find out their attenuation trends. The muscle fatigue trend parameter refers to a parameter obtained through time-domain attenuation analysis for measuring the muscle fatigue trend of the target user. As the training time increases, the degree of muscle fatigue will gradually increase, and the muscle synergy features will attenuate. By identifying the muscle fatigue trend parameters, the fatigue state of the target user can be predicted. In actual operation, linear regression and other methods can be used to perform time-domain attenuation analysis on the muscle synergy features. For example, linear regression is performed on the time series data of the muscle synergy features, and its slope is calculated, and the slope is used as a measurement indicator of the muscle fatigue trend parameter. The larger the slope is and the larger the absolute value is, the more obvious the muscle fatigue trend is.

[0055] Step S4140: Perform multimodal fusion processing on the action execution consistency parameters and the muscle fatigue trend parameters to generate a comprehensive fatigue state coefficient. Multimodal fusion processing refers to the operation of integrating and processing the action execution consistency parameters and the muscle fatigue trend parameters to generate a coefficient that can comprehensively reflect the fatigue state of the target user. The comprehensive fatigue state coefficient refers to a coefficient obtained through multimodal fusion processing for measuring the fatigue state of the target user. Both action execution consistency and muscle fatigue trend are related to the fatigue state of the target user. By performing multimodal fusion processing on them, the fatigue state of the target user can be evaluated more comprehensively. In actual operation, the action execution consistency parameters and muscle fatigue trend parameters can be multimodally fused using methods such as weighted summation. For example, weights are assigned to the action execution consistency parameters and the muscle fatigue trend parameters respectively, and then they are weighted summed to obtain the comprehensive fatigue state coefficient.

[0056] Step S4150: Adjusting the dynamic evaluation weights in the posture state recognition results based on the comprehensive fatigue state coefficient. Dynamic evaluation weights refer to weights used to measure the importance of different factors in the posture state recognition results. Adjusting the dynamic evaluation weights in the posture state recognition results based on the comprehensive fatigue state coefficient can more accurately reflect the actual posture state of the target user. When the comprehensive fatigue state coefficient is high, it indicates that the target user may be in a relatively fatigued state. In this case, the requirements for movement standardization can be appropriately lowered, the tolerance for factors such as movement deformation can be increased, and the dynamic evaluation weights can be adjusted accordingly. When the comprehensive fatigue state coefficient is low, it indicates that the target user is in a relatively good state. The requirements for movement standardization can be increased, and the dynamic evaluation weights can be reallocated. For example, a mapping table between the comprehensive fatigue state coefficient and the dynamic evaluation weights can be pre-set. Based on the calculated comprehensive fatigue state coefficient, the corresponding dynamic evaluation weight is searched from the mapping table, and the weights in the posture state recognition results are adjusted to make the posture state recognition results more consistent with the actual situation of the target user.

[0057] Step S500: Generate fitness assessment parameters based on the posture state recognition results and feed them back to the fitness training guidance system. Fitness assessment parameters are a series of parameters generated by comprehensively considering the posture state recognition results to assess the target user's physical fitness. They can reflect the target user's performance during fitness training, such as the degree of movement standardization, fatigue status, joint load, and body balance ability. By feeding these parameters back to the fitness training guidance system, the system can develop a more scientific and reasonable training plan for the target user based on these assessment results. When generating fitness assessment parameters, it is necessary to comprehensively consider multiple aspects of the posture state recognition results, such as movement deformation, abnormal energy consumption areas, posture stability assessment indicators, movement execution consistency parameters, and muscle fatigue trend parameters. For example, these indicators can be weighted and summed according to set weights to obtain a comprehensive fitness assessment score, which serves as part of the fitness assessment parameters.

[0058] Optionally, step S500 may specifically include the following steps S510-S550: Step S510: Performing action type classification processing on the posture state recognition results to determine the current training action category, and matching the current training action category with baseline physiological parameters in a preset action standard library. Action type classification processing refers to accurately classifying the target user's current training action into a preset action category based on the feature information in the posture state recognition results. The current training action category may be various common fitness exercises, such as running, squats, and push-ups. The preset action standard library is a pre-established database containing baseline physiological parameters for various training actions. These baseline physiological parameters are derived through research and analysis of a large number of professional athletes or standard action samples and represent the physiological indicators of the movements under standard execution. For example, for the squat movement, the baseline physiological parameters may include a standard pressure distribution range for joints, a standard muscle activation sequence, etc. In actual operation, a pattern recognition algorithm can be used to analyze the posture state recognition results to determine the current training action category, and then search the preset action standard library for the corresponding baseline physiological parameters. For example, the support vector machine (SVM) algorithm is used to classify the feature vectors in the posture state recognition results, determine that the current training action is a squat, and then retrieve the baseline physiological parameters of the squat action from the preset action standard library.

[0059] Step S520: Perform joint load distribution analysis on the posture state recognition results to generate actual joint load parameters. Joint load distribution analysis refers to an operation that deeply analyzes the joint-related information in the posture state recognition results and calculates the actual load conditions borne by each joint during training. The actual joint load parameters are parameters that reflect the actual load size and distribution borne by each joint of the target user during training. They are of great significance for assessing the health status of joints and training risks. During physical fitness training, different movements will produce different loads on the joints. By analyzing the joint load distribution, potential joint injury risks can be discovered in a timely manner.

[0060] Optionally, step S520 may specifically include the following steps S521~S526: Step S521: Perform joint position data extraction processing on the posture state recognition result to obtain the three-dimensional space coordinates and movement direction vector of each joint. Joint position data extraction processing refers to the operation of extracting specific position information and movement direction information about each joint from the posture state recognition result. The three-dimensional space coordinates can accurately describe the position of the joint in the three-dimensional space, and the movement direction vector represents the direction of the joint during the movement. In actual operation, the three-dimensional space coordinates of each joint can be directly extracted from the previously generated set of spatial key point coordinates, and the movement direction vector can be calculated based on the change of the joint coordinates between adjacent frames. For example, for the knee joint, the three-dimensional space coordinates of the knee joint key points in two adjacent frames are compared, the coordinate difference is calculated, and the difference vector is normalized to obtain the movement direction vector of the knee joint.

[0061] Step S522: Perform joint contact area detection based on the three-dimensional spatial coordinates and the motion direction vector to identify the contact pressure distribution area of ​​each joint during the motion process. Joint contact area detection refers to the operation of analyzing and determining the area of ​​contact between each joint and the surrounding tissues or objects during the motion process based on the three-dimensional spatial coordinates and the motion direction vector of the joint, and further identifying the pressure distribution of these contact areas. In physical fitness training, the contact pressure distribution of the joints will affect the health and movement efficiency of the joints. For example, in a squat movement, the contact area between the knee joint and the ground or support and the contact pressure distribution are very important for judging the stress condition of the knee joint. In actual operation, the finite element analysis method can be used in combination with the biomechanical model to perform joint contact area detection. First, a mechanical model of the joint is constructed based on the three-dimensional spatial coordinates and the motion direction vector of the joint, and then the stress condition of the joint during the motion process is simulated, and the pressure distribution of the joint contact area is calculated by finite element analysis software.

[0062] Step S523: Perform dynamic pressure threshold comparison processing on the contact pressure distribution area, and extract the set of abnormal contact areas that exceed the pressure threshold. Dynamic pressure threshold comparison processing refers to the operation of comparing the pressure value in the contact pressure distribution area with a pre-set dynamic pressure threshold. The dynamic pressure threshold is a pressure limit that is dynamically adjusted according to factors such as different training movements, joint types, and the physical condition of the target user. The set of abnormal contact areas refers to the set of areas in the contact pressure distribution area where the pressure value exceeds the dynamic pressure threshold. By extracting the set of abnormal contact areas, it is possible to promptly discover possible excessive stress conditions in the joints, providing a basis for subsequent joint injury risk assessment. In actual operation, for each contact pressure distribution area, its pressure value is compared with the dynamic pressure threshold. If the pressure value exceeds the threshold, the area is added to the set of abnormal contact areas. For example, for the contact pressure distribution area of ​​the knee joint, a dynamic pressure threshold is set to 100N / cm. 2 When the pressure value of a certain area reaches 120N / cm 2 , the area is marked as an abnormal contact area and added to the abnormal contact area set.

[0063] Step S524: Perform pressure gradient analysis on the abnormal contact area set to generate joint local pressure distribution characteristics. Pressure gradient analysis refers to the operation of analyzing and calculating the pressure changes in the abnormal contact area set. Joint local pressure distribution characteristics refer to information obtained through pressure gradient analysis that can reflect the law and characteristics of joint local pressure changes. The pressure gradient reflects the rate of change of pressure in space. By analyzing the pressure gradient, the concentration degree and change trend of joint local pressure can be understood. In actual operation, the numerical differentiation method can be used to calculate the ratio of the pressure difference to the distance difference between adjacent points in the abnormal contact area set to obtain the pressure gradient. Then, the joint local pressure distribution characteristics are generated based on information such as the size and direction of the pressure gradient. For example, for an abnormal contact area, the pressure gradient between adjacent points inside it is calculated, and the areas with larger pressure gradients are marked as pressure concentration areas. The position, size and pressure gradient value of these pressure concentration areas and other information are used as joint local pressure distribution characteristics.

[0064] Step S525: Perform muscle group activation degree analysis on the three-dimensional spatial coordinates and the motion direction vector, and calculate the muscle synergistic activation parameter based on the correlation between the motion direction vector and the preset muscle group action direction. The muscle group activation degree analysis refers to the operation of analyzing and calculating the activation degree of the muscle group related to the joint based on the three-dimensional spatial coordinates and the motion direction vector of the joint. The muscle synergistic activation parameter is a parameter that reflects the collaborative work between muscle groups. It is obtained by calculating the correlation between the motion direction vector and the preset muscle group action direction. In physical fitness training, different muscle groups will work together to complete various movements. The muscle synergistic activation parameter can evaluate whether the cooperation between muscle groups is coordinated. In actual operation, it is first necessary to predefine the action direction of each muscle group, and then calculate the cosine value of the angle between the motion direction vector and the preset muscle group action direction. The cosine value is used as a correlation index, and the muscle synergistic activation parameter is calculated based on the correlation index. For example, for the movement of the knee joint, the direction of action of the quadriceps femoris is defined in advance, and the cosine value of the angle between the knee joint movement direction vector and the quadriceps femoris action direction vector is calculated. If the cosine value is large, it means that the quadriceps femoris plays a greater synergistic role in the knee joint movement, and the muscle synergistic activation parameter will be correspondingly higher.

[0065] Step S526: The local pressure distribution characteristics of the joints and the muscle co-activation parameters are subjected to biomechanical model fusion processing to generate actual joint load parameters. Biomechanical model fusion processing refers to the operation of combining the local pressure distribution characteristics of the joints and the muscle co-activation parameters with the biomechanical model for comprehensive analysis and calculation. The actual joint load parameter is a parameter obtained through biomechanical model fusion processing that can fully reflect the actual load situation of the joints. In physical fitness training, the load on the joints is not only related to the contact pressure, but also to the synergistic effect of the muscles. By fusing the local pressure distribution characteristics of the joints and the muscle co-activation parameters, the actual load of the joints can be more accurately assessed. In actual operation, a biomechanical model, such as a multi-rigid body dynamics model, can be used, with the local pressure distribution characteristics of the joints and the muscle co-activation parameters as inputs, and the actual joint load parameters can be obtained through model calculation. For example, in the multi-rigid body dynamics model, the local pressure distribution characteristics of the joints are converted into external force input, and the muscle co-activation parameters are converted into muscle force input. The actual load parameters of the joints are obtained by solving the dynamic equations of the model.

[0066] Step S530: Calculate the load deviation rate between the actual joint load parameters and the baseline physiological parameters, and perform risk level mapping processing on the load deviation rate to generate a joint injury risk index. The load deviation rate refers to the degree of difference between the actual joint load parameters and the baseline physiological parameters. By calculating the load deviation rate, it is possible to understand the deviation of the target user's joint load from the standard during training. Risk level mapping processing refers to mapping the load deviation rate to different risk levels in order to intuitively assess the risk level of joint injury. The joint injury risk index is an indicator for assessing the possibility of joint injury obtained based on the load deviation rate and risk level mapping processing.

[0067] Optionally, step S530 may specifically include the following steps S531 to S536: Step S531: Retrieve the baseline physiological parameters that match the current training action category from the preset action standard library, the baseline physiological parameters include the standard joint pressure distribution range and the standard muscle activation sequence. As mentioned above, the preset action standard library is a database that stores various training action baseline physiological parameters. Retrieve the baseline physiological parameters that match the current training action category from the database, which are crucial for evaluating the rationality of joint loads. The standard joint pressure distribution range specifies the normal pressure range of various parts of the joint under standard training actions, and the standard muscle activation sequence describes the activation order and intensity of the muscles in the standard action. For example, for the push-up action, the baseline physiological parameters retrieved from the preset action standard library may include the standard pressure distribution range of the shoulder joint and the standard activation sequence of muscles such as the pectoralis major and deltoid muscles.

[0068] Step S532: Perform joint-by-joint comparison processing on the actual joint load parameters and the baseline physiological parameters, and calculate the bidirectional deviation of each joint in the pressure distribution dimension and the muscle activation dimension. Joint-by-joint comparison processing refers to the operation of performing a detailed comparative analysis of the actual joint load parameters and the baseline physiological parameters of each joint. The bidirectional deviation refers to the degree of difference between the actual parameters and the baseline parameters in the pressure distribution dimension and the muscle activation dimension. In actual operation, for the pressure distribution dimension, the difference between the actual joint contact pressure distribution area and the standard joint pressure distribution range is calculated, such as the difference in pressure values, the overlap rate of the pressure distribution areas, etc.; for the muscle activation dimension, the actual muscle co-activation parameters are compared with the standard muscle activation sequence, and the difference in activation order and activation intensity are calculated. For example, for the knee joint, the average difference between the pressure value of each point in the actual contact pressure distribution area and the pressure value of the corresponding point in the standard pressure distribution range, as well as the difference between the actual muscle co-activation parameters and the corresponding muscle activation intensity in the standard muscle activation sequence are calculated to obtain the bidirectional deviation of the knee joint in the pressure distribution dimension and the muscle activation dimension.

[0069] Step S533: Perform weighted summation on the bidirectional deviations to generate the comprehensive load deviation rate of each joint. Weighted summation refers to the operation of assigning different weights to the bidirectional deviations of the pressure distribution dimension and the muscle activation dimension, and then summing them. The comprehensive load deviation rate is a parameter obtained through weighted summation that can comprehensively reflect the degree of deviation between the actual load of each joint and the baseline load. The determination of weights needs to consider the degree of influence of different dimensions on the risk of joint injury. For example, for some joints that are more sensitive to pressure, such as the knee joint, the weight of the pressure distribution dimension can be set higher; for some joints that mainly rely on muscle synergy, such as the shoulder joint, the weight of the muscle activation dimension can be set higher.

[0070] Step S534: Perform inter-joint distribution pattern analysis on the comprehensive load deviation rate, and identify continuously distributed joint areas whose deviation rates exceed a preset threshold. Inter-joint distribution pattern analysis refers to the operation of analyzing the distribution of the comprehensive load deviation rates of each joint between joints and finding joint areas with similar deviation characteristics. The preset threshold is a pre-set boundary for judging whether the joint load deviation is abnormal. The continuously distributed joint area refers to the area composed of adjacent joints whose comprehensive load deviation rates exceed the preset threshold. By identifying the continuously distributed joint area, it is possible to find compensatory movements or local overload situations. In actual operation, the comprehensive load deviation rates of each joint are arranged in order according to the position of the joints to form an inter-joint deviation rate distribution sequence, and then the sequence is traversed to find the continuous joint areas whose deviation rates exceed the preset threshold. For example, in a squat movement, if it is found that the comprehensive load deviation rates of the hip joint, knee joint and ankle joint all exceed the preset threshold, and they are adjacent joints, the area composed of these three joints is identified as a continuously distributed joint area.

[0071] Step S535: Perform kinematic chain relevance detection on the continuously distributed joint area to determine the main risk joint set that may cause compensatory movements. The kinematic chain relevance detection refers to analyzing the kinematic relevance between the joints in the continuously distributed joint area to find out the operations of the key joints that may cause compensatory movements during the movement. Compensatory movements refer to when a joint or muscle cannot function normally, other joints or muscles will excessively participate to complete the movement, and this kind of movement may increase the risk of joint injury. The main risk joint set refers to the set of key joints that may cause compensatory movements determined by the kinematic chain relevance detection. In actual operation, kinematic chain relevance detection can be performed using kinematic analysis methods and biomechanical models. For example, the movement direction, movement amplitude and force conditions of each joint in the continuously distributed joint area are analyzed to find out the joints that bear too much load or move in an uncoordinated manner during the movement, and these joints are determined as the main risk joint set.

[0072] Step S536: Perform damage probability mapping processing based on the comprehensive load deviation rate of each joint in the main risk joint set to generate a joint damage risk index. Damage probability mapping processing refers to the operation of mapping the comprehensive load deviation rate of each joint in the main risk joint set to the corresponding joint damage probability. The joint injury risk index is an indicator that can intuitively reflect the possibility of joint damage obtained through damage probability mapping processing. In actual operation, a mapping table of comprehensive load deviation rate and joint damage probability can be established, and the corresponding damage probability can be found from the mapping table based on the comprehensive load deviation rate of each joint in the main risk joint set. Then, these damage probabilities are weighted averaged or other comprehensive calculations are performed to obtain the joint injury risk index.

[0073] Step S540: Perform center of gravity trajectory analysis on the posture state recognition results, extract body balance parameters, and perform stability quantification on the body balance parameters to generate a dynamic balance coefficient. Center of gravity trajectory analysis refers to the operation of analyzing and processing the human body center of gravity position information in the posture state recognition results to understand the movement trajectory and changes of the center of gravity of the target user during training. The body balance parameter is a parameter obtained through center of gravity trajectory analysis that can reflect the body balance ability of the target user. Stability quantification refers to the operation of converting the body balance parameter into a specific numerical indicator to more intuitively evaluate the stability of body balance. The dynamic balance coefficient is a coefficient obtained through stability quantification for measuring the body balance stability of the target user during dynamic movement.

[0074] Optionally, in step S540, the posture state recognition result is subjected to center of gravity trajectory analysis and processing to extract body balance parameters, which may specifically include the following steps S541 to S549: Step S541: The posture state recognition result is subjected to human center of mass coordinate extraction processing, and a sequence of three-dimensional coordinates of the center of mass of consecutive frames is generated based on the three-dimensional spatial coordinates of each joint and the preset human body mass distribution parameters. The human center of mass coordinate extraction processing refers to the operation of extracting the specific position information of the human body center of mass from the posture state recognition result. The preset human body mass distribution parameters are the mass ratios of each joint or body part pre-set according to the physiological structure and mass distribution law of the human body. The sequence of three-dimensional coordinates of the center of mass of consecutive frames refers to a sequence composed of the three-dimensional spatial coordinates of the human body center of mass corresponding to each frame in a multi-frame continuous image sequence. In actual operation, the three-dimensional spatial coordinates of the human body center of mass are calculated using a weighted average method based on the three-dimensional spatial coordinates of each joint and the preset human body mass distribution parameters. For example, different parts of the human body, such as the upper limbs, lower limbs, and torso, are assigned different mass weights. The 3D coordinates of each joint are multiplied by the corresponding mass weight, added together, and then divided by the total mass to obtain the 3D coordinates of the body's center of mass. This calculation process is repeated for each frame of the image, generating a sequence of 3D coordinates of the center of mass for consecutive frames.

[0075] Step S542: Perform horizontal support surface projection on the centroid three-dimensional coordinate sequence, mapping the centroid three-dimensional coordinate sequence to a two-dimensional plane coordinate system to generate a two-dimensional projection point sequence of the centroid. Horizontal support surface projection refers to the process of projecting the three-dimensional coordinates in the centroid three-dimensional coordinate sequence onto the horizontal support surface, converting the three-dimensional information into two-dimensional information. The two-dimensional plane coordinate system is a plane coordinate system on the horizontal support surface, and the two-dimensional projection point sequence of the centroid three-dimensional coordinate sequence is the sequence of projection points of the centroid three-dimensional coordinate sequence onto the two-dimensional plane coordinate system. In actual operation, the vertical coordinate (z coordinate) in the centroid three-dimensional coordinate sequence is ignored, and only the horizontal coordinates (x and y coordinates) are retained. These coordinates are used as the projection points of the centroid on the two-dimensional plane coordinate system to generate the two-dimensional projection point sequence of the centroid. For example, for each point (x, y, z) in the centroid three-dimensional coordinate sequence, it is projected onto the horizontal support surface to obtain a two-dimensional projection point (x, y). All projection points are arranged in order to generate the two-dimensional projection point sequence of the centroid.

[0076] Step S543: Perform displacement vector analysis on the two-dimensional projection point sequence of the centroid, calculate the displacement direction and displacement amount between adjacent projection points, and generate a set of centroid movement trajectory vectors. Displacement vector analysis refers to the operation of analyzing and calculating the position changes between adjacent projection points in the two-dimensional projection point sequence of the centroid. The displacement direction refers to the movement direction between adjacent projection points, and the displacement amount refers to the distance between adjacent projection points. The set of centroid movement trajectory vectors refers to a set consisting of displacement vectors between adjacent projection points. In actual operation, for two adjacent projection points (x1, y1) and (x2, y2) in the two-dimensional projection point sequence of the centroid, the coordinate difference (x2-x1, y2-y1) is calculated, the difference vector is normalized to obtain the displacement direction vector, and the modulus of the difference vector is calculated as the displacement amount. The displacement direction vectors and displacement amounts between all adjacent projection points are combined into a set of centroid movement trajectory vectors. For example, for two adjacent two-dimensional projection points of the centroid (3, 5) and (6, 8), the displacement direction vector is calculated as ((6-3) / √((6-3) 2 +(8-5) 2 ), (8-5) / √((6-3) 2 +(8-5) 2 ))=(0.707, 0.707), the displacement is √((6-3) 2 +(8-5) 2 )=4.24, and add this information to the center of mass moving trajectory vector set.

[0077] Step S544: Perform mutation point detection on the set of centroid movement trajectory vectors to identify abnormal trajectory segments where the displacement direction and displacement amount deviate from a preset smoothing threshold. Mutation point detection refers to the operation of analyzing and detecting the displacement direction and displacement amount in the set of centroid movement trajectory vectors to find points where sudden and significant changes occur. The preset smoothing threshold is a pre-set boundary for determining whether the changes in displacement direction and displacement amount are abnormal. An abnormal trajectory segment refers to a continuous portion of the set of centroid movement trajectory vectors where the displacement direction and displacement amount deviate from the preset smoothing threshold. In actual operation, for each displacement vector in the set of centroid movement trajectory vectors, calculate the direction angle and displacement difference between it and the adjacent displacement vectors. When the angle and difference exceed the preset smoothing threshold, the point is marked as a mutation point. Then, find the continuous portion between adjacent mutation points and use it as an abnormal trajectory segment. For example, the preset smoothing threshold is that the displacement direction angle is greater than 30° and the displacement difference is greater than 1 cm. When the direction angle between two adjacent displacement vectors in the center of mass moving trajectory vector set is 40° and the displacement difference is 1.5 cm, the point is marked as a mutation point, and the adjacent mutation points are continued to be searched to determine the abnormal trajectory segment.

[0078] Step S545: Perform action phase matching processing on the abnormal trajectory segment to determine the training action execution stage corresponding to the abnormal trajectory segment. Action phase matching processing refers to the operation of associating and matching the abnormal trajectory segment with the target user's training action execution process. The training action execution stage refers to the different stages of the training action from the beginning to the end, such as the squatting stage and the standing up stage of the squat action. Through the action phase matching processing, it is possible to understand in which training action execution stage the abnormal trajectory segment appears, which is helpful for analyzing the cause of the abnormality. In actual operation, the position corresponding to the abnormal trajectory segment in the standardized action cycle sequence is found in combination with the time information of the abnormal trajectory segment generated previously, and the corresponding training action execution stage is determined. For example, by analyzing the timestamp of the abnormal trajectory segment, it is found that it is in the standing up stage of the squat action, thereby determining that the training action execution stage corresponding to the abnormal trajectory segment is the standing up stage.

[0079] Step S546: Perform spatial distribution clustering on the two-dimensional projection point sequence of the centroid to generate a centroid projection density distribution map. Boundary contour extraction is then performed on the centroid projection density distribution map to construct a polygonal boundary of the centroid activity area. Spatial distribution clustering refers to the process of grouping and clustering points in the two-dimensional projection point sequence of the centroid according to their spatial location. The centroid projection density distribution map is a graph obtained through spatial distribution clustering that reflects the density distribution of the centroid on a two-dimensional plane. Boundary contour extraction refers to the process of extracting the boundary contour from the centroid projection density distribution map. The polygonal boundary of the centroid activity area is a polygon obtained through boundary contour extraction that describes the activity range of the centroid on a two-dimensional plane. In practice, the DBSCAN (density-based spatial clustering application) algorithm can be used to perform spatial distribution clustering on the two-dimensional projection point sequence of the centroid. Points with similar densities are grouped into the same cluster, and the density value of each cluster is calculated to generate a centroid projection density distribution map. Then, an edge detection algorithm, such as the Canny edge detection algorithm, is used to process the centroid projection density distribution map, extract its boundary contour, fit the boundary contour into a polygon, and construct the polygonal boundary of the centroid activity area.

[0080] Step S547: Perform geometric morphological analysis on the polygonal boundary of the centroid activity area to extract the polygonal coverage area and boundary curvature parameters. Geometric morphological analysis refers to the operation of analyzing and calculating the geometric shape and characteristics of the polygonal boundary of the centroid activity area. The polygonal coverage area refers to the area enclosed by the polygonal boundary of the centroid activity area, which reflects the size of the centroid's activity range on the two-dimensional plane. The boundary curvature parameter refers to the degree of curvature of the polygonal boundary, which can reflect the smoothness of the centroid's movement trajectory. In actual operation, a polygonal area calculation algorithm, such as the shoelace formula, is used to calculate the coverage area of ​​the polygonal boundary of the centroid activity area; for the boundary curvature parameter, the curvature of each vertex on the polygonal boundary is calculated, and the average value is taken as the boundary curvature parameter.

[0081] Step S548: Perform vertical axis fluctuation analysis on the three-dimensional center of mass coordinate sequence to generate a center of mass vertical height change rate sequence. Oscillation pattern recognition is then performed on the center of mass vertical height change rate sequence to extract the alternating vertical acceleration variation feature. Vertical axis fluctuation analysis involves analyzing and calculating the vertical coordinate (z-coordinate) in the three-dimensional center of mass coordinate sequence to obtain the center of mass vertical height change rate. A center of mass vertical height change rate sequence is a sequence of the center of mass vertical height change rates corresponding to each frame in a multi-frame continuous image sequence. Oscillation pattern recognition involves analyzing and identifying the center of mass vertical height change rate sequence to identify the oscillation pattern within the sequence. The alternating vertical acceleration variation feature is a feature obtained through oscillation pattern recognition that reflects the alternating vertical acceleration variation of the center of mass. In actual operation, the difference between the vertical coordinates of two adjacent frames in the three-dimensional center of mass coordinate sequence is calculated and divided by the time interval to obtain the center of mass vertical height change rate. All of these change rates are then arranged in sequence to generate the center of mass vertical height change rate sequence. Then, using spectrum analysis methods such as fast Fourier transform (FFT), the center of mass vertical height change rate sequence is analyzed to identify oscillation patterns and extract the characteristics of alternating vertical acceleration changes. For example, FFT analysis of the center of mass vertical height change rate sequence reveals an oscillation pattern with a frequency of 2 Hz. This frequency and oscillation amplitude are used as the characteristics of alternating vertical acceleration changes.

[0082] Step S549: Multidimensional feature fusion processing is performed on the polygon coverage area, boundary curvature parameters, and alternating vertical acceleration characteristics to generate a body balance parameter. Multidimensional feature fusion processing integrates and processes the polygon coverage area, boundary curvature parameters, and alternating vertical acceleration characteristics to generate a parameter that comprehensively reflects the target user's body balance. The body balance parameter is a comprehensive indicator obtained through multidimensional feature fusion processing. In practice, these three features can be fused using methods such as weighted summation.

[0083] Optionally, in step S540, stability quantification processing is performed on the body balance parameters to generate a dynamic balance coefficient. This may specifically include steps S5410 to S5415: Step S5410: Time-series segmentation is performed on the body balance parameters, dividing the analysis time intervals based on the training movement execution phase. Time-series segmentation refers to the operation of dividing and segmenting the body balance parameters in chronological order. Analysis time intervals refer to different time periods divided according to the training movement execution phase. In physical fitness training, different training movement execution phases may have different requirements and impacts on body balance. Time-series segmentation allows for a more detailed analysis of changes in body balance during different phases. In actual operation, based on the previously determined training movement execution phase and the time series of the body balance parameters, the body balance parameters are segmented by training movement execution phase to obtain body balance parameters within different analysis time intervals. For example, for a squat, the body balance parameters are segmented into the squat phase, the rise phase, and the stationary phase to obtain body balance parameters within the three analysis time intervals.

[0084] Step S5411: Perform stability measurement calculation processing on the body balance parameters within each analysis time interval to generate an interval stability score. Stability measurement calculation processing refers to the operation of analyzing and calculating the body balance parameters within each analysis time interval to obtain a score that can measure the body balance stability within that interval. The interval stability score is a numerical value obtained through the stability measurement calculation processing and is used to evaluate the body balance stability within each analysis time interval. In actual operation, the standard deviation of the body balance parameters within each analysis time interval can be calculated, and the inverse of the standard deviation can be used as the interval stability score. The smaller the standard deviation, the smaller the fluctuation of the body balance parameters within the interval, the higher the stability, and the higher the interval stability score.

[0085] Step S5412: Perform trend correlation analysis on the interval stability scores of adjacent analysis time intervals to identify the time nodes where the scores decrease. Trend correlation analysis refers to the operation of comparing and analyzing the interval stability scores of adjacent analysis time intervals to find the trend of score changes. The time node where the score decreases refers to the moment when the interval stability score of the adjacent analysis time interval decreases. In physical fitness training, a decrease in score may mean that the target user's body balance stability begins to deteriorate and requires timely attention. In actual operation, the interval stability scores of adjacent analysis time intervals are compared. If the score of the latter interval is lower than the score of the previous interval, the time node is marked as the time node where the score decreases.

[0086] Step S5413: Perform joint motion trajectory backtracking processing on the time node where the score drops, and match the corresponding joint angular velocity change data. Joint motion trajectory backtracking processing refers to the operation of backtracking and checking the joint motion trajectory information of the target user at that moment according to the time node where the score drops. Joint angular velocity change data refers to the change of the angular velocity of the joint over time during the movement process. Through joint motion trajectory backtracking processing and matching the corresponding joint angular velocity change data, the relationship between the score drop and the joint movement can be analyzed. In actual operation, according to the time node where the score drops, the three-dimensional spatial coordinates of each joint at that moment are extracted from the previously generated spatial key point coordinate set, and the change of coordinates between adjacent frames is calculated to obtain the motion trajectory of the joint. At the same time, the angular velocity change data of the joint is calculated in combination with the movement time. For example, for the time node where the score drops, the three-dimensional spatial coordinates of the shoulder joint key points in two adjacent frames are compared, the coordinate difference and the time interval are calculated, and the angular velocity change data of the shoulder joint at that moment are obtained.

[0087] Step S5414: Perform motion coupling analysis on the joint angular velocity change data and the center of mass movement trajectory vector set to determine the influence weight of the joint movement on the center of mass offset. Motion coupling analysis refers to the operation of analyzing the relationship and coupling degree between the joint movement and the center of mass movement. The influence weight of the joint movement on the center of mass offset refers to a parameter obtained through motion coupling analysis for measuring the degree of influence of the joint movement on the center of mass offset. In physical fitness training, the movement of the joint may cause the center of mass to shift. By determining the influence weight, the influence of different joint movements on the center of mass balance can be understood. In actual operation, correlation analysis methods, such as the Pearson correlation coefficient, can be used to calculate the correlation between the joint angular velocity change data and the displacement direction and displacement amount in the center of mass movement trajectory vector set, and the correlation coefficient can be used as the influence weight of the joint movement on the center of mass offset.

[0088] Step S5415: Balance decay coefficient modeling is performed based on the influence weights and interval stability scores to generate a dynamic balance coefficient. Balance decay coefficient modeling involves building a model to calculate the balance decay coefficient based on the influence weights of joint motion on center of mass shift and the interval stability scores. The dynamic balance coefficient, derived through balance decay coefficient modeling, is a comprehensive indicator used to measure the target user's body balance stability during dynamic motion. In practice, a balance decay coefficient model can be constructed using methods such as weighted product. For example, the influence weight of each joint motion on center of mass shift is multiplied by the corresponding interval stability score, and all results are added together to obtain the balance decay coefficient. The dynamic balance coefficient is then subtracted from the balance decay coefficient from 1. For example, given three joints with influence weights of 0.3, 0.4, and 0.3, respectively, and corresponding interval stability scores of 2, 3, and 2, the balance decay coefficient = 0.3 × 2 + 0.4 × 3 + 0.3 × 2 = 2.4, and the dynamic balance coefficient = 1 - 2.4 = -1.4 (this is just an example; in practice, the calculated results may require normalization or other processing).

[0089] Step S550: The joint injury risk index and the dynamic balance coefficient are weighted and fused to generate a fitness assessment parameter. Weighted fusion involves assigning different weights to the joint injury risk index and the dynamic balance coefficient, and then performing a weighted summation. The fitness assessment parameter, derived through this weighted fusion process, comprehensively reflects the target user's fitness status. In practice, the weightings must be determined based on the importance of joint injury risk and body balance to fitness.

[0090] Optionally, after generating the fitness assessment parameters in step S550, the following steps may be performed: Step S560: Performing historical trend comparison processing on the fitness assessment parameters to obtain a historical assessment parameter sequence for the same exercise. Historical trend comparison processing refers to comparing and analyzing the currently generated fitness assessment parameters with the assessment parameters of the target user when they previously performed the same exercise. A historical assessment parameter sequence for the same exercise refers to a sequence of fitness assessment parameters generated when the target user previously performed the same exercise. Through historical trend comparison processing, changes in the target user's fitness for the exercise can be understood. In actual operation, historical assessment parameters for the target user performing the same exercise can be retrieved from a database storing historical data, and these parameters can be arranged in chronological order to obtain a historical assessment parameter sequence. For example, if the target user performs squats multiple times, the fitness assessment parameters for each squat can be retrieved from the database to form a historical assessment parameter sequence.

[0091] Step S570: Calculate the rate of change of the historical evaluation parameter sequence to generate a fitness improvement trend curve. Calculating the rate of change involves performing difference and ratio calculations on adjacent parameters in the historical evaluation parameter sequence to determine the rate of change of the fitness evaluation parameter. A fitness improvement trend curve is a curve formed by connecting the rates of change in chronological order. It intuitively reflects the target user's fitness improvement trend for the exercise. In practice, for two adjacent parameters in the historical evaluation parameter sequence, their difference and ratio are calculated, and the ratio is used as the rate of change. For example, if the historical evaluation parameter sequence is [0.4, 0.5, 0.6], the rates of change of the adjacent parameters are calculated to be (0.5-0.4) / 0.4 = 0.25 and (0.6-0.5) / 0.5 = 0.2, respectively. These rates of change are then connected in chronological order to generate a fitness improvement trend curve.

[0092] Step S580: Perform inflection point detection on the fitness progress trend curve to identify training effect plateaus. Inflection point detection involves analyzing and detecting points within the fitness progress trend curve where the slope changes significantly. A training effect plateau refers to a period of time in the fitness progress trend curve where the rate of change approaches zero or changes very slowly. During this period, the target user's fitness improvement is not significantly improved. In practice, a derivative analysis method can be used to calculate the first and second derivatives of the fitness progress trend curve. When the second derivative is zero and the first derivative changes significantly, the point is marked as an inflection point. Next, identify time periods where the rate of change between adjacent inflection points is small and identify these periods as training effect plateaus. For example, calculate the first and second derivatives of the fitness progress trend curve. When the second derivative is zero and the first derivative changes from positive to near zero, mark this point as an inflection point. Continue searching for time periods where the rate of change between adjacent inflection points is less than 0.05 and identify these periods as training effect plateaus.

[0093] Step S590: When a plateau period of training effect is detected, the spatiotemporal posture feature set is subjected to motion pattern optimization analysis and processing. Motion pattern optimization analysis and processing refers to an operation of conducting in-depth analysis and research on the spatiotemporal posture feature set, identifying possible unreasonable motion patterns, and proposing optimization suggestions. When a plateau period of training effect is detected, it indicates that there may be problems with the target user's current training motion pattern and that it needs to be optimized. In actual operation, a machine learning algorithm, such as a decision tree algorithm, can be used to classify and analyze the spatiotemporal posture feature set, identify key features related to the training effect, and then propose motion pattern optimization suggestions based on these features. For example, by using a decision tree algorithm to analyze the spatiotemporal posture feature set, it is found that the movement angle of a certain joint is closely related to the training effect, and a motion pattern optimization suggestion for adjusting the movement angle of the joint is proposed.

[0094] Step S5100: Identify compensatory action pattern features through action pattern optimization analysis and processing. Compensatory action pattern features refer to the features of alternative action patterns adopted when the target user is unable to complete a certain action normally during training. Through action pattern optimization analysis and processing, these compensatory action pattern features are identified so that targeted corrections can be made. In actual operation, the feature differences between the normal action pattern and the possible compensatory action pattern in the spatiotemporal posture feature set are compared to find representative features as compensatory action pattern features. For example, in a squat movement, the knee and hip joints are bent at the same time in the normal action mode, while the compensatory action mode may be manifested as excessive bending of the knee joint and insufficient bending of the hip joint. The abnormal difference in the bending angles of the knee and hip joints is used as a compensatory action pattern feature.

[0095] Step S5110: Perform root cause tracing processing on the compensatory action pattern characteristics to determine the muscle strength imbalance area. Root cause tracing processing refers to the operation of analyzing the causes of the compensatory action pattern characteristics and finding the fundamental factors that lead to this action pattern. The muscle strength imbalance area refers to the part of the body where the compensatory action pattern appears due to muscle strength imbalance. In actual operation, the relationship between the compensatory action pattern characteristics and muscle strength can be analyzed in combination with the biomechanical model and muscle activation data to find the area of ​​muscle strength imbalance. For example, by analyzing the compensatory action pattern characteristics in the squat action, it was found that this was caused by the excessive strength of the quadriceps femoris and the insufficient strength of the gluteus maximus. The area where the quadriceps femoris and gluteus maximus are located is determined as the muscle strength imbalance area.

[0096] Step S5120: Generate training program adjustment suggestion parameters based on the muscle strength imbalance area. The training program adjustment suggestion parameters refer to the specific parameters and suggestions for adjusting the training program based on the situation of the muscle strength imbalance area. In actual operation, for the muscle strength imbalance area, suggestions such as increasing or decreasing the training intensity of certain muscles, adjusting the difficulty and frequency of training movements, etc. can be made. For example, for the situation of imbalance in the strength of the quadriceps and gluteus maximus, suggestions are made to increase the intensity of gluteus maximus training, such as adding gluteus maximus training movements such as hip bridges, and at the same time appropriately reducing the participation of the quadriceps in squats, and adjusting the training program adjustment suggestion parameters.

[0097] Step S5130: Integrate the training program adjustment suggestion parameters into the physical fitness assessment parameters. The integration operation refers to the operation of merging and unifying the training program adjustment suggestion parameters and the physical fitness assessment parameters. By integrating the training program adjustment suggestion parameters into the physical fitness assessment parameters, the physical fitness assessment parameters can contain more comprehensive information and provide a more accurate reference for the physical fitness training guidance system. In actual operation, the training program adjustment suggestion parameters can be used as an additional part of the physical fitness assessment parameters, or they can be weighted and fused with the original physical fitness assessment parameters. For example, the training program adjustment suggestion parameters are added to the physical fitness assessment parameters in text form, or the training intensity adjustment coefficient in the suggestion parameters is weightedly summed with the original physical fitness assessment parameters to obtain the integrated physical fitness assessment parameters.

[0098] Optionally, after step S500, the following steps S600 to S1300 may also be included: Step S600: Receive real-time training feedback instructions through the physical fitness training guidance system. Real-time training feedback instructions refer to instructions for guiding training issued by the physical fitness training guidance system based on the physical fitness assessment parameters and the real-time training situation of the target user. These instructions may include suggestions for adjusting the intensity, frequency, posture, etc. of training movements. In actual operation, the physical fitness training guidance system will analyze and process the physical fitness assessment parameters, and generate real-time training feedback instructions in combination with the real-time training data of the target user. For example, when the physical fitness assessment parameters show that the target user has a higher risk of joint injury, the physical fitness training guidance system will issue a real-time training feedback instruction to reduce the training intensity.

[0099] Step S700: Activate the posture correction guidance mode according to the real-time training feedback instruction. The posture correction guidance mode is a mode used to help the target user correct the posture of the training movement. When receiving the real-time training feedback instruction, the system will activate the posture correction guidance mode to provide the target user with guidance on posture correction. In actual operation, the system will determine the posture part that needs to be corrected and the correction target based on the specific content in the real-time training feedback instruction. For example, when the real-time training feedback instruction indicates that the target user's knee joint moves forward excessively during the squat movement, the system activates the posture correction guidance mode and uses the correct position and movement trajectory of the knee joint as the correction target.

[0100] Step S800: In the posture correction guidance mode, target deviation analysis processing is performed on the currently generated spatiotemporal posture feature set. Target deviation analysis processing refers to the operation of analyzing the differences and deviations between the currently generated spatiotemporal posture feature set and the correction target. Through target deviation analysis processing, the specific content and direction that the target user needs to correct in the training movement can be clarified. In actual operation, the joint position, angle and other features in the spatiotemporal posture feature set are compared with the correction target, and the difference and error between them are calculated to obtain target deviation information. For example, in the squat movement posture correction guidance mode, the angle of the knee joint in the spatiotemporal posture feature set is compared with the knee joint angle in the correction target, and the angle difference is calculated as the target deviation information.

[0101] Step S900: Identify the main movement defect parts through target deviation analysis. The main movement defect parts refer to the body parts that deviate greatly from the correction target during the target deviation analysis. By identifying the main movement defect parts, the target user can be given targeted posture correction guidance. In actual operation, based on the target deviation information, find the joints or body parts whose deviation values ​​exceed the preset threshold and identify them as the main movement defect parts. For example, the preset threshold is 5°. In the squat movement, it is found that the knee joint angle deviation is 8°, which exceeds the preset threshold. The knee joint is identified as the main movement defect part.

[0102] Step S1000: Generate a visual correction guidance signal based on the main action defect site. The visual correction guidance signal is a signal presented in a visual manner to guide the target user to correct the posture. It can be in the form of graphics, text, arrows, etc., which intuitively tell the target user how to adjust the action posture. In actual operation, the corresponding visual correction guidance signal is generated according to the main action defect site and the target deviation information. For example, in the case where the main action defect site is the knee joint, an arrow is drawn on the training scene display interface to indicate the direction in which the knee joint should move, and the angle value that needs to be adjusted is displayed at the same time as a visual correction guidance signal.

[0103] Step S1100: Superimpose the visual correction guidance signal on the training scene display interface in real time. The training scene display interface is the interface that the target user sees when performing physical fitness training. Superimposing the visual correction guidance signal on this interface in real time allows the target user to see the correction guidance information in time during the training process. In actual operation, graphics processing technology is used to superimpose the visual correction guidance signal with the real-time image of the training scene to ensure that the guidance signal can be accurately displayed on the corresponding body part. For example, a graphics library such as OpenGL is used to superimpose the arrows and text in the visual correction guidance signal on the position of the target user's knee joint in the training scene display interface.

[0104] Step S1200: Establish an instant mapping relationship between action execution and guidance signals in the training scene display interface. The instant mapping relationship refers to the real-time association and correspondence between the actual action execution of the target user and the visual correction guidance signal, so that the target user can adjust the action in time according to the guidance signal. In actual operation, the system will monitor the action posture of the target user in real time, compare the features in the spatiotemporal posture feature set with the visual correction guidance signal, and update the display of the guidance signal accordingly when the action posture of the target user changes. For example, when the target user gradually adjusts the position of the knee joint during a squat, the system will compare the actual position of the knee joint with the target position in the guidance signal in real time, update the display of the guidance signal according to the comparison result, and maintain the instant mapping relationship between the action execution and the guidance signal.

[0105] Step S1300: When it is detected that the degree of overlap between the motion trajectory of the defective part and the visual correction guidance signal reaches a preset standard, the posture correction guidance mode is turned off. The overlap refers to the degree of similarity between the actual motion trajectory of the defective part and the motion trajectory indicated by the visual correction guidance signal. The preset standard is a pre-set limit for judging whether the posture correction is successful. When the overlap reaches the preset standard, it means that the target user's posture has basically met the correction target. At this time, the posture correction guidance mode is turned off and the normal training mode is restored. In actual operation, the system will calculate the overlap between the motion trajectory of the defective part and the visual correction guidance signal in real time. When the overlap exceeds the preset standard, such as reaching 90%, the posture correction guidance mode is turned off. For example, during the squat posture correction process, the system monitors the overlap between the motion trajectory of the knee joint and the target motion trajectory of the knee joint in the visual correction guidance signal in real time. When the overlap reaches 90%, the posture correction guidance mode is turned off.

[0106] It is understandable that the various algorithms involved in the above content of the embodiment of the present invention, such as the Euclidean distance algorithm, the cosine distance algorithm, the edge detection algorithm, the clustering algorithm, etc., can be learned from the relevant content in the prior art. In order to save space, they will not be expanded too much in the embodiment of the present invention. In addition, when implementing the scheme of the present invention, those skilled in the art can supplement the details according to the common knowledge in this field. For example, according to the common knowledge in this field, normalization can be used to eliminate the dimension conflict before feature fusion, interpolation can be used to eliminate the dimension difference, and the threshold value can be reasonably set based on historical data, experience or business scenario requirements. The model can be trained based on a general model training method, and the number of layers in the model structure can be set based on actual needs, the activation function can be selected, etc. The present invention will no longer provide redundant introductions to the overly detailed implementation process.

[0107] See also Figure 2 , Figure 2This is a schematic diagram of the structure of a posture detection device provided by an embodiment of the present invention. In practical applications, such as an all-in-one fitness machine, the posture detection device includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 may be connected via a bus or other means. The processor 101 (also known as the Central Processing Unit (CPU)) is the computing and control core of the posture detection device, capable of parsing various instructions within the posture detection device and processing various data from the posture detection device. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi or a mobile communication interface), which can be used to send and receive data under the control of the processor 101. The communication interface 102 may also be used for data transmission and interaction within the posture detection device. The memory 103 is a storage device within the posture detection device, used to store programs and data. It is understood that the memory 103 herein may include both the built-in memory of the posture detection device and, of course, the extended memory supported by the posture detection device. Memory 103 provides storage space, and this storage space has stored the operating system of gesture detection device, and may include but not limited to: Android system, iOS system, Windows Phone system etc., and the present invention does not limit this.In one embodiment, processor 101 performs the human body posture detection method based on machine vision provided above in the embodiment of the present invention by the computer program in running memory 103.In addition, this gesture detection device can also include camera, or be connected by communication interface 102 with external camera, to gather the image of user.

Claims

1. A human posture detection method based on machine vision, applied to posture recognition in physical fitness testing, characterized in that: include: Acquire a multi-frame continuous image sequence of a physical fitness training scene; Performing human body region positioning processing on the multi-frame continuous image sequence to generate a dynamic human body region set of the target user; Performing posture feature extraction processing on the dynamic human body region set to obtain a spatiotemporal posture feature set of the target user; Based on a preset physical fitness assessment model, performing posture state recognition processing on the spatiotemporal posture feature set to generate a posture state recognition result of the target user; Performing action type classification processing on the posture state recognition result to determine the current training action category, and matching the baseline physiological parameters in the preset action standard library according to the current training action category; Performing joint load distribution analysis on the posture state recognition result to generate actual joint load parameters; Calculating a load deviation rate between the actual joint load parameter and a baseline physiological parameter, and performing risk level mapping processing on the load deviation rate to generate a joint injury risk index; Performing body center of mass coordinate extraction processing on the posture state recognition result, and generating a three-dimensional coordinate sequence of the center of mass of consecutive frames based on the three-dimensional spatial coordinates of each joint and preset body mass distribution parameters; Performing horizontal support surface projection processing on the three-dimensional coordinate sequence of the center of mass, mapping the three-dimensional coordinate sequence of the center of mass to a two-dimensional plane coordinate system, and generating a two-dimensional projection point sequence of the center of mass; Performing displacement vector analysis on the two-dimensional projection point sequence of the center of mass, calculating the displacement direction and displacement between adjacent projection points, and generating a center of mass movement trajectory vector set; Performing mutation point detection processing on the mass center movement trajectory vector set to identify abnormal trajectory segments whose displacement direction and displacement amount deviate from a preset smoothing threshold; Performing action phase matching processing on the abnormal trajectory segment to determine the training action execution phase corresponding to the abnormal trajectory segment; Performing spatial distribution clustering processing on the two-dimensional projection point sequence of the centroid to generate a centroid projection density distribution map, and performing boundary contour extraction processing on the centroid projection density distribution map to construct a polygonal boundary of the centroid activity area; Performing geometric morphological analysis on the polygonal boundary of the centroid activity area to extract polygonal coverage area and boundary curvature parameters; Performing vertical axis fluctuation analysis on the three-dimensional coordinate sequence of the center of mass to generate a vertical height change rate sequence of the center of mass, performing oscillation pattern recognition on the vertical height change rate sequence of the center of mass to extract alternating change characteristics of vertical acceleration; Performing multi-dimensional feature fusion processing on the polygon coverage area, boundary curvature parameters, and alternating change characteristics of vertical acceleration to generate body balance parameters, and performing time series segmentation processing on the body balance parameters to divide analysis time intervals based on the execution stages of the training action; Perform stability measurement calculation on the body balance parameters within each analysis time interval to generate an interval stability score; Perform trend correlation analysis on the interval stability scores of adjacent analysis time intervals to identify the time nodes when the scores decrease; Perform joint motion trajectory backtracking processing on the time node where the score decreases, and match the corresponding joint angular velocity change data; Performing motion coupling analysis on the joint angular velocity change data and the center of mass movement trajectory vector set to determine the influence weight of the joint motion on the center of mass offset; Performing balance attenuation coefficient modeling based on the impact weight and interval stability score to generate a dynamic balance coefficient; The joint injury risk index and the dynamic balance coefficient are weightedly fused to generate physical fitness evaluation parameters, and the physical fitness evaluation parameters are fed back to the physical fitness training guidance system.

2. The method according to claim 1, characterized in that The performing human body region positioning processing on the multi-frame continuous image sequence to generate a dynamic human body region set of the target user includes: Performing background separation processing on the multi-frame continuous image sequence to obtain a foreground moving object set; Performing biometric matching processing on the foreground moving object set to determine the target moving object outline of the target user; Performing bounding box fitting processing on the target moving object contour to generate an initial human body region bounding box set; Performing a temporal continuity check on the initial human body region bounding box set to detect a region displacement change between adjacent frames; If the region displacement change is less than a preset displacement threshold, performing spatial calibration on the initial human region bounding box set to generate a calibrated human region bounding box set; Performing key frame sampling processing on the calibrated human body region bounding box set, and selecting a sampling frame set that meets the posture change amplitude condition; Performing multi-scale region expansion processing on each sampling frame in the sampling frame set to generate an expanded human body region set including a complete limb extension range; The extended human body region set is subjected to spatiotemporal alignment processing with the calibrated human body region bounding box set of the unsampled frame to generate a dynamic human body region set.

3. The method according to claim 1, characterized in that The performing posture feature extraction processing on the dynamic human body region set to obtain the spatiotemporal posture feature set of the target user includes: Performing joint key point detection processing on each dynamic human body region in the dynamic human body region set to generate a basic key point coordinate set; Performing three-dimensional space mapping processing on the basic key point coordinate set, converting the two-dimensional image coordinates into three-dimensional space coordinates, and generating a space key point coordinate set; Performing motion trajectory modeling on the spatial key point coordinate set to construct a key point motion trajectory curve; Perform differential feature calculation processing on the key point motion trajectory curve to extract key point motion speed features and motion acceleration features; Performing time dimension fusion processing on the key point motion velocity feature and motion acceleration feature to generate a short-term motion feature vector; Calculating the relative position relationship of the spatial key point coordinate set to generate adjacent joint distance features and joint angle features; Performing spatial topological coding processing on the adjacent joint distance features and the joint angle features to generate a static posture feature vector; The short-term motion feature vector and the static posture feature vector are subjected to spatiotemporal feature fusion processing to generate the spatiotemporal posture feature set.

4. The method according to claim 1, wherein The step of performing posture state recognition processing on the spatiotemporal posture feature set based on a preset physical fitness assessment model to generate a posture state recognition result of the target user includes: Inputting the spatiotemporal posture feature set into the posture feature encoding module of the physical fitness assessment model, performing high-dimensional feature mapping processing, and generating an abstract posture feature space; Performing classification boundary division processing in the abstract posture feature space to determine the standard action feature cluster center; Calculating the feature deviation between the abstract posture feature space and the standard action feature cluster center; Performing dynamic threshold comparison processing on the feature deviation, and generating a motion deformation indicator when the feature deviation exceeds a first deviation threshold; Performing energy consumption pattern recognition processing in the abstract posture feature space to extract joint motion energy distribution features; performing pattern matching processing on the joint motion energy distribution characteristics to identify abnormal energy consumption areas; Performing correlation analysis on the motion deformation marker and the energy consumption abnormality area to generate a posture stability evaluation index; The posture stability evaluation index is compared with a preset physical fitness standard interval to generate a posture state recognition result.

5. The method according to claim 4, characterized in that The physical fitness assessment model includes a movement standardization analysis module and a fatigue state prediction module; after generating the posture state recognition result, the method further includes: Performing action cycle decomposition processing on the spatiotemporal posture feature set by the action standardization analysis module to identify repeated action units; performing phase alignment processing on the repetitive action units to generate a standardized action cycle sequence; Performing amplitude fluctuation analysis on the standardized action cycle sequence to calculate action execution consistency parameters; Performing muscle activation pattern extraction processing on the spatiotemporal posture feature set by the fatigue state prediction module to generate muscle synergy features; Performing time domain attenuation analysis on the muscle synergy characteristics to identify muscle fatigue trend parameters; Performing multimodal fusion processing on the movement execution consistency parameter and the muscle fatigue trend parameter to generate a comprehensive fatigue state coefficient; The dynamic evaluation weight in the posture state recognition result is adjusted according to the comprehensive fatigue state coefficient.

6. The method according to claim 1, characterized in that After generating the physical fitness assessment parameters, the method further includes: Performing historical trend comparison processing on the physical fitness assessment parameters to obtain a historical assessment parameter sequence for the same action; Calculating the change rate of the historical evaluation parameter sequence to generate a fitness improvement trend curve; Performing inflection point detection on the fitness improvement trend curve to identify a plateau period of training effect; When a plateau of training effect is detected, performing motion pattern optimization analysis on the spatiotemporal posture feature set; Identifying compensatory movement pattern characteristics through the movement pattern optimization analysis process; Conduct root cause tracing of the compensatory movement pattern characteristics to identify areas of muscle strength imbalance; generating a training program to adjust recommended parameters based on the muscle strength imbalance area; The training program adjustment suggestion parameters are integrated into the physical fitness assessment parameters.

7. The method according to claim 1, characterized in that Before acquiring a multi-frame continuous image sequence of a physical fitness training scene, the method further includes: Configuring a multi-view visual acquisition device array, wherein the multi-view visual acquisition device array comprises at least three spatially distributed image acquisition devices; Performing spatiotemporal synchronization processing on the multi-view visual acquisition device array to establish a unified timestamp system; Performing frame rate calibration processing on each image acquisition device through the unified timestamp system; Calculate the overlapping area of ​​the field of view of each calibrated image acquisition device to determine the effective acquisition area; Setting up a physical fitness training motion capture space within the effective acquisition area; Performing light adaptive adjustment processing on the physical fitness training motion capture space to eliminate ambient light interference factors; When the target user enters the physical fitness training motion capture space, the multi-view visual acquisition device array is triggered to synchronously capture images; When performing human body region positioning processing on the multi-frame continuous image sequence, the method further includes: Acquire a synchronous multi-view image set through the multi-view visual acquisition device array; Performing disparity calculation processing on the synchronized multi-view image set to generate a depth information map; Performing spatial point cloud reconstruction processing on the depth information map to generate a three-dimensional human body surface model; Performing motion artifact detection processing on the three-dimensional human body surface model to identify model distortion areas caused by rapid motion; Performing multi-frame data compensation processing on the distorted area of ​​the model and performing data repair using the three-dimensional human body surface model of adjacent frames; Projecting the restored 3D human body surface model onto a 2D image plane to generate an optimized human body contour map; The optimized human body contour map is input into the human body region positioning processing process to assist in generating a dynamic human body region set.

8. The method according to claim 1, characterized in that After feeding back the physical fitness assessment parameters to the physical fitness training guidance system, the method further includes: Receive real-time training feedback instructions through the physical fitness training guidance system; activating a posture correction guidance mode according to the real-time training feedback instruction; In the posture correction guidance mode, performing target deviation analysis processing on the currently generated spatiotemporal posture feature set; Identify the main action defect areas through the target deviation analysis process; Generate a visual correction guidance signal based on the main action defect parts; Superimposing the visual correction guidance signal on the training scene display interface in real time; Establishing an instant mapping relationship between action execution and guidance signals in the training scene display interface; When it is detected that the movement trajectory of the defective part and the degree of coincidence of the visual correction guidance signal reach a preset standard, the posture correction guidance mode is turned off.

9. A posture detection device, characterized in that: include: a memory storing a computer program; A processor, configured to load the computer program to implement the human posture detection method based on machine vision as described in any one of claims 1 to 8.

Images