A sports training system and method based on body measurement machine deep learning

Abstract

The application discloses a kind of sports training system and method based on body measuring machine deep learning, system includes: first acquisition module, for based on the reader and passive RFID tag matrix included in body side machine, the phase information of user is collected to carry out sports training;Second acquisition module, for based on the shooting module included in body side machine, video information of user is collected to carry out sports training;First determination module, for determining the first action information of user based on phase information;Second determination module, for determining the second action information of user based on video information;Adjustment optimization module, for determining the actual action of user according to first action information and second action information;Actual action is compared with standard action, and according to the suggestion information of comparison result is given adjustment optimization.It is adjusted and optimized to training action, improve the accuracy of action when user carries out sports training, further improve training effect.

Description

Technical Field

[0001] This invention relates to the field of training technology, and in particular to a sports training system and method based on deep learning for body composition analysis machines. Background Technology

[0002] With the rapid development of competitive sports and the popularization of the concept of fitness for all, the demand for scientific and precise sports training is increasing. Traditional training methods mainly rely on the coach's experience and the athlete's subjective feelings, which is difficult to meet the high requirements of modern sports for training effectiveness and athlete health protection. At the same time, they cannot evaluate and adjust the training movements of users during sports training, thus affecting the training effect and the accuracy of training movements. Summary of the Invention

[0003] This invention aims to at least partially solve one of the technical problems in the aforementioned technologies. Therefore, the purpose of this invention is to propose a sports training system and method based on deep learning for motion testing machines, which adjusts and optimizes training movements, improves the accuracy of users' movements during sports training, and further enhances training effectiveness.

[0004] To achieve the above objectives, embodiments of the present invention propose a sports training system based on deep learning for motion testing machines, comprising:

[0005] The first acquisition module is used to acquire phase information of the user during sports training based on the reader and passive RFID tag matrix included in the body measurement machine;

[0006] The second acquisition module is used to acquire video information of the user's sports training based on the shooting module included in the body measurement machine;

[0007] The first determining module is used to determine the user's first action information based on phase information;

[0008] The second determining module is used to determine the user's second action information based on video information;

[0009] The adjustment and optimization module is used to determine the user's actual action based on the first action information and the second action information; compare the actual action with the standard action; and provide adjustment and optimization suggestions based on the comparison results.

[0010] According to some embodiments of the present invention, the first acquisition module is used to read signals from a passive RFID tag matrix distributed on different parts of the user's body based on the reader included in the body measurement machine, and to acquire phase information of the user's physical training.

[0011] According to some embodiments of the present invention, the first determining module includes:

[0012] The calculation module is used for:

[0013] Calculate the phase difference between adjacent tags based on the phase information to determine the relative motion information of different parts of the user's body;

[0014] Based on the phase information, determine the rate of phase change over time, and thus determine the instantaneous speed of the user's action;

[0015] The user's body posture is determined based on the spatial distribution pattern of phase difference in the passive RFID tag matrix.

[0016] The recognition module is used to input the relative motion information of various parts of the user's body, instantaneous velocity, and body posture into the deep learning model to determine the user's first motion information.

[0017] According to some embodiments of the present invention, the second determining module includes:

[0018] The frame segmentation module is used to process video information into frames to obtain several frames of image information.

[0019] Extraction module, used for:

[0020] The image information is input into the Faster R-CNN network, which includes a fully convolutional deep network and a Fast R-CNN detector. The image information is processed to obtain feature maps. The feature maps are then input into the fully convolutional deep network to obtain candidate regions. The candidate regions and feature maps are then input into the fully convolutional layer for convolutional classification, and the detection results are output. Based on the detection results, the target occupying the largest area in the image information is extracted to obtain the user's skeleton information.

[0021] Key skeletal points and joints of the human body are extracted based on skeletal information; the key skeletal points include at least the shoulder, elbow, head, torso, and knee.

[0022] The calculation module is used for:

[0023] Based on the camera parameters of the shooting module and the attribute information of the training site, calculate the spatial position of key bone points and joints in three-dimensional space.

[0024] Determine the motion information of the joints based on their spatial location;

[0025] Based on the spatial location and the operation information of the joints, the user's posture information is determined, and the user's second action information is determined based on the posture information.

[0026] According to some embodiments of the present invention, the adjustment and optimization module compares the actual action with the standard action and provides adjustment and optimization suggestions based on the comparison results, including:

[0027] Add a module for:

[0028] Obtain the joint point information when the user completes the actual action, and model the joint points to obtain the joint diagram structure;

[0029] Extract the joint constraint rules of the actual action based on the action label corresponding to the actual action, add the joint constraint rules to the joint graph structure, and obtain the action constraint display diagram.

[0030] The third determining module is used for:

[0031] The motion constraint visualization is input into a graph convolutional neural network to determine the evaluation metrics for the joints. The evaluation metrics include the core role of the joints in the motion, the amount of information they provide, and whether they conform to the physical constraints of the motion.

[0032] The key scores of the joints are determined based on the evaluation indicators. The joints with key scores greater than the preset score threshold are selected as key joints. The standard movements corresponding to the actual movements are determined based on the key joints.

[0033] The alignment module is used to align the actual action with the standard action using a dynamic time warping algorithm, calculate the similarity score between the actual action and the standard action, obtain the comparison result, and provide adjustment and optimization suggestions based on the comparison result.

[0034] According to some embodiments of the present invention, it further includes: a processing module, configured to adjust and enhance the brightness of the image information and filter and denoise the image information before the extraction module inputs the image information into the Faster R-CNN network, to obtain enhanced image information.

[0035] According to some embodiments of the present invention, the processing module filters and denoises the image information to obtain enhanced image information, including:

[0036] The fourth determining module is used for:

[0037] In the image information, an arbitrary pixel is selected as the processing pixel, the gray value of the processing pixel is determined, and the average gray value of all pixels in the neighborhood corresponding to the processing pixel is determined.

[0038] When the absolute value of the difference between the gray value and the gray mean is greater than a preset difference threshold, the pixel to be processed is treated as a noise pixel.

[0039] The filtering and denoising module is used to treat the neighborhood of the noise pixel as the noise region, mark each noise region on the image information, and perform filtering and denoising processing to obtain enhanced image information.

[0040] According to some embodiments of the present invention, the filtering and denoising module includes:

[0041] The cropping module is used for:

[0042] Determine the grayscale histogram for each noise region and identify the peak value of the histogram;

[0043] Set a contrast limit threshold, crop the histogram peaks that are greater than the contrast limit threshold, and redistribute them to other gray levels;

[0044] The bilinear interpolation module is used for:

[0045] Equalize the cropped grayscale histogram to generate a local mapping table;

[0046] For noise pixels in a noise region, bilinear interpolation is performed based on the local mapping table of the noise region and adjacent noise regions to obtain the corresponding target gray value. The noise pixels are then processed based on the target gray value. This process is applied to all noise pixels in each noise region to obtain enhanced image information.

[0047] According to some embodiments of the present invention, a training method for a sports training system based on deep learning for motion testing includes:

[0048] Based on the reader and passive RFID tag matrix included in the body measurement machine, phase information of users during sports training is collected;

[0049] Based on the shooting module included in the body measurement machine, video information of users conducting sports training is collected;

[0050] Determine the user's first action information based on phase information;

[0051] Determine the user's second action information based on video information;

[0052] Based on the first action information and the second action information, determine the user's actual action; compare the actual action with the standard action, and provide suggestions for adjustment and optimization based on the comparison results.

[0053] This invention proposes a sports training system and method based on deep learning for motion testing machines, which adjusts and optimizes training movements to improve the accuracy of users' movements during sports training and further enhance training effectiveness.

[0054] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0055] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a block diagram of a sports training system based on deep learning for body composition analysis according to an embodiment of the present invention.

[0058] Figure 2 This is a block diagram of a first determining module according to an embodiment of the present invention;

[0059] Figure 3 This is a flowchart of a sports training method based on deep learning for body composition analysis according to an embodiment of the present invention. Detailed Implementation

[0060] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0061] like Figure 1 As shown in the figure, this embodiment of the invention proposes a sports training system based on deep learning for body composition analysis, comprising:

[0062] The first acquisition module is used to acquire phase information of the user during sports training based on the reader and passive RFID tag matrix included in the body measurement machine;

[0063] The second acquisition module is used to acquire video information of the user's sports training based on the shooting module included in the body measurement machine;

[0064] The first determining module is used to determine the user's first action information based on phase information;

[0065] The second determining module is used to determine the user's second action information based on video information;

[0066] The adjustment and optimization module is used to determine the user's actual action based on the first action information and the second action information; compare the actual action with the standard action; and provide adjustment and optimization suggestions based on the comparison results.

[0067] The working principle of the above technical solution is as follows: The first acquisition module uses the passive RFID reader and tag matrix built into the motion testing machine to capture the user's movement trajectory through the phase change of electromagnetic wave reflection signals. The second acquisition module uses a multi-camera array or depth camera to extract the user's skeletal key points and movement features through computer vision algorithms. The first determination module analyzes the movement patterns based on time-frequency analysis (such as short-time Fourier transform) and trajectory reconstruction algorithms of RFID phase data. The second determination module uses open-source frameworks such as OpenPose or MediaPipe, combined with deep learning models (such as HRNet) for pose estimation. The adjustment and optimization module integrates phase and video data through multimodal data fusion (such as Kalman filtering), compares it with a standard movement library (such as the International Gymnastics Federation scoring standards), and generates optimization suggestions. The optimization suggestions include displaying movement deviations.

[0068] The beneficial effects of the above technical solution are as follows: Based on the first acquisition module and the first determination module, it is convenient to achieve high-precision capture of rapid movements and minute actions. Based on the second acquisition module and the second determination module, it is convenient to achieve complex action decomposition. By using a Bayesian fusion algorithm and combining the confidence weights of the two types of data, the accuracy of action recognition is improved, making it easier to accurately determine the user's actual actions. The actual actions are compared with standard actions, and suggestions for adjustment and optimization are given based on the comparison results. The training actions are adjusted and optimized to improve the accuracy of the user's actions during sports training, thereby further improving the training effect.

[0069] According to some embodiments of the present invention, the first acquisition module is used to read signals from a passive RFID tag matrix distributed on different parts of the user's body based on the reader included in the body measurement machine, and to acquire phase information of the user's physical training.

[0070] The working principle and beneficial effects of the above technical solution are as follows: Passive RFID tags are affixed or embedded in key movement areas of the user (such as shoulders, elbows, hips, knees, etc.). The tags use the UHF band (e.g., 860-960MHz), supporting long-distance reading (>5 meters) and requiring no battery power. The reader emits electromagnetic waves, and the phase of the tag's reflected signal changes with the user's movements. Phase Difference of Arrival (PDOA) technology is used to extract the tag's three-dimensional spatial coordinate changes. Time-series phase data (e.g., 100 samples per second) is output, including the tag's position coordinates (x, y, z), movement speed, and acceleration. The user's motion trajectory and motion parameters are extracted by fitting the phase data. This achieves high-precision, low-latency capture of the user's motion trajectory.

[0071] like Figure 2 As shown, according to some embodiments of the present invention, the first determining module includes:

[0072] The calculation module is used for:

[0073] Calculate the phase difference between adjacent tags based on the phase information to determine the relative motion information of different parts of the user's body;

[0074] Based on the phase information, determine the rate of phase change over time, and thus determine the instantaneous speed of the user's action;

[0075] The user's body posture is determined based on the spatial distribution pattern of phase difference in the passive RFID tag matrix.

[0076] The recognition module is used to input the relative motion information of various parts of the user's body, instantaneous velocity, and body posture into the deep learning model to determine the user's first motion information.

[0077] The working principle and beneficial effects of the above technical solution are as follows: The relative motion information, instantaneous velocity, and body posture of various parts of the user's body are input into a deep learning model to determine the user's first motion information. Examples: For raising an arm: the shoulder label shows significant phase change, the wrist label shows phase lag, and the phase difference exhibits a gradient from top to bottom. For kicking a leg: the thigh label shows drastic phase change, the calf label shows smaller phase change, and the phase difference reaches its peak near the knee joint. The deep learning model includes a recurrent neural network. This facilitates accurate determination of the user's first motion information.

[0078] According to some embodiments of the present invention, the second determining module includes:

[0079] The frame segmentation module is used to process video information into frames to obtain several frames of image information.

[0080] Extraction module, used for:

[0081] The image information is input into the Faster R-CNN network, which includes a fully convolutional deep network and a Fast R-CNN detector. The image information is processed to obtain feature maps. The feature maps are then input into the fully convolutional deep network to obtain candidate regions. The candidate regions and feature maps are then input into the fully convolutional layer for convolutional classification, and the detection results are output. Based on the detection results, the target occupying the largest area in the image information is extracted to obtain the user's skeleton information.

[0082] Key skeletal points and joints of the human body are extracted based on skeletal information; the key skeletal points include at least the shoulder, elbow, head, torso, and knee.

[0083] The calculation module is used for:

[0084] Based on the camera parameters of the shooting module and the attribute information of the training site, calculate the spatial position of key bone points and joints in three-dimensional space.

[0085] Determine the motion information of the joints based on their spatial location;

[0086] Based on the spatial location and the operation information of the joints, the user's posture information is determined, and the user's second action information is determined based on the posture information.

[0087] The working principle of the above technical solution is as follows: Frame segmentation module: decomposes video information into a sequence of image frames. Extraction module: extracts human skeleton information and key skeletal points based on Faster R-CNN. Calculation module: combines camera parameters and site information to achieve 3D spatial positioning and action recognition. The frame segmentation module uses OpenCV or FFmpeg for video decoding and frame extraction. The Faster R-CNN network architecture includes a fully convolutional deep network (RPN): generating candidate regions (Region Proposals). The Fast R-CNN detector: classifies and regresses the candidate regions. The target occupying the largest area of ​​the image is selected to obtain the user's skeleton information; key skeletal points and joints of the human body are extracted based on the skeleton information; this facilitates accurate identification of key skeletal points and joints in the human body, facilitating subsequent action recognition. Camera parameters include intrinsic parameters (such as focal length, principal point coordinates, distortion coefficients, etc.) and extrinsic parameters (such as rotation matrix, translation vector), used to transform image coordinates to the world coordinate system. The attribute information of the training site includes the size, shape, and location of landmarks, used to establish the site coordinate system and assist in determining the relative position of the human body within the site. Human pose estimation algorithms (such as OpenPose and AlphaPose) detect the two-dimensional coordinates of key skeletal points and joints from images. Camera intrinsics are used to correct image distortion, eliminating the impact of lens distortion on coordinate measurement accuracy. Using triangulation principles and combined with camera extrinsic parameters, the two-dimensional coordinates are converted to three-dimensional coordinates. Based on the geometric information of the training ground, the three-dimensional coordinates are transformed from the camera coordinate system to the ground coordinate system for subsequent analysis, thereby determining the spatial positions of key skeletal points and joints in three-dimensional space. Joints are the intersections or connections of adjacent bone segments. Based on the spatial position, the motion information of the joints is determined, including joint angles and relative positions. Specific steps include: for each joint, defining the direction vectors of adjacent bone segments. For example, for the knee joint, the direction vectors of the thigh and calf bone segments can be defined. The angle between the two direction vectors, i.e., the joint angle, is calculated using the dot product formula. The calculated cosine value is converted into an angle value to obtain the joint angle. The Euclidean distance between different skeletal points or joints is calculated to describe their relative positional relationships. Analyze the directional relationships between skeletal points or joints, for example, determine whether a joint is oriented forward, backward, left, or right relative to another joint. Combine joint angles, relative position information, and spatial location to describe and analyze human posture. Posture information includes: timestamps: recording the time of data acquisition; spatial position vectors: the three-dimensional coordinates of key skeletal points or joints (e.g., shoulder, elbow, wrist); joint angle vectors: the rotation angles of each joint (e.g., shoulder, elbow, knee); and relative position vectors: describing the relative positional relationships between skeletal points or joints (e.g., the offset of the elbow relative to the shoulder).Each attitude point is added to the sequence in chronological order to form a continuous attitude change trajectory.

[0088] The beneficial effect of the above technical solution is that it accurately determines the user's second action information based on posture information.

[0089] According to some embodiments of the present invention, the adjustment and optimization module compares the actual action with the standard action and provides adjustment and optimization suggestions based on the comparison results, including:

[0090] Add a module for:

[0091] Obtain the joint point information when the user completes the actual action, and model the joint points to obtain the joint diagram structure;

[0092] Extract the joint constraint rules of the actual action based on the action label corresponding to the actual action, add the joint constraint rules to the joint graph structure, and obtain the action constraint display diagram.

[0093] The third determining module is used for:

[0094] The motion constraint visualization is input into a graph convolutional neural network to determine the evaluation metrics for the joints. The evaluation metrics include the core role of the joints in the motion, the amount of information they provide, and whether they conform to the physical constraints of the motion.

[0095] The key scores of the joints are determined based on the evaluation indicators. The joints with key scores greater than the preset score threshold are selected as key joints. The standard movements corresponding to the actual movements are determined based on the key joints.

[0096] The alignment module is used to align the actual action with the standard action using a dynamic time warping algorithm, calculate the similarity score between the actual action and the standard action, obtain the comparison result, and provide adjustment and optimization suggestions based on the comparison result.

[0097] The working principle of the above technical solution is as follows: It acquires the joint information of the user when performing actual actions, including the 3D joint coordinates of the target user (such as shoulder, elbow, wrist, etc.) and the connection relationships of the joints (skeletal chain). Each joint is treated as a graph node, and the skeletal connections are treated as undirected edges, constructing an undirected graph G = (V, E), where V is the set of nodes and E is the set of edges. Feature vectors (such as joint coordinates, velocity, acceleration) are added to each node to enhance the information representation capability of the graph structure. Joint constraint rules for the action are extracted based on action labels (such as "walking," "running," "jumping"). For example, when running, the knee joint angle must be within a specific range, and the arm swing amplitude must conform to kinematic laws. Joint constraint rules facilitate the determination of initial nodes. Examples include: Rule-driven: defining the starting node rules for each action based on action classification information. For example: Walking: using the hip joint as the initial node (starting point of movement). Jumping: using the ankle joint as the initial node (support point). Throwing: using the shoulder joint as the initial node (force point). The physical constraints of the movement (such as joint angle range and movement speed limit) are encoded as graph constraints, i.e., a movement constraint representation graph, for subsequent GCN calculations. This movement constraint representation graph is input into a graph convolutional neural network to determine the evaluation metrics for joints. These metrics include the core role of the joint in the movement (e.g., the weight-bearing function of the hip joint in walking), the amount of information provided (e.g., the crucial role of hand joints in grasping), and whether it conforms to the physical constraints of the movement (e.g., the range of ankle joint angle changes during running). Based on the evaluation metrics, key scores are determined for each joint. Joints with key scores greater than a preset threshold are selected as key joints. A standard movement corresponding to the actual movement is determined based on these key joints. An alignment module uses a dynamic time warping algorithm to align the actual movement with the standard movement, calculates the similarity score between the two movements, obtains the comparison results, and provides suggestions for adjustment and optimization based on the comparison results.

[0098] The beneficial effects of the above technical solution are: it facilitates the accurate comparison between actual actions and standard actions, thereby obtaining accurate suggestions.

[0099] According to some embodiments of the present invention, it further includes: a processing module, configured to adjust and enhance the brightness of the image information and filter and denoise the image information before the extraction module inputs the image information into the Faster R-CNN network, to obtain enhanced image information.

[0100] The beneficial effects of the above technical solution are: improved accuracy of image information.

[0101] According to some embodiments of the present invention, the processing module filters and denoises the image information to obtain enhanced image information, including:

[0102] The fourth determining module is used for:

[0103] In the image information, an arbitrary pixel is selected as the processing pixel, the gray value of the processing pixel is determined, and the average gray value of all pixels in the neighborhood corresponding to the processing pixel is determined.

[0104] When the absolute value of the difference between the gray value and the gray mean is greater than a preset difference threshold, the pixel to be processed is treated as a noise pixel.

[0105] The filtering and denoising module is used to treat the neighborhood of the noise pixel as the noise region, mark each noise region on the image information, and perform filtering and denoising processing to obtain enhanced image information.

[0106] The working principle and beneficial effects of the above technical solution are as follows: Noise detection: Based on the difference between the gray value of a pixel and the mean value of its neighborhood, noisy pixels are identified. Noise region marking: The neighborhood of a noisy pixel is defined as a noise region, and its position is marked. Filtering and denoising: The noise region is filtered to eliminate noise and enhance image details. The neighborhood can be 5*5. When the absolute value of the difference between the gray value and the mean gray value is greater than a preset difference threshold, the pixel is treated as a noise pixel; for each noise pixel P(x,y), a noise region is defined centered on it. Each noise region is marked on the image information, and filtering and denoising are performed to obtain enhanced image information. This facilitates accurate identification of each noise region, and filtering and denoising are performed on a unit basis. Through gray-level difference analysis and adaptive filtering, efficient noise detection and denoising are achieved, resulting in accurate image information.

[0107] According to some embodiments of the present invention, the filtering and denoising module includes:

[0108] The cropping module is used for:

[0109] Determine the grayscale histogram for each noise region and identify the peak value of the histogram;

[0110] Set a contrast limit threshold, crop the histogram peaks that are greater than the contrast limit threshold, and redistribute them to other gray levels;

[0111] The bilinear interpolation module is used for:

[0112] Equalize the cropped grayscale histogram to generate a local mapping table;

[0113] For noise pixels in a noise region, bilinear interpolation is performed based on the local mapping table of the noise region and adjacent noise regions to obtain the corresponding target gray value. The noise pixels are then processed based on the target gray value. This process is applied to all noise pixels in each noise region to obtain enhanced image information.

[0114] The working principle and beneficial effects of the above technical solution are as follows: Noise region grayscale histogram optimization: By cropping high-contrast peaks, the uniformity of grayscale distribution in the noise region is enhanced. Local grayscale mapping generation: A local grayscale mapping table is generated based on the cropped histogram to improve the contrast of the noise region. Bilinear interpolation enhancement: Combining the neighborhood mapping table, the target grayscale value of the noise pixels is calculated through interpolation to achieve smooth transition and detail preservation. For all pixels within the noise region R, the frequency H(g) of grayscale level g∈[0,255] is counted. The cumulative distribution function of the histogram is calculated. Where H(i) is the frequency of the i-th gray level.

[0115] Contrast Limit Threshold T c for:

[0116] T c =μ R +k·σ R

[0117] Where, μ R σ is the mean gray level of the noise region. R is the standard deviation; k is the adjustment coefficient;

[0118] For all g greater than T c The grayscale level is cropped to T based on its frequency. c And distribute the excess frequencies evenly to other gray levels:

[0119]

[0120] in, The number of gray levels to be cropped.

[0121] The cropped histogram is equalized to generate a local grayscale mapping table:

[0122]

[0123] Where CDF′(g) is the cumulative distribution function of the clipped histogram; CDF′ min N represents the minimum value of the cumulative distribution function of the clipped histogram; pixels This represents the total number of pixels in the noise region.

[0124] For each noisy pixel p(x, y), determine its noise region R and adjacent noise regions R1, R2, ... R. n Obtain the local mapping table for each region.

[0125] Based on the position (x,y) of P(x,y) and the boundaries of each region, calculate the interpolation weights W1, W2, ... W n :

[0126]

[0127] in, For region R i The center coordinates of d; x and d y This represents the interpolation range.

[0128] Calculate the target grayscale value;

[0129]

[0130] Replace the original grayscale value of P(x,y) with G′(x,y)′. Repeat the above steps for all noisy pixels to generate the enhanced image. Set multiple contrast limit thresholds for different grayscale ranges to improve the flexibility of histogram cropping. Improve the smoothness of grayscale value calculation. Dynamically adjust the interpolation weights based on the edge intensity of the noisy region to avoid edge blurring. Pre-calculate the integral image to quickly calculate the grayscale histogram and cumulative distribution function. Generate local mapping tables and interpolation weights in parallel for different noisy regions. Through grayscale histogram cropping and bilinear interpolation, local contrast enhancement and smooth grayscale transition in the noisy region are achieved.

[0131] like Figure 3 As shown, according to some embodiments of the present invention, a training method for a sports training system based on deep learning of a motion testing machine includes steps S1-S5:

[0132] S1. Based on the reader and passive RFID tag matrix included in the body measurement machine, collect phase information of the user during sports training;

[0133] S2. Based on the shooting module included in the body measurement machine, collect video information of the user's sports training;

[0134] S3. Determine the user's first action information based on phase information;

[0135] S4. Determine the user's second action information based on video information;

[0136] S5. Based on the first action information and the second action information, determine the user's actual action; compare the actual action with the standard action, and provide suggestions for adjustment and optimization based on the comparison results.

[0137] The beneficial effects of the above technical solution are as follows: By combining the confidence weights of two types of data through a Bayesian fusion algorithm, the accuracy of action recognition is improved, making it easier to accurately determine the user's actual actions. The actual actions are compared with standard actions, and suggestions for adjustment and optimization are provided based on the comparison results. Adjustments and optimizations to training actions improve the accuracy of users' movements during sports training, further enhancing training effectiveness.

[0138] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A sports training system based on body measurement machine deep learning, characterized in that, The method comprises the following steps: A first acquisition module is configured to acquire phase information of a user performing physical training based on a reader and a passive RFID tag matrix included in the body-side machine; A second acquisition module is configured to acquire video information of the user performing physical training based on a shooting module included in the body-side machine; A first determination module is configured to determine first action information of the user based on the phase information; A second determination module is configured to determine second action information of the user based on the video information; An adjustment and optimization module is configured to determine actual action of the user according to the first action information and the second action information; The actual action is compared with standard action, and adjustment and optimization suggestion information is given according to a comparison result; The second determination module comprises: A frame division module is configured to perform frame division processing on the video information to obtain a plurality of frame image information; An extraction module is configured to: input the image information into a Faster R-CNN network, the Faster R-CNN network comprising a full convolutional deep network and a Fast R-CNN detector, process the image information to obtain a feature map, input the feature map into the full convolutional deep network to obtain a candidate region, and input the candidate region and the feature map into a full convolutional layer for convolution classification to output a detection result; according to the detection result, extract a target occupying the largest area in the image information to obtain skeleton information of the user; extract key skeleton points and joint points of the human body according to the skeleton information; the key skeleton points at least include shoulders, elbows, heads, trunks and knees; A calculation module is configured to: calculate spatial positions of the key skeleton points and the joint points in a three-dimensional space according to camera parameters of the shooting module and attribute information of a training site; determine motion information of the joint points according to the spatial positions; determine posture information of the user according to the spatial positions and the motion information of the joint points, and determine the second action information of the user according to the posture information; The adjustment and optimization module compares the actual action with the standard action, and gives adjustment and optimization suggestion information according to a comparison result, which comprises: An adding module is configured to: obtain joint point information of the user when completing the actual action, and obtain a joint graph structure according to joint point modeling; extract joint point constraint rules of the actual action based on action labels corresponding to the actual action, add the joint point constraint rules to the joint graph structure to obtain an action constraint display graph; A third determination module is configured to: input the action constraint display graph into a graph convolutional neural network to determine evaluation indexes of the joint points; the evaluation indexes comprise a core role of the joint points in the action, an information amount provided by the joint points and whether the joint points meet physical constraints of the action; determine key scores of the joint points according to the evaluation indexes, filter out joint points with key scores greater than a preset score threshold as key joint points, and determine a standard action corresponding to the actual action based on the key joint points; An alignment module is configured to align the actual action with the standard action by using a dynamic time warping algorithm, calculate a similarity score between the actual action and the standard action to obtain a comparison result, and give adjustment and optimization suggestion information according to the comparison result. 2.The body measurement machine deep learning-based physical training system of claim 1, wherein, The first acquisition module is used to read signals from a matrix of passive RFID tags distributed on different parts of the user's body based on the reader included in the body measurement machine, and to collect phase information of the user during sports training. 3.The sports training system based on body measurement machine deep learning according to claim 2, wherein, The first determining module includes: The calculation module is used for: Calculate the phase difference between adjacent tags based on the phase information to determine the relative motion information of different parts of the user's body; Based on the phase information, determine the rate of phase change over time, and thus determine the instantaneous speed of the user's action; The user's body posture is determined based on the spatial distribution pattern of phase difference in the passive RFID tag matrix. The recognition module is used to input the relative motion information of various parts of the user's body, instantaneous velocity, and body posture into the deep learning model to determine the user's first motion information. 4.The sports training system based on body measurement machine deep learning according to claim 1, wherein, Also includes: The processing module is used to adjust and enhance the brightness of the image information and filter and denoise the image information before the extraction module inputs the image information into the Faster R-CNN network, so as to obtain enhanced image information. 5.The sports training system based on body measurement machine deep learning according to claim 4, wherein, The processing module filters and denoises the image information to obtain enhanced image information, including: The fourth determining module is used for: In the image information, an arbitrary pixel is selected as the processing pixel, the gray value of the processing pixel is determined, and the average gray value of all pixels in the neighborhood corresponding to the processing pixel is determined. When the absolute value of the difference between the gray value and the gray mean is greater than a preset difference threshold, the pixel to be processed is treated as a noise pixel. The filtering and denoising module is used to treat the neighborhood of the noise pixel as the noise region, mark each noise region on the image information, and perform filtering and denoising processing to obtain enhanced image information. 6.The sports training system based on body measurement machine deep learning according to claim 5, wherein, The filtering and denoising module includes: The cropping module is used for: Determine the grayscale histogram for each noise region and identify the peak value of the histogram; Set a contrast limit threshold, crop the histogram peaks that are greater than the contrast limit threshold, and redistribute them to other gray levels; The bilinear interpolation module is used for: Equalize the cropped grayscale histogram to generate a local mapping table; For noise pixels in a noise region, bilinear interpolation is performed based on the local mapping table of the noise region and adjacent noise regions to obtain the corresponding target gray value. The noise pixels are then processed based on the target gray value. This process is applied to all noise pixels in each noise region to obtain enhanced image information. 7.The training method of the body measurement machine deep learning-based sports training system according to any one of claims 1-6, wherein, include: Based on the reader and passive RFID tag matrix included in the body measurement machine, phase information of users during sports training is collected; Based on the shooting module included in the body measurement machine, video information of users conducting sports training is collected; Determine the user's first action information based on phase information; Determine the user's second action information based on video information; Based on the first action information and the second action information, determine the user's actual action; compare the actual action with the standard action, and provide suggestions for adjustment and optimization based on the comparison results.

Images