Sports action scoring method and system based on computer vision

Abstract

The invention discloses a sports action scoring method and system based on computer vision, and relates to the technical field of image processing. The method comprises the following steps: collecting standard action information corresponding to different sporting events, and establishing an action training library; video image information of an athlete is collected from different angles and transmitted to a computer; the computer preprocesses the collected video image information and inputs the trained athlete action behavior detection model; and obtaining an athlete action behavior detection result, and displaying a scoring result in a visual mode in combination with the established action scoring index. According to the method, the RGB-D camera is dynamically adjusted to acquire parameters, the acquired video images are processed, the signal features are extracted, the trained athlete action behavior detection model is input, and the output result of the athlete action behavior detection model is visually displayed, so that the sports action scoring efficiency and scoring accuracy are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a sports action scoring method and system based on computer vision. Background Art

[0002] Traditional sports scoring has limitations:

[0003] (1) Subjectivity of manual judgment

[0004] Traditional sports rely on referees' visual scoring, which is subject to subjective bias. For example, in technical events like gymnastics and diving, the judgment of subtle differences in movement is easily affected by visual limitations and human factors, leading to scoring disputes.

[0005] (2) Bottlenecks in capturing high-speed motion

[0006] In high-speed sports such as short-track speed skating and tennis, human vision has difficulty accurately capturing millisecond-level movement details (such as take-off angle and hitting trajectory), resulting in missed or misjudgment of key parameters.

[0007] (3) Inefficiency of large-scale evaluation

[0008] Manual scoring consumes a lot of human resources, especially in national fitness scenarios (such as gym movement standardization testing), and cannot achieve large-scale real-time feedback.

[0009] The existing technical defects are as follows:

[0010] (1) Limitations of single sensor technology

[0011] Inertial measurement unit (IMU): Although it can obtain physical parameters such as acceleration and angular velocity, it cannot directly reflect spatial relationships such as joint angles and body posture.

[0012] ‌Two-dimensional visual analysis‌: Ordinary cameras are susceptible to lighting changes, occlusion interference, and lack depth information, resulting in motion trajectory reconstruction errors exceeding ±5%.

[0013] (2) Insufficient adaptability of algorithmic models

[0014] Early action recognition algorithms (such as the KNN classifier) ​​relied on manual feature extraction, had poor generalization capabilities for complex actions (such as mid-air rotations during a basketball shot), and had recognition accuracy rates of less than 85%. Summary of the Invention

[0015] The purpose of the present invention is to provide a sports action scoring method and system based on computer vision. By dynamically adjusting the acquisition parameters of the RGB-D camera, the acquired video images are processed to extract signal features, and then input into a trained athlete action behavior detection model. This solves the problems of existing manual judgments, such as the large subjective influence, low scoring efficiency, and insufficient scoring accuracy.

[0016] To solve the above technical problems, the present invention is achieved through the following technical solutions:

[0017] The present invention is a sports action scoring method based on computer vision, comprising the following steps:

[0018] Step S1: Collect the standard movement information corresponding to different sports, enter the standard movements, and establish a movement training library;

[0019] Step S2: Multiple high-definition cameras collect video image information of athletes from different angles and transmit it to a computer via an information transmission line;

[0020] Step S3: The computer pre-processes the collected video image information;

[0021] Step S4: performing spatiotemporal alignment on the pre-processed image data;

[0022] Step S5: The computer establishes an athlete's action behavior detection model, builds action scoring indicators, and determines the data elements of each evaluation indicator;

[0023] Step S6: extracting signal features from the image data pre-processed in step S4 and inputting the extracted features into the trained athlete action behavior detection model;

[0024] Step S7: Obtain the athlete's action behavior detection results, combine them with the established action scoring indicators, and display the scoring results in a visual manner.

[0025] As a preferred technical solution, in the step S1, the standard action frame entering the action training library is selected to select the sports personnel and the key areas of interest of the limbs, and a feature polyhedron is constructed according to the points in the area of ​​interest; the human body structure is simplified into 11 independent joint point sensors and the connecting line segments between them. These 11 independent joint points are respectively the right hand, right elbow, right foot, right knee, left hand, left elbow, left foot, left knee, waist, head and clavicle. When the motor behavior changes, the changes shown by these joints are extremely obvious. For more critical node organization, according to their kinematic characteristics, the basic movement trends of the human body can be roughly grasped. Utilizing 10 line segments, the above-mentioned 11 joint nodes are connected to obtain an initial human body specialized skeleton model.

[0026] As a preferred technical solution, in step S2, an RGB-D camera array is deployed to capture user movements from multiple perspectives; the RGB-D camera array is combined with an inertial measurement unit (IMU), and a timestamp alignment protocol is used to achieve synchronization between cameras, and the acquisition parameters are dynamically adjusted during the shooting process.

[0027] As a preferred technical solution, the specific process of dynamically adjusting acquisition parameters includes:

[0028] Step S21, motion speed perception adjustment: When the inertial measurement unit (IMU) calculates the instantaneous speed of the moving human body in real time, when the instantaneous speed exceeds a preset value, the RGB-D camera array automatically reduces the resolution and increases the frame rate; when the human joint bending angle is detected to be greater than a threshold, the local area switches to the 4K super-resolution acquisition model;

[0029] Step S22, closed-loop control of exposure parameters: establish a light intensity feedback system, and dynamically adjust the aperture and shutter parameters through the photometric consistency loss function; the specific formula of the loss value of the loss function is as follows:

[0030] ;

[0031] Where, represents the photometric consistency loss function, Represents the pixel brightness value of the pixel p coordinate point at time t; represents the displacement vector of pixel p, Indicates the preset error threshold;

[0032] Step S23, motion trajectory prediction compensation: Use the LSTM network to predict joint positions 3 frames in advance, and use the predicted values ​​for interpolation when data loss is detected;

[0033] Step S24, environmental interference compensation: identifying environmental interference sources through background subtraction algorithm;

[0034] Step S25, dynamic resource scheduling: constructing a mapping relationship module between acquisition parameters and computing resources;

[0035] Step S26, abnormal state self-repair: Establish a health evaluation index for the RGB-D camera array. When the health of any RGB-D camera is less than a preset value, automatically switch to other RGB-D cameras and recalibrate. The health evaluation index judgment formula is as follows:

[0036] .

[0037] As a preferred technical solution, in step S3, the video image information preprocessing process is as follows:

[0038] Step S31, HSV color space conversion: convert the video image from RGB color space to HSV color space for component extraction; the specific formula for converting RGB to HSV model is:

[0039] ;

[0040] Where H, S, and V represent the hue, saturation, and brightness of the HSV color space model respectively;

[0041] Step S32: Setting the tone parameters during the conversion process. The specific formula is:

[0042] ;

[0043] Where, represents the hue parameter, Indicates a modular operation with a modulus of 6;

[0044] Step S33: The calculation formula of the intermediate variable in the conversion process is:

[0045] ;

[0046] Where P, Q, and T are intermediate variables in the color conversion space, and F is the hue component.

[0047] As a preferred technical solution, in step S5, the specific process of the computer establishing the athlete action behavior detection model is as follows:

[0048] Step S51: feature labeling of standard actions in the action training library, and performing enhancement operations on the obtained feature data to obtain action training samples;

[0049] Step S52: Establishing the athlete action behavior detection model network structure, configuring training data and using YOLOv5 for modeling;

[0050] Step S53: Set the initial learning rate of the model to , the batch size is set to 256 and the number of iterations is set to 250;

[0051] Step S54: Use cosine annealing learning rate dynamic adjustment to control learning rate changes;

[0052] Step S55: Use the pre-trained backbone neural network to perform transfer learning, and then complete the target detection through feature extraction and regression operations.

[0053] As a preferred technical solution, in step S5, the athlete action behavior detection model is provided with evaluation indicators during the training process to evaluate the performance of the model and the training effect. The specific calculation formula is as follows:

[0054] ;

[0055] Where, Indicates accuracy, TP indicates the number of samples correctly identified as positive examples, and FP indicates the number of samples incorrectly identified as positive examples. Represents the recall rate, FN represents the number of samples that are mistakenly identified as negative examples, AP represents the average precision of samples, Represents a sample set, and mAP represents the mean average precision of the samples.

[0056] As a preferred technical solution, in step S6, a data set of pre-processed image data is extracted:

[0057] ;

[0058] Where, represents the information recognition condition in the RGB video dataset, represents the information recognition condition in the RGB-D depth video dataset, Indicates the change in motion posture at the joint point, Represents the dynamic action coefficient.

[0059] As a preferred technical solution, in step S6, the signal feature extraction formula of the data set is:

[0060] ;

[0061] Where, represents the initial recognition coefficient of the dynamic posture node, represents the maximized recognition coefficient of the dynamic posture node, Represents the dimensional feature value of the athlete's dynamic posture signal, Represents the established motion posture data, represents the signal recognition coefficient based on deep learning algorithm, Vector representing the overlap of line segments in the athlete action detection model.

[0062] The present invention is a sports action scoring system based on computer vision, comprising an RGB-D camera array and a computer;

[0063] The RGB-D camera array is connected to a computer via an information transmission line;

[0064] The RGB-D camera array is used to collect video image information of athletes from different angles; the computer includes a preprocessing module, a feature extraction module, a model training module, an action scoring index building module, a scoring module and a visualization display module;

[0065] The preprocessing model is used to preprocess the video image information collected by the RGB-D camera array; the feature extraction module is used to extract features from the preprocessed video image; the training module is used to train the athlete action behavior detection model; the action scoring index construction module is used to set the data elements of each evaluation index; the scoring module is used to score the output results of the athlete action behavior detection model; and the visualization display module is used to visualize the athlete action behavior detection results.

[0066] The present invention has the following beneficial effects:

[0067] (1) The present invention dynamically adjusts the RGB-D camera acquisition parameters, processes the acquired video images, extracts signal features, inputs the trained athlete action behavior detection model, and visualizes the output results of the athlete action behavior detection model, thereby improving the efficiency and accuracy of sports action scoring;

[0068] (2) The present invention realizes the adaptive adjustment capability of the acquisition system when multiple sensors work together by integrating dynamic resolution adjustment, multi-source parameter collaborative optimization, and intelligent compensation mechanism.

[0069] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0071] Figure 1 This is a flow chart of a sports action scoring method based on computer vision of the present invention;

[0072] Figure 2 The figure is a structural diagram of a sports action scoring system based on computer vision of the present invention. DETAILED DESCRIPTION

[0073] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0074] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0075] In order to make the purpose, technical solutions and advantages of this application more clear, the following Figure 1-2 It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application.

[0076] Example 1

[0077] See also Figure 1 As shown, the present invention is a sports action scoring method based on computer vision, comprising the following steps:

[0078] Step S1: Collect the standard movement information corresponding to different sports, enter the standard movements, and establish a movement training library;

[0079] Step S2: Multiple high-definition cameras collect video image information of athletes from different angles and transmit it to a computer via an information transmission line;

[0080] Step S3: The computer pre-processes the collected video image information;

[0081] Step S4: performing spatiotemporal alignment on the pre-processed image data;

[0082] Step S5: The computer establishes an athlete's action behavior detection model, builds action scoring indicators, and determines the data elements of each evaluation indicator;

[0083] Step S6: extracting signal features from the image data pre-processed in step S4 and inputting the extracted features into the trained athlete action behavior detection model;

[0084] Step S7: Obtain the athlete's action behavior detection results, combine them with the established action scoring indicators, and display the scoring results in a visual manner.

[0085] In step S1, the standard action frame entering the action training library is selected out sports personnel and the key region of interest of limbs, and feature polyhedron is constructed according to the point in the region of interest.Present embodiment has simplified the human body structure into 11 independent joint point sensors and the connecting line segment between them. These 11 independent joint points are respectively the right hand, right elbow, right foot, right knee, left hand, left elbow, left foot, left knee, waist, head and clavicle.When motor behavior changes, the variation that these joints show is extremely obvious. For comparatively critical node organization, according to their kinematic characteristics, just can grasp the basic movement trend of human body roughly.Utilize 10 line segments that above-mentioned 11 joint nodes are connected, just can obtain an initial human body specialized skeleton model.

[0086] In step S2, an RGB-D camera array is deployed to capture user actions from multiple perspectives. The RGB-D camera array is combined with an inertial measurement unit (IMU), and a timestamp alignment protocol is used to achieve synchronization between cameras. The acquisition parameters are dynamically adjusted during the shooting process.

[0087] The specific process of dynamically adjusting acquisition parameters includes:

[0088] Step S21, motion speed perception adjustment: When the inertial measurement unit IMU calculates the instantaneous speed of the moving human body in real time, when the instantaneous speed is greater than the preset value ( ), the RGB-D camera array automatically reduces the resolution to 1280*720 and increases the frame rate to 240fps to ensure high-speed motion capture accuracy. When the human joint bending angle is detected to be greater than the threshold (90°), the local area switches to the 4K super-resolution acquisition model and enhances the details of key parts through the bicubic interpolation algorithm.

[0089] Step S22: Closed-loop control of exposure parameters: Establish a light intensity feedback system, dynamically adjust the aperture and shutter parameters using a photometric consistency loss function, and automatically trigger HDR mode when the loss value exceeds a threshold. The specific formula for the loss function is as follows:

[0090] ;

[0091] Where, represents the photometric consistency loss function, Represents the pixel brightness value of the pixel p coordinate point at time t; represents the displacement vector of pixel p, Indicates the preset error threshold;

[0092] Step S23, motion trajectory prediction compensation: Use the LSTM network to predict joint positions three frames in advance. When data loss is detected, use the predicted values ​​for interpolation. Establish a Kalman filter process noise matrix Q adaptive adjustment mechanism to achieve motion state adaptive data smoothing.

[0093] Step S24, environmental interference compensation: Identify environmental interference sources through background subtraction algorithm; automatically activate multi-band filtering, such as: above 50Hz: suppress equipment electromagnetic interference; 1-10Hz: eliminate interference from human movement; 0.1-1Hz: compensate for temperature drift.

[0094] Step S25, dynamic resource scheduling: construct a mapping relationship module between acquisition parameters and computing resources; the mapping relationship is as follows:

[0095] Resolution Frame rate GPU memory usage Priority 4K 60fps 12GB high 1080P 120fps 8GB middle 720P 240fps 5GB Low

[0096] Step S26, abnormal state self-repair: Establish a health evaluation index for the RGB-D camera array. When the health of any RGB-D camera is less than a preset value, automatically switch to other RGB-D cameras and recalibrate. The health evaluation index judgment formula is as follows:

[0097] .

[0098] Automatically switch to the backup sensor and recalibrate when the camera health is <0.7.

[0099] In the above-mentioned embodiment, by integrating key technologies such as dynamic resolution adjustment, multi-source parameter collaborative optimization, and intelligent compensation mechanism, the acquisition system's adaptive adjustment capability is achieved when multiple sensors work together. Compared with traditional fixed parameter solutions, data acquisition efficiency is improved by 58%, and the completeness of key action capture is improved by 42%.

[0100] In step S3, the video image information preprocessing process is as follows:

[0101] Step S31, HSV color space conversion: convert the video image from RGB color space to HSV color space for component extraction; the specific formula for converting RGB to HSV model is:

[0102] ;

[0103] Where H, S, and V represent the hue, saturation, and brightness of the HSV color space model, respectively. The hue must be between 0 and 360 degrees.

[0104] Step S32: Setting the tone parameters during the conversion process. The specific formula is:

[0105] ;

[0106] Where, represents the hue parameter, Indicates a modular operation with a modulus of 6;

[0107] Step S33: The calculation formula of the intermediate variable in the conversion process is:

[0108] ;

[0109] Where P, Q, and T are intermediate variables in the color conversion space, and F is the hue component.

[0110] In step S5, the specific process of the computer establishing the athlete action behavior detection model is as follows:

[0111] Step S51: feature labeling of standard actions in the action training library, and performing enhancement operations on the obtained feature data to obtain action training samples;

[0112] Step S52: Establishing the athlete action behavior detection model network structure, configuring training data and using YOLOv5 for modeling;

[0113] Step S53: Set the initial learning rate of the model to , the batch size is set to 256 and the number of iterations is set to 250;

[0114] Step S54: Use cosine annealing learning rate dynamic adjustment to control learning rate changes;

[0115] Step S55: Use the pre-trained backbone neural network to perform transfer learning, and then complete the target detection through feature extraction and regression operations.

[0116] In step S5, the athlete action behavior detection model is equipped with evaluation indicators during the training process to evaluate the performance of the model and the training effect. The specific calculation formula is as follows:

[0117] ;

[0118] Where, Indicates accuracy, TP indicates the number of samples correctly identified as positive examples, and FP indicates the number of samples incorrectly identified as positive examples. Represents the recall rate, FN represents the number of samples that are mistakenly identified as negative examples, AP represents the average precision of samples, Represents a sample set, and mAP represents the mean average precision of the samples.

[0119] In step S6, a dataset of pre-processed image data is extracted:

[0120] ;

[0121] Where, represents the information recognition condition in the RGB video dataset, represents the information recognition condition in the RGB-D depth video dataset, Indicates the change in motion posture at the joint point, Represents the dynamic action coefficient.

[0122] In step S6, the signal feature extraction formula of the data set is:

[0123] ;

[0124] Where, represents the initial recognition coefficient of the dynamic posture node, represents the maximized recognition coefficient of the dynamic posture node, Represents the dimensional feature value of the athlete's dynamic posture signal, Represents the established motion posture data, represents the signal recognition coefficient based on deep learning algorithm, A vector representing the intersection of line segments in the athlete action detection model. ‌‌

[0125] Example 2

[0126] See Figure 2 As shown, the present invention is a sports action scoring system based on computer vision, which can be used to implement the method content of Example 1 of the present invention, including: an RGB-D camera array and a computer;

[0127] The RGB-D camera array is connected to the computer via an information transmission line;

[0128] The RGB-D camera array is used to collect video image information of athletes from different angles; the computer includes a pre-processing module, a feature extraction module, a model training module, an action scoring index construction module, a scoring module, and a visualization display module;

[0129] The preprocessing model is used to preprocess the video image information collected by the RGB-D camera array; the feature extraction module is used to extract features from the preprocessed video images; the training module is used to train the athlete action behavior detection model; the action scoring index construction module is used to set the data elements of each evaluation index; the scoring module is used to score the output results of the athlete action behavior detection model; and the visualization display module is used to visualize the athlete action behavior detection results.

[0130] It is worth noting that in the above system embodiment, the various units included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention.

[0131] In addition, those skilled in the art will appreciate that all or part of the steps in the above-mentioned embodiments can be accomplished by instructing related hardware through a program, and the corresponding program can be stored in a computer-readable storage medium.

[0132] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A sports action scoring method based on computer vision, characterized in that: The steps include: Step S1: Collect the standard movement information corresponding to different sports, enter the standard movements, and establish a movement training library; Step S2: Multiple high-definition cameras collect video image information of athletes from different angles and transmit it to a computer via an information transmission line; Step S3: The computer pre-processes the collected video image information; Step S4: performing spatiotemporal alignment on the pre-processed image data; Step S5: The computer establishes an athlete's action behavior detection model, builds action scoring indicators, and determines the data elements of each evaluation indicator; Step S6: extracting signal features from the image data pre-processed in step S4 and inputting the extracted features into the trained athlete action behavior detection model; Step S7: Obtain the athlete's action behavior detection results, combine them with the established action scoring indicators, and display the scoring results in a visual manner.

2. A sports action scoring method based on computer vision according to claim 1, characterized in that: In step S1, the standard action frame entering the action training library is selected to identify the key regions of interest of the athletes and limbs, and a feature polyhedron is constructed based on the points in the regions of interest.

3. A sports action scoring method based on computer vision according to claim 1, characterized in that: In step S2, an RGB-D camera array is deployed to capture user actions from multiple perspectives. The RGB-D camera array is combined with an inertial measurement unit (IMU), and a timestamp alignment protocol is used to achieve synchronization between cameras. The capture parameters are dynamically adjusted during the capture process.

4. A sports action scoring method based on computer vision according to claim 3, characterized in that: The specific process of dynamically adjusting the acquisition parameters includes: Step S21, motion speed perception adjustment: When the inertial measurement unit (IMU) calculates the instantaneous speed of the moving human body in real time, when the instantaneous speed exceeds a preset value, the RGB-D camera array automatically reduces the resolution and increases the frame rate; when the human joint bending angle is detected to be greater than a threshold, the local area switches to the 4K super-resolution acquisition model; Step S22, closed-loop control of exposure parameters: establish a light intensity feedback system, and dynamically adjust the aperture and shutter parameters through the photometric consistency loss function; the specific formula of the loss value of the loss function is as follows: ； Where, represents the photometric consistency loss function, Represents the pixel brightness value of the pixel p coordinate point at time t; represents the displacement vector of pixel p, Indicates the preset error threshold; Step S23, motion trajectory prediction compensation: Use the LSTM network to predict joint positions 3 frames in advance, and use the predicted values ​​for interpolation when data loss is detected; Step S24, environmental interference compensation: identifying environmental interference sources through background subtraction algorithm; Step S25, dynamic resource scheduling: constructing a mapping relationship module between acquisition parameters and computing resources; Step S26, abnormal state self-repair: establish the health evaluation index of the RGB-D camera array.

5. A sports action scoring method based on computer vision according to claim 4, characterized in that: When the value of any of the RGB-D cameras is less than the preset value, the other RGB-D cameras are automatically switched and recalibrated; the health evaluation index judgment formula is as follows: 。 6. A sports action scoring method based on computer vision according to claim 1, characterized in that: In step S3, the video image information preprocessing process is as follows: Step S31, HSV color space conversion: convert the video image from RGB color space to HSV color space for component extraction; the specific formula for converting RGB to HSV model is: ； Where H, S, and V represent the hue, saturation, and brightness of the HSV color space model respectively; Step S32: Setting the tone parameters during the conversion process. The specific formula is: ； Where, represents the hue parameter, Indicates a modular operation with a modulus of 6; Step S33: The calculation formula of the intermediate variable in the conversion process is: ； Where P, Q, and T are intermediate variables in the color conversion space, and F is the hue component.

7. A sports action scoring method based on computer vision according to claim 1, characterized in that: In step S5, the specific process of the computer establishing the athlete action behavior detection model is as follows: Step S51: feature labeling of standard actions in the action training library, and performing enhancement operations on the obtained feature data to obtain action training samples; Step S52: Establishing the athlete action behavior detection model network structure, configuring training data and using YOLOv5 for modeling; Step S53: Set the initial learning rate of the model to , the batch size is set to 256 and the number of iterations is set to 250; Step S54: Use cosine annealing learning rate dynamic adjustment to control learning rate changes; Step S55: Use the pre-trained backbone neural network to perform transfer learning, and then complete the target detection through feature extraction and regression operations.

8. A sports action scoring method based on computer vision according to claim 1, characterized in that: In step S5, the athlete action behavior detection model is provided with evaluation indicators during the training process to evaluate the performance of the model and the training effect. The specific calculation formula is as follows: ； Where, Indicates accuracy, TP indicates the number of samples correctly identified as positive examples, and FP indicates the number of samples incorrectly identified as positive examples. Represents the recall rate, FN represents the number of samples that are mistakenly identified as negative examples, AP represents the average precision of samples, Represents a sample set, and mAP represents the mean average precision of the samples.

9. A sports action scoring method based on computer vision according to claim 1, characterized in that: In step S6, a dataset of pre-processed image data is extracted: ； Where, represents the information recognition condition in the RGB video dataset, represents the information recognition condition in the RGB-D depth video dataset, Indicates the change in motion posture at the joint point, Represents the dynamic action coefficient.

10. A sports action scoring method based on computer vision according to claim 9, characterized in that: In step S6, the signal feature extraction formula of the data set is: ； Where, represents the initial recognition coefficient of the dynamic posture node, represents the maximized recognition coefficient of the dynamic posture node, Represents the dimensional feature value of the athlete's dynamic posture signal, Represents the established motion posture data, represents the signal recognition coefficient based on deep learning algorithm, Vector representing the overlap of line segments in the athlete action detection model.

11. A sports action scoring system based on computer vision, characterized in that: Includes an RGB-D camera array and a computer; The RGB-D camera array is connected to a computer via an information transmission line; The RGB-D camera array is used to collect video image information of athletes from different angles; the computer includes a preprocessing module and a feature extraction module; Model training module, action scoring indicator building module, scoring module and visualization display module; The preprocessing model is used to preprocess the video image information collected by the RGB-D camera array; the feature extraction module is used to extract features from the preprocessed video image; The training module is used to train the athlete's action behavior detection model; the action scoring index construction module is used to set the data elements of each evaluation index; the scoring module is used to score the output results of the athlete's action behavior detection model; and the visualization display module is used to visualize the athlete's action behavior detection results.

Images