AI parallel bars exercise teaching system and method

Abstract

The application relates to the technical field of artificial intelligence and sports teaching, in particular to an AI parallel bars movement teaching system and method. The AI parallel bars movement teaching system comprises a parallel bar positioning sub-module, a protective pad detection sub-module, a judgment module and an ROI region detection module; the parallel bar positioning sub-module is used for recognizing the 2D coordinates of the parallel bars by adopting a neural network model, and calculating whether the parallel bars are inclined and whether the parallel bars are within a safety threshold range; the protective pad detection sub-module is used for detecting the coverage area of the protective pad by using a semantic segmentation network, and judging whether the coverage area is within the safety threshold range; the judgment module is used for judging whether environmental safety detection is passed; and the ROI region detection module is used for confirming the ROI region in image collection of the parallel bar test link. The AI parallel bars movement teaching system and method provided by the application are used for environmental safety detection-ROI region detection-human body posture detection, and a time sequence action posture analysis judgment module is used to output the number of standard parallel bar actions completed by a user and non-standard action feedback.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and physical education, and in particular to an AI parallel bars exercise teaching system and method. Background Technology

[0002] In recent years, physical education has received increasing attention, and physical fitness tests have become an essential part of student assessments. With the development of AI technology, some AI-based methods for detecting physical movements have been applied to sports such as the horizontal bar and calisthenics. However, existing technologies have the following drawbacks:

[0003] The equipment is complex: it requires the combination of multiple sensors such as motion capture equipment and infrared cameras, resulting in high deployment costs. For example, a smart coaching system and method in Chinese Patent Publication No. CN109011508A specifically collects user movement posture information and vital sign index data during user movement. It can establish a movement model and compare and analyze it in real time, and provide training evaluation and suggestions for user movement. Thus, it can provide real-time training and guidance for user movement without the need for human intervention.

[0004] Poor scenario adaptability: Existing systems do not consider limb obstruction during parallel bar exercises (such as the torso being obstructed when gripping the bar with both hands) and interference from multiple people training simultaneously. For example, a smart guidance system for pull-ups published in Chinese Patent Publication No. CN113694501A specifically captures the user's real-time motion data when performing pull-ups, then analyzes it based on stored standard motion data to determine whether the user's motion is standard. When the user's motion is not standard, it guides the user to adjust the motion through display enhancement technology. This ensures that the user can promptly identify and correct their own motion errors during training, improve their training effect, avoid physical injury caused by non-standard motion, and protect the user's health and safety.

[0005] Lack of safety inspection: Intelligent detection was not performed on safety factors such as the fixed state of the parallel bars and the arrangement of protective pads;

[0006] Insufficient robustness of algorithms: Traditional posture comparison algorithms are easily affected by differences in perspective (such as the standard action library being unable to cover individual body shape differences, leading to misjudgment), such as a comprehensive monitoring method and system for motor ability in Chinese patent publication number CN110464356A and an intelligent coaching system and method in Chinese patent publication number CN109011508A.

[0007] To address this, an AI-powered parallel bar exercise teaching system and method were designed, primarily focusing on the intelligent detection of parallel bar movements, including forward and backward swings, forward swing down, and arm flexion and extension. Traditional physical fitness tests rely on manual judgment by teachers. However, parallel bar movements are relatively fast, and manual judgment and counting are prone to misjudgment and errors, resulting in low efficiency and a large workload. Existing AI parallel bar detection methods only detect arm flexion and extension (as described in Chinese Patent Publication No. CN116392798B, "An Automatic Testing Method, Device, Equipment, and Medium for Arm Flexion and Extension on Parallel Bars"). Due to issues such as obstruction, the judgment of lower body movements during parallel bar exercises is even more difficult, and current work lacks in-depth analysis and research. Summary of the Invention

[0008] Therefore, it is necessary to provide an AI parallel bar exercise teaching system and method to address the aforementioned technical problems.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0010] An AI-powered parallel bars exercise teaching system includes a parallel bars positioning submodule, a protective pad detection submodule, a judgment module, and a Region of Interest (ROI) detection module.

[0011] The parallel bar positioning submodule is used to identify the 2D coordinates of the parallel bars using a neural network model, and to obtain the 3D coordinates of the parallel bars by combining the information provided by the binocular / depth camera, and to calculate whether the parallel bars are tilted and whether they are within the safe threshold range.

[0012] The protective pad detection submodule is used to detect the coverage area of ​​the protective pad using a semantic segmentation network, calculate its area ratio with the standard protective area, and determine whether it is within the safety threshold range.

[0013] The judgment module is used to determine whether the environmental safety test has passed based on the detection results of the double bar positioning submodule and the protective pad detection submodule.

[0014] The ROI region detection module is used to identify the ROI region during image acquisition in the double bar test phase.

[0015] It also includes a human posture detection module for detecting students' postures. The human posture detection module uses the YOLOv8-Pose neural network to detect human postures and generate a human skeleton map.

[0016] An AI-powered method for teaching parallel bars exercise, with the following steps:

[0017] S1: Based on the information collected by the double bar positioning submodule, calculate whether the double bars are tilted and whether they are within the safe threshold range;

[0018] S2: Determine whether the area ratio calculated by the protective pad detection submodule is within the safety threshold range;

[0019] S3: Based on the image acquisition by the ROI region detection module, determine the target student for the double bar test;

[0020] S4: Input the normalized coordinates of different joints into the timing motion discrimination module to judge the score of the parallel bars motion.

[0021] In a preferred embodiment of the AI ​​parallel bars exercise teaching method provided by the present invention, step S1 includes the following steps:

[0022] a. Collect images of the double bars and label the target bounding boxes and two endpoints of the double bars as input. After fine-tuning and training based on the Yolov8-pose model, the detection bounding boxes and two endpoints of the double bars can be output.

[0023] b. Given the intrinsic and extrinsic parameters of the binocular camera, perform stereo correction on the images generated by the left and right binocular cameras;

[0024] c. Input the corrected left and right images into the fine-tuned model, and the four endpoints of the double bar can be obtained from each image;

[0025] d. For each pair of points in the left and right images, perform triangulation to obtain their 3D points;

[0026] e. After obtaining the coordinates of the four endpoints, calculate the direction vector of the double bars;

[0027] f. Using the triangulation operation in step d, perform 3D reconstruction of the entire scene, then perform planar fitting on the point cloud of the entire scene to extract the ground normal vector. ;

[0028] g. Calculate the direction vector of the double bars and ground normal vector The included angle.

[0029] In a preferred embodiment of the AI ​​parallel bar exercise teaching method provided by the present invention, in step c, the four endpoints of the parallel bars are the left endpoints of the first bar in the left image of the binoculars. and the right endpoint The left end point of the second bar and the right endpoint The left end point of the first bar in the right image of the binoculars. and the right endpoint The left end point of the second bar and the right endpoint .

[0030] As a preferred embodiment of the AI ​​parallel bar exercise teaching method provided by the present invention, in step d, taking the left end point of the first bar as an example, the input pair of points is... and Calculate the disparity between two points. Then, the parallax is converted into three-dimensional coordinates using the principle of triangulation. Repeat the steps to obtain the three-dimensional coordinates of the other three endpoints of the double bars.

[0031] In a preferred embodiment of the AI ​​parallel bar exercise teaching method provided by the present invention, after obtaining the coordinates of the four endpoints in step e, the direction vector of the parallel bars is calculated, as shown in the following expression:

[0032] .

[0033] In a preferred embodiment of the AI ​​parallel bars exercise teaching method provided by the present invention, step S2 includes the following steps:

[0034] A. Obtain the 3D coordinates of the four endpoints of the parallel bars and the normal vector of the ground. Project the 3D coordinates of the four endpoints of the parallel bars onto the ground. Connect the four projection points to form a rectangle. Extend the rectangle outward by 0.6-1 meters as the standard protection rectangle area A.

[0035] B. Input the left image of the image, use the U-Net network to segment out the protective pad area, then calculate the bounding rectangle to obtain the 3D coordinates of the four vertices of the rectangle, and use the method in step A to obtain its projection rectangle B on the ground;

[0036] C. Calculate the overlapping area of ​​the protected area rectangle A and the protected pad area rectangle B, and calculate the ratio of its area to the area of ​​rectangle A to determine whether it is safe.

[0037] As a preferred embodiment of the AI ​​parallel bar exercise teaching method provided by the present invention, in step S3, the ROI region detection module is used to confirm the ROI region in the image acquisition of the parallel bar test, and the steps are as follows:

[0038] 1) Use the Yolov8 face detection network to find multiple faces in an image;

[0039] 2) Inferring the location of a person based on the position of their face;

[0040] 3) Calculate the overlapping area between the rectangular projection and the parallel bars, and find the bounding box projection with the largest overlapping area, which is the target student to be tested on the parallel bars.

[0041] It is clear without a doubt that the technical solution described above in this application can solve the technical problem that this application aims to address.

[0042] Meanwhile, through the above technical solutions, the present invention has at least the following beneficial effects:

[0043] 1. The present invention provides an AI parallel bar exercise teaching system and method, which includes environmental safety detection, ROI region detection, and human posture detection. It also uses a time-series motion posture analysis and discrimination module to output the number of standard parallel bar movements completed by the user and feedback on non-standard movements. In addition, for occluded environments, it combines the parallel bar position and face detection algorithm to delineate ROI regions to resist the influence of dynamic background environments.

[0044] 2. This invention proposes a multi-layer detection framework for environment and target to ensure the safe conduct of parallel bar exercises. Compared with single-frame discrimination, the results are more robust through temporal action discrimination. At the same time, interpretable action judgment analysis is helpful to quantify the shortcomings of students' parallel bar movements. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0048] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0049] It should be noted that, unless otherwise specified, the embodiments and features and technical solutions in the embodiments of the present invention can be combined with each other.

[0050] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0051] Example 1

[0052] Reference Figure 1An AI parallel bar exercise teaching system and method, including a parallel bar positioning submodule, a protective pad detection submodule, a judgment module, and an ROI region detection module;

[0053] The parallel bar positioning submodule uses a neural network model (such as the YOLO model) to identify the 2D coordinates of the parallel bars, combines this with information from binocular / depth cameras to obtain the 3D coordinates of the parallel bars, and calculates whether the parallel bars are tilted and whether they are within a safe threshold range. The specific calculation steps are as follows:

[0054] a. Acquire images of the parallel bars and label their bounding boxes and endpoints as input. After fine-tuning and training based on the Yolov8-pose model, the detection boxes and endpoints of the parallel bars can be output. Although both the detection boxes and endpoints are output, only the endpoints need to be processed in the following steps.

[0055] b. Given the intrinsic parameters (focal length) of the stereo camera. , light center Using distortion parameters and extrinsic parameters (rotation matrix R and translation matrix T), stereo correction is performed on the images generated by the left and right binocular cameras;

[0056] c. Input the corrected left and right images into the fine-tuned model. The four endpoints of the double bar can be obtained from each image. That is, the left endpoint of the first bar in the left image of the binocular view. and the right endpoint The left end point of the second bar and the right endpoint The left end point of the first bar in the right image of the binoculars. and the right endpoint The left end point of the second bar and the right endpoint .

[0057] d. For each pair of points in the left and right images, perform triangulation to obtain its 3D points. Taking the left endpoint of the first bar as an example, the input pair of points is... and Calculate the disparity between two points. Then, the parallax is converted into three-dimensional coordinates using the principle of triangulation. The expression is as follows:

[0058] ;

[0059] ;

[0060] ;

[0061] Using the same method, the three-dimensional coordinates of the other three endpoints of the parallel bars can be obtained.

[0062] e. After obtaining the coordinates of the four endpoints, the direction vector of the double bars can be calculated.

[0063]

[0064] f. Through the triangulation operation in step d, the entire scene can be reconstructed in 3D. Then, the point cloud of the entire scene is fitted with a plane to extract the ground normal vector. .

[0065] g. Calculate the direction vector of the double bars and ground normal vector The angle between the parallel bars should be as close to 90 degrees as possible. If the angle is less than a certain threshold, it indicates that the parallel bars are tilted, posing a safety risk and requiring adjustment.

[0066] The protective pad detection submodule is used to detect the coverage area of ​​the protective pad using a semantic segmentation network (such as U-Net), calculate its area ratio with the standard protective area, and determine whether it is within the safety threshold range.

[0067] A. The 3D coordinates of the four endpoints of the parallel bars and the normal vector of the ground were obtained. The 3D coordinates of the four endpoints of the parallel bars were projected onto the ground. The four projection points were connected to form a rectangle. The rectangle was expanded outward by a certain distance (different schools have different regulations) as the standard protected rectangular area A.

[0068] Specifically, the rectangle is expanded outward by 0.6 to 1.0 meters, preferably 0.8 meters, along the outer perimeter of the double bars, as the standard protected rectangular area A. The expansion distance can be parameterized according to the usage scenario.

[0069] B. Input the left image of the image, segment the protective pad area using the U-Net network, then calculate the bounding rectangle, and use a similar method in the double-bar positioning submodule to obtain the 3D coordinates of the four vertices of the rectangle. Then, use the method in step A to obtain its projection rectangle B on the ground.

[0070] C. Calculate the overlapping area of ​​the protected area rectangle A and the protected pad area rectangle B, and calculate the ratio of the overlapping area to the area of ​​rectangle A. If this ratio is greater than 80%, the area is considered safe. The overlapping area can be calculated using the following method:

[0071] The coordinates of the top left corner of rectangle A bottom right corner coordinates The coordinates of the top left corner of rectangle B bottom right corner coordinates ;

[0072] Top left corner coordinates of the overlapping region , The coordinates of the lower right corner are , .

[0073] The area of ​​the overlapping region is .

[0074] The judgment module is used to determine whether the environmental safety test has passed based on the detection results of the double bar positioning submodule and the protective pad detection submodule. If the environmental safety test fails, feedback will be sent to the administrator via the display screen. If it passes, the test will proceed to the area detection module.

[0075] The ROI (Region of Interest) detection module is used to identify ROIs in the image acquisition during the parallel bar test. Considering that there may be two safety officers and students who do not leave the testing area promptly after the previous test, multiple students may be present within the image acquisition range during the actual parallel bar test. The steps to determine the ROI in the image are as follows:

[0076] 1) Use the Yolov8 face detection network to find multiple faces in an image. (Input the image into the Yolov8 network, and the output is a bounding box with confidence and target category label. Extract the bounding box labeled "Face" with a confidence score greater than 0.6 to detect faces in the image.)

[0077] 2) Estimate the approximate size of the human body based on the position of the face. Based on camera parameters and the relative pose of the camera and the parallel bars, give an approximate 2D bounding box size of h*w for the human body. The calculation of the 3D coordinates of the parallel bars has already been introduced in the environmental safety inspection steps. Assuming the 3D bounding box of the human body is 1.8 meters high, and 1 meter wide and thick, calculate the center coordinates based on the coordinates of the four endpoints of the parallel bars. The coordinates of the eight vertices of the 3D bounding box of the human body are obtained based on the center point. The coordinates of the bottom four vertices are as follows: , , , The coordinates of the top four vertices are respectively , , , These vertices are projected onto the image based on the camera's intrinsic and extrinsic parameter matrices, using the following formula:

[0078] ;

[0079] Where the rotation matrix The size of the matrix is ​​3x3, and the size of the translation matrix t is 3x1. This represents any point among the aforementioned vertices. Projecting this onto the image yields eight points, which can be connected to form a polygon. The bounding rectangle of this polygon is calculated (using a common method in computer vision, already implemented in OpenCV, and will not be elaborated here), yielding its size h*w. The units of h and w are pixels.

[0080] Although the detected face may not be exactly in the center of the double bars, the size of its bounding box on the image can be approximated by the h*w dimension.

[0081] 3) Calculate the overlapping area between the projection and the parallel bars, and find the bounding box projection with the largest overlapping area, which is the target student for the parallel bar test. Expand the ROI region for human detection based on the union of the bounding box of the target student and the parallel bar region.

[0082] The above steps yield approximate sizes of the face detection bounding box and the body bounding box on the image, which can be used to infer the body bounding box for each person in the image. Assume the top-left pixel of the face detection bounding box is... The pixel in the upper right corner is The coordinates of the four vertices of the human body bounding box are respectively , , , .

[0083] The method for calculating the overlapping area of ​​the human bounding box rectangle and the double-bar rectangle is the same as step B in the environment detection framework, and will not be repeated here. Find the human bounding box with the largest overlapping area, and extend its column coordinates to the left and right endpoints of the double bars. , , , .

[0084] 4) Considering that the safety officer is also within the ROI area, after human posture detection, the student performing the parallel bar exercise is finally identified by the relative position of their arms to the parallel bars. Specifically, the student performing the parallel bar exercise is the one whose arms are perpendicular to the parallel bars.

[0085] After human pose detection, the pixel coordinates of each skeletal key point can be obtained. The vector formed by connecting the pixels of the wrist and shoulder is used as the arm vector. In the environmental safety detection step, the pixel coordinates of the left and right endpoints of the double bars have been obtained. Subtracting these coordinates yields the direction vector of the double bars in the pixel plane. ,calculate and If the angle between the arms and the parallel bars is close to 90 degrees (the error between the two angles is less than 10 degrees), then the arms are considered to be perpendicular to the parallel bars.

[0086] 5) If the ROI area is too small (less than one-quarter of the image), the region detection is considered to have failed and the result is uploaded and reported to the administrator.

[0087] Example 2

[0088] Based on the above embodiment one, the method also includes a human posture detection module for detecting the student's posture, and the specific steps are as follows:

[0089] Step 1: Data input and preprocessing, the steps are as follows:

[0090] The input data is a video stream with a defined ROI region acquired from a fixed-view binocular / depth camera.

[0091] Human pose estimation uses a lightweight and accurate multi-person pose estimation network (such as YOLOv8-Pose).

[0092] Output the 2D pixel coordinates and confidence level of 17 key points of the human body (including: nose, eyes, ears, shoulders, elbows, wrists, hips, knees, ankles, heels, toes, etc.) frame by frame;

[0093] To eliminate the influence of different student heights, body shapes, and distances from the camera, all joint coordinates were converted to relative coordinates in a human coordinate system with the hip center as the origin, and then scaled.

[0094] Step 2: Temporal action recognition and sub-action segmentation;

[0095] Step 1: Construct a time series model, the steps are as follows:

[0096] Construct a Transformer encoder or a bidirectional LSTM (Bi-LSTM) network as a temporal classifier;

[0097] A normalized sequence of coordinates for all joints within a time window (e.g., 30 frames, approximately 1 second). The input for each time step is a vector containing the (x, y) coordinates of all joints.

[0098] The probability that the time window is classified into a certain sub-action category (such as: take-off, straddle sit on the bar, forward swing, backward swing, landing, etc.).

[0099] Sub-motion segmentation is specifically achieved by calculating the geometric relationships between joints based on the principles of human motion biomechanics, for example:

[0100] The take-off action is calculated based on:

[0101] When the vertical velocity of the center of gravity changes abruptly: calculate the rate of change of height at the midpoint of the ankles and hips, and an upward peak will appear;

[0102] When both feet are off the ground: the height of both ankle joints is higher than the ground threshold;

[0103] The auxiliary criterion is that the wrist joint is close to the parallel bars area;

[0104] The calculation basis for sitting with legs apart on the bar is:

[0105] Leg posture: Calculate the distance between the two knee joints, which should exceed a certain proportion of shoulder width (e.g., 1.2 times).

[0106] Body posture: The hip height is stable near the height of the parallel bars, and the angle between the torso (the line connecting the shoulders and hips) and the horizontal plane is small.

[0107] The auxiliary criterion is that the angle between the thigh (hip-knee line) and the torso is close to 90 degrees.

[0108] The calculation basis for the forward swing is:

[0109] Hip trajectory: The hip joint swings from back to front, reaching its highest point.

[0110] Leg posture: The knee joint angle is increased (leg straight), and the ankle joint trajectory is a forward and upward arc.

[0111] The auxiliary criterion is that the angle between the arm (the line connecting the shoulder and wrist) and the direction vector of the parallel bars should be close to perpendicular.

[0112] The calculation basis for the backswing is:

[0113] Hip trajectory: The hip joint swings from front to back, reaching its highest point.

[0114] Body posture: The angle of the shoulder joint changes, and the body presents a slight "reverse arch" posture.

[0115] The auxiliary criterion is that the legs tend to swing backward.

[0116] Step 2: Keyframe extraction, the steps are as follows:

[0117] Within a sub-action time period identified by the time-series model, keyframes are automatically extracted based on the scoring criteria of that sub-action.

[0118] Taking the previous swing as an example:

[0119] Starting frame: The instant when the hip begins to move forward from its lowest point.

[0120] In-between frame: The posture when the body is positioned vertically below the bar, checking whether the body is straight.

[0121] Peak frame: Swing your hips to the highest point, and check the leg opening angle and body extension at this time.

[0122] End frame: The moment when the forward swing ends and the backward swing begins.

[0123] Step 3: Score based on the extracted keyframes, as follows:

[0124] Extension: At the highest point, is the body fully extended (are the shoulder, hip, and ankle approximately in a straight line)? Calculate the degree to which the angle between the shoulder, hip, and ankle joints deviates from 180 degrees.

[0125] Swing height: The absolute height of the hip joint at the highest point of the forward swing (or relative to the height of the parallel bars).

[0126] Leg posture: The average flexion angle of the knee joint throughout the entire forward swing.

[0127] Movement stability: The variance of the angle between the arms and the parallel bars; the smaller the variance, the more stable the support.

[0128] Based on the extracted keyframe locations and the degree of deviation from the ideal value, a score of 0-10 is given;

[0129] Each factor is assigned a different weight based on its importance in the action, and the weighted average score of the sub-action is calculated. For example, if the weight of the swing height is 2, then the weighted average score needs to be calculated by multiplying the swing height score by 2.

[0130] To meet the needs of instructors and students, in addition to keyframes, every frame in the entire video was also labeled with a human skeleton diagram and then output for frame-by-frame playback.

[0131] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An AI-powered parallel bars exercise teaching system, characterized in that, This includes a double-bar positioning submodule, a protective pad detection submodule, a judgment module, and an ROI area detection module; The parallel bar positioning submodule is used to identify the 2D coordinates of the parallel bars using a neural network model, and to obtain the 3D coordinates of the parallel bars by combining the information provided by the binocular / depth camera, and to calculate whether the parallel bars are tilted and whether they are within the safe threshold range. The protective pad detection submodule is used to detect the coverage area of ​​the protective pad using a semantic segmentation network, calculate its area ratio with the standard protective area, and determine whether it is within the safety threshold range. The judgment module is used to determine whether the environmental safety test has passed based on the detection results of the double bar positioning submodule and the protective pad detection submodule. The ROI region detection module is used to identify the ROI region during image acquisition in the double bar test phase. It also includes a human posture detection module for detecting students' postures. The human posture detection module uses the YOLOv8-Pose neural network to detect human postures and generate a human skeleton map.

2. An AI parallel bars exercise teaching method, used in the AI ​​parallel bars exercise teaching system described in claim 1, characterized in that, The steps are as follows: S1: Based on the information collected by the double bar positioning submodule, calculate whether the double bars are tilted and whether they are within the safe threshold range; S2: Determine whether the area ratio calculated by the protective pad detection submodule is within the safety threshold range; S3: Based on the image acquisition by the ROI region detection module, determine the target student for the double bar test; S4: Input the normalized coordinates of different joints into the timing motion discrimination module to judge the score of the parallel bars motion.

3. The AI ​​parallel bars exercise teaching method according to claim 2, characterized in that, In step S1, the steps are as follows: a. Collect images of the double bars and label the target bounding boxes and two endpoints of the double bars as input. After fine-tuning and training based on the Yolov8-pose model, the detection bounding boxes and two endpoints of the double bars can be output. b. Given the intrinsic and extrinsic parameters of the binocular camera, perform stereo correction on the images generated by the left and right binocular cameras; c. Input the corrected left and right images into the fine-tuned model, and the four endpoints of the double bar can be obtained from each image; d. For each pair of points in the left and right images, perform triangulation to obtain their 3D points; e. After obtaining the coordinates of the four endpoints, calculate the direction vector of the double bars; f. Using the triangulation operation in step d, perform 3D reconstruction of the entire scene, then perform planar fitting on the point cloud of the entire scene to extract the ground normal vector. ; g. Calculate the direction vector of the double bars and ground normal vector The included angle.

4. The AI ​​parallel bars exercise teaching method according to claim 3, characterized in that, In step c, the four endpoints of the double bars are the left endpoints of the first bar in the left image of the binoculars. and the right endpoint The left end point of the second bar and the right endpoint The left end point of the first bar in the right image of the binoculars. and the right endpoint The left end point of the second bar and the right endpoint .

5. The AI ​​parallel bars exercise teaching method according to claim 3, characterized in that, In step d, taking the left endpoint of the first bar as an example, the input pair of points is: and Calculate the disparity between two points. Then, the parallax is converted into three-dimensional coordinates using the principle of triangulation. Repeat the steps to obtain the three-dimensional coordinates of the other three endpoints of the double bars.

6. The AI ​​parallel bars exercise teaching method according to claim 3, characterized in that, After obtaining the coordinates of the four endpoints in step e, the direction vector of the double bars is calculated as follows: 。 7. The AI ​​parallel bars exercise teaching method according to claim 2, characterized in that, In step S2, the steps are as follows: A. Obtain the 3D coordinates of the four endpoints of the parallel bars and the normal vector of the ground. Project the 3D coordinates of the four endpoints of the parallel bars onto the ground. Connect the four projection points to form a rectangle. Extend the rectangle outward by 0.6-1 meters as the standard protection rectangle area A. B. Input the left image of the image, use the U-Net network to segment out the protective pad area, then calculate the bounding rectangle to obtain the 3D coordinates of the four vertices of the rectangle, and use the method in step A to obtain its projection rectangle B on the ground; C. Calculate the overlapping area of ​​the protected area rectangle A and the protected pad area rectangle B, and calculate the ratio of its area to the area of ​​rectangle A to determine whether it is safe.

8. The AI ​​parallel bars exercise teaching method according to claim 7, characterized in that, In step S3, the ROI region detection module is used to confirm the ROI region in the image acquisition during the double-bar test. The steps are as follows: 1) Use the Yolov8 face detection network to find multiple faces in an image; 2) Inferring the location of a person based on the position of their face; 3) Calculate the overlapping area between the rectangular projection and the parallel bars, and find the bounding box projection with the largest overlapping area, which is the target student to be tested on the parallel bars.

Images