Self-service table tennis ball learning method and system

Abstract

The invention provides a self-service table tennis ball learning method and system, and relates to the technical field of physical training, personalized training is provided through a complete user journey from reservation to feedback, automatic supervision and material management are realized by using a visual technology, a recovery task is gamified to improve user stickiness, and user experience is improved. And a social sharing function is combined to enhance the propagation effect, so that the training specialty is improved and the operation cost is remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of sports training technology, specifically to a self-help method and system for learning table tennis. Background Technology

[0002] Current table tennis training primarily relies on manual coaching or traditional automatic ball-serving machines. The former is costly and inflexible in terms of timing; the latter has limited functionality and lacks guidance, engagement, and intelligent management. Furthermore, collecting scattered table tennis balls after training increases the burden on users and administrators, and the consumption of materials during training (such as missing table tennis balls) cannot be detected and replenished in a timely manner, impacting the user experience. Summary of the Invention

[0003] To solve the above-mentioned technical problems, this invention provides a self-help method and system for learning to play table tennis.

[0004] The technical solution adopted in this invention is as follows:

[0005] A self-taught method for learning table tennis includes the following steps:

[0006] S1. In response to the user's identity verification request initiated by the user at the entrance of the training cabin through the identity recognition and access control system, obtain the user's identity information; bind the identity information with the current training session, and control the access control system to open;

[0007] S2. Receive the user's selection instruction for the teaching video of the target technical action on the interactive display terminal; based on the selection instruction, query the standard serving parameter set that matches the target technical action from the pre-stored technical action and serving parameter mapping database;

[0008] S3. Send the standard serving parameter set to the automatic serving machine; control the automatic serving machine to serve a ping-pong ball to the table according to the standard serving parameter set;

[0009] S4. Through the smart camera deployed in the learning chamber, video stream data including the user's hitting action and the landing point of the return ball is collected in real time; based on the video stream data, the landing state of each return ball is identified; according to the identification result of the landing state of each return ball, real-time audio and visual feedback is provided to the user through the feedback device.

[0010] S5. Perform human joint point recognition on the user's hitting action in the video stream data to generate a user motion skeleton model; compare and analyze the user motion skeleton model with a preset standard motion model corresponding to the target technical action to obtain motion deviation analysis results; generate a training report based on the ball return recognition results and the motion deviation analysis results.

[0011] S6. At a preset time before the end of the current training session, output a reminder message to the user to return the scattered ping-pong balls to the ball storage compartment of the automatic ball serving machine;

[0012] S7. After the current training session is detected to have ended, the environment inside the learning ball cabin is captured by the smart camera; the environment image is analyzed to identify the number of ping-pong balls scattered on the floor and table of the learning ball cabin; if the number of identified ping-pong balls is lower than the preset return completion threshold, a preset reward is issued to the user account bound to the identity information.

[0013] S8. Associate the training report, the action deviation analysis results, and the ball return recognition results with the identity information and upload them to the cloud server; based on the smart camera and / or the sensors in the automatic ball serving machine, monitor the total number of ping-pong balls in the ball chamber and send an alarm message to the management backend when the total number of ping-pong balls is lower than a preset threshold.

[0014] In step S1, obtaining the user's identity information specifically involves receiving user reservation data forwarded by the reservation server, which includes the verified user's identity identifier and the usage period.

[0015] Step S2 specifically includes: determining the target action ID according to the selection instruction, and querying the mapping database using the target action ID as an index to obtain a record of serving parameters including at least the landing point, spin type, spin speed, and serving frequency.

[0016] Step S4, which identifies whether each return of the ball successfully lands on the table based on the video stream data, specifically includes:

[0017] S401: Extract the spatiotemporal trajectory sequence of the ping-pong ball and the user's action sequence from the video stream data; S402: Determine whether the ping-pong ball touches the table within the effective area of ​​the table; S403: If so, further determine, based on the spatiotemporal trajectory sequence and the action sequence, whether there is an illegal secondary interaction event from the first ball-bouncing interaction to the second touch of the table, the illegal secondary interaction event includes: the ping-pong ball contacting a non-paddle part of the user's body, or the user's paddle performing a consecutive hitting action on the ping-pong ball; S404: If the illegal secondary interaction event exists, determine that the return shot is invalid; if the illegal secondary interaction event does not exist, determine that the return shot is valid; if the ping-pong ball does not touch the table within the effective area of ​​the table, determine that the return shot is not valid.

[0018] Step S403, determining whether the user's racket has performed a double-hit action on the ping-pong ball, includes:

[0019] S4031: Identify the first time point of contact between the ping-pong ball and the racket from the video stream data. S4032: Track the trajectory of the ping-pong ball after its initial separation, and continuously monitor the spatial pose of the racket based on the user's action sequence; S4033: Determine whether there is a second time point in the trajectory of the ping-pong ball before its second contact with the table that is abnormally close to or makes re-contact with the trajectory of the racket. S4034: If the second time point exists and time interval The number of consecutive hits is less than the preset combo threshold, and the ping-pong ball is in to If the trajectory changes during the stroke do not conform to the physical dynamics model of a single shot, it is determined that a double shot has occurred.

[0020] Step S4034, determining whether the trajectory change conforms to the physical dynamics model of a single shot, includes:

[0021] Based on the first time point The velocity vector of the incoming ping-pong ball The velocity vector of the racket at the moment of impact And a preset collision recovery coefficient, to calculate the ping-pong ball in Theoretical launch velocity vector after time step The theoretical launch velocity vector Compared with what is actually observed from video stream data, in The initial velocity vector of the ping-pong ball after time step Compare; calculate and Direction angle Ratio of speed to magnitude If the included angle greater than the preset angle deviation threshold and / or the ratio of the speed magnitudes Not within the preset reasonable range If the trajectory change does not conform to the physical dynamics model of a single shot, then it is determined that the change does not conform to the physical dynamics model of a single shot.

[0022] Step S403, determining whether the ball is in contact with a non-racket part of the user's body, includes: constructing a real-time 3D skeletal model of the user; performing spatiotemporal collision detection on the trajectory of the ball and the set of body joints defined as non-racket contactable parts; if spatial overlap is detected after the first ball-bouncing interaction and before the second touch of the table, it is determined that the body is touching the ball.

[0023] Step S4 provides real-time audio-visual feedback based on the stage recognition result, including configuring different audio-visual feedback signals for the three results of valid stage entry, no stage entry, and invalid stage entry.

[0024] When generating the training report in step S5, the number, proportion, and suspected cause classification of invalid return shots are included in the report as independent statistical indicators.

[0025] It also includes S9, receiving a sharing instruction initiated by the user through the interactive display terminal or a mobile terminal associated with the identity information; responding to the sharing instruction, publishing the training report or summary information generated based on the training report to a designated social platform; and / or, uploading the score calculated based on the ball return recognition result to the global or friend leaderboard in the cloud server.

[0026] A self-service table tennis learning system includes:

[0027] The identity verification and access control module is used to respond to the identity verification request initiated by the user at the entrance of the training cabin through the identity recognition and access control system, obtain the user's identity information, bind the identity information with the current training session, and control the access control system to open;

[0028] The instructional video selection and parameter query module is used to receive the user's selection instruction for the instructional video of the target technical action on the interactive display terminal; based on the selection instruction, it queries the standard set of serving parameters that match the target technical action from the pre-stored technical action and serving parameter mapping database;

[0029] An automatic ball-serving machine control module is used to send the standard ball-serving parameter set to the automatic ball-serving machine; and to control the automatic ball-serving machine to serve ping-pong balls to the table according to the standard ball-serving parameter set.

[0030] The ball return status recognition and real-time feedback module is used to collect video stream data containing the user's hitting action and the landing point of the ball in real time through a smart camera deployed in the learning chamber; based on the video stream data, the on-table status of each ball return is recognized; and based on the on-table status recognition result of each ball return, real-time audio-visual feedback is provided to the user through a feedback device.

[0031] The motion analysis and training report generation module is used to identify human joint points in the user's hitting motion in the video stream data and generate a user motion skeleton model; compare and analyze the user motion skeleton model with a preset standard motion model corresponding to the target technical motion to obtain motion deviation analysis results; and generate a training report based on the ball return recognition results and the motion deviation analysis results.

[0032] The reminder module is used to output a reminder message to the user at a preset time before the end of the current training session, prompting them to return the scattered ping-pong balls to the ball storage compartment of the automatic ball serving machine;

[0033] The environmental inspection and reward distribution module is used to collect environmental images inside the learning ball cabin through the smart camera after the current training session is detected to have ended; analyze the environmental images to identify the number of ping-pong balls scattered on the floor and table of the learning ball cabin; if the number of identified ping-pong balls is lower than a preset return completion threshold, a preset reward is issued to the user account bound to the identity information.

[0034] The data upload and device monitoring module is used to associate the training report, the action deviation analysis results, and the ball return recognition results with the identity information and upload them to the cloud server; based on the smart camera and / or the sensors in the automatic ball serving machine, it monitors the total number of ping-pong balls in the ball chamber and sends an alarm message to the management backend when the total number of ping-pong balls is lower than a preset threshold.

[0035] The beneficial effects of this invention are:

[0036] This invention provides personalized training through a complete user journey from appointment to feedback, and uses visual technology to achieve automatic supervision and material management. It gamifies recycling tasks to enhance user engagement and combines social sharing functions to enhance the dissemination effect, thereby significantly reducing operating costs while improving the professionalism of training. Attached Figure Description

[0037] Figure 1 This is a flowchart of a self-learning method for table tennis according to an embodiment of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] like Figure 1 As shown in the figure, a self-learning method for playing table tennis according to an embodiment of the present invention includes the following steps:

[0040] S1. In response to the user's identity verification request initiated by the identification and access control system at the entrance of the training cabin, obtain the user's identity information; bind the identity information with the current training session, and control the access control system to open. Specifically, obtaining the user's identity information in step S1 involves receiving user reservation data forwarded by the reservation server, which includes the verified user's identity identifier and the usage period.

[0041] Specifically, the identity recognition and access control system is deployed at the entrance of the learning cabin, encompassing an identity recognition module (supporting ID card readers, QR code scanning, facial recognition, etc.) and an access control execution module (composed of an electromagnetic lock, infrared sensor, and audio-visual prompt unit). The former is responsible for collecting user verification data, while the latter executes access control actions and provides feedback on the results. The data interaction and association unit is responsible for data transmission and processing with the reservation server and local control system, realizing functions such as receiving reservation data, generating session identifiers, and sending access control commands.

[0042] Upon arrival at the entrance to the learning cabin, users must initiate an active verification request through the identity recognition module. This can be triggered by placing their ID card near the card reader, displaying the reservation verification code on their mobile terminal, or completing facial recognition by facing the face capture device. These actions will prompt the identity recognition module to send an identity verification request signal containing the trigger method identifier and the original verification data to the local control system. The identity information is not obtained directly from the collected raw data, but rather from verified user data forwarded by the reservation server. The process is as follows: the user submits their identity identifier and usage time slot in advance through the reservation channel. After verification by the reservation server, structured reservation data is generated. When the user initiates a verification request, the local control system queries the reservation server for matching data. After successful verification, the reservation server forwards the encrypted reservation data to the local control system, completing the identity information acquisition. The acquired identity information is structured data, containing a unique identity identifier, reserved usage time slot, order validity identifier, and optional auxiliary verification information. The local control system will further perform time validity verification (the current time must be within the reserved time slot and its grace period) and identifier consistency verification (such as facial feature comparison). Only after both verifications are passed can the user proceed to the next stage.

[0043] After successful identity verification, the user must be bound to the current training session. The current training session refers to the complete process from the user entering the training chamber through the access control to the end of the training. The local control system will generate a unique training session for each successful verification. The generation rules can use the SHA-256 hash function combined with user identity, current timestamp, and random number to ensure uniqueness. Binding is achieved by creating a temporary session table in the local database, storing the mapping relationship between session identifier, user identity, reserved time period, and session state, and synchronizing it to the temporary storage area of ​​the cloud server. This binding operation establishes the association between user, session, and device, avoiding confusion in subsequent training data and ensuring data attribution.

[0044] The access control system requires three conditions to be met before it can be opened: successful secondary verification of identity information, completion of identity and session binding, and no ongoing training session in the learning chamber. Once these conditions are met, the local control system can send a high-level signal to the electromagnetic lock for 5 seconds, and the access control will automatically open.

[0045] In a specific embodiment of the present invention, the reservation server and the local control system use the HTTPS encrypted transmission protocol, and the identity information is encrypted with AES-256 to prevent data theft or tampering; the local system only stores the encrypted identity hash value or desensitized information to protect user privacy; at the same time, by limiting the access control opening time and deploying infrared anti-tailgating sensors, unauthorized personnel are prevented from tailgating into the system.

[0046] S2. Receive the user's selection instruction for the teaching video of the target technical action on the interactive display terminal; based on the selection instruction, query the standard serving parameter set that matches the target technical action from the pre-stored technical action and serving parameter mapping database.

[0047] Specifically, step S2 includes: determining the target action ID according to the selection instruction, and querying the mapping database using the target action ID as an index to obtain a record of serving parameters including at least the landing point, spin type, spin speed and serving frequency.

[0048] In this invention, the key technical components of step S2 include an interactive display terminal, a technical action and serve parameter mapping database, and an instruction parsing and parameter query unit. The interactive display terminal, serving as the user's interaction entry point with the system, is deployed in an easily accessible location within the learning cabin. It displays a categorized list of technical action instruction videos (including basic, advanced, and specialized actions), supports touch or voice selection commands, and provides real-time feedback on selection results and core training points. The technical action and serve parameter mapping database is a structured database pre-stored in the local control system. It can employ a key-value pair storage structure of target action ID - serve parameter set, with each record being... ,in This specifies the storage path for the corresponding technical action instruction videos. The instruction parsing and parameter query unit is responsible for parsing the user-selected instructions, generating the target action ID, querying the matching parameter set, and completing parameter verification and transmission.

[0049] After the training session binding is complete (step S1 ends), the local control system will automatically send a loading command to the interactive display terminal. The terminal reads the teaching video paths and names of all technical movements from the mapping database and displays them categorized into basic movements, advanced movements, and specialized movements. Difficulty levels (1-5 stars) can be simultaneously marked to assist users in selecting according to their skill level. Users initiate a selection command by clicking the video icon / name on the terminal, and the terminal generates a message containing the movement name, terminal identifier, and session information. ( The original instructions are transmitted to the local control system via wired or wireless means. After receiving the instructions, the instruction parsing unit generates the target action using the action name-action ID mapping table. ( The corresponding mapping rule expression is: ;

[0050] In the formula, The action category prefix (basic action = 01, advanced action = 02, specialized action = 03) should be a two-digit number. This is the sequence number of the action within its corresponding category, with a two-digit value. For example, the forehand attack is ranked 1st in the basic actions category, so the sequence number is 01. Example: Basic Action - Forehand Attack Advanced moves - corresponding to forehand topspin This mapping table is pre-stored in the instruction parsing unit and is linked to the mapping database. Maintain consistency to ensure index uniqueness.

[0051] The generated target action As an index, the parameter query unit initiates a retrieval from the mapping database to extract the corresponding standard serve parameter set. Among them, the landing point parameters The table tennis table is described using a Cartesian coordinate system, with the bottom left corner of the table closest to the user as the origin. The X-axis corresponds to the width of the table (0-2.74m), and the Y-axis corresponds to the length of the table (0-1.525m). The coordinate error is ≤ ±0.05m. (Spin type) It uses digital coding (0=no rotation, 1=weak topspin, 2=strong topspin, 3=weak backspin, 4=strong backspin, 5=sidespin) to adapt to the training needs of different movements; rotation speed The unit is revolutions per second (r / s), set according to spin type (weak spin 20-50 r / s, strong spin 80-120 r / s), with an accuracy error ≤ ±5 r / s; serve frequency. The unit is times per minute, categorized by difficulty level (basic 10-30, intermediate 30-50, specialized 50-60). After querying and extracting the parameters, the completeness and rationality of the parameter set are checked to ensure that the landing point is within the valid area and the parameters conform to the corresponding type range. If the check passes, the data is encoded and transmitted to the automatic ball-launching machine; if the check fails, a prompt and alarm are issued.

[0052] In a specific embodiment of the present invention, the interactive display terminal can be configured with a gamified training mode, which includes the following steps:

[0053] In response to the user's selection command for the gamified training mode on the interactive display terminal, at least two game types and corresponding difficulty setting options are displayed; the system receives the user's game type selection and difficulty setting commands, and generates a game parameter set including serving parameters, game rule parameters, and interactive feedback parameters; based on the selected game type, the system controls the interactive devices in the training chamber to enter the corresponding game environment; the interactive devices include at least a projector, a smart camera, an automatic ball-serving machine, and an audio-visual feedback device; the system controls the automatic ball-serving machine to serve the ball to the user according to the serving parameters in the game parameter set, and simultaneously starts the game timer; the system uses the smart camera to collect real-time data including the trajectory of the ping-pong ball and the user's... The system collects video stream data of user actions; based on game rule parameters, it identifies game events in the video stream data in real time, including at least: a ping-pong ball hitting a virtual target, the user successfully hitting or dodging the ball, and the user being hit by the ball; based on the game event identification results, it controls the interactive device to provide real-time interactive feedback; the interactive feedback includes at least: changes in projected content, changes in sound and light effects, and changes in score; after the game time ends or the game objective is achieved, it collects game performance data, including at least: score, hit rate, dodge success rate, and time taken; it binds the performance data with the user's identity information and uploads it to the cloud server.

[0054] Based on performance data, update the game leaderboard in the cloud server; the leaderboard includes at least one or more of the following: global leaderboard, friend leaderboard, and historical performance leaderboard; in response to a user's game performance sharing instruction initiated through the interactive display terminal or mobile terminal, publish the performance data or leaderboard information to the designated social platform.

[0055] This invention supports a gamified training mode on an interactive display terminal, aiming to enhance training fun and user engagement through interactive games. After step S2, users can select to enter the gamified training mode. The interactive display terminal provides a game selection interface, displaying icons and descriptions of various game types, such as: target shooting games (e.g., variations of "Shoot the Light Bulb"), where the core gameplay involves hitting virtual targets with a ball; dodgeball games (e.g., "Dodgeball Master"), where users dodge balls fired from an automatic ball-serving machine; challenge games (e.g., "Level Challenge"), which combine serving sequences and rules to complete challenges; and competitive games (e.g., "Man vs. Machine"), which simulates a competition scenario and calculates scores. After selecting a game type, users can further set difficulty parameters, such as serving speed, frequency, number of virtual targets, and game duration. The system generates a structured game parameter set based on the selection. ,in Encode the game type. Difficulty level For a subset of serve parameters, These are the game rule parameters (such as hit score, dodge score, and penalty rules). This includes parameters for interactive feedback (such as sound effects, light colors, and projection animations). Depending on the game type, the system initializes the corresponding interactive devices: for target shooting games, the projector projects virtual target patterns onto the table surface or background wall; the pattern positions can be randomly generated or distributed according to difficulty rules. For dodging games, a smart camera creates a user skeletal model and continuously tracks it, providing real-time aiming coordinates for the automatic ball-launching machine. For level-based or competitive games, the projector and automatic ball-launching machine can be activated simultaneously to create a multimodal interactive scene. After the game starts, the automatic ball-launching machine... The ball is served, and a smart camera captures the video stream in real time. Virtual target hit recognition: A transformation matrix is ​​established between the projection coordinate system and the camera coordinate system to map the virtual target pattern into three-dimensional space; spatial collision detection is performed between the ping-pong ball trajectory and the target area to determine if a hit has occurred. Upon a hit, the system controls the projector to switch patterns (such as a broken light bulb or a mole retreating) and triggers corresponding sound effects and lights.

[0056] S3. Send the standard serving parameter set to the automatic serving machine; control the automatic serving machine to serve ping-pong balls to the table according to the standard serving parameter set.

[0057] The key technical components of step S3 encompass four core modules: a data transmission interface module, an automatic ball-serving machine control module, an automatic ball-serving machine actuator, and a detection and feedback module. The data transmission interface module is responsible for parameter transmission between the local control system and the automatic ball-serving machine. It uses RS485 serial port protocol or Ethernet protocol, transmitting data in a binary encoding format of parameter type identifier + parameter value + check bit, and incorporates a built-in CRC-16 cyclic redundancy check algorithm to prevent data tampering or loss. The automatic ball-serving machine control module, as the control core of the ball-serving machine, integrates parameter parsing, command allocation, coordinated control, and status monitoring functions. It can accurately allocate decoded parameter commands to each actuator and synchronize the action sequence. The automatic ball-serving machine actuator consists of a landing point adjustment mechanism, a spin control mechanism, a ball-serving trigger mechanism, and a power supply mechanism, respectively responsible for controlling the ball landing point, spin state, launch frequency, and power output, completing the ball serve through the coordinated action of mechanical components. The detection and feedback module is deployed inside the automatic ball-serving machine and in the ball-learning chamber. Through position detection sensors, speed detection sensors, and ball-serving count sensors, it collects the action status data of the actuators in real time, providing feedback support for closed-loop control.

[0058] The local control system transmits the standard ball-serving parameter set to the data transmission interface module. The data transmission interface module encodes each parameter one by one according to a preset encoding format, generating binary data frames and sending them to the control module of the automatic ball-serving machine via a transmission protocol. After receiving the data frames, the automatic ball-serving machine control module first verifies the checksum using the CRC-16 algorithm. If the verification passes, it performs decoding. The decoding logic processes the parameters separately according to their types. The landing point coordinates are calculated using the encoded value, offset, and scaling factor to obtain the actual coordinates. The spin type is directly mapped using the encoded value, while the spin speed and ball-serving frequency are obtained by multiplying the encoded value by the corresponding scaling factor, ensuring that the decoded parameters are consistent with the original parameters.

[0059] The actuator achieves precise execution of various parameters through mathematical models. Specifically, for landing point control, a mapping model is established between the serve angle and the landing point on the table, using horizontal rotation... and vertical tilt angle Calculation formula , Adjust the angle of the serving nozzle so that the ping-pong ball touches the table according to the preset coordinates. The initial horizontal position of the serve opening. The horizontal distance from the ball's mouth to the near end of the table. The height of the serve opening, The height of the table tennis table. This is the vertical tilt angle correction coefficient; rotation control achieves different rotation types through the start-stop combination of the rotating motor, and uses PWM speed regulation to control the rotation speed, based on the mapping relationship between the motor speed N and the PWM duty cycle D. ( Motor speed control coefficient, (Minimum starting speed of the motor), combined with rotational speed Conversion formula with motor speed N ( (for the transmission ratio), ensuring rotational speed accuracy; serve frequency control is achieved through serve intervals. ( The trigger interval is calculated for the mechanical action time of a single serve. The electromagnetic push rod is periodically controlled by a timer to ensure a stable serve frequency. Each actuator operates in a coordinated sequence of landing point adjustment, rotation start, and serve trigger to avoid parameter conflicts. In the initial state, all actuators are reset, and then angle adjustment and rotation motor start are completed in sequence. Finally, the serve is triggered at the preset interval.

[0060] For landing point calibration, the intelligent camera can capture the touch point position once for every 10 balls launched, and calculate the actual landing point. Deviation from the preset landing point ( , If the deviation exceeds the threshold, then the angle correction formula is used. , Adjust the angle of the serve, among which, , For correction coefficients; rotational speed calibration can be performed once every 5 balls launched. If the actual rotational speed deviates from the preset value by more than ±3 r / s, the PWM duty cycle is adjusted by 0.5%; frequency calibration counts the actual number of balls launched every minute and adjusts the launching interval according to the deviation from the preset frequency to ensure that the execution accuracy of each parameter is always within the allowable range.

[0061] S4. Through the smart camera deployed in the learning chamber, video stream data including the user's hitting action and the landing point of the return ball is collected in real time; based on the video stream data, the landing state of each return ball is identified; according to the identification result of the landing state of each return ball, real-time audio-visual feedback is provided to the user through the feedback device.

[0062] In this invention, the key technical components of step S4 include a smart camera module, a video stream processing unit, a ball return status determination module, and an audio-visual feedback device. The smart cameras are deployed in key locations within the learning chamber, with no fewer than two units. Their hardware parameters must meet the following requirements: resolution ≥ 1920×1080px, frame rate ≥ 60fps, and a built-in TOF depth sensor supporting three-dimensional spatial coordinate acquisition with a latency ≤ 30ms. Multiple cameras are synchronized via the PTP precise time protocol with a time deviation ≤ 1ms. The video stream processing unit is integrated into the local control system, employing GPU-accelerated computing (such as NVIDIA Jetson Xavier NX). It is responsible for video stream preprocessing, target extraction, and trajectory and motion sequence generation. Gaussian filtering is used for noise reduction, and adaptive contrast adjustment improves data quality. The YOLOv8 algorithm is then used to separate the three core targets: the ping-pong ball, the user's body, and the racket. The ball return status determination module is the core processing unit, incorporating trajectory analysis, violation detection, and validity determination logic. It can complete touch-the-table determination, violation event detection, and final status output. The audio-visual feedback device is deployed in a prominent position inside the learning sphere chamber. It includes a speaker and an RGB LED light group. It can synchronously trigger differentiated audio-visual signals within 100ms after receiving the judgment result to ensure the timeliness of the feedback.

[0063] The initiation of real-time video stream acquisition is linked to the automatic ball-serving action in step S3. After the automatic ball-serving machine starts, the local control system sends an acquisition start command to the smart camera, and the camera focuses on the effective area of ​​the table according to preset parameters. , , ) and user activity area ( , , The system collects data, which is transmitted to the video stream processing unit via gigabit Ethernet in H.265 encoding format, balancing low latency and high compression ratio. Video streams from multiple cameras are aligned using timestamps, and the time offset of other data is adjusted based on the timestamp of the first camera. This fusion process generates 3D spatial data, eliminating single-view occlusion issues such as a user's body obscuring a ping-pong ball, ensuring the integrity of trajectory and motion capture.

[0064] In one embodiment of the present invention, step S4, which identifies the on-table status of each return based on the video stream data, specifically includes:

[0065] S401: Extract the spatiotemporal trajectory sequence of the ping-pong ball and the user's action sequence from the video stream data.

[0066] Among them, the extraction of the spatiotemporal trajectory sequence of the ping-pong ball requires continuously capturing the three-dimensional coordinates of the ping-pong ball from the video stream to generate a spatiotemporal trajectory sequence. ,in For table tennis in time The three-dimensional coordinates (unit: m). The timestamp is in seconds. Ping-pong ball detection first uses a lightweight YOLOv8 model to detect ping-pong balls in each frame, outputting the coordinates of the detection bounding boxes. The two-dimensional coordinates of the ping-pong ball are initially located based on the center position of the detection frame. Combined with the depth values ​​acquired by the TOF depth sensor Converted into three-dimensional coordinates using the camera intrinsic parameter matrix The transformation model is as follows:

[0067] ;

[0068] In the formula, This is the camera's focal length (default 8mm, which is equivalent to 3500px in pixels). The coordinates of the camera principal point (default image center, i.e.) Finally, the Kalman filter algorithm is used to smooth the three-dimensional coordinates of consecutive frames, eliminate detection noise, and complete the coordinates of occluded frames. The filtering model is as follows:

[0069] Equations of state: ; Observation equation ;

[0070] In the formula, The state vector contains position and velocity. ;

[0071] Here is the state transition matrix. , This is the frame interval, corresponding to a frame rate of 60fps.

[0072] , To control the input matrix and control vector (default 0, no active control);

[0073] For the observation matrix, ;

[0074] , The process noise and observation noise are Gaussian distributed with variances of respectively. , ).

[0075] Ultimately, a continuous spatiotemporal trajectory sequence is generated. It covers the entire process of a ping-pong ball being launched from the ball machine, hit by the user, and then hitting the table again / landing.

[0076] User action sequence extraction uses the MediaPipe Pose model to detect the three-dimensional coordinates of 24 core joints of the user (such as head, shoulder, elbow, wrist, racket grip, etc.), focusing on the time window from the serve to the user hitting the ball to the second contact of the ping-pong ball with the table. , For the time of serving, (For the second touch time of the ping-pong ball on the table), extract the skeletal model sequence within this window, i.e., the user action sequence. , .

[0077] S402: Determine whether the ping-pong ball touches the table within the effective area of ​​the table.

[0078] The core of determining whether a ping-pong ball touches the table within its effective contact area is detecting the presence of contact feature points in the trajectory sequence. The effective contact area of ​​the table is defined as a three-dimensional spatial rectangle. ,in, (Standard table height) To determine the tolerance for contact with the table, the elastic deformation of the ping-pong ball is considered. Trajectory sequence is then traversed. Find a point that satisfies the following conditions: height, planar region, and direction of motion (the Z-axis component of the velocity vector changes from negative to positive before and after touching the platform). Among them, the height condition is ;

[0079] Planar region conditions are For the direction of motion, the velocity vectors of the points in the preceding and following frames satisfy the following conditions: and The condition for the direction of motion is that the ping-pong ball moves downwards from the air and bounces upwards after hitting the table. This is a unit vector in the positive Z-axis direction. If a touch point that meets the conditions exists, it is judged as a suspected stepping onto the platform and proceeds to subsequent violation detection; if no touch point exists, or the touch point feature is... If the object is outside the designated area, it is considered not to have gone on stage. ).

[0080] S403: If so, further based on the spatiotemporal trajectory sequence and the action sequence, determine whether there is an illegal secondary interaction event from the first ball-bouncing interaction to the second touch of the table. The illegal secondary interaction event includes: the ping-pong ball contacting a non-paddle part of the user's body, or the user performing a consecutive hitting action on the ping-pong ball with the paddle. The illegal secondary interaction event detection targets suspected table-play situations, including two types of detection: consecutive paddle hits and body contact with the ball.

[0081] In one embodiment of the present invention, determining whether the user's racket has performed a double-hit action on the ping-pong ball in step S403 includes:

[0082] S4031: Identify the first time point of contact between the ping-pong ball and the racket from the video stream data. .

[0083] S4032: Track the trajectory of the ping-pong ball after its initial separation and continuously monitor the spatial pose of the racket based on the user's action sequence.

[0084] S4033: Determine whether there is a second time point in the trajectory of the ping-pong ball before its second contact with the table that is abnormally close to or makes contact with the trajectory of the racket. .

[0085] S4034: If the second time point exists and time interval Less than the preset combo threshold Meanwhile, the ping-pong ball in to If the trajectory changes during the stroke do not conform to the physical dynamics model of a single shot, it is determined that a double shot has occurred.

[0086] In one embodiment of the present invention, determining whether the trajectory change conforms to the physical dynamics model of a single shot in step S4034 includes: based on the first time point The velocity vector of the incoming ping-pong ball The velocity vector of the racket at the moment of impact And a preset collision recovery coefficient, to calculate the ping-pong ball in Theoretical launch velocity vector after time step The theoretical launch velocity vector Compared with what is actually observed from video stream data, in The initial velocity vector of the ping-pong ball after time step Compare; calculate and Direction angle Ratio of speed to magnitude If the included angle greater than the preset angle deviation threshold and / or the ratio of the speed magnitudes Not within the preset reasonable range If the trajectory change does not conform to the physical dynamics model of a single shot, then it is determined that the change does not conform to the physical dynamics model of a single shot.

[0087] In this embodiment of the invention, the time point of the first contact between the ping-pong ball and the racket is first detected. Traversing the trajectory sequence With user action sequence Find a table tennis ball and racket grip point where the spatial distance is less than the contact threshold. The first point in time, namely:

[0088] ;

[0089] in The three-dimensional coordinates of the racket grip point (derived from the fit of the wrist and finger joints).

[0090] Then track The trajectory of the table tennis ball At the same time, through action sequences Real-time monitoring of the racket's spatial pose (position and attitude angle). Searching for (Second touch time) and satisfy time point If it exists, proceed to the next verification step;

[0091] Finally, the angle between the actual velocity and the time interval verification, the physical dynamics model verification, and the actual velocity was verified. Ratio of speed If the triple check passes, it is considered a combo. Specifically, time interval verification is... ,in (The combo detection threshold is set based on the physical characteristics of table tennis ball collisions).

[0092] The physical dynamics model verification specifically involves: calculation The velocity vector of the incoming ping-pong ball at any given moment:

[0093] ( );

[0094] The formula for calculating the velocity vector of the racket at the moment of impact is:

[0095] ;

[0096] Based on the collision recovery coefficient k (k∈[0.5,0.8], determined by the racket material), the theoretical launch velocity vector is calculated:

[0097] ;

[0098] extract The actual initial velocity vector of the ping-pong ball after the stroke: ;

[0099] The formula for calculating the direction angle is: ;

[0100] The formula for calculating the speed ratio is: ;

[0101] like ( (angle deviation threshold) or ( , If the speed is within a reasonable range, then the trajectory change is determined to be inconsistent with the single-hit model.

[0102] When the time interval verification passes but the dynamic model verification fails, it is judged as a "double hit" (or "combo attack"). ).

[0103] In one embodiment of the present invention, step S403, determining that the ball is in contact with a non-racket part of the user's body, includes: constructing a real-time three-dimensional skeletal model of the user; performing spatiotemporal collision detection on the trajectory of the ball and a set of body joints defined as non-racket contactable parts; if spatial overlap is detected after the first ball-bouncing interaction and before the second touch of the table, it is determined that the body is touching the ball.

[0104] Determining body contact with the ball requires constructing a set of non-permitted contact joints. (Such as joints in the head, torso, and legs, excluding joints related to racket grip such as the wrist and fingers), traverse If the trajectory points and key points within the time window exist satisfy ( If the collision detection threshold is reached, then it is determined that the body has touched the ball ( ).

[0105] S404: If the aforementioned illegal secondary interaction event exists, the return ball is determined to be invalid; if the aforementioned illegal secondary interaction event does not exist, the return ball is determined to be valid; if the ping-pong ball does not touch the table within the valid area of ​​the table, the return ball is determined to be non-returning.

[0106] Based on the touch panel detection and violation detection results, the final state is output through logical judgment:

[0107] ;

[0108] In one embodiment of the present invention, providing real-time audio-visual feedback based on the stage recognition result in step S4 includes configuring different audio-visual feedback signals for the three results of valid stage entry, no stage entry, and invalid stage entry.

[0109] After the ball's return to the table is determined, the result is immediately transmitted to the audio-visual feedback device. Feedback is provided via a combination of sound and light signals, and the feedback mapping function is as follows: ;

[0110] in, For sound signals (frequency) Duration ); Light signal (color) flicker frequency ).

[0111] The specific mapping relationship is: effective on-stage ( )correspond (Pleasant notification sound) (Always-on lights); Invalid performance on stage ( )correspond (Warning sound) (Rapidly flashing lights); Not on stage ( )correspond (Prompt tone) (Slow-flashing lights). After the judgment result is generated, audio and visual signals are triggered synchronously with a delay of ≤100ms to ensure the user instantly perceives the quality of the ball return. Key data generated during the judgment process (spatiotemporal trajectory sequence) Action sequence On stage Illegal signage All are related to the training session. ( The data is bound and cached in the local database to provide raw data for the motion deviation analysis and training report generation in step S5.

[0112] S5. Perform human joint point recognition on the user's hitting action in the video stream data to generate a user motion skeleton model; compare and analyze the user motion skeleton model with a preset standard motion model corresponding to the target technical action to obtain motion deviation analysis results; generate a training report based on the ball return recognition results and the motion deviation analysis results.

[0113] In the specific implementation process, the first step is to identify and correct human joint points. Specifically, this is based on the user action sequence from step S4. Extract each time frame The three-dimensional coordinates of the following 24 joints form a complete joint coordinate sequence:

[0114] ;

[0115] In the formula, Indicates user number Each key node in time The three-dimensional coordinates (unit: m), k∈[1,24], i∈[1,m].

[0116] To eliminate noise and outliers during the recognition process, this invention employs Gaussian filtering to smooth the coordinate sequence. The filtering formula is as follows:

[0117] ;

[0118] in, (Filter coefficients), n=2 (Filter window size) These are the smoothed joint node coordinates.

[0119] To satisfy ( Anomalies (maximum permissible displacement of joints) are corrected based on human kinematic constraints:

[0120] ;

[0121] in This is a limiting function to ensure that the corrected joint displacement is within a reasonable range.

[0122] Based on the corrected key point coordinate sequence Constructing a dynamic skeletal model The model expression is: ;

[0123] in, Standard joint coordinate sequence ( , (The number of frames in the standard motion). It is a standard skeletal connection set with fixed bone side lengths (calibrated based on the body parameters of professional athletes).

[0124] To eliminate the speed difference between the user's and the standard action, the timing is aligned according to the keyframes of the shot. First, in the user model... Extract 3 core keyframes: lead frame (Lowest point of the movement), the frame of the hit

[0125] (The moment of contact between the racket and the ball), follow-through frame (The highest point of the movement); then in the standard model Extract the corresponding 3 keyframes , , Finally, establish the mapping relationship between the user timeline and the standard timeline. The time frame number of the user model is adjusted to be consistent with that of the standard model through linear interpolation. This ensures that the motion poses of each time frame can be directly compared.

[0126] Action deviation analysis is the core data processing step in step S5. This invention is based on the aligned user model. Compared with the standard model The deviation is calculated from three dimensions: joint deviation, posture deviation, and timing deviation. The larger the deviation value, the less standard the movement is.

[0127] At the level of joint spatial deviation, the Euclidean distance deviation of each joint in the same time frame is calculated using the following formula:

[0128] ;

[0129] In the formula, For the first Spatial deviation of a key point at time t (unit: m); The corrected coordinates of the user's k-th joint point; The coordinates of the k-th joint are given.

[0130] Further statistical analysis of the average deviation of a single joint throughout the entire motion cycle Overall average deviation from all relevant nodes The degree of deviation between a single part and the overall movement is quantified, and the corresponding calculation formula is as follows;

[0131] ;

[0132] ;

[0133] At the level of action posture similarity deviation, the joint coordinate sequence is flattened into a posture vector. The matching degree between the user posture and the standard posture is calculated using cosine similarity. The posture deviation is obtained by subtracting the similarity from 1, which intuitively reflects the posture difference. Specifically, the expression for the posture vector is:

[0134] ;

[0135] ;

[0136] The formula for calculating pose similarity is: ;

[0137] in, The closer the value is to 1, the more consistent the current pose is with the standard pose;

[0138] The formula for calculating attitude deviation is: ;

[0139] The expression for the average attitude deviation over the entire motion cycle is: ;

[0140] At the motion timing deviation level, the deviation between the user's and the standard motion in the execution time of three keyframes—backswing, hitting, and follow-through—is calculated. The consistency between the user's hitting rhythm and the standard rhythm is assessed by using the average timing deviation. Finally, the deviation analysis results are output in ensemble form, including the average deviation of each joint, the overall average deviation, the average posture deviation, and the keyframe timing deviation. The three joints with the largest deviations are marked, identifying the main problems in the user's motion. Specifically, the deviation between the user and the standard motion in keyframe execution time is calculated using the following formula:

[0141] ;

[0142] ;

[0143] ;

[0144] in, , , These represent the timing deviations (in seconds) for the backswing frame, the hitting frame, and the follow-through frame.

[0145] The formula for calculating the average timing deviation is: ;

[0146] The results of the motion deviation analysis are finally output in set form:

[0147] ;

[0148] At the same time, mark the top 3 joints with the largest deviations (such as "elbow, wrist, shoulder") to identify the main problems with the user's movements.

[0149] In one embodiment of the present invention, when generating the training report in step S5, the number, proportion, and suspected cause classification of invalid return balls are included in the report as independent statistical indicators.

[0150] This invention is based on the ball return status in step S4. (k=1,2,...,K, where K is the total number of returns) and violation indicators The calculation includes indicators such as the number of times a ball is played without being played, the proportion of balls played without being played, the classification of invalid play reasons, the effective play rate, and the non-play rate. Specifically, the corresponding calculation formulas are as follows:

[0151] Total number of invalid appearances: ,in, This is an indicator function; it takes the value 1 if the condition is met, and 0 otherwise.

[0152] Invalid appearances percentage: ;

[0153] In the category of invalid performance reasons, the number of invalid performances due to combos is as follows: Number of invalid attempts due to body contact with the ball: Reason percentage: , And satisfy .

[0154] Effective attendance rate: ;

[0155] Non-performance rate: .

[0156] In this invention, the training report generation process integrates and structures multi-source data. The core of the report includes a training overview section displaying the training session ID, target action name, total number of returns K, and effective ball landing rate. Invalid attendance rate Non-election rate The core statistical indicators include: The motion deviation analysis section visually presents the overall motion standardization, the joint point with the largest deviation, posture similarity, and temporal deviation, supplemented by a deviation radar chart to visualize the deviation of 24 joint points; the invalid bounce statistics section focuses on the total number, proportion, and cause classification of invalid bounces. It accurately calculates the number and proportion of invalid bounces caused by double hits and body contact with the ball using an indicator function, and analyzes the suspected causes of "double hits" and "body contact" based on motion deviation data. For example, double hits may be caused by the racket retracting too quickly after hitting the ball, corresponding to excessive deviation of the wrist joint; the improvement suggestion section provides targeted motion adjustment suggestions and error correction guidance based on the deviation analysis results and causes of invalid bounces. For example, if the average elbow joint deviation is 0.15m, it is recommended to maintain a distance of about 15cm between the elbow and body during training; if double hits account for 60%, it is recommended to control the racket follow-through after hitting the ball to avoid secondary contact, ensuring that users can quickly understand and apply this to subsequent training.

[0157] Training report, motion deviation analysis results Result of ball return on the table All are related to the training session ( ) and user identification ( The data is bound to a local database and cached, and then synchronized to a cloud server to ensure that the data is traceable and can be analyzed a second time.

[0158] S6. At a preset time before the end of the current training session, output a reminder message to the user to return the scattered ping-pong balls to the ball storage compartment of the automatic ball serving machine.

[0159] Preset reminder time This refers to the time interval between the trigger time of the reminder signal and the end time of the training session. The value range can be set to [30s, 60s], with a default value of 45s. This is based on the end time of the training session. With preset reminder time Through formula Calculate reminder trigger time When the current system time If the user initiates an early termination of training, a reminder will be triggered immediately. If the user actively terminates training early, the remaining time will be calculated in real time; if the remaining time is less than [a certain amount], [the reminder will be triggered]. If an alert is triggered immediately, the original model will be executed, ensuring the flexibility of the timing logic.

[0160] S7. After the current training session is detected to have ended, the environment inside the learning ball cabin is captured by the smart camera; the environment image is analyzed to identify the number of ping-pong balls scattered on the floor and table of the learning ball cabin; if the number of identified ping-pong balls is lower than a preset return completion threshold, a preset reward is issued to the user account bound to the identity information.

[0161] The first step is environmental image acquisition, triggered by events including the normal completion of the training session (current system time). , The session end time bound to step S1), the user actively ends the training early (by clicking "End Training" on the interactive terminal, and the system detects that the user has left the training chamber), or the session times out without any action ( And the grace period has been exceeded. (Without user intervention), multiple cameras work synchronously during data collection (time deviation ≤ 1ms), covering all areas within the learning chamber where ping-pong balls may be scattered, including the table surface and edges. ); Sphere cabin ground ( The area to be captured includes the entire floor area inside the cabin and the area within 1 meter around the automatic ball-launching machine (to avoid missing balls scattered near the ball storage compartment). Each camera captures 3 frames of images (0.5s interval), and selects the clear and unblurred frame for identification to improve anti-interference capabilities.

[0162] After data acquisition, the acquired environmental images undergo preprocessing such as filtering and image enhancement to lay the foundation for subsequent recognition. The denoising process uses a median filtering algorithm to eliminate image noise; the filtering formula is as follows: ;in, These are the original image pixel values. The pixel value is after denoising, k=1 (filter window size 3×3).

[0163] Image enhancement improves image contrast by using histogram equalization to highlight the difference between the ping-pong ball and the background. The equalization formula is: ;in, Original image grayscale histogram, This is the histogram after equalization, where M and N are the width and height (number of pixels) of the image.

[0164] Ping-pong ball recognition and counting can employ a combination of the YOLOv8 object detection algorithm and morphological screening. Specifically, the preprocessed image is first input into the YOLOv8 lightweight model to detect ping-pong ball candidate regions, and the bounding box coordinates of each candidate region are output. (Coordinates of the top left corner of the bounding box and its width and height), confidence level ;Preserve confidence level In a specific embodiment of the present invention, the candidate region is filtered to remove low-confidence interfering targets. The value can be set to 0.85; next, for the filtered candidate regions, verify whether their shape conforms to the characteristics of a ping-pong ball (round, aspect ratio close to 1). The verification formula is: ,like If a ping-pong ball is nearly perfectly circular, it is considered a valid ping-pong ball; otherwise, it is considered debris and discarded. Finally, the number of valid ping-pong balls is counted, which is the total number of scattered ping-pong balls. The mathematical expression is:

[0165] ;

[0166] in, This represents the total number of candidate regions detected by YOLOv8. This is an indicator function (it takes the value 1 if the condition is met, otherwise it takes the value 0).

[0167] The standard for determining whether a property has been properly placed is based on the threshold for completion of the placement process. With this as the core, the threshold is defined as the maximum number of scattered ping-pong balls allowed to remain. The default value can be set to 10 (based on the maximum capacity of the ball-collecting chamber of 200 balls, with a residual rate of ≤5%), and the management backend can dynamically adjust it within the range of [5, 15]. The mathematical expression for determining whether the ball return meets the standard is:

[0168] ;

[0169] In the formula, The indicator is for compliance (1 = compliant, 0 = non-compliant). The number of scattered ping-pong balls identified (integer, ≥0); The threshold for completing the relocation (default 10).

[0170] Among them, when (Meets standards) This indicates that the user has returned most of the ping-pong balls to the ball storage compartment, and the number of remaining scattered balls is within the allowable range; when (Not up to standard) indicates that there are too many scattered balls and they have not been effectively put back in place.

[0171] The reward distribution process after achieving the target follows preset rules and procedures. Reward types include training points, discount coupons, and membership level growth value bonuses, which can be configured in the management backend. Each user can receive a maximum of three rewards per day to prevent malicious reward fraud. Rewards are first distributed based on the user identity identifier bound in step S1. Match the corresponding account, then call the reward distribution interface to write the reward, generate a detailed record including user ID, number of scattered balls, reward type and distribution time, and finally push the payment notification to the interactive display terminal and the bound mobile terminal simultaneously to ensure that the user is aware of the payment in a timely manner.

[0172] S8. Associate the training report, the action deviation analysis results, and the ball return recognition results with the identity information and upload them to the cloud server; based on the smart camera and / or the sensors in the automatic ball serving machine, monitor the total number of ping-pong balls in the ball chamber and send an alarm message to the management backend when the total number of ping-pong balls is lower than a preset threshold.

[0173] In this embodiment of the invention, the core of data association is to establish a unique mapping relationship between users, sessions, and data. The mapping model is as follows:

[0174] in, Provide a unique identifier for each user (such as an encrypted ID number or member number) to ensure that data belongs to a single user. A unique identifier is generated for each training session (generated in step S1) to ensure that data from different training periods for the same user can be independently distinguished. The integrated and complete training dataset is defined as:

[0175] in, The training report generated in step S5 (including statistics on invalid presentations, improvement suggestions, etc.); The results of motion deviation analysis (including numerical values ​​for joint deviation, posture deviation, timing deviation, etc.); The result of the ball return identification in step S4 (including each ball return) , (identifier sequence); Metadata (training start time) End time Target Action ID Serving parameter set wait).

[0176] The integrated complete training data set (including training reports, motion deviation analysis results, ball return on the table recognition results, and metadata) is bound together. First, the data is formatted and converted to a standardized JSON format. Then, a field validation algorithm is used to verify the existence and validity of the core fields, ensuring that no critical fields are missing. The validation formula is as follows: ;

[0177] in, For the core field set, such as , ,wait, This is an indicator function (1 if the field exists and is valid, 0 otherwise). This indicates that the verification has passed.

[0178] After successful verification, the data is encrypted and uploaded to the cloud server via HTTPS protocol. The cloud server then sends a confirmation signal upon receiving the data.

[0179] Infrared beam sensors inside the ball storage compartment of the automatic ball-serving machine collect the number of balls in the compartment. The smart camera then uses image recognition to count the number of ping-pong balls scattered on the ground and table. (i.e., the recognition result of step S7). The total number is calculated using a weighted fusion algorithm, with the formula as follows:

[0180] ;

[0181] in, The total number of ping-pong balls in the ball storage (the result is rounded to the nearest integer). ), , These are the weighting coefficients. .

[0182] Alarm threshold ( The minimum number of ping-pong balls required to trigger the alarm is specified, and the value is determined according to the following rules: ( , (This refers to the maximum capacity of the sphere cabin), that is... One (default). When At that time, a structured alarm containing information such as alarm type, learning pod number, and current number of pods is generated and pushed through multiple channels, including pop-up windows in the operation and maintenance APP, SMS, and pop-up windows in the management backend.

[0183] S9. Receive a sharing instruction initiated by the user through the interactive display terminal or a mobile terminal associated with the identity information; in response to the sharing instruction, publish the training report or summary information generated based on the training report to a designated social platform; and / or, upload the score calculated based on the ball return recognition result to the global or friend leaderboard in the cloud server.

[0184] In this invention, users can initiate sharing via the in-cabin terminal or mobile terminal. The system first verifies whether the user is the owner of the training session. After verification, the system generates the corresponding content based on the selected option. The full report only supports sharing without anonymization (hiding sensitive identity information and detailed deviation data). The summary extracts key highlights such as action names, effective ball landing rate, and comprehensive score. The anonymized content is adapted to the target platform format, such as WeChat Moments image and text cards, Weibo long text, and charts. The training comprehensive score is based on the ball landing recognition result in step S4. Analysis results of the action deviation from step S5 Calculate the overall score The formula is: ;

[0185] Effective on-stage rate (%); The average deviation (m) across all nodes; This represents the maximum permissible average deviation. , Here, are the weighting coefficients; When the calculation result is outside the range, the boundary value is used, such as... Then take 0. Then take 100.

[0186] Once the calculation is complete, the score will be uploaded to the cloud server along with the user's identity, and a global leaderboard will be generated in descending order, displaying the top 100, the user's own ranking, and the friends' leaderboard.

[0187] According to the table tennis self-learning method of the present invention, the present invention provides personalized training through the complete user journey from appointment to feedback, and uses visual technology to realize automatic supervision and material management, gamifies the recycling task to enhance user stickiness, and combines social sharing function to enhance the dissemination effect, thereby significantly reducing operating costs while improving the professionalism of training.

[0188] Corresponding to the above example of a self-learning method for table tennis, this invention also proposes a self-learning system for table tennis.

[0189] A self-learning table tennis system according to an embodiment of the present invention includes:

[0190] The identity verification and access control module is used to respond to the identity verification request initiated by the user at the entrance of the learning cabin through the identity recognition and access control system, obtain the user's identity information, bind the identity information with the current training session, and control the access control system to open.

[0191] The instructional video selection and parameter query module is communicatively connected to the authentication and access control module. It is used to receive the user's selection instruction for the instructional video of the target technical action on the interactive display terminal. Based on the selection instruction, it queries the standard set of serving parameters that match the target technical action from the pre-stored technical action and serving parameter mapping database.

[0192] The automatic ball-serving machine control module is communicatively connected to the teaching video selection and parameter query module, and is used to send the standard ball-serving parameter set to the automatic ball-serving machine; and control the automatic ball-serving machine to serve ping-pong balls to the table according to the standard ball-serving parameter set.

[0193] The ball return status recognition and real-time feedback module is used to collect video stream data containing the user's hitting action and the landing point of the ball in real time through a smart camera deployed in the learning chamber; based on the video stream data, the landing status of each ball return is recognized; and based on the landing status recognition result of each ball return, real-time audio-visual feedback is provided to the user through a feedback device.

[0194] The motion analysis and training report generation module is used to identify human joint points in the user's hitting motion in the video stream data and generate a user motion skeleton model; compare and analyze the user motion skeleton model with a preset standard motion model corresponding to the target technical motion to obtain motion deviation analysis results; and generate a training report based on the ball return recognition results and the motion deviation analysis results.

[0195] The reminder module is connected to the motion analysis and training report generation module. It is used to output a reminder message to the user at a preset time before the end of the current training session, prompting the user to return the scattered ping-pong balls to the ball storage compartment of the automatic ball serving machine.

[0196] The environmental inspection and reward distribution module is used to collect environmental images inside the learning ball cabin through the smart camera after the current training session is detected to have ended; analyze the environmental images to identify the number of ping-pong balls scattered on the floor and table of the learning ball cabin; if the number of identified ping-pong balls is lower than a preset return completion threshold, a preset reward is issued to the user account bound to the identity information.

[0197] The data upload and device monitoring module is used to associate the training report, the action deviation analysis results, and the ball return recognition results with the identity information and upload them to the cloud server; based on the smart camera and / or the sensors in the automatic ball serving machine, it monitors the total number of ping-pong balls in the ball chamber and sends an alarm message to the management backend when the total number of ping-pong balls is lower than a preset threshold.

[0198] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A self-taught method for learning to play table tennis, characterized in that, Includes the following steps: S1. In response to the user's identity verification request initiated by the user at the entrance of the training cabin through the identity recognition and access control system, obtain the user's identity information; bind the identity information with the current training session, and control the access control system to open; S2. Receive the user's selection instruction for the teaching video of the target technical action on the interactive display terminal; based on the selection instruction, query the standard serving parameter set that matches the target technical action from the pre-stored technical action and serving parameter mapping database; S3. Send the standard serving parameter set to the automatic serving machine; control the automatic serving machine to serve a ping-pong ball to the table according to the standard serving parameter set; S4. Through the smart camera deployed in the learning pod, video stream data including the user's hitting action and the landing point of the return ball is collected in real time; Based on the video stream data, the on-table state of each return shot is identified; based on the identification results of the on-table state of each return shot, real-time audio-visual feedback is provided to the user through a feedback device. S5. Perform human joint point recognition on the user's hitting action in the video stream data to generate a user motion skeleton model; compare and analyze the user motion skeleton model with a preset standard motion model corresponding to the target technical action to obtain motion deviation analysis results; generate a training report based on the ball return recognition results and the motion deviation analysis results. S6. At a preset time before the end of the current training session, output a reminder message to the user to return the scattered ping-pong balls to the ball storage compartment of the automatic ball serving machine; S7. After the current training session is detected to have ended, the environment inside the learning ball cabin is captured by the smart camera; the environment image is analyzed to identify the number of ping-pong balls scattered on the floor and table of the learning ball cabin; if the number of identified ping-pong balls is lower than the preset return completion threshold, a preset reward is issued to the user account bound to the identity information. S8. Associate the training report, the action deviation analysis results, and the ball return recognition results with the identity information and upload them to the cloud server; based on the smart camera and / or the sensors in the automatic ball serving machine, monitor the total number of ping-pong balls in the ball chamber and send an alarm message to the management backend when the total number of ping-pong balls is lower than a preset threshold.

2. The self-learning method for table tennis according to claim 1, characterized in that, The specific steps for obtaining user identity information in step S1 are as follows: receiving user reservation data forwarded by the reservation server, which includes the verified user identity identifier and the usage period; Step S2 specifically includes: determining the target action ID according to the selection instruction, and querying the mapping database using the target action ID as an index to obtain a record of serving parameters including at least the landing point, spin type, spin speed, and serving frequency.

3. The self-learning method for table tennis according to claim 1, characterized in that, Step S4, which identifies whether each return of the ball successfully lands on the table based on the video stream data, specifically includes: S401: Extract the spatiotemporal trajectory sequence of the ping-pong ball and the user's action sequence from the video stream data; S402: Determine whether the ping-pong ball touches the table within the effective area of ​​the table; S403: If so, then based on the spatiotemporal trajectory sequence and the action sequence, determine whether there is an illegal secondary interaction event from the first ball-bouncing interaction to the second touch of the table. The illegal secondary interaction event includes: the ping-pong ball contacting a non-paddle part of the user's body, or the user's paddle performing a series of hits on the ping-pong ball. S404: If the aforementioned illegal secondary interaction event exists, the return ball is determined to be invalid; if the aforementioned illegal secondary interaction event does not exist, the return ball is determined to be valid; if the ping-pong ball does not touch the table within the valid area of ​​the table, the return ball is determined to be non-returning.

4. The self-learning method for table tennis according to claim 3, characterized in that, Step S403, determining whether the user's racket has performed a double-hit action on the ping-pong ball, includes: S4031: Identify the first time point of contact between the ping-pong ball and the racket from the video stream data. ; S4032: Track the trajectory of the ping-pong ball after its initial separation and continuously monitor the spatial pose of the racket based on the user's action sequence; S4033: Determine whether there is a second time point in the trajectory of the ping-pong ball before its second contact with the table that is abnormally close to or makes contact with the trajectory of the racket. ; S4034: If the second time point exists and time interval The number of consecutive hits is less than the preset combo threshold, and the ping-pong ball is in to If the trajectory changes during the stroke do not conform to the physical dynamics model of a single shot, it is determined that a combo has occurred.

5. The self-learning method for table tennis according to claim 4, characterized in that, Step S4034, determining whether the trajectory change conforms to the physical dynamics model of a single shot, includes: Based on the first time point The velocity vector of the incoming ping-pong ball The velocity vector of the racket at the moment of impact And a preset collision recovery coefficient, to calculate the ping-pong ball in Theoretical launch velocity vector after time step ; The theoretical ejection velocity vector Compared with what is actually observed from video stream data, The initial velocity vector of the ping-pong ball after time step Compare; calculate and Direction angle Ratio of speed to magnitude ; If the included angle greater than the preset angle deviation threshold and / or the ratio of the speed magnitudes Not within the preset reasonable range If the trajectory change does not conform to the physical dynamics model of a single shot, then it is determined that the change does not conform to the physical dynamics model of a single shot.

6. The self-learning method for table tennis according to claim 3, characterized in that, Step S403, determining whether the ball is in contact with a non-racket part of the user's body, includes: constructing a real-time 3D skeletal model of the user; performing spatiotemporal collision detection on the trajectory of the ball and the set of body joints defined as non-racket contactable parts; if spatial overlap is detected after the first ball-bouncing interaction and before the second touch of the table, it is determined that the body is touching the ball.

7. The self-learning method for table tennis according to claim 3, characterized in that, Step S4 provides real-time audio-visual feedback based on the stage recognition result, including configuring different audio-visual feedback signals for the three results of valid stage entry, no stage entry, and invalid stage entry.

8. The self-learning method for table tennis according to claim 1, characterized in that, When generating the training report in step S5, the number, proportion, and suspected cause classification of invalid return shots are included in the report as independent statistical indicators.

9. The self-learning method for table tennis according to claim 1, characterized in that, It also includes S9, receiving a sharing instruction initiated by a user through the interactive display terminal or a mobile terminal associated with the identity information; and in response to the sharing instruction, publishing the training report or summary information generated based on the training report to a designated social platform; And / or, the score calculated based on the ball return recognition result is uploaded to the global or friend leaderboard in the cloud server.

10. A self-service table tennis learning system, characterized in that, include: The identity verification and access control module is used to respond to the identity verification request initiated by the user at the entrance of the training cabin through the identity recognition and access control system, obtain the user's identity information, bind the identity information with the current training session, and control the access control system to open; The instructional video selection and parameter query module is used to receive the user's selection instruction for the instructional video of the target technical action on the interactive display terminal; based on the selection instruction, it queries the standard set of serving parameters that match the target technical action from the pre-stored technical action and serving parameter mapping database; An automatic ball-serving machine control module is used to send the standard ball-serving parameter set to the automatic ball-serving machine; and to control the automatic ball-serving machine to serve ping-pong balls to the table according to the standard ball-serving parameter set. The ball return status recognition and real-time feedback module is used to collect video stream data containing the user's hitting action and the ball's landing point in real time through a smart camera deployed in the learning chamber; Based on the video stream data, the on-table state of each return shot is identified; based on the identification results of the on-table state of each return shot, real-time audio-visual feedback is provided to the user through a feedback device. The motion analysis and training report generation module is used to identify human joint points in the user's hitting motion in the video stream data and generate a user motion skeleton model; compare and analyze the user motion skeleton model with a preset standard motion model corresponding to the target technical motion to obtain motion deviation analysis results; and generate a training report based on the ball return recognition results and the motion deviation analysis results. The reminder module is used to output a reminder message to the user at a preset time before the end of the current training session, prompting them to return the scattered ping-pong balls to the ball storage compartment of the automatic ball serving machine; The environmental inspection and reward distribution module is used to collect environmental images inside the learning ball cabin through the smart camera after the current training session is detected to have ended; analyze the environmental images to identify the number of ping-pong balls scattered on the floor and table of the learning ball cabin; if the number of identified ping-pong balls is lower than a preset return completion threshold, a preset reward is issued to the user account bound to the identity information. The data upload and device monitoring module is used to associate the training report, the action deviation analysis results, and the ball return recognition results with the identity information and upload them to the cloud server; based on the smart camera and / or the sensors in the automatic ball serving machine, it monitors the total number of ping-pong balls in the ball chamber and sends an alarm message to the management backend when the total number of ping-pong balls is lower than a preset threshold.

Images