Intelligent trial method for standing long jump foul behavior integrating AI and rule engine

By integrating AI and rules engine methods, combining video signal processing and logic analysis, we construct enhanced action feature descriptors, and solve the problem of misjudgment and misjudgment of foul behavior judgments in standing long jump competitions, achieving more accurate and consistent judgments result.

CN118781527BActive Publication Date: 2024-11-15ZHUHAI XINWEI TECHNOLOGY CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202411262506.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-10
Publication Date
2024-11-15
Estimated Expiration
2044-09-10

AI Technical Summary

Technical Problem

The prior art has misjudgment and misjudgment in the judgment of foul behavior in a standing long jump competition, and traditional methods are difficult to capture the diversity and uncertainty of the actions, and real-time and consistency are difficult to guarantee.

Method used

Using a method of integrating AI and rules engines, an enhanced action feature descriptor is constructed through video signal acquisition, preprocessing, spatio-temporal feature conversion and artificial intelligence-driven logical analysis, and logically compared with the rules engine to generate foul judgment results.

Benefits of technology

It improves the diversity and uncertainty of the opposing long jump action, reduces misjudgment and misjudgment, and improves the real-time and consistency of the judgment results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118781527B_ABST
    Figure CN118781527B_ABST
Patent Text Reader

Abstract

The present application relates to an intelligent adjudication method for standing long jump foul behavior that integrates AI and a rule engine. The method includes: sampling from an action video signal at a preset sampling frequency to obtain non-integer cycle sampling data; preprocessing and standardizing the non-integer cycle sampling data to obtain integer cycle sampling data; optimizing the integer cycle sampling data using a compensation curve; obtaining the amplitude, phase and scale features that reflect the long jump characteristics of the standing long jumper; extracting the key action features of the standing long jumper based on the amplitude, phase and scale features; obtaining the comprehensive action performance index of the standing long jumper; extracting the high-frequency action features of the standing long jumper based on the amplitude; constructing an enhanced action feature descriptor; performing a logical comparison between the enhanced action feature descriptor and the preset foul judgment standard in the rule engine to generate a foul judgment result. The present application can cope with the diversity and uncertainty of students' standing long jump actions and avoid misjudgment or missed judgment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of data recognition technology, and in particular to an intelligent adjudication method for standing long jump foul behavior that integrates AI and a rule engine. Background Art

[0002] In the standing long jump competition on campus, accurately judging students' fouls is crucial to maintaining the fairness of the competition. Traditional foul judgment mainly relies on the subjective judgment of the referee, which is easily affected by factors such as experience and perspective, and it is difficult to ensure real-time and consistency. Especially for students, their movements may not be as standardized as those of professional students, which increases the difficulty of judgment.

[0003] With the development of computer vision and artificial intelligence technology, the automatic detection and objective judgment of foul behaviors can be achieved by intelligently analyzing the video data of the standing long jump process and combining it with the rule engine for logical reasoning. Existing technologies usually use a single camera for video acquisition, which may result in a limited viewing angle and fail to fully capture the details of the student's movements. In addition, traditional methods often only focus on the analysis of key frames, ignoring the continuity and integrity of the movements.

[0004] In terms of feature extraction, existing technologies mainly focus on time domain analysis and lack full use of frequency domain features. This limits the ability to identify small movements and potential violations. At the same time, traditional feature extraction methods often use linear models, which are difficult to capture complex nonlinear relationships, which may lead to inaccurate expressions of student movement features.

[0005] In terms of foul determination, existing methods usually use simple threshold methods or hard-coded rules, which lack flexibility and adaptability. This method is difficult to cope with the diversity and uncertainty of students' actions, and may lead to misjudgment or missed judgment. Summary of the invention

[0006] Based on this, it is necessary to provide an intelligent trial method for standing long jump fouls that integrates AI and rule engines and can avoid misjudgments and missed judgments in response to the above technical problems.

[0007] An intelligent trial method for standing long jump foul behavior integrating AI and rule engine, comprising the following steps:

[0008] The action video signal of the standing long jumper is collected, and sampling is performed from the action video signal according to a preset sampling frequency to obtain non-integer cycle sampling data; the non-integer cycle sampling data is the sampling data that does not completely cover the entire long jump action cycle;

[0009] Preprocess and standardize the non-integer cycle sampling data to obtain the whole cycle sampling data; the whole cycle sampling data is the sampling data that completely covers the entire long jump action cycle;

[0010] A compensation curve is constructed based on the full-cycle sampling data, and the full-cycle sampling data is optimized using the compensation curve;

[0011] The optimized full-cycle sampling data is transformed into time-space features to obtain the amplitude, phase and scale features reflecting the long jump characteristics of standing long jumpers.

[0012] Based on amplitude, phase and scale features, the key action features of standing long jumpers are extracted; the key action features include the spatiotemporal features of the take-off phase, the motion trajectory features of the air phase, the dynamic change features of the body posture and the spatiotemporal features of the landing phase;

[0013] Artificial intelligence-driven logic analysis and data technology are used to process key action features to obtain comprehensive action performance indicators of standing long jumpers; comprehensive action performance indicators are used to judge the standardization of the standing long jumper's take-off phase, the action quality in the air phase, and the effectiveness of the landing phase;

[0014] Based on the amplitude, the high-frequency motion features of the standing long jumper are extracted; the high-frequency motion features are used to capture the small and rapid motion changes of the standing long jumper;

[0015] Based on key action features, comprehensive action performance indicators and high-frequency action features, an enhanced action feature descriptor is constructed;

[0016] The enhanced action feature descriptor is input into the rule engine, and a logical comparison is performed with the preset foul judgment criteria in the rule engine to generate a foul judgment result.

[0017] In one embodiment, the step of collecting the action video signal of the standing long jumper and sampling from the action video signal according to a preset sampling frequency to obtain non-integer period sampling data includes the steps of:

[0018] Deploy a convolutional neural network to analyze the action video signal in real time and perform intelligent discretization processing according to a preset sampling frequency; the preset sampling frequency is dynamically adjusted based on the change in semantic differences between image frames in the action video signal;

[0019] The dynamic window technology based on long short-term memory network is used to process the discretized action video signal to capture the jumping cycle of the standing long jumper; the jumping cycle is the whole process from take-off to landing of the standing long jumper;

[0020] The sliding buffer data segments and jump cycles are intelligently spliced ​​into non-integer cycle sampling data; the sliding buffer data segments are high-value video signals of standing long jumpers before and after take-off and landing cached by the AI ​​system.

[0021] In one embodiment, the step of preprocessing and standardizing the non-integer cycle sampling data to obtain the full cycle sampling data includes the steps of:

[0022] Computer vision algorithms are used to identify key frames of non-integer period sampling data; key frames include the preparation for take-off posture, the moment of leaving the ground, and the moment of landing;

[0023] Based on the timing, add additional sampling points before and after the key frame;

[0024] A cubic spline interpolation algorithm is used to process the sampling points and supplementary sampling points corresponding to the preset sampling frequency to obtain a cubic spline interpolation function;

[0025] Generate continuous trajectory data based on cubic spline interpolation function;

[0026] Correct the trajectory of the body center of mass of the long jumper in the air flight phase in the continuous trajectory data;

[0027] The Kalman filter algorithm is used to perform continuity enhancement processing on the corrected continuous trajectory data to obtain a continuity enhanced data sequence;

[0028] Normalizing the spatial coordinates in the continuity enhanced data sequence to obtain a standardized continuity enhanced data sequence containing dimensionless standardized coordinates;

[0029] The standardized continuity enhanced data sequence is resampled at intervals to obtain whole-cycle sampling data; the whole-cycle sampling data includes multiple standardized cycles.

[0030] In one embodiment, the step of constructing a compensation curve based on the whole cycle sampling data and optimizing the whole cycle sampling data using the compensation curve includes the steps of:

[0031] Analyze the end part of the standardized cycle in the full cycle sampling data to obtain the end motion characteristics;

[0032] Based on the end motion characteristics, generating transition interval data;

[0033] The compensation curve is constructed using the transition interval data, and the compensation curve is used to optimize the whole cycle sampling data.

[0034] In one embodiment, the step of performing time-space feature conversion on the optimized full-cycle sampling data to obtain amplitude, phase and scale features reflecting the long jump characteristics of the standing long jumper includes the steps of:

[0035] The spatial coordinates in the optimized full-cycle sampling data are subjected to discrete Fourier transformation to obtain the amplitude and phase reflecting the long jump characteristics of the standing long jumper;

[0036] The spatial coordinates in the optimized whole-cycle sampling data are transformed by continuous wavelet transform to obtain the scale features reflecting the long jump characteristics of standing long jumpers.

[0037] In one embodiment, the step of extracting key action features of a standing long jumper based on amplitude, phase and scale features includes the steps of:

[0038] The spatiotemporal characteristics of the take-off phase are obtained based on the following steps:

[0039] Based on the amplitude and phase, the vertical spectrum amplitude and take-off duration of the standing long jumper are calculated;

[0040] Based on the vertical spectrum amplitude and take-off duration, the take-off force index is defined to characterize the spatiotemporal characteristics of the take-off phase.

[0041] The motion trajectory characteristics of the aerial phase are obtained based on the following steps:

[0042] Based on the scale feature, the trajectory of the center of mass of the standing long jumper is tracked;

[0043] Based on the wavelet coefficient of the center of mass position of the body center of mass trajectory, a trajectory smoothness index is defined to characterize the motion trajectory characteristics in the air phase;

[0044] The dynamic change characteristics of body posture are obtained based on the following steps:

[0045] According to the phases, the upper body movement phase and the lower body movement phase of the standing long jumper are obtained;

[0046] According to the upper body movement phase and the lower body movement phase, the posture coordination index is defined to characterize the dynamic change characteristics of body posture;

[0047] The spatiotemporal characteristics of the landing phase are obtained based on the following steps:

[0048] Based on the amplitude and phase, the vertical spectrum amplitude and balance recovery time of the standing long jumper in the high frequency range are obtained;

[0049] Based on the vertical spectrum amplitude in the high-frequency range and the balance recovery time, a landing stability index is defined to characterize the spatiotemporal characteristics of the landing stage.

[0050] In one embodiment, the steps of using artificial intelligence driven logic analysis and data technology to process key action features and obtain comprehensive action performance indicators of standing long jumpers include the following steps:

[0051] The multi-layer perceptron neural network model is used to process key action features to obtain the take-off standardization, aerial action quality and landing effectiveness of standing long jumpers.

[0052] The standardization of take-off, the quality of aerial movements and the effectiveness of landing are weighted to obtain the comprehensive performance index of standing long jumpers.

[0053] In one embodiment, the step of extracting high-frequency motion features of a standing long jumper based on amplitude includes the steps of:

[0054] According to the amplitude, the high-frequency energy ratio and high-frequency fluctuation index are calculated;

[0055] The high-frequency motion characteristics of standing long jumpers are quantified based on high-frequency energy ratio and high-frequency fluctuation index.

[0056] In one embodiment, the step of constructing an enhanced motion feature descriptor according to key motion features, comprehensive motion performance indicators and high-frequency motion features includes the steps of:

[0057] According to key action features, comprehensive action performance indicators and high-frequency action features, an initial enhanced action feature descriptor is constructed;

[0058] The principal component analysis algorithm is used to reduce the dimension of the initial enhanced action feature descriptor to obtain the enhanced action feature descriptor.

[0059] In one embodiment, the step of inputting the enhanced action feature descriptor into a rule engine and performing a logical comparison with a preset foul determination standard in the rule engine to generate a foul determination result includes the following steps:

[0060] Mapping the enhanced motion feature descriptor to the original feature space corresponding to the full-cycle sampling data to obtain a reconstructed feature vector;

[0061] The reconstructed feature vector is input into the rule engine, and a logical comparison is performed with the preset foul judgment criteria in the rule engine to generate a foul judgment result.

[0062] One of the above technical solutions has the following advantages and beneficial effects:

[0063] The present application comprises the following steps: collecting an action video signal of a standing long jumper, and sampling from the action video signal at a preset sampling frequency to obtain non-integer-cycle sampling data; preprocessing and standardizing the non-integer-cycle sampling data to obtain full-cycle sampling data; constructing a compensation curve based on the full-cycle sampling data, and optimizing the full-cycle sampling data using the compensation curve; performing spatiotemporal feature conversion on the optimized full-cycle sampling data to obtain amplitude, phase and scale features reflecting the long jump characteristics of the standing long jumper; extracting key action features of the standing long jumper based on the amplitude, phase and scale features; processing the key action features using artificial intelligence-driven logic analysis and data technology to obtain comprehensive action performance indicators of the standing long jumper; extracting high-frequency action features of the standing long jumper based on the amplitude; constructing an enhanced action feature descriptor based on the key action features, comprehensive action performance indicators and high-frequency action features; inputting the enhanced action feature descriptor into a rule engine, performing a logical comparison with a preset foul judgment standard in the rule engine, and generating a foul judgment result. Therefore, the present application can cope with the diversity and uncertainty of students' standing long jump movements, accurately identify various standing long jump movements, improve the accuracy of identifying foul behaviors, and avoid misjudgment or missed judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 The present invention is a flowchart of an intelligent adjudication method for standing long jump foul behavior that integrates AI and a rule engine in an embodiment of the present application. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0066] In order to achieve the above object, in one embodiment, Figure 1 As shown, an intelligent trial method for standing long jump foul behavior integrating AI and rule engine is provided, comprising the following steps:

[0067] Step S110, collecting the action video signal of the standing long jumper, and sampling from the action video signal according to a preset sampling frequency to obtain non-integer cycle sampling data, wherein the non-integer cycle sampling data means that the sampling data does not completely cover the entire long jump action cycle.

[0068] In order to fully capture the standing long jump action of students (standing long jumpers), the system arranges multiple high-speed cameras equipped with edge computing units around the long jump venue. Specifically, one AI-optimized camera is deployed on each side of the take-off line, and one is deployed on each side of the long jump bunker, for a total of four. The selected camera integrates a neural network processor (NPU) with high frame rate, high resolution and real-time AI processing capabilities. Preferably, this example uses an intelligent high-speed camera with a frame rate of 240 frames per second and a resolution of 1920x1080 pixels. This configuration can perform target detection and posture estimation algorithms in real time to capture subtle changes in movements. The camera also integrates deep learning-based automatic exposure adjustment and AI-driven high dynamic range (HDR) technology to cope with the complex light changes that may occur in outdoor competition environments, ensuring clear images under various lighting conditions.

[0069] In one example, the steps of collecting action video signals of a standing long jumper, and sampling from the action video signals according to a preset sampling frequency to obtain non-integer-cycle sampling data include the following steps: deploying a convolutional neural network to perform real-time analysis on the action video signals, and performing intelligent discretization processing according to a preset sampling frequency; the preset sampling frequency is dynamically adjusted based on changes in semantic differences between image frames in the action video signal; applying a dynamic window technology based on a long short-term memory network to process the discretized action video signals to capture the jumping cycle of the standing long jumper; the jumping cycle is the entire process from take-off to landing of the standing long jumper; intelligently splicing the sliding buffer data segments and the jumping cycle into non-integer-cycle sampling data; the sliding buffer data segments are high-value video signals of the standing long jumper before take-off and after landing cached by the AI ​​system.

[0070] Specifically, after acquiring the action video signal, the system deploys a deep neural network based on the ResNet architecture to perform intelligent discretization processing on the video signal according to the preset sampling frequency. This application selects 60 Hz as the base sampling frequency, which means that 60 frames of images are extracted from the original video per second for AI analysis. This frequency is optimized through machine learning algorithms, balancing the accuracy of motion capture and computing resource requirements.

[0071] During the sampling process, the system uses a dynamic sampling strategy based on the visual transformer model to obtain finer action details. The specific implementation is as follows: Deploy the visual transformer model to monitor the semantic differences between image frames in real time. When the difference between consecutive frames exceeds the threshold dynamically calculated by AI, the system determines that it has entered the critical action stage and automatically increases the sampling frequency to 120 Hz. When the difference decreases and remains below the threshold for a period of time, the AI ​​controller restores the sampling frequency to the 60 Hz baseline.

[0072] Since the duration of the standing long jump varies from person to person, the system uses dynamic window technology based on recurrent neural networks (RNN) to accurately capture the complete process from take-off to landing. In specific implementation, the AI ​​system analyzes the action sequence in the action video signal in real time, and uses the posture estimation network to identify the start (standing still state) and end (stable state after landing) signs of the action. The recognition of the action start sign uses a spatiotemporal attention model: when it is detected that the student maintains a still posture in multiple consecutive frames and his position is near the take-off line, the point is marked as the start of the action. The recognition of the action end sign uses a 3D convolutional network: when it is detected that the player has completed the landing action and remains stable in the sand pit, the point is marked as the end of the action.

[0073] Furthermore, based on these AI-generated markers, the system dynamically adjusts the sampling window size. Specifically, the deep reinforcement learning algorithm sets the starting point of the sampling window to 0.5 seconds before the action start mark and the ending point to 0.5 seconds after the action end mark. This AI-driven setting ensures that the entire jumping action is fully captured, while including a small amount of buffer time before and after the action, providing complete contextual information for subsequent analysis.

[0074] In actual operation, the system maintains an intelligent sliding buffer based on a graph neural network (GNN) to continuously store and analyze the high-value sliding buffer data segments of the last 6 seconds. When the AI ​​detects a complete jump cycle, the corresponding sliding buffer data segments are extracted from the sliding buffer to form non-integer cycle sampling data. After the acquisition is completed, the deep learning model is deployed to perform a comprehensive integrity and quality check on the non-integer cycle sampling data. This includes using anomaly detection algorithms to verify whether the video length meets expectations, using optical flow estimation networks to check whether there is frame loss, and applying key point detection models to confirm whether all key action frames (take-off, highest point, landing) are accurately captured. Only data that passes the strict quality control of the AI ​​system will be transmitted to subsequent analysis modules.

[0075] Through the above AI-driven signal acquisition and intelligent sampling process, the system obtains high-quality, high-precision non-integer period sampling data. These data accurately reflect the entire process of the student's jump, including the preparatory action before take-off, the change of body shape at the moment of take-off, the posture adjustment in the air, and the buffering action when landing.

[0076] Step S120, preprocessing and standardizing the non-integer cycle sampling data to obtain the entire cycle sampling data, wherein the entire cycle sampling data refers to the sampling data completely covering the entire long jump action cycle.

[0077] In one example, the steps of preprocessing and standardizing non-integer-period sampling data to obtain integer-period sampling data include the following steps: using a computer vision algorithm to identify key frames of the non-integer-period sampling data; the key frames include the preparation for take-off posture, the moment of leaving the ground and the moment of landing; based on the time sequence, adding supplementary sampling points before and after the key frames; using a cubic spline interpolation algorithm to process the sampling points and supplementary sampling points corresponding to the preset sampling frequency to obtain a cubic spline interpolation function; based on the cubic spline interpolation function, generating continuous trajectory data; correcting the body center of mass trajectory of a standing long jumper in the air flight phase in the continuous trajectory data; using a Kalman filter algorithm to perform continuity enhancement processing on the corrected continuous trajectory data to obtain a continuity enhanced data sequence; normalizing the spatial coordinates in the continuity enhanced data sequence to obtain a standardized continuity enhanced data sequence containing dimensionless standardized coordinates; performing interval resampling on the standardized continuity enhanced data sequence to obtain integer-period sampling data; the integer-period sampling data contains multiple standardized periods.

[0078] Specifically, the key frames of the action cycle in the non-integer period sampling data are first accurately identified by a computer vision algorithm: the preparation for take-off posture, the moment of lifting off the ground, and the moment of landing. Preferably, the identification of the preparation for take-off posture uses a posture estimation algorithm. When it is detected that the athlete's feet are together, the forward leaning angle θ satisfies 5°<θ<15°, and the swing angle of the arms is >30°, the frame is marked as the take-off posture. The recognition of the moment of taking off is based on the contact state of the athlete's feet with the ground. When the gap d between the foot and the ground is greater than the threshold (usually set to 2cm) for the first time, it is marked as the moment of taking off. The recognition of the moment of landing is based on the contact state of the foot with the sand pit. When the foot descent speed v suddenly drops to 0 and the deformation of the sand pit surface is detected, it is marked as the moment of landing.

[0079] After identifying the key frames, add additional sampling points before and after these key frames. Starting from 0.5 seconds before the take-off posture, add sampling points at an interval of 1 / 4 of the original sampling frequency f until the moment of lifting off. If the original sampling interval is Δt, the time interval of the additional sampling points is 4Δt. Keep the original sampling frequency unchanged between the moment of lifting off and the moment of landing. Within 1 second after the moment of landing, use 1 / 2 of the original sampling frequency for supplementation, that is, the sampling interval is 2Δt.

[0080] Based on the existing data (non-integer period sampling data, preset sampling frequency, sampling points corresponding to the preset sampling frequency) and supplementary sampling points, a continuity enhanced data sequence is generated. The process is as follows:

[0081] First, all sampling points (including original sampling points (sampling points corresponding to the preset sampling frequency) and supplementary sampling points) are processed using the cubic spline interpolation algorithm. For any body key point P(x, y, z), the cubic spline interpolation function S(t) in the time interval [ti, ti+1] is defined as follows:

[0082] S(t) = ai(t-ti)³ + bi(t-ti)² + ci(t-ti) + di

[0083] Where ai, bi, ci, di are unknown coefficients, which are determined by solving a system of linear equations to ensure the continuity of the values, first-order derivatives, and second-order derivatives of the cubic spline interpolation function at the nodes. These coefficients are determined by solving a system of equations involving the positions of adjacent sampling points and the estimated velocities.

[0084] Based on these cubic spline interpolation functions, the initial continuous trajectory data P_initial(t) is generated. For each body key point j, its trajectory can be expressed as Sj(t) = (Sjx(t), Sjy(t), Sjz(t)), where Sjx(t), Sjy(t), Sjz(t) are the cubic spline interpolation functions in the x, y, and z directions respectively. Therefore, P_initial(t) can be expressed as the set of cubic spline interpolation functions of all body key points:

[0085] P_initial(t) = [S1(t), S2(t), ..., Sm(t)]

[0086] Where m is the number of body key points tracked. In this way, P_initial(t) becomes a continuous function, which can provide the three-dimensional coordinates of all body key points at any time t, realizing the conversion from discrete sampling points to continuous trajectories.

[0087] Next, the trajectory of the body's center of mass is corrected during the airborne phase. Specifically, assuming that the initial velocity of the center of mass is v0 and the take-off angle is α, the position of the center of mass at time t P(t) = (x(t), y(t)) can be expressed as:

[0088] x(t) = v0 * cos(α) * t

[0089] y(t) = v0 * sin(α) * t - 0.5 * g * t²

[0090] Where g is the acceleration of gravity, x(t) represents the horizontal displacement, and y(t) represents the vertical displacement. This model is used to fine-tune the initial trajectory P_initial(t) to ensure that the air trajectory conforms to the parabolic characteristics. The correction process is as follows:

[0091] P_corrected(t) = P_initial(t) - λ * (P_initial(t) - [x(t), y(t), P_initial_z(t)])

[0092] Where λ is a correction factor less than 1 used to smooth the correction process, and P_initial_z(t) is the z-coordinate component of the initial trajectory. This step ensures that the air trajectory is closer to the ideal parabola while retaining the characteristics of the full-cycle sampling data.

[0093] Furthermore, the Kalman filter algorithm is applied to enhance the continuity of the corrected continuous trajectory data. For each key body joint in the standing long jump action, an independent Kalman filter model is established. In this model, the state vector X(k) contains the three-dimensional position coordinates x, y, z of the joint point and the velocity components vx, vy, vz in the three directions of x, y, and z, that is, X(k) = [x, y, z, vx, vy, vz], a total of 6 variables. The state transfer matrix A adopts a uniform motion model in the form of a 6x6 matrix, in which the relationship between position and velocity is represented by the time interval dt. Considering the characteristics of the standing long jump action, the process noise covariance matrix Q is introduced and its value is dynamically adjusted at different stages. Q is a 6x6 symmetric matrix whose diagonal elements represent the noise variance of each state variable. For example, in the take-off and landing stages, the value of Q will increase accordingly to adapt to the rapid state changes that may occur in these stages. The observation matrix H is used to map the state vector to the observation space. Here, we mainly focus on the position information, so H is a 3x6 matrix, and only the position part of the state vector is extracted. The P_corrected(t) obtained in the previous step is input into the Kalman filter as the observation value Z(k). The filtering process includes two main steps: prediction and update. In the prediction step, the state at the next moment is predicted based on the current state estimate and the state transition model, and the prediction error covariance is updated at the same time. In the update step, the prediction result is compared with the new observation data, the Kalman gain K is calculated, and then the state estimate and the posterior error covariance are updated. The Kalman gain K is a key factor that balances the weights of the predicted value and the observed value. Its value will be dynamically adjusted according to the prediction error covariance and the observation noise covariance. Through this iterative optimization, the estimation of the trajectory of each joint point is continuously improved, and finally a smooth trajectory P_smooth(t) is obtained.

[0094] Through the above steps, a continuity enhanced data sequence is generated. This sequence can be expressed as:

[0095] P_smooth = {P_smooth(t) | t ∈ [tstart, tend]}

[0096] Among them, P_smooth(t) represents the enhanced data point at time t, which contains the three-dimensional coordinate information of all key points of the body.

[0097] Finally, the obtained continuity enhanced data sequence P_smooth is adjusted to full-cycle sampling data. The specific process is as follows:

[0098] First, define the standardized time τ∈[0,1] to map the original time t to the τ space:

[0099] τ = (t - tstart) / (tend - tstart)

[0100] Where tstart and tend are the start and end time of the action respectively.

[0101] Then, the spatial coordinates in P_smooth are normalized by dividing all coordinate values ​​by the player's height h to obtain dimensionless standardized coordinates. The continuity-enhanced data sequence after this processing can be expressed as:

[0102] E'(τ) = P_smooth(tstart + τ * (tend - tstart)) / h

[0103] Where h is the height of the player. E'(τ) is a three-dimensional vector function representing the normalized position at the standardized time τ.

[0104] Preferably, the standardized enhanced sequence E' is resampled at equal intervals to ensure that each standardized period contains 1000 sampling points. The resampling uses a linear interpolation method:

[0105] P(τi) = E'(τj) + (E'(τj+1) - E'(τj)) * (τi - τj) / (τj+1 - τj)

[0106] in:

[0107] - τi = i / 999, i = 0, 1, ..., 999 is the normalized time of the target sampling point

[0108] - τj and τj+1 are the normalized times of the two known sampling points in E' closest to τi, satisfying τj ≤ τi<τj+1

[0109] - E'(τj) and E'(τj+1) are the position vectors in E' corresponding to τj and τj+1

[0110] - P(τi) is the new position vector obtained after resampling, which is a three-dimensional vector like E'(τ)

[0111] Through this series of processing, the original non-integer cycle sampling data is converted into standardized full cycle sampling data. The final full cycle sampling data can be expressed as:

[0112] D_original(τ) = {P(τi) | τi = i / 999, i = 0, 1, ..., 999}

[0113] This D_original(τ) is the final standardized full-cycle sampling data, which contains 1000 equally spaced sampling points, and each point P(τi) is a three-dimensional position vector, representing the normalized three-dimensional coordinates of the key point of the body at the standardized time τi. This processing method not only achieves data standardization and full-cycle, but also ensures the continuity and smoothness of the action data.

[0114] Step S130, constructing a compensation curve based on the full-cycle sampling data, and optimizing the full-cycle sampling data using the compensation curve.

[0115] In one example, the steps of constructing a compensation curve based on whole-cycle sampling data and optimizing the whole-cycle sampling data using the compensation curve include the steps of: analyzing the end portion of the standardized cycle in the whole-cycle sampling data to obtain the end motion feature; generating transition interval data based on the end motion feature; constructing a compensation curve using the transition interval data, and optimizing the whole-cycle sampling data using the compensation curve.

[0116] Specifically, we first analyze the end of each standardized cycle in the full cycle sampling data, focusing on the interval τ∈[0.9, 1]. This interval is chosen to strike a balance between ensuring motion continuity and computational efficiency, while capturing the key features of the transition between motion cycles. In this interval, key feature parameters are calculated, including the position vector P(τ), velocity vector V(τ), acceleration vector A(τ), and the Euler angle Θ(τ) and angular velocity ω(τ) describing the body direction. These parameters (i.e., the end motion features) fully describe the motion state and trend at the end of the cycle.

[0117] First, extract the position vector P(τ) from D_original obtained in step S120:

[0118] P(τ) = D_original(τ) for τ ∈ [0.9, 1]

[0119] Wherein D_original(τ) directly corresponds to the full-cycle sampling data defined in step S120 .

[0120] Next, calculate the other parameters:

[0121] 1. The velocity vector V(τ) is calculated using the central difference method:

[0122] V(τ) = (P(τ+Δτ) - P(τ-Δτ)) / (2Δτ)

[0123] 2. The acceleration vector A(τ) is calculated using the second-order central difference:

[0124] A(τ) = (P(τ+Δτ) - 2P(τ) + P(τ-Δτ)) / (Δτ²)

[0125] 3. The Euler angle θ(τ) is calculated by analyzing the relative positions of key points of the body (such as shoulders and hips):

[0126] Θ(τ) = calculateEulerAngles(P_shoulder(τ), P_hip(τ))

[0127] where calculateEulerAngles is a function that calculates the Euler angles of the body based on the relative positions of the shoulders and hips.

[0128] 4. The angular velocity ω(τ) is obtained by the time derivative of the Euler angle:

[0129] ω(τ) = (Θ(τ+Δτ) - Θ(τ-Δτ)) / (2Δτ)

[0130] In these calculations, Δτ is the normalized time step size, typically taken as 0.001 (since D_original contains 1000 sample points).

[0131] Based on these end motion features, a transition interval data is generated so that it can maintain the motion features at the end of the current cycle and naturally transition to the starting state of the next cycle. The length of the transition interval is set to Δτ = 0.05. The generation process uses a dynamic model combined with a spline interpolation method. First, the motion trajectory of the transition interval is predicted using the end motion feature parameters:

[0132] P_pred(τ') = P(1) + V(1) * (τ' - 1) + 0.5 * A(1) * (τ' - 1)²

[0133] Θ_pred(τ') = Θ(1) + ω(1) * (τ' - 1)

[0134] where τ' ∈ [1, 1.05]. This preliminary prediction is based on the uniformly accelerated motion model and provides a good starting point for subsequent fine-tuning.

[0135] Then, the predicted trajectory is adjusted using quintic spline interpolation, taking into account the starting state P(0) and θ(0) of the next cycle. The interpolation function S(τ) is in the form of:

[0136]

[0137] coefficient arrive The position, velocity and acceleration constraints of the start and end points are determined by solving a system of linear equations. This six-parameter model provides enough degrees of freedom to generate smooth and natural transition curves.

[0138] Using S(τ), generate transition interval data T(τ):

[0139] T(τ) = S(τ) for τ ∈ [1, 1.05]

[0140] Here, S(τ) is actually a smooth adjustment of P_pred(τ') and Θ_pred(τ'), which ensures the continuity of transition data and full-cycle sampling data.

[0141] Next, the compensation curve C(τ) is constructed using the generated transition interval data:

[0142] C(τ) = w(τ) * D_original(τ) + (1 - w(τ)) * T(τ)

[0143] Where D_original(τ) is the full cycle sampling data, T(τ) is the transition interval data, and w(τ) is the weight function:

[0144] w(τ) = 0.5 * (1 + cos(π * (τ - 0.9) / 0.15)), τ ∈ [0.9, 1.05]

[0145] This cosine weight function ensures that the compensation curve is completely consistent with the full cycle sampling data at the start and end points, and has a smooth transition in the middle. Its derivative is zero at the end points, further ensuring the smoothness of the transition.

[0146] Finally, the constructed compensation curve is applied to the full-cycle sampling data:

[0147] D_optimized(τ) = { D_original(τ), τ ∈ [0, 0.9) C(τ), τ ∈ [0.9, 1.05]}

[0148] Where D_optimized(τ) is the optimized data. The meaning of this expression is:

[0149] 1. When τ is in the interval [0, 0.9) (including 0 and excluding 0.9), D_optimized(τ) is equal to the full-cycle sampling data D_original(τ).

[0150] 2. When τ is in the interval [0.9, 1.05] (inclusive), D_optimized(τ) is equal to the compensation curve C(τ). This interval not only includes the end of the current cycle (0.9 to 1), but also extends to the beginning of the next cycle (1 to 1.05). By using the compensation curve C(τ), possible instabilities at the end of the current cycle can be smoothed and a continuous and natural starting point can be provided for the next cycle. This method ensures the coherence and smoothness of the entire action sequence at the junction of the cycles, avoiding sudden changes or discontinuities.

[0151] To illustrate the effect of this processing, a specific standing long jump action is taken as an example. Assume that a player has a slight unstable jitter at the end of a jump action cycle, which may be caused by the imbalance when landing. In the whole cycle sampling data, this jitter is manifested as abnormal fluctuations in the vertical position data, with an amplitude of about 2cm. After applying the above curve compensation processing, this jitter is successfully smoothed. Specifically, in the transition interval, the standard deviation of the vertical position data is reduced from the original 1.5cm to 0.5cm, while the horizontal movement trend is maintained. At the same time, the change of the attitude angle is also smoother, and the maximum value of the angular velocity is reduced from the original 15° / s to 5° / s. This optimization not only improves the quality of the data, but also more accurately reflects the actual action of the player. For example, before optimization, due to the jitter at the end, the slight shake after landing may be mistakenly judged as the start of a new action. After optimization, the possibility of such misjudgment is greatly reduced, and the real action cycle boundary can be identified more accurately.

[0152] Step S140, performing time-space feature conversion on the optimized full-cycle sampling data to obtain amplitude, phase and scale features reflecting the long jump characteristics of the standing long jumper.

[0153] The spatial-temporal feature conversion mainly includes two parts: Fourier transform and continuous wavelet transform. These two transformation methods can capture the time-frequency characteristics of the action from different angles. In one example, the step of performing spatial-temporal feature conversion on the optimized whole cycle sampling data to obtain the amplitude, phase and scale characteristics reflecting the long jump characteristics of the standing long jumper includes the steps of: performing discrete Fourier transform on the spatial coordinates in the optimized whole cycle sampling data to obtain the amplitude and phase reflecting the long jump characteristics of the standing long jumper; performing continuous wavelet transform on the spatial coordinates in the optimized whole cycle sampling data to obtain the scale characteristics reflecting the long jump characteristics of the standing long jumper.

[0154] The spatial-temporal feature conversion mainly includes two parts: Fourier transform and continuous wavelet transform. These two transformation methods can capture the time-frequency characteristics of the action from different angles. Specifically, firstly, the three-dimensional spatial coordinate data in the optimized full-cycle sampling data D_optimized(τ) obtained in step S130 is subjected to discrete Fourier transform (DFT). For each spatial dimension, the calculation formula of DFT is as follows:

[0155] X(k) = Σ[n=0 to N-1] x(n) * e^(-j2πkn / N)

[0156] Y(k) = Σ[n=0 to N-1] y(n) * e^(-j2πkn / N)

[0157] Z(k) = Σ[n=0 to N-1] z(n) * e^(-j2πkn / N)

[0158] Where N is the total number of sampling points, and k = 0, 1, ..., N-1 represents the index of the frequency component. Preferably, a fast Fourier transform (FFT) algorithm is used to improve computational efficiency. The FFT algorithm reduces computational complexity by decomposing the DFT into smaller DFTs, reducing the time complexity from O(N^2) to O(N log N), greatly improving processing speed, especially for a large number of data points.

[0159] Through this transformation, we get the representation of the action in the frequency domain. The complex values ​​X(k), Y(k), Z(k) corresponding to each frequency component k contain amplitude and phase information:

[0160] Amplitude: A_x(k) = |X(k)|, A_y(k) = |Y(k)|, A_z(k) = |Z(k)|

[0161] Phase:

[0162] The amplitudes A_x(k), A_y(k), and A_z(k) reflect the strength of each frequency component, while the phase , , The relative time delay of each component is represented by . This information describes the frequency characteristics and temporal structure of the action.

[0163] Furthermore, we pay special attention to the amplitude A_y(k) in the vertical direction (y-axis) and divide it into two parts:

[0164] 1. A_y(k): overall vertical spectrum amplitude.

[0165] 2. A_y(k_high): vertical spectrum amplitude in the high frequency range, where k_high corresponds to a frequency greater than 10 Hz.

[0166] This classification allows for separate analysis of low-frequency components, which are often associated with overall movement trends, and high-frequency components, which may reflect rapid adjustments and minor movements.

[0167] For example, when analyzing the take-off phase, the focus is on the A_y(k) values ​​in the range of 2-5 Hz. This frequency range usually corresponds to the activity frequency of the large muscle groups of the human body. Assuming that a student's A_y(k) at 3 Hz reaches a peak value of 0.8 (normalized units), this may indicate that the student has good take-off explosiveness. The take-off ability of different students can be evaluated by comparing their peak values ​​in this frequency range.

[0168] When analyzing the landing phase, we mainly focus on A_y(k_high). For example, if A_y(k_high) peaks at 0.3 at 15 Hz and decays rapidly in the higher frequency range, this may indicate that the student has good landing cushioning ability. Rapid decay of high-frequency components usually means effective impact absorption and fast body stabilization.

[0169] In order to analyze the coordination of various parts of the body, the body is divided into the upper and lower parts, and their phase information is calculated separately:

[0170]

[0171]

[0172] where Yupper(k) and Ylower(k) are the Fourier transform results of the vertical motion of the upper and lower body, respectively. This division allows the evaluation of the coordination of the upper and lower body movements. By comparing and The difference between the upper and lower body frequencies can be used to quantify the coordination between the upper and lower body. For example, if the difference between the upper and lower body frequencies is 0.01 and 0.02, respectively, at the take-off frequency (about 3 Hz), and A difference close to 0 or π may indicate good coordination. A difference close to 0 indicates that the upper and lower body move in phase, while a difference close to π indicates opposite movement, both of which may reflect good technical movements, depending on the specific requirements of the movement.

[0173] Continuous wavelet transform (CWT) is applied to capture the multi-scale characteristics of the action. The calculation formula of CWT is as follows:

[0174] W(a,b) = (1 / √a) ∫[t=-∞ to ∞] x(t)ψ*((tb) / a)dt

[0175] Where a is the scale parameter, b is the translation parameter, and ψ is the mother wavelet function. Preferably, the Morlet wavelet is selected as the mother wavelet because it performs well in both time and frequency localization. The Morlet wavelet is a complex exponential function multiplied by a Gaussian window, which makes it have good localization characteristics in both time domain and frequency domain, and is particularly suitable for analyzing signals with transient characteristics.

[0176] The result of the wavelet transform W(a,b) provides a detailed representation of the action at different time scales. Special attention is paid to the wavelet coefficients representing the position of the center of mass of the body, which is used for subsequent analysis of the trajectory characteristics of the aerial phase. By analyzing the wavelet coefficients at different scales, the time and frequency information of the signal can be obtained simultaneously, which is not provided by the Fourier transform.

[0177] For example, when analyzing the aerial phase, if W(a,b) shows smooth changes on a large scale (a>32, corresponding to low-frequency components), such as forming a parabola-like shape in the vertical direction, this may indicate that the overall trajectory conforms to the ideal jump curve. The quality of the jump trajectory can be evaluated by fitting these large-scale coefficients. At the same time, if the fluctuations on a small scale (a<32, corresponding to high-frequency components) are small (for example, the standard deviation is less than 0.05), this may mean that the student maintains a stable posture in the air. The stability of the aerial posture can be evaluated by calculating the statistical properties of these small-scale coefficients.

[0178] Through these spatiotemporal feature transformations, detailed amplitude and phase information reflecting the characteristics of students' standing long jump movements were obtained. These transformation results can reveal the frequency components, time-varying characteristics, and multi-scale characteristics of the movements, which helps to more comprehensively and accurately describe and evaluate students' long jump techniques.

[0179] Step S150, based on the amplitude, phase and scale features, extract the key action features of the standing long jumper; the key action features include the spatiotemporal features of the take-off phase, the motion trajectory features of the air phase, the dynamic change features of the body posture and the spatiotemporal features of the landing phase.

[0180] In one example, the spatiotemporal characteristics of the take-off phase are obtained based on the following steps: based on the amplitude and phase, the vertical spectrum amplitude and the take-off duration of the standing long jumper are calculated; based on the vertical spectrum amplitude and the take-off duration, a take-off force index is defined to characterize the spatiotemporal characteristics of the take-off phase.

[0181] For the spatiotemporal characteristics of the take-off phase, we focus on the vertical acceleration change and horizontal displacement control. Using the Fourier transform result of step S140, we calculate the vertical spectrum amplitude and take-off duration. We define the take-off force index (TPI) to characterize the spatiotemporal characteristics of the take-off phase: TPI = max(A_y(k)) * Δt_takeoff

[0182] Where A_y(k) is the vertical spectrum amplitude, and Δt_takeoff is the take-off duration. TPI reflects the take-off explosiveness and sustainability. The higher the value, the greater the take-off power and the more reasonable the duration.

[0183] When calculating TPI, focus on the A_y(k) values ​​in the 2-5 Hz frequency range, which usually corresponds to the main activity frequency of the large muscle groups in the human body. By analyzing the amplitude distribution in this frequency range, the student's force output characteristics can be evaluated. For example, if a peak appears at a lower frequency (such as 2-3 Hz), it may mean that the student's force output is relatively slow; while a peak at a higher frequency (such as 4-5 Hz) may indicate a faster burst of force.

[0184] The measurement of Δt_takeoff involves precisely identifying the moments when the takeoff begins and ends. These two critical moments are determined by analyzing the changing pattern of vertical acceleration, allowing the takeoff duration to be accurately calculated. This time reflects the student's ability to convert push-off force into jump height and is critical for evaluating takeoff efficiency.

[0185] The TPI calculation combines the magnitude of the force (reflected by max(A_y(k))) and the duration of the force (reflected by Δt_takeoff), providing a comprehensive assessment of the quality of the take-off. A higher TPI value generally indicates that the student is able to generate and maintain a large thrust for an appropriate period of time, which is the basis for a good long jump performance.

[0186] In one example, the motion trajectory characteristics of the air phase are obtained based on the following steps: based on the scale characteristics, the body center of mass trajectory of the standing long jumper is tracked; based on the center of mass position wavelet coefficient of the body center of mass trajectory, a trajectory smoothness index is defined to characterize the motion trajectory characteristics of the air phase.

[0187] For the motion trajectory characteristics in the air phase, the wavelet transform result of step S140 is used to track the trajectory of the body center of mass. The trajectory smoothness index (TSI) is defined to characterize the motion trajectory characteristics in the air phase:

[0188] TSI = 1 / (1 + var(dW(a,b) / db))

[0189] Where W(a,b) is the wavelet coefficient of the center of mass position, dW(a,b) / db is its time derivative, and var represents the function for calculating the variance. The larger the TSI value, the smoother the trajectory, reflecting the body's ability to control in the air.

[0190] When calculating TSI, wavelet coefficients of different scales are considered. Large-scale coefficients (such as a>32) mainly reflect the overall trajectory shape, while small-scale coefficients (such as a<32) capture local minor adjustments. By comprehensively analyzing these different scales of information, students' attitude control ability in the air can be comprehensively evaluated.

[0191] The calculation of TSI involves numerically differentiating W(a,b) to obtain dW(a,b) / db. This derivative reflects the rate and direction of trajectory change. By calculating the variance of dW(a,b) / db, the degree of irregularity in trajectory change can be quantified. Smaller variance means a smoother trajectory, generally indicating better air control capabilities.

[0192] A high TSI value indicates that the student is able to maintain a stable posture in the air and reduce unnecessary body sway. This stability is essential for maximizing jump distance and maintaining a good landing posture. By comparing TSI values ​​at different time periods, you can evaluate the changes in a student's control ability during the initial jump, the highest point in the air, and before landing.

[0193] In one example, the dynamic change characteristics of body posture are obtained based on the following steps: according to the phase, the upper body movement phase and the lower body movement phase of the standing long jumper are obtained; according to the upper body movement phase and the lower body movement phase, a posture coordination index is defined to characterize the dynamic change characteristics of body posture.

[0194] For the dynamic change characteristics of body posture, the coordination of various parts of the body and the overall posture stability are analyzed. Using the phase information of step S140, a posture coordination index (PCI) is defined to characterize the dynamic change characteristics of body posture:

[0195] PCI = |1 / N Σ[n=0 to N-1] e^j(φupper(n) - φlower(n))|

[0196] Among them, φupper(n) and φlower(n) are the upper body movement phase and lower body movement phase respectively. The closer the PCI value is to 1, the better the coordination between the upper and lower body is, which reflects the degree of whole body movement coordination.

[0197] The calculation of PCI involves a complex analysis of the phase difference between the upper and lower body movements. First, the phase difference φupper(n) - φlower(n) is calculated for each frequency component n, and then this difference is converted into a complex exponential form. This conversion allows the periodic nature of the phase difference to be captured, even when the phase difference exceeds 2π.

[0198] By averaging the complex exponentials of all frequency components, a comprehensive coordination index is obtained. This index not only reflects the overall coordination level, but also retains the contribution of different frequency components. For example, special attention can be paid to the phase difference near the take-off frequency (about 3 Hz), because this frequency range is particularly important for evaluating the coordination of the take-off action.

[0199] A high PCI value indicates that the upper and lower body movements are highly coordinated, which is usually associated with more efficient force transfer and better technical execution. The PCI values ​​in different frequency ranges can be analyzed to evaluate the coordinated performance of students in different movement phases (such as knee flexion, extension, air adjustment, etc.).

[0200] In one example, the spatiotemporal characteristics of the landing phase are obtained based on the following steps: based on the amplitude and phase, the vertical spectrum amplitude in the high-frequency range and the balance recovery time of the standing long jumper are obtained; based on the vertical spectrum amplitude in the high-frequency range and the balance recovery time, a landing stability index is defined to characterize the spatiotemporal characteristics of the landing phase.

[0201] For the spatiotemporal characteristics of the landing phase, the impact force and body stability are analyzed. Using the Fourier transform result of step 140, a landing stability index (LSI) is defined to characterize the spatiotemporal characteristics of the landing phase:

[0202] LSI = 1 / (max(|A_y(k_high)|) * BRT)

[0203] A_y(k_high) is the vertical spectrum amplitude in the high frequency range, and BRT is the balance recovery time. The larger the LSI value, the more stable the landing and the faster the balance recovery.

[0204] When calculating LSI, the focus is on the spectrum amplitude in the high-frequency range (usually >10 Hz). These high-frequency components reflect the impact intensity and the body's rapid response at the moment of landing. By analyzing the distribution characteristics of A_y(k_high), the student's ability to absorb the impact of landing can be evaluated. For example, if the high-frequency components decay quickly, it may indicate good cushioning ability; while continuous high-frequency vibrations may indicate insufficient landing techniques.

[0205] The measurement of BRT involves the precise identification of the time interval from foot contact to body position stabilization. This time point is determined by analyzing the changes in the position and velocity of the body's center of mass, which takes into account both vertical and horizontal movements. A shorter BRT generally indicates better balance control and faster postural adjustments.

[0206] LSI is a comprehensive indicator that combines the impact absorption capacity (reflected by max(|A_y(k_high)|)) and the balance recovery speed (reflected by BRT). A high LSI value indicates that the student can effectively control the impact of landing and quickly recover to a stable posture. This not only helps to achieve better results in competitions, but also reduces the risk of injury caused by improper landing.

[0207] By extracting and analyzing these key action features, we can comprehensively and quantitatively describe each stage of the student's standing long jump action. These features not only reflect the basic parameters of the action, but also contain advanced information such as technical level, coordination and stability, which helps to more accurately evaluate the student's performance. Coaches can provide targeted guidance to students based on these indicators, such as improving take-off techniques, optimizing air posture control or strengthening landing cushioning ability, thereby helping students improve their overall long jump results.

[0208] Step S160, using artificial intelligence driven logic analysis and data technology to process key action features, and obtain comprehensive action performance indicators of the standing long jumper. The comprehensive action performance indicators are used to judge the standardization of the standing long jumper's take-off phase, the action quality of the air phase, and the effectiveness of the landing phase.

[0209] In one example, the steps of using artificial intelligence-driven logical analysis and data technology to process key action features and obtain comprehensive action performance indicators of standing long jumpers include: using a multi-layer perceptron neural network model to process key action features to obtain the take-off standardization, aerial action quality and landing effectiveness of the standing long jumpers; weighted processing of the take-off standardization, aerial action quality and landing effectiveness to obtain the comprehensive action performance indicators of the standing long jumpers.

[0210] Specifically, a multi-layer perceptron (MLP) neural network model is first constructed to process and analyze the key action features extracted in step S150. The network contains three hidden layers, each with 64, 32 and 16 neurons, respectively, using the ReLU activation function. The input layer receives the key action features extracted in step S150, including the take-off force index (TPI), trajectory smoothness index (TSI), posture coordination index (PCI) and landing stability index (LSI). The output layer corresponds to three main evaluation dimensions: take-off norms, aerial action quality and landing effectiveness.

[0211] For normative assessment of the take-off phase, the Take-off Position Deviation Index (TPDI) is introduced:

[0212] TPDI = |d_takeoff - d_line| / L_foot

[0213] Where d_takeoff is the actual take-off point, d_line is the take-off line, and L_foot is the length of the student's foot. TPDI combines TPI with neural network analysis to output the take-off standardization score S_takeoff.

[0214] By analyzing the image sequences captured by the high-speed camera, d_takeoff and d_line are precisely located. At the same time, image processing technology is used to measure the student's foot length L_foot. These precise measurements ensure the accuracy of TPDI, thereby improving the reliability of the take-off normative assessment.

[0215] The evaluation of the action quality in the air phase mainly considers the smoothness and stability of the trajectory of the body's center of gravity. The air posture change rate (APCR) is introduced:

[0216] APCR = max(|dW(a,b) / db|)

[0217] Wherein, W(a,b) is the wavelet coefficient obtained in step S140. The neural network comprehensively considers TSI, APCR and PCI, and outputs the aerial action quality score S_aerial.

[0218] When calculating APCR, wavelet coefficients at different scales are considered, which allows it to capture everything from large-scale overall trajectory changes to small-scale subtle posture adjustments. Through the comprehensive analysis of these multi-scale information, the quality of students' movements in the air can be comprehensively evaluated.

[0219] The effectiveness of the landing phase is evaluated by focusing on the relative position of the landing point to the take-off point. The Effective Jump Distance Index (EJDI) is defined as:

[0220] EJDI = (d_landing - d_takeoff) / H

[0221] Among them, d_landing is the landing point, d_takeoff is the take-off point, and H is the height of the student. The neural network integrates EJDI, LSI and BRT (balance recovery time) and outputs the landing effectiveness score S_landing.

[0222] By analyzing the image sequences of the high-speed camera, d_landing and d_takeoff are accurately measured. Combined with the previously measured student height H, EJDI provides a standardized jump distance indicator, allowing the performance of students of different heights to be fairly compared.

[0223] Finally, the Comprehensive Performance Index (CPI) is calculated by weighting the three output scores of the neural network:

[0224] CPI = w1 * S_takeoff + w2 * S_aerial + w3 * S_landing

[0225] Among them, w1, w2, and w3 are weight coefficients, which are determined by expert knowledge and data analysis and are 0.3, 0.4, and 0.3 respectively. These weights reflect the importance of different phases in the overall performance, among which the aerial phase is given a slightly higher weight because it directly affects the distance and quality of the jump.

[0226] Step S170, extracting high-frequency motion features of the standing long jumper based on the amplitude; the high-frequency motion features are used to capture the tiny and rapid motion changes of the standing long jumper.

[0227] The main focus is on A_y(k_high) defined in step S140, that is, the vertical spectrum amplitude with a frequency higher than 10 Hz. These high-frequency components are usually associated with rapid and subtle body adjustments, which have an important impact on the quality and standardization of movements. The analysis of high-frequency components is particularly important for fine movements such as standing long jump, because it can reveal the student's power output, balance control, and technical details in a short period of time.

[0228] In one example, the step of extracting high-frequency motion features of a standing long jumper based on amplitude includes the steps of: calculating a high-frequency energy ratio and a high-frequency fluctuation index according to the amplitude; and quantifying the high-frequency motion features of the standing long jumper based on the high-frequency energy ratio and the high-frequency fluctuation index.

[0229] In order to quantitatively describe these high-frequency characteristics, two key indicators are introduced: high-frequency energy ratio (HER) and high-frequency fluctuation index (HFI). The calculation formula of HER is:

[0230] HER = Σ|A_y(k_high)|^2 / Σ|A_y(k)|^2

[0231] In this formula, the numerator Σ|A_y(k_high)|^2 represents the sum of the energy of high-frequency components (frequency greater than 10 Hz), and the denominator Σ|A_y(k)|^2 represents the sum of the energy of all frequency components. This ratio reflects the proportion of high-frequency components in the overall movement. A higher HER value means that the movement contains more high-frequency components, which may mean faster movement adjustments or stronger muscle activity.

[0232] By calculating the HER value within each time window, a time-varying HER curve is obtained to analyze the changing trend of high-frequency components throughout the long jump. Specifically, a 0.1-second sliding window is used, with each movement of 0.02 seconds to ensure that rapidly changing high-frequency features are captured. This sliding window technique can provide continuous HER value changes, which helps to analyze the instantaneous changes of the action more finely. For example, at the moment of take-off, the HER value may rise rapidly, reflecting the rapid contraction of muscles and the burst of power.

[0233] The HER value shows a specific change pattern in different stages of the standing long jump. In the preparation stage, the HER value is usually maintained at a low level, about 0.05, reflecting the student's ability to keep the body still and stable before taking off. The low HER value indicates that the student can control the body well and reduce unnecessary small movements. The take-off stage is the stage where the HER value rises sharply, and the rising rate and peak value of the HER value are accurately calculated. This rapid rise reflects the explosive power and rapid contraction ability of the muscles. The HER value may rise rapidly from 0.05 to 0.3 or higher in 0.1 seconds. In the air phase, it is expected to see the HER value drop rapidly and remain at a low level, usually below 0.1, which reflects the student's ability to keep the body stable in the air and reduce unnecessary movements. The change of HER value in the landing phase also has typical characteristics. At the moment of contact with the ground, the HER value will rise rapidly again, reflecting the impact and rapid posture adjustment when landing, and then drop rapidly, indicating that the body is gradually stabilizing.

[0234] On the other hand, the calculation formula of High Frequency Fluctuation Index (HFI) is:

[0235] HFI = std(A_y(k_high))

[0236] Here std represents the standard deviation calculation, which quantifies the degree to which the high-frequency component amplitude A_y(k_high) deviates from the mean. HFI describes the degree of change in the amplitude of the high-frequency component, which directly reflects the stability and coordination of the movement. A higher HFI value indicates that the amplitude of the high-frequency component varies greatly, which may mean that the movement is not stable or coordinated enough. The HFI value is also calculated using a 0.1 second sliding window to form a continuous HFI curve. This method can capture instantaneous changes in movement stability. For example, the HFI value may rise rapidly when taking off, reflecting rapid force output and posture changes.

[0237] The HFI also exhibits specific characteristics in different phases. The preparation phase usually has very low HFI values, indicating that the student is able to maintain a stable posture. The HFI value will rise rapidly during the take-off phase, reflecting the intense muscle contraction and burst of power. The HFI should drop rapidly and remain at a low level during the air phase, indicating good air posture control. The landing phase will see another rapid rise and fall in the HFI value, reflecting the impact absorption and rapid posture adjustment when landing.

[0238] Based on the above analysis, the eigenvalues ​​of high-frequency energy ratio (HER) and high-frequency fluctuation index (HFI) were calculated for the four stages of standing long jump (preparation, take-off, air, and landing). For the preparation and air stages, since these two stages need to maintain a relatively stable state, the average values ​​of HER and HFI are calculated as eigenvalues ​​to reflect the overall stability of the entire stage. For the take-off and landing stages, since these two stages involve rapid state changes and power bursts, the maximum values ​​of HER and HFI are taken as eigenvalues ​​to capture the peak performance of these critical moments.

[0239] In this way, eight high-frequency characteristic indicators, HER_prep, HER_takeoff, HER_air, HER_land, and HFI_prep, HFI_takeoff, HFI_air, and HFI_land, were obtained. These indicators reflect the high-frequency characteristics of each stage, including static stability, explosive power, coordination, air posture stability, and impact absorption and stability when landing. For example, lower HER_prep and HFI_prep values ​​indicate a good take-off preparation state, higher HER_takeoff and HFI_takeoff values ​​may mean strong take-off explosive power, lower HER_air and HFI_air values ​​indicate good air posture control, and the rapid rise and fall of HER_land and HFI_land reflect effective landing cushioning.

[0240] Step S180, constructing an enhanced action feature descriptor according to key action features, comprehensive action performance indicators and high-frequency action features.

[0241] The results of step S150, step S160 and step S170 are combined to obtain a comprehensive and informative feature descriptor. The specific reasons are as follows:

[0242] 1. Step S150 provides key action features (f1-f4), which capture the basic quality indicators of the action, such as take-off force, trajectory smoothness, posture coordination, and landing stability. These features are the core indicators for evaluating the overall quality of the action.

[0243] 2. Step S160 provides a comprehensive action performance index (CPI), which is a high-level indicator derived from an AI model that comprehensively considers multiple aspects of performance. Incorporating CPI into feature descriptors can introduce a higher level of action evaluation.

[0244] 3. Step S170 provides high-frequency action features (HER and HFI), which can capture small, rapid action changes that are difficult to identify with traditional methods. These high-frequency features are very important for detecting potential illegal actions or technical details.

[0245] By combining the results of these three steps, the enhanced action feature descriptor is able to simultaneously incorporate basic action quality, overall performance evaluation, and fine-scale action features, thereby providing a comprehensive, multi-scale action representation.

[0246] In an example, the steps of constructing an enhanced action feature descriptor based on key action features, comprehensive action performance indicators and high-frequency action features include the steps of: forming an initial enhanced action feature descriptor based on key action features, comprehensive action performance indicators and high-frequency action features; and using a principal component analysis algorithm to reduce the dimension of the initial enhanced action feature descriptor to obtain an enhanced action feature descriptor.

[0247] Specifically, the initial enhanced action feature descriptor x is constructed as follows:

[0248] x = [f1, f2, f3, f4, CPI, HER_prep, HER_takeoff, HER_air, HER_land,HFI_prep, HFI_takeoff, HFI_air, HFI_land]

[0249] Among them, f1 to f4 represent the key action features extracted in step S150. CPI is the comprehensive action performance index obtained in step S160. HER_prep, HER_takeoff, HER_air, and HER_land represent the high-frequency energy ratios of the four stages of preparation, take-off, air, and landing, respectively, and HFI_prep, HFI_takeoff, HFI_air, and HFI_land represent the high-frequency fluctuation indexes of these four stages. These high-frequency feature indicators come from the analysis results of step S170.

[0250] Next, the kernel principal component analysis (KPCA) algorithm is applied to perform nonlinear dimensionality reduction on the enhanced action feature descriptor x. KPCA implicitly maps the data points in the original feature space to a higher-dimensional (possibly infinite-dimensional) feature space through a kernel function. In this high-dimensional feature space, nonlinear relationships may become linear relationships, so that traditional principal component analysis can be applied. This implicit mapping allows the capture of nonlinear patterns in the data without actually calculating high-dimensional features, which is particularly suitable for processing complex action data such as standing long jump. The specific steps of KPCA are as follows:

[0251] 1. Calculate the feature mean X_mean:

[0252] First, calculate the average feature vector of all samples, that is, X_mean = (1 / N) * Σ[i=1 to N] xi, where N is the total number of samples and xi is the feature vector of the i-th sample. X_mean is a 13-dimensional vector corresponding to the average value of each dimension in the original feature space.

[0253] 2. Centralize data processing:

[0254] Each sample is centrally processed, that is, This step moves the centroid of the dataset to the origin, eliminating the average level differences between features, allowing subsequent analysis to focus on the changing patterns of the data rather than the absolute values.

[0255] 3. Construct the kernel matrix K and centralize it:

[0256] For each pair of centered samples and , calculate the kernel function value This step implements feature mapping, mapping the original 13-dimensional feature vector to a high-dimensional feature space. Next, the kernel matrix is ​​centered: Kc = K - 1N K - K 1N + 1N K 1N, where 1N is an N×N matrix with all elements being 1 / N.

[0257] 4. Perform eigenvalue decomposition on the centralized kernel matrix Kc:

[0258] Perform eigenvalue decomposition Kc = UΛUT, where Λ is a diagonal matrix containing eigenvalues ​​λ1 ≥ λ2 ≥ ...≥ λN, and U is the corresponding eigenvector matrix.

[0259] 5. Select the eigenvectors corresponding to the first k largest eigenvalues ​​to form the projection matrix P:

[0260] The eigenvectors corresponding to the first k largest eigenvalues ​​are selected to form the projection matrix P = [u1, u2, ..., uk]. The selection of k is based on the cumulative explained variance ratio, and the number of principal components that can explain 95% of the total variance is usually selected.

[0261] 6. Calculate the features after dimensionality reduction:

[0262] Finally, the reduced-dimensional feature Y is obtained by calculating Y = KcP. Each column in Y represents a principal component, and each row corresponds to the representation of a sample in the new feature space.

[0263] Through KPCA, the original 13-dimensional feature space is successfully compressed into a lower-dimensional space (usually less than 13 dimensions) while retaining the key nonlinear information in the data. This dimensionality reduction not only reduces the computational complexity of subsequent processing, but also helps to remove noise in the data and highlight key action features. The final Y is the enhanced action feature descriptor, which captures the main nonlinear patterns of the original high-dimensional features in a low-dimensional space, providing a more compact and informative feature description for subsequent action analysis and evaluation.

[0264] Step S190, inputting the enhanced action feature descriptor into the rule engine, performing a logical comparison with the preset foul determination criteria in the rule engine, and generating a foul determination result.

[0265] In one example, the enhanced action feature descriptor is input into a rule engine, and a logical comparison is performed with a preset foul determination standard in the rule engine to generate a foul determination result, including the steps of: mapping the enhanced action feature descriptor to the original feature space corresponding to the whole cycle sampling data to obtain a reconstructed feature vector; inputting the reconstructed feature vector into the rule engine, and a logical comparison is performed with the preset foul determination standard in the rule engine to generate a foul determination result.

[0266] Specifically, first, the enhanced action feature descriptor Y obtained in step S180 is remapped to the original feature space corresponding to the full-cycle sampling data:

[0267] X_reconstructed = Y * P^T + X_mean

[0268] Where P is the projection matrix in step S180, P^T is the transposed matrix of P, X_mean is the mean of the original data, X_mean = (1 / N) * Σ[i=1 to N] xi where N is the total number of samples, and xi is the original feature vector of the i-th sample. In this way, a reconstructed feature vector X_reconstructed corresponding to the original feature space is obtained, and its dimension is the same as the original feature vector x:

[0269] X_reconstructed = [X_f1, X_f2, X_f3, X_f4, X_CPI, X_HER_prep, X_HER_takeoff, X_HER_air, X_HER_land,

[0270] Among them, X_f1 to X_f4 correspond to the reconstructed values ​​of the key action features extracted in step 150, X_CPI corresponds to the reconstructed value of the comprehensive action performance index obtained in step S160, and X_HER_* and X_HFI_* correspond to the reconstructed values ​​of the high-frequency energy ratio and high-frequency fluctuation index of each stage respectively.

[0271] Next, the reconstructed feature vector X_reconstructed is input into the rule engine. The rule engine contains a series of predefined rule functions, which are based on expert knowledge and game rules. The rule engine executes the following rule functions in sequence:

[0272] 1. Take-off phase rules:

[0273] R_takeoff(X) = max(0, min(1, (w1*X_f1 + w2*X_HER_takeoff + w3*X_HFI_takeoff + w4*X_CPI - θ_takeoff) / δ_takeoff))

[0274] This rule evaluates compliance during the take-off phase. X_f1 represents the take-off force index (TPI), which reflects the explosiveness of the take-off; X_HER_takeoff and X_HFI_takeoff reflect the high-frequency energy ratio and volatility of the take-off phase, respectively, which can detect whether there is an illegal run-up or multiple take-off attempts; X_CPI is also taken into account as an indicator of overall performance. θ_takeoff is the violation threshold of the take-off phase, and δ_takeoff is the normalization factor.

[0275] 2. Rules for aerial posture:

[0276] R_posture(X) = max(0, min(1, (w5*X_f2 + w6*X_HER_air + w7*X_HFI_air +w8*X_CPI - θ_posture) / δ_posture))

[0277] This rule evaluates posture control in the air phase. X_f2 represents the trajectory smoothness index (TSI), which reflects the stability of the air posture; X_HER_air and X_HFI_air reflect the high-frequency characteristics of the air phase, which can detect unnecessary body swings; X_CPI is also taken into account as an indicator of overall performance. θ_posture is the violation threshold of the air posture, and δ_posture is the normalization factor.

[0278] 3. Rules for the landing phase:

[0279] R_landing(X) = max(0, min(1, (w9*X_f4 + w10*X_HER_land + w11*X_HFI_land + w12*X_CPI - θ_landing) / δ_landing))

[0280] This rule evaluates the stability and regularity of the landing phase. X_f4 represents the Landing Stability Index (LSI), which reflects the stability of the landing; X_HER_land and X_HFI_land reflect the high-frequency characteristics of the landing phase, which can detect extra steps or instability; X_CPI is also taken into account as an indicator of overall performance. θ_landing is the violation threshold of the landing phase, and δ_landing is the normalization factor.

[0281] 4. Rules for the preparation phase:

[0282] R_prep(X) = max(0, min(1, (w13*X_HER_prep + w14*X_HFI_prep - θ_prep) / δ_prep))

[0283] This rule evaluates the static stability of the preparation phase. X_HER_prep and X_HFI_prep reflect the high-frequency characteristics of the preparation phase and can detect the presence of small run-up movements or improper body swings. θ_prep is the violation threshold of the preparation phase and δ_prep is the normalization factor.

[0284] 5. Overall coordination rules:

[0285] R_coord(X) = max(0, min(1, (w15*X_f3 + w16*X_CPI - θ_coord) / δ_coord))

[0286] This rule evaluates the coordination of the entire action process. X_f3 represents the Postural Coordination Index (PCI), which reflects the coordination of the whole body action; X_CPI is taken into account as an indicator of overall performance. θ_coord is the violation threshold of overall coordination, and δ_coord is the normalization factor.

[0287] In these rules, wi (i = 1, 2, ..., 16) is the weight of each feature, which is used to adjust the importance of different features in the judgment.

[0288] The output of each rule function is limited to between 0 and 1, where 0 indicates no violation at all, 1 indicates a serious violation, and intermediate values ​​indicate possible violations to varying degrees. The output of each rule function is fuzzified and an S-type membership function is used to enhance the robustness of the judgment.

[0289] After the rule engine executes these rule functions, it uses the weighted sum method to combine the outputs of each rule:

[0290]

[0291] where αi (i = 1, 2, ..., 5) is the corresponding weight.

[0292] Finally, the rule engine compares V(X) with the preset foul judgment standard θ_decision to obtain the final foul judgment result:

[0293] If V(X)>θ_decision, it is considered a foul (Foul = True). If V(X) ≤ θ_decision, it is considered not a foul (Foul = False).

[0294] The present application comprises the following steps: collecting the action video signal of a standing long jumper, and sampling from the action video signal according to a preset sampling frequency to obtain non-integer cycle sampling data; preprocessing and standardizing the non-integer cycle sampling data to obtain full cycle sampling data; constructing a compensation curve based on the full cycle sampling data, and optimizing the full cycle sampling data using the compensation curve; performing spatiotemporal feature conversion on the optimized full cycle sampling data to obtain amplitude, phase and scale features reflecting the long jump characteristics of the standing long jumper; extracting key action features of the standing long jumper based on the amplitude, phase and scale features; processing the key action features using artificial intelligence-driven logic analysis and data technology to obtain comprehensive action performance indicators of the standing long jumper; extracting high-frequency action features of the standing long jumper based on the amplitude; the high-frequency action features are used to capture the tiny and rapid action changes of the standing long jumper; constructing an enhanced action feature descriptor based on the key action features, the comprehensive action performance indicators and the high-frequency action features; inputting the enhanced action feature descriptor into a rule engine, performing a logical comparison with the preset foul judgment criteria in the rule engine, and generating a foul judgment result. Therefore, the present application can cope with the diversity and uncertainty of students' standing long jump movements, accurately identify various standing long jump movements, improve the accuracy of identifying foul behaviors, and avoid misjudgment or missed judgment.

[0295] This application uses multiple high-speed cameras for all-round video acquisition, combined with dynamic sampling strategies and dynamic window technology, to improve the accuracy of capturing the entire process of students' standing long jump. Through Fourier transform and wavelet analysis, the system analyzes the time domain and frequency domain characteristics, especially the high-frequency components, to enhance the ability to identify small movements and potential violations. The introduced indicators such as high-frequency energy ratio (HER) and high-frequency fluctuation index (HFI) provide more dimensions for motion analysis. This method is used for nonlinear feature mapping, combined with principal component analysis dimensionality reduction, to construct an enhanced descriptor that expresses the characteristics of the action, which is helpful for subsequent foul judgments.

[0296] This application combines signal processing technology and rule engines to achieve intelligent foul determination. By mapping enhanced action feature descriptors to the original feature space, the system can apply preset determination rules. The rule engine covers many aspects of standing long jump, including take-off line violations, aerial posture violations, unstable landing, etc. The weighted summation method enables the system to evaluate the severity of different violations and provide a comprehensive determination result. The determination threshold mechanism combined with the comprehensive violation score helps to improve the accuracy of the final determination and may reduce misjudgments and missed determinations.

[0297] This application not only provides foul determination results, but also generates action analysis reports. By quantifying key action features, such as take-off power index (TPI), trajectory smoothness index (TSI), etc., it provides specific technical analysis for coaches and students. High-frequency feature analysis can reveal some action details that are difficult to observe directly, and this information may be helpful in improving training methods. The system's determination process has a certain degree of explainability. Each rule function has a specific calculation basis and can point out potential violations, which may help the referee's judgment and provide students with a reference for action improvement. This analysis method aims to improve the objectivity of game judgments, while providing new analysis tools for sports teaching and training.

[0298] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0299] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0300] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent application shall be subject to the attached claims.

Claims

1. An intelligent judgment method for standing long jump foul behavior integrating AI and rule engine, characterized in that: The following steps are involved: Collecting an action video signal of a standing long jumper, and sampling from the action video signal according to a preset sampling frequency to obtain non-integer cycle sampling data; the non-integer cycle sampling data is sampling data that does not completely cover the entire long jump action cycle; Preprocessing and standardizing the non-integer cycle sampling data to obtain the entire cycle sampling data; the entire cycle sampling data is the sampling data that completely covers the entire long jump action cycle; Constructing a compensation curve based on the full-cycle sampling data, and optimizing the full-cycle sampling data using the compensation curve; Performing time-space feature conversion on the optimized full-cycle sampling data to obtain amplitude, phase and scale features reflecting the long jump characteristics of the standing long jumper; Based on the amplitude, the phase and the scale features, extracting key action features of the standing long jumper; the key action features include spatiotemporal features of the take-off phase, motion trajectory features of the air phase, dynamic change features of body posture and spatiotemporal features of the landing phase; The key action features are processed by artificial intelligence-driven logic analysis and data technology to obtain comprehensive action performance indicators of the standing long jumper; the comprehensive action performance indicators are used to judge the standardization of the standing long jumper's take-off phase, the action quality of the air phase, and the effectiveness of the landing phase; Based on the amplitude, extracting high-frequency motion features of the standing long jumper; the high-frequency motion features are used to capture the small and rapid motion changes of the standing long jumper; constructing an enhanced action feature descriptor according to the key action features, the comprehensive action performance index and the high-frequency action features; Inputting the enhanced action feature descriptor into a rule engine, performing a logical comparison with a preset foul determination standard in the rule engine, and generating a foul determination result; The step of collecting the action video signal of the standing long jumper and sampling from the action video signal according to a preset sampling frequency to obtain non-integer period sampling data includes the following steps: Deploy a convolutional neural network to perform real-time analysis on the action video signal and perform intelligent discretization processing according to a preset sampling frequency; the preset sampling frequency is dynamically adjusted based on changes in semantic differences between image frames in the action video signal; The dynamic window technology based on long short-term memory network is applied to process the discretized action video signal to capture the jumping cycle of the standing long jumper; the jumping cycle is the whole process from take-off to landing of the standing long jumper; The sliding buffer data segment and the jumping period are intelligently spliced ​​into the non-integer period sampling data; the sliding buffer data segment is a high-value video signal of a standing long jumper before taking off and after landing cached by the AI ​​system.

2. The intelligent adjudication method for standing long jump foul behavior integrating AI and rule engine according to claim 1 is characterized in that: The step of preprocessing and standardizing the non-integer cycle sampling data to obtain the full cycle sampling data includes the following steps: Using a computer vision algorithm to identify key frames of the non-integer period sampling data; the key frames include a preparation for take-off posture, a moment of leaving the ground, and a moment of landing; Based on the timing, adding supplementary sampling points before and after the key frame; Using a cubic spline interpolation algorithm to process the sampling points corresponding to the preset sampling frequency and the supplementary sampling points to obtain a cubic spline interpolation function; Based on the cubic spline interpolation function, generating continuous trajectory data; Correcting the trajectory of the body center of mass of the standing long jumper in the air flight phase in the continuous trajectory data; Using a Kalman filter algorithm to perform continuity enhancement processing on the corrected continuous trajectory data to obtain a continuity enhanced data sequence; Normalizing the spatial coordinates in the continuity enhanced data sequence to obtain a standardized continuity enhanced data sequence containing dimensionless standardized coordinates; The standardized continuity enhanced data sequence is resampled at intervals to obtain the whole cycle sampling data; the whole cycle sampling data includes a plurality of standardized cycles.

3. The intelligent judging method for standing long jump foul behavior integrating AI and rule engine according to claim 1 is characterized in that: The step of constructing a compensation curve based on the full-cycle sampling data and optimizing the full-cycle sampling data using the compensation curve comprises the following steps: Analyzing the end part of the standardized period in the full-period sampling data to obtain the end motion feature; Based on the end motion characteristics, generating transition interval data; The compensation curve is constructed by using the transition interval data, and the full-cycle sampling data is optimized by using the compensation curve.

4. The intelligent adjudication method for standing long jump foul behavior integrating AI and rule engine according to claim 1 is characterized in that: The step of performing time-space feature conversion on the optimized full-cycle sampling data to obtain amplitude, phase and scale features reflecting the long jump characteristics of the standing long jumper includes the following steps: Performing discrete Fourier transformation on the space coordinates in the optimized full-cycle sampling data to obtain the amplitude and phase reflecting the long jump characteristics of the standing long jumper; The space coordinates in the optimized whole-cycle sampling data are subjected to continuous wavelet transformation to obtain scale features reflecting the long jump characteristics of the standing long jumper.

5. The intelligent adjudication method for standing long jump foul behavior integrating AI and rule engine according to claim 1 is characterized in that: The step of extracting key action features of the standing long jumper based on the amplitude, the phase and the scale features comprises the steps of: The spatiotemporal characteristics of the take-off phase are obtained based on the following steps: Based on the amplitude and the phase, calculating the vertical frequency spectrum amplitude and take-off duration of the standing long jumper; Based on the vertical spectrum amplitude and the take-off duration, a take-off force index is defined to characterize the spatiotemporal characteristics of the take-off phase; The motion trajectory characteristics of the aerial stage are obtained based on the following steps: Based on the scale feature, tracking the trajectory of the body center of mass of the standing long jumper; Based on the wavelet coefficient of the center of mass position of the body center of mass trajectory, a trajectory smoothness index is defined to characterize the motion trajectory characteristics of the aerial stage; The dynamic change characteristics of the body posture are obtained based on the following steps: According to the phases, acquiring the upper body movement phase and the lower body movement phase of the standing long jumper; According to the upper body movement phase and the lower body movement phase, a posture coordination index is defined to characterize the dynamic change characteristics of the body posture; The spatiotemporal characteristics of the landing stage are obtained based on the following steps: Based on the amplitude and the phase, obtaining the vertical frequency spectrum amplitude and balance recovery time of the standing long jumper in a high frequency range; Based on the vertical spectrum amplitude of the high-frequency range and the balance recovery time, a landing stability index is defined to characterize the spatiotemporal characteristics of the landing stage.

6. The intelligent judging method for standing long jump foul behavior integrating AI and rule engine according to claim 1 is characterized in that: The step of using artificial intelligence driven logic analysis and data technology to process the key action features and obtain the comprehensive action performance index of the standing long jumper includes the following steps: A multi-layer perceptron neural network model is used to process the key action features to obtain the take-off standardization, aerial action quality and landing effectiveness of the standing long jumper; The take-off standardization, the aerial action quality and the landing effectiveness are weighted to obtain a comprehensive action performance index of the standing long jumper.

7. The intelligent judging method for standing long jump foul behavior integrating AI and rule engine according to claim 1 is characterized in that: The step of extracting the high-frequency action features of the standing long jumper based on the amplitude includes the steps of: calculating a high frequency energy ratio and a high frequency fluctuation index based on the amplitude; Based on the high-frequency energy ratio and the high-frequency fluctuation index, the high-frequency motion characteristics of the standing long jumper are quantified.

8. The intelligent judging method for standing long jump foul behavior integrating AI and rule engine according to claim 1 is characterized in that: The step of constructing an enhanced action feature descriptor according to the key action features, the comprehensive action performance index and the high-frequency action features comprises the following steps: According to the key action features, the comprehensive action performance index and the high-frequency action features, an initial enhanced action feature descriptor is constructed; The principal component analysis algorithm is used to perform dimension reduction processing on the initial enhanced action feature descriptor to obtain the enhanced action feature descriptor.

9. The intelligent judging method for standing long jump foul behavior integrating AI and rule engine according to claim 1 is characterized in that: The step of inputting the enhanced action feature descriptor into a rule engine and performing a logical comparison with a preset foul determination standard in the rule engine to generate a foul determination result comprises the following steps: Mapping the enhanced action feature descriptor to the original feature space corresponding to the full-cycle sampling data to obtain a reconstructed feature vector; The reconstructed feature vector is input into the rule engine, and a logical comparison is performed with a preset foul determination standard in the rule engine to generate the foul determination result.

Citation Information

Patent Citations

  • Intelligent agent trajectory prediction method, system and device under frequency domain perspective, and medium

    CN115661199A

  • Standing long jump action normalization evaluation method based on attitude estimation

    CN116189301A

  • System and method for analysis of spatio-temporal data

    US20100098289A1