Phase segmentation method, system and storage medium for competitive equestrian sports

Abstract

The present invention discloses a phase segmentation method, system, and storage medium for competitive equestrian sports. The phase segmentation method comprises acquiring motion signals of a rider and a horse; preprocessing the motion signals and labeling the preprocessed motion signals; constructing a time series using the motion signals and their corresponding labeled labels; inputting the time series into a phase segmentation model; performing phase segmentation on the rider and horse during competitive equestrian sports to obtain phase features. The present invention can segment the phases of the rider and horse during competitive equestrian sports to obtain phase features.

Description

Technical Field

[0001] The present invention relates to a competitive equestrian sports phase segmentation method, system and storage medium, belonging to the technical field of competitive sports evaluation. Technical Background

[0002] Equestrianism, a highly artistic and spectacular competitive sport, places great emphasis on the interaction between rider and horse. Competitive equestrianism is primarily divided into dressage and show jumping, with dressage being the most spectacular. According to the dressage rules published by the International Equestrian Federation, during the entire dressage riding process, riders in formal attire and horses performing dressage steps become one, simultaneously displaying strength and beauty, tension and rhythm, coordination and unrestrainedness, making it highly entertaining.

[0003] Dressage comprises four basic gaits: the walk, the trot, the trot, and the canter. These four fundamental steps contribute to the spectacular performance of dressage. Furthermore, these four basic gaits generally exhibit regular, cyclical movements, and therefore contain a wealth of phase information. Studying the phase characteristics of these four basic gaits can, to a certain extent, reveal a rider's riding proficiency and provide valuable insights for equestrian training.

[0004] Due to the difficulty in obtaining professional equestrian competition data, current phase segmentation research in competitive equestrianism is limited, and most studies focus on the periodic kinematics of different basic gaits, such as limb joint angles and speed. Regarding phase segmentation within different gaits in equestrianism, the only simple description of phase division within these gaits is found in the rules published by the International Equestrian Federation.

[0005] Early research on competitive equestrian monitoring primarily relied on visual observation for analysis and assessment. With the advancement of sensing, communication, and computing technologies, modern equestrian monitoring technologies, such as video imaging, optical tracking, and inertial measurement units (IMUs), have emerged. These technologies, each with its own unique characteristics, can be used for equestrian monitoring and analysis, but their accuracy and adaptability vary. Analysis methods based on video imaging generally have limited adaptability and require specific environments. Analysis systems based on optical tracking require the placement of optical markers at the measurement site and use high-speed, high-precision cameras to capture the position of specific light points. While these systems offer high accuracy, they are generally expensive and susceptible to environmental influences such as lighting and occlusion. With the continuous development of micro-electromechanical systems (MEMS) technology, motion monitoring and analysis methods using micro-inertial measurement units (MIMUs) have gradually emerged. Compared to video images, the optical capture technology of equestrian motion analysis methods based on inertial sensors (IMUs) has many advantages. For example, it is not affected by temporal and spatial factors such as lighting and occlusion, and can accurately reflect the entire motion process in three-dimensional space. In addition, inertial sensors (IMUs) are affordable, easy to wear, and portable.

[0006] Currently, phase segmentation methods for time series are mainly divided into three categories: early neural network phase segmentation methods based on feature extraction, and the emerging methods based on deep recurrent neural networks and convolutional networks in recent years. Among them, feature-based neural network phase segmentation methods mainly include k-nearest neighbors (KNN), support vector machine (SVM), decision tree C4.5, naive Bayes, etc. The accuracy of these methods is relatively dependent on features, and the initial feature engineering is relatively time-consuming. Deep recurrent neural network-based methods mainly include long short-term memory neural network (LSTM) and gated recurrent unit (GRU). These methods eliminate the complex feature extraction process and can observe the forward and backward information of the sequence, generally with higher accuracy. Convolutional network methods mainly include temporal convolutional network (TCN). TCN is a special convolutional network specifically designed for processing time series data. It uses dilated convolution to capture long-term dependency information in convolutional neural networks and has high accuracy in many time series data processing.

[0007] To address the labeling issues associated with walk, trot, saddle trot, and running in competitive equestrianism, most current research using inertial sensors for motion recognition uses methods such as thresholding and comparison. The threshold method primarily uses the characteristics of sensor signals during motion to assign labels to sequences by setting a threshold interval. The comparison method uses video or other related technologies to simultaneously record video images during data collection and compare the images against the sequence to assign labels. However, these methods often suffer from low accuracy.

[0008] Furthermore, in the process of reconstructing three-dimensional equestrian movements, pose calculation is a key component for achieving accurate motion reconstruction. In current research, commonly used inertial sensor pose estimation methods include complementary filtering, particle filtering, and Kalman filtering. For example, Chinese Patent Publication No. 201310520233.X discloses a Kalman filter pose estimation algorithm based on dual MEMS-IMUs to estimate foot pose during human gait; Chinese Patent Application No. 201710493018.3 discloses a multi-sensor data fusion method based on particle filtering, which uses particle filtering fusion technology to calculate the orientation of pedestrians during movement; and Chinese Patent Application No. 201410247629.6 discloses a data fusion method based on complementary filtering to calculate the pose of the vehicle. Although many scholars have conducted related research, most are still at the experimental and theoretical stage, and most current algorithms suffer from problems such as data drift and poor adaptability.

[0009] Therefore, the present application provides a phase segmentation method, system and storage medium for competitive equestrian sports. Summary of the Invention

[0010] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a phase segmentation method, system and storage medium for competitive equestrian sports, which can segment the phases of the rider and horse during competitive equestrian sports and obtain phase characteristics.

[0011] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0012] In one aspect, the present invention provides a phase segmentation method for competitive equestrian sports, comprising the following steps:

[0013] Obtain movement signals from riders and horses;

[0014] Preprocess the motion signal and label the preprocessed motion signal.

[0015] The motion signal and its corresponding labeling tags are used to construct a time series, which is then input into the phase segmentation model to perform phase segmentation on the rider and horse during competitive equestrian sports and obtain phase features.

[0016] Furthermore, obtaining the motion signals of the rider and the horse includes:

[0017] Use wearable wireless body area network data acquisition system to obtain the motion signals of riders and horses;

[0018] The wearable wireless body area network data acquisition system includes a plurality of sensor nodes;

[0019] Each of the sensor nodes includes a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer.

[0020] Furthermore, the motion signal includes three-axis acceleration signals, three-axis angular velocity signals and three-axis geomagnetic field intensity signals of the joints of the rider and the horse at m consecutive moments.

[0021] Furthermore, the pre-processing of the motion signal includes:

[0022] The three-axis angular velocity signals are processed to remove zero drift and obtain the corrected angular velocity;

[0023] Perform ellipsoid fitting processing on the three-axis acceleration signal and the three-axis geomagnetic field intensity signal to obtain the corrected acceleration and the corrected geomagnetic field intensity;

[0024] The attitude quaternion between the sensing node and the binding joint point is determined using the corrected angular velocity, corrected acceleration and corrected geomagnetic field strength.

[0025] Furthermore, the marking labels include marking labels for four basic gaits: walk, trot, trot and canter.

[0026] Furthermore, the slow-moving marking label includes:

[0027] The right hind hoof leaves the ground,

[0028] The right front hoof leaves the ground, the left front hoof and hind hoof hit the ground,

[0029] The right hind hoof strikes the ground, and only the right front hoof leaves the ground.

[0030] The right front hoof hits the ground, only the left hind hoof leaves the ground,

[0031] The left front hoof is off the ground, and only the right front hoof and hind hoof strike the ground.

[0032] The left hind hoof leaves the ground, and the other three hooves hit the ground;

[0033] The labeling of the up and down quick step includes:

[0034] The interval between the left front hoof leaving the ground and striking the ground Ⅰ,

[0035] The interval between the right hind hoof leaving the ground and striking the ground Ⅰ,

[0036] The first vacant period I,

[0037] The interval between the right front hoof and the left hind hoof from leaving the ground to striking the ground Ⅰ,

[0038] Second Vacancy Period I;

[0039] The marking labels for the saddle trot include:

[0040] The interval between the left front hoof leaving the ground and striking the ground Ⅱ,

[0041] The interval between the right hind hoof leaving the ground and striking the ground Ⅱ

[0042] The first vacant period II,

[0043] The interval between the right front hoof and the left hind hoof from leaving the ground to striking the ground Ⅱ,

[0044] The second empty phase II;

[0045] The running marking tags include:

[0046] The left hind hoof strikes the ground, while the other hooves leave the ground.

[0047] The right hind hoof and the left front hoof hit the ground at the same time, and only the right front hoof leaves the ground.

[0048] The right front hoof hits the ground, the left hind hoof leaves the ground,

[0049] The left front hoof and the right hind hoof leave the ground at the same time, and only the right front hoof hits the ground.

[0050] In the airborne state, all four horseshoes are off the ground.

[0051] Furthermore, the phase segmentation model includes an input layer, a temporal convolution layer, a bidirectional long short-term memory neural network layer, a Dropout layer, a fully connected layer and an output layer arranged in sequence.

[0052] Furthermore, the phase segmentation model includes the following formula:

[0053]

[0054] Where Softmax() is the Softmax activation function, which mainly maps the output of the bidirectional long short-term memory neural network layer as the input of the function to the output layer, thereby outputting the phase division result, that is, the phase feature, x i represents the i-th element in x, x j represents the jth element in x, ∑ j exp(x j ) is the sum of the derivatives of all elements in x, exp(x i ) is the element x i Find the derivative.

[0055] In a second aspect, the present invention provides a competitive equestrian phase segmentation system, comprising

[0056] A sampling module for acquiring motion signals of riders and horses;

[0057] The marking module is used to pre-process the motion signal and mark the pre-processed motion signal.

[0058] The phase segmentation module is used to construct a time series using motion signals and their corresponding labeled labels, and input the time series into the phase segmentation model to perform phase segmentation on the rider and horse during competitive equestrian sports to obtain phase features.

[0059] In a third aspect, the present invention provides a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the above methods.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] The present invention enables three-dimensional visualization of competitive motion signals by preprocessing motion signals, and labels the motion signals through three-dimensional visualization; the present invention inputs the motion signals and the corresponding labeled labels into a phase segmentation model based on deep network fusion, and extracts the phase characteristics of the rider and the horse in different interaction modes from an objective and quantitative perspective, that is, the phases of the slow walk, the saddle trot, the trot and the canter during competitive equestrian sports are segmented to obtain the phase characteristics of the rider at each moment in equestrian training; in competitive equestrian sports, each basic gait contains rich kinematic phase characteristics. By analyzing the phase characteristics, the degree of interaction and coordination between the rider and the horse and the riding level can be understood to a certain extent. Therefore, the present invention's prevention of phase segmentation in competitive equestrian sports can provide quantitative data reference for the subsequent improvement of competitive equestrian training methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The accompanying drawings further illustrate the present invention and constitute a part of the specification, and are used to explain the present invention together, but do not constitute a limitation of the present invention. In the accompanying drawings:

[0063] Figure 1 A schematic structural diagram of an embodiment of a phase segmentation method for competitive equestrian sports according to the present invention;

[0064] Figure 2 The motion signal of the 9-axis inertial sensor in competitive equestrian sports collected by the data acquisition system of the present invention;

[0065] Figure 3 This is a diagram showing the ellipsoid fitting effect of the three-dimensional accelerometer in the present invention;

[0066] Figure 4 This is the effect diagram before and after the ellipsoid fitting of the three-dimensional magnetometer in the present invention;

[0067] Figure 5 A schematic diagram of an embodiment of the initial posture alignment process of the present invention;

[0068] Figure 6Schematic diagram of the structure of an embodiment of the lower limb kinematics model of the present invention;

[0069] Figure 7 A schematic diagram of an embodiment of phase label division for competitive equestrianism according to the present invention;

[0070] Figure 8 The present invention is a competitive equestrian slow walk process using three-dimensional reconstruction of the phase label and the corresponding motion signal of the sensor node of the pelvis;

[0071] Figure 9 Schematic diagram of the structure of an embodiment of the TCN of the present invention;

[0072] Figure 10 This is a schematic diagram of the structure of a single LSTM unit according to an embodiment of the present invention;

[0073] Figure 11 This is a schematic diagram of the structure of an embodiment of the Bi-LSTM network of the present invention;

[0074] Figure 12 The confusion matrix for the four gaits of competitive equestrian sports according to the present invention;

[0075] Figure 13 This is a comparison diagram of phase division results of various algorithms in the case of slow walking in competitive equestrian sports of the present invention;

[0076] Figure 14 This is a comparison diagram of phase division results of various algorithms in the case of trotting in competitive equestrian sports according to the present invention;

[0077] Figure 15 This is a comparison diagram of phase division results of various algorithms in the case of trotting in competitive equestrian sports according to the present invention;

[0078] Figure 16 This is a comparison chart of the phase division results of various algorithms in the running condition of competitive equestrian sports according to the present invention. DETAILED DESCRIPTION

[0079] The following describes an embodiment of the present invention in conjunction with the accompanying drawings. Please note that the preferred embodiments described herein are only used to explain the present invention and are not limiting.

[0080] This application can be applied to a wearable wireless body area network data acquisition system. The data acquisition system is composed of multiple wireless sensor nodes connected through wireless communication.

[0081] In the application, the sensor nodes are installed on the rider's waist, chest, upper arms, forearms, thighs, calves and the horse's back to obtain motion signals during competitive equestrian sports.

[0082] Each sensor node in this application is primarily composed of a 9-axis inertial sensor, a computing core MCU, a wireless communication WiFi module, and a data storage SD card. The 9-axis inertial sensor integrates a 3-axis accelerometer, a 3-axis gyroscope, and a 3-axis magnetometer. It is encapsulated in a 3D-printed resin housing using a printed circuit board (PCB) and mounted on the moving parts of the rider and horse using straps or other fixing methods.

[0083] In application, reference Figure 2 ,The motion signal includes the three-axis acceleration signal, three-axis angular velocity signal and three-axis geomagnetic field intensity signal of the rider and horse at m consecutive moments.

[0084] In this application, the host computer sends a motion signal collection request to the sensor node through the communication module of the USB interface. After the sensor node receives the motion signal collection instruction, it starts collecting motion signals. The collected motion signals are uploaded to the host computer through the node's onboard WiFi module or stored in real time on an SD memory card. When the host computer sends a command to stop collecting motion signals, the sensor node stops collecting motion signals.

[0085] In the application, the three-axis gyroscope of each sensor node is de-zeroed, and the three-axis accelerometer and three-axis magnetometer are subjected to ellipsoid fitting to reduce the impact of noise on subsequent phase segmentation results.

[0086] Example 1

[0087] This embodiment provides a phase segmentation method for competitive equestrian sports.

[0088] The phase segmentation method for competitive equestrian sports in this embodiment includes the following steps:

[0089] S1 acquires motion signals from the rider and horse.

[0090] During application, a wearable wireless body area network data acquisition system is used to obtain motion signals of the rider and the horse. In addition, the wearable wireless body area network data acquisition system includes multiple sensor nodes, each of which includes a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer.

[0091] In practical applications, the motion signal includes three-axis acceleration signals, three-axis angular velocity signals, and three-axis geomagnetic field intensity signals of the joints of the rider and the horse at m consecutive moments.

[0092] S2 preprocesses the motion signal, reduces the noise of the signal, and labels the preprocessed motion signal.

[0093] In practice, 3D gyroscopes typically have zero bias errors, 3D accelerometers typically have inter-axis alignment errors, and 3D magnetometers are generally susceptible to interference from external magnetic fields. This embodiment uses a calibration method based on least squares ellipsoid fitting to correct for inter-axis errors in 3D accelerometers and external magnetic field interference in 3D magnetometers.

[0094] In practical applications, preprocessing motion signals includes the following steps:

[0095] S21 removes the zero drift of the three-axis angular velocity signal to eliminate the measurement error of the three-axis gyroscope and obtain the corrected angular velocity;

[0096] S22 performs ellipsoid fitting processing on the three-axis acceleration signal and the three-axis geomagnetic field intensity signal to eliminate the measurement errors of the three-axis accelerometer and the three-axis magnetometer, and obtain the corrected acceleration and the corrected geomagnetic field intensity;

[0097] S23 determines the attitude quaternion between the sensing node and the binding joint point using the corrected angular velocity, the corrected acceleration, and the corrected geomagnetic field strength.

[0098] In addition, the marking tags include marking tags for the four basic gaits: walk, trot, trot and canter.

[0099] Among them, the marking labels for slow walking include: right hind hoof off the ground A1; right front hoof off the ground, left front hoof and hind hoof hitting the ground A2; right hind hoof hitting the ground, only right front hoof off the ground A3; right front hoof hitting the ground, only left hind hoof off the ground A4; left front hoof off the ground, only right front hoof and hind hoof hitting the ground A5; left hind hoof off the ground, the other three hooves all hit the ground A6.

[0100] The marking tags for saddle trot include:

[0101] The interval between the left front hoof leaving the ground and striking the ground is ⅡB1; the interval between the right hind hoof leaving the ground and striking the ground is ⅡB2; the first airborne period is ⅡB3; the interval between the right front hoof and the left hind hoof leaving the ground and striking the ground is ⅡB4+B5; the second airborne period is ⅡB6;

[0102] The labeling of the Up and Down Quickstep includes:

[0103] The interval between the left front hoof leaving the ground and striking the ground is I C1; the interval between the right hind hoof leaving the ground and striking the ground is I C2; the first air period is I C3; the interval between the right front hoof and the left hind hoof leaving the ground and striking the ground is I C4 + C5; the second air period is I C6;

[0104] The marking labels for running include: the left hind hoof strikes the ground while the other hooves are off the ground (D1); the right hind hoof and the left front hoof strike the ground at the same time, with only the right front hoof off the ground (D2); the right front hoof strikes the ground, with the left hind hoof off the ground (D3); the left front hoof and the right hind hoof leave the ground at the same time, with only the right front hoof striking the ground (D4); and the horse is in the air with all four hooves off the ground (D5).

[0105] S3 uses motion signals and their corresponding labeled labels to construct a time series, and inputs the time series into a phase segmentation model to perform phase segmentation on riders and horses during competitive equestrian sports to obtain phase features.

[0106] In the application, the phase segmentation model includes an input layer, a temporal convolution layer, a bidirectional long short-term memory neural network layer, a dropout layer, a fully connected layer, and an output layer arranged in sequence.

[0107] In practical applications, the phase division model includes the following formula:

[0108]

[0109] Where Softmax() is the Softmax activation function, which mainly maps the output of the bidirectional long short-term memory neural network layer as the input of the function to the output layer, thereby outputting the phase division result, that is, the phase feature, x i represents the i-th element in x, x j represents the jth element in x, ∑ j exp(x j ) is the sum of the derivatives of all elements in x, exp(x i ) is the element x i Find the derivative.

[0110] The present invention enables three-dimensional visualization of competitive motion signals by preprocessing motion signals, and labels the motion signals through three-dimensional visualization; the present invention inputs the motion signals and the corresponding labeled labels into a phase segmentation model based on deep network fusion, and extracts the phase characteristics of the rider and the horse in different interaction modes from an objective and quantitative perspective, that is, the phases of the slow walk, the saddle trot, the trot and the canter during competitive equestrian sports are segmented to obtain the phase characteristics of the rider at each moment in equestrian training; in competitive equestrian sports, each basic gait contains rich kinematic phase characteristics. By analyzing the phase characteristics, the degree of interaction and coordination between the rider and the horse and the riding level can be understood to a certain extent. Therefore, the present invention's prevention of phase segmentation in competitive equestrian sports can provide quantitative data reference for the subsequent improvement of competitive equestrian training methods.

[0111] Example 2

[0112] Based on Example 1, this example introduces in detail the method for training the kinematic mechanism model.

[0113] Step 1: Obtain the three-axis acceleration signals, three-axis angular velocity signals, and three-axis geomagnetic field intensity signals of the rider's waist, chest, upper arm, forearm, thigh, calf, and horse's back for m consecutive moments, and record them as motion signals.

[0114] Step 2: training a kinematics mechanism model and labeling the motion signal using the kinematics mechanism model.

[0115] In the application, the data fusion and three-dimensional reconstruction methods of state switching extended Kalman filter are used to preprocess the motion signals, and the processed motion signals are used to train the kinematic mechanism model to divide the spatial phases of various basic gaits.

[0116] Step 2.1 calibrates the 9-axis inertial sensor to remove external interference, thereby ensuring the accuracy of subsequent phase segmentation.

[0117] The 9-axis inertial sensor in this embodiment is a MEMS inertial sensor.

[0118] In this application, the acquired three-axis angular velocity signals are de-zeroed, and the three-axis acceleration and geomagnetic field strength signals are fitted with an ellipsoid to reduce the impact of noise on subsequent phase segmentation results. When calibrating a three-axis gyroscope or magnetometer, the gyroscope or magnetometer is rotated in multiple directions in space and the output data is fitted with an ellipsoid to obtain the fitted calibrated acceleration or geomagnetic field strength.

[0119] Step 2.1.1 De-zeroing of the three-axis angular velocity signal

[0120] The zero-drift removal processing of the three-axis angular velocity signal can be performed by using the weighted average method in a static state to calculate the zero-degree drift, and the three-axis angular velocity output by the three-axis gyroscope is subtracted from the calculated zero-degree drift to obtain the calibrated angular velocity after zero-drift removal.

[0121] Step 2.1.2 Ellipsoid fitting of triaxial acceleration signals

[0122] The triaxial accelerometer is only affected by the static gravity. In three-dimensional space, the reference Figure 3 , the output modulus of the three-dimensional accelerometer in each motion posture is located on a virtual sphere with a radius of gravity, as shown in the following formula:

[0123]

[0124] Where, is the three-axis acceleration signal, α x,0 , α y,0 , α z,0 is the calibration value of each axis, p, l, j are the calibration amplitudes of each axis after ellipsoid fitting, and g is the acceleration due to gravity.

[0125] Since the center of the ellipsoid is the offset of each axis of acceleration, the problem of acceleration calibration is: how to find the center of the ellipsoid when a point on the ellipsoid is given.

[0126] In applications, the three-axis error between the three-axis acceleration signal and the calibration value of each axis is used to represent the deviation between the calibration value and the ideal value, as shown in the following formula:

[0127] e i ={e x ,e y ,e z}

[0128]

[0129] Where, e i is the three-axis error, e x 、e y 、e z is the error of each axis.

[0130] In practical applications, obtain sufficient motion signals and set the summation function Based on the least squares method, the optimal calibration parameter β is obtained by using the summation function. x,0 ,α y,0 ,α z,0 ,p,l,j), reference Figure 3 (b).

[0131] Step 2.1.3 Ellipsoid fitting of the three-axis geomagnetic field intensity signal

[0132] In the absence of deviation, the output of the three-axis magnetometer appears as a virtual spherical surface in three-dimensional space. However, in practice, there are often various interferences that distort the spherical surface into an ellipsoid. Figure 4 .

[0133] In applications, the error caused by the interference of the geomagnetic field is divided into soft magnetic influence, and the error caused by the interference of ferromagnetic materials in the environment is divided into hard magnetic influence. In actual applications, the soft magnetic influence is very small and generally negligible, while the hard magnetic influence will cause the center of the circle to deviate. Figure 4 (a) The output modulus of the three-axis magnetometer in each motion posture is located on a virtual ellipsoid, as shown in the following formula:

[0134]

[0135] Where, It is the deviation of the ellipsoid center caused by the influence of hard magnetism. is the three-axis geomagnetic field strength signal, a, b, c are the semi-axis lengths of the ellipsoid axes, and R is the geomagnetic field modulus constant.

[0136] Refer to step 2.1.2 to find the calibration parameters and fit the ellipsoid, that is, determine the deviation of the ellipsoid center caused by the hard magnetic effect. And the semi-axis lengths a, b, and c of the ellipsoid's axes. Figure 4 Before ellipsoid fitting, the magnetometer measurement value is affected by the magnetic field interference and appears as an ellipsoid whose center is not at the origin. After ellipsoid fitting, the external magnetic field interference is removed and it appears as a spherical surface with the center at the origin.

[0137] Step 2.2: Initially align the sensing node with the binding joint point to obtain the quaternion correspondence between the sensing node and the binding joint point.

[0138] In the application, the three-dimensional accelerometer and three-dimensional magnetometer are used to jointly solve the initial alignment of the sensor node and the binding joint point posture.

[0139] Step 2.2.1 sets a special calibration posture and uses the fusion of three-dimensional acceleration signals and three-dimensional geomagnetic field strength signals in this special state to obtain the posture reference between the sensor node and the joint point.

[0140] In this embodiment, the navigation coordinate system is set to the North-East-Earth coordinate system. Figure 5 , let the rider stand still facing north for about 10-20 seconds, so that the rider's joint coordinate system coincides with the navigation coordinate system, and the initial alignment between the sensor node and the binding joint point is achieved. At this time, the Euler angle between the sensor node and the navigation coordinate system includes the following formula:

[0141]

[0142] Where, α x , α y , α z represents the calibration acceleration after ellipsoid fitting processing, g represents the acceleration due to gravity, m x 、m y 、m z represents the calibrated geomagnetic field intensity after ellipsoid fitting, φ roll Rotation angle expressed as Euler angle θ pitch Represents Euler angles Nutation Angle, ψ yaw represents the precession angle of the Euler angles.

[0143] Step 2.2.2 Convert the Euler angle into the rotation quaternion between the sensor node and the navigation coordinate system using the following formula:

[0144]

[0145] Where, A quaternion representing the rotation between the sensor node and the navigation coordinate system.

[0146] Step 2.2.3 At this time, the joint coordinate system and the navigation coordinate system coincide with each other, so the rotation quaternion between the sensor node and the navigation coordinate system is equal to the attitude quaternion between the sensor node and the joint coordinate system. The attitude quaternion of the joint point in the navigation coordinate system in the initial state is determined by the following formula:

[0147]

[0148]

[0149] in, Represents the attitude quaternion between the sensor node and the joint point coordinate system, Represents the attitude quaternion between the joint point and the navigation coordinate system, and * represents the conjugate matrix.

[0150] Step 2.3 Fusion of the corrected motion signals of each sensor node

[0151] By fusing the corrected motion signals of each sensor node, the attitude quaternion between the sensor node and the navigation coordinate system can be determined in dynamic situations.

[0152] In this application, a state-switching extended Kalman data fusion model for 9-DOF inertial sensors is developed to estimate the spatial posture of each joint of the rider and horse during competitive equestrian sports. In practice, an extended Kalman filter (ESKF) is designed that switches based on the magnetic field observation state, tailored to the magnitude of magnetic field interference. This state-switching EKF fuses inertial data to determine the attitude quaternion between the sensor node and the navigation coordinate system.

[0153] The data fusion of multiple sensor nodes in this embodiment is mainly divided into two parts: state update and measurement update, and is based on the rotation quaternion. and the bias of the three-dimensional gyroscope are used as the state observations of the system.

[0154] Step 2.3.1 Status Update

[0155] The state update of this embodiment includes the following formula:

[0156] ω g =[ω gx ω gy ω gz ]

[0157]

[0158] Where, ω g is the calibration angular velocity after the gyroscope is debiased, f(X,ω g ) is the state update function, X is the state observation vector of the system;

[0159] Since the bias of the 3D gyroscope changes very little over time, it can be ignored:

[0160] b g =[b x b y b z ],

[0161]

[0162] Where b g Indicates the bias of the 3D gyroscope.

[0163] The differential of the rotation quaternion is obtained by the following formula:

[0164]

[0165]

[0166] Where, represents the differential of the rotation quaternion, Ω(ω g ) is expressed as the quaternion form of the gyroscope;

[0167] Since Ω(ω g ) includes the following formula:

[0168]

[0169] The differential of the rotation quaternion includes the following formula:

[0170]

[0171] Where 0 represents the zero matrix.

[0172] In application, the discrete state equation includes the following:

[0173] X k =X k-1 +f(X k-1 ,ω gk )Δt=f'(X k-1 ,ω gk )

[0174] Where, X k is the posterior estimate of the system state, X k-1 is the prior estimate of the system state, Δt represents the duration of obtaining the motion signal, ω gk represents the true angular velocity, f(X k-1 ,ω gk ) represents the state update function, f'(X k-1 ,ω gk ) represents the discrete state equation.

[0175] Linearize the discrete state equation to obtain the following equation:

[0176]

[0177] Where, is the previous step prediction value of the state vector, A represents the first Jacobian matrix, is the true value of the previous state, W k-1 is the process noise.

[0178] In the application, the first Jacobian matrix A includes the following formula:

[0179]

[0180] Where τ represents the sampling frequency, I represents the identity matrix, and since the discrete state equation is in matrix form with 7 rows, f'0, f'1, f'2, f'3, f'4, f'5, and f'6 here refer to the expressions of each row of the state update equation.

[0181] The state update of this embodiment is to use the process model of the system to predict the next state of the system. Here, the result of the state update is the quaternion of the next state.

[0182] Step 2.3.2 Measurement Update

[0183] The measurement update mainly includes two parts: gravitational acceleration measurement update and geomagnetic field measurement update.

[0184] Generally, the gravity acceleration vector can be used to observe the roll angle and pitch angle, while the geomagnetic field direction vector can be used to observe the yaw angle. However, when the sensor node is in a dynamic situation, the modulus output by the three-dimensional accelerometer includes not only the gravity acceleration, but also the external acceleration. In this case, using an inaccurate gravity acceleration measurement equation will reduce the accuracy of attitude estimation. Therefore, the Kalman gain must be used to weight the residual. In addition, ellipsoid fitting is suitable for the calibration of three-axis magnetometers in a fixed magnetic field interference environment. When the magnetic interference in the environment is more complex, it is usually no longer applicable. Therefore, the following formula is used to monitor magnetic field interference and determine whether to update the geomagnetic field measurement:

[0185]

[0186] Where, P t m Indicates the result of whether to execute the geomagnetic field update. is the modulus of the triaxial magnetic field, m n is the local standard geomagnetic field strength, and ρ is the preset empirical value.

[0187] In application, when P t m= 0, the geomagnetic field is not updated. t m =1, the geomagnetic field update is executed.

[0188] (1) Gravity acceleration update

[0189] Taking the acceleration of gravity as the observation vector, the measurement equation is established as follows:

[0190] Z1=h1(X)+V1

[0191] Where Z1 represents the measured value of gravity acceleration, h1(X) represents the true value of gravity acceleration, and V1 represents the measurement noise of gravity acceleration.

[0192] In the application,

[0193] Where, Represents the rotation matrix from the sensor coordinate system to the navigation coordinate system, g n The matrix representation of the gravitational acceleration relative to the navigation coordinate system, here is [0 0 -g], where g represents the gravitational acceleration.

[0194] In actual application,

[0195] Linearize the measurement equation to obtain the following formula:

[0196]

[0197] Where Z 1k represents the linearized measurement value of the accelerometer at time k, is the true value of the accelerometer after linearization, V 1k is the measurement noise after linearization of the accelerometer, H1 is the second Jacobian matrix, X k-1 Represents a prior estimate of the system state.

[0198] The second Jacobian matrix H1 of this embodiment is:

[0199]

[0200] (2) Update of geomagnetic field measurements

[0201] Taking the magnetic field as the observation vector, the measurement equation is established, including the following formula:

[0202] Z2=h2(X)+V2

[0203] Where Z2 represents the magnetic field measurement value, h2(X) represents the true value of the magnetic field, and V2 represents the magnetic field measurement noise.

[0204] In the application,

[0205] Linearize the measurement equation to obtain the following formula:

[0206]

[0207] Where Z 2k is the linearized magnetometer measurement value, V 2k is the measurement noise of the magnetometer after linearization, is the true value of the linearized magnetometer, H2 represents the third Jacobian matrix,

[0208] In application, the third Jacobian matrix H2 includes the following formula:

[0209]

[0210] This embodiment obtains a prediction result of the current state through state update, and then collects the measurement value of the current state. Combining the prediction value and the measurement value, an optimized estimate of the current state can be obtained, that is, the optimized quaternion is obtained through measurement update.

[0211] Step 2.4 Use the fused signal to train the kinematic mechanism model

[0212] Based on the state switching extended Kalman filter, the posture of each limb is estimated, and a kinematic mechanism model driven by posture quaternion is established to reconstruct the three-dimensional spatial posture of the rider in competitive equestrian sports, and realize the signal mapping between the spatial posture of the joints of the rider and the horse and the kinematic mechanism model.

[0213] The kinematic mechanism model can predict the spatial position of each joint of the rider and the horse, estimate the rider's limb posture, visualize the rider's posture at each moment during competitive equestrianism, and divide the phase labels of the rider and horse in competitive equestrianism based on the visualization results.

[0214] When applied, the motion cycles of walk, trot, trot and canter in competitive equestrian sports are analyzed. Combined with the phase division standard of competitive equestrian sports, the reconstructed three-dimensional spatial posture is used to mark the phase labels of walk, trot, trot and canter for the motion signals.

[0215] This embodiment uses the rider's pelvis as the zero point for training the kinematic mechanism model applied to the lower limb joints. Figure 6 .

[0216] Step 2.4.1 Obtain the spatial positions of the left and right pelvis a, e, thigh b, f, calf c, g, and foot joints d, h using the following formula:

[0217]

[0218]

[0219]

[0220]

[0221] Where, P a (t), P b (t), P c (t), P d (t) represents the spatial position of joint points a, b, c, and d, Represents the rotation quaternion of the pelvis at time t; Represents the rotation quaternion of the right thigh at time t; Represents the rotation quaternion of the right calf at time t; represents the rotation quaternion of the right foot at time t; * represents the quaternion conjugate matrix; V oa Represents the space vector corresponding to the right pelvis; V ab V represents the space vector corresponding to the right thigh; bc V represents the space vector corresponding to the right calf; cd Represents the space vector corresponding to the right foot.

[0222] When applied, the spatial positions of the joint points e, f, g, and h can also be obtained by traversing the above formula.

[0223] Step 2.4.2: Determine the four horse hoof contact events with the ground based on the spatial position of each joint point, and divide the labels based on the four horse hoof contact events with the ground. Figure 7 .

[0224] When applied, the label division process is combined with the phase division criteria for competitive equestrian sports published by the International Equestrian Federation to divide the phase labels of the walk, trot, trot and canter processes in equestrian sports.

[0225] (1) Slow walk

[0226] The slow walk is the slowest gait in equestrianism. In this gait, the horse's four hooves strike the ground in sequence, mainly consisting of four beats. According to the contact between the horse's hooves and the ground, a complete cycle can be divided into six stages and six phases. Figure 7 .

[0227] The various stages of walking are labeled A1, A2, A3, A4, A5, and A6. A1 indicates that the right hind hoof is off the ground; A2 indicates that the right front hoof is off the ground and the left front and hind hooves strike the ground; A3 indicates that the right hind hoof strikes the ground and only the right front hoof is off the ground; A4 indicates that the right front hoof strikes the ground and the left hind hoof is off the ground; A5 indicates that the left front hoof is off the ground and only the right front and hind hooves are on the ground; A6 indicates that the left hind hoof is off the ground and the other three hooves are on the ground.

[0228] (2) Saddle trot

[0229] During the saddle trot, the rider's hips are close to the horse's back, rising and falling with the horse's movement. The saddle trot is divided into five stages and six phases. Figure 7 .

[0230] The various phases of the trot are labeled B1, B2, B3, B4+B5, and B6. The interval between the left front hoof leaving the ground and striking the ground is IIB1; the interval between the right hind hoof leaving the ground and striking the ground is IIB2; the first airborne phase is IIB3; the interval between the right front hoof and the left hind hoof leaving the ground and striking the ground is IIB4+B5; and the second airborne phase is IIB6.

[0231] In this embodiment, B1 and B2 are separated for further subdivision.

[0232] (3) Quick steps up and down

[0233] In the trot, the horse's gait is the same as in the saddle trot. The difference is that in the trot, the rider will do a rising and falling movement on the horse's back. Similarly, the entire trot cycle is divided into five stages and six phases, refer to Figure 7 .

[0234] The various stages of the trot are labeled C1, C2, C3, C4+C5, and C6. The interval between the left front hoof leaving the ground and striking the ground is I C1; the interval between the right hind hoof leaving the ground and striking the ground is I C2; the first airborne period is I C3; the interval between the right front hoof and the left hind hoof leaving the ground and striking the ground is I C4+C5; and the second airborne period is I C6.

[0235] (4) Running

[0236] There are three rhythms in running, and the whole cycle can be divided into five stages and five phases. Figure 7 .

[0237] The running phases are labeled D1, D2, D3, D4, and D5. D1 indicates that the left hind hoof strikes the ground while the other hooves leave the ground, which initiates the entire pulse; D2 indicates that the right hind hoof and the left front hoof (diagonal hoof) strike the ground at the same time, with only the right front hoof leaving the ground; D3 indicates that the right front hoof strikes the ground and the left hind hoof leaves the ground; D4 indicates that the left front hoof and the right hind hoof (diagonal hoof) leave the ground at the same time, with only the right front hoof in contact with the ground; D5 indicates the airborne state, with all four hooves off the ground.

[0238] Example 3

[0239] Based on Example 1 or 2, this example introduces in detail the phase segmentation process of competitive equestrian sports.

[0240] refer to Figure 1 ,This implementation utilizes the phase segmentation model to achieve the phase segmentation of walk, trot, trot and canter in competitive equestrian ,sports, thereby obtaining the phase characteristics of the rider in equestrian training.

[0241] The phase feature here specifically refers to the phase ratio contained in the periodic motion process of each of the four periodic gaits of the rider, namely, the slow walk, the saddle trot, the trot, and the canter. In the implementation process, taking the slow walk process as an example, it is only necessary to input the acquired motion signal and its corresponding label into the phase segmentation model TCN-Bi-LSTM for training. Once the model training is completed, it can be used for real-time segmentation of the slow walk motion phase of the rider during equestrian sports. The segmented phase here specifically refers to a step-shaped time series, refer to Figure 8 The implementation of other saddle trots, ascending and descending trots, and cantering is similar to that of the slow walk. During the implementation process, the rider only needs to wear the sensor and perform the slow walk. After the PC receives the sensor data of the slow walk, it can use the software equipped with this model to divide the phases of the equestrian movement process, thereby analyzing whether the rider's technique is standard and whether the technical level is up to standard. This quantitative indicator can be used to assist daily training and improve competitive level.

[0242] The phase segmentation model implemented in this paper is a TCN-Bi-LSTM deep fusion model.

[0243] The phase segmentation model implemented in this embodiment includes an input layer having multiple input nodes. When applied, a time series is constructed using the motion signal and its labeled labels, and the time series is input into the input layer.

[0244] In practical application, the time series is Xp=[X1,X2,X3,…,X k ].

[0245] The phase segmentation model of this implementation includes a temporal convolutional network (TCN) layer, which extracts sequence features of the time series.

[0246] In application, reference Figure 9 ,TCN consists of three residual blocks. The first residual block contains two causal dilated convolutional layers with corresponding kernel size, dilation factor, and number of filters of 9, 1, and 4 respectively; the dilation factors of the second and third residual blocks are 2 and 4 respectively, and the other parameters are similar to those of the first residual block.

[0247] The phase segmentation model implemented in this paper sets a bidirectional long short-term memory neural network (Bi-LSTM) layer after the TCN layer, and uses a Bi-LSTM network with 300 hidden units to learn forward and backward information to process the output of TCN and obtain the output vector

[0248] In application, reference Figure 10 , the iterative process of a single LSTM unit can be expressed as:

[0249] f t =σ(W f [h t-1 ,y t ]+b f )

[0250] i t =σ(W i [h t-1 ,y t ]+b i )

[0251]

[0252]

[0253] o t =σ(W o [h t-1 ,y t ]+b o )

[0254] h t =o t ⊙tanh(a t )

[0255] where f t ,i t ,o t They are the outputs corresponding to the input gate, forget gate and output gate respectively. For memory cells, h t is the hidden state at time t, h t-1 is the hidden state at time t-1, σ represents the sigmoid activation function, tanh represents the hyperbolic tangent function, b f 、bi 、b c 、b o Represent the bias of input gate, forget gate, memory cell and output gate respectively, W f 、W i 、W c 、W o Represents the weight matrices corresponding to the input gate, forget gate, memory cell and output gate, y t Represents the input vector.

[0256] Bi-LSTM is composed by stacking two LSTM layers in opposite directions. Figure 11 This bidirectional structure can preserve the information of the forward and backward directions of the sequence, and the information considered is more comprehensive than that of the unidirectional structure. In Bi-LSTM, the final output They are affected by the output of the bidirectional network respectively, which can be expressed as:

[0257]

[0258] Where, and Represent the weights of the forward and backward LSTM networks respectively, and Represent the output of the forward and backward LSTM networks, b y represents the bias of the input vector, Represents the output of the Bi-LSTM network at time t.

[0259] During the training of the phase segmentation model, the phase segmentation model optimizer is Adam, and the maximum number of iterations, minimum batch size, and gradient threshold are 80, 128, and 3, respectively.

[0260] The phase segmentation model implemented in this paper uses a Dropout layer after the Bi-LSTM layer to prevent overfitting, and sets a fully connected layer after the Dropout layer and an output layer after the fully connected layer to achieve phase segmentation.

[0261] In applications, phase division includes the following formula

[0262]

[0263] In the formula, Softmax() is the Softmax activation function, which mainly maps the output of the network as the input of the function to the output end, thereby outputting the phase division result, x i represents the i-th element in x, x j represents the jth element in x, ∑ j exp(x j) is the sum of the derivatives of all elements in x, exp(x i ) is the element x i Find the derivative.

[0264] In practical applications, the activation function uses the output of the bidirectional long short-term memory neural network as the input of the activation function, thereby mapping it to the output layer as a phase segmentation result, namely, a phase feature.

[0265] The present invention uses a phase segmentation model to segment the phases of slow walk, trot, trot and canter in competitive equestrian sports, thereby obtaining the phase characteristics of the rider at each moment in equestrian training. Figure 12 The phase segmentation model based on the TCN-Bi-LSTM network proposed in this paper has generally high phase segmentation accuracy and can accurately segment the rich phase information contained in the four gaits used in equestrianism. Analyzing the phase characteristics in competitive equestrianism can provide a certain understanding of the interaction between riders and horses, as well as their riding skills. Therefore, phase segmentation in competitive equestrianism can provide quantitative data reference for the subsequent improvement of competitive equestrian training methods.

[0266] Example 4

[0267] Based on any one of Examples 1-3, this example compares the accuracy of phase feature segmentation of competitive equestrian sports using the method of the present application with methods such as the bidirectional long short-term memory neural network Bi-LSTM, the gated recurrent neural unit GRU, the long short-term memory neural network LSTM, the support vector machine SVM, and the K-nearest neighbor KNN.

[0268] refer to Figure 13-16 It can be seen that in the phase segmentation of slow walk, trot, trot and canter in competitive equestrian sports, the accuracy of phase feature segmentation at each moment of the TCN-Bi-LSTM method of the present application is higher than that of other methods.

[0269] In summary, the present invention uses wearable MEMS inertial sensors to collect motion signals of riders and horses during competitive equestrian sports. The motion signals are reconstructed in three dimensions, and the phase labels of the four basic gaits in competitive equestrian sports are divided according to the three-dimensional reconstruction results. Finally, the motion signals and the corresponding phase labels are input into the TCN-Bi-LSTM deep fusion model to perform fine segmentation of the phase features of competitive equestrian sports.

[0270] Example 5

[0271] This embodiment provides a phase segmentation system for competitive equestrian sports, comprising:

[0272] A sampling module for acquiring motion signals of riders and horses;

[0273] The marking module is used to pre-process the motion signal and mark the pre-processed motion signal.

[0274] The phase segmentation module is used to construct a time series using motion signals and their corresponding labeled labels, and input the time series into the phase segmentation model to perform phase segmentation on the rider and horse during competitive equestrian sports to obtain phase features.

[0275] For the implementation of the specific functions of the above functional modules, please refer to the relevant contents of the methods described in Examples 1-4.

[0276] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0277] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0278] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0279] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0280] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which are all protected by the present invention.

Claims

1. A phase segmentation method for competitive equestrian sports, characterized in that: The following steps are involved: Obtain movement signals from riders and horses; Preprocess the motion signal and label the preprocessed motion signal. The motion signal and its corresponding label are used to construct a time series, which is then input into a phase segmentation model to perform phase segmentation on the rider and horse during competitive equestrian sports and obtain phase features. The obtaining of the motion signals of the rider and the horse comprises: Use wearable wireless body area network data acquisition system to obtain the motion signals of riders and horses; The wearable wireless body area network data acquisition system includes a plurality of sensor nodes; Each of the sensing nodes includes a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer; The phase segmentation model includes an input layer, a temporal convolution layer, a bidirectional long short-term memory neural network layer, a Dropout layer, a fully connected layer and an output layer arranged in sequence.

2. The competitive equestrian phase segmentation method according to claim 1, characterized in that: The motion signal includes three-axis acceleration signals, three-axis angular velocity signals and three-axis geomagnetic field intensity signals of the joints of the rider and the horse at m consecutive moments.

3. The competitive equestrian phase segmentation method according to claim 2, characterized in that: The pre-processing motion signal comprises: The three-axis angular velocity signals are processed to remove zero drift and obtain the corrected angular velocity; Perform ellipsoid fitting processing on the three-axis acceleration signal and the three-axis geomagnetic field intensity signal to obtain the corrected acceleration and the corrected geomagnetic field intensity; The attitude quaternion between the sensing node and the binding joint point is determined using the corrected angular velocity, corrected acceleration and corrected geomagnetic field strength.

4. The competitive equestrian phase segmentation method according to claim 1, characterized in that: The marking tags include marking tags for four basic gaits: walk, trot, trot and canter.

5. The competitive equestrian phase segmentation method according to claim 4, characterized in that: The slow-step marking labels include: The right hind hoof leaves the ground, The right front hoof leaves the ground, the left front hoof and hind hoof hit the ground, The right hind hoof strikes the ground, and only the right front hoof leaves the ground. The right front hoof hits the ground, only the left hind hoof leaves the ground, The left front hoof is off the ground, and only the right front hoof and hind hoof strike the ground. The left hind hoof leaves the ground, and the other three hooves hit the ground; The labeling of the up and down quick step includes: The interval between the left front hoof leaving the ground and striking the ground Ⅰ, The interval between the right hind hoof leaving the ground and striking the ground Ⅰ, The first vacant period I, The interval between the right front hoof and the left hind hoof from leaving the ground to striking the ground Ⅰ, Second Vacancy Period I; The marking labels for the saddle trot include: The interval between the left front hoof leaving the ground and striking the ground Ⅱ, The interval between the right hind hoof leaving the ground and striking the ground Ⅱ The first vacant period II, The interval between the right front hoof and the left hind hoof from leaving the ground to striking the ground Ⅱ, The second empty phase II; The running marking tags include: The left hind hoof strikes the ground, while the other hooves leave the ground. The right hind hoof and the left front hoof hit the ground at the same time, and only the right front hoof leaves the ground. The right front hoof hits the ground, the left hind hoof leaves the ground, The left front hoof and the right hind hoof leave the ground at the same time, and only the right front hoof hits the ground. In the airborne state, all four horseshoes are off the ground.

6. The competitive equestrian phase segmentation method according to claim 1, characterized in that: The phase segmentation model includes the following formula: ， Where Softmax( ) is the Softmax activation function, which mainly maps the output of the bidirectional long short-term memory neural network layer as the input of the function to the output layer, thereby outputting the phase division result, that is, the phase feature, express The i-th element in express The jth element in Yes The sum of the derivatives of all elements in , It is the element Find the derivative.

7. A phase segmentation system for competitive equestrian sports, characterized in that: The method for implementing the phase segmentation method of competitive equestrian sports according to any one of claims 1 to 6 comprises: A sampling module for acquiring motion signals of riders and horses; The marking module is used to pre-process the motion signal and mark the pre-processed motion signal. The phase segmentation module is used to construct a time series using motion signals and their corresponding labeled labels, and input the time series into the phase segmentation model to perform phase segmentation on the rider and horse during competitive equestrian sports to obtain phase features.

8. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any one of the methods according to claims 1 to 6.

Images