Method for evaluating children's object control ability based on nonlinear dynamics

Abstract

The present invention discloses a method for evaluating children's object control ability based on nonlinear dynamics. First, the Kinect sensor is used to collect the spatial position information of the motion skeleton points of the tester's hand, elbow and shoulder, and the inertial sensor is used to collect the motion biomechanical signals of the tester's hand; then the time-frequency domain features and sample entropy features of human motion are extracted from the spatial position information of the skeleton points and the motion biomechanical signals; the principal component analysis method is used to fuse the time-frequency domain features and sample entropy features; finally, an RF-based evaluation model for children's object control ability is constructed to realize the classification of children's object control ability. The present invention solves the problems of identifying children's object control ability and low recognition accuracy caused by the inability to quantify the target in the prior art, and realizes the hierarchical evaluation of children's object control ability.

Description

Technical Field

[0001] The present invention belongs to the technical field of the evaluation of children's physical movement coordination ability, and particularly relates to a method for evaluating children's object control ability based on nonlinear dynamics. Background Technique

[0002] As an important part of children's development, the movement coordination ability is a key basic condition for children's physical and mental health, motor skills, and growth and development. Children with Developmental Coordination Disorder (DCD) usually have their abilities in aspects such as language and cognition affected, which may affect their self-care ability and social-related functions in both childhood and the long term in adulthood. Therefore, exploring scientific and efficient methods for children's movement coordination ability and early diagnosis is of great significance for improving the recognition and diagnosis ability and early detection and intervention.

[0003] The object control ability is one of the important contents for measuring the coordinated development level of the large muscle movements of the human body, generally referring to action skills such as throwing, kicking, and hitting. These actions can define whether the coordination between the child's torso and limbs is coordinated. Therefore, the development of the object control ability plays a crucial role in promoting children to reach a natural and smooth level of motor skills.

[0004] With the rapid development of sensor technology, researchers can use sensors to measure features that cannot be observed by the naked eye from different perspectives such as physiology and kinematics. The measurement results have high sensitivity, effectiveness, and repeatability, providing a way for a more accurate quantification of children's movement coordination ability assessment. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for evaluating children's object control ability based on nonlinear dynamics, which solves the problems of the recognition of children's object control ability and the low recognition accuracy caused by the inability to quantify the target in the prior art, and realizes the hierarchical evaluation of children's object control ability.

[0006] The technical solution adopted by the present invention is that the method for evaluating children's object control ability based on nonlinear dynamics is specifically implemented according to the following steps:

[0007] Step 1: Use a Kinect sensor to collect the spatial position information of the movement bone points of the tester's hand, elbow, and shoulder, and use an inertial sensor to collect the movement biomechanical signals of the tester's hand;

[0008] Step 2: Extract the time-frequency domain features and sample entropy features of human movement from the spatial position information of the bone points and the movement biomechanical signals;

[0009] Step 3: Use the principal component analysis method to fuse the time-frequency domain features and sample entropy features;

[0010] Step 4: Construct an RF-based evaluation model for children's object control ability to achieve the classification of children's object control ability.

[0011] The features of the present invention also lie in that

[0012] Step 1 is specifically implemented according to the following steps:

[0013] Use an inertial sensor to collect acceleration and angular velocity signals, and extract time-domain signal features including signal peak value, peak discrete coefficient, and signal period. At the same time, frequency-domain signal features including signal peak power, peak frequency, and peak total power can also be extracted;

[0014] Use Kinect to obtain the spatial position data of the human body bones, so as to extract some spatial position features and joint angle features.

[0015] Step 2 is specifically implemented according to the following steps:

[0016] Step 2.1: Extract the convex hull feature of the ball landing point and the elbow angle feature;

[0017] Step 2.2: Extract the sample entropy:

[0018] The specific calculation of the sample entropy is as follows:

[0019] Step 2.2.1: Assume that the set of X-direction sequences of the sample hand trajectory is {a x (1), a x (2),..., a x (n)}, then the calculation of the sample entropy of this sequence is as follows:

[0020] According to the sequence {a x (1), a x (2),..., a x (n)}, form a group of m-dimensional vectors A(i) in order:

[0021] A(i) = [a x (1), a x (2),..., a x (i + m - 1)], i = 1, 2,..., i + m - 1 (14)

[0022] Reconstruct the m-dimensional vectors A(1), A(i),..., A(z), z = n - m + 1

[0023] Step 2.2.2: Calculate the Chebyshev distance d[A(i), A(j)] between the vectors A(i) and A(j):

[0024]

[0025] Given a threshold r, calculate the number n of d[A(i), A(j)] < r when i < n - m + 1 i , divide by the total number of distances n - m + 1, and denote this result as The calculation formula is as follows:

[0026]

[0027] where m represents the embedding dimension; r represents the similarity tolerance, r ∈ [0.1*std, 0.25*std], and std represents the standard deviation of the X-direction sequence of the hand trajectory;

[0028] Step 2.2.3: Expand the dimension to m + 1, repeat Steps 2.2.2 - 2.2.4, and similarly calculate

[0029] Then the sample entropy of the X-direction sequence of the hand trajectory is:

[0030]

[0031] Step 2.1 is specifically implemented according to the following steps:

[0032] The envelope formed by the projection points of the hand bone points on the ground is approximately regarded as the envelope formed by the ball landing points. The size and dispersion degree of the envelope area reflect the level of the hand's control ability over the ball. The calculation methods for the envelope area and dispersion degree are as follows:

[0033] Step 2.1.1: Assume that a sample projection point set is {(x1, z1), (x2, z2),..., (x n , z n )}, the coordinate of the j-th projection point is (x j , z j ). Use the Euclidean distance formula to obtain the length C of the projection trajectory of the center of gravity of this sample XOZ , and the calculation is as follows:

[0034]

[0035] Step 2.1.2: Reorder the point set according to the magnitude of the vertical coordinate z value, and relabel it as Q1, Q2, Q3....Q n ; if there are points with the same minimum z coordinate, select the point with the minimum abscissa x as Q1;

[0036] Step 2.1.3: Sort the other points in ascending order according to the included angle value formed by Q1Q i and the horizontal line where Q1 is located; when the included angles are the same, use Q1Q iSort the distances from small to large to complete the re - numbering of the remaining points Q2, Q3....Q n ;

[0037] Step 2.1.4, Q i+1 Concavity and convexity judgment: Assume that Q i , Q i+1 , Q i+2 The coordinates of the three points are (x i , z i ), (x i+1 , z i+1 ), (x i+2 , z i+2 ). Calculate the value of the concavity - convexity discriminant variable K. The formula is:

[0038] K = (x i+1 z i+2 - x i+2 z i+1 )+(x i+2 , z i - x i , z i+2 )+(x i , z i+1 - x i+1 , z i ) (7)

[0039] If K > 0, Q i+1 is a convex point;

[0040] If K < 0, it is a concave point, delete Q i+1 ;

[0041] If K = 0, it is a neutral point, choose to delete this point;

[0042] Step 2.1.5. After traversing all the points in the point set, obtain k vertices of the convex hull graph, and mark them counter - clockwise as q1(x1, z1), q2(x2, z2),..., q k (x k , z k ). Connect the k vertices to obtain the convex hull. Calculate it using the Euclidean distance formula according to the following formula:

[0043] for i = 0, 1, 2,..., k - 2:

[0044]

[0045]

[0046]

[0047] Obtain k - 2 groups of distance combinations. The set is represented as follows:

[0048] {(d1, d 1_k , d 2_k ), (d2, d 2_k , d 3_k ),..., (d k-2 , d k-2_k , d (k-1)_k )}

[0049] Among them, d i represents the distance between adjacent vertices, d i_k represents the distance between two points q1q k , and d (i+1)_k represents the distance between two points q i+1 q k ;

[0050] Step 2.1.6: Use Heron's formula and the combined triangle method to find the convex hull area S XOZ , and the calculation is as follows:

[0051]

[0052]

[0053] Similarly, the envelope area and the trajectory length of the projection of the centroid data on the YOX plane are respectively denoted as S YOX , C YOX , and the envelope area and the trajectory length of the projection of the centroid on the YOZ plane are denoted as S YOZ , C YOZ ;

[0054] The elbow angle is defined as follows:

[0055] Using (a x , a y , a z ), (b x , b y , b z ), (c x , c y , c z ) to represent the three-dimensional spatial coordinates of the hand, elbow, and shoulder respectively, the spatial vector of the elbow - hand is denoted as H = (a x - b x , a y - b y , a z - b z ), and the spatial vector of the elbow - shoulder is denoted as G = (c x - b x , c y - b y , c z - bz ), the elbow angle is expressed as:

[0056]

[0057] Step 3 is specifically implemented according to the following steps:

[0058] The method of principal component analysis is used to fuse multiple features and extract representative comprehensive evaluation features. The specific calculation method is as follows:

[0059] Suppose there are M object control ability samples {X 1 , X 2 ,..., X m}, and each sample has N-dimensional features Each feature x j has its own eigenvalue;

[0060] Center all features and calculate the mean μ of the jth eigenvalue j

[0061]

[0062] Find the covariance matrix D

[0063]

[0064]

[0065] where cov(x p , x q ) is the covariance of any two features, and p, q ∈ ∈ [1, N];

[0066] The relationship between the eigenvalues λ of the covariance matrix D and the corresponding eigenvectors α is as follows:

[0067] Dα = λα (21)

[0068] There are N corresponding eigenvalues λ in the formula, denoted as {λ1...λ i ...λ N}, and each λ i corresponds to an eigenvector α i . Sort the eigenvalues λ from smallest to largest, select the top k largest of these N λ, and obtain the corresponding eigenvectors. Each eigenvector is denoted as (λ i , α i ), where the value of i ranges from 1 to k, thus obtaining a set of {(λ1, α1), (λ2, α2),...,(λ i , α i ),...,(λ k , αk )};

[0069] For each sample X i , project the original features onto the selected feature vectors to obtain the k-dimensional new features after dimensionality reduction The calculation method is as follows:

[0070]

[0071] Finally, obtain the k-dimensional new features Y of M samples, which is expressed as follows:

[0072]

[0073] The matrix Y is the result of fusing the time-frequency domain features and sample entropy features.

[0074] Step 4 is specifically implemented according to the following steps:

[0075] Evaluate the children's object control ability based on a random forest. The random forest integrates multiple trees. The basic unit of the random forest is a decision tree. The construction process of the decision tree is as follows:

[0076] Step 4.1: Take the output result Y of the above step 3 as the sample set. The i-th dimension feature of the sample set Y is denoted as y i ,

[0077] Sort the feature y i in ascending order of value, denoted as Based on the partition point t, divide Y into subsets Y t - and Y t + , where Y t - contains those samples whose values on the feature y i are not greater than t, and Y t + contains those samples whose values on the feature y i are greater than t;

[0078] For the feature y i , take the midpoint of the interval as the candidate partition point to obtain a set of m - 1 partition points:

[0079]

[0080] Step 4.2: Traverse and examine the set Select the optimal partition point to divide the sample set. The formula is as follows:

[0081]

[0082] Among them, Ent(Y) is the information entropy calculated by the sample set based on the label values, and Gain(Y, y i , t) is the information gain after the sample set is bisected based on the partitioning point t, and Gain(Y, y i ) is the maximum information gain after the sample set is partitioned based on t;

[0083] Step 4.3: Repeat Steps 4.1 - 4.2 to obtain the information gains of the k-dimensional features, and the set is represented as follows:

[0084] G = max{Gain(Y, y1), Gain(Y, y2),..., Gain(Y, y k )} (26)

[0085] Then the feature corresponding to G is selected as the best splitting feature of the dimension and used as the root node. By repeating the above process recursively, a decision tree can be constructed;

[0086] Step 4.4: The process of constructing a random forest is as follows:

[0087] The sample set Y is represented with new symbols as follows:

[0088]

[0089] The input is the training data set Y = {(X1, y1), (X2, y2),..., (X m , y m )};

[0090] The output is the result H(X) of the final strong classifier.

[0091] Step 4.4 is specifically implemented according to the following steps:

[0092] Step 4.4.1: Given that m samples form the training set Y, use sampling with replacement to generate a new training set Y';

[0093] Step 4.4.2: Randomly select n features from all features T, where n < T, to form a new training set Then select the best splitting feature from the n features as the node and use to construct a decision tree;

[0094] Step 4.4.3: Iterate Steps 4.4.1 - 4.4.2 d times, and mark the training set generated each time as where i = (1, 2,..., d), and construct d decision trees;

[0095] Step 4.4.4: These d decision trees together constitute a random forest. The sample category is voted on by each tree. Equation 28 represents the class with the most votes, that is, the final classification:

[0096]

[0097] Among them, H(X) represents the combined classification model, h i (x) represents a single decision tree classification model, y represents the output variable, and I(x) is an indicator function;

[0098] Step 4.5: Evaluation level division. According to the children's basic motor skills evaluation tool, the evaluation level of object control ability is determined to be strong, general, and weak, and the corresponding score intervals are [0, 59], [60, 79], and [80, 100]. The probability interval is mapped to the corresponding score interval through a linear function;

[0099] The mapping function is F = kf + c, where f is the probability corresponding to the classification category, F is the score, and the mapping relationship is as follows:

[0100] (1) If the sample classification result is weak in object control ability, then the probability interval [1 / 3 - 1] is mapped to the interval [0, 59]:

[0101] F = 88.5*f - 29.5 (29)

[0102] (2) If the sample classification result is general in object control ability, then the probability interval [1 / 3 - 1] is mapped to the interval [60, 79]:

[0103] F = 28.5*f + 50.5 (30)

[0104] (3) If the sample classification result is strong in object control ability, then the probability interval [1 / 3 - 1] is mapped to the interval [80, 100]:

[0105] F = 30*f + 70 (31).

[0106] The beneficial effect of the present invention is a method for evaluating children's object control ability based on nonlinear dynamics. The nonlinear dynamics algorithm is introduced to extract the sample entropy features of motion signals, realizing the multi-feature acquisition of human motion signals; based on the multi-sensor data fusion theory, the method of principal component analysis is used to realize the fusion of multi-features; finally, the RF classification model is constructed using the random forest method, realizing the classification of children's object control ability. Brief Description of the Drawings

[0107] Figure 1 It is the convex hull feature of the ball landing point and the elbow angle feature in the method for evaluating children's object control ability based on nonlinear dynamics of the present invention;

[0108] Figure 2 This is the sample entropy calculation process in the method for evaluating children's object control ability based on nonlinear dynamics of the present invention;

[0109] Figure 3 This is the schematic diagram of the random forest model in the method for evaluating children's object control ability based on nonlinear dynamics of the present invention. Detailed implementation manners

[0110] The present invention will be described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0111] The method for evaluating children's object control ability based on nonlinear dynamics of the present invention combines Figures 1 to 3 , and is specifically implemented according to the following steps:

[0112] Step 1: Use the Kinect sensor to collect the spatial position information of the motion skeleton points of the tester's hand, elbow, and shoulder, and use the inertial sensor to collect the motion biomechanical signals of the tester's hand;

[0113] Step 1 is specifically implemented according to the following steps:

[0114] Use the inertial sensor to collect the acceleration and angular velocity signals, and extract the time-domain signal features including the signal peak value, peak discrete coefficient, and signal period. At the same time, the frequency-domain signal features including the signal peak power, peak frequency, and peak total power can also be extracted;

[0115] Use the Kinect to obtain the spatial position data of the human skeleton, so as to extract some spatial position features and joint angle features.

[0116] Step 1.1, resultant acceleration

[0117] Acceleration is often used to reflect the vector of the speed change rate of a moving object. According to Newton's second law, the magnitude of the acceleration determines the magnitude of the ball-patting force, that is, the greater the magnitude of the acceleration, the greater the energy reflected by the ball-patting motion. At the same time, in order to avoid the errors caused by the analysis of the acceleration and angular velocity in a single-axis direction, the resultant acceleration is introduced to supplement the reflection of the intensity of hand movement. First, calculate the sum of the squares of the three-axis accelerations, and then take the square root of the result. The calculation formula is as follows:

[0118]

[0119] In the formula: J t is the resultant acceleration, and ax, ay, and az are the acceleration values of each dimension respectively.

[0120] Step 1.2, root mean square

[0121] RMS is a commonly used feature in time-domain feature extraction methods, which has the advantage of fast calculation speed. Its core idea is square filtering, which has a certain smoothing effect and can remove some noise in acceleration and angular velocity. RMS reflects the energy value of each hand slap, which helps to identify the level of object control ability. The calculation formula is as follows:

[0122]

[0123] In the formula: m is the length of the signal sequence, and a i is the acceleration value.

[0124] Step 1.3, the mean value of the peak signal

[0125] The peak refers to the extreme difference value of the signal within one period, and this index describes the signal change range. In each ball-patting action, the acceleration and angular velocity signals will generate a relatively obvious peak and valley. The present invention extracts the peaks of the acceleration and angular velocity signals to construct a new signal sequence, and calculates the mean value of the peak signals of the angular velocity and acceleration to reflect the situation of the signal energy concentration trend. First, sum the peak sequence, and then perform an averaging operation. The calculation formula is as follows:

[0126]

[0127] Peak_avg is the mean value of the peak sequence of the acceleration signal, n is the number of peak points extracted from the acceleration signal, and a i _peak i represents the peak of the acceleration signal in the i-th action cycle.

[0128] Step 1.4, the peak signal coefficient of variation

[0129] The coefficient of variation is often used to compare the degree of dispersion of two sets of data. When the units or averages of two variables are different, the standard deviation cannot be used to compare their degrees of dispersion, but the ratio of the standard deviation to the average, that is, the coefficient of variation, needs to be used to compare, so as to eliminate the influence of the measurement scale and dimension. The calculation formula is as follows:

[0130]

[0131] Peak_cv represents the peak signal coefficient of variation of the acceleration signal, Peak_avg represents the mean value of the peak sequence of the acceleration signal, and Peak_std represents the standard deviation of the peak sequence of the acceleration signal.

[0132] Step 1.5, the mean value of the signal period

[0133] The period of the signal can indirectly reflect the frequency of the signal. In the ball-bouncing motion, the time the subject takes to bounce the ball once is defined as the ball-bouncing period T. Assume that the frame index corresponding to the acceleration peak point is (p0, p1..., p n ), then define the nth period T n =(p n -p n-1 ) / f. The cycle average T_avg is calculated as follows:

[0134]

[0135] Step 1.6, power spectrum characteristics

[0136] While time-domain feature extraction methods offer computational simplicity and can more intuitively capture the effective information from bouncing activity data, some energy information cannot be detected using time-domain methods alone. Therefore, frequency-domain features can be used to supplement the effective information in acceleration and angular velocity signals. Fast Fourier transforms can convert signals into frequency-domain signals. The energy information contained in the signals can be captured by peak power, peak frequency, and total signal power, as calculated from the power spectral density.

[0137] Step 2: extracting the time-frequency domain features and sample entropy features of human motion from the spatial position information of the skeleton points and the motion biomechanical signals;

[0138] Step 2 is implemented as follows:

[0139] Step 2.1, extract the convex hull feature of the ball landing point and the elbow angle feature;

[0140] Step 2.2, extract sample entropy:

[0141] Sample entropy is a nonlinear dynamic parameter used to quantify the complexity of time series. It has the advantages of high sensitivity and strong anti-interference ability.

[0142] The object control motion signal is a time series. In the object control motion, the smaller the sample entropy value, the lower the complexity of the object control motion signal; on the contrary, the larger the sample entropy value, the higher the complexity of the object control motion signal. The sample entropy calculation process is as follows: Figure 2 The specific calculation is as follows:

[0143] Step 2.2.1, assume that the set of sample hand trajectory X-direction sequences is {a x (1),a x (2),...,a x (n)}, the sample entropy of the sequence is calculated as follows:

[0144] According to the sequence {a x (1),a x(2),...,a x (n)} are sequentially composed into a group of m-dimensional vectors A(i):

[0145] A(i) = [a x (1), a x (2),..., a x (i + m - 1)], i = 1, 2,..., i + m - 1 (14)

[0146] Reconstruct the m-dimensional vectors A(1), A(i),..., A(z), z = n - m + 1

[0147] Step 2.2.2, calculate the Chebyshev distance d[A(i), A(j)] between vectors A(i) and A(j):

[0148]

[0149] Given a threshold r, calculate the number n of d[A(i), A(j)] < r when i < n - m + 1 i , divide by the total number of distances n - m + 1, and record this result as The calculation formula is as follows:

[0150]

[0151] Among them, m represents the embedding dimension, usually taking 1 or 2; r represents the similarity tolerance, r ∈ [0.1*std, 0.25*std], where std represents the standard deviation of the X-direction sequence of the hand trajectory;

[0152] Step 2.2.3, expand the dimension to m + 1 dimensions, repeat Steps 2.2.2 - 2.2.4, and similarly calculate

[0153] Then the sample entropy of the X-direction sequence of the hand trajectory is:

[0154]

[0155] Step 2.1 is specifically implemented according to the following steps:

[0156] The manipulation of the ball is not only related to the kinematic characteristics of the hand movement, but also some spatial positions and angular changes in the overall coordinated cooperation process of the hand-forearm-humerus should be fully considered, such as Figure 1 , for example, the more concentrated the ball landing points are, the more stable the hand's control of the ball is. The envelope formed by the projection points of the hand bone points on the ground is approximately regarded as the envelope formed by the ball landing points. The size and dispersion degree of the envelope area also reflect the hand's control ability level to a certain extent. The calculation methods of the envelope area and dispersion degree are:

[0157] Step 2.1.1, assume that a sample projection point set is {(x1,z1),(x2,z2),...,(x n ,z n )}, the coordinates of the j-th projection point are (x j ,z j ), the Euclidean distance formula is used to calculate the length C of the sample centroid projection trajectory XOZ , calculated as follows:

[0158]

[0159] Step 2.1.2, reorder the point set according to the size of the vertical coordinate z value, and use Q1, Q2, Q3....Q n Re-label. If there are points with the same minimum z coordinate, select the point with the minimum x coordinate and record it as Q1.

[0160] Step 2.1.3, press Q1Q on the other points i The angles formed with the horizontal line where Q1 is located are sorted in ascending order; if the angles are the same, Q1Q i Sort the distances from small to large, and complete the remaining points Q2, Q3....Q n renumbering of

[0161] Step 2.1.4, Q i+1 Convexity judgment: Assumption Q i , Q i+1 , Q i+2 The coordinates of the three points are (x i ,z i ), (x i+1 ,z i+1 ), (x i+2 ,z i+2 ), calculate the value of the concave-convex discriminant variable K, the formula is:

[0162] K=(x i+1 z i+2 -x i+2 z i+1 )+(x i+2 ,z i -x i ,z i+2 )+(x i ,z i+1 -x i+1 ,z i ) (7)

[0163] K>0,Q i+1 is a convex point;

[0164] K<0, it is a concave point, delete Q i+1 ;

[0165] K = 0, neutral point, select to delete this point;

[0166] Step 2.1.5. After traversing all points in the point set, obtain k vertices of the convex hull graph, and label them counterclockwise as q1(x1, z1), q2(x2, z2),..., q k (x k , z k ). Connect the k vertices to obtain the convex hull, and calculate it using the Euclidean distance formula according to the following formula:

[0167] for i = 0, 1, 2,..., k - 2:

[0168]

[0169]

[0170]

[0171] Obtain k - 2 groups of distance combinations, and the set is expressed as follows:

[0172] {(d1, d 1_k , d 2_k ), (d2, d 2_k , d 3_k ),..., (d k-2 , d k-2_k , d (k-1)_k )}

[0173] Among them, d i represents the distance between adjacent vertices, d i_k represents the distance between q1q k two points, d (i+1)_k represents the distance between q i+ 1q k two points;

[0174] Step 2.1.6. Use Heron's formula and the combination triangle method to find the convex hull area S XOZ , and calculate it according to the formula:

[0175]

[0176]

[0177] Similarly, the envelope area and trajectory length of the projection of the centroid data on the YOX plane are respectively expressed as S YOX , C YOX , and the envelope area and trajectory length of the projection of the centroid on the YOZ plane are expressed as S YOZ , C YOZ ;

[0178] Meanwhile, the strength of the control ability not only affects the range of the ball landing point, but also causes varying degrees of changes in the elbow angle sequence. The degree of dispersion can not only reflect the concentration degree of the ball landing point, but also reflect the rhythm and stability of the changes in the elbow angle sequence. The larger the coefficient of dispersion, the greater the deviation between the angle sequences, indicating that the rhythm and stability of dribbling are weaker. On the contrary, the smaller the coefficient of dispersion, the smaller the deviation between the angle sequences. The elbow angle is defined as follows:

[0179] Let (a x , a y , a z ), (b x , b y , b z ), (c x , c y , c z ) represent the three-dimensional spatial coordinates of the hand, elbow, and shoulder respectively. The spatial vector from the elbow to the hand is expressed as H = (a x - b x , a y - b y , a z - b z ), and the spatial vector from the elbow to the shoulder is expressed as G = (c x - b x , c y - b y , c z - b z ). Then the elbow angle is expressed as:

[0180]

[0181] Step 3: Use the principal component analysis method to fuse the time-frequency domain features and sample entropy features;

[0182] Step 3 is specifically implemented according to the following steps:

[0183] The present invention uses multi-sensors to collect multivariate signal features. Due to the complex internal coupling relationship between muscles in object control movement, there is an easy problem of feature redundancy. The principal component analysis method is used to fuse the multivariate features and extract representative comprehensive evaluation features. The specific calculation method is as follows:

[0184] Suppose there are M object control ability samples {X 1 , X 2 ,..., X m}, and each sample has N-dimensional features Each feature x j has its own eigenvalue;

[0185] Center all features and calculate the mean μ of the j-th eigenvalue j

[0186]

[0187] Find the covariance matrix D

[0188]

[0189]

[0190] where cov(x p , x q ) is the covariance of any two features, and p, q ∈ ∈ [1, N];

[0191] The relationship between the eigenvalues λ and the corresponding eigenvectors α of the covariance matrix D is as follows:

[0192] Dα = λα (21)

[0193] There are N corresponding eigenvalues λ in the formula, denoted as {λ1...λ i ...λ N}, and each λ i corresponds to an eigenvector α i . Sort the eigenvalues λ in ascending order, select the largest k of these N λ, and obtain the corresponding eigenvectors. Each eigenvector is denoted as (λ i , α i ), where i ranges from 1 to k, thus obtaining a set {(λ1, α1), (λ2, α2),...,(λ i , α i ),...,(λ k , α k )};

[0194] For each sample X i , project the original features onto the selected eigenvectors to obtain the k-dimensional new features after dimensionality reduction The calculation method is as follows:

[0195]

[0196] Finally, obtain the k-dimensional new features Y of M samples, which is expressed as follows:

[0197]

[0198] Matrix Y is the result of fusing time-frequency domain features and sample entropy features.

[0199] Step 4: Construct an RF-based evaluation model for children's object control ability to achieve the classification of children's object control ability.

[0200] Step 4 is specifically implemented according to the following steps:

[0201] Evaluate children's object control ability based on random forest. The random forest integrates multiple trees, and the basic unit of the random forest is the decision tree. The construction process of the decision tree is as follows:

[0202] Step 4.1: Use the output result Y of Step 3 as the sample set, and the i-th dimension feature is denoted as y i ,

[0203] Sort the feature y i in ascending order of values, denoted as Based on the partition point t, divide Y into subsets Y t - and Y t + , where Y t - contains those samples whose values on the feature y i are not greater than t, and Y t + [[ID=?]]contains those samples whose values on the feature y i are greater than t;

[0204] For the feature y i , take the midpoint of the interval as the candidate partition point to obtain a set of m - 1 partition points:

[0205]

[0206] Step 4.2: Traverse and examine the set Select the optimal partition point to divide the sample set. The formula is as follows:

[0207]

[0208] where Ent(Y) is the information entropy calculated by the sample set based on the label value, Gain(Y, y i , t) is the information gain after the sample set is bisected based on the partition point t, and Gain(Y, y i ) is the maximum information gain after the sample set is divided based on t;

[0209] Step 4.3: Repeat Steps 4.1 - 4.2 to obtain the information gain of the k-dimensional feature. The set is represented as follows:

[0210] It should be noted that there seems to be a missing number in the original text at "are greater than t; ". It is translated as "are greater than t;" for now, but it might need to be adjusted according to the correct content.G = max{Gain(Y, y1), Gain(Y, y2),..., Gain(Y, y k )} (26)

[0211] Then the feature corresponding to G is selected as the best splitting feature and used as the root node. By repeating the above process recursively, a decision tree can be constructed;

[0212] Step 4.4: Construction of the random forest. The random forest model is as Figure 3 . The process of constructing the random forest is as follows:

[0213] To avoid confusion and for convenient expression, the sample set Y is represented by a new symbol as follows:

[0214]

[0215] The input is the training data set Y = {(X1, y1), (X2, y2),..., (X m , y m )};

[0216] The output is the final strong classifier result H(X).

[0217] Step 4.4 is specifically implemented according to the following steps:

[0218] Step 4.4.1: Given a training set Y consisting of m samples, a new training set Y' is generated by sampling with replacement;

[0219] Step 4.4.2: Randomly select n features from all features T, where n < T, to form a new training set Then select the best splitting feature from the n features as the node and use to construct a decision tree;

[0220] Step 4.4.3: Iterate steps 4.4.1 to 4.4.2 d times, and mark the training set generated each time as where i = (1, 2,..., d), and construct d decision trees;

[0221] Step 4.4.4: These d decision trees together form a random forest. The sample category is voted on by each tree. Equation 28 represents the class with the most votes, that is, the final classification:

[0222]

[0223] Among them, H(X) represents the combined classification model, h i (x) represents a single decision tree classification model, y represents the output variable, and I(x) is an indicator function;

[0224] Step 4.5, Evaluation level classification. According to the Children's Gross Motor Development Measure (TGMD-3), determine the evaluation level of object control ability as strong, average, or weak, with the corresponding score intervals being [0, 59], [60, 79], and [80, 100]. Map the probability interval to the corresponding score interval through a linear function;

[0225] The mapping function is F = kf + c, where f is the probability corresponding to the class to which the classification belongs, and F is the score. The mapping relationship is as follows:

[0226] (1) If the sample classification result is weak in object control ability, then map the probability interval [1 / 3 - 1] to the interval [0, 59]:

[0227] F = 88.5*f - 29.5 (29)

[0228] (2) If the sample classification result is average in object control ability, then map the probability interval [1 / 3 - 1] to the interval [60, 79]:

[0229] F = 28.5*f + 50.5 (30)

[0230] (3) If the sample classification result is strong in object control ability, then map the probability interval [1 / 3 - 1] to the interval [80, 100]:

[0231] F = 30*f + 70 (31).

Claims

1. A method for evaluating children's object control ability based on nonlinear dynamics, characterized in that, The implementation is specifically carried out according to the following steps: Step 1: Use the Kinect sensor to collect the spatial position information of the skeletal points of the tester's hand, elbow, and shoulder, and use the inertial sensor to collect the motion biomechanical signals of the tester's hand; Step 1 is specifically implemented according to the following steps: Use the inertial sensor to collect acceleration and angular velocity signals, and extract time-domain signal features including signal peak value, peak discrete coefficient, and signal period. At the same time, frequency-domain signal features including signal peak power, peak frequency, and peak total power can also be extracted; Use Kinect to obtain the spatial position data of the human body bones, so as to extract some spatial position features and joint angle features; Step 2: Extract the time-frequency domain features and sample entropy features of human motion from the skeletal point spatial position information and motion biomechanical signals; Step 3: Use the principal component analysis method to fuse the time-frequency domain features and sample entropy features; Step 3 is specifically implemented according to the following steps: Use the principal component analysis method to fuse multi-features and extract representative comprehensive evaluation features. The specific calculation method is as follows: Suppose there is object control ability samples , and each sample has dimensional features = , and each feature has its own feature value; Center all features and calculate the mean of the eigenvalue of the （18） Find the covariance matrix （19） （20） Among them, is the covariance of any two features, ϵ ; Covariance matrix eigenvalues and corresponding eigenvectors are as follows: （21） The corresponding eigenvalue in the formula have , recorded as , each Corresponding to a feature vector , the eigenvalue Sort by smallest to largest, select this indivual The largest front and obtain the corresponding eigenvectors, each eigenvector is recorded as ,in The value of is 1~k, thus obtaining a set of ; For each sample , project the original features onto the selected feature vectors to obtain the dimension-reduced new features , and the calculation method is as follows: （22） Finally obtain samples of Vistide , shown as follows: （23） Matrix is the result of fusing time-frequency domain features and sample entropy features; Step 4: Build an RF-based evaluation model for children's object control ability to realize the classification of children's object control ability; Step 4 is specifically implemented according to the following steps: Evaluate children's object control ability based on random forest. The random forest integrates multiple trees. The basic unit of the random forest is the decision tree. The construction process of the decision tree is as follows: Step 4.1: Take the output result of the said Step 3 as a sample set. For the nth dimensional feature of the sample set, denote it as , Sort the features in ascending order of values, denoted as . Based on the splitting point , divide Y into subsets and , where contains those samples whose values on the feature are not greater than , and contains those samples whose values on the feature are greater than . For the feature , take the midpoint of the interval as the candidate division point to obtain a set of division points: （24） Step 4.

2. Traverse the inspection set , select the optimal splitting point to split the sample set, and the formula is as follows: （25） Among them, is the information entropy calculated by the sample set based on the label value, is the information gain after the sample set is dichotomized based on the partitioning point , is the maximum information gain after the sample set is partitioned based on the partitioning; Step 4.

3. Repeat steps 4.1 - 4.2 to obtain the information gain of the dimensional features, which is represented in a set as follows: （26） Then The corresponding feature is selected as the best splitting feature for the root node, and by repeating the above process recursively, a decision tree can be constructed; The process of constructing the random forest in Step 4.4 is as follows: For the sample set It is represented by new symbols as follows: （27） The input is the training dataset = ; The output is the result of the final strong classifier ; Step 4.5: Evaluation grade division. According to the children's basic motor skills evaluation tool, determine that the evaluation grades of object control ability are strong, general, and weak. The corresponding score intervals are [0, 59], [60, 79], and [80, 100]. Map the probability interval to the corresponding score interval through a linear function; Mapping function , where is the probability corresponding to the class to which the classification belongs, F is the score, and the mapping relationship is as follows: (1) If the sample classification result is that the object control ability is weak, then map the probability interval [1 / 3 - 1] to the interval [0, 59]: （29） (2) If the sample classification result is that the object control ability is general, then map the probability interval [1 / 3 - 1] to the interval [60, 79]: （30） (3) If the sample classification result is that the object control ability is strong, then map the probability interval [1 / 3 - 1] to the interval [80, 100]: （31）。 2. The method for evaluating a child's object control ability based on nonlinear dynamics according to claim 1, wherein Step 2 is specifically implemented according to the following steps: Step 2.1: Extract the convex hull feature of the ball landing point and the elbow angle feature; Step 2.2, extract the sample entropy: The specific calculation of the sample entropy is as follows: Step 2.2.

1. Assume that the set of X-direction sequence of the sample hand trajectory is , then the sample entropy of this sequence is calculated as follows: According to the sequence order to form a group of m-dimensional vectors : （14） Reconstruction dimensional vector , Step 2.2.2, calculate and the Chebyshev distance between vectors : （15） Given a threshold r, calculate when the number of , and divide it by the total number of distances . Denote the result as . The calculation formula is as follows: （16） Among them, represents the embedding dimension; represents the similarity tolerance, , represents the standard deviation of the sequence of the hand trajectory in the X direction; Step 2.2.3, extend the dimension to dimensions, repeat steps 2.2.2 - 2.2.4, and similarly calculate Then the sample entropy of the hand trajectory in the X direction sequence is: （17）。 3. The method for evaluating a child's object control ability based on nonlinear dynamics according to claim 2, wherein Step 2.1 is specifically implemented according to the following steps: The envelope formed by the projection points of the hand bone points on the ground is approximately regarded as the envelope formed by the ball landing point. The size and dispersion degree of the envelope area reflect the level of the hand's control ability over the ball. The calculation methods of the envelope area and dispersion degree are as follows: Step 2.1.

1. Assume that a sample projection point set is , and the coordinates of the -th projection point are ([[]] ). Use the Euclidean distance formula to obtain the length of the centroid projection trajectory of the sample. The calculation is as follows: （6） Step 2.1.2: Reorder the point set according to the magnitude of the vertical coordinate z value, and relabel it. If there are points with the same minimum z coordinate, select the point with the minimum horizontal coordinate x and denote it as ; Step 2.1.3: Sort the included angle values formed by other points and the horizontal line where is located in ascending order; when the included angles are the same, sort the distances from the smallest to the largest with respect to to complete the renumbering of the remaining points ; ​ Step 2.1.4, Concavity and convexity judgment: Assume that The coordinates of three points are respectively ( ), ( ), ( ), calculate the value of the concavity and convexity discrimination variable The formula is: （7） , is a bump; , is a concave point, delete ; , neutral point, select to delete this point; Step 2.1.

5. After traversing all the points in the point set, obtain the vertices of the convex hull graph, and label them counterclockwise as . Connect the vertices to obtain the convex hull, and calculate it using the Euclidean distance formula according to the following formula: （8） （9） （10） Obtain The group distance combination is represented as follows: Among them, represents the distance between adjacent vertices, represents the distance between two points, represents the distance between two points; Step 2.1.6: Use Heron's formula and the combined triangle method to find the area of the convex hull , and the calculation is as per the formula: （11） （12） Similarly, the envelope area and the trajectory length of the projection of the center-of-gravity data on the YOX plane are respectively expressed as , and the envelope area and the trajectory length of the projection of the center of gravity on the YOZ plane are expressed as ; The elbow angle is defined as follows: Use , respectively represent the three-dimensional spatial coordinates of the hand, elbow, and shoulder. The spatial vector from the elbow to the hand is denoted as , and the spatial vector from the elbow to the shoulder is denoted as . Then the elbow angle is expressed as: （13）。 4. The method for evaluating a child's object control ability based on nonlinear dynamics according to claim 3, wherein Step 4.4 is specifically implemented according to the following steps: Step 4.4.

1. Given samples to form a training set , a new training set is generated by sampling with replacement ; Step 4.4.2, randomly select from all features and features to form a new training set < . Then select the best splitting feature as a node from features and construct a decision tree using ; ​ Step 4.4.3: Iterate Steps 4.4.1 to 4.4.2 for times, and mark the training set generated each time as , where decision trees are constructed; Step 4.4.4, these decision trees together form a random forest, and the sample category is voted on by each tree. Equation 28 represents the class with the most votes, that is, the final classification: （28） Among them, represents the combined classification model, represents a single decision tree classification model, represents the output variable, is the indicator function.

Images