A study of student movement characteristics in college soccer teaching based on motion capture technology

Motion analysis solutions based on inertial motion capture technology is a popular research direction recently, among which the more representative brands of inertial sensors are Xsens MVN motion capture system and Noitom’s Perception Neuron motion capture system. Inertial motion capture technology has the following advantages: simple installation and operation, good portability of the equipment; it can collect attitude data with high sampling rate and capture richer detail information; and with the continuous development of microelectromechanical technology, the cost of the capture equipment is gradually reduced. In the experiments of this paper, we used the Perception Neuron inertial motion capture system developed by Beijing Noiton Technology Company as the motion capture sensor device in the experiments, which is able to acquire accurate and effective human motion and posture data, and provide a reliable data base for the subsequent bone modeling and motion analysis.

Perception Neuron is a motion capture system based on inertial sensors, which consists of sensors, signal collection and transmission devices, and a signal processing system. Perception Neuron is a motion capture system based on inertial sensors, which consists of sensors, signal collection and transmission devices, and a signal processing system, which uses the Perception Neuron software on the PC to process the signals from the sensors and calculate the relative offset positions of the human body’s joints, thus obtaining and reconstructing three-dimensional motion data. Neuron system combines navigation and orientation system, which has the advantages of simple operation and high accuracy of data acquisition.

The Perception Neuron motion capture system utilizes wearable IMU sensors, including accelerometers, gyroscopes, and magnetometers, to measure data such as body acceleration, angular velocity, and magnetic field strength. The system can collect data from multiple sensors simultaneously to obtain more accurate data on the body’s movements and posture. During data acquisition, the sensors are placed at key areas of the body, such as the head, arms, waist, and feet. As the human body moves, the sensors measure a number of movement-related information, such as acceleration, angular velocity, and the strength of the geomagnetic field, which is then transmitted wirelessly to a computer for processing and storage.The Perception Neuron system uses high-precision data-processing algorithms to filter out noise and calibrate the data, ensuring that it is accurate. The Perception Neuron system uses highly accurate data processing algorithms to filter out noise and calibrate the collected data to ensure the accuracy and reliability of the collected data. The data that is obtained can be saved in either raw data format or standard motion capture format for later signal processing and 3D reconstruction applications.

Data acquisition based on inertial motion capture devices is a fully automated process that automatically senses the rotation angle and offset information of human motion through motion capture sensors and transmits this information to a computer for processing. When measuring different types of motion postures of the experimenter, the Perception Neuron device is equipped with lightweight IMU sensors that sense and capture the motion data of the experimenter. The data is transferred to the PC via wireless transmission technology or USB interface, and then relevant processing is carried out on the PC. Finally, it is displayed in real-time.

The operation process of data acquisition is as follows: 1)

Installation of hardware equipment. Wear a full set of IMU sensors of Perception Neuron device on the subject, tightly fix each IMU sensor in each specific part of the subject according to the standard position, and turn on the power of each sensor one by one.

The BVH (Bounding Volume Hierarchy) data structure, a generalized human feature animation file format, uses a tree structure for representing the hierarchical structure and relative positional data of the human skeleton, including the rotation data of the experimenter’s skeleton and limb joints. The motion of each joint node is represented by the joint nodes, which combine offset and rotation information.

The BVH file contains two parts: the definition of the human joint tree and the human motion data. In the definition of the joint tree, it includes the name of each joint point, the number of channels, the relative position between the connected joints, and the length of each part, and this part defines the skeletal relationship connectivity between each part of the human body. And the human motion data section records the duration of the human motion data in that experiment, i.e., the number of frames and the time interval between each frame. According to the joint order in the joint tree definition, the rotation data and translation data of each joint node in each frame are recorded, which can be used to create and animate a 3D human model or as input data for human movement recognition and analysis tasks.

The data in BVH format allows for reproducing the movements performed by the subject during motion capture. For this purpose, it is necessary to model the corresponding skeletal nodes and read the BVH spatial data so that the 3D human model can render the movements as close as possible to those of the actual subject.

The 3D human pose estimation problem can be formulated using probabilistic reasoning. First, we denote the 3D pose, the 2D pose, and the corresponding image as X ∈ R^{3 × J}, y ∈ R^{2 × J}, and I, respectively.The 3D pose estimation problem can be formulated in the following joint probabilistic form: (1)P(X,Y,I)=P(X|Y,I)P(Y|I)P(I)

In other words, to estimate the 3D pose from a picture, the 2D pose is first estimated from the picture to obtain P(Y|I). P(Y|I) denotes the likelihood function of the observed image and the 2D pose. Many 2D pose estimation methods can be used to solve this problem. However, the P(X|Y, I) term in Eq. (1) is difficult to obtain. The reason is that reconstruction dichotomy leads to the relationship that 2D pose and 3D pose are not one-to-one correspondence, i.e., the 3D pose is not independent of the image condition under the given 2D pose condition, i.e., P(X|Y, I) ≠ P(X|Y). Although the human brain can solve the dichotomy problem by the local visual cues provided by the image and combined with the visual experience, it is difficult for computers to extract and interpret the local visual cues of the image. Therefore, in this paper, we adopt another strategy, i.e., using temporal cues to solve the sequential pose estimation problem: (2)P(Xt,Yt,Xt−1)=P(Yt|Xt,Xt−1)P(Xt|Xt−1)P(Xt−1)

Equation (2) illustrates that the attitude X_t at the t moment can be estimated from the attitude X_t−1 at the t − 1 moment and the observation Y_t at the t moment. In contrast to P(X|Y, I) indivisibility, in Eq. (2), P(Yt|Xt,Xt−1) can be simplified to P(Yt|Xt) . Thus, 3D attitude estimation is transformed into solving the attitude transfer problem P(Xt|Xt−1) . i.e., if it can be accurately modeled P(Xl|Xt−1) , it is possible to infer the 3D attitude at any moment given the initial conditions based on the chain rule.

The Riemannian Popper human skeletal pose model is shown in Figure 1.

Figure 1.

Riemann Pop represents human posture

As shown in Fig. 1(a), the rigid skeletal parts in the human skeleton model are connected over the joints, and the kinematic equations of the human kinematic chain can be used to represent the relative motion between two joints. In this paper, it is assumed that the upper limb trunk is divided into rigid bodies, so the root node in each frame can be accurately estimated by the method of Rigid Structure from Motion (Rigid Structure from Motion). As shown in Fig. 1(b), the relative motions of adjacent joints in the human skeleton can be expressed as a series of vectors normalized to the unit sphere xtij=x1i−x1j‖x1x−x12‖2∈S2 . Similarly, the relative motions of 2D joints can be expressed as vectors on the unit circumference ytij=y1i−y1j‖yt′−y2l‖2∈S1 , as shown in Fig. 1(c) and Fig. 1(d).

The next state xtij of the manifold can be obtained by an exponential mapping if the initial value of the tangent plexus (xt−1ij,vt−1ij)∈TS2 is given: (3)xtij=expxt−1″(vt−1ij)

However, during human motion, there are difficult to estimate time-varying accelerations between neighboring frames, resulting in an inability to accurately estimate the current joint motion velocity. Simply moving the tangent vector v_t−1 of the previous moment in parallel to the tangent space of the next state x_t will lead to deformation of the estimation result. The human body is affected by external forces and other physical and other factors (e.g., weight, gravity, contact with the ground, friction, etc.) during motion. Therefore, a single kinematic model is not able to simulate complex human motion processes. However, for short-period motion with high frame rate, it can be simulated by a second-order stochastic dynamic model. Based on this, this paper proposes to use the following nonlinear stochastic differential equation to estimate the joint motion state: (4)(xt+1ij,vt+1ij)=f(xtij,vtij)+qti

where qti~N(0,Qti) is additive process noise with zero mean. The nonlinear function f in Eq. (4) can be expressed in the following form: (5){ xt+1ij = expxiij(xtij) vt+1ij = Pxiijt+t+t(vtij)

Based on the skeletal constants and the following measurement model, we can establish a nonlinear relationship between the 2D observation and the 3D hidden state: (6)yt+1i=H(xt+1i,vt+1i)+rt+1i

where the standard perspective function H is the observation matrix in the measurement model and rt+1i~N(0,Rt+1i) is the measurement noise.

As shown in Fig. 2, the computing process of the model in the soccer action recognition and evaluation system can be divided into five steps: data preprocessing, attitude angle modeling, feature extraction and selection, downscaling and classification.

Figure 2.

Flow of football action recognition and evaluation system

Soccer is more complex than other sports (e.g., badminton, tennis, and volleyball): the completion of the action mainly relies on the lower limbs and the pattern of the action is variable, and there are more invalid actions in the lower limbs during the process of the movement, and the similarity between the actions is higher. Therefore, the posture angle feature based on the ankle area plays a key role in improving the accuracy of action classification, and more implicit action feature information can be obtained through the posture angle model solving to reflect the differences between the actions, which helps the model learning. The model uses PCA to balance complexity and accuracy, followed by an SVM-based classification algorithm to recognize soccer movements. In data preprocessing, a 3-point moving average filter is used to reduce the noise interference of the raw data, and then the motion signal is automatically segmented by locating the peaks of the signal.

SVM algorithm is used to recognize soccer movements and improve the accuracy of sports data analysis after differential movements are obtained through the pose angle model.

SVM algorithm is widely used in the field of machine learning and pattern recognition.The basic model of SVM algorithm aims to maximize the distance between the samples called support vectors and the hyperplane on the feature space. The generalization ability of the learning machine can be improved by the SVM algorithm, which reduces the empirical risk and confidence interval with small samples. The training points are mapped to points in the space and the width of the gap between two or more classes is maximized by the SVM algorithm. Compared with other supervised algorithms such as LR algorithm or Naive Bayes algorithm, SVM algorithm is more suitable to deal with high-dimensional and linearly indivisible problems, it is free to choose the parameter model and use the support vectors as the basis of classification in the hyperplane, the problem in this study belongs to a small-sample linearly indivisible problem, so the SVM algorithm model is chosen for solving.

Let the training dataset be T={(Xk,yk)|X∈Rm,yk∈{0,1}nk−1} , which contains n samples, X represents a m-dimensional matrix, the classification label y_k has a value of 0 or 1, and the current sample is X_k. Therefore, the optimization objective function is represented as Equation (7): (7)minω,b12‖ω‖2 s.t.yk(ω⋅xk+b)−1≥0,k=1,2,…,N

Here, ω is a vector on the hyperplane and the offset b of the hyperplane is along ω from the origin. Equation (7) is a convex quadratic programming problem. According to the theory of convex optimization, Eq. (7) is transformed into an unconstrained problem. The optimization function can be expressed as Eq. (8): (8)L(ω,b,α)=12‖ω‖2−∑k=1Nαkyk(ω⋅xk+b)+∑k=1Nαk

where α_k is the Lagrange multiplier and αk≥0(k=1,2,3,…,n) . The original problem can be expressed as equation (9): (9)maxαminωbL(ω,b,α)

Solving ω and b as a minimal problem gives the values of ω and b, i.e., there is equation (10): (10){ ω=∑k=1Nαkykxk ∑k=1Nαkyk=0

The obtained solution is substituted into the Lagrangian function of the minimization problem and the substitution leads to the optimization function (11): (11)minα12∑k=1N∑j=1Nαiαjykyj(xk⋅xj)−∑k=1Nαk s.t.∑k=1Nαkyk=0 αk≥0,k=1,2,…,N

Then some anomalous sample points with a slack variable ζ_k of (xk,yk) are introduced to make the training set linearly inseparable, with a penalty parameter C ≥ 0. The original problem is described as Equation (12): (12)minω,b,ζ12‖ω‖2+C∑k=1Nζk s.t.yk(ω⋅xk+b)≥1−ζk,k=1,2,…,N ζk≥0,k=1,2,…,N

The final function is Eq. (13) obtained after the transformation: (13)minα12∑k=1N∑j=1Nαkαjykyj(xk⋅xj)−∑k=1Nαk s.t.∑k=1Nαkyk=0 0≤αk≤C,k=1,2,…,N

As a result, the relevant motion recognition data can be processed using Eq. (13) obtained after the conversion to reduce the interference of abnormal data on the conclusion and improve the accuracy of motion recognition.

In order to make the collected data of soccer movements more accurate and to ensure a reliable and effective data base for subsequent movement recognition, this paper designs a scheme for soccer experiments. Firstly, the inertial sensor was worn on the ankle of the commonly used kicking side, and the action information of the foot during the movement was collected. In the process of experimental data collection, the same person in a fixed 7 meters away from the use of the same force to the athlete pass the ball, the athlete is passed to the ball stopping action, the whole process of the motion camera is used to record the tester’s movements, as the basis of the later data analysis.

A male tester (height 175cm, weight 71Kg, non-physical education major) was invited for data collection in this experiment, and the sensor was worn on the athlete’s right leg at the ankle because the invited tester was accustomed to using his right foot to perform stopping maneuvers. The triaxial acceleration and triaxial angular velocity data of the athlete’s leg were collected throughout the experiment. Meanwhile, in order to ensure the quality of each stopping motion, the athletes were required to warm up for 10 minutes before the experiment, and take a 5-minute rest after every 10 stopping motions during the experiment in order to maintain sufficient physical strength to ensure the quality of the experiment.

After the collection of raw data, in order to facilitate the subsequent data processing, this paper manually split the continuous data of multiple actions into single stopping actions. In the experimental process, the athlete consciously took a short rest after each stopping action, and this paper relied on this basis to segment the data. Through the analysis of the three-axis acceleration and three-axis angular velocity data, it was found that the difference between the three-axis acceleration data and the angular acceleration data of the X-axis and the Y-axis was relatively small, so in order to ensure the accuracy of the recognition of the differences in the later movements, the Z-axis angular velocity data was chosen as the basis for segmentation. The segmented Z-axis angular velocity data is shown in Figure 5. According to Figure 5, it can be seen that the stopped ball and the unstopped ball have more significant data feature differences.

Figure 5.

Z-axis angular velocity data of football stopping action

Five traditional machine models were used for recognition in soccer stop motion recognition, namely SVM, NBC, DA, K-NN and CART.In the soccer stop motion experiment, a total of 60 sets of data were selected for analysis, including 30 sets of stopped data and 30 sets of unstopped data, so 50 sets of data were randomly selected for training on the model, and the remaining 10 sets of data were used for testing. Because of the relatively small amount of data, random selection was used for classification model training, and a total of five sets of experiments were conducted, and the average of the results of the five sets of experiments was taken as the final results. Table 2 shows the results of the soccer stoppage experiments.

It can be seen that the classifier SVM performs best in recognizing stops and non-stops with 100% accuracy. This shows that we can accurately recognize differences in movements not only at the wrist but also at the legs.