Sport Skill Evaluation with Wearable Sensors and Statistical Analysis

This research discusses sport skill evaluation with wearable sensor and statistical analysis. Skateboarding is selected as the target sport to analyze because it will be an official competition in Tokyo 2020 Olympic Games. Skateboarding is one of the difficult sports because of controlling balance to move forward and to get speed on flat ground. The balance control is a basic trick named Tic-Tac, while the trick is difficult for beginners. To make data set for analyzing Tic-Tac skill, we have provided a basic lesson to research participants. After giving them enough self-training time, we put two inertial sensors on a skateboard and waist of a research participant and got total 41 running data with Tic-Tac. According to the result of statistical analysis on the data, we confirmed that swinging a skateboard left and right is the most important motion to generate moving forward driving forth. This result means that inertial sensors are one of the effective tools to evaluate sport skill for sports science and physical education.


Introduction
This paper proposes an evaluation method of sport skill with wearable inertial sensors and statistical analysis. Skateboarding is one of the attractive sports for young generation and then the competition will be official events in 2020 Tokyo summer Olympic Games. To lean action sports skill, there is few scientific training methodology, although, video sharing services, e.g. YouTube and Youku, are effective tools to share skill and knowledge visually. To tackle the lack of scientific training methodology for action sports training, we have already de-  [1].
In recent years, many researchers have employed smartphones with micro electro-mechanical systems (MEMS) to collect and analyze this type of human behavior. Due to the widespread use of smartphones in our daily lives, we can use them to record human activities. Ueda et al. [2] used smartphones to collect movement data of human bodies at desks and estimated the status of test subjects as either concentrated on a task or relaxed without any task. However, we need purpose-specific devices to collect sensor data for healthcare and sports applications. Morita et al. [3] developed a measurement device based on a three-axis accelerometer, three-axis gyroscope, and three-axis digital compass coupled with a Bluetooth modem to record body activity. Avci et al. [4] surveyed a wide range of research papers concerning inertial sensing to analyze actions concerning healthcare and sports. Concerning action sports, Harding et al. [5] used an accelerometer and gyroscope to analyze the variance in the aerial dy-

Wearable Sensor System
Due to the wide range of movements in action sports, it is difficult to capture complete actions with a motion capture system, which deploys several high-speed cameras at fixed places. On the other hand, due to the small footprint and light weight of MEMS, it is easy to put inertial sensors on parts of a human body to record all the motion data of action sports in terms of acceleration and angular velocity. A smartphone is an available device that contains various types of sensors. However, in order to get synchronized data sets from multiple devices, we needed precise time stamps on the data collected from each device. In addition, due to the aggressive motions of action sports, the wearable device had to be impact resistant. Due to these very specific needs, we developed our own recording device.

Hardware
The developed motion recording device is composed of a microcomputer (Arduino), accelerometer and gyroscope (MPU6050), digital compass (HMC5883L),

Software
The recording data is processed by pre-processing software. The software con- Hence, the manual detection and extraction could be quite a time-consuming task. We did automatic detection and extraction of all actions based on no motion durations in sensor value. Each action data set was written to a separate file.

Experiment, Sensor Data, and Statistics
We have recorded motions of research participants' upper and lower bodies on Tic-Tac runs. To gather inertial sensor data for analyzing Tic-Tac skill, we have provided a basic lesson to research participants and gave enough self-training time. We put two inertial sensors on a skateboard and waist of a research partic-

Basic Motion to Move Forward
A skateboard does not have any direct mechanism to move forward. Hence a skateboarder has to do Tic-Tac, which is a basic trick to move forward on a skateboard. A skateboarder has to select his or her stance to do Tic-Tac. Before trying to do Tic-Tac, a skateboarder has to learn clockwise and counter clockwise turns, respectively. These turns lead to yaw rotations of skateboard ( Figure   1). When the skateboarder turns his or her skateboard, he or she has to stand on two back wheels of skateboard with his or her back foot to float two front wheels. The floating motion leads to pitch rotation of skateboard.
Tic-Tac motion is composed of alternate rotations of clockwise and counter clockwise ( Figure 2). Because of difficulty of Tic-Tac motion, some of the beginners tend to swing their bodies on X axis forward and backward to make driving force of skateboard. Unfortunately, they are going to stay same location, because the swing motion cannot generate sufficient force. Tic-Tac or alternate rotations of skateboard is right way to go ahead with a skateboard. Figure 3 shows examples of accelerometers and gyroscopes on skateboard. We can observe alternate rotations on the sensors data. To identify effective motions to generate driving force, we will discuss relation between running time and each sensor data in 3.2.

Average and Standard Deviation of Time Series Data
To apply statistical method to analyze time series data of skateboarding, averages and standard deviations are derived from time series data with an inertial sensor.
The sensor is composed of three axis accelerometer and three axis gyroscope. Effect of each independent variable on the running speed is also analyzed with simple linear regression analysis. Table 1

Correlation Analysis
To understand key factors of tic-tac motion, three types of correlations are employed, i.e. correlation between each pair of sensor's axes on skateboard, auto-correlations of body and skateboard, and cross-correlations between body and skateboard in terms of accelerometer and gyroscope.

Correlation between Each Pair of Sensor's Axes on Skateboard
Correlation between sensor axis pair on skateboard are calculated to analyze skateboard motion in detail. For example, the correlation between ax2 and ay2 is labeled as ax2_ay2. There are 15 pairs of axis of sensor. The multiple regression ax2_gz2**, ay2_az2**, ay2_gz2****, gx2_gy2**, and gy2_gz2*. The correlation between the actual and predicted running time is 0.932. To estimate effect of lower body or skateboard motion on running time, the multiple regression function is estimated with the 12 independent variables derived from sensor on skateboard in terms of average and standard deviation and 15 independent variables in terms of correlation, and then the adjusted R 2 is 0.921. Nine variables have high statistical significance, i.e. s_ax2**, s_ay2*, a_az2*, a_gy2*, s_gz2*, ax2_gz2*, ay2_az2**, ay2_gx2**, and gx2_gy2*. The correlation between the actual and predicted running time is 0.987.
To identify important factors to get fast running speed, 41 running data is divided into two category, i.e. fast and slow with a threshold 20 seconds running time. The fast and slow cases have 21 and 20 runs, respectively. The linear discriminant model is estimated with the 12 and 15 independent variables. The model can properly separate 39 cases in 41, which means that successful rate is 0.951. Evaluation index for all independent variables is defined by each parameter multiplied by each average of independent variable. The difference between fast and slow cases in terms of the index is defined to find important independent variables to achieve fast running speed. The top six variables (and index) are ay2_gz2 (2.497), s_gz2 (1.694), s_ay2 (1.184), gx2_gz2 (0.365), a_ay2 (0.2733), and s_az2 (0.270). Significance of swing motion for getting fast running speed is supported by ay2_gz2, s_gz2, and s_ay2. Tic-tac also needs leaning skateboard (gx2) to turn for swing motion (gz2), which is explained by gx2_gz2.

Auto-Correlation of Body and Skateboard
To identify relation between running speed and periodicity of motion concerning upper and lower body, 2 average auto-correlations of all six sensor axes are calculated, i.e. acor1 and acor2, respectively. The auto-correlations are converted to base-10 logarithm to identify relation with base 10 logarithmic running time. The multiple regression function is estimated by the 2 variables on upper and lower body, and then the adjusted R 2 is 0.965. The variable on skateboard has high statistical significance, i.e. acor2 ****. The other variable on body (acor1) is not statistically significant. The correlation between the actual and predicted running time is 0.983.
To separate effects of upper and lower body on running speed, acor1, which is the auto-correlation of upper body is employed in a regression formula, and then the adjusted R-squared of the model is 0.594. The correlation between the actual and predicted running time is 0.777. In the other hand, acor2, which is the auto-correlation of lower body is employed in a regression formula, and then the adjusted R 2 is 0.964. The correlation between the actual and predicted running time is 0.982 (Table 2).

Cross-Correlation between Body and Skateboard
To understand how to generate swing motion of skateboard, two cross-correlations between upper and lower body are calculated in terms of accelerometer and gy- The maximum and minimum cross-correlations tend to be derived around zero of time lag. Table 3 summarizes basic statistics of them with time unit of 10 micro second. The outlier means the rate of runs far from the average. The table tells us that all lags are small, which is under 1.5 second and the percentage of outlier is also small, which is under 14%.
To sum up, the regression and discriminant analysis indicates the minimum cross-correlation on accelerometer (CaMin) is the most significant for fast running speed. CaMin is negative value, i.e. −0.698, which describes that counter motion between upper and lower body. The high significance of the variable seems to be meant that counter motion between upper and lower body is effective to generate fast running speed (Table 4).

Discussion
According

Conclusion
This research discussed sport skill with statistical analysis on wearable sensor data. We selected skateboarding as the target sport and Tic-Tac as a basic skill to move forward with a skateboard. According to the result of statistical analysis on sensor data, we identified three important factors to get fast speed are identified, i.e. swing motion of skateboard, periodicity of swing motion, and counter motion between upper and lower body. This result means that wearable sensors and statistical analysis are effective tools to evaluate sport skill and to find important factors for sports science and physical education.

Conflicts of Interest
The author declares no conflicts of interest regarding the publication of this paper.