Keywords

1 Introduction

People can fully enjoy golf at any age or skill level, for its widespread and increasing popularity. In order to increase Golf-swing playability and make it more suitable for an amateur to practice, the analytical of the golf swing is imperative. Golf professionals aim to educate a golfer on the best approach to utilize the body and club during the swing which will transfer the most amount of energy into the ball and maximize the driving distance [1]. In this paper, Kinect was used to assess the joints’ visibility (skeletal joints tracking status information) and extracted joints’ coordinate 3D-dimension data.

The capture and analysis of human behaviors such as jumping, running, swing, are common in some domains, including sports science, musculoskeletal injury management and the human-computer interaction [2, 3]. It requires highly accurate motion capture in the analysis of joint angle, position, and angular velocities. It is outside the reach of most users of the highly accurate motion capture systems, whether based on camera or inertia sensor. To popularize motion capture technology, Microsoft cooperation has rolled out the Kinect which bases on depth camera, as low-cost alternatives. The Kinect uses an infrared based active stereo vision system to get a depth map of the observed scene, and it was designed to recognize human gestures and skeleton joints. There are still challenging problem due to strong noise in-depth data and self-occlusion when using Kinect to analyze the golf swing motion. Yeung et al. [4] evaluated Kinect as a clinical assessment tool for Total body center of mass sway measurement, and their results revealed the Kinect system produced a highly correlated measurement of Total body center of mass and comparable intra-session reliability to Vicon (Motion Capture Systems). Previous studies have indicated a positive relationship between Microsoft Kinect and OptiTrack motion tracking system. Høilund et al. [5] investigated the precision between the Microsoft Kinect and OptiTrack and found that the OptiTrack and the Kinect has been shown to deliver approximately the same results, with some restrictions. Wang et al. [6] examined the first generation Kinect and the second generation Microsoft Kinect and found that the second generation Kinect provides better accuracy in joint estimation while providing skeletal tracking that is more robust to occlusion and body rotation than the first generation Kinect released at 2010. Kinect has limited visible area and could not track the human limbs which behind others limbs. Thus, just one Kinect is not enough for tracking the golf swing motion. However, multiple Kinect sensors have broader visible area even 360 degrees, and users do not need facing the Kinect directly. Williamson et al. [7] using multi-Kinect tracking for dismounted soldier training, and within the Microsoft Kinect Software Development Kit that can be merged using commercially available tools and advanced fusion algorithms to produce better quality representations of users in the real world within a virtual environment. Furthermore, several Kinect coupled with inference algorithms can produce a much better tracked representation as users move around.

In order to figure out accurately joint 3D-dimensional data, Kinect was using to track the joint and figure out each joint visibility ratio during one full golf swing. Tracking a golf swing with only one Kinect sensor could cause tracking all the swing motion to be difficult since swing speed is typically fast, both arms overlap each other, and the self-occlusion in the motions being captured. For obtaining larger visible area, we prepare to use Multiple-Kinect to capture the golf swing.

2 Materials and Methods

2.1 Experimental Equipment

Kinect for Windows 2.0 (Xbox one), Dell M8600 mobile workstation, tripods, iphone7 take photos, Cougar as the hardware. We used the Kinect SDK-v2.0_1409-Setup.exe, skeletal Joint analysis program as the software.

2.2 Experimental Scheme

In this experiment, the subjects are the sophomore professional golf students. The Microsoft Kinect v2 was used to collect four feature golf swing motions which including golf swing setup posture, the top of the swing, impact time and finish swing motions and captured ten students including eight golf swing directions. Golf swing directions take turns every 45-degree from facing directly to the Kinect to −45-degree angle to Kinect coordinate Y-axis, as showed in Fig. 1.

Fig. 1.
figure 1

(a) Human skeletal joints by Kinect. (b) Eight swing directions

Full size image

2.3 Experimental Procedures

Before the experiment, the eight swing directions position was marked on the ground. Every participant posed the four feature golf swing motion and did the next swing direction. All participants took turns to pose the eight swing directions exactly. Then we collected the skeletal data when giving a sign (Fig. 2).

Fig. 2.
figure 2

Four feature swing postures. (a) Swing setup, prepare to upon the club. (b) Top of swing position. (c) Time of shots the ball. (d) End of golf swing motion

Full size image

3 Data Analyses

Kinect may perform differently on the same swing posture. In swing directions, the participants were asked to hold on one swing motion posture until we collected 30 frames data. To improve the data reliability, the least joint amount of frames will be extracted as the current swing posture data.

3.1 Joints Moving Distance Analysis of One Full Golf Swing Motion

Each joint has different moving distance on full golf swing. Thus, some joints will be weighted up, and some joints will be weighted down. Figure 3 shows which joint should be weighted up or weighted down using the joints moving distance.

Fig. 3.
figure 3

Joint moving distance on full golf swing motion

Full size image

Figure 3 describes that joints of SpineMid, Neck, Head, ShoulderLeft, ShoulderRight, HipLeft, KneeLeft, FootLeft, AnkleLeft, HipRight, KneeRight, AnkleRight, FootRight and SpineShoulder’s performance are relatively stable. SpineBase, ElbowLeft, WristLeft, HandLeft, ElbowRight, WristRight, HandRight, HandLeft, ThumbLeft, HandRight and ThumbRight joints’ performance are unstable. Joints’moving distance averages show how long they go through during full swing motion. The longer distance averages, the more vitality of joint. Combine with joints tracking status of Table 3, some joints will be weighted up or weighted down among these joints.

3.2 Joint Tracking Status Data on Four Swing Postures of Eight Swing Directions

The two following figures demonstrate joint tracking status data which was collected from 10 participants. And the X-axle value means eight swing directions; the Y-axle value means joint accurate tracking status amount. The joints visibility status was presented graphically, as shown in Fig. 4. Total information was shown in Tables 4 and 5. For example, visibility of eight swing directions of ElbowLeft swing_setup posture was presented on scatter diagram. The scatter diagram illustrates that on the swing directions of 3 and 4, the joint tracking status amount of ElbowLeft achieve nine while others are below seven. This statistics data of swing directions 3 and 4 shows high tracking confidence level, while averages the joints data possess highly accurate.

Fig. 4.
figure 4

ElbowLeft Swing_Setup tracking status amount

Full size image

According to the Fig. 4 analytical process, the joints tracking status data was categorized into high tracking confident level and low tracking confident level.

3.3 Joint Average and Variance

Statistic method of the average value and variance value is used to assess the joints tracking data performance. Joint tracking status data on four swing postures of eight swing directions of ten participants are presented in the two following table.

To assess each joint tracking amount on four swing postures of eight swing directions, the average is used for evaluating that. The average for each joint tracking amount is computed as

$$ \bar{x} = \frac{1}{8}\mathop \sum \limits_{i = 1}^{8} xi $$
(1)

\( xi \) value means each joint tracking status amount of eight swing directions, \( \bar{x} \) value means the joint tracking status amount average value.

The variance for each joint tracking amount is computed as

$$ \sigma^{2} = \frac{1}{8}\mathop \sum \limits_{i = 1}^{8} xi - \bar{x} $$
(2)

\( \sigma^{2} \) value means the variance of each joint tracking status amount of eight swing directions.

Joint tracking status data of four swing postures of eight swing directions’ average and variance is presented in Table 3. Table 3 illustrates that joints of SpineBase, SpineMid, Neck, Head, HipLeft, KneeLeft, HipRight, KneeRight, and SpineShoulder maximum average and minimize variance receive high average value and variance. The average of those joints’ average is between 8.875 and 10, and the variance is between 0 and 1.6875. Statistic data of average value and variance value represents high tracking confident level. And joints of ShoulderLeft, ElbowLeft, WristLeft, HandLeft, ShoulderRight, ElbowRight, WristRight, HandRight, AnkleLeft, FootLeft, AnkleRight, FootRight, HandTipRight, ThumbLeft, HandTipRight, and ThumbRight are categorized into low tracking confident level since receiving low average value and high variance value.

Combine with the above three tables, joints of SpineBase, SpineMid, Neck, Head, HipLeft, KneeLeft, HipRight, KneeRight and SpineShoulder perform high tracking confident level, while could apply these joints 3D-dimension data to golf swing analytical system. Take ElbowLeft for example. The statistical data shows that the swing posture of swing_setup, ElbowLeft joint obtains nine out of 10 times accurate tracking data on swing directions of 3 and 4. On the swing posture of swing_Top, ElbowLeft joint obtains 10 times accurate tracking data on the swing directions of 1 and 9 out of 10 times on the swing directions of 4. On the swing posture of swing_impact, ElbowLeft joint obtains 10 times accurate tracking data on the swing directions of 1, 2 and 8, 9 out of 10 times on the swing directions of 3 and 4. On the swing posture of swing_Finish, ElbowLeft obtains 10 times accurate tracking data on the swing directions of 4, 5, and 6, and 9 out of 10 times on the swing directions of 3.

Without description all of the joints tracking status information, the two following tables show how to extract to value joint information. These two tables distinguish the joint tracking status data confident level from the different geometric figure. The circle ( ○) means joint tracking data receive high tracking confident level, the square (□) means joint tracking data receive moderate tracking confident level, triangle ( ▵) means joint tracking data receive low tracking confident level and the null cells means the data of the joint can’t be applied to the application.

3.4 Joints Weight Analysis

From the above discussion results (Fig. 5 and Tables 1, 2, and 3), it shows that joints of WristLeft, WristRight, HandLeft and HandRight have a big contribution to the full swing motion, but have poor performance for the small average value. Joints of WristLeft, WristRight, HandLeft, and HandRight should be weighted down when applying to the golf swing motion analysis system. Joints of SpineBase, SpineMid, Neck, Head, HipLeft, HipRight, SpineShoulder have highly visible and stable performance on the full swing motion. On the basis of golf characteristic, pelvis has important significance on a swing [8]. Thus, HipLeft and HipRight should be weighted up. Joints of ElbowLeft, ElbowRight, KneeLeft, KneeRight, AnkleLeft, AnkleRight, FootLeft, and FootRight receive high average value, also make an important contribution to golf swing should be weighted down. While HandTipLeft, HandTipRight, ThumbLeft, ThumbRight make little contribution to golf swing according to sports biomechanical principles [9].

Fig. 5.
figure 5

Total joint tracking data of eight swing directions

Full size image
Table 1. 10 participants joint tracking status data on Swing_Setup and Swing_Top postures of eight swing directions
Full size table
Table 2. 10 participants’ joint visibility status data on Swing_Impact and Swing_Finish postures of eight swing directions
Full size table
Table 3. Joints tracking status average and variance value
Full size table
Table 4. Value of joint swing posture on setup and top swing postures
Full size table
Table 5. Value of joint swing posture on impact and finish swing postures
Full size table

3.5 Total Joint Tracking Data of Eight Swing Directions

From what has been discussed above, the invalid joint of HandLeft and HandRight are wiped out of the statistical table. And the value joint tracking data is presented in Fig. 5. Figure 5 shows that swing directions 8 receive the highest joint tracking rate of 13.51%, while swing directions 5 receives the lowest joint tracking rate of 11.92%. This result can be applied to Multiple Kinect golf swing analytical system. When using two Kinects, we can locate the Kinect to swing directions 8 and 4, while Kinect should face directly to the user. While using three Kinect, we can locate the Kinect to swing directions 8, 4 and 1, while Kinect should face directly to the user. The more Kinect used, we can locate the Kinect following the law which shows in Fig. 5.

4 Discussion

This study evaluated the human skeletal joint tracking status information with Microsoft Kinect sensor. Low cost, makerless, high tracking accurately makes Kinect be most popular motion captures sensor. Though Kinect does the good performance in the game field, still contains the disadvantage, like low frames, self-occlusion, data jitter.

Destelle et al. [10] proposed to fuse the joint position information obtained from the popular Kinect sensor with more precise estimation of body segment orientations provided by a small number of wearable inertial sensors. The use of inertial sensors can help to address many of the well-known limitations of the Kinect sensor and enhance the joint angle measurement accuracy. We assessed the skeletal joint tracking amount of different swing directions and extract the valuable joint data to form an integral skeleton. To ascertain the joint accurate position is common of two studies. Zhang et al. [11] displayed a method of scoring time-sequential postures of the golf swing. Unlike their study, they extracted the time-sequential posture of golf swing features when swing was performed and used HMM-NF models for scoring. They proposed methods can be implemented to identify and score the golf swing effectively with up to 80% accuracy rate.

The results presented here demonstrate the feasibility of using Kinect to analyze golf swing motion. Above analysis and conclusion indicate that human body skeletal joints can be capture most of the joints which extract from Kinect sensor. Tables 1 and 2 show that, different swing directions due to different joint tracking amount. Statistical analyses show Kinect can be placed an ideal location for obtaining high joint tracking confident, while this conclusion can be applied to the golf swing.

5 Conclusions

The overall goal of this study was to investigate visibility of each joint obtained by Kinect on four swing postures of eight swing directions. The results indicate that different swing directions due to different joint tracking amount of swing posture. Our study found that valuable joints can be extracted from various swing directions posture to fill a whole human skeleton. We propose a method of complementary joints which can enhance the valuable joint amount. This research can be applied to Multiple-Kinect golf swing analytical system to enhance the system reliability and robustness.