A Quantitative Model To Evaluate Wrist-Rotation In Golf (P6)
In this section, we describe our methods of data collection, model generation and validation to provide feedback on the quality of movements with respect to wrist rotation.
We conducted our experiments to express the quality of the golf swing with respect to the wrist rotation. The experiments were conducted on three male subjects and one female subject all aged between 20 and 35. Each of the subjects wore three on-body sensors. In addition, two sensor nodes were placed on the golf club: one on the club head and one on the grip as shown in Fig. 4. The subjects were asked to perform the golf swing ten times for each of the variations listed in Table 1.
Our subjects performed swings after ﬁrst addressing the ball with 20◦, 40◦, 60◦ and 80◦ clockwise and counter-clockwise rotation of the wrists. Each subject also performed a perfect golf swing that has no wrist rotation or out-of-plane movements. For each movement, the amount of wrist rotation was controlled by ﬁxing the location of the nodes placed on the golf club. The subjects must grip the club aligned with the nodes on the club. They were asked to keep their wrist ﬁxed throughout the movements. This allows the system to control the swing plane while achieving consistent angles in different segments of the swing. All of the swings were performed in the absence of a golf ball. The subjects were also asked to perform the swings at a speciﬁed speed for experimental consistency.
An extra mote was connected to a laptop via USB port to collect data from all sensor nodes.The data was collected using our tool developed in MATLAB. We followed the procedure for data collection, preprocessing, feature extraction, model generation, and validation as described previously. We processed collected data ofﬂine using our tools developed in MATLAB.
For each trial, the data collected from four subjects was ﬁrst preprocessed using a ﬁve-point moving average ﬁlter to remove the effect of noise. Each trial was divided into four major segments consisting of takeaway, backswing, downswing and follow-through. The manual segmentation was performed with the help of the video recorded during data collection. An exhaustive set of features was extracted from each segment. The features include statistical and morphological features as shown in Table 2 in which the ﬁrst eleven features represent statistical features obtained from each signal segment, and the next four features are morphological features extracted from ten evenly distributed samples over each segment. We used 50% of the trials for the training to build our quantitative model,andthe rest to evaluate performance of the model.
For each of the major segments, a separate quantitative model was built. The features extracted from ﬁve sensors (x,y,z accelerometer, and x,y gyroscope) formed a 215-dimensional feature space for each sensor node. Data fusion was used to combine features from all sensor nodes to form a 1075 dimensional feature space which was used for subsequent processing. The features were fed to the PCA block for dimension reduction. Only a small number of principal components obtained from PCA were used to ﬁnd LDA projections. The number of principal components was set to the rank of the within-class scatter matrix.
Given nine different groups of wrist rotation, LDA creates eight discriminant functions in the form of linear combinations of the input. In Fig. 8 we illustrate projections of the training trials using the ﬁrst two dimensions for takeaway, backswing, downswing and follow-through. The group 1 indicated by green color corresponds to perfect swings while red represented by groups 2, 3 …5 and magenta colors annotated by 6, 7 …9 show clockwise and counter clockwise rotations respectively. These ﬁgures demonstrate the effectiveness of our technique in distinguishing different variations of the wrist rotation. Furthermore, the graphs would clearly describe the angular rotation.
The projections obtained by applying LDA were used to build a linear regression as described previously. We used the validation set to measure the degree of wrist rotation based on the model acquired. The values of error in terms of RMSE and MAE are shown in Table 3 and Table 4 respectively. In overall, the amount of root mean squared error was 15.5, 10.7, 8.9 and 9.1 for takeaway (TA), backswing (BS), downswing (DS) and follow-through (FT) respectively. The overall value of absolute mean error was reported as 9.2, 7.7, 6.6 and 6.5 degrees for TA, BS, DS and FT respectively which introduces an average error of less than 10 degrees for all segments.
Throughout our experiments, we used a sampling frequency of 50Hz which provides good resolution in capturing motions of golf swing. Reducing the sampling frequency can potentially reduce the complexity of processing. However, over-reduction may eliminate important details of the signal. In an effort to address this issue, we further adjusted our sampling rate with respect to the performance of our model. Recall the performance of our model expressed in terms of RMSE and MAE, our adjustment processtends to ﬁnd a minimum sampling frequency that maintains approximately similar performance to 50Hz.
For the purpose of frequency adjustment, we measured RMSE and MAE errors for different sampling frequencies between 5Hz and 50Hz. The results are illustrated in Fig. 9 and Fig. 10. For each segment, the error remained almost constant beyond certain frequency. This threshold varied from one segment to another.The lowest threshold was obtainedfor takeaway (10Hz) and the highest frequency belonged to downswing and follow-through (30Hz).
The difference between minimum sampling frequencies is mainly a factor of changes in speed of swing motions from one segment to another. According to the analysis performed using highs peedcine-ﬁlms of tournament professionals, the golf club can move four times faster during downswing than it usually does during takeaway and backswing. As a result, faster motions require higher sampling frequencies to ensure the collected data has acceptable resolution. Considering the worst case (i.e. frequency required for downswing and follow-through), our system allows a frequency of 30Hz while maintaining the same amount of error as reported at 50Hz.