A Quantitative Model To Evaluate Wrist-Rotation In Golf (P3)
We use a BSN consisting of several sensor units placed on the body and the golf club to capture the physical movements of the golf swing. Each sensor node, also called a mote, is equipped with a custom designed sensor board consisting of several inertial sensors as shown in Fig. 3. We use the TelosB mote which is commercially available from XBow R °. The mote has a micro controller for process ingandstorage, and a radio for communication. Embedded with our custom-designed sensor board, a tri-axial accelerometer and a bi-axial gyroscope are interfaced with the mote platform. The mote and the sensor board are powered by a Li-Ion battery integrated with each node.
Our body-worn sensor nodes are placed on the upper body and arms to capture signiﬁcant motions during the swing. The movements of the golf club are captured by the two nodes attached to the club. This conﬁguration ensures that the system captures inertial information associated with the major parts of the body involved in the golf swing.We placed two nodes on the golf club (one near the club head and another near the grip, asshown in Fig.4), one on the right wrist,one on the left arm,and one on the backat waist level. We will demonstrate effectiveness of this sensor setup through our experiments.The optimal sensor conﬁguration, including the best senor placement and smallest set of sensors required for our system, is not investigated in this paper.
The processing unit of each node samples sensor readings at 50Hz. This tentative sampling rate is experimentally chosen to provide sufﬁcientre solution while compensating for bandwidth constraints of our sensor platform. We will investigate more efﬁcient rates later in this paper. Each sensor node can perform local processing on the inertial data and transmit the result wirelessly to a base-station. The base-station can be either another mote or a PDA which collects local information from all other nodes, performs ﬁnal processing, and provides the user with a feed back on the quality of the swing.
Our model for assessing quality of golf swing saims to utilize processing capability of each sensor node and combine local information obtained from all sensor nodes to achieve a measure of quality. This process consists of several steps as illustrated in Fig.5 and explained below.
The preprocessing consists of filtering and segmentation to facilitate subsequent operations without losing relevant information. The data collected at each sensor node is locally ﬁltered. We use a ﬁve-point moving average ﬁlter to reduce the effect of noise. The number of points used to average the signal is experimentally chosen to maintain sharp step response while as mooth out put signal can be obtained. For segmentation, we determine parts of the signal that represent swing segments. That is, each signal segment corresponds to one of take away, backswing, downswing and follow-through. Currently, we perform this process manually to avoid introducing errors by automatic segmentation. W take advantage of video which captures experimental procedure to perform ﬁne-grain manual segmentation. This video is speciﬁcally used in our prototype to isolate an didentify segments of the golf swing.
In feature extraction, an exhaustive set of features is considered to ensure capturing as much useful information as possible for each movement segment. We extract an exhaustive set of time-domain features including statistical and morphological features. Each statistical feature is a mathematical function taken over a complete segment. Morphological features,however, are calculated from m uniformly distributed samples over a complete segment.
The quantitative model performs further analysison the features extracted from all sensor nodes in order to obtain a quality metric.The quality of each segment of a golf swing can be measured with respect to different criteria. Examples of such criteria include the amount of wrist rotation and how out of plane a swing is. We develop our model for quantifying golf swings with respect to several criteria. Our model employs feature conditioning techniques to reﬁne features contributing to the quality of the swing.Although the exhaustive set o features maintains relevant information on the quality the swing, it contains relatively large number of redundant features. On the other hand, curse of dimensionality is an impediment for our system as our sensor nodes are constrained interms of computational capabilities, communication bandwidth and memory. A high dimensional feature space requires more bandwidth for trans mission and more computation for quality analysis. Furthermore, quality of a golf swing with respect to each criterion can be expressed by certain properties of the physical movement. Therefore, speciﬁc tools are required to extract such attributes from thesignal. We use several signal processing techniques including PCA (Principal Component Analysis) and LDA (Local Discriminant Analysis) to obtain signiﬁcant information with respect to each criterion.