TWI664550B - Golf player swing posture detection system - Google Patents

Abstract

A golf player swing posture detection system includes a plurality of wearable devices, an electronic device, an image capture device, a display device, a first processing server that obtains an incorrect posture parameter, and a second processing server Device. The second processing server receives the hand triaxial movement data, the foot triaxial movement data, the waist triaxial movement data, the human skeleton data, the exercise time, the preference parameter, the club type parameter, and The wrong posture parameter is an ontology data structure to be detected, and a humanized ontology data structure is input to a Huffman tree model for comparison to output a detection result. The second processing server executes a weighted imbalanced learning algorithm to calculate the personalized ontology data structure.

Description

Golf player swing posture detection system

Golf player swing posture detection system

Golf is a high-tech sport that emphasizes body coordination and balance. Athlete's swing posture and body rhythm are very important issues. Golf swing can be divided into preparatory period, upper period, lower period, and acceleration. There are six phases including the period, the sending period and the closing period. Each stage is prone to the problem of incorrect posture, and the non-standard swing posture is easy to cause sports injury to the body.

Building a computer system to implement athlete-assisted training services has always been a research topic that is easy to plan but difficult to carry out. Considering the measurement technology of athletes, it is necessary to wear a large number of sensors for measurement, or use image analysis to analyze athletes' movements, which causes a lot of inconvenience in operation. Athletes use wearable sensors to collect sensing data during sports, analyze and infer the data, and provide customized services for athletes. However, each athlete has different habits and standards, so in the process of data analysis and inference, it cannot provide athletes with a more complete training task. So began to use big data technology to collect sensory data of different athletes during sports, so as to obtain perfect data for analysis and inference, and provide training tasks suitable for athletes. Athletes' training task services can be roughly divided into: 1. Personal sensing analysis method: mainly use wearable sensors to be installed on athletes or sports equipment, and devices to capture relevant values of athletes during sports, or use pictures or Video method to record the golfer ’s movements during the swing Analyze and compare the threshold value of the sports model to infer the training task results of the athletes. 2. Mass sensing analysis method: It mainly collects sensory data from different athletes during sports, analyzes, classifies and infers the collected data. The inference results will help athletes to correct errors or assist beginners to learn, thereby realizing suitable athletes. Training task.

Regardless of individual or mass sensing analysis methods, the prior art counting methods for data analysis can be mainly divided into two types: lightweight analysis and giant quantitative analysis. Among them, lightweight analysis captures swing movements for athletes through wearable sensing devices, thereby obtaining relevant motion factors, analyzing and inferring based on the obtained relevant factors, and constructing a customized sports model to provide athletes with ubiquitous personalization service. However, although lightweight analysis takes into account issues such as the individualization of athletes' movements, coordination and immediacy of movements, they often lack scarcity and rationality.

In addition, the quantitative analysis collects relevant information from various channels as the basis for data exploration, inference and analysis, and helps users obtain relevant information of interest. Although issues such as user coordination and integrity are taken into account, they often lack customization and scarcity.

As mentioned above, Wulun is a personal or mass-sensing analysis method. Because the data of erroneous swing movements is relatively scarce, to detect incorrect postures to avoid sports injuries, the existing technology is still insufficient.

Therefore, an object of the present invention is to provide a swing posture detection system that can detect a golf player's swing posture to prevent sports injuries.

Therefore, the golf player swing posture detection system of the present invention is suitable for detecting a golf player's swing posture. The detection system includes a plurality of wearable devices 1, an electronic device 2, an image capture device 3, and a display. Device 4, a first processing server 5, and a 第二 处理 SERVER6。 The second processing server 6.

The plurality of wearable devices capture three-axis motion data of one hand and three-axis motion data of one foot.

The electronic device is connected to the wearable devices to obtain the three-axis movement data of the hand and the three-axis movement data of the foot, and receive a movement time, a preference parameter, and a club type parameter, and retrieve One waist three axis movement data.

The image capturing device includes a compound lens camera for capturing a human skeleton data.

The display device includes a screen unit.

The first processing server is connected to the electronic device, the image capturing device, and the display device. The first processing server executes an inverted transfer neural network model, and receives the three-axis motion data of the hand, the The three-axis motion data of the foot, the three-axis motion data of the waist, the human skeleton data, the exercise time, the preference parameters, and the club type parameters are input to the inverted transfer neural network model to obtain a wrong posture parameter.

The second processing server is connected to the first processing server, stores a humanized ontology data structure and a Huffman tree model, and receives the three-axis motion data of the hand, the three-axis motion data of the foot, the The waist triaxial movement data, the human skeleton data, the movement time, the preference parameter, the club type parameter, and the wrong posture parameter are an ontology data structure to be detected, and the second processing server is based on the personalized ontology The data structure and the ontology data structure to be tested are input to the Huffman tree model for comparison to output a test result, the test result is transmitted to the display device, and the test result is displayed on the screen unit.

Wherein, the second processing server executes a weighted imbalance learning algorithm, Input an original hand triaxial motion data, an original foot triaxial motion data, an original waist triaxial motion data, an original human skeleton data, an original exercise time, an original preference parameter, an original club type parameter, And an original wrong posture parameter to calculate the personalized ontology data structure.

The effect of the present invention is that a more reasonable detection result is calculated by using the weighted imbalance learning algorithm executed by the second processing server, and even if the collected wrong swing posture is rare, the problem posture can still be detected, and The display device can be viewed by a golfer to prevent sports injuries.

2‧‧‧ plural wearable devices

2‧‧‧ electronic device

3‧‧‧Image capture device

31‧‧‧ composite lens camera

32‧‧‧Computer host

4‧‧‧ display device

41‧‧‧Screen unit

5‧‧‧First Processing Server

6‧‧‧Second Processing Server

Other features and effects of the present invention will be clearly presented in the embodiment with reference to the drawings, in which: FIG. 1 is a system block diagram illustrating a golf player's swing posture detection system of the present invention; FIG. 2 is a body A schematic diagram of the data structure, illustrating a set of multiple parameters in the embodiment of the golf player swing posture detection system of the present invention; and FIG. 3 is a schematic diagram of a Huffman tree illustrating the golf player swing of the present invention In an embodiment of the posture detection system, the detection result is calculated.

Referring to FIG. 1, an embodiment of a golf player swing posture detection system according to the present invention is suitable for detecting a golf player's swing posture. The detection system includes a plurality of wearable devices 1, an electronic device 2, an image capture device 3, a display device 4, a first processing server 5, and a second processing server 6.

In this embodiment, there are two wearable devices 1 for golf players to wear on their wrists and ankles respectively. When the golfer makes a swing posture, the three-axis movement data of one hand and one foot are captured. Tri-axial motion data. Taking the three-axis movement data of the hand as an example, the wearable device 1 can obtain the value of the gravity acceleration (G value) in three directions (X-axis, Y-axis, and Z-axis), for example, the X-axis is 3.5G and the Y-axis is The 3.1G and Z axes are 2.9G.

The electronic device 2 is a smart phone in this embodiment, and the wearable devices 1 can be connected through a Bluetooth communication interface to obtain the triaxial motion data of the hand and the triaxial motion data of the foot. At the same time, the user can operate an application program executed by the electronic device 2 to input, so that the electronic device 2 receives a movement time (such as: longer, appropriate, or shorter), a preference parameter (such as: left-hand stick or right-hand stick). ), And a club type parameter (such as: Code: "Drive" or "Wood"). Because the golfer wears the electronic device 2 at the waist, and the existing smart phones are equipped with a gravity sensor, the electronic device 2 can also capture a waist triaxial motion data.

The image capturing device 3 includes a compound lens camera 31 and a computer host 32 executing a Kinect application program. The compound lens camera 31 is an existing Kinect camera. After the composite lens camera 31 obtains a plurality of images of a golf player's dynamics, a human skeleton data is calculated by a Kinect application program executed by the computer host 32. Among them, the human skeleton data is a combined data recording the joints (points) of the human limbs, the trunk (line) of the body, and the coordinates of the limbs (line), and has been commonly used for the detection of human action posture.

The display device 4 is a display in this embodiment, and includes a screen unit 41 for displaying pictures.

The first processing server 5 is connected to the electronic device 2 and the image capturing device 3 And this display device 4. The first processing server 5 executes a Back Propagation Neural Network model, and receives the triaxial motion data of the hand, the triaxial motion data of the foot, the triaxial motion data of the waist, and the human body. Skeleton data, the movement time, the preference parameter, and the club type parameter are input to the inverted transfer neural network model to obtain a wrong posture parameter.

Among them, the incorrect posture parameters are plural false posture data from the following table 1. There are eighteen types from EP01 to EP18:

The aforementioned inverted transfer neural network model is one of the existing neural network-like models, including an input layer, a hidden layer, and an output layer. The first processing server 5 includes the hand triaxial motion data, the foot triaxial motion data, the waist triaxial motion data, the human skeleton data, the exercise time, the preference parameter, and the club type parameter. Input to the input layer of the backward transitive neural network model, after calculation by the hidden layer, output one of the wrong posture data from the wrong posture data to the input layer as the wrong posture parameter.

The second processing server 6 is connected to the first processing server 5 through the Internet, and stores a personal ontology data structure corresponding to the golfer and a Huffman Tree. )model. In particular, the so-called "ontological data architecture" is a tree-like collection of multiple parameters as shown in Figure 2. In this embodiment, the ontological data architecture includes dynamic events and fixed events. The event has three-axis hand data, three-axis foot data, three-axis waist data, wrong posture parameters, and human skeleton data; fixed events have exercise time, preference parameters, and club type parameters. The Huffman tree is a kind of existing special binary tree as shown in FIG. 3. In this embodiment, an ontology data structure to be tested and the personalized ontology data structure are compared in the Huffman tree. Ordinary binary trees can more quickly find the user's motion state corresponding to the ontology data structure to be detected, such as "the waist is too hard" and "the hands are not straightened" recorded at the end leaf nodes of the Huffman tree Etc., and "muscle strain" and "ligament strain" recorded by the leaf nodes above it.

In this embodiment, the second processing server 6 executes a weight-added class-imbalance learning algorithm. The weight-added class-imbalance learning algorithm includes a processing stage and a feature acquisition. Phase, a build candidate phase, and an integration phase. An original hand triaxial motion data, an original foot triaxial motion data, an original waist triaxial motion data, an original human skeleton data, an original exercise time obtained from the golfer's previous swing posture, An original preference parameter, an original club type parameter, and an original wrong posture parameter are input into the weighted imbalance learning algorithm, and sequentially pass through the processing stage, the feature taking stage, the establishment candidate stage, and the During the integration phase, the personal ontology data structure is calculated.

In this processing stage of the augmented imbalanced learning algorithm, the original hand triaxial motion data, the original foot triaxial motion data, the original waist triaxial motion data, the original human skeleton data, The original motion time, the original preference parameter, the original club type parameter, and the original wrong posture parameter are firstly based on an existing AND (Automatic Neighborhood size Determination) algorithm, and then based on an existing SMOTE (Synthetic Minority Over- (sampling Technique) algorithm to calculate a candidate inferred hand triaxial motion data, a candidate inferred foot triaxial motion data, a candidate inferred waist triaxial motion data, a candidate inferred human skeleton data, a candidate inferred exercise time, a Candidate inference preference parameters, a candidate inference club type parameter, and a candidate inference error posture parameter.

In the feature extraction phase of the weighted imbalanced learning algorithm, based on an existing Fast Independent Component Analysis (FastICA) algorithm, the original hand triaxial motion data and the candidate inferred hand Take one of the three-axis motion data as a characteristic hand three-axis motion data, and take one of the original foot three-axis motion data and the candidate inferred foot three-axis motion data as a characteristic foot three Axis motion data, one of the original waist triaxial motion data and the candidate inferred waist triaxial motion data is taken as a characteristic waist triaxial motion data, from the original human skeleton data and the candidate inferred human skeleton data Take one of them as a characteristic human skeleton data, take one of the original exercise time and the candidate inference exercise time as a characteristic exercise time, and take one of the original preference parameter and the candidate inference preference parameter as A feature preference parameter, one of which is a characteristic club type parameter from the original club type parameter and the candidate inferred club type parameter, and one of the original wrong posture parameter and the candidate inferred error posture parameter This is a characteristic error posture parameter.

In the establishment candidate stage of the weighted imbalanced learning algorithm, the characteristic hand triaxial motion data, the characteristic foot triaxial motion data, the characteristic waist triaxial motion data, and the characteristic human skeleton data , The characteristic movement time, the characteristic preference parameter, the characteristic club type parameter, and the characteristic error posture parameter are input into the complex hypothesis algorithm, and the complex hypothesis ontology data structure is calculated respectively. In this embodiment, there are three kinds of hypothetical algorithms, namely an existing Sparse RBMs algorithm modified from Gaussian-binary RBMs (Restricted Boltzmann Machines), and a density-based OPTICS (Ordering Points to Identify the Clustering Structure) ) Clustering algorithm, and a clustering algorithm (Affinity Propagation) algorithm that clusters the feature points of each data. The three hypothesis algorithms can calculate their own hypothetical ontology data structures.

In the integration phase of the weighted imbalanced learning algorithm, it includes a learning sequence with an integrated input layer, an integrated hidden layer, and an integrated output layer ( (Learn to Rank, LTR) neural network model, and the ranking learning neural network model is one of the existing neural network model. The hypothetical ontology data framework is input to the input layer of the sorting learning neural network model, and after calculation by the hidden layer, the personalized ontology data framework is output.

When the golf player makes a swing, the second processing server 6 passes The first processing server 5 receives the real-time tri-axis motion data, the foot tri-axis motion data, the waist tri-axis motion data, the human skeleton data, the exercise time, the preference parameter, and the cue. The type parameter and the wrong posture parameter are an ontology data structure to be detected, and the second processing server 6 is input to the Huffman tree model according to the personalized ontology data structure and the ontology data structure to be detected. The middle comparison is to output a detection result, for example, the detection result is "the waist is too hard" recorded by the terminal leaf node and the "muscle strain" recorded by the leaf node above it. The detection result will be transmitted to the display device 4 through the first processing server 5 and displayed on the screen unit 41. In this embodiment, the screen unit 41 displays "Swing posture detection: The waist is too hard. Please note! It may cause muscle strain. "For golf players to check the test results.

In summary, the golf player's swing posture detection system of the present invention can detect a golf ball through the weighted imbalance learning algorithm of the second processing server 6 even if the collected incorrect swing posture is rare. Athlete's swing action, and then find problematic postures to avoid possible sports injuries.

Claims (7)

A golf player swing posture detection system is suitable for detecting a golf player's swing posture. The detection system includes: a plurality of wearable devices, which captures one hand triaxial motion data and one foot triaxial motion data; An electronic device connected to the wearable devices to obtain the three-axis movement data of the hand and the three-axis movement data of the foot, and receive a movement time, a preference parameter, and a club type parameter, and retrieve Take a waist triaxial motion data; an image capture device, including a compound lens camera to capture a human skeleton data; a display device, including a screen unit; a first processing server, connected to the The electronic device, the image capturing device, and the display device. The first processing server executes an inverted transfer neural network model, and receives the triaxial motion data of the hand, the triaxial motion data of the foot, and the waist three. The axis motion data, the human skeleton data, the motion time, the preference parameters, and the club type parameters are input to the inverted transfer neural network model to obtain Wrong posture parameters, where the inverted transfer neural network model includes an input layer, a hidden layer, and an output layer, the first processing server stores the triaxial motion data of the hand, the triaxial motion data of the foot, The waist triaxial movement data, the human skeleton data, the movement time, the preference parameters, and the club type parameters are input to the input layer of the inverted transfer neural network model. After calculation by the hidden layer, the complex number is incorrect. In the pose data, one of the wrong pose data is output at the input layer as the wrong pose parameter; and a second processing server connected to the first processing server, storing a humanized ontology data structure and a Huffman Tree model, and receive the hand triaxial motion data, the foot triaxial motion data, the waist triaxial motion data, the human skeleton data, the exercise time, the preference parameter, the club type parameter, and the error The posture parameter is a to-be-detected ontology data structure, and the second processing server inputs The Huffman tree model is compared to output a detection result, the detection result is transmitted to the display device, and the detection result is displayed on the screen unit; wherein the second processing server executes a weighted imbalance Learn algorithms, input a raw hand triaxial motion data, a raw foot triaxial motion data, a raw waist triaxial motion data, a raw human skeleton data, a raw exercise time, a raw preference parameter, a raw ball Rod type parameters, and an original wrong posture parameter to calculate the personalized ontology data structure. According to the golf player's swing posture detection system according to item 1 of the patent application scope, wherein the weighted imbalance learning algorithm includes a processing stage, a feature acquisition stage, a candidate establishment stage, and an integration stage, and the Original hand triaxial movement data, the original foot triaxial movement data, the original waist triaxial movement data, the original human skeleton data, the original movement time, the original preference parameters, the original club type parameters, and the The original wrong posture parameters are input into the weighted imbalanced learning algorithm, and sequentially pass through the processing stage, the feature taking stage, the establishment candidate stage, and the integration stage to calculate the personalized ontology data structure. According to the golf player's swing posture detection system according to item 2 of the patent application scope, the processing stage of the weighted imbalance learning algorithm is to input the original hand triaxial motion data and the original foot triaxial Motion data, the original waist triaxial motion data, the original human skeleton data, the original motion time, the original preference parameters, the original club type parameters, and the original wrong posture parameters to calculate a candidate inference hand triaxial Exercise data, a candidate inference foot triaxial motion data, a candidate inference lumbar triaxial motion data, a candidate inferential human skeleton data, a candidate inferential exercise time, a candidate inferential preference parameter, a candidate inferential club type parameter, and A candidate infers wrong posture parameters. The golfer's swing posture detection system according to item 3 of the patent application scope, wherein the feature acquisition phase of the weighted imbalance learning algorithm is from the original hand triaxial motion data and the candidate inferred hand One of the three-axis motion data is a characteristic hand three-axis motion data, and one of the original foot three-axis motion data and the candidate inferred foot three-axis motion data is a characteristic foot three-axis data Take one of the original waist triaxial motion data and the candidate inferred waist triaxial motion data as a characteristic waist triaxial motion data, and take one of the original human skeleton data and the candidate inferred human skeleton data. One is a characteristic human skeleton data. One of the original motion time and the candidate inference motion time is a characteristic motion time, and one of the original preference parameter and the candidate inference motion parameter is a feature. Preference parameter, taking one of the original club type parameter and the candidate inferred club type parameter as a characteristic club type parameter, from the original One of the initial error posture parameter and the candidate inference error posture parameter is a characteristic error posture parameter. According to the golf player's swing posture detection system according to item 4 of the patent application scope, the establishment candidate stage of the weighted imbalanced learning algorithm is based on the characteristic hand triaxial motion data, the characteristic foot three The axis motion data, the characteristic waist triaxial motion data, the characteristic human skeleton data, the characteristic motion time, the characteristic preference parameters, the characteristic club type parameters, and the characteristic error posture parameters are input into the complex hypothesis algorithm and calculated Develop the ontology data structure of the plural hypothesis. The golfer's swing posture detection system according to item 5 of the patent application, wherein the integration phase of the weighted imbalanced learning algorithm includes an integrated input layer, an integrated hidden layer, and an integrated output layer. Sequential learning neural network model. The hypothetical ontology is input to the input layer of the sequencing learning neural network model. After the hidden layer is calculated, the personalized ontology data structure is output from the hypothetical ontology. . According to the golf player ’s swing posture detection system according to item 1 of the scope of patent application, the image capture device further includes a computer host executing a Kinect application program, and the compound lens camera is a Kinect camera, and the compound lens camera obtains After the plurality of images, the human skeleton data is calculated by the computer host.