CN108171278A - A kind of recognizing model of movement method and system based on training data - Google Patents
A kind of recognizing model of movement method and system based on training data Download PDFInfo
- Publication number
- CN108171278A CN108171278A CN201810074757.3A CN201810074757A CN108171278A CN 108171278 A CN108171278 A CN 108171278A CN 201810074757 A CN201810074757 A CN 201810074757A CN 108171278 A CN108171278 A CN 108171278A
- Authority
- CN
- China
- Prior art keywords
- players
- information
- base station
- acceleration
- speed
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of recognizing model of movement method and system based on training data, including step:Obtain the three dimensional local information of players;According to the three-dimensional coordinate information of acquisition, the speed and acceleration information of players are calculated;Collect training sample data;Speed in training sample data and acceleration time series are pre-processed;Using binary decision tree support vector machines decision Tree algorithms, BP neural network algorithm, RBF neural recognizer this three classes recognizer, to treated, training sample data are trained;The motor pattern of players is identified using above-mentioned three classes trained recognizer, and passes through Weighted Fusion method and obtains final pattern recognition result.The present invention more succinctly, conveniently, can realize recognizing model of movement under training and competition scene complicated and changeable, improve the precision and robustness of recognizing model of movement in information collection side's plane system structure.
Description
Technical field
The invention belongs to exercise data analysis fields, are related to exercise data acquisition and recognizing model of movement, more particularly to one
Recognizing model of movement method and system of the kind based on training data.
Background technology
With the rapid development of computer hardware technique and network, video camera, sensor and wireless sensor network are utilized
Network etc. acquires movement position information and technique and tactics data in in-situ match, has become a kind of for the competitiveness of lifter motion person
Trend, motion information acquisition and analytical equipment play an important role wherein.It can be adopted by motion tracking and positioning device
Collect the kinematics and dynamics data of sportsman, these data are handled by data mining technology, obtain motor pattern feature, it can
For accurately moving technical-tactics analyzing.
Accurate training monitoring is carried out, be unable to do without the support of motion tracking and positioning device.It is domestic at present common
Indoor and outdoor wireless location technology be broadly divided into two big technology category of imaging and non-imaged.Imaging class can not interfere fortune completely
Kinematic data acquisition is completed under the premise of mobilization, is monitored in real time suitable for competition field.But such product technology is complicated, realizes difficulty
Height, it is expensive, it is difficult to promote.Non-imaged class technology mainly includes:Global positioning system (GPS), BEI-DOU position system, auxiliary
Global positioning system (A-GPS), infrared, ultrasonic, WLAN, bluetooth, radio frequency identification (RFID), purple honeybee (ZigBee) and
Ultra wide band (UWB).Compared with other technologies, UWB wireless location technologies have that low in energy consumption, anti-multipath effect is good, registration, safety
Property it is high, system structure is simple the advantages that.But UWB label individuals at present on the market are larger, are unfavorable for sportsman and wear in play
It wears.
Currently used algorithm for pattern recognition mainly has two classes:Statistical-simulation spectrometry classification and neural network filter
Classification.
The correlative study in recognizing model of movement field at present is concentrated mainly on the mould under relatively simple scene in daily life
Formula identifies, mainly to walking, running or the simple actions such as stair activity carry out pattern-recognition, for supporting national physical training and old man automatic
The demands such as nurse, the recognizing model of movement research under complicated competition and training scene, which is not much, to be seen, existing mode identifies standard
True rate is also less high, and the higher supercomputer that needs to employ of indivedual discriminations is handled with more complicated algorithm, no
Suitable for promoting and applying.
In terms of recognizing model of movement, laboratory research is more, and is concentrated mainly on simple motion pattern in living scene
Identification.The geometric properties pair of acceleration information that Chen Lei, Yang Jie etc. are acquired using the acceleration transducer for being worn on human body crotch
Upstairs, downstairs, run, stand and the identification of the motion states such as walking, discrimination is between 70% to 95%;Sa-kwang
Song et al. using three dimension acceleration sensor embedded in mobile phone the common action state in daily life is identified with
Realize the monitoring to daily routines, the activity involved in research includes running, sitting down, standing up, falling, and uses multilayer perceptron
Activity is identified, discrimination has reached 97.9% to 99.30%;Yu-Jin Hong et al. combine acceleration transducer with
Radio frequency identification guards the life of old man, its cardinal principle is that electronics mark is sticked on daily necessities at home
Label are worn over being integrated with acceleration transducer and the device of radio-frequency apparatus and can perceive these articles on arm, so as to old
The action of year people judges, for example answers the call, cleaning etc., on the other hand utilizes the acceleration transducer being placed on leg
Some action states are identified, thereby realize the basic monitoring of the daily life to the elderly;Wang Changxi etc. is by two
A acceleration transducer is fixed on human upper limb, by acquiring the acceleration information of two three dimension acceleration sensors, logarithm
It is mainly acceleration, image and myoelectricity etc. by the main information that wavelet transformation is used for recognizing model of movement after being pre-processed.
Although Yi Shang researcher realizes higher pattern-recognition rate in simple scenario, these algorithms can not support more multiple
It is miscellaneous, and motor pattern changes rapid competition and training application.
In conclusion 1. existing motion positions equipment are broadly divided into based on optics and GPS+ inertial navigation two types, it is preceding
Person due to itself it is intrinsic the defects of, move that positioning accuracy is not high, and system itself is made when being blocked under compound movement environment
High price is expensive, high to computer hardware requirement, and the latter is mainly used for the positioning of outdoor sport project, and positioning accuracy is not high.2.
Existing recognizing model of movement algorithm can not adapt to be accurately identified in the environment of compound movement based on mostly differentiating with threshold value,
Subsequent technical-tactics analyzing and exercise load can not more preferably be supported to calculate.3. at present existing indoor positioning mode as based on RFID and
Either positioning accuracy is not high or label is excessive is unfavorable for carrying by WIFI etc., and is only capable of realizing two-dimensional space positioning, does not prop up
Hold three-dimensional fix.
Invention content
In view of the deficiencies of the prior art, the present invention provides a kind of recognizing model of movement method based on training data,
It realizes and real time position tracking and recognizing model of movement is carried out to training and match, be that subsequent technical-tactics analyzing and training are supervised
Control provides the scientific and technological level of accurate real-time data supporting, training for promotion and match.
In order to achieve the above object, technical solution provided by the invention is:A kind of movement mould based on training data
Formula recognition methods, includes the following steps,
Step 1, the three dimensional local information of players is obtained;
Step 2, the three-dimensional coordinate information obtained according to step 1 calculates the speed and acceleration information of players;
Step 3, training sample data, speed and acceleration information including players are collected;
Step 4, the speed in training sample data and acceleration time series are pre-processed;
Step 5, known using binary decision tree support vector machines decision Tree algorithms, BP neural network algorithm, RBF neural
To treated, training sample data are trained this three classes recognizer of other algorithm;
Step 6, the motor pattern of players is identified, and pass through using above-mentioned three classes trained recognizer
Weighted Fusion method obtains final pattern recognition result.
Further, the three dimensional local information of players, the alignment system packet are obtained in step 1 by alignment system
Include positioning label, locating base station, server, client;
The positioning label includes micro-control unit (MCU), transmitting antenna and power supply module, and positioning label is led to by UWB
Believe that agreement emits signal to locating base station;
The locating base station includes reception antenna, MCU, network communication module and power supply module, and locating base station passes through reception
Label wireless signal is positioned, the number of the locating base station is at least 4;
Location processor is connected with locating base station, for the information received according to locating base station, using the time difference to fixed
Position label is positioned, and obtains (x, y, the z) three dimensional local information and time series of positioning label;
The server is connected by WIFI or cable network with location processor, and label information is positioned for storing;
The client obtains the three dimensional local information and time series of positioning label by accessing server.
Further, speed and the acceleration of players are calculated with difference algorithm according to three-dimensional coordinate information in step 2
Information is spent, realization method is as follows,
If Δ t is the speed V of time interval, then playersiAnd acceleration AiFor,
Wherein, Xi, Yi, ZiFor three-dimensional coordinate.
Further, the realization method pre-processed in step 4 to the Velocity Time sequence in training sample data is such as
Under,
According to the feature of the rate conversion frequency, the minimum value N and maximum value M of a length of window are defined, then from speed
First point of sequence starts, and with N point lines, calculates each point error of fitting, if error total value is less than given threshold values R, fitting
Success, increases the length of window, and line digital simulation error again if error of fitting is less than R, continues to increase length of window always
It is M to length of window;If error of fitting is more than R, first segment terminates, the starting point of the point as new window, continues same
The process of sample, until sequence division finishes.
Further, the realization method for obtaining final pattern recognition result in step 6 by Weighted Fusion method is as follows,
Wherein, wi(i=1,2,3) is weight of algorithms of different, i=1, when 2,3, correspond to respectively RBF neural algorithm,
RP neural network algorithms and binary decision tree support vector machines decision Tree algorithms;μiFor variable, imitated according to the identification of algorithms of different
Fruit (0,0.5] in the range of value;X represents input data, fu(x) it is the corresponding final pattern recognition result of input data,
Its maximum value corresponds to classification y, and y ∈ Y, wherein Y are motor pattern category set;fi(x) it is that i-th of pattern-recognition is calculated
Method as a result, corresponding weights are wi, [fi(x)=yi] represent corresponding pattern recognition classifier yi。
The present invention also provides a kind of recognizing model of movement system based on training data, including following module:
Three dimensional local information acquisition module, for obtaining the three dimensional local information of players;
Speed and acceleration calculation module for the three-dimensional coordinate information obtained according to step 1, calculate the speed of players
Degree and acceleration information;
Training sample data collection module, for collecting training sample data, speed and acceleration including players
Information;
Preprocessing module, for being pre-processed to the speed in training sample data and acceleration time series;
Training module, for utilizing binary decision tree support vector machines decision Tree algorithms, BP neural network algorithm, RBF god
Through Network Recognition algorithm this three classes recognizer, to treated, training sample data are trained;
Pattern recognition module, for being carried out using the trained recognizer of above-mentioned three classes to the motor pattern of players
Identification, and pass through Weighted Fusion method and obtain final pattern recognition result.
Further, the three-dimensional position for obtaining players in three dimensional local information acquisition module by alignment system is believed
Breath, the alignment system include positioning label, locating base station, server, client;
The positioning label includes micro-control unit (MCU), transmitting antenna and power supply module, and positioning label is led to by UWB
Believe that agreement emits signal to locating base station;
The locating base station includes reception antenna, MCU, network communication module and power supply module, and locating base station passes through reception
Label wireless signal is positioned, the number of the locating base station is at least 4;
Location processor is connected with locating base station, for the information received according to locating base station, using the time difference to fixed
Position label is positioned, and obtains (x, y, the z) three dimensional local information and time series of positioning label;
The server is connected by WIFI or cable network with location processor, and label information is positioned for storing;
The client obtains the three dimensional local information and time series of positioning label by accessing server.
Further, it calculates and moves with difference algorithm according to three-dimensional coordinate information in speed and acceleration calculation module
The speed and acceleration information of personnel, realization method is as follows,
If Δ t is the speed V of time interval, then playersiAnd acceleration AiFor,
Wherein, Xi, Yi, ZiFor three-dimensional coordinate.
Further, the realization side pre-processed in preprocessing module to the Velocity Time sequence in training sample data
Formula is as follows,
According to the feature of the rate conversion frequency, the minimum value N and maximum value M of a length of window are defined, then from speed
First point of sequence starts, and with N point lines, calculates each point error of fitting, if error total value is less than given threshold values R, fitting
Success, increases the length of window, and line digital simulation error again if error of fitting is less than R, continues to increase length of window always
It is M to length of window;If error of fitting is more than R, first segment terminates, the starting point of the point as new window, continues same
The process of sample, until sequence division finishes.
Further, the realization side of final pattern recognition result is obtained in pattern recognition module by Weighted Fusion method
Formula is as follows,
Wherein, wi(i=1,2,3) is weight of algorithms of different, i=1, when 2,3, correspond to respectively RBF neural algorithm,
RP neural network algorithms and binary decision tree support vector machines decision Tree algorithms;μiFor variable, imitated according to the identification of algorithms of different
Fruit (0,0.5] in the range of value;X represents input data, fu(x) it is the corresponding final pattern recognition result of input data,
Its maximum value corresponds to classification y, and y ∈ Y, wherein Y are motor pattern category set;fi(x) it is that i-th of pattern-recognition is calculated
Method as a result, corresponding weights are wi, [fi(x)=yi] represent corresponding pattern recognition classifier yi。
Compared with prior art, advantages of the present invention:The present invention is based on the recognizing model of movement methods of training data
Contain the monitoring positioned to exercise data and the classification and information extraction of motor pattern feature.
Compared with prior art, the present invention is more succinct, conveniently in information collection side's plane system structure, and sportsman only needs to take
Band positioning label, you can obtain the three dimensional local information of sportsman and sports apparatus, number in real time with new UWB wireless computings technology
According to can be as accurate as Centimeter Level;And no longer need to carry the sensors such as three-dimensional acceleration and heart rate, portability is increased, is reduced
System complexity and energy consumption.
On recognizing model of movement, present invention employs complexity higher binary decision tree support vector machines and neural network
Blending algorithm increases the precision of pattern-recognition under compound movement environment, so as under training and competition scene complicated and changeable
Recognizing model of movement is realized, and improves the precision and robustness of recognizing model of movement.
Description of the drawings
Fig. 1 is positioning system structure schematic diagram in the embodiment of the present invention.
Fig. 2 is alignment system work flow diagram in the embodiment of the present invention.
Fig. 3 is typical action type Binominal Tree in the embodiment of the present invention.
Fig. 4 is support vector machines Decision Tree Algorithm flow chart in the embodiment of the present invention.
Fig. 5 is binary decision tree support vector machines and neural network fusion algorithm flow chart in the embodiment of the present invention.
Specific embodiment
Technical scheme of the present invention is described further with reference to the accompanying drawings and examples.
Technical solution of the present invention includes two stages, and training monitors the classification and extraction of stage and valid data in real time
Stage.Training monitors the stage using the effective training data of alignment system acquisition in real time, then real-time to training data
The realization of computer automatic running can be used in analysis, the stage;Alignment system is by positioning label, locating base station, server, computer
It is formed with mobile terminal.Main acquisition, storage and the classification of valid data for realizing data of training data analysis in real time and
Extraction, main algorithm are Motion feature extraction algorithm.
1. alignment system:Position label, locating base station, server, client (structure chart is shown in attached drawing 1).
Wherein positioning label is mainly made of micro-control unit (MCU), transmitting antenna and power supply module.Position the master of label
It act as emitting signal to communication base station by UWB communication protocols.
Locating base station is mainly by reception antenna, MCU, network communication module and power supply module composition.Locating base station (at least 4
It is a) by receiving the wireless signal for positioning label and emitting by UWB communication protocols, the main function of locating base station is received for record
Time and tag serial number to positioning label signal.
Location processor is connected with locating base station, for the information received according to locating base station, using the time difference to fixed
Position label is positioned, and obtains (x, y, the z) three dimensional local information and time series of positioning label.
The main function of server is:It is connected by WIFI or cable network with location processor, is positioned for storing
Label information, including the number for positioning label and corresponding position time series;
Client obtains the position time series of positioning label by accessing server, and the reception terminal is computer
Or mobile terminal device, such as iPad.
Alignment system work flow diagram is as shown in Figure 2:
S1, positioning label power on action;
S2, positioning label enter working region;
S3, location processor receives the beacon signal that positioning label is sent out, if entering step S4, if not, judging
Positioning label is in non-wake-up states, prompts to restart positioning label;
S4, positioning label are successfully waken up, and send out UWB positioning signals;
S5, locating base station receives UWB signal, and is handled;
The UWB signal of receiving and receiving time are uploaded to location processor by S6, locating base station;
S7, the time difference that location processor receives information according to different base station show that the exact position of positioning label is believed
Breath;
Accurate location information is uploaded to server by S8, location processor;
2. motor pattern feature extraction algorithm:With binary decision tree support vector machines and neural network fusion algorithm, root
The real-time position information, speed and the acceleration information that are obtained according to alignment system realize the accurate judgement of motor pattern, and in insertion
Formula system realizes the algorithm, ensures real-time and the low-power consumption of algorithm.The specific steps are:
(1) according to three-dimensional coordinate sequence obtained by alignment system, with difference algorithm calculate sportsman or instrument speed and
Acceleration information, Δ t are time interval, and 0.02s~0.1s is taken in the present embodiment.
(2) the typical action mode division based on binary tree, action be divided into away in the present embodiment, run, jump, fighting,
Non- this five kinds of typical action types of confrontation, binary tree, which divides, sees attached drawing 3, and each action data is as root node;Then it is classified as
The left subtree that depth is 1 runs, walks, jumping and right subtree is static;The left subtree that depth is 1 is further divided into the left subtree that depth is 2
It runs, walk, right subtree is jumped, and the right subtree that depth is 1 is divided into leaf node confrontation and non-confrontation;The left subtree that depth is 2 is finally divided into
Leaf node is walked and is run.When it is implemented, practical training data, can be divided according to sports events type;
(3) typical training sample data are collected, are mainly included per the data characteristics in class training sample, position, adds speed
Velocity characteristic (comprising itself and with other sportsmen and the relative information of instrument), training stage, players speed and acceleration
It spends known to corresponding motor pattern classification.
(4) it in order to reduce error, needs first to pre-process speed and acceleration time series.Pass through medium filtering first
The isolated noise point in two sequences is filtered out, time series point is then carried out according to sliding window algorithm according to rate curve feature
It cuts.Specially:According to the feature of the rate conversion frequency, the minimum value N and maximum value M of a length of window are defined, then rapidly
First point of degree series start, and with N point lines, calculate each point error of fitting, if error total value is less than given threshold values R, intend
Work(is synthesized, increases the length of window, line digital simulation error again if error of fitting is less than R, continues to increase length of window one
Until length of window is M.If error of fitting is more than R, first segment terminates, the starting point of the point as new window, continues
Same process, until sequence division finishes.In the same way can acceleration time series pre-processed.
(5) Separatory measure between class.The present embodiment to training action be divided into five classes, x in training sample datak, k
=1,2,3,4,5 correspond to, run, jump, fighting, non-confrontation respectively;Then class i in sample data, j (i, j take 1,2,3,4,5) between
Separatory measure formula it is as follows:
For linear sample collection
Wherein dijRepresent class i center ciWith class j center cjBetween Euclidean distance;
dij=| | ci-cj||
ci、cj、σi、σjIt is calculated according to training sample data, with class xiFor
niFor class xiThe number of middle training sample.
For non-linear training sample set, Nonlinear Mapping Φ is to it, is converted into the linear problem of higher dimensional space,
Then Separatory measure calculating is carried out in specific high latitude feature space, calculation formula is same as above.
(6) modeling is trained to training sample data with binary decision tree support vector machines decision Tree algorithms.Training
Start in the process from the class being most difficult to point, that is, separation is selected between class to estimate minimum two classes m and n and is trained, obtains SVM (N-
1), and merged a new class cluster m '.Further in class cluster m ' and during other do not participate in trained class or class cluster, searching is most difficult to point
Two classes or class cluster p and q, be trained to obtain SVM (N-2), merge a new class cluster p '.Can finally obtain (N-1) it is a support to
Amount machine grader.This training process is from leaf node, is condensed to minor matters by merging to be most difficult to classify, ultimately forms root node.
Specific algorithm flow is shown in attached drawing 4:Initialization leaf node (being equivalent to motor pattern classification) and training sample data first;Judge leaf
Whether node number is more than 2, if it is greater, then the separation calculated between arbitrary two class or class cluster is estimated, chooses class spacing minimum
Using the sample data of m and n as training sample, disaggregated model is obtained by training by two classes or class cluster m and n;It is then combined with class
Or class cluster m and n obtain new class cluster m ', and update decision tree structure body variable, and with other class cluster m ' is not further participated in instruction
Experienced class or class cluster is compared, and is judged that separation is estimated, is obtained final support vector machine classifier;If leaf node number is
2, directly using the sample data of m and n as training sample, disaggregated model is obtained by training, algorithm terminates.
Training sample data are instructed using existing BP neural network and RBF neural algorithm for pattern recognition simultaneously
Practice.
(7) algorithm fusion.On the basis of binary decision tree algorithm of support vector machine, binary decision tree supporting vector is constructed
Machine adds neural network to mix algorithm.Different algorithm for pattern recognitions has different specialities, incorporates the multiple of each feature
The fusion of algorithm for pattern recognition can allow pattern-recognition effect and robustness further to improve.This algorithm fusion uses decision level
Fusion, convergence strategy Organization Chart as shown in Figure 5, first input exercise data, and movement mould is obtained using above-mentioned three classes algorithm
Then formula classification obtains final classification results, and the convergence strategy for exporting this algorithm is, by adding using fusion decision function
Power merges multiple algorithm for pattern recognitions:
Wherein wi(i=1,2,3) is weight of algorithms of different, i=1, when 2,3, correspond to respectively RBF neural algorithm,
RP neural network algorithms and support vector cassification algorithm;μiFor variable, existed according to the recognition effect of different mode recognizer
(0,0.5] in the range of value.Final pattern recognition result is calculated by the result of each algorithm for pattern recognition according to weighted formula
Go out:
In formula, x represents input data, fu(x) it is the corresponding final pattern recognition result of input data, and it is maximum
Value correspond to classification y, and y ∈ Y,
Wherein Y is motor pattern category set.fi(x) it is i-th of algorithm for pattern recognition as a result, corresponding weights are wi,
[fi(x)=yi] represent corresponding pattern recognition classifier yi。
The embodiment of the present invention also provides a kind of recognizing model of movement system based on training data, including such as lower die
Block:
Three dimensional local information acquisition module, for obtaining the three dimensional local information of players;
Speed and acceleration calculation module for the three-dimensional coordinate information obtained according to step 1, calculate the speed of players
Degree and acceleration information;
Training sample data collection module, for collecting training sample data, the three dimensional local information including players,
The motor pattern classification of speed and acceleration information and players;
Preprocessing module, for being pre-processed to the speed in training sample data and acceleration time series;
Training module, for utilizing binary decision tree support vector machines decision Tree algorithms, BP neural network algorithm, RBF god
Through Network Recognition algorithm this three classes recognizer, to treated, training sample data are trained;
Pattern recognition module, for being carried out using the trained recognizer of above-mentioned three classes to the motor pattern of players
Identification, and pass through Weighted Fusion method and obtain final pattern recognition result.
Wherein, the three dimensional local information of players, institute are obtained in three dimensional local information acquisition module by alignment system
It states alignment system and includes positioning label, locating base station, server, client;
The positioning label includes micro-control unit (MCU), transmitting antenna and power supply module, and positioning label is led to by UWB
Believe that agreement emits signal to locating base station;
The locating base station includes reception antenna, MCU, network communication module and power supply module, and locating base station passes through reception
Label wireless signal is positioned, the number of the locating base station is at least 4;
Location processor is connected with locating base station, for the information received according to locating base station, using the time difference to fixed
Position label is positioned, and obtains (x, y, the z) three dimensional local information and time series of positioning label;
The server is connected by WIFI or cable network with location processor, and label information is positioned for storing;
The client obtains the three dimensional local information and time series of positioning label by accessing server.
Wherein, players are calculated with difference algorithm according to three-dimensional coordinate information in speed and acceleration calculation module
Speed and acceleration information, realization method is as follows,
If Δ t is the speed V of time interval, then playersiAnd acceleration AiFor,
Wherein, Xi, Yi, ZiFor three-dimensional coordinate.
Wherein, the realization method pre-processed in preprocessing module to the Velocity Time sequence in training sample data is such as
Under,
According to the feature of the rate conversion frequency, the minimum value N and maximum value M of a length of window are defined, then from speed
First point of sequence starts, and with N point lines, calculates each point error of fitting, if error total value is less than given threshold values R, fitting
Success, increases the length of window, and line digital simulation error again if error of fitting is less than R, continues to increase length of window always
It is M to length of window;If error of fitting is more than R, first segment terminates, the starting point of the point as new window, continues same
The process of sample, until sequence division finishes.
Wherein, the realization method of final pattern recognition result is obtained such as by Weighted Fusion method in pattern recognition module
Under,
Wherein, wi(i=1,2,3) is weight of algorithms of different, i=1, when 2,3, correspond to respectively RBF neural algorithm,
RP neural network algorithms and binary decision tree support vector machines decision Tree algorithms;μiFor variable, imitated according to the identification of algorithms of different
Fruit (0,0.5] in the range of value;X represents input data, fu(x) it is the corresponding final pattern recognition result of input data,
Its maximum value corresponds to classification y, and y ∈ Y, wherein Y are motor pattern category set;fi(x) it is that i-th of pattern-recognition is calculated
Method as a result, corresponding weights are wi, [fi(x)=yi] represent corresponding pattern recognition classifier yi。
Each module specific implementation is corresponding with each step, and it will not go into details by the present invention.
The present invention carries out movement mould with BP neural network, RBF neural, SVM (support vector machines) and blending algorithm
Formula feature recognition is tested, and is carried out 10 times altogether, discrimination is as shown in table 1.The discrimination of wherein blending algorithm is best, in 10 experiments
In achieve 9 suboptimums, and relatively stable, average recognition rate highest minimum in 4 kinds of algorithm Plays difference.
The motor pattern feature recognition rate (%) of algorithms of different in 1 10 experiments of table
Specific embodiment described herein is only an example for the spirit of the invention.Technology belonging to the present invention is led
The technical staff in domain can do various modifications or additions to described specific embodiment or replace in a similar way
In generation, however, it does not deviate from the spirit of the invention or beyond the scope of the appended claims.
Claims (10)
- A kind of 1. recognizing model of movement method based on training data, which is characterized in that include the following steps:Step 1, the three dimensional local information of players is obtained;Step 2, the three-dimensional coordinate information obtained according to step 1 calculates the speed and acceleration information of players;Step 3, training sample data, speed and acceleration information including players are collected;Step 4, the speed in training sample data and acceleration time series are pre-processed;Step 5, it is calculated using binary decision tree support vector machines decision Tree algorithms, BP neural network algorithm, RBF neural identification To treated, training sample data are trained this three classes recognizer of method;Step 6, the motor pattern of players is identified using above-mentioned three classes trained recognizer, and passes through weighting Fusion method obtains final pattern recognition result.
- 2. a kind of recognizing model of movement method based on training data as described in claim 1, it is characterised in that:Step In 1 by alignment system obtain players three dimensional local information, the alignment system include positioning label, locating base station, Server, client;The positioning label includes micro-control unit (MCU), transmitting antenna and power supply module, and positioning label passes through UWB communication protocols It discusses to locating base station and emits signal;The locating base station includes reception antenna, MCU, network communication module and power supply module, and locating base station is positioned by receiving Label wireless signal, the number of the locating base station are at least 4;Location processor is connected with locating base station, and for the information received according to locating base station, positioning is marked using the time difference Label are positioned, and obtain (x, y, the z) three dimensional local information and time series of positioning label;The server is connected by WIFI or cable network with location processor, and label information is positioned for storing;The client obtains the three dimensional local information and time series of positioning label by accessing server.
- 3. a kind of recognizing model of movement method based on training data as described in claim 1, it is characterised in that:Step According to three-dimensional coordinate information in 2, the speed and acceleration information of players are calculated with difference algorithm, realization method is as follows,If Δ t is the speed V of time interval, then playersiAnd acceleration AiFor,Wherein, Xi, Yi, ZiFor three-dimensional coordinate.
- 4. a kind of recognizing model of movement method based on training data as described in claim 1, it is characterised in that:Step The realization method pre-processed in 4 to the Velocity Time sequence in training sample data is as follows,According to the feature of the rate conversion frequency, the minimum value N and maximum value M of a length of window are defined, then from velocity series First point starts, and with N point lines, calculates each point error of fitting, if error total value is less than given threshold values R, is fitted successfully, Increase the length of window, line digital simulation error again, if error of fitting is less than R, continue to increase length of window until window Mouth length is M;If error of fitting is more than R, first segment terminates, the starting point of the point as new window, continues same Process, until sequence division finishes.
- 5. a kind of recognizing model of movement method based on training data as described in claim 1, it is characterised in that:Step The realization method for obtaining final pattern recognition result in 6 by Weighted Fusion method is as follows,Wherein, wi(i=1,2,3) is weight of algorithms of different, and when 2,3, it is refreshing correspond to RBF neural algorithm, RP respectively by i=1 Through network algorithm and binary decision tree support vector machines decision Tree algorithms;μiFor variable, existed according to the recognition effect of algorithms of different (0,0.5] in the range of value;X represents input data, fu(x) it is the corresponding final pattern recognition result of input data, most Big value corresponds to classification y, and y ∈ Y, wherein Y are motor pattern category set;fi(x) it is i-th of algorithm for pattern recognition As a result, corresponding weights are wi, [fi(x)=yi] represent corresponding pattern recognition classifier yi。
- 6. a kind of recognizing model of movement system based on training data, which is characterized in that including following module:Three dimensional local information acquisition module, for obtaining the three dimensional local information of players;Speed and acceleration calculation module, for the three-dimensional coordinate information obtained according to step 1, calculate players speed and Acceleration information;Training sample data collection module, for collecting training sample data, speed and acceleration information including players;Preprocessing module, for being pre-processed to the speed in training sample data and acceleration time series;Training module, for utilizing binary decision tree support vector machines decision Tree algorithms, BP neural network algorithm, RBF nerve nets To treated, training sample data are trained this three classes recognizer of network recognizer;Pattern recognition module, for being known using the trained recognizer of above-mentioned three classes to the motor pattern of players Not, and pass through Weighted Fusion method and obtain final pattern recognition result.
- 7. a kind of recognizing model of movement system based on training data as claimed in claim 6, it is characterised in that:It is three-dimensional The three dimensional local information of players is obtained in position information acquisition module by alignment system, the alignment system includes positioning Label, locating base station, server, client;The positioning label includes micro-control unit (MCU), transmitting antenna and power supply module, and positioning label passes through UWB communication protocols It discusses to locating base station and emits signal;The locating base station includes reception antenna, MCU, network communication module and power supply module, and locating base station is positioned by receiving Label wireless signal, the number of the locating base station are at least 4;Location processor is connected with locating base station, and for the information received according to locating base station, positioning is marked using the time difference Label are positioned, and obtain (x, y, the z) three dimensional local information and time series of positioning label;The server is connected by WIFI or cable network with location processor, and label information is positioned for storing;The client obtains the three dimensional local information and time series of positioning label by accessing server.
- 8. a kind of recognizing model of movement system based on training data as claimed in claim 6, it is characterised in that:Speed With according to three-dimensional coordinate information, calculate the speed of players in acceleration calculation module with difference algorithm and acceleration believed Breath, realization method is as follows,If Δ t is the speed V of time interval, then playersiAnd acceleration AiFor,Wherein, Xi, Yi, ZiFor three-dimensional coordinate.
- 9. a kind of recognizing model of movement system based on training data as claimed in claim 6, it is characterised in that:Pre- place The realization method pre-processed in reason module to the Velocity Time sequence in training sample data is as follows,According to the feature of the rate conversion frequency, the minimum value N and maximum value M of a length of window are defined, then from velocity series First point starts, and with N point lines, calculates each point error of fitting, if error total value is less than given threshold values R, is fitted successfully, Increase the length of window, line digital simulation error again, if error of fitting is less than R, continue to increase length of window until window Mouth length is M;If error of fitting is more than R, first segment terminates, the starting point of the point as new window, continues same Process, until sequence division finishes.
- 10. a kind of recognizing model of movement system based on training data as claimed in claim 6, it is characterised in that:Mould The realization method for obtaining final pattern recognition result in formula identification module by Weighted Fusion method is as follows,Wherein, wi(i=1,2,3) is weight of algorithms of different, and when 2,3, it is refreshing correspond to RBF neural algorithm, RP respectively by i=1 Through network algorithm and binary decision tree support vector machines decision Tree algorithms;μiFor variable, existed according to the recognition effect of algorithms of different (0,0.5] in the range of value;X represents input data, fu(x) it is the corresponding final pattern recognition result of input data, most Big value corresponds to classification y, and y ∈ Y, wherein Y are motor pattern category set;fi(x) it is i-th of algorithm for pattern recognition As a result, corresponding weights are wi, [fi(x)=yi] represent corresponding pattern recognition classifier yi。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810074757.3A CN108171278B (en) | 2018-01-25 | 2018-01-25 | Motion pattern recognition method and system based on motion training data |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810074757.3A CN108171278B (en) | 2018-01-25 | 2018-01-25 | Motion pattern recognition method and system based on motion training data |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108171278A true CN108171278A (en) | 2018-06-15 |
CN108171278B CN108171278B (en) | 2020-02-14 |
Family
ID=62515935
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810074757.3A Active CN108171278B (en) | 2018-01-25 | 2018-01-25 | Motion pattern recognition method and system based on motion training data |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108171278B (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109242149A (en) * | 2018-08-13 | 2019-01-18 | 华中师范大学 | A kind of student performance early warning method and system excavated based on educational data |
CN109472281A (en) * | 2018-09-10 | 2019-03-15 | 平安科技(深圳)有限公司 | Motion recording processing method and system, terminal and readable storage medium storing program for executing |
CN110135583A (en) * | 2019-05-23 | 2019-08-16 | 北京地平线机器人技术研发有限公司 | The generation method of markup information, the generating means of markup information and electronic equipment |
CN111694643A (en) * | 2020-05-12 | 2020-09-22 | 中国科学院计算技术研究所 | Task scheduling execution system and method for graph neural network application |
CN111887859A (en) * | 2020-08-05 | 2020-11-06 | 安徽华米智能科技有限公司 | Fall behavior recognition method and device, electronic device and medium |
CN111898933A (en) * | 2019-05-06 | 2020-11-06 | 上海新微技术研发中心有限公司 | Checking equipment and checking method |
CN112513995A (en) * | 2018-06-01 | 2021-03-16 | 弗劳恩霍夫应用研究促进协会 | System for determining game scenes in a sports game |
CN112770251A (en) * | 2020-12-30 | 2021-05-07 | 山东大学 | Self-adaptive human behavior data acquisition and identification system and method based on UWB and LoRa |
CN113628721A (en) * | 2021-07-15 | 2021-11-09 | 武汉中体智美科技有限公司 | Child motion development level evaluation method and system based on motion cycle |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103150016A (en) * | 2013-02-20 | 2013-06-12 | 兰州交通大学 | Multi-person motion capture system fusing ultra wide band positioning technology with inertia sensing technology |
WO2016100369A1 (en) * | 2014-12-15 | 2016-06-23 | i4c Innovations Inc. | Improved distance measuring in a monitoring system for animals |
-
2018
- 2018-01-25 CN CN201810074757.3A patent/CN108171278B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103150016A (en) * | 2013-02-20 | 2013-06-12 | 兰州交通大学 | Multi-person motion capture system fusing ultra wide band positioning technology with inertia sensing technology |
WO2016100369A1 (en) * | 2014-12-15 | 2016-06-23 | i4c Innovations Inc. | Improved distance measuring in a monitoring system for animals |
Non-Patent Citations (1)
Title |
---|
肖玲: "无线体域网中人体动作监测与识别若干方法研究", 《中国博士学位论文全文数据库 信息科技辑,2014年第9期, I136-20》 * |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112513995A (en) * | 2018-06-01 | 2021-03-16 | 弗劳恩霍夫应用研究促进协会 | System for determining game scenes in a sports game |
CN109242149A (en) * | 2018-08-13 | 2019-01-18 | 华中师范大学 | A kind of student performance early warning method and system excavated based on educational data |
CN109242149B (en) * | 2018-08-13 | 2021-08-13 | 华中师范大学 | Student score early warning method and system based on education data mining |
CN109472281A (en) * | 2018-09-10 | 2019-03-15 | 平安科技(深圳)有限公司 | Motion recording processing method and system, terminal and readable storage medium storing program for executing |
CN111898933A (en) * | 2019-05-06 | 2020-11-06 | 上海新微技术研发中心有限公司 | Checking equipment and checking method |
CN110135583A (en) * | 2019-05-23 | 2019-08-16 | 北京地平线机器人技术研发有限公司 | The generation method of markup information, the generating means of markup information and electronic equipment |
CN111694643A (en) * | 2020-05-12 | 2020-09-22 | 中国科学院计算技术研究所 | Task scheduling execution system and method for graph neural network application |
CN111694643B (en) * | 2020-05-12 | 2023-04-11 | 中国科学院计算技术研究所 | Task scheduling execution system and method for graph neural network application |
CN111887859A (en) * | 2020-08-05 | 2020-11-06 | 安徽华米智能科技有限公司 | Fall behavior recognition method and device, electronic device and medium |
CN112770251A (en) * | 2020-12-30 | 2021-05-07 | 山东大学 | Self-adaptive human behavior data acquisition and identification system and method based on UWB and LoRa |
CN113628721A (en) * | 2021-07-15 | 2021-11-09 | 武汉中体智美科技有限公司 | Child motion development level evaluation method and system based on motion cycle |
CN113628721B (en) * | 2021-07-15 | 2024-07-12 | 武汉中体智美科技有限公司 | Child motion development level evaluation method and system based on motion cycle |
Also Published As
Publication number | Publication date |
---|---|
CN108171278B (en) | 2020-02-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108171278A (en) | A kind of recognizing model of movement method and system based on training data | |
CN106096662B (en) | Human motion state identification based on acceleration transducer | |
US9978425B2 (en) | Method and device for associating frames in a video of an activity of a person with an event | |
Yao et al. | A robust step detection and stride length estimation for pedestrian dead reckoning using a smartphone | |
KR101690649B1 (en) | Activity classification in a multi-axis activity monitor device | |
Hu et al. | A human body posture recognition algorithm based on BP neural network for wireless body area networks | |
CN103974887B (en) | Method and apparatus for elevator motion detection | |
CN111144217A (en) | Motion evaluation method based on human body three-dimensional joint point detection | |
CN109643499A (en) | The system and method for swimming analysis | |
Altini et al. | Self-calibration of walking speed estimations using smartphone sensors | |
CN103970271A (en) | Daily activity identifying method with exercising and physiology sensing data fused | |
CN105678222A (en) | Human behavior identification method based on mobile equipment | |
CN108958482B (en) | Similarity action recognition device and method based on convolutional neural network | |
Jensen et al. | Classification of kinematic swimming data with emphasis on resource consumption | |
CN106210269A (en) | A kind of human action identification system and method based on smart mobile phone | |
CN116981401A (en) | Motion monitoring method and system | |
Wang et al. | Electromyography-based locomotion pattern recognition and personal positioning toward improved context-awareness applications | |
CN113663312B (en) | Micro-inertia-based non-apparatus body-building action quality evaluation method | |
Fereidoonian et al. | Human activity recognition: From sensors to applications | |
Wu et al. | Real-time physical activity classification and tracking using wearble sensors | |
CN102156994B (en) | Joint positioning method for single-view unmarked human motion tracking | |
CN110163264A (en) | A kind of walking mode recognition methods based on machine learning | |
Braojos et al. | A wireless body sensor network for activity monitoring with low transmission overhead | |
CN115346272A (en) | Real-time tumble detection method based on depth image sequence | |
CN114847931A (en) | Human motion tracking method, device and computer-readable storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |