CN107812343B - A kind of vault sports training method - Google Patents
A kind of vault sports training method Download PDFInfo
- Publication number
- CN107812343B CN107812343B CN201710753357.0A CN201710753357A CN107812343B CN 107812343 B CN107812343 B CN 107812343B CN 201710753357 A CN201710753357 A CN 201710753357A CN 107812343 B CN107812343 B CN 107812343B
- Authority
- CN
- China
- Prior art keywords
- vault
- weight
- population
- hidden
- chromosome
- 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.)
- Active
Links
- 238000012549 training Methods 0.000 title claims abstract description 49
- 238000000034 method Methods 0.000 title claims abstract description 38
- 239000013598 vector Substances 0.000 claims abstract description 41
- 238000012360 testing method Methods 0.000 claims abstract description 11
- 239000011159 matrix material Substances 0.000 claims abstract description 4
- 210000000349 chromosome Anatomy 0.000 claims description 24
- 230000005484 gravity Effects 0.000 claims description 23
- 210000003127 knee Anatomy 0.000 claims description 19
- 238000013528 artificial neural network Methods 0.000 claims description 17
- 210000003423 ankle Anatomy 0.000 claims description 13
- 108090000623 proteins and genes Proteins 0.000 claims description 12
- 230000037396 body weight Effects 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 claims description 8
- 239000000203 mixture Substances 0.000 claims description 5
- 239000000284 extract Substances 0.000 claims description 4
- 230000035772 mutation Effects 0.000 claims description 4
- 230000002860 competitive effect Effects 0.000 claims description 3
- 238000006073 displacement reaction Methods 0.000 claims description 3
- 230000005284 excitation Effects 0.000 claims description 3
- 230000006870 function Effects 0.000 claims description 3
- 238000010276 construction Methods 0.000 abstract description 2
- 230000000977 initiatory effect Effects 0.000 abstract 1
- 238000005516 engineering process Methods 0.000 description 7
- 210000002414 leg Anatomy 0.000 description 6
- 230000000694 effects Effects 0.000 description 3
- 210000003484 anatomy Anatomy 0.000 description 2
- 238000005452 bending Methods 0.000 description 2
- 239000003550 marker Substances 0.000 description 2
- 230000001537 neural effect Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000002360 preparation method Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000013210 evaluation model Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000010187 selection method Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
Classifications
-
- A—HUMAN NECESSITIES
- A63—SPORTS; GAMES; AMUSEMENTS
- A63B—APPARATUS FOR PHYSICAL TRAINING, GYMNASTICS, SWIMMING, CLIMBING, OR FENCING; BALL GAMES; TRAINING EQUIPMENT
- A63B5/00—Apparatus for jumping
- A63B5/02—High-jumping posts
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- A—HUMAN NECESSITIES
- A63—SPORTS; GAMES; AMUSEMENTS
- A63B—APPARATUS FOR PHYSICAL TRAINING, GYMNASTICS, SWIMMING, CLIMBING, OR FENCING; BALL GAMES; TRAINING EQUIPMENT
- A63B2244/00—Sports without balls
- A63B2244/08—Jumping, vaulting
- A63B2244/085—Pole vaulting
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Theoretical Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Physiology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Biology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Physical Education & Sports Medicine (AREA)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
The invention discloses a kind of vault sports training methods, comprising the following steps: step 1: obtaining human body key position motion profile during vault;Step 2: multiple motion profiles are extracted into feature vector;Step 3: feature vector and the amount for characterizing bar success or not are merged into new feature vector;Step 4: using the empirical value estimation technique initiation parameter NP, F, CR, generate initial population at random;Step 5: calculating hidden layer output matrix and obtain output weight and export the corresponding root-mean-square error of weight institute, root-mean-square error is as fitness index;Step 6: carrying out variation and crossing operation obtains new individual;Step 7: being judged whether to continue model training according to the number of iterations;Step 8: model being tested using test sample, vault training model construction is finally completed and is trained using the model;Where the method for the present invention can quickly and effectively find out training problem, thus the nonstandard technical movements of corrective exercise person.
Description
Technical field
The present invention relates to training technical field, in particular to a kind of vault sports training method.
Background technique
Pole vault is one of the project that technology is the most complicated in track and field event.In pole vault campaign, take-off, group
Body, bar excessively are several important basic technologies.To cross the bar, sportsman must be obtained certain by flying jump
Kinetic energy completes action in the air in the stage of bunching up the body so that crossing bar.Take-off bunch up the body movement complete quality directly affect sportsman's energy
It is no at merits and demerits bar.During take-off, bunching up the body, crossing bar, speed, flight angle, motion profile, position and and the cross bar of sportsman
Distance be very important.
Pole vault sports achievement in China's improves comparatively fast at present, but integral level lags behind always the world, Asia at
Achievement ranking is also undesirable.For physical fitness, China pole-vaulter and the foreign top pole-vaulter in the world
And apparent difference is not present.The objective factor of China's athletes ' performance is so influenced, first electing is exactly the tactical level raced.
And take-off, bunch up the body, cross pole technology then become pole vault tactics raising very important aspect.
Currently, Toshihiko Fukushima et al. utilizes pole vault robot, have studied more positive on bar
Bunch up the body, pole vault achievement can be improved.GuanyuLiu et al. is proposed by Dynamics Optimization Controlling model, in conjunction with measurement
Relevant parameter, to analyze the performance of pole-vaulter, and show holding rod height, take-off angle and strut
The relationship of degree of flexibility and pole vault achievement.Satoshi Nishikawa et al. utilizes robot research pole vault process
In effectively swing opportunity.
Above-mentioned research to pole vault or it only considered some link or to have ignored human body specific on bar
Action form.Therefore how to propose that effective training pattern is extremely important.
Summary of the invention
The present invention provides a kind of vault sports training method, realizes and determined bar according to the motion profile of sportsman
Success or not effectively improves training effect to realize to the guidance for skipping bar process is acted.
A kind of vault sports training method, comprising the following steps:
Step 1: obtaining human body key position motion profile during vault;
Motion profile captures system using OptiTrack three-dimensional infrared moving, by arranging mark point on human body, by height
Fast camera receives the infrared light of label point reflection, is obtained by identification mark point.The high speed camera used in the present embodiment
For 12 Prime13 (abbreviation p13), most high frame rate is 240fps.
Step 2: multiple motion profiles that step 1 is obtained extract feature vector;
Step 3: by feature vector that step 2 obtains and characterized bar success or not amount be merged into new feature to
Amount, as training sample, obtains N number of training sample, if excitation function is g (x), maximum number of iterations GmaxAnd hidden node
Number L;
Step 4: using the ginseng in the empirical value estimation technique initialization very fast learning machine list hidden layer feedforward neural network of canonical
NP, F, CR are counted, parameter NP refers to the individual number of composition initial population, and parameter F refers to that the zoom factor of mutation operation, CR are to intersect
The crossover probability of operation, CR ∈ [0,1];
Random to generate initial population, preparation starts to train, i.e., initial population is x (t)={ ai,bi, wherein ai=[ai1,
ai2,…,ain] it is the input weight for connecting i-th of hidden node, wherein aiSubscript n in vector expression indicates each input
The dimension of sample, biIt is the deviation of i-th of hidden node, ai、biIn subscript i value be 1 to arrive L, L indicates as described in step 3
The number of hidden node;
Step 5: for each population, (population when first time iteration herein refers to the initial population in step 4, changes for the first time
The population of generation later herein refers to the population in last iteration after the cross and variation that step 6 obtains), it is defeated to calculate hidden layer
Matrix obtains output weight and exports the corresponding root-mean-square error of weight institute out,
Root-mean-square error is as fitness index;
Step 6: population (population when first time iteration herein refers to the initial population in step 4, after first time iteration this
The population at place refer in last iteration by step 6 obtain cross and variation after population) in individual according to difference into
Change algorithm, advanced row variation obtains g+1 generation individual vi(g+1), then crossing operation acquisition g+1 is carried out for new individual ui(g+1), it presses
Individual of the competitive individual as next-generation population (this population is known as the population after cross and variation) is chosen according to following formula,
Wherein, E () is the fitness index in step 5, the calculating of the root-mean-square error in calculation such as step 5
Formula.
Step 7: being judged whether to continue model training according to the number of iterations, repeat to walk if within the scope of the number of iterations
Rapid 5 and step 6, until target is completed to obtain vault training model, obtained if being more than the number of iterations range of setting
Export weightComplete training;
Step 8: being tested using the model that test sample obtains step 7, be finally completed vault training mould
Type is constructed and is trained using the model.
The realization process of single hidden layer feedforward neural network into step 8 of above-mentioned steps 1 is known as DE_RELM algorithm, that is, adopts
RELM algorithm (i.e. regularization extreme learning machine) is improved with DE algorithm (i.e. differential evolution algorithm).
In order to improve the measuring accuracy of model, it is preferred that in step 1, human body key position is transported during obtaining vault
Specific step is as follows for dynamic rail mark:
1-1 adheres to mark point in two hips, left ankle, left knee of sportsman;
The location information of mark point at left ankle and left knee is used to describe sportsman and lifts the leg stage in i.e. layback of bunching up the body,
The bending degree of both legs and at a distance from body.It is blocked since centre of body weight mark point may occur during the motion
The phenomenon that, center of gravity can not directly be tracked, so our two hips of selected marker, with the midpoint approximating anatomy weight of its line
The heart.Each vault result is sequentially recorded in file, and result, quality are not fed back to sportsman.
1-2 acquires mark point motion profile, acquires n group altogether, wherein crossing bar successful trail n1Group crosses bar failure track n2Group,
Wherein n1+n2=n;
During collecting sample, if not completing bar movement or the touched landing of cross bar, then it is assumed that be bar
Failure;Otherwise it is denoted as into merits and demerits bar.
The n group echo point motion profile that step 1-2 is acquired is divided into two groups by 1-3: training group and test group.
In order to improve the measuring accuracy of model, it is preferred that in step 2, extracted by multiple motion profiles that step 1 acquires special
It is as follows to levy vector process:
2-1 records ankle in the minimum point of the track y of world coordinate system vertical direction Y-axis, as ski-jump, if it is sat
It is designated as (xmin,ymin,zmin), and the coordinate for recording first maximum point that the track y after ski-jump reaches is (xmax,ymax,
zmax);
2-2 carrys out approximate fits people using the motion profile line midpoint of two hips as the motion profile of two hip central points, with this
The motion profile of the weight heart;
2-3 calculates two frame position of centre of gravitys after following ski-jump closely in world coordinate system X, Y, the displacement of Z-direction and two interframe
The time interval inverse of frame speed (time interval of two interframe be) ratio, as sportsman at ski-jump in Y direction
SpeedThe speed of X-directionWith the speed on Z axisThe corresponding flight angle θ of ski-jump according to the following formula1:
It willAnd θ1Feature when as take-off;
Above-mentioned flight angle can reflect the case where " Free taking off " technology is grasped indirectly.The theoretical basis of " Free taking off "
It is: keeps horizontal velocity when sportsman's take-off as far as possible, and this speed is transformed into subsequent action in the air, obtain good
Pendulum mass effect.
2-4 calculates the mark point at left knee at a distance from gravity center of human body, when the distance reaches minimum, by left hip and
The line and left ankle of left knee and the line of left knee calculate the left knee angle θ of human body at this time according to the cosine law2, and choose θ2、
The speed of center of gravityPosition of centre of gravity information p20=(x20,y20,z20) as course movement feature of bunching up the body;
It bunches up the body the stage, it is desirable that sportsman's both legs are bent and draw close to chest, and back is almost parallel to the ground.Therefore can lead to
When the mark point crossed at left knee reaches minimum at a distance from gravity center of human body, determine that sportsman is in the state of bunching up the body.
Centre of body weight reaches the coordinate p for highest point of prancing during 2-5 record pole vault3m=(x3m,y3m,z3m), and will
Barycentric velocity when peakingThe horizontal distance l of centre of body weight and cross barxAnd x3mRatio
lx0, vertical range lyAnd y3mRatio ly0As the feature for spending the bar stage;
Good pole technology of crossing should be that centre of body weight reaches when prancing highest point, and center of gravity is exactly in the surface of cross bar.
In order to improve the measuring accuracy of model, it is preferred that in step 3, the feature vector that step 2-3~2-5 is obtained, with
And characterized the amount of bar success or not and be merged into new feature vector, obtain the training sample of 16 following dimensional vectors:
X={ v1,θ1,υ2,θ2,p20,υ3m,lx0,ly0}。
In order to improve the measuring accuracy of model, it is preferred that the position of centre of gravity information p in step 2-420=(x20,y20,z20)
Calculation be record stage of bunching up the body world coordinate system in position of centre of gravity coordinate p2=(x2,y2,z2) and step 2-5
The ratio of middle respective coordinates, it may be assumed that
In order to improve the measuring accuracy of model, it is preferred that the parameter NP in step 4 refers to the individual of composition initial population
Number, parameter F refers to the zoom factor of mutation operation in step 6, decides the method ratio of bias vector, in step 6, is become
Different is that (population when first time iteration herein refers to the initial population in step 4, the population of first time iteration later herein from population
Refer to the population in last iteration after the cross and variation that step 6 obtains) 3 individuals of middle random selection: xk1, xk2,
xk3, and k1 ≠ k2 ≠ k3 ≠ i, then obtain following vi(g+1) calculation formula:
vi(g+1)=xk1(g)+F·(xk2(g)-xk3(g))
xi(g) g i-th article of chromosome of generation in population is indicated, wherein i value is k1, k2, k3, vi(g+1) it indicates to make a variation
To g+1 for i-th article of chromosome of chromosome.
Preferably, in step 6, crossing operation is carried out by following formula:
Wherein jrandFor the random number on [1,2 ..., D];
D is the dimension of initialization population;
uj,i(g+1) j-th strip gene of the g+1 for i-th article of chromosome of chromosome for intersecting acquisition is indicated;
vj,i(g+1) indicate that variation obtains g+1 for the j-th strip gene of i-th article of chromosome of chromosome;
xj,i(g) the j-th strip gene of g i-th article of chromosome of generation in initialization population is indicated;
Rand (0,1) is indicated in the equally distributed random number in (0,1) section.
In order to improve the measuring accuracy of model, it is preferred that in step 5, aiAnd biFor hidden layer feed forward neural single in step 4
Hidden node parameter in network, aiFor the input weight for connecting i-th of hidden node, biFor the deviation of i-th of hidden node, βi
Indicate the outer weight vector between i-th of hidden node of connection and network output, vector dimension m, m are single hidden layer feed forward neural
The number exported in network, G (ai,bi,xj) indicate that single hidden layer i-th of hidden node of feedforward neural network corresponds to sample x the
The hidden node output of j gene, tjIndicating, there is the neural networks with single hidden layer of L hidden node, which to correspond to n-th sample, corresponds to
Output in j-th of output, expression formula is as follows:
Wherein N indicates input sample number.
In order to improve the measuring accuracy of model, it is preferred that in step 7, export weightCalculation formula it is as follows:
Wherein L is the number of hidden node;
M is the number exported in single hidden layer feedforward neural network,Indicate that i-th of hidden node of connection and network are defeated
Outer weight vector final output weight between out.
The present invention analyze original very limited learning machine hide parameter selection method on the basis of, by Regularization Theory with
And differential evolution algorithm carries out parameter optimization to ELM, proposes DE_RELM and skips rod model.By will be during vault
The kinematic parameter and cross bar success or not composition sample vector that sportsman's key position extracts, training DE_RELM, which rises, skips bar mould
Type realizes technological guidance whether strut skips bar, and then improves the technical level of vault.
Beneficial effects of the present invention:
The method of the present invention constructs vault kinematic parameter Evaluation model, refers to after model foundation to vault training
It leads, if model determines vault result to cross bar failure, the feature vector that this is extracted rises with model Plays skips bar
The feature vector of extraction compares and analyzes, and finds out the place of problem, thus the nonstandard technical movements of corrective exercise person.
Detailed description of the invention
Fig. 1 is the wire frame flow chart of the method for the present invention.
Fig. 2 is ankle Y-axis trajectory diagram before run-up final stage to pendulum mass obtained in the method for the present invention.
Specific embodiment
As shown in Figure 1, the vault sports training method of the present embodiment, comprising the following steps:
Step 1: obtaining human body key position motion profile during vault;
Motion profile captures system using OptiTrack three-dimensional infrared moving, by arranging mark point on human body, by height
Fast camera receives the infrared light of label point reflection, is obtained by identification mark point.The high speed camera used in the present embodiment
For 12 Prime13 (abbreviation p13), most high frame rate is 240fps.
Step 2: extracting feature vector by multiple motion profiles that step 1 acquires;
Step 3: the feature vector that step 2 is obtained, and characterized bar success or not amount be merged into new feature to
Amount, as training sample, obtains N number of training sample.And set excitation function as g (x), maximum number of iterations GmaxAnd hidden layer section
The number L of point;
The present embodiment acquires 200 groups of data samples altogether, wherein crossing bar at merits and demerits bar 153 times and failing 47 times, wherein 150 groups
For training, i.e. N=150, another 50 groups for testing.In 150 groups of training sample, having 115 groups of data is at merits and demerits bar 35
Group data were bar failure;In 50 groups of test samples, having 38 groups of data is into merits and demerits bar, and 12 groups of data were bar failure.
Step 4: using parameter NP, F, CR of empirical value estimation technique initialization RELM, and one group of initial population is generated at random,
Preparation starts to train, i.e., initial population is x (t)={ ai,bi, wherein ai=[ai1,ai2,…,ain] it is i-th of hidden layer section of connection
The input weight of point, wherein aiSubscript n in vector expression indicates the dimension of each input sample, biIt is i-th of hidden node
Deviation, ai、biIn subscript i value be 1 to arrive L, L indicates the number of hidden node as described in step 3, takes 7 in the present embodiment,
Its determination process is as follows:
Empirically formula determines the number of hidden nodes:
Wherein, L, n, m respectively represent hidden layer, input layer, output layer number of nodes.Because of the input that the present embodiment finally determines
The corresponding vector dimension of sample is 16, so n=16, output, which is then only indicated, is into merits and demerits bar or failure, so m=1, generation
Enter above formula and round up to show that the number of hidden nodes is L=7.
NP=200 in the present embodiment for the training sample of acquisition and the summation of test sample number, while passing through many experiments
CR=0.6, F=0.8 are chosen,
Step 5: for each population, calculating hidden layer output matrix and obtain output weight, and find out its corresponding root mean square
ErrorAs fitness index.
Step 6: for the individual in population according to DE algorithm, advanced row variation obtains individual vi(g+1), then intersection acquisition is carried out
New individual ui(g+1), the individual that competitive individual enters next-generation population is then chosen, is chosen shown in process such as formula (1):
Wherein, E () represents its fitness index, the calculating formula of the root-mean-square error in calculation such as step 5.
Step 7: being judged whether to continue model training according to the number of iterations.It repeats to walk if within the scope of the number of iterations
Rapid 5 and step 6, until target is completed, otherwise obtain output weightComplete training;
The maximum number of iterations of algorithm is set as 500 in the present embodiment, and algorithm executes 10 times.
Step 8: being tested using the model that test sample obtains step 7, be finally completed model construction.
In above-mentioned steps 1, specific step is as follows for human body key position motion profile during obtaining vault:
1-1 adheres to mark point in two hips, left ankle, left knee of sportsman;
The location information of mark point at left ankle and left knee is used to describe sportsman and lifts the leg stage in i.e. layback of bunching up the body,
The bending degree of both legs and at a distance from body.It is blocked since centre of body weight mark point may occur during the motion
The phenomenon that, center of gravity can not directly be tracked, so our two hips of selected marker, with the midpoint approximating anatomy weight of its line
The heart.Each vault result is sequentially recorded in file, and result, quality are not fed back to sportsman.
1-2 acquires mark point motion profile, acquires n group altogether, wherein crossing bar successful trail n1Group crosses bar failure track n2Group,
Wherein n1+n2=n;
During collecting sample, if not completing bar movement or the touched landing of cross bar, then it is assumed that be bar
Failure use -1 indicates;Otherwise it is denoted as into merits and demerits bar, use 1 indicates.
As previously mentioned, data sample number n=200, wherein at merits and demerits bar n1=153 times, cross bar failure n2=47 times,
The n group echo point motion profile that step 1-2 is acquired is divided into two groups by 1-3: training group and test group.
In above-mentioned steps 2, it is as follows that feature vector process is extracted by multiple motion profiles that step 1 acquires:
2-1 records ankle in the minimum point of the track y of world coordinate system vertical direction Y-axis, as ski-jump, if it is sat
It is designated as (xmin,ymin,zmin), and the coordinate for recording first maximum point that the track y after ski-jump reaches is (xmax,ymax,
zmax);
Ankle Y-axis track is as shown in Figure 2 before run-up final stage to pendulum mass.
2-2 carrys out approximate fits people using the motion profile line midpoint of two hips as the motion profile of two hip central points, with this
The motion profile of the weight heart;
2-3 calculates two frame position of centre of gravitys after following ski-jump closely in world coordinate system X, Y, the displacement of Z-direction and two interframe
The time interval inverse of frame speed (time interval of two interframe be) ratio, as sportsman at ski-jump in Y direction
SpeedThe speed of X-directionWith the speed on Z axisThe corresponding flight angle θ of ski-jump according to the following formula1:
It willAnd θ1Feature when as take-off;
Above-mentioned flight angle can reflect the case where " Free taking off " technology is grasped indirectly.The theoretical basis of " Free taking off "
It is: keeps horizontal velocity when sportsman's take-off as far as possible, and this speed is transformed into subsequent action in the air, obtain good
Pendulum mass effect.
2-4 calculates the mark point at left knee at a distance from gravity center of human body, when the distance reaches minimum, utilizes left hip
With the line and left ankle of left knee and the line of left knee, the left knee angle θ of human body at this time is calculated according to the cosine law2, and select
Take θ2, center of gravity speedPosition of centre of gravity information p20=(x20,y20,z20) as course movement feature of bunching up the body;
It bunches up the body the stage, it is desirable that sportsman's both legs are bent and draw close to chest, and back is almost parallel to the ground.Therefore can lead to
When the mark point crossed at left knee reaches minimum at a distance from gravity center of human body, determine that sportsman is in the state of bunching up the body.
Centre of body weight reaches the coordinate p for highest point of prancing during 2-5 record pole vault3m=(x3m,y3m,z3m), and will
Barycentric velocity when peakingThe horizontal distance l of centre of body weight and cross barxAnd x3mRatio
lx0, vertical range lyAnd y3mRatio ly0As the feature for spending the bar stage.
Good pole technology of crossing should be that centre of body weight reaches when prancing highest point, and center of gravity is exactly in the surface of cross bar.
The feature vector that above-mentioned steps 2 are obtained, and characterized the amount of bar success or not and be merged into new feature vector,
It is 16 following dimensional vectors as the feature vector in training sample:
X={ v1,θ1,υ2,θ2,p20,υ3m,lx0,ly0}
Position of centre of gravity information p in above-mentioned steps 2-420=(x20,y20,z20) calculation be record bunch up the body the stage
World coordinate system in position of centre of gravity coordinate p2=(x2,y2,z2) and step 2-5 in respective coordinates ratio, it may be assumed that
Parameter NP in above-mentioned steps 4 refers to the individual number of composition initial population, and parameter F refers in claim 1 step 6
The zoom factor of mutation operation, decides the method ratio of bias vector, and variation is 3 individuals of random selection from population: xk1,
xk2, xk3, and k1 ≠ k2 ≠ k3 ≠ i, then
vi(g+1)=xk1(g)+F·(xk2(g)-xk3(g))
xi(g) g i-th article of chromosome of generation in population is indicated, wherein i value is k1, k2, k3, vi(g+1) it indicates to make a variation
To g+1 for i-th article of chromosome of chromosome.
Parameter CR in above-mentioned steps 4 refers to the crossover probability of crossing operation in step 6, and CR ∈ [0,1], crossing operation is such as
Shown in lower:
Wherein jrandFor the random number on [1,2 ..., D], D is the dimension of initialization population, uj,i(g+1) it indicates to intersect to obtain
J-th strip gene of the g+1 obtained for i-th article of chromosome of chromosome, vj,i(g+1) indicate that variation obtains g+1 for chromosome i-th
The j-th strip gene of chromosome, xj,i(g) indicate initialization population in g generation i-th article of chromosome j-th strip gene, rand (0,
1) it indicates in the equally distributed random number in (0,1) section;
A in above-mentioned steps 5iAnd biJoin for hidden node in the mono- hidden layer feedforward neural network of RELM described in step 4
Number is respectively the input weight of i-th of hidden node of connection and the deviation of i-th of hidden node, βiIndicate i-th of connection
Outer weight vector between hidden node and network output, vector dimension m, m are defeated in the mono- hidden layer feedforward neural network of RELM
Number out, G (ai,bi,xj) indicate that mono- i-th of the hidden node of hidden layer feedforward neural network of RELM corresponds to sample x j-th strip
The hidden node of gene exports, tjIt is corresponding defeated to indicate that the neural networks with single hidden layer with L hidden node corresponds to n-th sample
J-th of output in out, expression formula are as follows:
Wherein N indicates input sample number.
In above-mentioned steps 7It is shown below:
Wherein L is the number of hidden node, and m is the number exported in the mono- hidden layer feedforward neural network of RELM,It indicates
Connect the outer weight vector final output weight between i-th of hidden node and network output.
The realization process of single hidden layer feedforward neural network into step 8 of above-mentioned steps 1 is known as DE_RELM algorithm, that is, adopts
(i.e. differential evolution algorithm) method is calculated with DE to improve RELM algorithm (i.e. regularization extreme learning machine).
In order to compare, traditional ELM, SVM support vector machines is respectively adopted, DE_RELM algorithm proposed by the present invention carries out
Sort operation.Algorithm all executes 10 times respectively, and average value is recorded.As shown in the table, it can be seen that in iteration time
Number, one timing of hidden node, the error rate of DE_RELM algorithm is lower, and experimental result is ideal.
1 three kinds of algorithm performance comparisons of table
From upper table, it can be seen that for vault track classification problem, the standard that DE_RELM algorithm classifies to test sample
True rate is better than other two kinds of algorithms.
Claims (9)
1. a kind of vault sports training method, which comprises the following steps:
Step 1: obtaining human body key position motion profile during vault;
Step 2: multiple motion profiles that step 1 is obtained extract feature vector;
Step 3: feature vector that step 2 obtains and the amount for characterizing bar success or not being merged into new feature vector, made
For training sample, N number of training sample is obtained, if excitation function is g (x), maximum number of iterations GmaxAnd the number of hidden node
L;
Step 4: using the empirical value estimation technique initialization the very fast learning machine list hidden layer feedforward neural network of canonical in parameter NP,
F, CR, parameter NP refer to the individual number of composition initial population, and parameter F refers to that the zoom factor of mutation operation, CR are crossing operations
Crossover probability, CR ∈ [0,1];
Random to generate initial population, initial population is x (t)={ ai,bi, wherein ai=[ai1,ai2,…,ain] it is i-th of connection
The input weight of hidden node, wherein aiSubscript n in vector expression indicates the dimension of each input sample, biIt is hidden i-th
The deviation of node layer, ai、biIn subscript i value be 1 to arrive L, L indicates the number of hidden node as described in step 3;
Step 5: for each population, calculating hidden layer output matrix and obtain output weight and export the corresponding root mean square of weight institute
Error, root-mean-square error is as fitness index;
Step 6: the advanced row variation of individual in population obtains g+1 generation individual vi(g+1), then crossing operation acquisition g+1 generation is carried out newly
Individual ui(g+1), individual of the competitive individual as next-generation population is chosen according to the following equation,
Wherein, E () is the fitness index in step 5;
Step 7: being judged whether to continue model training according to the number of iterations, step 5 is repeated if within the scope of the number of iterations
It is exported if being more than the number of iterations range of setting with step 6 until target is completed to obtain vault training model
WeightComplete training;
Step 8: being tested using the model that test sample obtains step 7, be finally completed vault training model structure
It builds and is trained using the model.
2. vault sports training method according to claim 1, which is characterized in that in step 1, obtain strut jump process
Specific step is as follows for middle human body key position motion profile:
1-1 adheres to mark point in two hips, left ankle, left knee of sportsman;
1-2 acquires mark point motion profile, acquires n group altogether, wherein crossing bar successful trail n1Group crosses bar failure track n2Group, wherein
n1+n2=n;
The n group echo point motion profile that step 1-2 is acquired is divided into two groups by 1-3: training group and test group.
3. vault sports training method according to claim 1, which is characterized in that in step 2, acquired by step 1
It is as follows that multiple motion profiles extract feature vector process:
2-1 records ankle in the minimum point of the track y of world coordinate system vertical direction Y-axis, as ski-jump, if its coordinate is
(xmin,ymin,zmin), and the coordinate for recording first maximum point that the track y after ski-jump reaches is (xmax,ymax,zmax);
2-2 carrys out approximate fits human body weight using the motion profile line midpoint of two hips as the motion profile of two hip central points, with this
The motion profile of the heart;
Two frame position of centre of gravitys after 2-3 calculating ski-jump are in world coordinate system X, Y, between the displacement of Z-direction and the time of two interframe
Every ratio, as sportsman at ski-jump Y direction speedThe speed of X-directionWith the speed on Z axis
The corresponding flight angle θ of ski-jump according to the following formula1:
It willAnd θ1Feature when as take-off;
2-4 calculates the mark point at left knee at a distance from gravity center of human body, when the distance reaches minimum, by left hip and left knee
Line and the line of left ankle and left knee the left knee angle θ of human body at this time is calculated according to the cosine law2, and choose θ2, center of gravity
SpeedPosition of centre of gravity information p20=(x20,y20,z20) as course movement feature of bunching up the body;
Centre of body weight reaches the coordinate p for highest point of prancing during 2-5 record pole vault3m=(x3m,y3m,z3m), and be up to
Barycentric velocity when highest pointThe horizontal distance l of centre of body weight and cross barxAnd x3mRatio lx0, hang down
Directly distance lyAnd y3mRatio ly0As the feature for spending the bar stage.
4. vault sports training method according to claim 3, which is characterized in that in step 3, by step 2-3~2-5
The feature vector of acquisition, and characterized the amount of bar success or not and be merged into new feature vector, obtain 16 following dimensional vectors
Training sample:
X={ v1,θ1,υ2,θ2,p20,υ3m,lx0,ly0}。
5. vault sports training method according to claim 3, which is characterized in that the position of centre of gravity letter in step 2-4
Cease p20=(x20,y20,z20) calculation be record stage of bunching up the body world coordinate system in position of centre of gravity coordinate p2=
(x2,y2,z2) and step 2-5 in respective coordinates ratio, it may be assumed that
6. vault sports training method according to claim 1, which is characterized in that in step 6, carrying out variation is from kind
3 individuals: x are randomly choosed in groupk1, xk2, xk3, and k1 ≠ k2 ≠ k3 ≠ i, then obtain following vi(g+1) calculation formula:
vi(g+1)=xk1(g)+F·(xk2(g)-xk3(g))
xi(g) g i-th article of chromosome of generation in population is indicated, wherein i value is k1, k2, k3, vi(g+1) indicate that variation obtains g
+ 1 chromosome of generation chromosome i-th.
7. vault sports training method according to claim 1, which is characterized in that in step 6, by following formula into
Row crossing operation:
Wherein jrandFor the random number on [1,2 ..., D];
D is the dimension of initialization population;
uj,i(g+1) j-th strip gene of the g+1 for i-th article of chromosome of chromosome for intersecting acquisition is indicated;
vj,i(g+1) indicate that variation obtains g+1 for the j-th strip gene of i-th article of chromosome of chromosome;
xj,i(g) the j-th strip gene of g i-th article of chromosome of generation in initialization population is indicated;
Rand (0,1) is indicated in the equally distributed random number in (0,1) section.
8. vault sports training method according to claim 1, which is characterized in that in step 5, aiAnd biFor in step 4
Hidden node parameter in single hidden layer feedforward neural network, aiFor the input weight for connecting i-th of hidden node, biIt is hidden for i-th
The deviation of node layer, βiIndicate the outer weight vector between i-th of hidden node of connection and network output, vector dimension m, m are
The number exported in the very fast learning machine list hidden layer feedforward neural network of canonical, G (ai,bi,xj) indicate the very fast learning machine list of canonical
The hidden node that i-th of hidden node of hidden layer feedforward neural network corresponds to sample x j-th strip gene exports, tjIndicate that there is L
The neural networks with single hidden layer of a hidden node corresponds to j-th of output in the corresponding output of n-th sample, the following institute of expression formula
Show:
Wherein N indicates input sample number.
9. vault sports training method according to claim 1, which is characterized in that in step 7, export weightMeter
It is as follows to calculate formula:
Wherein L is the number of hidden node;
M is the number exported in single hidden layer feedforward neural network,Indicate that i-th of hidden node of connection and network export it
Between outer weight vector final output weight.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710753357.0A CN107812343B (en) | 2017-08-29 | 2017-08-29 | A kind of vault sports training method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710753357.0A CN107812343B (en) | 2017-08-29 | 2017-08-29 | A kind of vault sports training method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107812343A CN107812343A (en) | 2018-03-20 |
CN107812343B true CN107812343B (en) | 2019-12-03 |
Family
ID=61600490
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710753357.0A Active CN107812343B (en) | 2017-08-29 | 2017-08-29 | A kind of vault sports training method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107812343B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111144260A (en) * | 2019-12-19 | 2020-05-12 | 北京文安智能技术股份有限公司 | Detection method, device and system of crossing gate |
CN113128108B (en) * | 2021-04-07 | 2023-05-26 | 汕头大学 | Method for determining diameter of jet grouting pile based on differential evolution artificial intelligence |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7244216B2 (en) * | 2003-04-05 | 2007-07-17 | Omnitek Partners Llc | Method and system for preventing pole vault fall injuries |
CN105303252A (en) * | 2015-10-12 | 2016-02-03 | 国家计算机网络与信息安全管理中心 | Multi-stage nerve network model training method based on genetic algorithm |
CN105615890B (en) * | 2015-12-24 | 2018-08-10 | 西安交通大学 | Human body lower limbs walking joint angles myoelectricity continuous decoding method |
CN106964137B (en) * | 2017-02-23 | 2019-05-28 | 浙江理工大学 | A kind of ball service behavior rating method based on image procossing |
CN106945059B (en) * | 2017-03-27 | 2019-11-22 | 中国地质大学(武汉) | A kind of gesture tracking method based on population random disorder multi-objective genetic algorithm |
-
2017
- 2017-08-29 CN CN201710753357.0A patent/CN107812343B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN107812343A (en) | 2018-03-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN102682302B (en) | Human body posture identification method based on multi-characteristic fusion of key frame | |
CN101558996B (en) | Gait recognition method based on orthogonal projection three-dimensional reconstruction of human motion structure | |
CN105868779B (en) | A kind of Activity recognition method based on feature enhancing and Decision fusion | |
CN109086698A (en) | A kind of human motion recognition method based on Fusion | |
CN107812343B (en) | A kind of vault sports training method | |
CN106714220A (en) | WSN (Wireless Sensor Network) anomaly detection method based on MEA-BP neural network | |
CN105912991B (en) | Activity recognition based on 3D point cloud and crucial bone node | |
CN107944431A (en) | A kind of intelligent identification Method based on motion change | |
CN109062962A (en) | A kind of gating cycle neural network point of interest recommended method merging Weather information | |
CN111931804B (en) | Human body action automatic scoring method based on RGBD camera | |
CN109101911B (en) | Visual analysis method for football match formation change and personnel flow | |
CN106778796A (en) | Human motion recognition method and system based on hybrid cooperative model training | |
CN108363959A (en) | One kind being directed to table tennis or badminton action identification method | |
CN108549866A (en) | Remote sensing aeroplane recognition methods based on intensive convolutional neural networks | |
CN108211318B (en) | Based on the race walking posture analysis method perceived in many ways | |
CN110110686A (en) | Based on the human motion recognition methods for losing double-current convolutional neural networks more | |
CN110490109A (en) | A kind of online human body recovery action identification method based on monocular vision | |
CN110222977A (en) | One kind movement sport methods of marking based on computer vision and device | |
CN108717548B (en) | Behavior recognition model updating method and system for dynamic increase of sensors | |
KR102369945B1 (en) | Device and method to discriminate excersice stance using pressure | |
CN106073793A (en) | Attitude Tracking based on micro-inertia sensor and recognition methods | |
CN108211268A (en) | Exercise load monitoring and sports fatigue method for early warning and system based on training data | |
CN108229435A (en) | Method for pedestrian recognition | |
CN110287896A (en) | A kind of Human bodys' response method based on heterogeneous layering PSO and SVM | |
CN101826155A (en) | Method for identifying act of shooting based on Haar characteristic and dynamic time sequence matching |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
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 | ||
TR01 | Transfer of patent right | ||
TR01 | Transfer of patent right |
Effective date of registration: 20201015 Address after: 226200 No. 46 Industrial Park, Nanyang Town, Nantong, Jiangsu, Qidong Patentee after: Qidong plant medical equipment factory Address before: 310018, No. 928, No. 2, Poplar Street, Hangzhou economic and Technological Development Zone, Zhejiang Patentee before: Zhejiang University of Technology |