CN106909891B - A kind of Human bodys' response method based on self feed back gene expression programming - Google Patents
A kind of Human bodys' response method based on self feed back gene expression programming Download PDFInfo
- Publication number
- CN106909891B CN106909891B CN201710059525.6A CN201710059525A CN106909891B CN 106909891 B CN106909891 B CN 106909891B CN 201710059525 A CN201710059525 A CN 201710059525A CN 106909891 B CN106909891 B CN 106909891B
- Authority
- CN
- China
- Prior art keywords
- human body
- artis
- motion
- modelling
- gene expression
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
-
- 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
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Psychiatry (AREA)
- Social Psychology (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of Human bodys' response methods based on self feed back gene expression programming, this method is directed to human body behavior depth image, the three-dimensional time sequence data of the multiple artis of human body is therefrom extracted as sample, sample is modeled using by gene expression programming intersecting, TIS is added after mutation operation insert string operation and construct to obtain self feed back gene expression programming, modelling of human body motion is obtained, TIS inserts string operation and refers to being inserted into functor string on artis motion sequence head；Then gradient information is extracted as the aspect of model.The aspect of model of the modelling of human body motion of training sample is input to neural network, training obtains neural network model as human body behavior classifier；The aspect of model of the corresponding modelling of human body motion of test sample is input in the human body behavior classifier of above-mentioned acquisition, Human bodys' response result is obtained.The present invention has the advantages that Human bodys' response accuracy is high and recognition speed is fast.
Description
Technical field
The present invention relates to Human bodys' response field, in particular to a kind of human body based on self feed back gene expression programming
Activity recognition method.
Background technique
In recent years, Human bodys' response research was one of the Some Questions To Be Researched that Computer Subject is most popular in the world, tool
There is important theoretical research value, research topic is abundant in content, is related to image processing and analysis, machine vision, Human Physiology
The multidisciplinary domain knowledge such as, human cinology, pattern-recognition and artificial intelligence.Human bodys' response research has not only widened
Research direction between section, and the development of related discipline is driven, while being also that smart city and public safety are in the urgent need to address
Key technology, be widely used in human-computer intellectualization, intelligent vision monitoring, content based video retrieval system, virtual reality
Equal fields.The research in the big portion of human bioequivalence at present also only stops at static identification, the complexity and mutability of human motion
So that the Accuracy and high efficiency of identification is unable to satisfy the real requirement of relevant industries.
Currently, mainly using artis wearable sensors either using more the identification of human body behavior in the prior art
Camera carries out multi-angle of view monitoring, and artis wearable sensors, which refer to, puts sensor, this mode at some position of human body
Identify that the behavior type of human body is limited, and recognition accuracy is not high, recognition speed is slow, identification behavior is single and poor expandability；
Multi-angle of view monitoring is carried out using multi-cam and refers to that, in each view position installation camera, the camera for passing through each visual angle is clapped
The picture taken the photograph determines the behavior of human body, this mode there are at high cost, data precision is low, it is inconvenient for use, by environmental factor shadow
Ring the defects of big.
Summary of the invention
The purpose of the present invention is to overcome the shortcomings of the existing technology and deficiency, provides a kind of based on self feed back gene expression
The Human bodys' response method of programming, this method establish modelling of human body motion using self feed back gene expression (SGEP), then
Each anthropoid behavior is accurately portrayed using the gradient information of modelling of human body motion as modelling of human body motion feature, is finally led to
It crosses neural network to identify human body behavior, has the advantages that Human bodys' response accuracy is high and recognition speed is fast.
The purpose of the invention is achieved by the following technical solution: a kind of human body row based on self feed back gene expression programming
For recognition methods, which is characterized in that steps are as follows:
S1, human body behavior depth image is obtained, then extracts human body N pass respectively from human body behavior depth image
The corresponding three-dimensional time sequence data of node, as a sample；
S2, sample is modeled using self feed back gene expression programming, gets the corresponding each artis fortune of sample
Dynamic sequence, to obtain various kinds based on the modelling of human body motion of artis；Wherein the self feed back gene expression programming by
Gene expression programming adds the slotting string operation of TIS after intersection, mutation operation and constructs to obtain, and wherein TIS inserts what string operation referred to
It is to be inserted into functor string on artis motion sequence head, the length for increasing artis motion sequence significance bit increases；
S3, the gradient information of modelling of human body motion is extracted as the aspect of model；
S4, training sample set is got by step S1, and each training sample is got respectively by way of step S2
Then the modelling of human body motion of training sample extracts the model of the modelling of human body motion of each training sample by way of step S3
Feature；
S5, each training sample modelling of human body motion input of the aspect of model as neural network, to neural network into
Row training, the neural network model that training is obtained, as final human body behavior classifier；
S6, test sample is got by step S1, and test sample gets test specimens by way of step S2
This modelling of human body motion, then extracts the aspect of model of the modelling of human body motion of test sample by way of step S3；
S7, the aspect of model of the modelling of human body motion of test sample is input to the human body behavior point got in step S5
In class device, Human bodys' response result is obtained by human body behavior classifier.
Preferably, it is calculated in the step S1 using Microsoft's somatosensory device Kinect combination second generation SDK and OpenCV
Machine vision library obtains human body behavior depth image；TOF (Time Of Flight, flight time) technology is wherein utilized, is calculated micro-
The infrared light that infrared sensor emits in soft Kinect sensor reaches the phase difference of human body back reflection, obtains based on human body
Human body behavior depth image；
The corresponding three-dimensional time sequence data of each artis is to be with the infrared sensor of Microsoft somatosensory device Kinect
What the rectangular coordinate system in space that origin is established obtained, which includes x-axis, y-axis and z axis, and wherein x-axis is being just
Direction is to be parallel to Microsoft's somatosensory device Kinect horizontal direction to the left, and positive direction of the y-axis is perpendicular to Microsoft's somatosensory device
Kinect inclined direction is upward, and z-axis positive direction is Microsoft's somatosensory device Kinect direction.
Preferably, N is 25 or more in the step S1.
It preferably, will be in human body behavior depth image in the step S1 after getting human body behavior depth image
Human body is split with background, is then converted to skeleton tracing system, to obtain multiple artis three-dimensional time sequences of human body
Column data.
Preferably, self feed back gene expression programming models sample in the step S2, obtains based on artis
Modelling of human body motion detailed process is as follows:
S21, determine modelling of human body motion needed for functor collection and terminal symbol collection and terminal symbol number, and set circulation
Termination condition；
S22, the random artis motion sequence initial population of creation, each artis motion sequence is expressed as gene expression
Tree；Artis motion sequence is made of functor collection and terminal symbol collection, and it includes artis three-dimensional time sequence number that terminal symbol, which is concentrated,
According to middle time t and the coordinate components of artis three-dimensional time sequence data；
S23, the adaptive value for calculating each artis motion sequence in population, then by relative error estimation model or absolutely
Error model assesses the adaptability of environment artis motion sequence each in population, qualified in assessment result
In the case of, a certain number of individuals are selected as male parent from population；And judge whether to meet loop termination condition, if full
Foot, then jump to S26, otherwise enter step S24；
S24, the male parent selected according to step S23 carry out gene expression volume to the artis motion sequence in population
The intersection of journey, mutation operation；
S25, the slotting string operation of TIS is carried out to the artis motion sequence of population, obtain next-generation population at individual, return to step
S23；
S26, output model obtain function expression represented by optimal chromosome, the i.e. modelling of human body motion of artis.
Further, functor collection F in step S21 are as follows:
F={ S, C, T, I, R, N, P, Q, E, A, G, L, B, D, F, H }
Wherein it is as follows to represent meaning for each element in functor collection F:
S indicates that SIN function Sin, C indicate that cosine function cos, T indicate that tangent function tan, I indicate arcsin function
Asin, R indicate that inverse cosine function acos, N indicate that arctan function atan, P indicate that power function pow, Q indicate square root function
Sqrt, E indicate that exponential function exp, A that natural number is bottom indicate that ABS function abs, G indicate that sign function sign, L are indicated
Natural logrithm function ln, B indicate that denary logarithm function lg, D indicate that hyperbolic sine function sinh, F indicate hyperbolic cosine
Function cosh, H indicate hyperbolic tangent function tanh；
Terminal symbol collection described in step S21 includes multidimensional terminating character and constant terminating character collection；
Wherein multidimensional terminating character T are as follows:
T={ a, b, c, d, e, f, g, h, i }；
Wherein each element in multidimensional terminal symbol collection T respectively represents each independent variable in multidimensional variable, is examining human body fortune
Actual value corresponding to each variable is substituted into modelling of human body motion when movable model；
Wherein constant terminal symbol collection C are as follows: C={ 1000, -1000 }.
Preferably, which is characterized in that the process that TIS inserts string operation is as follows:
As artis motion sequence effective length m=1, i.e. the root position of artis motion sequence is a constant, this
When, the functor string since artis motion sequence root position to the position [L/4] variation to get at random；Wherein [L/4] table
Show L/4 round numbers part, L is the length of artis motion sequence；While other characters on artis motion sequence head are successively
It moves back, the character string beyond head length is deleted；Wherein L is the length of artis motion sequence, and h is artis motion sequence head
The length in portion；
When artis motion sequence imitates 1 < m of length < h, in the last one functor of artis motion sequence significance bit
It is inserted into the h-m functor strings got at random below, other characters behind the insertion point of artis motion sequence head successively move back,
Character string beyond head length is deleted.
Preferably, in the step S3, for each artis Ji(i=1,2..., N), N are the total of artis in sample
Number, the modelling of human body motion remembered are zi=Fi(x,y,t)；The aspect of model that modelling of human body motion is extracted in step S3 is specific
Process is as follows:
K moment is equidistantly taken to be denoted as tj, j=1,2 ..., k；
Calculate the gradient information of each artisWithI=1,2 ..., N；J=1,2 ..., k；WhereinI-th of artis is respectively indicated in tjMoment corresponding x, y-coordinate value；
By the above-mentioned each artis gradient information being calculatedWithI=1,2 ..., N；J=1,
2 ..., k is z as modelling of human body motioni=FiThe aspect of model of (x, y, t).
Further, it is obtained in the step S5 by the aspect of model training of the modelling of human body motion of each training sample
Detailed process is as follows for neural network model:
S51, neural networks with single hidden layer frame is initially set up, the modelling of human body motion of each training sample will be obtained in step S4
The aspect of modelI=1,2 ..., N；J=1,2 ..., input training sample of the k as neural network
Collection, is provided with N × 2k input neural network, and setting hidden neuron number is N × 4k+1；According to human body behavior packet
The classification number M setting output neuron number included is M, i.e., output vector is { c1,c2,…,cs,…,cM, wherein cs∈ R,
And value interval is [0,1], works as cs=max { c1,c2,…,cMWhen, csBe taken as 1, i.e., neural network output vector be 0,0 ...,
1 ..., 0 }, indicate that human body behavior belongs to s class behavior；
S52, setting neural network learning rate η are the random number between 0~1, and neural network transformation function f (x) is arranged
For Sigmoid function, learning rules are continuous perceptron learning rules, train neural network using error backpropagation algorithm,
Obtain neural network model.
Preferably, which is characterized in that in the step S5, the model for being directed to the modelling of human body motion of each training sample is special
Sign obtains neural network model by the training of ten folding cross-validation methods, as final human body row as the input of neural network
For classifier.
The present invention has the following advantages and effects with respect to the prior art:
(1) Human bodys' response method of the present invention is directed to the human body behavior depth image got, therefrom extracts respectively
Then the corresponding three-dimensional time sequence data of the multiple artis of human body utilizes self feed back gene expression as a sample
Programming models sample, obtains various kinds based on the modelling of human body motion of artis, and extract modelling of human body motion
Gradient information is as the aspect of model.After getting training sample set, training sample is concentrated into the corresponding human motion of training sample
The aspect of model of model is input to neural network, and training is obtained neural network model as human body behavior classifier；It gets
After test sample, the human body behavior that the aspect of model of the corresponding modelling of human body motion of test sample is input to above-mentioned acquisition is classified
In device, Human bodys' response result is obtained by human body behavior classifier.The present invention is used by gene table in the prior art
The slotting string operation of TIS is added after intersection, mutation operation up to formula programming (GEP) to construct to obtain self feed back gene expression programming pair
Sample is modeled, and wherein TIS inserts string operation and refers to being inserted into functor string on artis motion sequence head, is closed for increasing
The length of joint movements sequence significance bit converge to it can more quickly so as to the uniformity to increase population at individual
The globally optimal solution of problem avoids falling into locally optimal solution.Therefore use self feed back gene expression programming to sample in the present invention
This carries out the speed that modeling can effectively improve accuracy and the identification of Human bodys' response.
(2) using the gradient information representative for specific behavior as the model in modelling of human body motion in the present invention
Feature, compared to the prior art using global or local motion feature as the aspect of model, the aspect of model of the invention can
More accurately reflect human body behavior campaign, greatly reduce feature complexity, improves feature representativeness and recognition correct rate.
(3) present invention is obtained using Microsoft's somatosensory device Kinect combination second generation SDK and OpenCV computer vision library
Human body behavior depth image is taken, data precision can be effectively improved, is solved to a certain extent in the prior art using tradition camera shooting
Equipment recognition efficiency is low, recognition effect is big by such environmental effects and to wear joint sensors at high cost, inconvenient to use etc.
Technical problem.
Detailed description of the invention
Fig. 1 is the method for the present invention flow chart.
Fig. 2 is that self feed back gene expression programming establishes modelling of human body motion flow chart in the method for the present invention.
Fig. 3 is that self feed back gene expression programming establishes the expression established during modelling of human body motion in the method for the present invention
Formula tree graph.
Fig. 4 is expression tree map generalization process in Fig. 3.
Specific embodiment
Present invention will now be described in further detail with reference to the embodiments and the accompanying drawings, but embodiments of the present invention are unlimited
In this.
Embodiment
Present embodiment discloses a kind of Human bodys' response method based on self feed back gene expression programming, feature exists
In as shown in Figure 1, steps are as follows:
S1, human body behavior depth image is obtained, then extracts human body N pass respectively from human body behavior depth image
The corresponding three-dimensional time sequence data of node includes the three of the N number of artis of human body in an i.e. sample as a sample
Tie up time series data；Microsoft's somatosensory device Kinect combination second generation SDK and OpenCV are wherein used in the present embodiment
Computer vision library obtains human body behavior depth image, using TOF (Time Of Flight, flight time) technology, calculates micro-
The infrared light that infrared sensor emits in soft Kinect sensor reaches the phase difference of human body back reflection, obtains based on artis
Human body behavior depth image.The corresponding three-dimensional time sequence data of each artis is with the red of Microsoft somatosensory device Kinect
What the rectangular coordinate system in space that outer sensor is established by origin obtained, which includes x-axis, y-axis and z-axis,
Wherein positive direction of the x-axis is to be parallel to Microsoft's somatosensory device Kinect horizontal direction to the left, and positive direction of the y-axis is perpendicular to Microsoft's body-sensing
Equipment Kinect inclined direction is upward, and z-axis positive direction is Microsoft's somatosensory device Kinect direction.It is being got in this step
After human body behavior depth image, the human body in human body behavior depth image is split with background, skeleton is then converted to and chases after
Track system, to obtain multiple artis three-dimensional time sequence datas of human body, N is 25 in the present embodiment, i.e., from human body behavior
The corresponding three-dimensional time sequence data of 25 artis of human body is extracted in depth image respectively.
S2, sample is modeled using self feed back gene expression programming, gets the corresponding each artis fortune of sample
Dynamic sequence, to obtain various kinds based on the modelling of human body motion of artis；The wherein self feed back gene expression programming
(SGEP) the slotting string operation of TIS is added after intersection, mutation operation by gene expression programming (GEP) to construct to obtain, wherein TIS
Slotting string operation refers to being inserted into functor string on artis motion sequence head, for increasing artis motion sequence significance bit
Length increases.
S3, the gradient information of modelling of human body motion is extracted as the aspect of model；For each artis Ji(i=1,2...,
N), the modelling of human body motion remembered is zi=Fi(x,y,t)；The aspect of model that modelling of human body motion is extracted in step S3 is specific
Process is as follows:
K moment is equidistantly taken to be denoted as tj, j=1,2 ..., k；
Calculate the gradient information of each artisWithI=1,2 ..., N；J=1,2 ..., k；WhereinI-th of artis is respectively indicated in tjMoment corresponding x, y-coordinate value；
By the above-mentioned each artis gradient information being calculatedWithI=1,2 ..., N；J=1,
2 ..., k is z as modelling of human body motioni=FiThe aspect of model of (x, y, t).
S4, training sample set is got by step S1, and each training sample is got respectively by way of step S2
Then the modelling of human body motion of training sample extracts the model of the modelling of human body motion of each training sample by way of step S3
Feature.
S5, using the aspect of model of the modelling of human body motion of each training sample as the input of neural network, pass through setting mind
The number of the number of plies, neuron through network, initialization weighting parameter and selection gradient optimization algorithm, to neural network into
Training is obtained neural network model as final human body behavior classifier by row training；Detailed process is as follows:
S51, neural networks with single hidden layer frame is initially set up, the modelling of human body motion of each training sample will be obtained in step S4
The aspect of modelI=1,2 ..., N；J=1,2 ..., k are input to neural network, be provided with N ×
2k input neural network, setting hidden neuron number are N × 4k+1；The classification sum M for including according to human body behavior is set
Setting output neuron number is M, i.e., output vector is { c1,c2,…,cs,…,cM, wherein cs∈ R, 1,2 ..., s ..., M
Human body behavior generic is respectively indicated, and value interval is [0,1], works as cs=max { c1,c2,…,cMWhen, csIt is taken as 1, i.e.,
Neural network output vector is { 0,0 ..., 1 ..., 0 }, indicates that human body behavior belongs to s class behavior；
S52, setting neural network learning rate η are the random number between 0~1, and neural network transformation function f (x) is arranged
For Sigmoid function, i.e. f (x)=1/ (1+e-x), learning rules are continuous perceptron learning rules, utilize error back propagation
Algorithm trains neural network, obtains neural network model.
In this step, for the aspect of model of the modelling of human body motion of each training sample, as the input of neural network,
Neural network model is obtained by the training of ten folding cross-validation methods, as final human body behavior classifier.
S6, test sample is got by step S1, and test sample gets test specimens by way of step S2
This modelling of human body motion, then extracts the aspect of model of the modelling of human body motion of test sample by way of step S3.
S7, the aspect of model of the modelling of human body motion of test sample is input to the human body behavior point got in step S5
In class device, Human bodys' response result is obtained by human body behavior classifier.
In the present embodiment above-mentioned steps S2, as shown in Fig. 2, self feed back gene expression programming models sample, obtain
To various kinds, based on the modelling of human body motion of artis, detailed process is as follows:
S21, determine modelling of human body motion needed for functor collection and terminal symbol collection and terminal symbol number, and set circulation
Termination condition, subsequently into step S22；Wherein functor collection F in this step are as follows:
F={ S, C, T, I, R, N, P, Q, E, A, G, L, B, D, F, H }
Wherein it is as follows to represent meaning for each element in functor collection F:
S indicates that SIN function Sin, C indicate that cosine function cos, T indicate that tangent function tan, I indicate arcsin function
Asin, R indicate that inverse cosine function acos, N indicate that arctan function atan, P indicate that power function pow, Q indicate square root function
Sqrt, E indicate that exponential function exp, A that natural number is bottom indicate that ABS function abs, G indicate that sign function sign, L are indicated
Natural logrithm function ln, B indicate that denary logarithm function lg, D indicate that hyperbolic sine function sinh, F indicate hyperbolic cosine
Function cosh, H indicate hyperbolic tangent function tanh；
That is functor collection F are as follows: F=Sin, cos, tan, asin, acos, atan, pow, sqrt, exp, abs, sign,
Ln, lg, sinh, cosh, tanh }
Terminal symbol collection includes multidimensional terminating character and constant terminating character collection in this step；
Wherein multidimensional terminating character T are as follows:
T={ a, b, c, d, e, f, g, h, i }；
Wherein each element in multidimensional terminal symbol collection T respectively represents each independent variable in multidimensional variable, is examining human body fortune
Actual value corresponding to each variable is substituted into modelling of human body motion when movable model；Wherein constant terminal symbol collection C are as follows: C=
{ 1000, -1000 }.
Loop termination condition is usually arranged as one of the following conditions in this step:
(1) the evolution number for representing the population of the entire human motion motion model of construction has reached certain value；
(2) adaptive value for representing the population optimum individual of entire human motion motion model reaches preset value, or
Do not change in certain evolution number.
S22, the random artis motion sequence initial population of creation, are gene expression by artis motion sequence each in population
Formula tree, i.e. ET tree；Wherein initial population is initial modelling of human body motion, is made of each artis motion sequence, and joint
Point motion sequence is made of functor collection and terminal symbol collection again；It includes artis three-dimensional time sequence number that wherein terminal symbol, which is concentrated,
According to middle time t, artis three-dimensional time sequence data coordinate components etc.；
The process that each artis motion sequence expression formula is changed into gene expression tree in this step is as follows:
Using the first symbol of the artis motion sequence generated at random as the root node of expression formula；
The character of artis motion sequence is successively passed through into the method for breadth traversal as each element set functor in order
Or terminal symbol generate tree node, wherein the number of child node be the node answer containing parameter number (terminal symbol and without ginseng function
Required number of parameters is 0, and the typical value of child node is according to symbol sequence filling polishing in sequence.
Artis motion sequence in the present embodiment represents the gene in evolution algorithmic, for example i-th of artis is Ji,
At a time the coordinate data of the t artis is under some behaviorIt is closed by i-th
The three-dimensional time sequence data of node is available to be arrived, such asIt is one
A legal sequence, wherein Q representative function accords with evolution, and the part with underscore is gene tail portion, and the part without underscore is
Gene head.The corresponding expression tree of Fig. 3.Wherein Fig. 4 A to 4G show the generating process of the expression tree of Fig. 3.
S23, the adaptive value for calculating each artis motion sequence in population, then by relative error estimation model or absolutely
Error model assesses the adaptability of environment artis motion sequence each in population, qualified in assessment result
In the case of, a certain number of individuals are selected as male parent from population；And judge whether to meet loop termination condition, if full
Foot, then jump to S26, otherwise enter step S24；
Wherein relative error used in this step estimates model are as follows:
Wherein absolute error used in this step estimates model are as follows:
Wherein n is that the artis motion sequence extracted in population verifies the sum of sample, fiIndicate artis JiMotion sequence
Fitness value, R be a constant, (0, R) be fitness value control range；P(i,j′)It is to utilize artis JiCorresponding movement
Jth calculated by sequence ' a artis motion sequence verifying sample functional value, Yj′It is jth′A artis motion sequence is tested
Demonstrate,prove the true value of sample；
Using two norms (linear distance for referring to spatially two vector matrixs), shaped like:It is current to evaluate
The quality of model, whereinIt is predicted vector, Y is actual value vector.Work as f in this stepiWhen being substantially equal to n*R, i.e., relatively
In model of error estimate and absolute error estimation model | P(i,j′)-Yj′| it is substantially equal in the case of 0, then in this step
Judge that assessment result is eligible.
S24, the male parent selected according to step S23 carry out the intersection of gene expression programming to the individual in population, become
ETTHER-OR operation；
S25, the slotting string operation of TIS is carried out to population at individual, obtain next-generation population at individual, return to step S23；
S26, output model obtain function expression represented by optimal chromosome, the i.e. modelling of human body motion of artis.
Artis motion sequence is made of head and tail portion, and the length of artis motion sequence refers to the length of head and tail portion
The sum of degree, effective length is generally less than the length of artis motion sequence in artis motion sequence.If artis motion sequence
The length on head is h, and the length of tail portion is I, then the length h on artis motion sequence head and artis motion sequence tail portion
Relationship between length I can be with is defined as:
I=h (n-1)+1；
Wherein, in above formula, the parameter mesh number quantity of the most functor of institute's containing parameter mesh number in n representative function symbol set.
So regulation is to guarantee that the length of artis motion sequence is enough and sequence will not lead to illegal show due to length
As.
Wherein steps are as follows for the calculating of effective length in artis motion sequence:
1) first symbol that two pointers p, q are directed toward artis motion sequence is established；
2) what if pointer p was directed toward is functor, pointer q is moved back, and the digit moved back is parameter needed for the functor
Number；
3) pointer p moves back 1；
If 4) pointer p is not after pointer q, the 2) step is returned；
5) the effective length digit for calculating artis motion sequence, plays finger from first sign bit of artis motion sequence
Sign bit pointed by needle q is only ordered sequence, and length is the effective length of sequence；
The slotting string operation process of above-mentioned TIS includes the following: in the present embodiment
As artis motion sequence effective length m=1, i.e. the root position of artis motion sequence is a constant, this
When, the functor string since artis motion sequence root position to the position [L/4] variation to get at random；Wherein [L/4] table
Show L/4 round numbers part；Other characters on artis motion sequence head successively move back simultaneously, the character string beyond head length
It deletes；Wherein L is the length of artis motion sequence, and h is the length on artis motion sequence head.
Such as certain artis motion sequence are as follows: ab*/aba+*aababbbbbaba, wherein band horizontal line part is artis fortune
The tail portion of dynamic sequence, and with horizontal line part it is not the head of artis motion sequence, a, b, c respectively represent artis JiCoordinate
DataThe length L=21 of the artis motion sequence, head length h=10, tail length I=11 lead to
The sequence is crossed it can be concluded that its effective length m=1, meets m=1, string operation is inserted according to above-mentioned TIS, to the since root position
The position [L/4] variation is the functor string got at random, i.e., starts first position of artis motion sequence to the 5th position
The function string that variation is set to be randomly derived carries out TIS and inserts string operation if the function string being randomly derived is * -+QS
Afterwards, which becomes * -+QS ab*/aababbbbbaba, wherein first position of artis motion sequence
While beginning to the 5th position variation is * _ * SQ, the ab*/aba+* on head is moved back, beyond head length after moving back
Character string ba+* is deleted.Artis motion sequence * -+QS ab*/a after inserting string operation by TISababbbbbaba, can
To learn that the significance bit of the pass artis motion sequence has become 13.
When artis motion sequence imitates 1 < m of length < h, in the last one functor of artis motion sequence significance bit
It is inserted into the h-m functor strings got at random below, other characters behind the insertion point of artis motion sequence head successively move back,
Character string beyond head length is deleted；
Such as certain artis motion sequence are as follows:+b/ababa+*ababbbbaba, wherein band horizontal line part is artis fortune
The tail portion of dynamic sequence, and with horizontal line part it is not the head of artis motion sequence, the length L=of the artis motion sequence
21, head length h=11, tail length I=10, by the sequence it can be concluded that its effective length m=5, significance bit are+b/
Ab meets 1 < m < h, inserts string operation according to above-mentioned TIS, each random in the last one functor/rear insertion h-m=5 of significance bit
Obtained function string, if the function string being randomly derived is * _ * SQ, after carrying out the slotting string operation of TIS, artis fortune
Dynamic sequence becomes+b/*_*SQ abababbbbaba, wherein after * _ * SQ insertion, the ababa+* on head is moved back, and is exceeded after moving back
The character string aba+* of head length is deleted.The artis motion sequence after inserting string operation by TIS is+b/*_*SQ
abababbbbaba, can learn that the significance bit of the artis motion sequence becomes 14, the effective length the long more can guarantee kind
The uniformity of group's individual distribution, can preferably avoid algorithm from sucking locally optimal solution.
When artis motion sequence effective length be m >=h, do not do any operation；
It is grasped by gene expression programming in the prior art (GEP) in intersection, variation from the foregoing, it can be seen that the present embodiment is used
TIS is added after work insert string operation and construct to obtain self feed back gene expression programming and sample is modeled, in modeling process
The length that string operation increases artis motion sequence significance bit is inserted by TIS, so as to the uniformity to increase population at individual,
The globally optimal solution of problem can quickly be converged to by making it more, avoid falling into locally optimal solution.Therefore the present embodiment method uses
Self feed back gene expression programming models the accuracy and identification that can effectively improve Human bodys' response to sample
Speed.Meanwhile the present embodiment gradient information representative for specific behavior is special as the model in modelling of human body motion
Sign, the aspect of model extracted compared to the prior art using global or local motion feature as the aspect of model, the present embodiment
It can more accurately reflect human body behavior campaign, greatly reduce feature complexity, improve feature representativeness and identification is correct
Rate.
The above embodiment is a preferred embodiment of the present invention, but embodiments of the present invention are not by above-described embodiment
Limitation, other any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present invention,
It should be equivalent substitute mode, be included within the scope of the present invention.
Claims (9)
1. a kind of Human bodys' response method based on self feed back gene expression programming, which is characterized in that steps are as follows:
S1, human body behavior depth image is obtained, it is each then extracts the N number of artis of human body respectively from human body behavior depth image
Self-corresponding three-dimensional time sequence data, as a sample；
S2, sample is modeled using self feed back gene expression programming, gets the corresponding each artis motion sequence of sample
Column, to obtain various kinds based on the modelling of human body motion of artis；Wherein the self feed back gene expression programming is by gene
Expression formula is programmed in after intersection, mutation operation to add TIS and insert string operation and constructs to obtain, and wherein TIS inserts string operation and refers to
Functor string is inserted on artis motion sequence head, and the length for increasing artis motion sequence significance bit increases；
Self feed back gene expression programming models sample in the step S2, obtains the human motion mould based on artis
Detailed process is as follows for type:
S21, determine modelling of human body motion needed for functor collection and terminal symbol collection and terminal symbol number, and set loop termination
Condition；
S22, the random artis motion sequence initial population of creation, each artis motion sequence is expressed as gene expression tree；
Artis motion sequence is made of functor collection and terminal symbol collection, and it includes in artis three-dimensional time sequence data that terminal symbol, which is concentrated,
The coordinate components of time t and artis three-dimensional time sequence data；
S23, the adaptive value for calculating each artis motion sequence in population, then estimate model or absolute error by relative error
Model assesses the adaptability of environment artis motion sequence each in population, in the eligible situation of assessment result
Under, a certain number of individuals are selected as male parent from population；And judge whether to meet loop termination condition, if satisfied, then
S26 is jumped to, S24 is otherwise entered step；
S24, the male parent selected according to step S23 carry out gene expression programming to the artis motion sequence in population
Intersect, mutation operation；
S25, the slotting string operation of TIS is carried out to the artis motion sequence of population, obtain next-generation population at individual, return to step S23；
S26, output model obtain function expression represented by optimal chromosome, the i.e. modelling of human body motion of artis；
S3, the gradient information of modelling of human body motion is extracted as the aspect of model；
S4, training sample set is got by step S1, and each training sample gets each training by way of step S2
The modelling of human body motion of sample, the model that the modelling of human body motion of each training sample is then extracted by way of step S3 are special
Sign；
S5, each training sample modelling of human body motion input of the aspect of model as neural network, neural network is instructed
Practice, the neural network model that training is obtained, as final human body behavior classifier；
S6, test sample is got by step S1, and test sample gets test sample by way of step S2
Then modelling of human body motion extracts the aspect of model of the modelling of human body motion of test sample by way of step S3；
S7, the aspect of model of the modelling of human body motion of test sample is input to the human body behavior classifier got in step S5
In, Human bodys' response result is obtained by human body behavior classifier.
2. the Human bodys' response method according to claim 1 based on self feed back gene expression programming, feature exist
In, in the step S1 using combination second generation SDK and OpenCV the computer vision library Microsoft's somatosensory device Kinect acquisition
Human body behavior depth image；TOF (Time Of Flight, flight time) technology is wherein utilized, Microsoft Kinect sensing is calculated
The infrared light that infrared sensor emits in device reaches the phase difference of human body back reflection, obtains the human body behavior depth based on human body
Image；
The corresponding three-dimensional time sequence data of each artis is using the infrared sensor of Microsoft somatosensory device Kinect as origin
What the rectangular coordinate system in space established obtained, which includes x-axis, y-axis and z-axis, and wherein positive direction of the x-axis is
It is parallel to Microsoft's somatosensory device Kinect horizontal direction to the left, positive direction of the y-axis is to tilt perpendicular to Microsoft somatosensory device Kinect
Direction is upward, and z-axis positive direction is Microsoft's somatosensory device Kinect direction.
3. the Human bodys' response method according to claim 1 based on self feed back gene expression programming, feature exist
In N is 25 or more in the step S1.
4. the Human bodys' response method according to claim 1 based on self feed back gene expression programming, feature exist
In, in the step S1 after getting human body behavior depth image, by human body behavior depth image human body and background into
Row segmentation, is then converted to skeleton tracing system, to obtain multiple artis three-dimensional time sequence datas of human body.
5. the Human bodys' response method according to claim 1 based on self feed back gene expression programming, feature exist
In functor collection F in step S21 are as follows:
F={ S, C, T, I, R, N, P, Q, E, A, G, L, B, D, F, H }
Wherein it is as follows to represent meaning for each element in functor collection F:
S indicates that SIN function Sin, C indicate that cosine function cos, T indicate that tangent function tan, I indicate arcsin function asin, R
Indicate that inverse cosine function acos, N indicate that arctan function atan, P indicate that power function pow, Q indicate square root function sqrt, E
Indicate that exponential function exp, A that natural number is bottom indicate that ABS function abs, G indicate that sign function sign, L indicate nature pair
Number function ln, B indicate that denary logarithm function lg, D indicate that hyperbolic sine function sinh, F indicate hyperbolic cosine function
Cosh, H indicate hyperbolic tangent function tanh；
Terminal symbol collection described in step S21 includes multidimensional terminating character and constant terminating character collection；
Wherein multidimensional terminating character T are as follows:
T={ a, b, c, d, e, f, g, h, i }；
Wherein each element in multidimensional terminal symbol collection T respectively represents each independent variable in multidimensional variable, is examining human motion mould
Actual value corresponding to each variable is substituted into modelling of human body motion when type；
Wherein constant terminal symbol collection C are as follows: C={ 1000, -1000 }.
6. according to claim 1 or 5 based on the Human bodys' response method of self feed back gene expression programming, feature
It is, the process that TIS inserts string operation is as follows:
As artis motion sequence effective length m=1, i.e. the root position of artis motion sequence is a constant, at this point, from
The functor string that artis motion sequence root position starts to the position [L/4] to make a variation to get at random；Wherein [L/4] indicates L/4
Round numbers part, L are the length of artis motion sequence；Other characters on artis motion sequence head successively move back simultaneously,
Character string beyond head length is deleted；Wherein L is the length of artis motion sequence, and h is artis motion sequence head
Length；
When artis motion sequence imitates 1 < m of length < h, behind the last one functor of artis motion sequence significance bit
The h-m functor strings got at random are inserted into, other characters behind the insertion point of artis motion sequence head successively move back, exceed
The character string of head length is deleted.
7. the Human bodys' response method according to claim 1 based on self feed back gene expression programming, feature exist
In in the step S3, for each artis J of samplei(i=1,2..., N), N are the sum of artis in sample, are remembered
The modelling of human body motion arrived is zi=Fi(x,y,t)；The aspect of model detailed process of modelling of human body motion is extracted such as in step S3
Under:
K moment is equidistantly taken to be denoted as tj, j=1,2 ..., k；
Calculate the gradient information of each artisWithWhereinPoint
Do not indicate i-th of artis in tjMoment corresponding x, y-coordinate value；
By the above-mentioned each artis gradient information being calculatedWithAs
Modelling of human body motion is zi=FiThe aspect of model of (x, y, t).
8. the Human bodys' response method according to claim 7 based on self feed back gene expression programming, feature exist
In the aspect of model training for passing through the modelling of human body motion of each training sample in the step S5 obtains the tool of neural network model
Body process is as follows:
S51, neural networks with single hidden layer frame is initially set up, the mould of the modelling of human body motion of each training sample will be obtained in step S4
Type featureAs the input training sample set of neural network, wherein
N × 2k input neural network is set, and setting hidden neuron number is N × 4k+1；The classification for including according to human body behavior
Number M setting output neuron number is M, i.e., output vector is { c1,c2,…,cs,…,cM, wherein cs∈ R, and value interval
For [0,1], work as cs=max { c1,c2,…,cMWhen, csIt is taken as 1, i.e. neural network output vector is { 0,0 ..., 1 ..., 0 }, table
Body behavior of leting others have a look at belongs to s class behavior；
S52, setting neural network learning rate η are the random number between 0~1, and neural network transformation function f (x) is set as
Sigmoid function, learning rules are continuous perceptron learning rules, using error backpropagation algorithm training neural network, are obtained
To neural network model.
9. the Human bodys' response method according to claim 1 based on self feed back gene expression programming, feature exist
In, in the step S5, it is directed to the aspect of model of the modelling of human body motion of each training sample, as the input of neural network,
Neural network model is obtained by the training of ten folding cross-validation methods, as final human body behavior classifier.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710059525.6A CN106909891B (en) | 2017-01-24 | 2017-01-24 | A kind of Human bodys' response method based on self feed back gene expression programming |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710059525.6A CN106909891B (en) | 2017-01-24 | 2017-01-24 | A kind of Human bodys' response method based on self feed back gene expression programming |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106909891A CN106909891A (en) | 2017-06-30 |
| CN106909891B true CN106909891B (en) | 2019-11-15 |
Family
ID=59207616
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710059525.6A Active CN106909891B (en) | 2017-01-24 | 2017-01-24 | A kind of Human bodys' response method based on self feed back gene expression programming |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106909891B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107480692A (en) * | 2017-07-06 | 2017-12-15 | 浙江工业大学 | A kind of Human bodys' response method based on principal component analysis |
| CN110325965B (en) * | 2018-01-25 | 2021-01-01 | 腾讯科技（深圳）有限公司 | Object processing method, device and storage medium in virtual scene |
| CN108805965B (en) * | 2018-05-18 | 2022-03-15 | 合肥工业大学 | Human physical motion generation method based on multi-objective evolution |
| CN111403039A (en) * | 2020-03-19 | 2020-07-10 | 中国科学院深圳先进技术研究院 | Dynamic balance evaluation method, device, equipment and medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102764167A (en) * | 2012-06-12 | 2012-11-07 | 天津大学 | Myoelectric prosthesis control source lead optimization method based on correlation coefficients |
| CN104850846A (en) * | 2015-06-02 | 2015-08-19 | 深圳大学 | Human behavior recognition method and human behavior recognition system based on depth neural network |
| CN105615890A (en) * | 2015-12-24 | 2016-06-01 | 西安交通大学 | Angle and myoelectricity continuous decoding method for human body lower limb walking joint |
-
2017
- 2017-01-24 CN CN201710059525.6A patent/CN106909891B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102764167A (en) * | 2012-06-12 | 2012-11-07 | 天津大学 | Myoelectric prosthesis control source lead optimization method based on correlation coefficients |
| CN104850846A (en) * | 2015-06-02 | 2015-08-19 | 深圳大学 | Human behavior recognition method and human behavior recognition system based on depth neural network |
| CN105615890A (en) * | 2015-12-24 | 2016-06-01 | 西安交通大学 | Angle and myoelectricity continuous decoding method for human body lower limb walking joint |
Non-Patent Citations (2)
| Title |
|---|
| NCRS-GEP编程算法及其在统计建模中的应用;李康顺等;《武汉理工大学学报》;20141231;第122-129页 * |
| 基因表达式编程的一种新遗传算子;莫海芳等;《计算机工程与应用》;20111231;第23-29页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106909891A (en) | 2017-06-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Duan et al. | A survey of embodied ai: From simulators to research tasks | |
| Zhong et al. | The application of virtual reality technology in the digital preservation of cultural heritage | |
| CN106909891B (en) | A kind of Human bodys' response method based on self feed back gene expression programming | |
| CN108334830A (en) | A kind of scene recognition method based on target semanteme and appearance of depth Fusion Features | |
| CN107423398A (en) | Exchange method, device, storage medium and computer equipment | |
| CN109508686A (en) | A kind of Human bodys' response method based on the study of stratification proper subspace | |
| CN107066979A (en) | A kind of human motion recognition method based on depth information and various dimensions convolutional neural networks | |
| CN109086664A (en) | A kind of polymorphic gesture identification method of sound state fusion | |
| Lv et al. | Generic content-based retrieval of marker-based motion capture data | |
| Zhu et al. | Unpaired image captioning by image-level weakly-supervised visual concept recognition | |
| Akhter | Automated posture analysis of gait event detection via a hierarchical optimization algorithm and pseudo 2D stick-model | |
| CN112785721A (en) | LeapMotion gesture recognition-based VR electrical and electronic experiment system design method | |
| Zhang et al. | An efficient method for automatic generation of labanotation based on bi-directional lstm | |
| CN114398485B (en) | Expert portrait construction method and device based on multi-view fusion | |
| Zhao et al. | Cross-modal knowledge distillation for fine-grained one-shot classification | |
| Wang | [Retracted] Long Jump Action Recognition Based on Deep Convolutional Neural Network | |
| CN116954365A (en) | Virtual human limb movement interaction system based on monocular video drive | |
| CN116958337A (en) | Virtual object animation generation method and device, electronic equipment and readable storage medium | |
| Niu et al. | [Retracted] Analysis of the 3D Application Evaluation System of Landscape Based on Hybrid Cooperation of VR and AI | |
| Chen et al. | Movement Evaluation Algorithm‐Based Form Tracking Technology and Optimal Control of Limbs for Dancers | |
| Han et al. | Computer Aided Dance Art Action Identification and Classification Algorithm Based on Machine Learning | |
| Zhang | Research on Dance Art Teaching System Based on Data Mining and Machine Learning | |
| Zeng | Gesture Recognition Technology of VR Piano Playing Teaching Game based on Hidden Markov Model | |
| Liang et al. | Interactive Experience Design of Traditional Dance in New Media Era Based on Action Detection | |
| Li et al. | Characteristic Behavior of Human Multi-Joint Spatial Trajectory in Slalom Skiing |
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 |