CN101692284B - Three-dimensional human body motion tracking method based on quantum immune clone algorithm - Google Patents

Abstract

The invention discloses a three-dimensional human body motion tracking method based on quantum immune clone algorithm. The tracking process comprises: detecting articulation points of the human body in a two dimension picture in a unary picture sequence with human body gesture existing, using a kalman filter to predict the articulation points not detected, and simultaneously establishing a three-dimensional human body model; applying quantum immune clone algorithm in human body tracking, initializing species group, setting initial parameters of human body motion, then carrying out clone operation, performing operation of immunizing genes and immunoselection to generate a new species group; substituting the state parameters of the new species group in the three-dimensional human body model, using likelihood function to calculate the distance affinity, retaining the optimal solution, performing multiple times of substitution and calculation to obtain the ideal human body motion parameters to restore the three-dimensional human body gesture. The invention has the advantages of lower calculation cost and capable of obtaining three-dimensional human body gesture quickly, and is applicable to tracking of human body motion gesture in scene monitoring, motion analysis and health assessment.

Description

Three-dimensional human body motion tracking method based on quantum immune clone algorithm

Technical field

The invention belongs to technical field of image processing, relate to human body tracing method, can be used for the human motion in image or the image sequence is followed the tracks of, realize scene monitoring, virtual reality and game making and sports analysis and health assessment.

Background technology

On the basis of Flame Image Process; Can understand image in order to make computing machine; Usually people introduce computer vision; Its fundamental purpose is the means that a kind of quantification is provided for vision, through the anthropoid visual performance of artificial intelligence system type of realization, utilizes monocular, many orders camera to carry out structure or motion feature of vision guided navigation, scene monitoring, the various types of objects of extraction, people and animal etc.Yet the visual capacity that lets computing machine have the people is a very thing of difficulty; Because human living is at the geometric space of a three-dimensional; Vision sensor can only obtain the projection information of three-dimensional world; Be two dimensional image, this process causes great deal of information to be lost, and makes Computer Vision Task become complicated more through the dynamic scene information that vision sensor obtained that moves.

In research,, can be divided into based on the joint motions analysis of monumented point and the motion analysis of no marks point according to having or not the joint monumented point based on the human motion analysis of vision.Wherein in motion analysis based on monumented point; The joint monumented point is accurately detected by stereoscopic vision, can obtain the articulation point three-dimensional position easily, thereby carries out the three-dimensional reconstruction of human skeleton; Commercial relatively exemplary systems is a Motion Capture equipment, and price comparison is expensive.In the motion analysis of no marks point, can be divided into again based on the motion analysis of many orders camera with based on the human motion analysis of monocular camera.

In the human motion analysis based on monocular image, a lot of researchs in the past all are to the first two field picture manual markings initial point, realize initial tracking or estimation.Forefathers also have a lot of employing optimized Algorithm to carry out the precedent of human body tracking, and traditional optimal algorithms such as evolution algorithm and genetic algorithm have been applied to human body tracking.The dimension of human body tracking is higher and the fitness curved surface is complicated, and it is unusual difficult that optimization problem becomes, and a lot of traditional optimization all are difficult to be competent at.But evolution algorithm can address these problems preferably; Be that mainly it does not need the domain knowledge of problem basically; And the shape to CWinInetConnection type and search volume has no restriction, but the subject matter that perplexs evolution algorithm is that algorithm possibly be absorbed in Local Extremum, and the dimension of human body tracking is higher and the fitness curved surface is complicated; In case be absorbed in Local Extremum, just can not well recover human body attitude.

Summary of the invention

The objective of the invention is to overcome the shortcoming of above-mentioned prior art, propose a kind of three-dimensional human body motion tracking method,, realize the human motion tracking with the accuracy of further raising 3 d human motion attitude based on quantum immune clone algorithm.

The technical scheme that realizes the object of the invention is: utilize quantum immune clone algorithm to have the efficient advantage of seeking globally optimal solution in high bit space; The crucial articulation point of human body that obtains with the crucial articulation point automatic testing method of two-dimension human body is the basis; Through the crucial articulation point of two dimension and tripleplane's point apart from similarity function; Obtain overall optimum condition parameter, finally recover robust, stable 3 D human body attitude.Detailed process is following:

(1) in the monocular image sequence that has human body attitude to exist, detects the crucial articulation point of body surface in the two dimensional image; And use classical kalman wave filter, and the crucial articulation point of detected two-dimension human body to be blocked a little and the omission point prediction, the motion that makes the crucial articulation point of two-dimension human body is more rationally with stable;

(2) based on the crucial artis of the two-dimension human body after detection and the prediction processing, set up a virtual human body three-dimensional skeleton pattern, make it in tracing process, to realize dynamic stance adjustment and coupling;

(3) quantum immune clone algorithm is introduced in the human motion tracking, at first the initialization population is provided with the human motion initial parameter, carries out clone operations then, with the search volume of increase human body attitude parameter to be estimated;

(4) to carrying out the population after the clone operations, carry out immunogene operation and Immune Selection, generate new population;

(5) with the state parameter of new population, the substitution three-dimensional (3 D) manikin produces crucial articulation point three-dimensional coordinate P _i=(P _Ix, P _Iy, P _Iz), and the key point that this key point three-dimensional coordinate is projected in the plane of delineation is designated as p _i=(p _Ix, p _Iy), the crucial articulation point of detected two-dimension human body is designated as q _i=(q _Ix, q _Iy), use these two key points to construct distance function to be: $G\left(X\right)=\underset{i=1}{\overset{15}{\mathrm{\Σ}}}{\left|\right|p_i-q_i\left|\right|}_2;$

(6) according to distance function G (X) structure similarity function be: X=minG (X), utilize this similarity function calculate two-dimensional detection point and tripleplane point apart from affinity, and keep its optimum solution, if optimum solution satisfies the end condition of setting, termination calculating; Otherwise, get back to step (3), for calculating desirable human body sport parameter, recover the 3 D human body attitude through too much.

The present invention has the following advantages compared with prior art:

1, the present invention has set up the reasonable three-dimensional manikin owing to adopt the crucial articulation point of automatic human body two dimension, so three-dimensional model can be realized dynamic stance adjustment in tracing process, can efficiently mate with original turntable.

2, the present invention because adopt two-dimentional crucial artis and the crucial artis of tripleplane apart from similarity function, make human body tracking can with the good combination of quantum-inspired immune clone operator, obtain the body state parameter fast.

3, the present invention in search procedure, not only make the search volume of human parameters become big, and calculation cost is lower owing to adopt the chaos variation method.

Description of drawings

Fig. 1 is a schematic flow sheet of the present invention;

Fig. 2 is the human skeleton illustraton of model that uses among the present invention;

Fig. 3 is the 3 D human body attitude figure that the present invention realizes;

Fig. 4 be three-dimensional framework spot projection and two-dimentional corresponding check point among the present invention distance and the mean value curve map.

Embodiment

The present invention has proposed a kind of new method of following the tracks of based on the 3 d human motion of quantum immune clone algorithm from the continuity of image sequence, promptly utilizes the ability of the superpower search globally optimal solution of quantum immune clone algorithm, carries out three-dimensional human motion and follows the tracks of.The present invention is on some the improved bases to the quantum evolution algorithm; Established relevant optimized parameter and statistical nature through experiment; Utilize quantum renewal operator and quantum crossover operator that state parameter is induced then, and tracking results and corresponding two-dimensional detection point are compared.Describe in detail like Fig. 1, Fig. 3 and Fig. 4.

With reference to Fig. 1, concrete implementation procedure of the present invention is following:

Step 1 detects the crucial articulation point of human body in the two dimensional image, and uses the kalman wave filter that the omission point is estimated.

The image sequence that the present invention uses is from the video sequence that has human motion to exist, to obtain; 660 frames altogether, because the present invention does not study dividing method, the human motion of this video sequence is from not blocking and being blocked; And owing to the rotation for trunk detects the very big problem of existence; Human motion among the present invention does not embody the rotation of trunk, simultaneously in order to reduce the manikin dimension, supposes that here the articulation point on the trunk does not have degree of freedom usually.

From people's silhouette information, finding the crucial articulation point of human body, is very challenging must a job, mainly chooses no too many theory, and the three-dimensional position of crucial articulation point and unlike capturing movement equipment that works accurate the acquisition by experience.Before detecting articulation point, at first make following hypothesis: 1) suppose front cross frame image nobody side shadow from blocking or being blocked, this hypothesis can guarantee the articulation point that later having blocked, and can predict out according to certain rule; 2) suppose that the center of silhouette and the root node projection of three-dimensional framework hardly differ, this hypothesis can guarantee obtaining root node under the situation arbitrarily.The human body silhouette of prospect splits from image sequence, then with the method for digital morphological, is refined into people's silhouette zone the image framework of single pixel.Its survey process is following:

(1.1) head point detects: in the human body silhouette, the shape of head generally can not be blocked, and characteristic is also the most stable; Be generally circle; Therefore, the three-dimensional spherical shell template of head in the people detection voxel data before using for reference provides 2 dimension concentric circles templates that detect the sequence of video images head zone.Suppose concentrically ringed in the round heart be R ₁, the cylindrical center of circle is R ₂, the center of circle of concentric circles template is along the skeleton S={s of silhouette _iSearch, and i=1,2 ..., N, N are the number of present frame silhouette zone skeleton point, calculate the profile C={c that drops between interior circle and the cylindrical _jCount, j=1,2 ..., M is the number of present frame point, when falling into counting the most for a long time of inside and outside circle, s _iBe head point, more than 2 times of radius of circle in general exradius will be got;

(1.2) root node detects: suppose that human body silhouette point set is remembered and make A={a _k, the number of point is L, then center position does $c_a=\frac{1}{L}\underset{1}{\overset{L}{\mathrm{\Σ}}}{a}_k,$ The experience detection method of root node is to get skeleton point S={s _iOn from human body silhouette central point nearest a bit;

(1.3) shoulder joint and hip joint are detected: suppose that head point and root node finish after testing; With right and left shoulders joint and hip joint and the binding of whole trunk; Make it in testing process, keep relative position constant; Good the corresponding point matching on the head subpoint of three-dimensional framework and root node subpoint and the image, make the 3-D shoulder joint spot projection regional to people's silhouette;

(1.4) four limbs end-point detection: on the silhouette skeleton line, utilization end points search method, the end points of search skeleton line; Generally speaking, in the absence of burr, be easy to detect the position of four limbs usually; But this method will lose efficacy under the very big situation of noise, search the zone so provide a taboo, promptly at head part's point and root node; Delimit a rectangular area between the two shoulder joint nodes, the end points that searches in this zone will be dropped, and will greatly improve the detection accuracy rate of four limbs point like this;

(1.5) knee joint and elbow joint detect: at first the point set of silhouette skeleton line above root node is defined as above the waist; Remaining point is defined as the lower part of the body; Cross the Lower Half central point and make the parallel lines of bipod; The intersection point of these parallel lines and skeleton line is the knee joint point, elbow joint detect be seek on the skeleton line a bit, make the difference minimum of this point distance with shoulder in one's hands;

When human motion takes place to block in vain or because the incompleteness of detection algorithm when causing some key point correctly not detect, just must implement to block forecasting mechanism a little, blocks point prediction with the classic card Thalmann filter.

Below the basic theories of Kalman filtering is briefly described, is considered discrete time linear random dynamical system:

x _k+1＝F _kx _k+Γ _kw _k

z _k＝H _kx _k+v _k

Wherein k ∈ N is a time index, x _k∈ R ⁿBe k system state vector constantly, F _kBe the system state transition matrix, w _kThen be process evolution noise, Γ _kBe noise matrix, z _k∈ R ^mBe that k is vectorial to the measurement of system state constantly, H _kFor measuring matrix, v _kIt is measurement noise.

Just can estimate to obtain the crucial articulation point that lacks between two frames through simple extrapolation, simultaneously the noise that extracts in skeleton and the testing process carried out filtering, make that the position of the crucial articulation point of all detected two dimensions is more reasonable.

Step 2 according to the crucial articulation point of the two-dimension human body after detection and the prediction processing, is set up a virtual human body three-dimensional skeleton pattern.

(2.1) the human skeleton model that uses among the present invention is with the kinematic chain that connects the rigid body connection, and uses the constraint of some priori, so that generate the torso portion of health, and root node and crotch, neck joint and shoulder do not have freedom of motion, just along with the trunk motion;

(2.2) in order to reduce the search volume size as far as possible, in the model that adopts usually, human skeleton has 21 degree of freedom, comprises the translation of 3 integral body, 6 hip joints, 2 knee joints, 6 shoulder joint, 2 elbow joints and 2 neck joints;

(2.3) set up three-dimensional (3 D) manikin; At first roughly confirm the ratio that the partes corporis humani divides according to people's anatomy ratio, the center of manikin coordinate system is a web joint, is root node; Each articulation point all can have a local coordinate system, and its child node coordinate is used matrix M _{P, C}Move to the father node coordinate:

M _P，C＝T(t _x，t _y，t _z)·R _z(γ)·R _y(β)·R _x(α)

T (t wherein _x, t _y, t _z) represent that the child node coordinate is along x axle translation t _x, along y axle translation t _y, along z axle translation t _z, R _x(α) expression child node coordinate is around x axle rotation alpha degree, R _y(β) expression child node coordinate is around y axle rotation β degree, R _z(γ) expression child node coordinate is around z axle rotation γ degree, and their matrix representation formula is:

$T(t_x,t_y,t_z)=\left[\begin{array}{cccc}0& 0& 0& t_x\\ 0& 0& 0& t_y\\ 0& 0& 0& {\mathrm{<figure-callout\; id="0"\; label="t\; z"\; filenames="H2009100234183E00011.png,H2009100234183E00022.png"\; state="\{\{state\}\}">t</figure-callout>}}_z\\ 0& 0& 0& 1\end{array}\right]$

$R_x\left(\mathrm{\&alpha;}\right)=\left[\begin{array}{cccc}1& 0& 0& 0\\ 0& \cos \left(\mathrm{\&alpha;}\right)& -\sin \left(\mathrm{\&alpha;}\right)& 0\\ 0& \sin \left(\mathrm{\&alpha;}\right)& \cos \left(\mathrm{\&alpha;}\right)& 0\\ 0& 0& 0& 1\end{array}\right]$

$R_y\left(\mathrm{\&beta;}\right)=\left[\begin{array}{cccc}\cos \left(\mathrm{\&beta;}\right)& 0& \sin \left(\mathrm{\&beta;}\right)& 0\\ 0& 1& 0& 0\\ -\sin \left(\mathrm{\&beta;}\right)& 0& \cos \left(\mathrm{\&beta;}\right)& 0\\ 0& 0& 0& 1\end{array}\right]$

$R_z\left(\mathrm{\&gamma;}\right)=\left[\begin{array}{cccc}\cos \left(\mathrm{\&gamma;}\right)& -\sin \left(\mathrm{\&gamma;}\right)& 0& 0\\ \sin \left(\mathrm{\&gamma;}\right)& \cos \left(\mathrm{\&gamma;}\right)& 0& 0\\ 0& 0& 1& 0\\ 0& 0& 0& 1\end{array}\right]$

For any 1 p=(x, y, z, 1) in the child node local coordinate system ^T, the position in the father node local coordinate system is P=(X, Y, Z, 1) ^T, the mutual transforming relationship between them is:

$\left[\begin{array}{c}X\\ Y\\ Z\\ 1\end{array}\right]=\left[\begin{array}{cccc}0& 0& 0& t_x\\ 0& 0& 0& t_y\\ 0& 0& 0& {\mathrm{<figure-callout\; id="0"\; label="t\; z"\; filenames="H2009100234183E00011.png,H2009100234183E00022.png"\; state="\{\{state\}\}">t</figure-callout>}}_z\\ 0& 0& 0& 1\end{array}\right]\left[\begin{array}{cccc}\cos \mathrm{\&gamma;}& -\sin \left(\mathrm{\&gamma;}\right)& 0& 0\\ \sin \left(\mathrm{\&gamma;}\right)& \cos \left(\mathrm{\&gamma;}\right)& 0& 0\\ 0& 0& 1& 0\\ 0& 0& 0& 1\end{array}\right]\left[\begin{array}{cccc}\cos \left(\mathrm{\&beta;}\right)& 0& \sin \left(\mathrm{\&beta;}\right)& 0\\ 0& 1& 0& 0\\ -\sin \left(\mathrm{\&beta;}\right)& 0& \cos \left(\mathrm{\&beta;}\right)& 0\\ 0& 0& 0& 1\end{array}\right]$

$\left[\begin{array}{cccc}1& 0& 0& 0\\ 0& \cos \left(\mathrm{\&alpha;}\right)& -\sin \left(\mathrm{\&alpha;}\right)& 0\\ 0& \sin \left(\mathrm{\&alpha;}\right)& \cos \left(\mathrm{\&alpha;}\right)& 0\\ 0& 0& 0& 1\end{array}\right]\left[\begin{array}{c}x\\ \mathrm{<figure-callout\; id="1"\; label="y\; z"\; filenames="H2009100234183E00011.png,H2009100234183E00022.png"\; state="\{\{state\}\}">y</figure-callout>}& \\ z\\ 1\end{array}\right]$

Can just form a kinematic chain through continuous several conversion operation like this, give expression to the 3D coordinate of each point and the relation of each rotation angle, just can obtain the position of each point in world coordinate system, for example, each point coordinate of left lower extremity is:

P _LH＝T(t _x，t _y，t _z)R _0zR _0yR _0xT _LHR·[0，0，0，1] ^T

P _LK＝T(t _x，t _y，t _z)R _0zR _0yR _0xT _LHRR _1zR _1yR _1xT _LKH[0，0，0，1] ^T

P _LA＝T(t _x，t _y，t _z)R _0zR _0yR _0xT _LHRR _1zR _1yR _1xT _LKHR _2zR _2yR _2xT _LAK·[0，0，0，1] ^T

P _LF＝T(t _x，t _y，t _z)R _0zR _0yR _0xT _LHRR _1zR _1yR _1xT _LKH

R _2zR _2yR _2xT _LAKR _3zR _3yR _3xT _LFK·[0，0，0，1] ^T

Wherein, P _LH, P _LK, P _LAAnd P _LFRepresent left stern point respectively, left knee point, left ankle point and left foot point, T (t _x, t _y, t _z) be the displacement that root node arrives the world coordinate system initial point, R _IxBe the rotation matrix of i node around x, R _IyBe the rotation matrix of i node around y, R _IzBe the rotation matrix of i node around z, subscript 0x, 0y and 0z represent the integral body rotation of trunk; T _LHR, T _LKH, T _LAKAnd T _LFARepresent left buttocks to arrive the initial translation of left knee and the initial translation that left foot portion arrives left ankle to the initial translation of left stern node, left ankle respectively to the initial translation of root node, left knee; The method of asking of upper limbs point is similar to the lower limb point; Here list no longer one by one, obtain the coordinate of each articulation point in world coordinate system after, be easy to release the projection value of each points in image coordinate system according to various projection models; The three-dimensional (3 D) manikin of global design, as shown in Figure 2.

Step 3, the initialization population is provided with the human motion initial parameter, and carries out clone operations.

(3.1), be provided with and totally wait to estimate human parameters collection X and be according to the degree of freedom of each crucial articulation point:

X＝{x，y，z，θ ₁，θ ₂，…，θ ₁₈}

X wherein, y, z represent integral translation respectively along the translation distance of x, y, z axle, θ ₁, θ ₂..., θ ₁₈The anglec of rotation of representing the crucial articulation point of human body to be estimated; Because the continuity of human motion, the human body limb angle changing is generally little between per two two field pictures, according to the statistics of the continuous frame of this image sequence, the anglec of rotation θ of the crucial articulation point of human body to be estimated between two two field pictures is set ₁, θ ₂..., θ ₁₈Scope be [15 ,+15] degree; Because motion translation is also very little, and integral translation x is set here, y, the scope of z is [15 ,+15] pixel.

Each two field picture only need obtain the difference DELTA X of parameter between two frames in estimation procedure:

ΔX＝{Δx，Δy，Δz，Δθ ₁，Δθ ₂，…，Δθ ₁₈}

Δ x wherein, Δ y, Δ z represent integral translation between two frames respectively along the difference of x, y, the translation of z axle, and Δ θ ₁, Δ θ ₂..., Δ θ ₁₈The difference of representing 3 D human body state angle to be estimated between two frames.

According to the setting of human parameters above-mentioned to be estimated, initial population A (0) be set be:

Wherein each column element value is that from+15 to-15 even assignment are given this k+1 numeral; 21 parameter values to be estimated of each line display; K representes the size of the state value that each element to be estimated can be chosen; Each finite element correspondence is provided with an equal probability 1/k+1;

(3.2) carry out clone operations, generation clone population A ' (k).

According to initialization body state parameter population, in order to expand the search volume, initial population is carried out clone operations here, be expressed as T _c ^C:

$T_c^C\left(A\right)={\left[\begin{array}{cccc}{T}_c^C\left(a_1\right)& T_c^C\left(a_2\right),& ...,& T_c^C\left(a_N\right)\end{array}\right]}^T$

Wherein, $T_c^C\left(a_i\right)=I_i\&times;a_i$ (i=1,2 ..., N), N representes the number of elements of each row of A (k), I _iFor element value 1 q _iDimension row vector, q _iBe antibody a _iScale behind the clone, a _iThe capable all elements of i among the expression A (k), q _iBe taken as:

$q_i\left(k\right)=\mathrm{Int}(N_c\&times;\frac{D\left(a_i\left(k\right)\right)}{\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}D\left(a_j\left(k\right)\right)}),i=\mathrm{1,2},...,N$

Int (x) expression is more than or equal to the smallest positive integral of x, and N _cBe with the clone after the relevant setting value of scale and satisfy N _c＞N, D (a _i(k)) size of each element probability among the expression A (k),

After the clone, initial population has formed a plurality of search populations according to each element probability, and these populations totally are defined as A ' (k):

A′＝{A，A′ ₁，A′ ₂，...，A′ _N}

Wherein, A ' _iThe new population that expression generates after clone operations, N represent to clone the scale of back population.

Step 4 to carrying out the population after the clone operations, is carried out immunogene operation and Immune Selection, generates new population.

(4.1) sub-group behind the clone is carried out the immunogene operation, generate immune population A " (k);

Through after the clone operations; The search volume of body state parameter has been expanded several times according to human joint points apart from affinity; For the speed of convergence that prevents the body state function slow and too early sink into locally optimal solution, carry out the immunogene operation, it mainly comprises clone's reorganization and clonal vaviation.Use quantum to upgrade operator and realize clone's reorganization, use the quantum crossover operator to realize clonal vaviation; This quantum upgrades operator, adopts the accelerating convergence of quantum rotation door, and the chaos mutation operation is to prevent precocity.

Quantum upgrades the quantum rotation door in the operator, is expressed as U (θ):

$U\left(\mathrm{\&theta;}\right)=\left[\begin{array}{cc}\cos \left(\text{\&theta;}\right)& -\sin \left(\mathrm{\&theta;}\right)\\ \sin \left(\mathrm{\&theta;}\right)& \cos \left(\mathrm{\&theta;}\right)\end{array}\right]$

Wherein, θ is the angle of rotation renewal.

According to A ' (k) in the probability of element, judge the direction and the angle of rotation, realize upgrading operation through continuous renewal and rotation are final.

Quantum upgrades chaos variation in the operator in order to prevent that antibody from developing towards an aspect, and is absorbed in local optimum, defines the chaos variation here, the increase local search ability.For the antibody that upgrades through above-mentioned quantum rotation, with the variation Probability p _mSelect one or several positions at random, implement operation as follows:

$a_{\mathrm{ji}}^{\&prime;}=\left\{\begin{array}{ccc}{a}_{\mathrm{ji}}+(1-a_{\mathrm{ji}})\&times;{\mathrm{Logistic}}_{\mathrm{ji}}\left(k\right)& \mathrm{if}& \mathrm{rand}>0.5\\ a_{\mathrm{ji}}-a_{\mathrm{ji}}\&times;{\mathrm{Logistic}}_{\mathrm{ji}}\left(k\right)& \mathrm{if}& \mathrm{rand}\&le;0.5\end{array}\right.$

A wherein _JiBe the probable value of i quantum bit in preceding j the quantum antibody that makes a variation, a ' _JiProbable value for i quantum bit in j the quantum antibody in back that makes a variation.

The variation yardstick of chaos variation is made up of two parts, that is:

I.1-a _JiOr a _JiCan guarantee to make a variation back antibody variable automatically still between [0,1].

II.Logistic _Ji(k) be k sequential value of Logistic mapping, wherein the sequence length of Logistic mapping is q _i-1, promptly keep original antibody information, to implementing this mutation operation in the sub-group, the Logistic mapping is described below:

x _n+1＝μx _n(1-x _n) (n＝0，1，2，...)

Wherein choose μ=4, x _n=a _Ji

(4.2) " (k) carry out Immune Selection, form new population to population A.

After the immunogene operation, carry out Immune Selection, be expressed as T _S ^C, it is from filial generation separately after the operation of clone's immunogene and respective parent, to select outstanding antibody, thereby forms new population.That is:

$A(k+1)=T_s^C(A\left(k\right)\&cup;A^{\&prime;\&prime;}\left(k\right))$

According to the affinity degree, $\&ForAll;i=\mathrm{1,2},...N,$ If have

b _i(k)＝{a″ _ij(k)|maxD(a″ _ij)j＝1，2，...，q _i-1}

Make

D(a _i(k))＜D(b _i(k))，i＝1，2，...，N

Then

a _i(k+1)＝b _i(k)，i＝1，2，...，N

A wherein " _IjExpression is operated the probability of i quantum antibody j position, back through renewal, D (a " _Ij) expression a " _IjAffinity, D (a _iAnd D (b (k)) _i(k)) expression a _i(k) affinity and b _i(k) affinity, a _i(k+1) the follow-on i row element of expression.

Promptly use b _i(k) the replacement A " a (k) _i(k), if there is not such b _i(k), a then _i(k) constant, Immune Selection obtains antibody population A of future generation (k+1) after accomplishing.

Step 5 with the state parameter substitution three-dimensional (3 D) manikin in the population, obtains the crucial articulation point three-dimensional coordinate of human body, utilizes the distance of two-dimentional crucial articulation point and tripleplane's point, designs distance function.

In estimating based on the human motion of monocular image sequence; Projection model of hypothesis under the situation of unknown camera parameters; In the present invention, used weak perspective projection model, preset root node depth of field parameter D; Make three-dimensional (3 D) manikin can project to the plane of delineation, be complementary with the corresponding point that detect on the plane.Kinematic parameter to be estimated comprises the integral translation of manikin, and each joint anglec of rotation totally 21 parameter X={ x, y, z, θ ₁, θ ₂..., θ ₁₈, 15 key point P on the human 3d model ₁, P ₂..., P ₁₅Three-dimensional coordinate determine P wherein by these 21 parameters _i=(P _Ix, P _Iy, P _Iz) coordinate that projects to behind the plane of delineation is p _i=(p _Ix, p _Iy)

$p_{\mathrm{ix}}=f\frac{P_{\mathrm{ix}}}{D}$

$p_{\mathrm{iy}}=f\frac{P_{\mathrm{iy}}}{D}$

Key point in the detected image is designated as q _i=(q _Ix, q _Iy), make that G (X) is the corresponding weighted sum of two check points and tripleplane's point, then,

$G\left(X\right)=\underset{i=1}{\overset{17}{\mathrm{\&Sigma;}}}{\left|\right|p_i-q_i\left|\right|}_2.$

Step 6 is designed similarity function, and the computed range affinity is preserved optimum solution, recovers the 3 D human body attitude.

(6.1) set up distance function G (X) after, the parameter estimation problem is asked a nonlinear optimization problem exactly:

X＝minG(X)

Utilize similarity function X=argminG (X), calculate the weighted sum of the distance of two-dimentional crucial articulation point and tripleplane's point, ask its minimum value again, in each generation, all stored optimum solution, and the optimum solution in this generation is not more than not preserving of globally optimal solution; Set a minimal value δ,, then stop calculating, otherwise turn back to step (3.2) if globally optimal solution is worth smaller or equal to this.

(6.2) last according to the three-dimensional human skeleton model of designing in the past, use the optimum solution substitution model of preserving, utilization computer graphics relevant knowledge is rotated and translation the initial skeleton pattern of human body, obtains 3 D human body reconstruction figure.

Effect of the present invention can further specify through following emulation:

1. emulation content:

Adopt quantum immune clone algorithm proposed by the invention, carry out the experiment of 3 D human body pose recovery.Wherein image sequence comes from the auto heterodyne video sequence, collects number of image frames 660 frames, image sequence spacing 1/24s.In estimation procedure, per 5 frame samplings, one frame is estimated usually, obtains the result of 132 frames.

At first come test head point detection method with two sequences, initial radium is obtained under the suitable situation, and head detection rate of accuracy reached to 100% has exceeded the requirement that is commonly used to do initial point.The detection of other each articulation points meets this experimental requirements basically.For the omission point,, can be easy to the unique point of estimating that detection is not come out usually through Kalman filtering.

The present invention is summed up as the determinacy nonlinear optimal problem with the human motion estimation problem, generally is difficult to find optimum solution, and in theory, this problem has 2 ^NIndividual local minimal solution, N are the number of rotary freedom.Initial value choose and whether kinematic constraint suitable will greatly influence results of optimization, in this experiment, people's initial attitude is that health is stood, level is stretched out both hands, is reflected to numeric representation to do

X ₀＝(167，96，0，

15，0，0，-30，

15，0，0，-30，

0，0，

0，0，0，10，

0，0，0，-10)。

Wherein, first three first primitive unit cell is a pixel, other be angle; The the 17th and 21 element is the angle of left and right sides ancon, is set to 10 degree and-10 degree respectively, promptly lets arm little curved; Being provided with of the 4th and 8 element guarantees that thigh is little anteflexion, and being provided with of the 7th and 11 element guarantees to stretch behind the shank.Optimized Algorithm can be provided with the hunting zone of each parameter component, and the lower boundary of each parameter is designated as LB, and the coboundary is UB, and generally speaking, the setting of initial value is for the motion that guarantees whole people in the algorithm search process is reasonably, the counter-rotating of ancon and knee can not occur.

According to affinity evaluation value, the storage optimum solution recovers the human body three-dimensional attitude.

Hardware platform is: Intel Core2 Duo CPU E6550 2.33GHZ, 2GB RAM.Software platform is MATLAB 7.0.

2. simulation result and analysis

Simulation result such as Fig. 3 and shown in Figure 4.

Fig. 3 provides the part estimated result, and wherein Fig. 3 a is that right leg protracts, the action of last arm held upward; Fig. 3 b is the 3 D human body attitude of recover the view from the front; Fig. 3 c recovers the 3 D human body attitude view from the side of coming out; Fig. 3 d left side leg protracts last arm held upward and back bending; Fig. 3 e is the 3 D human body attitude of recover the view from the front; Fig. 3 f recovers the 3 D human body attitude view from the side of coming out; Fig. 3 g is that right leg protracts, the action of stretching behind the both arms; Fig. 3 h is the 3 D human body attitude of recover the view from the front; Fig. 3 i recovers the 3 D human body attitude view from the side of coming out.

As can beappreciated from fig. 3, utilize quantum immune clone algorithm can estimate to obtain satisfied 3 D human body attitude.As can beappreciated from fig. 4, the coupling of tripleplane's point and two-dimentional corresponding check point is very desirable.And it is lower that the algorithm of being introduced applies in the human body tracking calculation cost, is not easy to be absorbed in locally optimal solution.

Claims (2)

1. three-dimensional human body motion tracking method based on quantum immune clone algorithm comprises following process:

(1) in the monocular image sequence that has human body attitude to exist, detects the crucial articulation point of two-dimension human body in the two dimensional image; And use classical kalman wave filter, and the crucial articulation point of detected two-dimension human body to be blocked a little and the omission point prediction, the motion that makes the crucial articulation point of two-dimension human body is more rationally with stable;

(2) based on the crucial artis of the two-dimension human body after detection and the prediction processing, set up a virtual human body three-dimensional skeleton pattern, make it in tracing process, to realize dynamic stance adjustment and coupling;

(3) quantum immune clone algorithm is introduced in the human motion tracking, at first the initialization population is provided with the human motion initial parameter, carries out clone operations then, with the search volume of increase human body attitude parameter to be estimated;

(4) to carrying out the population after the clone operations; Use quantum renewal operator and quantum crossover operator to clone reorganization and clonal vaviation respectively; This quantum upgrades operator; Adopt the accelerating convergence of quantum rotation door; Adopt the chaos mutation operation preventing precocity, and according to the affinity degree,

is if having

b _i(k)＝{a″ _ij(k)|max?D(a″ _ij)j＝1，2，...，q _i-1}

Make D (a _i(k))＜D (b _i(k)), i=1,2 ..., N, then a _i(k+1)=b _i(k),

A wherein " _IjExpression is operated the probability of i quantum antibody j position, back through renewal, D (a " _Ij) expression a " _IjAffinity, D (a _iAnd D (b (k)) _i(k)) expression a _i(k) affinity and b _i(k) affinity, a _i(k+1) the follow-on i row element of expression, N representes the number of elements of each row of population A (k), a _i(k) the capable all elements of i in the expression population A (k), b _i(k)=a " _Ij(k) | max D (a " _Ij) j=1,2 ..., q _i-1}, wherein q _iExpression antibody a _iScale behind the clone, thus the antibody that the chosen distance affinity is high relatively from filial generation separately after the operation of clone's immunogene and respective parent generates new population as outstanding antibody;

(5) with the state parameter of new population, substitution three-dimensional framework model produces crucial articulation point three-dimensional coordinate P _i=(P _Ix, P _Iy, P _Iz), and the key point that should key articulation point three-dimensional coordinate projects in the plane of delineation is designated as p _i=(p _Ix, p _Iy), the crucial articulation point of detected two-dimension human body is designated as q ' _i=(q ' _Ix, q ' _Iy), construct distance function and be:

X is the degree of freedom according to each crucial articulation point, setting totally wait to estimate the human parameters collection;

(6) based on distance function G (X) structure similarity function be: X '=minG (X); Utilize this similarity function to calculate the weighted sum of the distance of the crucial artis of two-dimension human body and tripleplane's point; Minimize again; Keep as optimum solution, the optimum solution in this generation is not more than not preserving of globally optimal solution; If optimum solution satisfies the end condition of setting, stop calculating; Otherwise, get back to step (3), for calculating desirable human body sport parameter, recover the 3 D human body attitude through too much.

2. three-dimensional human body motion tracking method according to claim 1; The described crucial articulation point of two-dimension human body that detects in the two dimensional image of process (1) wherein; Comprise that head node detection, root node detection, four limbs end-point detection and knee joint and elbow joint detect, the step of this head node detection is following:

(2a) in the front face human body contour images, establish concentrically ringed in the round heart be R ₁, the cylindrical center of circle is R ₂,, constitute the concentric circles template;

(2b) with the center of circle of concentric circles template skeleton point set C={c along the human body silhouette _jSearch, and j=1,2 ..., N ', N ' are current human skeleton zone skeleton point number of pixels; Profile point set S={s in calculating drops between circle and the cylindrical _iNumber of pixels, i=1,2 ..., M, M are the number of pixels of present frame point, when the number of pixels that falls into inside and outside circle the most for a long time, c _jBe the head node.

Images