CN100585622C - Human body tracing method based on gauss hybrid models - Google Patents

Abstract

A kind of human body tracing method of technical field of image processing based on gauss hybrid models, the present invention at first sets up model with the human body linear model to human body, head, neck, barycenter and double-legged with human body as following the tracks of the position, in follow-up frame, human body is respectively followed the tracks of the position then and carry out the Kalman filtering tracking, wherein follow the tracks of based on gauss hybrid models for both legs, adopt Kalman filtering to follow the tracks of to left foot, and, simplify tracking to right crus of diaphragm by the state of left foot being judged the motion state of right crus of diaphragm.The present invention is by the parameter of Kalman filtering prediction parts at next frame, and use observed reading correction predicted value, has guaranteed the accuracy that parts are followed the tracks of.Use gauss hybrid models to be the tracing object modeling, improved the state of following the tracks of efficient and can identifying the both legs motion.

Description

Human body tracing method based on gauss hybrid models

Technical field

The present invention relates to a kind of method of technical field of image processing, specifically is a kind of human body tracing method based on gauss hybrid models.

Background technology

Human motion analysis is the important technology that contemporary biomechanics and computer vision combine.It all has a wide range of applications in fields such as intelligent monitoring, man-machine interaction, biological identification technology and virtual realities.The motion analysis of human body shank particularly is for researchs such as medical treatment, physical culture provide important reference data.

Find by prior art documents, Qi Zhao etc. are in International Conference onPattern Recognition (pattern-recognition international conference, 2006) among the Part Based HumanTracking In A Multiple Cues Fusion Framework that delivers on (the human body tracking under multi thread fusion framework) based on parts, propose to set up HMM (hidden Markov) model, adopt the criterion of maximum a posteriori probability that each several part is followed the tracks of for each tracking section of human body.Its weak point is that each is followed the tracks of position integral body follows the tracks of as tracing object, can not overcome occlusion issue effectively.

Summary of the invention

The objective of the invention is to overcome the deficiency of said method, a kind of human body tracing method based on gauss hybrid models (GMM) has been proposed, with the human body linear model human body is set up model, adopt Kalman filtering to follow the tracks of head, neck, barycenter and the both feet of human body respectively, and the motion state of both feet is set up the GMM model, not only tracking efficient can be improved, and double-legged state can be identified.

The present invention is achieved by the following technical solutions, the present invention at first sets up model with the human body linear model to human body, head, neck, barycenter and double-legged with human body as following the tracks of the position, in follow-up frame, human body is respectively followed the tracks of the position then and carry out the Kalman filtering tracking, wherein follow the tracks of based on the GMM model for both legs, adopt Kalman filtering to follow the tracks of to left foot, and, simplify tracking right crus of diaphragm by the state of left foot being judged the motion state of right crus of diaphragm.

Describedly human body is set up model with the human body linear model, be specially: head, neck, barycenter and the both feet of representing human body with circle, the human body that circle is represented is follows the tracks of the position, will follow the tracks of the position with line segment and couple together, and obtains the height ratio that corresponding partes corporis humani divides.

Described tracking, the eigenwert of its tracking is the pendulum angle of the head position of human body, neck location, centroid position and both legs.

Describedly human body is respectively followed the tracks of the position follow the tracks of, be specially: adopt Kalman filtering that each position of manikin is followed the tracks of, Kalman filtering comprises prediction and revises two parts, predicted portions is meant: the predictive equation group utilizes the state value of previous moment and predicated error to make prediction, and obtains the position of each tracking unit at current time; Retouch is meant: owing to predict the outcome and can have certain error, the observed reading correction of the current time that is obtained by the utilization of update equation group predicts the outcome.

Describedly carry out both legs based on the GMM model and follow the tracks of, be specially: barycenter and both feet for human body are set up the single pendulum model, and the pendulum angle by both legs positions both feet; Simultaneously, the motion state of both feet is set up the GMM model, the motion of human body both feet is divided into two states: state that left foot motion right crus of diaphragm is static and the static state of right crus of diaphragm motion left foot, set up a Gauss model respectively for these two states, which Gauss model the motion of judging right crus of diaphragm by the state of left foot swing is in, when right crus of diaphragm is in relative ground static the time, then need not it is carried out Kalman's prediction, only need to adopt update equation correction motion parameter just can guarantee the accuracy of following the tracks of.

Compared with prior art, the present invention has following beneficial effect: the present invention can obtain higher tracking accuracy rate, uses linear model to calculate simple, intuitive.By the parameter of Kalman filtering prediction parts, and use observed reading correction predicted value, guaranteed the accuracy that parts are followed the tracks of at next frame.Use GMM to be the tracing object modeling, improved the state of following the tracks of efficient and can identifying the both legs motion, the accuracy that the right crus of diaphragm motion state is judged is 88.8%.

Description of drawings

Fig. 1 is that the present invention sets up human body linear model synoptic diagram to the human body structure

Wherein: figure (a) is the human body structure, (b) is human body linear model synoptic diagram;

Fig. 2 is a both legs single pendulum model of the present invention;

Fig. 3 is the experimental result of the embodiment of the invention;

Fig. 4 is the tracking results curve map of right crus of diaphragm pendulum angle of the present invention.

Embodiment

Below in conjunction with accompanying drawing embodiments of the invention are elaborated: present embodiment is being to implement under the prerequisite with the technical solution of the present invention, provided detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

As shown in Figure 3, present embodiment is used for following the tracks of the pedestrian of one section video sequence.

Present embodiment comprises the steps:

Step 1, with the human body linear model human body is set up model, as shown in Figure 1, figure (a) is the human body structure, (b) be human body linear model synoptic diagram, represent with circle that head, neck, barycenter and the both feet of human body, the human body that circle is represented are and follow the tracks of the position, to follow the tracks of the position with line segment and couple together, position or pendulum angle according to various piece in image sequence are followed the tracks of as feature.Wherein, be separate to the tracking of head, neck and barycenter, for both legs, then utilize the angle of its swing to follow the tracks of as feature;

Step 2 to the position of head, neck, barycenter, adopts Kalman filtering to follow the tracks of, and the tracking of Kalman filtering is divided into prediction and revises two parts.

Predicted portions is specially: the predictive equation group is utilized the state value of previous moment and the predicted value that predicated error obtains current time, and the predictive equation group is:

x′ _k＝Ax _k-1

P ' _k=AP _K-1A ^T+ Q (formula 1)

Wherein: x ' _kBe k predicted state variable constantly, x _K-1Be k-1 state variable constantly, A is a state-transition matrix, P ' _kBe k moment predicated error correlation matrix, P _K-1Be the modified value of k-1 moment error correlation matrix, Q is the motion noise correlation matrix.

Retouch is specially: obtain after predicted value and the prediction correlated error, the update equation group is made correction by the observed reading of current time to predicted value and prediction correlated error, and the update equation group is:

K _k＝P′ _kH ^T(HP′ _kH ^T+R) ^-1

x _k＝x′ _k+K _k(z _k-Hx′ _k)

P _k=(I-K _kH) P ' _k(formula 2)

Wherein: K _kBe the Kalman filtering gain matrix, H is for measuring matrix, and R is for measuring noise correlation matrix, z _kObserved reading.

In the present embodiment, observed reading is set at the position at place, tracking unit center, and state variable is made as position, movement velocity and the acceleration at parts center, and the hypothesis parts do uniformly accelerated motion, then that parameter setting is as follows:

z＝[s _x，s _y] ^T，x＝[s _x，v _x，a _x，s _y，v _y，a _y] ^T，

Wherein: z is an observed reading, and subscript x, y represent horizontal ordinate and ordinate, and s, a, v be position, movement velocity and the acceleration at part center, finger respectively.

State-transition matrix $A=\left[\begin{array}{cccccc}1& 1& 0.5& 0& 0& 0\\ 0& 1& 1& 0& 0& 0\\ 0& 0& 1& 0& 0& 0\\ 0& 0& 0& 1& 1& 0.5\\ 0& 0& 0& 0& 1& 1\\ 0& 0& 0& 0& 0& 1\end{array}\right],$

Measure matrix $H=\left[\begin{array}{cccccc}1& 0& 0& 0& 0& 0\\ 0& 0& 0& 1& 0& 0\end{array}\right]$

Parameter P _k, Q and R initial value need utilize priori to determine.

The update equation group is by observed reading z _kRevise current predicted value, obtain revised state estimation value and Noise Variance Estimation value.

Described observed reading z _k, its computation process is as follows: the displacement coordinate that dopes (s ' _x, s ' _y) for taking the method for spiral search in the small neighbourhood at center, search first and meet ρ [p spiral _k(x, y), q _K-1The coordinate of]＜l (x, y) ^TAs z _k, p wherein _k(x, y) (x y) is the histogram of the small neighbourhood at center, q to expression k frame with coordinate _K-1The object color model histogram of expression k-1 frame, ρ [p _k(x, y), q _K-1] be the distance of the histogram and the object color model of current coordinate, l is a distance threshold.

Step 3, for both feet, follow the tracks of based on the GMM model:

At first, utilize Kalman filtering that left foot is followed the tracks of.As shown in Figure 2, line and the horizontal angle of establishing barycenter and left foot is θ _l, the line of barycenter and right crus of diaphragm and horizontal angle are θ _r, the coordinate of barycenter is (x _o, y _o), the coordinate of left foot is (x _l, y _l), the length of leg is l, so:

x _l＝x _o+l*cos?x _l

y _l=y _o+ l*sin x _l(formula 3)

The utilization Kalman filtering is to θ _lFollow the tracks of, then the observational variable z=[θ of Kalman filter _l] ^T, state variable x=[θ _l, v _{L θ}], v _{L θ}For the angular velocity of both legs swing, suppose θ _lMove with uniform velocity, then have:

$A=\left[\begin{array}{cc}1& 1\\ 0& 1\end{array}\right]$

$H=\left[\begin{array}{c}1\\ 0\end{array}\right]$

In the parametric procedure of revising Kalman filter, through type 3 obtains left foot coordinate (x _l, y _l), take the method for spiral search in the small neighbourhood around its coordinate figure, search first and meet ρ [p _k(x, y), q _K-1(x is y) as (x for the coordinate of]＜l _t, y _t), p wherein _k(x, y) (x y) is the histogram of the small neighbourhood at center, q to expression k frame with coordinate _K-1The object color model histogram of expression k-1 frame, ρ [p _k(x, y), q _K-1] be the distance of the histogram and the object color model of current coordinate, l is a distance threshold.

The pendulum angle θ of left foot _lCoordinate (x with left foot _t, y _t) and the pass be:

${\mathrm{\θ}}_l=\arctan \left(\frac{x_t-x_o}{y_t-y_o}\right).$ (formula 4)

Utilize Kalman filtering to obtain the parameter θ of left foot swing state _lAnd v _{L θ}Afterwards, the motion of using the GMM model to judge both feet by the kinematic parameter of left foot belongs to any motion state.

The motion of human body both feet is divided into two states, i.e. state that left foot motion right crus of diaphragm is static and the static state of right crus of diaphragm motion left foot.Set up a Gauss model respectively for these two states, by parameter lambda _iExpression, the task of identification are determines a state i ^*, the model that it is corresponding

Make sample set X have maximum posterior probability p (λ _i/ X).According to bayesian theory, $i^{*}=\arg \underset{i}{\max }p(X/{\mathrm{\λ}}_i),$ i ^*Be the double-legged motion state that identifies.

Under the single pendulum model, double-legged motion state i ^*Can judge by following formula:

v _{L θ}Angular velocity (formula 5) for the left foot swing

The v ' that dopes by the Kalman _{L θ}Meet Gaussian distribution, v ' _{L θ}=v _{L θ}+ Δ ∈ λ _i, i=1,2, Δ is a noise, meets parameter and is (μ _Δ, σ _Δ) Gaussian distribution.

Probability p (v when the left foot motion _{L θ}During＞0) greater than probability threshold value α, then left foot is in motion, promptly

$p(v_{\mathrm{l\θ}}>0)$

$=p(v_{\mathrm{l\θ}}+\mathrm{\Δ}>\mathrm{\Δ})$

$=p(v_{\mathrm{l\θ}}^{\′}>\mathrm{\Δ})$ (formula 6)

$=p(\frac{\mathrm{\&Delta;}-{\mathrm{\&mu;}}_{\mathrm{\&Delta;}}}{{\mathrm{\&sigma;}}_{\mathrm{\&Delta;}}}<\frac{v_l^{\&prime;}-{\mathrm{\&mu;}}_{\mathrm{\&Delta;}}}{{\mathrm{\&sigma;}}_{\mathrm{\&Delta;}}})>\mathrm{\&alpha;}$

Can draw by formula 6

(formula 7)

Wherein $\frac{1}{\sqrt{2\mathrm{\&pi;}}}\underset{-\&infin;}{\overset{\mathrm{\&beta;}}{\&Integral;}}{e}^{-\frac{x^2}{2}}\mathrm{dx}=\mathrm{\&alpha;}.$

When judging right crus of diaphragm, then the right crus of diaphragm pendulum angle is carried out Kalman filtering and follow the tracks of in motion; When right crus of diaphragm remains static, then only need the Kalman filtering parameter is revised.

Fig. 3,4 is the figure as a result of present embodiment.As shown in Figure 3, each of movement human in the video sequence followed the tracks of position can both accurately locate (among the embodiment in order to distinguish left foot and right crus of diaphragm, with left foot box indicating), figure (a) and (b), (c), (d), (e), (f), (g), (h), (i), (j) are respectively the video intercepting picture of the 48th frame, the 60th frame, the 71st frame, the 82nd frame, the 92nd frame, the 111st frame, the 125th frame, the 140th frame, the 171st frame, the 180th frame, wherein, among Fig. 3 (b), for both feet block situation mutually, present embodiment also can correctly be followed the tracks of.Among Fig. 3 (e), both feet separate, and still are correct to the judgement of bipod.

Fig. 4 is the statistical graph of right crus of diaphragm pendulum angle in each frame, and circle represents that the circle of band plus sige represents that then this frame right crus of diaphragm is in static state by the speed of left foot swing is judged the state that right crus of diaphragm is in motion.As shown in Figure 4, the accuracy that the right crus of diaphragm motion state is judged is 88.8%, but by the correction to the right crus of diaphragm position, still can guarantee the accuracy that right crus of diaphragm is followed the tracks of.

Claims (3)

1, a kind of human body tracing method based on gauss hybrid models, it is characterized in that, the head of at first representing human body with circle, neck, barycenter and both feet, the human body that circle is represented is follows the tracks of the position, to follow the tracks of the position with line segment couples together, obtain the height ratio that corresponding partes corporis humani divides, head with human body, neck, barycenter and both feet are as following the tracks of the position, in follow-up frame, human body is respectively followed the tracks of the position then and carry out the Kalman filtering tracking, this Kalman filtering is followed the tracks of and is comprised the Kalman filtering prediction and revise two parts, predicted portions is meant: the predictive equation group utilizes the state value of previous moment and predicated error to make prediction, and obtains the position of each tracking unit at current time; Retouch is meant: can have error owing to predict the outcome, the observed reading correction of the current time that is obtained by the utilization of update equation group predicts the outcome, wherein follow the tracks of based on gauss hybrid models for both legs, barycenter and both feet for human body are set up the single pendulum model, and the pendulum angle by both legs positions both feet; Simultaneously, motion state to both feet is set up gauss hybrid models, the motion of human body both feet is divided into two states: state that left foot motion right crus of diaphragm is static and the static state of right crus of diaphragm motion left foot, set up a Gauss model respectively for these two states, which Gauss model the motion of judging right crus of diaphragm by the state of left foot swing is in, when right crus of diaphragm is in relative ground static the time, then need not it is carried out the Kalman filtering prediction, only need to adopt update equation correction motion parameter, adopt Kalman filtering to follow the tracks of to left foot, and, simplify tracking to right crus of diaphragm by the state of left foot being judged the motion state of right crus of diaphragm;

Which Gauss model the motion that described state by the left foot swing is judged right crus of diaphragm is in is meant: under the single pendulum model, double-legged motion state i can judge by following formula:

The v ' that dopes by Kalman filtering _{L θ}Meet Gaussian distribution, v ' _{L θ}=v _{L θ}+ Δ ∈ λ _i, i=1,2, Δ is a noise, meets parameter and is (μ _Δ, σ _Δ) Gaussian distribution,

Probability p (v when the left foot motion _{L θ}During＞0) greater than probability threshold value α, then left foot is in motion, promptly

$p(v_{\mathrm{l\&theta;}}>0)$

$=p(v_{\mathrm{l\&theta;}}+\mathrm{\&Delta;}<\mathrm{\&Delta;})$

(formula 6)

$=p(v_{\mathrm{l\&theta;}}^{\&prime;}>\mathrm{\&Delta;})$

$=p(\frac{\mathrm{\&Delta;}-{\mathrm{\&mu;}}_{\mathrm{\&Delta;}}}{{\mathrm{\&sigma;}}_{\mathrm{\&Delta;}}}<\frac{v_l^{\&prime;}-{\mathrm{\&mu;}}_{\mathrm{\&Delta;}}}{{\mathrm{\&sigma;}}_{\mathrm{\&Delta;}}})>\mathrm{\&alpha;}$

Can draw by formula 6

Wherein $\frac{1}{\sqrt{2\mathrm{\&pi;}}}\underset{-\&infin;}{\overset{\mathrm{\&beta;}}{\&Integral;}}{e}^{-\frac{x^2}{2}}\mathrm{dx}=\mathrm{\&alpha;}.$

2, the human body tracing method based on gauss hybrid models according to claim 1 is characterized in that, described tracking, and the eigenwert of its tracking is the pendulum angle of the head position of human body, neck location, centroid position and both legs.

3, the human body tracing method based on gauss hybrid models according to claim 1 is characterized in that, the observation of described Kalman filtering is meant: the displacement coordinate that draws in Kalman filtering predictive equation group (s ' _x, s ' _y) for taking the method for spiral search in the small neighbourhood at center, find the coordinate of observed reading, search first and meet ρ [p spiral _k(x, y), q _K-1The coordinate of]＜l (x, y) ^TAs observed reading z _k, p wherein _k(x, y) (x y) is the histogram of the small neighbourhood at center, q to expression k frame with coordinate _K-1The object color model histogram of expression k-1 frame, ρ [p _k(x, y), q _K-1] be the distance of the histogram and the object color model of current coordinate, l is a distance threshold.

Images