CN104200485A - Video-monitoring-oriented human body tracking method - Google Patents

Abstract

The invention provides a video-monitoring-oriented human body tracking method. Specific to the high instantaneity and scene complexity in video monitoring, a kernel-function-based adaptive human body shape change tracking method which combines mean shift and Kalman filtering is adopted. The method comprises the following steps: acquiring a tracking target through a background difference method; establishing a human body tracking template and a color histogram, and initializing the state of a Kalman filter; acquiring the position of the target by using a kernel-function-based mean shift method in a predicted range in a next frame video image of a video according to the range of a motion target predicted according to the Kalman filter; calculating the size of a target region through a projection drawing of the histogram of a target color; performing Kalman filtering correction to obtain a final position.

Description

A kind of human body tracing method of facing video monitoring

Technical field

The present invention relates to the method for technical field of image processing, specifically a kind of human body tracing method of facing video monitoring.

Background technology

Along with society, constantly development is with progressive, and the large common people are more and more higher for the security requirement of personal property, because video monitoring has intuitively, conveniently and be not subject to be more and more subject to apart from supremacy clauses such as the restrictions of discrete time people's favor.And in video monitoring for the detection of moving object, identification and tracking are the popular research directions of current intelligent video monitoring research field always, tracking common in video sequence has: the snake algorithm of the tracking based on profile, particle filter algorithm based on motion model, and meanshift algorithm based on color probability.Because the calculating of meanshift algorithm is simple, real-time is good, can be applicable to real-time video monitoring.

Cheng has applied to meanshift the direction of image processing first.The meanshift that the people such as Comaniciu propose follows the tracks of overall framework and performing step, has proposed to utilize kernel function to carry out to target the color characteristic that the upper weighting of distance can better represent target.Bradski has proposed the camshift method of utilizing the hue component of hsv color space to follow the tracks of.

Meanshift algorithm Kernel Function window width is changeless, in the time that target exists certain dimensional variation, can cause target following inaccurate, especially target away from or approach process in this phenomenon more obvious.

Camshift algorithm, taking meanshift algorithm as basis, has proposed a kind of adaptive method of target sizes.The method has improved the robustness of meanshift, and the tracking effect under solid background is better, but for the tracking under complex background still exist follow the tracks of less than problem.

Two kinds of algorithms are all taking average drifting as basic following principle, mean that the speed of moving object can not be excessive, if mobile object speed, mean algorithm cannot search out accurately target in the drift number of times of regulation.

Human body in video monitoring is nonrigid moving object, in motion process, have major part deformation, and due to the out-of-shape of human body, in the time that being extracted, target easily brings the interference of background into, the progressively expanding of interference meeting of background colour in tracking afterwards.Therefore it is not ideal using separately two kinds of algorithm effects.

Summary of the invention

The present invention will overcome the above-mentioned shortcoming of prior art, provide a kind of find target accurately, the human body tracing method of the facing video monitoring of high robust.

The present invention is studying the advantage in conjunction with two kinds of algorithms after meanshift based on color and camshift algorithm, the average drifting function of utilization based on core determined the position of human body, utilize the perspective view of the H component in HSV space to calculate the size of human body, and further improved the robustness of track algorithm in conjunction with Kalman filtering.

Irregular due to body shape, edge background bring the easy histogram that affects human body color into, along with constantly carrying out of following the tracks of, background error can be increasing, therefore select kernel function, follow the tracks of different place, center for distance and carry out different weighted values, can make the template of target more concentrate on center, reduce the interference that edge background is set up body templates.Make target following more concentrated.

The meanshift method of syncaryon function is because the bandwidth of the kernel function of its tracking target is constant, therefore the size of the tracking to target can not change along with target deformation, the good adaptive change of size can not follow the tracks of frame with to(for) the human body just camera being seesawed.Therefore use the perspective view of the H component in the HSV space of image, calculate the length of tracking object and wide.Make also can adaptive tracking object along with the generation of human body deformation size.And set up template renewal strategy, made kernel function to change tracking bandwidth along with the big or small change of target.

Owing to calculating the position of human body and size not in a kind of color space, there is certain increase for calculated amount, therefore adopted the way of Kalman Prediction, in the environs of Kalman prediction position, human body target is searched for, reduce for the calculating without human space, increased the robustness of disturbing for Similar color simultaneously.

The present invention mainly comprises following content

1. a human body tracing method for facing video monitoring, comprises the following steps:

1) background modeling and target detection are extracted, first applied scene is carried out to background modeling, extract scene background picture from video monitoring image, carry out gray scale for static scene and be modeled as pFimage, after completing, modeling starts video to carry out the detection of moving object, current detection frame is pFrame, first pFrame is carried out to gaussian filtering with smoothed image, reduce the interference of noise, present frame and background subtraction value obtain the target to be tracked of motion, the target target to be tracked detecting is carried out to morphologic filtering and can remove noise, and fill hole and make the profile of object more approach real human body, for the variation of indoor light and entering of some non-human objects, adopt the context update strategy of turnover rate β, the formula wherein upgrading is

pFimage＝pFimage+βpFrame

For detected object to be tracked, utilize the size exclusion of area to fall non-human moving target, obtain needing the human body target of tracking;

2) Kalman filter initialization and prediction

If target appears at F first _kframe, it appears at the upper p of being set to of image _k(x _k, y _k), according to detecting that the position of target is to the preliminary examination of Kalman filter, and utilize the state equation of Kalman filtering to dope the position (x of next frame target _p, y _p), for F _kframe is by predicted position (x _p, y _p) be made as the actual position of human body, and near region future position is set as estimation range;

3) initialization of human body target tracking

By F _kframe is transformed into the color distribution histogram hist of the hsv color space target body H component that also statistics obtains according to target detection from rgb space, calculate the back projection figure bp (x of estimation range according to hist _ij);

4) foundation of template and renewal

If target is to occur first, set up the feature templates model of human body at RGB color space, and the frame number that statistics has been followed the tracks of, if the frame number of statistics is the multiple that upgrades coefficient, think that template changes, recalculate body templates model with the target area of current tracking, the template using new template as search;

5) search of estimation range

In order to reduce calculated amount and to improve robustness, in the next frame of the frame that target occurs, near the position of searching for estimation range after target moves, searching method is at former frame target location (x _k, y _k) locate to choose the candidate target region the same with target sizes, centered by the position of former frame target, set up candidate target region model1, obtain the color probability distribution maximum position p of candidate target region _n;

6) searching position of judgement final goal

The method that determines whether optimal estimation is the color probability distribution maximum position p with twice target _n+1(x _n+1, y _n+1) and p _n(x _n, y _n) displacement | p _n+1-p _n| whether be less than threshold value, if be less than threshold value, the position p of new target _n+1(x _n+1, y _n+1) as the searching position of final goal, if be greater than threshold value, with new p _n+1(x _n+1, y _n+1) be core, To Template size re-establishes new candidate target region, and calculates the model1 of new candidate target region, returns to 5) recalculate the core position p of fresh target _n+2(x _n+2, y _n+2), until be less than threshold value or reached the number of times circulating, return to the target location p finally obtaining _n+m(x _n+m, y _n+m);

7) self-adaptation is calculated human region

According to 3) the target projection figure bp (x that obtains _ij) calculate the zeroth order square of target location, first moment, second moment meter and calculate length and the wide l of image ₁and l ₂and the angle tilting;

8) correction of Kalman filtering and renewal

When 5) position and 2 of the target that obtains of search) in the residual error of predicted position exceed after certain threshold value, select the position of prediction of Kalman filtering as the final position of target, otherwise use Kalman filter to proofread and correct position later as final position, the state of renewal carry out to(for) Kalman filter comprises the current state value of covariance and target, makes Kalman filter in next frame, can continue to predict the position of human body target.

2. step 2) middle according to detecting that the position of target is to the preliminary examination of Kalman filter, the state equation of Kalman filtering is

X(k)＝A?X(k-1)+W(k)

Wherein X (k) is current state matrix, the state matrix that X (k-1) is previous moment, W (k) is system noise, its feature distributes and meets Gaussian distribution, A is the transition matrix of system, X (k) is the matrix of one 4 dimension, X (k)=(x _k, y _k, V _kx, V _ky) x wherein _k, y _krepresent initial position horizontal ordinate and the ordinate of object of which movement, V _kx, V _kyrepresent transverse velocity and the longitudinal velocity of the speed of object of which movement, original state x _k, y _kfor the position of target, the speed V of preliminary examination _kx, V _kybe 0; The prediction object of Kalman filtering when carrying out location finding for human body target and to calculate estimation range perspective view, is searched near Kalman prediction point, can reduce the scope of search, reduces calculated amount, the real-time of enhancing target following.

3. step 3) in the target body region that obtains according to target detection, the probability histogram hist of the color of object H component in statistics HSV space, the interval statistics value of each is

$b_u=\underset{i=0}{\overset{n}{\mathrm{\Σ}}}\underset{j=0}{\overset{m}{\mathrm{\Σ}}}\mathrm{\δ}(p_{\mathrm{ij}}-u)$

Wherein p _ijexpression pixel (i, j) is located the value of pixel, and u represents histogrammic u interval, and the scope of u is l is the discrimination arranging, m, and n represents the horizontal number of target body area pixel point and longitudinally counts;

The back projection figure bp (x of estimation range _ij) formula as follows:

$\mathrm{bp}\left(p_{\mathrm{ij}}\right)=\underset{i=0}{\overset{n}{\mathrm{\Σ}}}\underset{j=0}{\overset{m}{\mathrm{\Σ}}}{b}_u\mathrm{\δ}[b\left(p_{\mathrm{ij}}\right)-u]$

Wherein p _ijexpression pixel (i, j) is located the value of pixel, b (p _ij) be illustrated in the p on position (i, j) _ijbetween corresponding u Statistical Area of histogram, b _urepresent u the value between Statistical Area, m, n represents that in the horizontal number of estimation range pixel and above two formulas of longitudinal number, δ (x) function expression is as follows:

$\mathrm{\δ}(x-u)=\left\{\begin{array}{cc}1& x=u\\ 0& x\≠u\end{array}\right.$

If belong to color of object in the perspective view obtaining, its value can be larger, do not belong to color of object value be 0.

4. the human body tracing method of a kind of facing video monitoring according to claim 1 is characterized in that: step 4) in set up human body feature templates be the target body of utilizing target detection to obtain first, be p in target location _k(x _k, y _k) locate to adopt the way of kernel density estimation to set up the model model of human body, model represents it is the probability density estimation of the estimation of u eigenwert in To Template, set up the triple channel model model of the RGB model of human body, model is a three-dimensional matrix, be used for representing the color characteristic of target, color of object is characterized as 16*16*16, and object module can be expressed as:

$\mathrm{mode}l=C\underset{i=1}{\overset{n}{\mathrm{\Σ}}}K\left({\left|\right|\frac{x_k-x_i}{h}\left|\right|}^2\right)\mathrm{\δ}[b\left(x_i\right)-u]$

Wherein b (x _i) be x _ithe fiducial value of u feature of pixel at place, the value of u be 1...m}, δ is Kronecker function, h is kernel function bandwidth, the height of n target area, C is that normalization coefficient is expressed as:

$C=\frac{1}{{\mathrm{\Σ}}_{i=1}^{n}K\left({\left|\right|\frac{x_k-x_i}{h}\left|\right|}^2\right)}$

K (x) Ye Panieqi Nico husband kernel function expression formula is:

$K\left(x\right)=\left\{\begin{array}{cc}\frac{1}{2}\mathrm{cd}-1(d+2)(1-|\left|x\right|\left|2\right),& \mathrm{if}\left|\right|x\left|\right|<1\\ 0,& \mathrm{if}\left|\right|x\left|\right|>1\end{array}\right.$

The size that wherein cd is area of space, the dimension of d representation space;

5. the human body tracing method of a kind of facing video monitoring according to claim 1 is characterized in that: step 5) in the core calculations of candidate target region be the principle according to average drifting, set up the weight ratio value matrix ω of the object module based on kernel function _i

${\mathrm{\&omega;}}_i=\underset{u=1}{\overset{m}{\mathrm{\&Sigma;}}}\sqrt{\frac{\mathrm{mode}l}{\mathrm{mode}l1}}\mathrm{\&delta;}[b\left(x_i\right)-u]$

Wherein model is the object module in right 4, the model that model1 is candidate target region, and modeling pattern is identical with model, and the summation that m is target area point, when target former frame target location is p _k(x _k, y _k) time, the candidate target region of choosing due to present frame is taking former frame target location as core, and the expression formula of setting up candidate region model model1 is as follows

$\mathrm{mode}l1=C\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}K\left({\left|\right|\frac{x_k-x_i}{h}\left|\right|}^2\right)\mathrm{\&delta;}[b\left(x_i\right)-u]$

B (x _i) be x _ithe fiducial value of u feature of pixel at place, the value of u be 1...m}, δ is Kronecker function, h is kernel function bandwidth, the height that n is candidate target region, C is normalization coefficient, expression formula is identical with model model,

The core p of present frame candidate target region _nfor

$p_n=\frac{{\mathrm{\&Sigma;}}_{i=1}^{\mathrm{nh}}{x}_i{\mathrm{\&omega;}}_{i}g\left({\left|\right|\frac{x_k-x_i}{h}\left|\right|}^2\right)}{{\mathrm{\&Sigma;}}_{i=1}^{\mathrm{nh}}{\mathrm{\&omega;}}_{i}g\left({\left|\right|\frac{x_k-x_i}{h}\left|\right|}^2\right)}$

Wherein g (x) is the derivative of kernel function K (x), and K (x) adopts Ye Panieqi Nico husband kernel function, the height that n is candidate target region, the bandwidth that h is kernel function, i.e. candidate target region width.

6. step 7) according to perspective view bp (x _ij) moment of the orign of target area, first order and second order moments, and calculate size and the following M of its computing method of angle of inclination of target ₀₀for moment of the orign is M ₁₀, M ₀₁, M ₁₁for first moment M ₂₀, M ₀₂for second moment

$\begin{array}{cccccc}{M}_{00}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_i& M_{10}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_{i}x& M_{01}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_{i}y& M_{20}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_i{x}^2& M_{02}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_i{y}^2& M_{11}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_i\mathrm{xy}\end{array}$

$\begin{array}{cc}{x}_c=\frac{M_{10}}{M_{00}}& y_c=\frac{M_{01}}{M_{00}}\end{array}$

$\begin{array}{ccc}a=\frac{M_{20}}{M_{00}}-{x_c}^2& b=2(\frac{M_{11}}{M_{00}}-x_c{y}_c)& c=\frac{M_{02}}{M_{00}}-{y_c}^2\end{array}$

$\begin{array}{ccc}{l}_1=\sqrt{\frac{(a+c)+\sqrt{b^2+{(a-c)}^2}}{2}}& l2=\sqrt{\frac{(a+c)-\sqrt{b^2+{(a-c)}^2}}{2}}& \mathrm{\&theta;}=\frac{1}{2}{\tan }^{-1}\left(\frac{b}{a-c}\right)\end{array}$

Wherein l ₁, l ₂the length in the region of expression human body target and wide, θ represents the angle that target tilts, w _irepresent perspective view bp (x _ij) upper point (x, y) value of locating.

7. step 8) in the renewal of Kalman filter comprise the current state value of covariance and target, wherein upgrade step 5 in observed reading observed reading Z (k) right 1 of Karman equation) in the position of the human body searching,

Current state value is made to optimized being estimated as follows:

X(k|k)＝X(k|k-1)+Kg(k)(Z(k)-H?X(k|k-1))

The kalman gain (Kalman Gain) that wherein Kg (k) is current state, the optimal estimation of X (k|k) current state, X (k|k-1) is the optimal estimation to the k moment in the k-1 moment, and H is the given measurement matrix of initialization

Kg(k)＝P(k|k-1)H ^T/(H?P(k|k-1)H ^T+R)

Wherein P (k|k-1) is illustrated in the covariance of the optimal estimation of k-1 moment to the k moment, upgrades P (k|k-1),

P(k|k-1)＝A?P(k-1|k-1)A ^T+Q

Wherein A is the transition matrix of state equation in right 2, finally needs to upgrade the covariance of X under k state (k|k)

P(k|k)＝(I-Kg(k)H)P(k|k-1)

In above formula, Q and R represent respectively the covariance of motion artifacts and the measurement noise of system, are made as constant 1e-5 and 1e-1;

The correction of Kalman filter is the position (x (k) of the target that obtains of present frame searching algorithm, y (k)) and former frame Kalman prediction position (x (k-1), y (k-1)) residual error exceed after threshold value, select the position of prediction of Kalman filtering as the final position of target, wherein residual error is defined as:

$r\left(k\right)=\sqrt{(x\left(k\right)-{x(k-1)}^2)+(y\left(k\right)-y{(k-1)}^2)}$

Do not exceed threshold value and talk about the position adopting after Kalman filter is proofreaied and correct as final position.

Advantage of the present invention is: can adaptive calculating follow the tracks of the variation of human body size for the human body of following the tracks of 1.; 2. in the time that search needs the target of tracking, trace template can be according to following the tracks of the bandwidth of frame number to trace template self-adaptation adjustment kernel function; 3. predict the position that has adopted the forecasting characters of Kalman filtering to occur target, reduces the calculating in unnecessary region, has avoided the interference of other objects in guarded region, has improved real-time and the robustness of target following.

Brief description of the drawings

Fig. 1 is that target of the present invention is extracted binary map

Fig. 2 is the schematic diagram of kernel function of the present invention at higher dimensional space

Fig. 2 a is the schematic diagram of average kernel function of the present invention

Fig. 2 b is the schematic diagram of gaussian kernel function of the present invention

Fig. 3 is adaptive tracking method process flow diagram of the present invention

Fig. 4 is overall flow figure of the present invention

Embodiment

Just the concrete implementation process of the present invention is described in detail below, below setting parameter be the optimal value that test obtains under this experimental situation.Because color tracking algorithm has certain degree of association for light and surrounding environment, in other environment, also can further optimize and revise, example is below only an example of numerous experiments.The present invention, through multiple case verification, has proved its validity and practicality.

The test environment of this test is indoor environment, and human body tracking has selected single body to walk normally in laboratory, the fixing camera that camera is monocular, and human body is to advance in front and back, has certain human body deformation quantity.

1. a human body tracing method for facing video monitoring, comprises the following steps:

1) background modeling and target detection are extracted, first applied scene is carried out to background modeling, extract scene background picture from video monitoring image, carry out gray scale for static scene and be modeled as pFimage, after completing, modeling starts video to carry out the detection of moving object, current detection frame is pFrame, first pFrame is carried out to gaussian filtering with smoothed image, reduce the interference of noise, present frame and background subtraction value obtain the target to be tracked of motion, the target target to be tracked detecting is carried out to morphologic filtering and can remove noise, and fill hole and make the profile of object more approach real human body, for the variation of indoor light and entering of some non-human objects, adopt the context update strategy of turnover rate β, the formula wherein upgrading is

pFimage＝pFimage+βpFrame

For detected object to be tracked, utilize the size exclusion of area to fall non-human moving target, obtain needing the human body target of tracking, in actual implementation process, turnover rate β is 0.005;

2) Kalman filter initialization and prediction

If target appears at F first _kframe, it appears at the upper p of being set to of image _k(x _k, y _k), according to detecting that the position of target is to the preliminary examination of Kalman filter, and utilize the state equation of Kalman filtering to dope the position (x of next frame target _p, y _p), for F _kframe is by predicted position (x _p, y _p) be made as the actual position of human body, and near region future position is set as estimation range;

3) initialization of human body target tracking

By F _kframe is transformed into the color distribution histogram hist of the hsv color space target body H component that also statistics obtains according to target detection from rgb space, calculate the back projection figure bp (x of estimation range according to hist _ij);

4) foundation of template and renewal

If target is to occur first, set up the feature templates model of human body at RGB color space, and the frame number that statistics has been followed the tracks of, if the frame number of statistics is the multiple that upgrades coefficient, think that template changes, recalculate body templates model with the target area of current tracking, template using new template as search, upgrades coefficient and be made as 4 in actual implementation process;

5) search of estimation range

In order to reduce calculated amount and to improve robustness, in the next frame of the frame that target occurs, near the position of searching for estimation range after target moves, searching method is at former frame target location (x _k, y _k) locate to choose the candidate target region the same with target sizes, centered by the position of former frame target, set up candidate target region model1, obtain the color probability distribution maximum position p of candidate target region _n;

6) searching position of judgement final goal

The method that determines whether optimal estimation is the color probability distribution maximum position p with twice target _n+1(x _n+1, y _n+1) and p _n(x _n, y _n) displacement | p _n+1-p _n| whether be less than threshold value, if be less than threshold value, the position p of new target _n+1(x _n+1, y _n+1) as the searching position of final goal, if be greater than threshold value, with new p _n+1(x _n+1, y _n+1) be core, To Template size re-establishes new candidate target region, and calculates the model1 of new candidate target region, returns to 5) recalculate the core position p of fresh target _n+2(x _n+2, y _n+2), until be less than threshold value or reached the number of times circulating, return to the target location p finally obtaining _n+m(x _n+m, y _n+m), in actual enforcement, threshold value is made as 3, and cycle index is 20 times;

7) self-adaptation is calculated human region

According to 3) the target projection figure bp (x that obtains _ij) calculate the zeroth order square of target location, first moment, second moment meter and calculate length and the wide l of image ₁and l ₂and the angle tilting;

8) correction of Kalman filtering and renewal

When 5) position and 2 of the target that obtains of search) in the residual error of predicted position exceed after certain threshold value, select the position of prediction of Kalman filtering as the final position of target, otherwise use Kalman filter to proofread and correct position later as final position, the state of renewal carry out to(for) Kalman filter comprises the current state value of covariance and target, makes Kalman filter in next frame, can continue to predict the position of human body target.

2. the human body tracing method of a kind of facing video monitoring according to claim 1 is characterized in that: step 2) middle according to detecting that the position of target is to the preliminary examination of Kalman filter, the state equation of Kalman filtering is

X(k)＝A?X(k-1)+W(k)

Wherein X (k) is current state matrix, the state matrix that X (k-1) is previous moment, W (k) is system noise, its feature distributes and meets Gaussian distribution, A is the transition matrix of system, X (k) is the matrix of one 4 dimension, X (k)=(x _k, y _k, V _kx, V _ky) x wherein _k, y _krepresent initial position horizontal ordinate and the ordinate of object of which movement, V _kx, V _kyrepresent transverse velocity and the longitudinal velocity of the speed of object of which movement, original state x _k, y _kfor the position of target, the speed V of preliminary examination _kx, V _kybe 0; The prediction object of Kalman filtering when carrying out location finding for human body target and to calculate estimation range perspective view, is searched near Kalman prediction point, can reduce the scope of search, reduces calculated amount, and the real-time of enhancing target following wherein

$A=\left\{\begin{array}{c}1010\\ 0101\\ 0010\\ 0001\end{array}\right\},$ B is full 0 matrix.

3. step 3) in the target body region that obtains according to target detection, the probability histogram hist of the color of object H component in statistics HSV space, the interval statistics value of each is

$b_u=\underset{i=0}{\overset{n}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{m}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(p_{\mathrm{ij}}-u)$

Wherein p _ijexpression pixel (i, j) is located the value of pixel, and u represents histogrammic u interval, and the scope of u is l is the discrimination arranging, m, and n represents horizontal number and longitudinal number of target body area pixel point, in specific implementation process, l is made as 50;

The back projection figure bp (x of estimation range _ij) formula as follows:

$\mathrm{bp}\left(p_{\mathrm{ij}}\right)=\underset{i=0}{\overset{n}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{m}{\mathrm{\&Sigma;}}}{b}_u\mathrm{\&delta;}[b\left(p_{\mathrm{ij}}\right)-u]$

Wherein p _ijexpression pixel (i, j) is located the value of pixel, b (p _ij) be illustrated in the p on position (i, j) _ijbetween corresponding u Statistical Area of histogram, b _urepresent u the value between Statistical Area, m, n represents that in the horizontal number of estimation range pixel and above two formulas of longitudinal number, δ (x) function expression is as follows:

$\mathrm{\&delta;}(x-u)=\left\{\begin{array}{cc}1& x=u\\ 0& x\&NotEqual;u\end{array}\right.$

If belong to color of object in the perspective view obtaining, its value can be larger, do not belong to color of object value be 0.

4. the human body tracing method of a kind of facing video monitoring according to claim 1 is characterized in that: step 4) in set up human body feature templates be the target body of utilizing target detection to obtain first, be p in target location _k(x _k, y _k) locate to adopt the way of kernel density estimation to set up the model model of human body, model represents it is the probability density estimation of the estimation of u eigenwert in To Template, set up the triple channel model model of the RGB model of human body, model is a three-dimensional matrix, be used for representing the color characteristic of target, color of object is characterized as 16*16*16, and object module can be expressed as:

$\mathrm{mode}l=C\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}K\left({\left|\right|\frac{x_k-x_i}{h}\left|\right|}^2\right)\mathrm{\&delta;}[b\left(x_i\right)-u]$

Wherein b (x _i) be x _ithe fiducial value of u feature of pixel at place, the value of u be 1...m}, δ is Kronecker function, h is kernel function bandwidth, the height of n target area, C is that normalization coefficient is expressed as:

$C=\frac{1}{{\mathrm{\&Sigma;}}_{i=1}^{n}K\left({\left|\right|\frac{x_k-x_i}{h}\left|\right|}^2\right)}$

K (x) Ye Panieqi Nico husband kernel function expression formula is:

$K\left(x\right)=\left\{\begin{array}{cc}\frac{1}{2}\mathrm{cd}-1(d+2)(1-|\left|x\right|\left|2\right),& \mathrm{if}\left|\right|x\left|\right|<1\\ 0,& \mathrm{if}\left|\right|x\left|\right|>1\end{array}\right.$

The size that wherein cd is area of space, the dimension of d representation space;

5. the human body tracing method of a kind of facing video monitoring according to claim 1 is characterized in that: step 5) in the core calculations of candidate target region be the principle according to average drifting, set up the weight ratio value matrix ω of the object module based on kernel function _i

${\mathrm{\&omega;}}_i=\underset{u=1}{\overset{m}{\mathrm{\&Sigma;}}}\sqrt{\frac{\mathrm{mode}l}{\mathrm{mode}l1}}\mathrm{\&delta;}[b\left(x_i\right)-u]$

Wherein model is the object module in right 4, the model that model1 is candidate target region, and modeling pattern is identical with model, and the summation that m is target area point, when target former frame target location is p _k(x _k, y _k) time, the candidate target region of choosing due to present frame is taking former frame target location as core, and the expression formula of setting up candidate region model model1 is as follows

$\mathrm{mode}l1=C\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}K\left({\left|\right|\frac{x_k-x_i}{h}\left|\right|}^2\right)\mathrm{\&delta;}[b\left(x_i\right)-u]$

B (x _i) be x _ithe fiducial value of u feature of pixel at place, the value of u be 1...m}, δ is Kronecker function, h is kernel function bandwidth, the height that n is candidate target region, C is normalization coefficient, expression formula is identical with model model,

The core p of present frame candidate target region _nfor

$p_n=\frac{{\mathrm{\&Sigma;}}_{i=1}^{\mathrm{nh}}{x}_i{\mathrm{\&omega;}}_{i}g\left({\left|\right|\frac{x_k-x_i}{h}\left|\right|}^2\right)}{{\mathrm{\&Sigma;}}_{i=1}^{\mathrm{nh}}{\mathrm{\&omega;}}_{i}g\left({\left|\right|\frac{x_k-x_i}{h}\left|\right|}^2\right)}$

Wherein g (x) is the derivative of kernel function K (x), and K (x) adopts Ye Panieqi Nico husband kernel function, the height that n is candidate target region, the bandwidth that h is kernel function, i.e. candidate target region width.

6. the human body tracing method of a kind of facing video monitoring according to claim 1 is characterized in that: step 7) according to perspective view bp (x _ij) moment of the orign of target area, first order and second order moments, and calculate size and the following M of its computing method of angle of inclination of target ₀₀for moment of the orign is M ₁₀, M ₀₁, M ₁₁for first moment M ₂₀, M ₀₂for second moment

$\begin{array}{cccccc}{M}_{00}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_i& M_{10}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_{i}x& M_{01}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_{i}y& M_{20}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_i{x}^2& M_{02}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_i{y}^2& M_{11}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_i\mathrm{xy}\end{array}$

$\begin{array}{cc}{x}_c=\frac{M_{10}}{M_{00}}& y_c=\frac{M_{01}}{M_{00}}\end{array}$

$\begin{array}{ccc}a=\frac{M_{20}}{M_{00}}-{x_c}^2& b=2(\frac{M_{11}}{M_{00}}-x_c{y}_c)& c=\frac{M_{02}}{M_{00}}-{y_c}^2\end{array}$

$\begin{array}{ccc}{l}_1=\sqrt{\frac{(a+c)+\sqrt{b^2+{(a-c)}^2}}{2}}& l2=\sqrt{\frac{(a+c)-\sqrt{b^2+{(a-c)}^2}}{2}}& \mathrm{\&theta;}=\frac{1}{2}{\tan }^{-1}\left(\frac{b}{a-c}\right)\end{array}$

Wherein l ₁, l ₂the length in the region of expression human body target and wide, θ represents the angle that target tilts, w _irepresent perspective view bp (x _ij) upper point (x, y) value of locating.

7. the human body tracing method of a kind of facing video monitoring according to claim 1 is characterized in that: step 8) in the renewal of Kalman filter comprise the current state value of covariance and target, wherein upgrade step 5 in observed reading observed reading Z (k) right 1 of Karman equation) in the position of the human body searching

Current state value is made to optimized being estimated as follows:

X(k|k)＝X(k|k-1)+Kg(k)(Z(k)-H?X(k|k-1))

The kalman gain (Kalman Gain) that wherein Kg (k) is current state, the optimal estimation of X (k|k) current state, X (k|k-1) is the optimal estimation to the k moment in the k-1 moment, and H is the given measurement matrix of initialization

Kg(k)＝P(k|k-1)H ^T/(H?P(k|k-1)H ^T+R)

Wherein P (k|k-1) is illustrated in the covariance of the optimal estimation of k-1 moment to the k moment, upgrades P (k|k-1),

P(k|k-1)＝A?P(k-1|k-1)A ^T+Q

Wherein A is the transition matrix of state equation in right 2, finally needs to upgrade the covariance of X under k state (k|k)

P(k|k)＝(I-Kg(k)H)P(k|k-1)

In above formula, Q and R represent respectively the covariance of motion artifacts and the measurement noise of system, are made as constant 1e-5 and 1e-1;

The correction of Kalman filter is the position (x (k) of the target that obtains of present frame searching algorithm, y (k)) and former frame Kalman prediction position (x (k-1), y (k-1)) residual error exceed after threshold value, select the position of prediction of Kalman filtering as the final position of target, wherein residual error is defined as:

$r\left(k\right)=\sqrt{(x\left(k\right)-{x(k-1)}^2)+(y\left(k\right)-y{(k-1)}^2)}$

Do not exceed threshold value and talk about the position adopting after Kalman filter is proofreaied and correct as final position.

Content described in this instructions embodiment is only enumerating of way of realization to inventive concept; protection scope of the present invention should not be regarded as only limiting to the concrete form that embodiment states, protection scope of the present invention is also forgiven those skilled in the art and conceived the equivalent technologies means that can expect according to the present invention.

Claims (7)

1. A human body tracking method for video monitoring, comprising the following steps: 1) Background modeling and target detection and extraction. First, perform background modeling on the applied scene, extract scene background pictures from video surveillance images, and perform grayscale modeling for static scenes as pFimage. After the modeling is completed, start to move the video For object detection, the current detection frame is pFrame. First, Gaussian filtering is performed on pFrame to smooth the image and reduce noise interference. The difference between the current frame and the background is used to obtain the moving target to be tracked, and morphological filtering is performed on the detected target to be tracked. It can remove the noise and fill holes to make the outline of the object closer to the real human body. For the change of indoor light and the entry of some non-human objects, the background update strategy of the update rate β is adopted, and the update formula is pFimage＝pFimage+βpFrame For the detected objects to be tracked, use the size of the area to exclude non-human moving targets, and obtain the human targets that need to be tracked; 2) Kalman filter initialization and prediction If the target appears for the first time in the F _kth frame, and its position on the image is p _k (x _k , y _k ), the Kalman filter is initialized according to the position of the detected target, and the state of the Kalman filter is used The equation predicts the position (x _p , y _p ) of the target in the next frame. For the F _kth frame, the predicted position (x _p , y _p ) is set as the real position of the human body, and the area near the predicted point is set as the predicted area; 3) Initialization of human target tracking Convert the F _k frame from the RGB space to the HSV color space and count the color distribution histogram histogram of the H component of the target human body obtained according to the target detection, and calculate the back projection map bp(x _ij ) of the predicted area according to the hist; 4) Establishment and update of templates If the target appears for the first time, build a feature template model of the human body in the RGB color space, and count the number of frames that have been tracked. If the number of frames counted is a multiple of the update coefficient, the template is considered to have changed, and the current tracked target area is recalculated. Human body template model, using the new template as the search template; 5) Search for prediction area In order to reduce the amount of calculation and improve the robustness, in the next frame of the frame where the target appears, search for the moved position of the target near the predicted area. The search method is to select the target position (x _k , y _k ) in the previous frame and For the candidate target area with the same target size, the candidate target area model1 is established centering on the position of the target in the previous frame, and the maximum position p _n of the color probability distribution of the candidate target area is obtained; 6) Determine the search position of the final target The method of judging whether it is the best estimate is based on the maximum position p _n+1 (x _n+1 , y _n+1 ) and the moving distance of p _n (x _n ,y _n ) of the color probability distribution of the target | p _n Whether ₊₁ -p _n | is less than the threshold, if it is less than the threshold, the new target position p _n+1 (x _n+1 , y _n+1 ) is used as the search position of the final target, if it is greater than the threshold, the new p _n is used ₊₁ (x _n+1 , y _n+1 ) as the core, the size of the target template to re-establish a new candidate target area, and calculate the model1 of the new candidate target area, return 5) recalculate the core position p _n+ of the new target ₂ (x _n+2 ,y _n+2 ), until it is less than the threshold or reaches the number of cycles, return to the final target position p _n+m (x _n+m ,y _n+m ); 7) Adaptive calculation of human body area Calculate the zero-order moment, first-order moment, and second-order moment of the target position according to the target projection map bp(x _ij ) obtained in 3), and calculate the length and width l ₁ and l ₂ of the image and the angle of inclination; 8) Correction and update of Kalman filter When the residual error between the position of the target obtained in 5) and the predicted position in 2) exceeds a certain threshold, the predicted position of the Kalman filter is selected as the final position of the target, otherwise the corrected position of the Kalman filter is used as the final position Position, for the Kalman filter, the update of the state includes the covariance and the current state value of the target, so that the Kalman filter can continue to predict the position of the human target in the next frame. the 2. a kind of human body tracking method facing video monitoring according to claim 1 is characterized in that: in step 2) according to detecting the position of target to the initial test of Kalman filter, the state equation of Kalman filter is X(k)＝A X(k-1)+W(k) Among them, X(k) is the current state matrix, X(k-1) is the state matrix at the previous moment, W(k) is the system noise, and its characteristic distribution conforms to Gaussian distribution, A is the transition matrix of the system, X(k) It is a 4-dimensional matrix, X(k)=(x _k , y _k , V _kx , V _ky ) where x _k , y _k represent the abscissa and ordinate of the initial position of the object movement, V _kx , V _ky represent The horizontal velocity and longitudinal velocity of the moving speed of the object, the initial state x _k , y _k is the position of the target, and the initial velocity V _kx , V _ky is 0; the purpose of Kalman filter prediction is to search and calculate the position of the human target When predicting the regional projection map, searching near the predicted point of Kalman filter can reduce the search range, reduce the amount of calculation, and enhance the real-time performance of target tracking. 3. a kind of human body tracking method facing video monitoring according to claim 1 is characterized in that: step 3) in the target human body area that obtains according to target detection, the probability histogram hist of the target color H component of statistical HSV space, every An interval statistic value of Among them, p _ij represents the value of the pixel at the pixel point (i,j), u represents the uth interval of the histogram, and the range of u is l is the set discrimination degree, m and n represent the horizontal number and vertical number of pixels in the target human body area; The formula of the back projection map bp(x _ij ) of the prediction area is as follows: Among them, p _ij represents the value of the pixel point at the pixel point (i, j), b(p _ij ) represents the uth statistical interval of the histogram corresponding to p _ij at the position (i, j), and b _u represents the uth The value of the statistical interval, m, n represent the horizontal number and vertical number of pixels in the prediction area. The expression of the δ(x) function in the above two formulas is as follows: If the obtained projection image belongs to the target color, its value will be relatively large, and if the value does not belong to the target color, it will be 0. the 4. a kind of human body tracking method facing video monitoring according to claim 1 is characterized in that: step 4) sets up the feature template of human body to be the target human body that utilizes target detection to obtain for the first time, and is p _k (x _k at target position , y _k ) uses the method of kernel function density estimation to establish the model model of the human body. The model represents the estimated probability density estimation of the uth eigenvalue in the target template. The three-channel model model of the RGB model of the human body is established. The model is a The matrix of the three-dimensional space is used to represent the color feature of the target. The target color feature is 16*16*16. The target model can be expressed as: Where b( _xi ) is the reference value of the uth feature of the pixel point at x _i , the value of u is {1...m}, δ is the Kronecker function, h is the bandwidth of the kernel function, and n targets The height of the area, C is the normalization coefficient expressed as: The expression of K(x) Yepanichnikov kernel function is: where cd is the size of the spatial region, and d represents the dimension of the space. 5. a kind of human body tracking method facing video surveillance according to claim 1 is characterized in that: the core calculation of candidate target area in step 5) is to set up the weight ratio matrix based on the target model of kernel function according to the principle of mean value shift ω _i Where model is the target model in right 4, model1 is the model of the candidate target area, the modeling method is the same as model, m is the sum of points in the target area, and the target position in the previous frame of the target is p _k (x _k ,y _k ) When , since the candidate target area selected in the current frame is the core of the target position in the previous frame, the expression for establishing the candidate area model model1 is as follows b( _xi ) is the reference value of the uth feature of the pixel point at x _i , the value of u is {1...m}, δ is the Kronecker function, h is the bandwidth of the kernel function, and n is the candidate The height of the target area, C is the normalization coefficient, the expression is the same as the model model, The core p _n of the candidate target area in the current frame is Where g(x) is the derivative of the kernel function K(x), K(x) uses the Yepanichnikov kernel function, n is the height of the candidate target area, h is the bandwidth of the kernel function, that is, the width of the candidate target area . the 6. a kind of human body tracking method facing video monitoring according to claim 1 is characterized in that: in step 7), according to the origin moment of projection map bp (x _ij ) target area, first-order moment and second-order moment, and calculate The calculation method for the size and inclination angle of the target is as follows: M ₀₀ is the origin moment, M ₁₀ , M ₀₁ , M ₁₁ is the first-order moment M ₂₀ , and M ₀₂ is the second-order moment Among them, l ₁ and l ₂ represent the length and width of the area of the human target, θ represents the angle of target inclination, and w _i represents the value at point (x, y) on the projection map bp(x _ij ). 7. a kind of human body tracking method facing video surveillance according to claim 1 is characterized in that: the update of Kalman filter in step 8) comprises covariance and the current state value of target, wherein the observed value of updating Kalman equation The position of the human body searched in step 5) in the observation value Z(k) right 1, The optimal estimation of the current state value is as follows: X(k|k)＝X(k|k-1)+Kg(k)(Z(k)-H X(k|k-1))  Where Kg(k) is the Kalman Gain of the current state, X(k|k) is the optimal estimate of the current state, and X(k|k-1) is the optimal estimate of time k at time k-1 , H is to initialize the given measurement matrix Kg(k)＝P(k|k-1)H ^T /(H P(k|k-1)H ^T +R) Where P(k|k-1) represents the covariance of the optimal estimate at time k-1 to time k, update P(k|k-1), P(k|k-1)＝A P(k-1|k-1) ^AT +Q Where A is the transition matrix of the state equation in right 2, and finally the covariance of X(k|k) in state k needs to be updated P(k|k)＝(I-Kg(k)H)P(k|k-1) In the above formula, Q and R represent the covariance of the motion noise and measurement noise of the system respectively, and are set to constants 1e-5 and 1e-1; The deviation correction of the Kalman filter is the target position (x(k), y(k)) obtained by the current frame search algorithm and the predicted position of the Kalman filter in the previous frame (x(k-1), y(k-1) ) exceeds the threshold, the predicted position of the Kalman filter is selected as the final position of the target, where the residual is defined as: If the threshold is not exceeded, the position corrected by the Kalman filter is used as the final position. the