CN103823973B - Target tracking method based on MeanShift algorithm - Google Patents

Abstract

The invention discloses a target tracking method based on a MeanShift algorithm. The method cam improve reliability in the tracking process of a single target, increase tracking speed, and reduce the situation that a tracking frame shifts away from the target. The method comprises the steps: (1) calculating a characteristic value of each pixel point in an original image and generating a characteristic image, and establishing a characteristic model of the target by utilizing a double-core smooth function according to the characteristic image; (2) improving a single core smooth function in an original MeanShift algorithm into the double-core smooth function; (3) in order to ensure convergence of the algorithm, enabling the double-core smooth function to meet the limit condition of a binary convex function; (4) in order to ensure standardization of the characteristic model of the tracked target, obtaining a standardized constant formula.

Description

A kind of method for tracking target based on MeanShift algorithm

Technical field

The invention belongs to the technical field of Digital Image Processing, intelligent video monitoring and machine vision is and in particular to one kind Method for tracking target based on MeanShift algorithm.

Background technology

The basis of machine vision applications is usually to the real-time modeling method of video flowing.The target recognition in later stage, track are divided Analysis and target behavior understand etc. partly all using target following as the process of early stage.

MeanShift as a kind of efficient target pattern matching algorithm, because it does not need exhaustive search, can be adaptive Step-length and direction should be adjusted, be widely used in the target tracking domain higher to requirement of real-time.MeanShift algorithm Target following is modeling and the tracking process of the method performance objective being smoothed using core.It mainly has three key elements to determine tracking The robustness of process：Target scale self adaptation；The separability of target model features；Similarity measure.

First, during the continuous tracking to specific objective, with target movement in three dimensions, the chi of target Degree would generally constantly change.Object module based on constant tracking box wafts larger from the probability of target in the track.Therefore The object module of adaptive scale is an important channel improving tracking effect.

Secondly, target model features and the separability of background characteristics and other target characteristics are also that accuracy is followed the tracks of in impact One condition.Various features are blended and is an up one of method of characteristic model.In addition, the correlation reason based on pattern recognition By by the training of special algorithm, it is possible to achieve automatically choose the good characteristic model of separability.Space is added in characteristic model Information can also improve the separability of itself and local background.For example, the Density Estimator based on a sample characteristics.

MeanShift algorithm is fundamentally based on the similarity of feature in tracking box.In recent years, based on kernel function research On the basis of, the similarity measure algorithm sought becomes the hot issue strengthening MeanShift algorithm.Kernel function is to follow the tracks of framework Basis, it affects similarity to a great extent.When gradient at local extremum for the similarity curved surface improves, then MeanShift convergence of algorithm characteristic is better, and then improves convergence rate on the basis of improving tracking accuracy.

Content of the invention

The technology solve problem of the present invention is：Overcome the deficiencies in the prior art, provide a kind of based on MeanShift algorithm Method for tracking target, it can improve reliability during monotrack and tracking velocity, reduce tracking box and waft from mesh Target situation.

The technical solution of the present invention is：This method for tracking target based on MeanShift algorithm, the method includes Following steps：

（1）Each of original image pixel is calculated with its eigenvalue, generates characteristic image, for characteristic image profit Set up clarification of objective model with double-core smooth function；

（2）Monokaryon smooth function in original MeanShift algorithm is improved to double-core smooth function；

（3）In order to ensure convergence, double-core smooth function is made to meet the qualificationss of binary convex function；

（4）In order to ensure to follow the tracks of the standardization of target characteristic model, obtain new standardized constant formula.

Because the present invention enhances the spatial information of target using double-core function, improve the separability of target and background； Double-core function enhances the gradient in similarity curved surface local extremum region, have updated iterative formula, so that iterative search procedures is restrained Speed faster, more accurately positions extreme value place（I.e. target location）；So reliability is high, processing speed fast, it is to target Situations such as partial occlusion and yardstick convert within the specific limits all has stronger robustness, thus reduce tracking box wafing from target Situation.

Brief description

Fig. 1 realizes the idiographic flow schematic diagram of process for the present invention；

Fig. 2 is the similar contrast schematic diagram of writing music of the corresponding target of two examples；

Fig. 3 is the search iteration frequency curve figure of each frame in video sequence.

Specific embodiment

This method for tracking target based on MeanShift algorithm, the method comprises the following steps：

（1）Each of original image pixel is calculated with its eigenvalue, generates characteristic image, for characteristic image profit Set up clarification of objective model with double-core smooth function；

（2）Monokaryon smooth function in original MeanShift algorithm is improved to double-core smooth function；

（3）In order to ensure convergence, double-core smooth function is made to meet the qualificationss of binary convex function；

（4）In order to ensure to follow the tracks of the standardization of target characteristic model, obtain new standardized constant formula.

Because the present invention enhances the spatial information of target using double-core function, improve the separability of target and background； Double-core function enhances the gradient in similarity curved surface local extremum region, have updated iterative formula, so that iterative search procedures is restrained Speed faster, more accurately positions extreme value place（I.e. target location）；So reliability is high, processing speed fast, it is to target Situations such as partial occlusion and yardstick convert within the specific limits all has stronger robustness, thus reduce tracking box wafing from target Situation.

Preferably, step（1）In obtain its eigenvalue at each pixel of original image first, then use formula （1）Double-core smooth function characteristic model be calculated clarification of objective model,

${\hat{p}}_u\left(y\right)=c_h^{\′}\{\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}k\left({\left|\right|\frac{y-x_i}{h}\left|\right|}^2\right)\·\mathrm{\δ}[b\left(x_i\right)-u\left]\right\}\·\{\underset{j=1}{\overset{n_h}{\mathrm{\Σ}}}l\left({\left|\right|\frac{y-x_j}{h}\left|\right|}^2\right)\·\mathrm{\δ}[b\left(x_j\right)-u\left]\right\}---\left(1\right)$

Wherein, two kernel functions are respectivelyWithB (x) is at x-th point in image Eigenvalue, δ [b (x)-u] is impulse function, and h is the window width of target region, and u is characterized value size, c'_hFor model criteria Change coefficient, n_hFor the pixel number of the target area for h for the window width, y is the center point coordinate of target window, x_iFor target window The coordinate of middle ith pixel point.

Preferably, step（2）The middle iterative formula by original monokaryon smooth function MeanShift algorithm is improved to be based on The MeanShift iterative formula of double-core smooth function（2）

${\hat{y}}_1=\frac{\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_h}{\mathrm{\Σ}}}[k\left({\left|\right|\frac{{\hat{y}}_0-x_i}{h}\left|\right|}^2\right)\·l^{\′}\left({\left|\right|\frac{{\hat{y}}_0-x_j}{h}\left|\right|}^2\right)\·x_j+k^{\′}\left({\left|\right|\frac{{\hat{y}}_0-x_i}{h}\left|\right|}^2\right)\·l\left({\left|\right|\frac{{\hat{y}}_0-x_j}{h}\left|\right|}^2\right)\·x_i]\·w_{\mathrm{ij}}}{\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_h}{\mathrm{\Σ}}}[k\left({\left|\right|\frac{{\hat{y}}_0-x_i}{h}\left|\right|}^2\right)\·l^{\′}\left({\left|\right|\frac{{\hat{y}}_0-x_j}{h}\left|\right|}^2\right)+k^{\′}\left({\left|\right|\frac{{\hat{y}}_0-x_i}{h}\left|\right|}^2\right)\·l\left({\left|\right|\frac{{\hat{y}}_0-x_j}{h}\left|\right|}^2\right)]\·w_{\mathrm{ij}}}---\left(2\right)$

$w_{\mathrm{ij}}=\underset{u=1}{\overset{m}{\mathrm{\Σ}}}\sqrt{\frac{{\hat{q}}_u}{{\hat{p}}_u\left({\hat{y}}_0\right)}}\·\mathrm{\δ}[b\left(x_i\right)-u]\·\mathrm{\δ}[b\left(x_j\right)-u]---\left(3\right)$

Wherein,For the position of target in previous frame image,For the position of target in present frame, w_ijFor two core letters The joint weights of number, it passes through formula（3）Obtain, k'() and l'() be respectively k () and l () first derivative, For clarification of objective model,For coordinateThe characteristic model of place's candidate target, m is characterized the number of value, remaining ginseng Number implication is with claim 2.

Preferably, step（3）In if k (x) l (y) be binary convex function, then this MeanShift algorithm target with Track process can restrain, and iterative process otherwise can be made to dissipate, and leads to target to be lost.

Preferably, step（4）Middle new standardized constant formula is formula（4）

$c_h^{\′}=\frac{1}{\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_h}{\mathrm{\Σ}}}k\left({\left|\right|\frac{y-x_i}{h}\left|\right|}^2\right)\·l\left({\left|\right|\frac{y-x_j}{h}\left|\right|}^2\right)}---\left(4\right).$

Wherein each meaning of parameters is with claim 2.

With reference to Fig. 1, the present invention to implement step as follows：

Step 1, sets up the characteristic model of the double-core smooth function of target in initial frame.

（1a）Target in initial frame, target as to be tracked, need to complete the initialization to target to be tracked, set up Clarification of objective model.First, obtain its sample at each pixel of original gray level image and estimate eigenvalue, secondly Obtain the object module based on double-core smooth function, its formula is：

${\hat{p}}_u\left(y\right)=c_h^{\′}\{\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}k\left({\left|\right|\frac{y-x_i}{h}\left|\right|}^2\right)\·\mathrm{\δ}[b\left(x_i\right)-u\left]\right\}\·\{\underset{j=1}{\overset{n_h}{\mathrm{\Σ}}}l\left({\left|\right|\frac{y-x_j}{h}\left|\right|}^2\right)\·\mathrm{\δ}[b\left(x_j\right)-u\left]\right\}$

Wherein, y is position in initial frame for target's center's point, and kernel function k () chooses Epanechnikov Kernel, its profile function expression formula is

Kernel function l () chooses Uniform Kernel, and its profile function expression formula is

（1b）Because model formation is improved to double-core smooth function, for ensureing the standardization of model, new generalized constant Formula is $c_h^{\′}=\frac{1}{\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_h}{\mathrm{\Σ}}}k\left({\left|\right|\frac{y-x_i}{h}\left|\right|}^2\right)\·l\left({\left|\right|\frac{y-x_j}{h}\left|\right|}^2\right)}$

Wherein, y is position in initial frame for target's center's point.

（1c）In order to ensure the convergence of track algorithm, double-core function need to meet the qualificationss of binary convex function, i.e. double Kernel function k (x) l (y) is necessary for binary convex function.Only double-core function meets this condition, and guarantee target location is repeatedly Can restrain during generation,（1a）The k (x) l (y) of middle selection meets this qualifications.

Step 2, sets up the candidate family of the double-core smooth function of current location in the current frame.

Similar to the foundation of object module in step 1, its difference is candidate family to the computing formula of candidate family Position be present frame current location.

Step 3, obtains pixel weights in the target following window of present frame current location.

The joint weight computing formula of its double-core function is as follows：

$w_{\mathrm{ij}}=\underset{u=1}{\overset{m}{\mathrm{\Σ}}}\sqrt{\frac{{\hat{q}}_u}{{\hat{p}}_u\left({\hat{y}}_0\right)}}\·\mathrm{\δ}[b\left(x_i\right)-u]\·\mathrm{\δ}[b\left(x_j\right)-u]$

Wherein,For object module,For the candidate family of present frame current location,For present frame present bit Put the center position of target, i and j is the sequence number of each of target window pixel.

Step 4, obtains the more new position of present frame target.

Its location updating formula is as follows：

${\hat{y}}_1=\frac{\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_h}{\mathrm{\Σ}}}[k\left({\left|\right|\frac{{\hat{y}}_0-x_i}{h}\left|\right|}^2\right)\·l^{\′}\left({\left|\right|\frac{{\hat{y}}_0-x_j}{h}\left|\right|}^2\right)\·x_j+k^{\′}\left({\left|\right|\frac{{\hat{y}}_0-x_i}{h}\left|\right|}^2\right)\·l\left({\left|\right|\frac{{\hat{y}}_0-x_j}{h}\left|\right|}^2\right)\·x_i]\·w_{\mathrm{ij}}}{\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_h}{\mathrm{\Σ}}}[k\left({\left|\right|\frac{{\hat{y}}_0-x_i}{h}\left|\right|}^2\right)\·l^{\′}\left({\left|\right|\frac{{\hat{y}}_0-x_j}{h}\left|\right|}^2\right)+k^{\′}\left({\left|\right|\frac{{\hat{y}}_0-x_i}{h}\left|\right|}^2\right)\·l\left({\left|\right|\frac{{\hat{y}}_0-x_j}{h}\left|\right|}^2\right)]\·w_{\mathrm{ij}}}$

Wherein, W_ijValue obtained by step 3.

According to step 1（a）In two kernel functions definition, have

$k^{\′}\left({\left|\right|\frac{y-x_i}{h}\left|\right|}^2\right)=-1$ With $l^{\′}\left({\left|\right|\frac{y-x_i}{h}\left|\right|}^2\right)=0.$

Therefore, iterative formula can be reduced to

$y_1=\frac{\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_h}{\mathrm{\Σ}}}[x_i\·w_{\mathrm{ij}}]}{\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_h}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}}$

Step 5, judges the new target location obtaining in step 4 whether inside image, if inside image-region Then continuing step 6, if exceeding image-region, carrying out step 7.

Step 6, judges in present frame the need of stopping iteration more new target location.Then carry out if necessary to stop updating Step 7, if not needing to stop updating, carries out step 2.

Its judgment criterion is：The distance of the current position updating and position before is more than threshold value and then continues to update, and is less than Threshold value then stops updating.

Step 7, by the target location output of present frame, and is identified in current frame image, draws tracking box.

Step 8, judges whether present frame is last frame, then terminates if last frame, if not last frame Then it is loaded into next frame and carry out step 2.

Below in conjunction with accompanying drawing and specific embodiment, the present invention is described in further detail.

Fig. 2 gives the similarity surface chart of video sequence in two embodiments.In Fig. 2（a）In, contrast similarity curved surface The gradient of in figure extreme point local neighborhood, the MeanShift track algorithm of double-core smooth function is apparently higher than original MeanShift Algorithm；And in Fig. 2（b）In, also can be evident from identical conclusion.Higher gradient mean algorithm can faster, more accurate Converge at extreme point.Which obviate search iteration process and stop near gentle extreme value vertex neighborhood and lead to obtain Accurately extreme point position（I.e. target location）.

Fig. 3 gives the curve chart of single frames iterative search number of times.The reality of its experimentation and Dorin Comaniciu et al. Test process identical：With football video sequence as experimental subject, one two field picture size is 352 × 240 pixels, is transported with No. 75 Mobilize as following the tracks of target.With regard to experimental result, its maximum iteration time about 8 times/every frame, mean iterative number of time is 3.076 times/every Frame, and the experimental result of Dorin Comaniciu et al. is 4.19 times/every frame.This experimental result display inventive algorithm is far excellent In original MeanShift algorithm.

The above, be only presently preferred embodiments of the present invention, and not the present invention is made with any pro forma restriction, every according to Any simple modification, equivalent variations and modification above example made according to the technical spirit of the present invention, all still belongs to the present invention The protection domain of technical scheme.

Claims (4)

1. a kind of method for tracking target based on MeanShift algorithm, the method comprises the following steps：

(1) each of original image pixel is calculated with its eigenvalue, generates characteristic image, for characteristic image using double Core smooth function sets up clarification of objective model；

(2) the monokaryon smooth function in original MeanShift algorithm is improved to double-core smooth function；

(3) in order to ensure convergence, double-core smooth function is made to meet the qualificationss of binary convex function；

(4) in order to ensure to follow the tracks of the standardization of target characteristic model, obtain new standardized constant formula；It is characterized in that, step (1) obtain its eigenvalue in first at each pixel of original image, then use the double-core smooth function feature of formula (1) Model is calculated clarification of objective model,

${\hat{p}}_u\left(y\right)=c_h^{\′}\{\underset{i=1}{\overset{n_h}{\Σ}}k\left(\right|\left|\frac{y-x_i}{h}\right|{|}^2)\·\mathrm{\δ}\[b\left(x_i\right)-u\]\}\·\{\underset{j=1}{\overset{n_h}{\Σ}}l\left(\right|\left|\frac{y-x_j}{h}\right|{|}^2)\·\mathrm{\δ}\[b\left(x_j\right)-u\]\}---\left(1\right)$

Wherein, two kernel functions are respectivelyWithB (x) is the eigenvalue at x point in image, δ [b (x)-u] is impulse function, and h is the window width of target region, and u is characterized value size, c'_hFor model standardization coefficient, n_h For the pixel number of the target area for h for the window width, y is the center point coordinate of target window, x_iFor i-th picture in target window The coordinate of vegetarian refreshments.

2. the method for tracking target based on MeanShift algorithm according to claim 1 is it is characterised in that in step (2) The iterative formula of original monokaryon smooth function MeanShift algorithm is improved to the MeanShift based on double-core smooth function Iterative formula (2)

${\hat{y}}_1=\frac{\underset{i=1}{\overset{n_h}{\Σ}}\underset{j=1}{\overset{n_h}{\Σ}}\[k\left(\right|\left|\frac{{\hat{y}}_0-x_i}{h}\right|{|}^2)\·l^{\′}\left(\right|\left|\frac{{\hat{y}}_0-x_j}{h}\right|{|}^2)\·x_j+k^{\′}\left(\right|\left|\frac{{\hat{y}}_0-x_i}{h}\right|{|}^2)\·l\left(\right|\left|\frac{{\hat{y}}_0-x_j}{h}\right|{|}^2)\·x_i\]\·w_{ij}}{\underset{i=1}{\overset{n_h}{\Σ}}\underset{j=1}{\overset{n_h}{\Σ}}\[k\left(\right|\left|\frac{{\hat{y}}_0-x_i}{h}\right|{|}^2)\·l^{\′}\left(\right|\left|\frac{{\hat{y}}_0-x_j}{h}\right|{|}^2)+k^{\′}\left(\right|\left|\frac{{\hat{y}}_0-x_i}{h}\right|{|}^2)\·l\left(\right|\left|\frac{{\hat{y}}_0-x_j}{h}\right|{|}^2)\]\·w_{ij}}---\left(2\right)$

$w_{ij}=\underset{u=1}{\overset{m}{\Σ}}\sqrt{\frac{{\hat{q}}_u}{{\hat{p}}_u\left({\hat{y}}_0\right)}}\·\mathrm{\δ}\[b\left(x_i\right)-u\]\·\mathrm{\δ}\[b\left(x_j\right)-u\]---\left(3\right)$

Wherein,For the position of target's center's point in previous frame image,The position of the target's center's point for estimating in present frame, w_ijFor the joint weights of two kernel functions, it is obtained by formula (3), k'() and l'() it is respectively k () and l () First derivative,For clarification of objective model,For coordinateThe characteristic model of place's candidate target, m is characterized the individual of value Number.

3. the method for tracking target based on MeanShift algorithm according to claim 2 is it is characterised in that in step (3) If k (x) l (y) is binary convex function, this MeanShift algorithm can be restrained in object tracking process, otherwise can make to change Dissipate for process, lead to target to be lost.

4. the method for tracking target based on MeanShift algorithm according to claim 3 is it is characterised in that in step (4) New standardized constant formula is formula (4)

$c_h^{\′}=\frac{1}{\underset{i=1}{\overset{n_h}{\Σ}}\underset{j=1}{\overset{n_h}{\Σ}}k\left(\right|\left|\frac{y-x_i}{h}\right|{|}^2)\·l\left(\right|\left|\frac{y-x_j}{h}\right|{|}^2)}---\left(4\right).$