CN101551909B - Tracking method based on kernel and target continuous adaptive distribution characteristics - Google Patents

Abstract

The invention relates to a tracking method based on kernel and target continuous adaptive distribution characteristics in the technical field of computer vision. In mean shift self-frame, the invention is combined with adaptive distribution of target area and relatively stable target module, a candidate module is mapped in reverse direction through a column diagram so that the length of a trackingwindow is estimated by the zeroth moment of the distribution image of the tracking area. At last, in updating period, the adaptive updating of the target module and the adaptive updating of trackingwindow length are realized simultaneously according to the similarity of the target module and the candidate module. The target length and the tracking in the condition that greater changes appear canbe realized in dynamic environment.

Description

Tracking based on nuclear and target continuous adaptive distribution characteristics

Technical field

What the present invention relates to is a kind of tracking of technical field of computer vision, particularly a kind of tracking based on nuclear and target continuous adaptive distribution characteristics.

Background technology

The visual object tracking technique can be divided at present: the tracking of data-driven and the tracking of model-driven.The visual object tracking of data-driven often adopts the algorithm of local optimum, finds the gradient information of similarity to determine the direction (real-time is good) of target travel, and in fact, this method is to become tracking problem to be the local optimal searching problem; Perhaps directly carry out Block Matching Algorithm and follow the tracks of (real-time is poor), but utilize multiple dimensioned decomposition method, adopt, can improve the real-time of Block Matching Algorithm by thick tracking thinking to essence.More popular model driven method then is a motion state of utilizing dynamic motion model at random to come target of prediction, utilizes observation information to upgrade model state to determine the position of target then, and this method is called filtering method again.Because nonlinear often relation between the motion state of observation model and target is so the visual object tracking problem has just become a nonlinear filtering problem.The common ground of above-mentioned two class methods is exactly all to have utilized topographical view's measurement information to follow the tracks of.Also having another kind of method is that the global detection algorithm that depends on static state or dynamic image is realized following the tracks of.

Find by prior art documents, Comaniciu etc. are at " IEEE Transactions onPattern Analysis and Machine Intelligence " (pp.564-577,2003) deliver " Kernel based object tracking " (based on the target following of nuclear, " IEEE pattern analysis and machine intelligence magazine ") on.(brief note is: KBT) be called mean shift (average drifting) again and follow the tracks of based on the target following of examining.Mean shift method will be examined the description of histogram as target signature, promptly adopt the pixel color of target area or the intensity profile describing mode as characteristics of image.Mean shift is followed the tracks of the field as a kind of method for mode matching efficiently by the visual object that being applied in of success had relatively high expectations to real-time, has become the representative of the visual object tracking of data drive type.

In general, visual object is followed the tracks of and be can be described as: at initial frame, the feature of extracting the target area is as trace template, and in subsequent video frame, the image-region the most similar to the trace template coupling becomes the target following result.For the KBT method, the To Template feature of extracting in initial frame is called " object module ", and the feature that is used to mate in subsequent video frame is called " candidate's model ".

In the tracking application of reality, the variation of scale size, shape and attitude etc. will appear in target, add that various other factors in the environment are disturbed, and relevant matches is followed the tracks of and be can not get best matched position, certainly exists measuring error.When following the tracks of degree of confidence when dropping to threshold value less than template renewal, bigger drift has taken place in template.In addition, when the brightness and contrast of image changed and gray inversion occurs, the position of optimal match point will change with the variation of gray scale, and tracking error increases gradually.So, in tracing process, must in time obtain upgrading and revise as the template image of similarity measurement benchmark in the correlation tracking process.

Mean Shift method is the nonparametric technique that a kind of density function gradient is estimated, finds the extreme value of probability distribution to come localizing objects by the iteration optimizing.For mean shift tracking, object module can be regarded as trace template.When the tracking window is sneaked into a lot of background area gradually when comprising the target area, owing to follow the tracks of the centre of form that target is being indicated at the center of window all the time, so the space orientation error is very little.But when the continuous increased in size of target and when following the tracks of the yardstick of window, not only can cause the yardstick deviations, also cause the space orientation deviation.On the other hand, when the continuous minification of target,, finally must cause great yardstick positioning error because it is constant to follow the tracks of window size.The content that therefore, should comprise two aspects for the tracking masterplate self-adaptation of mean shift:

(1) object module adaptive updates;

(2) following the tracks of the window dimension self-adaption upgrades;

Cause the positioning error accumulation for fear of template renewal is untimely, and, must design effective template renewal and correction strategy to obtain sane tracking performance to follow-up tracing process transmission.

CAMSHIFT (the Continuously adaptive mean shift) method that Bradski G R proposes is used for the face tracking of perception user interface, it adopts the H component of HSV color space to set up the target histogram model, is limited to the target of following the tracks of particular color.It should be noted that to be different from KBT that the continuous adaptive distribution that the CAMSHIFT method is based on target realizes target following.The CAMSHIFT method can be regulated window size automatically to adapt to the size of tracked target in image.By comparative analysis KBT and CAMSHIFT advantage and shortcoming separately, KBT lacks necessary model modification as can be seen, its changeless kernel function window width, both followed the tracks of the big or small constant of window, influenced the accuracy of following the tracks of, when there is obvious dimensional variation in target, can causes the yardstick location inaccurate, even cause losing of target.And therefore the CAMSHIFT method under the extended background disturbed condition, can not realize effective tracking owing to lack a metastable object module as template.The present invention is intended to the static model and the dynamic continuous adaptive distribution advantage separately of combining target, obtains more sane tracking effect.

Summary of the invention

The objective of the invention is to the deficiency that exists at prior art, proposes a kind of tracking of the continuous adaptive distribution characteristics based on nuclear and target area, can be implemented in the dynamic scene, target scale and apparent generation be the tracking under the change condition greatly.

The present invention is achieved through the following technical solutions, the present invention is based on the target following of nuclear, in average drifting tracking self framework, the self-adaptation of combining target tracing area distributes and metastable object module, to candidate's model through behind the histogram to mapping, estimate to follow the tracks of the yardstick of window with the zeroth order square of tracing area distributed image.In update stage,, realized the adaptive updates of object module simultaneously and followed the tracks of the adaptive updates of window size at last according to the similarity of object module and candidate's model.

The present invention includes following steps:

1, tracking initiation: determine the target following zone at initial frame, that is: determine the initial parameter of target location and yardstick.In initial target following zone, extract To Template feature, i.e. object module.

2, after the tracking beginning,, extract the nuclear histogram feature of target, i.e. candidate's model in the candidate region.According to KBT algorithm computation mean shift vector, determine the position of target.With the target location that newly obtains is the parameter update tracing area, and extracts the nuclear histogram feature of target in the tracing area after renewal again.

3, be distribution with the target nuclear histogram that extracts again, tracing area is carried out behind the histogram to mapping transformation (HBP:Histogram Back-Projection).Can obtain the distributed image of tracing area by HBP.

4, calculate the zeroth order square of tracing area distributed image.According to the vertical scale of following the tracks of window and the scale parameter of breadth wise dimension, utilize the zeroth order square to be refreshing weight, calculate new tracking window scale parameter.

5, according to the location parameter of new tracking window scale parameter and the 2nd step gained, upgrade tracing area, and, extract target nuclear histogram feature as new candidate's model at tracing area.Calculate the similarity degree between object module and the new candidate's model: being higher than the similarity of former frame as if current similarity degree, is weights with current similarity then, and object module is weighted renewal.Simultaneously, use the scale parameter of the 4th step gained to upgrade the tracking window size.

After the present invention follows the tracks of beginning, at first extract candidate's model of target in the candidate region.Determine new target location parameter and upgrade tracing area by the KBT method with this.With new tracing area target nuclear histogram serves as to distribute to carry out behind the histogram to mapping transformation, thereby obtains the distributed image of new tracing area.Calculate the zeroth order square of new tracing area distributed image then, utilize the zeroth order square to be refreshing weight, calculate new tracking window scale parameter.Again extract new candidate's model according to new tracking window scale parameter and new target location parameter.At last, calculate the similarity between new candidate's model, candidate's model and object module, by the adaptive updates of relatively realizing object module of similarity and the adaptive updates of tracking window size.

The present invention has proposed a kind of visual tracking that distributes in conjunction with continuous adaptive and has had following beneficial effect in mean shift technological frame:

In the KBT method, at each width of cloth tracking frame, candidate's model all needs to be recomputated.Therefore the present invention directly adopts candidate's model to distribute as continuous adaptive.Because the tracing area distributed image is set up according to target nuclear histogram, therefore in tracing area, the pixel that the belongs to target demonstration that can be enhanced.But not object pixel can be suppressed, and therefore as long as object module can access in good time renewal and can keep higher similarity with candidate's model, the tracking performance of method just can be greatly improved.By calculating the zeroth order square of tracing area distributed image, can make full use of the information that candidate's model comprises, estimate the scale parameter of following the tracks of window simultaneously.In update stage, judge the condition of carrying out of upgrading according to candidate's model and object module similarity, original static object module in the KBT algorithm is transformed for the dynamic object model, make the renewal operation of yardstick renewal and object module have higher robustness and adaptability.Because the present invention realizes in mean shift technological frame, therefore kept the advantage that traditional KBT method is easy to realize, travelling speed is fast, have higher utility.

Description of drawings

Fig. 1 is the tracking synoptic diagram that the present invention is based on nuclear and target continuous adaptive distribution characteristics.

Fig. 2 is the comparison synoptic diagram of the present invention and KBT method and CAMSHIFT method tracking effect.

Fig. 3 is similarity comparative graph between candidate's model of the present invention and KBT method and the object module.

Embodiment:

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

As shown in Figure 1, present embodiment is at definite tracing area, after the extraction object module is trace template, operation KBT tracking obtains the target location, the result who utilizes KBT to obtain upgrade tracing area for the candidate region after, extract nuclear histogram feature, i.e. candidate's model in the candidate region.Based on candidate's model, the candidate region is carried out can obtaining this regional distributed image to map operation behind the histogram.Calculate the zeroth order square of this areal distribution image.With the update module on right side among candidate's model and the zeroth order square input figure, thereby the renewal operation of window size and object module is followed the tracks of in realization, makes tracking possess adaptive characteristic.

Present embodiment comprises the steps:

1, determines target initial position parameters:＜x _c ^K-1, y _c ^K-1And the initial gauges parameter: (S _x ^K-1, S _y ^K-1).

${\mathrm{TR}}^{k-1}=\mathrm{RECT}(x_c^{k-1},y_c^{k-1},S_x^{k-1},S_y^{k-1})---\left(1\right)$

Wherein, the frame number that subscript k representative is followed the tracks of, during initialization, k=1; (1) in the formula, (x _c ^K-1, y _c ^K-1, S _x ^K-1, S _y ^K-1) the center parameter of the initial selected target of representative, and target is at the tracking window scale parameter of x coordinate direction and y coordinate direction. ${\mathrm{TR}}^{k-1}=\mathrm{RECT}(x_c^{k-1},y_c^{k-1},S_x^{k-1},S_y^{k-1})$ Representative is by (x _c ^K-1, y _c ^K-1, S _x ^K-1, S _y ^K-1) rectangle that constitutes of parameter follows the tracks of window.

In initial tracing area, extract the To Template feature, i.e. object module:

Wherein, $q_u=\frac{1}{C_h}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{Kernel}(X_i-c^0)\·\mathrm{\δ}(b\left(X_i\right),u)---\left(2\right)$

According to the definition of KBT method, the initial center position of target note is done: c ⁰{ X _i} _{I=1 ..n}Be each location of pixels in the tracing area, Kernel is the weighting kernel function to the tracing area pixel value.δ () is the Kroneckerdelta function.B (X _i) be to finish with the function of each pixel to each respective components mapping of histogram.C _hBe normaliztion constant, in order to guarantee $\underset{u=1}{\overset{m}{\mathrm{\Σ}}}{q}_u=1$ Set up.The subscript of u representative nuclear histogram component.Similar with the definition of object module, candidate's model definition is following form:

Wherein, $p_u=\frac{1}{C_h}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{Kernel}(X_i-c^k)\·\mathrm{\δ}(b\left(X_i\right),u).---\left(3\right)$

2, after the tracking beginning, in current k frame, originally the tracing area of k-1 frame becomes the candidate region, extracts the nuclear histogram feature of target according to (3) in the candidate region, i.e. candidate's model.Calculate mean shift vector, can carry out according to following formula iteration:

Determine the target location: Δ c ^k=(Δ x ^k, Δ y ^k) (4)

Iterative formula is: $\mathrm{\Δ}{c}^k=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{Kernel}(X_i-c^k)\·w\left(X_i\right)\·(X_i-c^k)}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{Kernel}(X_i-c^k)\·w\left(X_i\right)}---\left(5\right)$

Wherein: $w\left(X_i\right)=\sqrt{\frac{q_u}{p_u\left(c^k\right)}}---\left(6\right)$

Use the second step result to upgrade and follow the tracks of the candidate region, be expressed as:

${\mathrm{CR}}^k=\mathrm{RECT}(x_c^{k-1}+{\mathrm{\Δx}}^k,y_c^{k-1}+{\mathrm{\Δy}}^k,S_x^{k-1},S_y^{k-1})---\left(7\right)$

Wherein, Δ c ^k=(Δ x ^k, Δ y ^k) representative target's center's position translation amount of adopting the average drifting algorithm to obtain.

$\mathrm{\Δ}{c}^k=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{Kernel}(X_i-c^k)\·w\left(X_i\right)\·(X_i-c^k)}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{Kernel}(X_i-c^k)\·w\left(X_i\right)}$ Represent the computing formula of average drifting algorithm.

${\mathrm{CR}}^k=\mathrm{RECT}(x_c^{k-1}+{\mathrm{\Δx}}^k,y_c^{k-1}+{\mathrm{\Δy}}^k,S_x^{k-1},S_y^{k-1})$ New target's center position is used in representative

Translational movement upgrades the result who obtains behind the target area.

Next, be the parameter update tracing area with the target location that newly obtains, and in the tracing area after renewal according to (3), extract the nuclear histogram feature of target again.

3, be distribution with the target nuclear histogram that extracts again, tracing area is carried out behind the histogram to mapping transformation (HBP:Histogram Back-Projection).Mapping mode is as follows:

$\hat{p}={\left[{\hat{p}}_u\right]}_{u=1...m}$ Wherein, ${\hat{p}}_u=\min {(\frac{255}{\max \left(\stackrel{V}{p}\right)}{p}_u,255)}_{u=1...m}---\left(8\right)$

According to the definition of (8), histogram component p _uSpan from

Be mapped to [0,255]. can obtain the distributed image of tracing area by HBP.

4, calculate the zeroth order square of tracing area distributed image: M ₀₀Computing formula is as follows:

$M_{00}=\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}I(x,y)---\left(9\right)$

Wherein, (x is that the distributed image of tracing area is at (x, y) gray-scale value of position through behind the HBP y) to I.

Utilize zeroth order square M ₀₀Be refreshing weight, calculate new tracking window scale parameter:

$S_x=\mathrm{\α}\·\sqrt{M_{00}}---\left(10\right)$

$S_y=\mathrm{\β}\·S_x^k---\left(11\right)$

Wherein, α, for complete coverage goal, follows the tracks of window size and need be slightly larger than target scale (α=1.1) in the practical application for following the tracks of the ratio of window size and target scale, and β is the vertical scale of tracking window and the scale parameter of breadth wise dimension.

5, according to new tracking window scale parameter (S _x, S _y) and the 2nd the step gained location parameter, upgrade current candidate region:

${\mathrm{CR}}^k=\mathrm{RECT}(x_c^{k-1}+{\mathrm{\Δx}}^k,y_c^{k-1}+{\mathrm{\Δy}}^k,S_x,S_y)---\left(12\right)$

Extract target nuclear histogram feature as new candidate's model according to (3).

Calculate the similarity degree between object module and the candidate's model:

Wherein:

Be the object module of previous frame and the Bhattacharyya coefficient between candidate's model,

Be the object module of previous frame and the Bhattacharyya coefficient between current candidate's model, computing formula is as follows:

$B(\stackrel{V_{k-1}}{q},\stackrel{V_{k-1}}{p})=\underset{u=1}{\overset{m}{\mathrm{\Σ}}}\sqrt{q_u^{k-1}\·p_u^{k-1}}$ $B(\stackrel{V_{k-1}}{q},\stackrel{V_k}{p})=\underset{u=1}{\overset{m}{\mathrm{\Σ}}}\sqrt{q_u^{k-1}\·p_u^k}$

According to the Bhattacharyya coefficient ratio, specifically renewal process is as follows:

If $B(\stackrel{V_{k-1}}{q},\stackrel{V_{k-1}}{p})<B(\stackrel{V_{k-1}}{q},\stackrel{V_k}{p})$

Then upgrade yardstick: $S_x^k=S_x,$ $S_y^k=S_y$

Upgrade object module: $\stackrel{V_k}{q}=B(\stackrel{V_{k-1}}{q},\stackrel{V_k}{p})*\stackrel{V_k}{p}+(1-B(\stackrel{V_{k-1}}{q},\stackrel{V_k}{q})*\stackrel{V_{k-1}}{q}$

Otherwise keep tracking window size and object module constant:

$S_x^k=S_x^{k-1},$ $S_y^k=S_y^{k-1},$ $\stackrel{V_k}{q}=\stackrel{V_{k-1}}{q}$

Upgrade and finish.

As shown in Figure 2, (a)-(c) is the tracking effect of traditional KBT method among Fig. 2, can't obtain suitable renewal because it follows the tracks of window size, and when target scale became big, tracking performance is variation progressively.(d)-(f) is the tracking effect of CAMSHIFT method among Fig. 2.Though CAMSHIFT can regulate window size automatically, but because the strong interference of dynamic scene lacks a relatively-stationary trace template again, the tracking window size that CAMSHIFT obtains is often excessive, thereby comprised too much background pixel, finally caused tracking performance to descend.(g)-(i) is the tracking effect of present embodiment among Fig. 2, and therefrom as can be seen, the tracking that present embodiment proposes has been realized in dynamic scene, the tracking under the big change condition of target scale and apparent generation.

As shown in Figure 3, the asterisk line has been represented the tracking results of present embodiment proposition method and the similarity between To Template; Solid line has been represented the tracking results of KBT method and the similarity between To Template.

In the present embodiment,, guaranteed that target can be locked all the time fully, therefore, can keep better consistance, realized sane tracking target with the To Template feature owing to introduced special target scale and model update method.

Claims (3)

1. one kind based on the nuclear and the tracking of target continuous adaptive distribution characteristics, it is characterized in that, based on the nuclear method for tracking target, in average drifting tracking self framework, the self-adaptation in combining target zone distributes and metastable object module, to candidate's model through behind the histogram to mapping, estimate to follow the tracks of the yardstick of window with the zeroth order square of tracing area distributed image, at last in update stage, similarity degree according to object module and candidate's model, realized the adaptive updates of object module simultaneously and followed the tracks of the adaptive updates of window size, described tracking specifically comprises the steps:

1., determine the target following zone at initial frame, extraction To Template feature;

2., follow the tracks of beginning after, in the candidate region, extract the nuclear histogram feature of target, determine the location parameter of target, be the parameter update tracing area with the target location that newly obtains, and extract the nuclear histogram feature of target in the tracing area after renewal again;

3., be distribution with the target nuclear histogram that extracts again, tracing area is carried out behind the histogram obtaining the distributed image of tracing area by mapping transformation to mapping transformation;

4., calculate the zeroth order square of tracing area distributed image, according to the vertical scale of following the tracks of window and the scale parameter of breadth wise dimension, utilize the zeroth order square to be refreshing weight, calculate new tracking window scale parameter;

5., 2. go on foot the location parameter of gained according to new tracking window scale parameter and, upgrade tracing area, and at tracing area, extract target nuclear histogram feature as new candidate's model, calculate the similarity degree between object module and the new candidate's model: being higher than the similarity of former frame as if current similarity degree, is weights with current similarity then, and object module is weighted renewal, simultaneously, use the scale parameter that 4. goes on foot gained to upgrade and follow the tracks of window size.

2. according to claim 1ly it is characterized in that, describedly determine the target following zone, be meant: the initial parameter of determining target location and yardstick at initial frame based on the nuclear and the tracking of target continuous adaptive distribution characteristics.

3. the tracking based on nuclear and target continuous adaptive distribution characteristics according to claim 1, it is characterized in that, described is weights with current similarity, object module is weighted renewal, simultaneously, use the scale parameter that 4. goes on foot gained to upgrade and follow the tracks of window size, be meant: if the similarity degree of present frame candidate model and object module is higher than the similarity degree of former frame candidate model and object module, be that refreshing weight is upgraded object module and tracking window then, otherwise keep object module constant with the tracking window size with the zeroth order square.

Images