CN101493889B - Method and apparatus for tracking video object - Google Patents

Abstract

The embodiment of the invention provides a method for tracking a video object and a device thereof, and relates to the technical field of image processing. The method and the device are invented for realizing the accurate tracking of the video object. The method comprises the steps of: picking feature points of the profile of the video object in a current frame of image; finding out matched feature points; detecting at least one alternative profile of the video object in a next frame of image; calculating the profile feature value of the video object in the current frame of image; calculating the profile feature value of the alternative profile; and comparing the profile feature value of the alternative profile with that of the video object in the current frame of image and taking the alternative profile as the profile of the video object in next frame of image if the two profile feature values are matched. The method and the device can improve the accuracy of tracking the video object.

Description

Method and device that object video is followed the tracks of

Technical field

The present invention relates to technical field of image processing, relate in particular to a kind of method and device that object video is followed the tracks of.

Background technology

Computer vision is meant with equipment such as video camera and computing machines and replaces human eye tracking target to be discerned, followed the tracks of and measurement etc.Wherein, the real time technique for tracking of object video is an important topic of computer vision field, and it is the basis of a series of activities such as video analysis, video understanding, object video identification, object video behavioural analysis.

At present, the method for object video tracking has a variety of.According to whether needs carry out pattern match in each interframe of image, the method that object video is followed the tracks of can be divided into based on the method that object video is followed the tracks of that detects with based on the method that object video is followed the tracks of of identification.

The method that object video is followed the tracks of based on detecting is directly to extract the profile of object video according to a certain feature of object video in each two field picture, and need carry out the transmission of object motion state parameter in each interframe of image, outline etc.Comprise the method for Differential Detection etc. based on the method that object video is followed the tracks of that detects.Based on the method that object video is followed the tracks of of identification, normally at first extract certain feature of object video, in each two field picture, hunt out the zone of mating the most then with described feature, the described zone of coupling the most is object video.

In above-mentioned two kinds of methods that object video is followed the tracks of, simple based on its algorithm of method that object video is followed the tracks of that detects, realize easily, but tracking effect is undesirable.Therefore the main direction of studying of current object video tracking technique has been transferred to the method based on identification.

In the method that object video is followed the tracks of based on identification, Condensation (ConditionalDensity propagation) track algorithm, promptly the conditional probability density propagation algorithm is a kind of contour tracing method that is most widely used.

The Condensation track algorithm is based on a kind of in the track algorithm of particle filter.Particle filter claims that again (Sequential Monte Carlo SMC), is a kind of method that realizes Bayes's Recursive Filtering with Monte Carlo method to sequential Monte Carlo method.It represents the posterior probability density p (x of system state vector with one group of random sample (particle) that has a weight _k| x _1:k), when sample size was abundant, this probability estimation was equal to the posterior probability density function.

At the Condensation track algorithm, be to adopt movable contour model and shape space to represent the profile characterizing method of object video, it characterizes the possible variation of described contour curve with the contour curve of the reference mark sign object video of B-Snake with shape space, as translation, rotation etc.

The motion state parameters T of object video profile can be expressed as: T=(TX, TY, θ, SX, SY), wherein TX and TY are respectively the central points of object video x direction and y direction, θ is the angle that the object video profile is rotated, and SX and SY are respectively the yardstick of object video in x direction and y direction.The shape space parameter S of object video in shape space is expressed as: S=(TX, TY, SX cos θ-1, SY cos θ-1 ,-SY sin θ, SX sin θ).

It is as follows to utilize the Condensation track algorithm to carry out the process that object video follows the tracks of.

1) obtains the state initial value T that object video moves by the initial frame image ₀, initialization N _sIndividual particle, initial weight

Be 1/N _s, the motion state of each particle and shape space parameter are respectively (i=1,2 ... N _s).When the k frame, each particle state is carried out state transitions.State transition equation is as shown in Equation (11):

TX _k ⁱ＝TX _k-1 ⁱ+B ₁×ξ _1-k ⁱ；

TY _k ⁱ＝TY _k-1 ⁱ+B ₂×ξ _2-k ⁱ；

θ _k ⁱ＝θ ⁱ _k-1+B ₃×ξ _3-k ⁱ；(11)

SX _k ⁱ＝SX ⁱ _k-1+B ₄×ξ _4-k ⁱ；

SY _k ⁱ＝SY ⁱ _k-1+B ₅×ξ _5-k ⁱ；

Wherein, B ₁, B ₂, B ₃, B ₄, B ₅Be constant, ξ is the random number of [1,1].

2) utilize the observed reading (motion state parameters T, shape space parameter S etc.) of current frame image that each candidate's particle is assessed, calculate the weighted value of each particle.

Detailed process is as follows:

21) to particle N ⁱ, calculate its motion state parameters T according to the method in the formula (1) ⁱWith the shape space parameter S ⁱ

22) according to described motion state parameters T ⁱWith the shape space parameter S ⁱTry to achieve particle N ⁱThe reference mark of B-Snake, simulate the contour curve of object video by the reference mark of described B-Snake.

23) N sampling point of sampling on the contour curve of object video tried to achieve the pixel of each sampling point gradient maximum on normal direction.

24) try to achieve each sampling point on the described contour curve, and the distance D IS between the pixel of gradient maximum on this sampling point normal ⁱ(n), (n=1,2 ... N), obtain particle N as the measurement factor with this ⁱThe observation probability density function And to particle N ⁱRight value update, its weights

Computing formula as follows:

Motion state parameters T and weights by each particle

Be weighted summation, obtain the motion state parameters of the expectation of each particle, so can be regarded as the intended shape spatial parameter S of object video _k, B-Snake reference mark and contour curve.So far just finished once tracing process to the object video contour of object.

In realizing process of the present invention, the inventor finds to exist in the prior art such problem:

The Condensation track algorithm can be realized the real-time follow-up to the object video profile of affine variation (as rotation, translation, scaling etc.).For example when object video is rigid body,, therefore can realize tracking accurately to described rigid body by the Condensation track algorithm because described rigid body can not produce separation between its each ingredient in motion process.But the object video that changes for nonaffine, as human body in the process of walking, when situation such as arm bending occurring, the Condensation track algorithm can not accurately be followed the tracks of.In addition, because the calculated amount complexity of Condensation track algorithm, so the Condensation track algorithm follows the tracks of object video, and tracking velocity is lower.

Summary of the invention

In order to solve the problem of prior art and poor accuracy low to the object video tracking velocity, embodiments of the invention provide a kind of method and apparatus that object video is followed the tracks of.

On the one hand, embodiments of the invention provide a kind of method that object video is followed the tracks of, and described method comprises the steps:

Get the unique point of the profile of object video described in the current frame image;

In the next frame image, find matching characteristic point with described Feature Points Matching;

According to described matching characteristic point, in the next frame image, detect at least one candidate's profile of described object video;

Calculate the contour feature value of the object video in the current frame image;

Calculate the contour feature value of described candidate's profile;

The contour feature value of described candidate's profile and the contour feature value of the object video in the current frame image are compared, if the two coupling, then described candidate's profile is the profile of described object video in the next frame image.

By the described method of the embodiment of the invention, at first determine the matching characteristic point of object video in the next frame image in the present frame, detect the candidate profile of described object video in the next frame image according to described matching characteristic point then.Then the contour feature value of the object video in forward and backward two two field pictures is mated, if the two coupling, then described candidate's profile is the profile of described object video in the next frame image.When object video carries out the nonaffine variation, owing to can extract the contour feature value of described object video in the next frame image, and by mating with the contour feature value of described object video in current frame image, obtain with current frame image in the contour feature value contour feature value of mating the most, describe thereby can make accurately the profile of described object video in the next frame image.Utilize the described method of the embodiment of the invention, having overcome in the prior art can not be to having the defective that object video that nonaffine changes is accurately followed the tracks of.In addition, utilize the described method of the embodiment of the invention, reduced object video is carried out operand in the tracing process, improved the speed of following the tracks of.

Therefore, the method that the embodiment of the invention is followed the tracks of object video, not only can follow the tracks of object video accurately, and can follow the tracks of, improve the accuracy that object video is followed the tracks of object video with nonaffine variation with affine variation.

On the other hand, embodiments of the invention provide a kind of device that object video is followed the tracks of, and described device comprises:

First positioning unit is used to obtain the unique point of the profile of object video described in the current frame image;

Second positioning unit is used for finding matching characteristic point with described Feature Points Matching at the next frame image;

The profile detecting unit, be used for according to described matching characteristic point, in the next frame image, detect at least one candidate's profile of described object video, comprise: the regional prediction module, being used for described matching characteristic point is the center, obtains the appearance zone of described object video profile in the next frame image by linear transformation; Profile is chosen module, is used for detecting in described appearance zone at least one candidate's profile of described object video;

First computing unit is used for calculating the contour feature value of the object video of current frame image;

Second computing unit is used to calculate the contour feature value of described candidate's profile;

The outline unit is used for the contour feature value of the object video of the contour feature value of described candidate's profile and current frame image is compared, if the two coupling, then described candidate's profile is the profile of described object video in the next frame image.

By the described device of the embodiment of the invention, at first determine the candidate profile of described object video in the next frame image by the profile detecting unit, first, second computing unit calculates the contour feature value of object video in forward and backward two two field pictures respectively, by the outline unit two contour feature values are mated, when object video carries out the nonaffine variation, also can obtain accurate description by the contour feature value of object video in two two field pictures before and after the coupling to described object video profile.When object video carries out the nonaffine variation, owing to can extract the contour feature value of described object video in the next frame image, and by mating with the contour feature value of described object video in current frame image, obtain with current frame image in the contour feature value contour feature value of mating the most, describe thereby can make accurately the profile of described object video in the next frame image.Utilize the described device of the embodiment of the invention, having avoided in the prior art can not be to having the defective that object video that nonaffine changes is accurately followed the tracks of.In addition, utilize the described device of the embodiment of the invention, reduced object video is carried out operand in the tracing process, improved the speed of following the tracks of.

Therefore, the device that the embodiment of the invention is followed the tracks of object video, not only can follow the tracks of object video accurately, and can follow the tracks of, improve the accuracy that object video is followed the tracks of object video with nonaffine variation with affine variation.

Description of drawings

The process flow diagram of Fig. 1 method that to be embodiments of the invention follow the tracks of object video;

The synoptic diagram of Fig. 2 method that to be embodiments of the invention follow the tracks of object video;

The synoptic diagram of the embodiment one of Fig. 3 method that to be embodiments of the invention follow the tracks of object video;

The displayed map as a result of Harr wavelet transformation in Fig. 4 method that to be embodiments of the invention follow the tracks of object video;

Under Fig. 5 method different resolution that to be embodiments of the invention follow the tracks of object video, the displayed map of small echo boundary descriptor;

Fig. 6 is the experimental result picture that utilizes the method that embodiments of the invention follow the tracks of object video;

Fig. 7 is the another experimental result picture that utilizes the method that embodiments of the invention follow the tracks of object video;

The schematic diagram of Fig. 8 device that to be embodiments of the invention follow the tracks of object video;

Fig. 9 is the schematic representation of apparatus that embodiments of the invention are followed the tracks of object video.

Embodiment

In order to be illustrated more clearly in the technical scheme of the embodiment of the invention, the accompanying drawing of required use is done an introduction simply in will describing embodiment below, apparently, accompanying drawing in describing below only is some embodiments of the present invention, for those of ordinary skills, under the prerequisite of not paying creative work, can also obtain other accompanying drawing according to these accompanying drawings.

In order can to follow the tracks of accurately the object video that affine variation and nonaffine change, the tracking that embodiments of the invention carry out object video is at first tried to achieve the contour feature value of object video described in the current frame image; Utilize the average drifting method to obtain the unique point of the object video in the current frame image then, try to achieve the matching characteristic point of described object video in the next frame image; Then, obtain the candidate profile of described object video in the zone occurring, try to achieve the contour feature value of described object video in the next frame image according to the matching characteristic point of described object video in the next frame image; At last,, mate, obtain the profile of described object video in the next frame image with the contour feature value of described object video in the next frame image with the contour feature value of object video described in the current frame image.

For the advantage of the technical scheme that makes the embodiment of the invention is clearer, embodiments of the invention are described in further detail below in conjunction with accompanying drawing.

As shown in Figure 1, embodiments of the invention method that object video is followed the tracks of may further comprise the steps:

S1, the unique point of getting the profile of object video described in the current frame image;

S2, in the next frame image, find the matching characteristic point with described Feature Points Matching;

S3, according to described matching characteristic point, in the next frame image, detect at least one candidate's profile of described object video;

The contour feature value of the object video in S4, the calculating current frame image;

The contour feature value of S5, the described candidate's profile of calculating;

S6, the contour feature value of described candidate's profile and the contour feature value of the object video in the current frame image are compared, if the two coupling, then described candidate's profile is the profile of described object video in the next frame image.

By the described method of the embodiment of the invention, at first determine the matching characteristic point of object video in the next frame image in the present frame, then according to the appearance zone of the described object video of described matching characteristic point prediction, and in the appearance zone of described prediction, detect candidate's profile of described object video.Then the contour feature value of the object video in forward and backward two two field pictures is mated, if the two coupling, then described candidate's profile is the profile of described object video in the next frame image.

When object video carries out the nonaffine variation, owing to can extract the contour feature value of described object video in the next frame image, and by mating with the contour feature value of described object video in current frame image, obtain with current frame image in the contour feature value contour feature value of mating the most, describe thereby can make accurately the profile of described object video in the next frame image.Utilize the described method of the embodiment of the invention, having avoided in the prior art can not be to having the defective that object video that nonaffine changes is accurately followed the tracks of.Therefore, the method that the embodiment of the invention is followed the tracks of object video, not only can follow the tracks of object video accurately, and can follow the tracks of, improve the accuracy that object video is followed the tracks of object video with nonaffine variation with affine variation.And because the algorithm of the embodiment of the invention is simple, thereby, compared to existing technologies, utilize the described method of the embodiment of the invention can improve the speed that object video is followed the tracks of.

As shown in Figure 2, above-mentioned steps S3 is described according to described matching characteristic point, and the step that detects at least one candidate's profile of described object video in the next frame image comprises:

S31:, predict the appearance zone of described object video profile in the next frame image according to described matching characteristic point;

S32: at least one the candidate's profile that in the appearance zone of described prediction, detects described object video.

Owing at first dope the appearance zone of object video in the next frame image, reduced the operand that object video is carried out outline, improved the efficient that object video is followed the tracks of.

In step S1, the unique point of object video profile described in the described current frame image can be, the central point of described object video in current frame image; Correspondingly, the matching characteristic point among the described step S2 is the coupling central point of described object video in the next frame image.

In addition, the profile of describing object video has a variety of modes, for example the length breadth ratio of invariant moments, excentricity, object video, form factor, small echo boundary descriptor etc.Wherein, the small echo boundary descriptor has advantages such as explicit physical meaning, retrieval performance are good, rotation, the neither change of convergent-divergent, can describe the contour feature value of object video accurately.Therefore, in an embodiment of the present invention, adopt the small echo boundary descriptor as the contour feature value of describing the object video profile.

Describe the specific implementation process of the method that the embodiment of the invention follows the tracks of object video in detail below in conjunction with Fig. 3.

T1: in current frame image, described object video is carried out profile detect, obtain the point of described object video.

The various calculating of carrying out in current frame image all are as the basis with the former frame of current frame image.And the various calculating of in the next frame image, being done, all based on present frame.Therefore, the various calculating principle in present frame, next frame image are identical, just the reference data difference.

The method that detects point is: in current frame image, check object indexing M _k ^jAll connection bitmap V in the scope of delineation _k, be 0 as long as in the upper and lower, left and right gray-scale value of any is arranged around certain point, just indicate that this point is a point.

T2: the contour vector that obtains described object video by described point.

Suppose that in current frame image object video has Np point, then its contour vector is defined as

After finding whole point, with described point ordering.The method that described point is sorted is: from checking object indexing M _k ^jThe coboundary in delineation zone begin horizon scan to the 1st point be the first point P ₀Then with first point P ₀Be the center, utilize the template of search pattern such as 3*3, the point that finds according to counterclockwise order is second point P ₁Again with the second point P ₁Being the center, utilizing search pattern, is the 3rd point P by the point that counterclockwise searches out ₂The rest may be inferred, and last point that finds is

So again with Be the center, utilize the search pattern of 3*3, first point that finds should be P ₀Above-mentioned method of searching point has been ignored the in-profile point of object video, and the contour vector of output only comprises the peripheral point of object video.

According to above-mentioned method with point Order sort, obtain described contour vector

Wherein, P _n=(P _Xn, P _Yn) (n=0 ... Np).

After obtaining above-mentioned contour vector, calculate the center-of-mass coordinate of described profile according to formula (1), (2)

${\mathrm{TX}}_k^j=\frac{1}{N_p}\underset{n=0}{\overset{N_p-1}{\mathrm{\Σ}}}{x}_n---\left(1\right)$

${\mathrm{TY}}_k^j=\frac{1}{N_p}\underset{n=0}{\overset{N_p-1}{\mathrm{\Σ}}}{y}_n---\left(2\right)$

Wherein, (x _n, y _n) be the coordinate of each point, (n=0,1......N _p-1).

T3:, calculate the normalization wheelspan vector of translation, rotation, the neither change of convergent-divergent according to described contour vector

Shown in computing formula following (3), (4), (5):

$r_n=\sqrt{{(x_n-{\mathrm{TX}}_k^j)}^2+{(y_n-{\mathrm{TY}}_k^j)}^2}---\left(3\right)$

r _max＝Max(r ₀，r ₁，...r _N-1) (4)

U _n＝r _n/r _max； (5)

Wherein, r _nBe the distance of each point to described barycenter, r _MaxBe the maximal value of each point in the described centroid distance, n=0,1......N _p-1.

T4: with the normalization wheelspan vector that obtains

Resequence, obtain directional wheel apart from vector

The method that described normalization wheelspan vector is resequenced is as follows:

From described normalization wheelspan vector

In, find out all maximal values and minimum value.Suppose to find J maximal value, K minimum value can be formed J*K " maximal value-minimum value to " so between these maximal values and the minimum value.In described J*K " maximal value-minimum value to ", find subscript between described maximal value, the minimum value maximum at interval a pair of " maximal value-minimum value to ".Because first in the described normalization wheelspan vector is adjacent with last on the profile of object video, so can make the interval of any two vectors remain on N _pIn/2.Therefore, if the interval d between certain two maximal value, minimum value greater than N _p/ 2, then make d=N _p/ 2.

If have only one " maximal value-minimum value to ", then with the most described directional wheel of minimum value in described " maximal value-minimum value to " apart from vector In first q ₀, and make maximal value at the preceding N of described directional wheel apart from vector _pIn/2, described normalization wheelspan vector is sorted, obtain directional wheel apart from vector according to the direction of " minimum value-maximal value "

If there are a plurality of " maximal value-minimum value to ", then determine which " maximal value-minimum value to " of selection, as the basis of the described directional wheel of calculating apart from vector by more described minimum value or peaked adjacency.For example, first to " maximal value-minimum value to " in, described peaked adjacency greater than second to peaked adjacency in " maximal value-minimum value to ", will be calculated the basis of described directional wheel apart from vector to " maximal value-minimum value to " with first so the most.If the maximal value of its adjacency equates that the profile that described object video then is described is symmetrical.

T5: with described directional wheel apart from vector

Carry out length normalization method, formation has the length normalization method directional wheel of regular length (for example length M=1024) apart from vector

Computing formula is as follows:

$a=\left[\frac{i}{M}{N}_p\right];---\left(6\right)$

b＝[a+1]； (7)

$c=\frac{i}{M}{N}_p-a;---\left(8\right)$

L _i＝(1-c)×q _a+c×q _b，(i＝0，1，......M-1)；(9)

Wherein, a, b are integer, and c is a floating number.

T6: by described length normalization method directional wheel apart from vector

Calculate the small echo boundary descriptor B of described object video in current frame image _k={ b ₀, b ₁... b _N-1.

To described length normalization method directional wheel apart from vector

Carry out Harr (Ha Er) wavelet transformation, obtain Harr wavelet transformation result

Concrete Harr wavelet transformation is achieved as follows:

Be provided with the one-dimension array L that a length is m, and m is 2 power power, then can realizes with following false code method the Harr wavelet transformation of this array:

By method recited above, can be with described length normalization method directional wheel apart from vector

Be transformed into

The equal in length of the two.The result schematic diagram of described Harr wavelet transformation as shown in Figure 4.

According to the difference of image resolution ratio N, described small echo boundary descriptor B _n={ b ₀ ⁿ, b ₁ ⁿ... b _N-1 ⁿCan obtain by formula (10):

B _n＝{b ₀ ⁿ，b ₁ ⁿ，...b _N-1 ⁿ}＝{w ₀ ⁿ，w ₁ ⁿ，...w _N-1 ⁿ} (10)

By formula (10) as can be seen, described small echo boundary descriptor B _n={ b ₀ ⁿ, b ₁ ⁿ... b _N-1 ⁿ, be the described Harr wavelet transformation result of intercepting

The top n coefficient and get.

Fig. 5 has shown that resolution is respectively 256,64 and at 16 o'clock, the result of the small echo boundary descriptor that calculates.In actual applications, in order to save object video profile operand relatively, resolution desirable 16.

The small echo boundary descriptor B of object video in calculating current frame image _n={ b ₀ ⁿ, b ₁ ⁿ... b _N-1 ⁿAfter, at first need to determine the appearance zone of described object video in the next frame image, and calculate the candidate profile of object video in described appearance zone, then at the small echo boundary descriptor B that calculates the candidate profile of described object video in the next frame image _N+1={ b ₀ ^N+1, b ₁ ^N+1... b _N-1 ^N+1.

Below, describe aforementioned calculation process in detail.

T7: obtain the appearance zone of described object video in the next frame image.

T71: the central point that obtains object video described in the current frame image.

T72: the central point of object video in current frame image according to obtaining, calculate the coupling central point of described object video in the next frame image.

In an embodiment of the present invention, adopt the average drifting method to calculate the coupling central point of described object video in the next frame image.So in calculating current frame image during the central point of object video, be that the former frame image with current frame image is that reference calculation obtains, it is identical with following described computation process to calculate principle.

Suppose

The normalization location of pixels of expression object video model, its center point coordinate is O; The color gray-scale value of described object video further is quantified as the m grade, and b (x) is the reflection of the pixel of position x to color index; The definition of probability that color u occurs is:

$\stackrel{\‾}{q_u}=\mathrm{\α}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}k\left({\left|\right|x_i^{*}\left|\right|}^2\right)\mathrm{\δ}[b\left(x_i^{*}\right)-u]---\left(11\right)$

Wherein: k (x) is a kernel function, and distance center point is less than the pixel weight of distant positions;

α is a constant, and its expression formula is:

$\mathrm{\α}=\frac{1}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}k\left({\left|\right|x_i^{*}\left|\right|}^2\right)}---\left(12\right)$

Then the object video model representation is:

$\stackrel{\‾}{q}={\left\{\stackrel{\‾}{q_u}\right\}}_{u=1,...,m},\underset{u=1}{\overset{m}{\mathrm{\Σ}}}\stackrel{\‾}{q_u}=1---\left(13\right)$

Suppose

Be the location of pixels of the object video of candidate in the present frame, its central point is C, and at the same kernel function k (x) of utilization in described central point scope h, then the probability that color u occurs in candidate's the object video can be expressed as:

$\stackrel{\‾}{p_u}\left(y\right)={\mathrm{\α}}_h\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}k\left({\left|\right|\frac{C-x_i}{h}\left|\right|}^2\right)\mathrm{\δ}[b\left(x_i\right)-u]---\left(14\right)$

Wherein: α _hBe constant, its expression formula is:

${\mathrm{\α}}_h=\frac{1}{\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}k\left({\left|\right|\frac{C-x_i}{h}\left|\right|}^2\right)}---\left(15\right)$

Then candidate's object video model representation is:

$\stackrel{\‾}{p}\left(y\right)={\left\{\stackrel{\‾}{p_u}\left(y\right)\right\}}_{u=1,...,m},\underset{u=1}{\overset{m}{\mathrm{\Σ}}}\stackrel{\‾}{p_u}=1---\left(16\right)$

By the object video model of above definition and candidate's object video model, can calculate between them apart from d (C):

$d\left(C\right)=\sqrt{1-\mathrm{\ρ}[\stackrel{\‾}{p}\left(C\right),\stackrel{\‾}{q}]}---\left(17\right)$

Wherein: $\stackrel{\‾}{\mathrm{\ρ}}\left(y\right)\≡\mathrm{\ρ}[\stackrel{\‾}{p}\left(C\right),\stackrel{\‾}{q}]=\underset{u=1}{\overset{m}{\mathrm{\Σ}}}\sqrt{\stackrel{\‾}{p_u}\left(C\right)\stackrel{\‾}{q_u}}.$

By top analysis as can be seen, the optimal candidate object video of the object video in the current image frame is the candidate object video nearest with described object video modal distance, just makes the candidate region minimum apart from d (C).Therefore obtain d (C) minimum value, just can determine the coupling central point of described object video in the next frame image.

The computing method of d (C) can be tried to achieve by iterative formula (18),

${\stackrel{\‾}{C}}_1=\frac{\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}{x}_i{w}_{i}g\left({\left|\right|\frac{{\stackrel{\‾}{C}}_0-x_i}{h}\left|\right|}^2\right)}{\underset{i=1}{\overset{n_h}{\mathrm{\Σ}}}{w}_{i}g\left({\left|\right|\frac{{\stackrel{\‾}{C}}_0-x_i}{h}\left|\right|}^2\right)}---\left(18\right)$

Wherein:

Be the current central point of object video,

Coupling central point for object video in the next frame image;

w _iExpression formula be:

$w_i=\underset{u=1}{\overset{m}{\mathrm{\Σ}}}\sqrt{\frac{\stackrel{\‾}{q_u}}{\stackrel{\‾}{p_u}\left(y_0\right)}}\mathrm{\δ}[b\left(x_i\right)-u]---\left(19\right)$

So, can be by in each two field picture, using this iterative formula (19), try to achieve and make d (C) obtain the candidate's object video and the central point thereof of minimum value, this candidate target is the optimal candidate object of object video, has so also just obtained the coupling central point of object video in the next frame image.Utilize the average drifting method, can improve the speed of determining matching characteristic point, improve the efficient of whole tracing process.Certainly, in the process of calculating the matching characteristic point of described object video in the next frame image, also can be without the average drifting method.Promptly after the definite unique point of object video in current frame image, determine that at first this unique point the zone may occur in the next frame image.And in the zone that may occur, the mating of individual element point, is until the unique point of being mated the most.

T73: after obtaining the coupling central point of described object video in the next frame image, utilizing linear method, is the center with described coupling central point, dopes the appearance zone of described object video in the next frame image.Described linear method can comprise translation, rotation etc.

For example, with

The object video profile scope that representative obtains from current frame image, and object video profile scope in the next frame image, the zone also promptly occurring can use

Expression.

The profile scope concrete predictor formula of described object video in the next frame image is as follows:

$\left\{\begin{array}{c}{\mathrm{Left}}_{k+1}={\mathrm{Left}}_k-w_k/2\\ {\mathrm{Top}}_{k+1}={\mathrm{Top}}_k-h_k/2\\ {\mathrm{Right}}_{k+1}={\mathrm{Right}}_k+w_k/2\\ {\mathrm{Bottom}}_{k+1}={\mathrm{bottom}}_k+h_k/2\end{array}\right.---\left(20\right)$

Wherein, (CX _k, CY _k) be the center point coordinate of object video in current frame image, (CX _K+1, CY _K+1) be the coupling center point coordinate of object video in the next frame image, can calculate acquisition according to above-mentioned average drifting method.

In formula (20),

$\left\{\begin{array}{c}{w}_k=({\mathrm{Right}}_k-{\mathrm{Left}}_k)\frac{{\mathrm{speed}}_k-\mathrm{speed}\_\min }{\mathrm{speed}\_\max -\mathrm{speed}\_\min }\\ h_k=({\mathrm{Bottom}}_k-{\mathrm{Top}}_k)\frac{{\mathrm{speed}}_k-\mathrm{speed}\_\min }{\mathrm{speed}\_\max -\mathrm{speed}\_\min }\end{array}---\left(21\right)\right.$

Here speed_min represents the minimum speed (generally getting 0) of object video motion, and speed_max represents the maximal rate (generally getting 1) of object video motion, speed _kRepresent the actual speed of object in the previous frame, its computing formula is as follows:

${\mathrm{speed}}_k=\frac{\sqrt{{({\mathrm{CX}}_k-{\mathrm{CX}}_{k-N})}^2+{({\mathrm{CY}}_k-{\mathrm{CY}}_{k-N})}^2}}{N({\mathrm{Right}}_{k-N}-{\mathrm{Left}}_{k-N})}---\left(22\right)$

N representative is used between two frames of computing velocity frame number (generally getting 10) at interval.

Certainly it should be noted that the method that obtains the coupling central point of described object video in the next frame image, the average drifting method that is not limited in the present embodiment to be mentioned.Any method that can obtain the central point of object video in image all can be applicable in the embodiments of the invention.

T8: after obtaining the coupling central point of described object video in the next frame image, in described appearance zone, in the next frame image, described object video is carried out profile to be detected, obtain candidate's profile of described object video, and calculate the small echo boundary descriptor of described object video in the next frame image

$B_{n+1}=\{{b_0}^{n+1},{b_1}^{n+1},...{b_{N-1}}^{n+1}\}.$

In this step, calculate the process of the small echo boundary descriptor of described object video in the next frame image, identical with the principle of step T1-T6 noted earlier, do not repeat them here.

T9: behind the small echo boundary descriptor of trying to achieve object video described in current image frame, next picture frame, described object video is carried out outline, obtain the profile of described object video in next picture frame, thereby finish tracking to object video.

The method of object video being carried out outline is as follows:

T91: with the small echo boundary descriptor B of candidate's profile in the described appearance zone _N+1={ b ₀ ^N+1, b ₁ ^N+1... b _N-1 ^N+1, with the small echo boundary descriptor B in the current frame image _n={ b ₀ ⁿ, b ₁ ⁿ... b _N-1 ⁿCompare, calculate the similarity between the two.Described similarity is calculated according to following formula:

$\mathrm{Similarity}\left(i\right)=1-\frac{1}{N}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}{(b_i^{n-1}-b_i^n)}^2---\left(23\right)$

If the similarity value of the medium and small wave contour descriptor of front and back two two field pictures surpasses the similarity threshold value, so at B _N+1={ b ₀ ^N+1, b ₁ ^N+1... b _N-1 ^N+1In, select the similarity value the highest as B _n={ b ₀ ⁿ, b ₁ ⁿ... b _N-1 ⁿIn corresponding tracking results.

Described similarity threshold value can freely define, and in the present embodiment, for the accuracy that guarantees object video is followed the tracks of, described similarity threshold value value is 80%.

If to B _n={ b ₀ ⁿ, b ₁ ⁿ... b _N-1 ⁿTracking results have a plurality ofly, will select the highest conduct of similarity to B so _n={ b ₀ ⁿ, b ₁ ⁿ... b _N-1 ⁿTracking results.

If to B _n={ b ₀ ⁿ, b ₁ ⁿ... b _N-1 ⁿFollow the tracks of failure, B then is described _n={ b ₀ ⁿ, b ₁ ⁿ... b _N-1 ⁿPairing object video is blocked or disappears; If B _N+1={ b ₀ ^N+1, b ₁ ^N+1... b _N-1 ^N+1Can not find source tracing object, b so _N+1={ b ₀ ^N+1, b ₁ ^N+1... b _N-1 ^N+1Pairing object video is for the object video of disappearance newly occurring or block in the next frame image.

At last with the pairing object video profile of optimum matching small echo boundary descriptor, mate with the profile of object video described in the current frame image, if its similarity surpasses threshold value, then illustrate the object video in the current frame image is followed the tracks of successfully, otherwise follow the tracks of failure.

By utilizing top method as can be seen, even deformation appears in tracked object video in motion process, the mode that the described method of the embodiment of the invention utilizes average drifting method and outline to combine still can be done accurately object video and follow the tracks of.

Utilize the described method that object video is followed the tracks of of the embodiment of the invention, the result that object video is followed the tracks of is shown in Fig. 6,7.From tracking results as can be seen, the tracking effect of the described method that object video is followed the tracks of of the embodiment of the invention is more satisfactory, can follow the tracks of the profile of object video more exactly, even during variations such as pedestrian's leg and arm bend, the true profile of the curve of candidate's profile and object video also is relatively to coincide, for example the c among Fig. 6) and d).

The tracing process of the described method that object video is followed the tracks of of the embodiment of the invention also is more stable, also can stably carry out profile and follows the tracks of even bigger variation takes place the movement velocity of target object.As shown in Figure 7, the automobile in the automobile video sails the parking lot from fast to slow into, and this algorithm has all been realized its stable tracking.

In addition, compare with the Condensation track algorithm, the described object video tracking of the embodiment of the invention, calculated amount is little, and tracking velocity improves a lot.Listed the tracking velocity value when utilizing two kinds of track algorithms to follow the tracks of respectively in the table 1.

Table 1

By top experiment as can be seen, the described method that object video is followed the tracks of of the embodiment of the invention, not only can follow the tracks of accurately the object video of affine variation or nonaffine variation, and because the algorithm of the embodiment of the invention is simple, thereby compare compared to prior art, the described method of the embodiment of the invention has improved the tracking velocity to object video.

Corresponding with the method that the embodiment of the invention is followed the tracks of object video, embodiments of the invention also provide a kind of device that object video is followed the tracks of.

As shown in Figure 8, the described device that object video is followed the tracks of of the embodiment of the invention comprises:

First positioning unit 801 is used to obtain the unique point of the profile of object video described in the current frame image;

First positioning unit 802 is used for finding matching characteristic point with described Feature Points Matching at the next frame image;

Profile detecting unit 803 is used for detecting at least one candidate's profile of described object video according to described matching characteristic point in the next frame image;

First computing unit 804 is used to calculate the contour feature value of described candidate's profile;

Second computing unit 805 is used for calculating the contour feature value of the object video of current frame image;

Outline unit 806 is used for the contour feature value of the object video of the contour feature value of described candidate's profile and current frame image is compared, if the two coupling, then described candidate's profile is the profile of described object video in the next frame image.

By the described device of the embodiment of the invention, at first determine the candidate profile of described object video in the next frame image by profile detecting unit 803, first computing unit 804, second computing unit 805 calculate the contour feature value of object video in forward and backward two two field pictures respectively, by outline unit 806 two contour feature values are mated, when object video carries out the nonaffine variation, also can obtain accurate description by the contour feature value of object video in two two field pictures before and after the coupling to described object video profile.When object video carries out the nonaffine variation, owing to can extract the contour feature value of described object video in the next frame image, and by mating with the contour feature value of described object video in current frame image, obtain with current frame image in the contour feature value contour feature value of mating the most, describe thereby can make accurately the profile of described object video in the next frame image.Utilize the described device of the embodiment of the invention, having avoided in the prior art can not be to having the defective that object video that nonaffine changes is accurately followed the tracks of.

Therefore, the device that the embodiment of the invention is followed the tracks of object video, not only can follow the tracks of object video accurately, and can follow the tracks of, improve the accuracy that object video is followed the tracks of object video with nonaffine variation with affine variation.

Equally, the profile of description object video has a variety of, for example the length breadth ratio of invariant moments, excentricity, object video, form factor, small echo boundary descriptor etc.Wherein, the small echo boundary descriptor has advantages such as explicit physical meaning, retrieval performance are good, rotation, the neither change of convergent-divergent, can describe the contour feature value of object video accurately.Therefore, in the device that object video is followed the tracks of, adopt the small echo boundary descriptor in an embodiment of the present invention as the contour feature value of describing the object video profile.

As shown in Figure 9, described profile detecting unit 803 comprises:

Regional prediction module 8031 is used for according to the appearance zone of the described object video profile of described matching characteristic point prediction at the next frame image;

Profile is chosen module 8032, is used for detecting in described appearance zone at least one candidate's profile of described object video.

At first dope the prediction of object video in the next frame image and the zone occurs, can carry out profile to object video targetedly and choose, reduced the calculated amount of object video being carried out outline, improved speed and efficient that object video is followed the tracks of.

Described first computing unit 804 comprises:

First profile detection module 8041 is used at current frame image, described object video is carried out profile detect, and obtains the point of described object video;

The first normalization wheelspan vector computing module 8042 is used for being obtained by described point the normalization wheelspan vector computing module of described object video;

The first directed profile vector meter is calculated module 8043, is used for being calculated by described normalization wheelspan vector computing module the directed contour vector of described object video;

The directed profile vector meter of first length normalization method is calculated module 8044, is used for described directed contour vector is carried out length normalization method, obtains the directed contour vector of length normalization method;

The first profile characteristic value calculating module 8045 is used for being obtained by the directed contour vector of described length normalization method the small echo boundary descriptor of described object video.

Described second computing unit 805 comprises:

Regional prediction module 8051 is used for obtaining the appearance zone of described object video at the next frame image;

Second profile detection module 8052 is used in described appearance zone, described object video is carried out profile detect, and obtains the point of described object video;

The second normalization wheelspan vector computing module 8053 is used for being obtained by described point the normalization wheelspan vector computing module of described object video;

The second directed profile vector meter is calculated module 8054, is used for being calculated by described normalization wheelspan vector computing module the directed contour vector of described object video;

The directed profile vector meter of second length normalization method is calculated module 8055, is used for described directed contour vector is carried out length normalization method, obtains the directed contour vector of length normalization method;

The second contour feature value computing module 8056 is used for being obtained by the directed contour vector of described length normalization method the small echo boundary descriptor of described object video.

Described first computing unit 804, second computing unit 805 are respectively formed module used algorithm in computation process, and identical with in the method embodiment that object video is followed the tracks of do not repeat them here.

In sum, utilize the method and the device that object video is followed the tracks of of the embodiment of the invention, not only can improve the accuracy that object video is followed the tracks of, and, improve the speed that object video is followed the tracks of because the algorithm of the embodiment of the invention is simple.

Certainly; embodiments of the invention also can have a variety of; under the situation that does not deviate from embodiments of the invention spirit and essence thereof; those of ordinary skills belong to the scope of protection of the invention not making under the creative work prerequisite by revising, be equal to, substituting the every other embodiment that is obtained.

Claims (21)

1. the method that object video is followed the tracks of is characterized in that, described method comprises the steps:

Get the unique point of the profile of object video described in the current frame image;

In the next frame image, find matching characteristic point with described Feature Points Matching;

According to described matching characteristic point, in the next frame image, detect at least one candidate's profile of described object video, comprise: with described matching characteristic point is the center, obtain the appearance zone of described object video profile in the next frame image by linear transformation, at described at least one candidate's profile that described object video occurs detecting in the zone;

Calculate the contour feature value of the object video in the current frame image;

Calculate the contour feature value of described candidate's profile;

The contour feature value of described candidate's profile and the contour feature value of the object video in the current frame image are compared, if the two coupling, then described candidate's profile is the profile of described object video in the next frame image.

2. the method that object video is followed the tracks of according to claim 1 is characterized in that, described contour feature value is the small echo boundary descriptor, or the invariant moments of profile, or excentricity, or form factor.

3. the method that object video is followed the tracks of according to claim 1 is characterized in that, describedly finds the process with the matching characteristic point of described Feature Points Matching to be specially in the next frame image:

Utilize the average drifting method, in the next frame image, find matching characteristic point with described Feature Points Matching.

4. the method that object video is followed the tracks of according to claim 1 is characterized in that, is specially in the described process that at least one candidate's profile of the described object video of detection in the zone occurs:

In described appearance zone, described object video is carried out profile detect, obtain the point of described object video;

Described point is sorted, obtain the contour vector of described object video

Wherein, described P ₀Be first point,

Be N _pIndividual point, N _pFor greater than 0 integer.

5. the method that object video is followed the tracks of according to claim 4 is characterized in that, described point is sorted, and obtains the contour vector of described object video

Process be specially:

The coboundary of the scope of being drawn a circle to approve from the object video index, first point that searches with horizontal direction is the first point P ₀

With the first point P ₀Be the center, utilize search pattern, in the scope that described search pattern is determined, search for, obtain the second point P according to counterclockwise direction ₁

According to obtain the identical step of second point by first point, up to finding N _pIndividual point

According to (the first point P ₀..., N _pIndividual point

) rank order, obtain the contour vector of described object video $P_{k+1}^i=(P_0,...,P_{N_p-1}).$

6. the method that object video is followed the tracks of according to claim 5 is characterized in that, the process of the contour feature value of described calculated candidate profile is specially:

By described contour vector

Obtain the normalization wheelspan vector of described object video Wherein, U ₀The distance of the barycenter of the profile of the object video that is first point in the current frame image and each point peaked merchant in the distance of described barycenter, U ₁The distance of the barycenter of the profile of the object video that is second point in the current frame image and each point peaked merchant in the distance of described barycenter,

Be N _pThe distance of the barycenter of the profile of the object video of individual point in the current frame image and each point peaked merchant in the distance of described barycenter;

By described normalization wheelspan vector

The directional wheel that calculates described object video is apart from vector

Wherein,

It is right to represent

Result after resequencing;

With described directional wheel apart from vector

Carry out length normalization method, the length normalization method directional wheel that obtains described object video is apart from vector

Wherein, L ₀... L _M-1It is right to represent

Carry out the result of length normalization method;

By described length normalization method directional wheel apart from vector

Obtain the small echo boundary descriptor B of described object video _K+1={ b ₀, b ₁... b _N-1; b ₀... b _N-1Expression L ₀, L ₁... L _M-1The small echo boundary descriptor, N represents the wavelet transformation result's that intercepts coefficient length;

Wherein, N _pFor forming described contour vector

The point number, M is that the length normalization method directional wheel is apart from vector

Length factor.

7. the method that object video is followed the tracks of according to claim 6 is characterized in that, by described contour vector

Obtain the normalization wheelspan vector of described object video

Process be specially:

Calculate the center-of-mass coordinate of described profile by described contour vector The computing formula of described center-of-mass coordinate is:

${\mathrm{TX}}_{k+1}^i=\frac{1}{N_p}\underset{n=0}{\overset{N_p-1}{\mathrm{\Σ}}}{x}_n,$ ${\mathrm{TY}}_{k+1}^j=\frac{1}{N_p}\underset{n=0}{\overset{N_p-1}{\mathrm{\Σ}}}{y}_n,$

Wherein, (x _n, y _n) be the coordinate of each point, (n=0,1......N _p-1);

Calculate normalization wheelspan vector $U_{k+1}^i=(U_0,U_1,......U_{N_{p}1}):$

$r_n=\sqrt{{(x_n-{\mathrm{TX}}_{k+1}^j)}^2+{(y_n-{\mathrm{TY}}_{k+1}^j)}^2}$

$r_{\max }=\mathrm{Max}(r_0,r_1,...r_{N_p-1})$

U _n＝r _n/r _max(n＝0，......N _p-1)；

Wherein, r _nBe the distance of each point to described barycenter, r _MaxBe the maximal value of each point in the described centroid distance.

8. the method that object video is followed the tracks of according to claim 6 is characterized in that, by described normalization wheelspan vector The directional wheel that calculates described object video is apart from vector $Q_{k+1}^i=(q_0,q_1\·\·\·q_{N_p-1})$ Process be specially:

At described normalization wheelspan vector

In, find maximal value and minimum value, form " maximal value-minimum value to ";

From described " maximal value-minimum value to ", find out maximum at interval a pair of of described maximal value and minimum value subscript;

According to the order of " minimum value-maximal value ", " maximal value-minimum value to " ordering that described subscript is maximum at interval obtains described directional wheel apart from vector

The order of wherein said " minimum value-maximal value " is meant that with described minimum value be initial, guarantees that maximal value is at preceding N _pOrdering in/2.

9. the method that object video is followed the tracks of according to claim 6 is characterized in that, with described directional wheel apart from vector

Carry out length normalization method, the length normalization method directional wheel that obtains described object video is apart from vector

Process be specially:

$a=\left[\frac{i}{M}{N}_p\right];$

b＝[a+1]；

$c=\frac{i}{M}{N}_p-a;$

L _i＝(1-c)×q _a+c×q _b，(i＝0，1，......M-1)；

Wherein, a, b, c is constant.

10. the method that object video is followed the tracks of according to claim 6 is characterized in that, by described length normalization method directional wheel apart from vector

Obtain the small echo boundary descriptor B of described object video _K+1={ b ₀, b ₁... b _N-1Process be specially:

To described length normalization method directional wheel apart from vector

Carry out wavelet transformation, obtain the wavelet transformation result

w ₀, w ₁... w _M-1Be expressed as L ₀, L ₁... L _M-1Carry out the result behind the wavelet transformation;

According to the difference of image resolution ratio, intercept described transformation results

Coefficient, obtain B _K+1={ b ₀, b ₁... b _N-1}=(w ₀, w ₁... w _N-1), the coefficient number that is intercepted is identical with the numerical value of described resolution.

11. the method that object video is followed the tracks of according to claim 2 is characterized in that, when described contour feature value was the small echo boundary descriptor, the process of calculating the contour feature value of the object video in the current frame image was specially:

In current frame image, described object video is carried out profile detect, obtain the contour vector of described object video $P_k^i=(P_0,...,P_{N_p-1}),$ Be each point coordinate;

By described contour vector Obtain the normalization wheelspan vector of described object video

Wherein, U ₀The distance of the barycenter of the profile of the object video that is first point in the current frame image and each point peaked merchant in the distance of described barycenter,

Be N _pThe distance of the barycenter of the profile of the object video of Jie's point in the current frame image and each point peaked merchant in the distance of described barycenter;

By described normalization wheelspan vector

The directional wheel that calculates described object video is apart from vector

Wherein, It is right to represent

Result after resequencing;

With described directional wheel apart from vector

Carry out length normalization method, the length normalization method directional wheel that obtains described object video is apart from vector

L ₀... L _M-1It is right to represent

Carry out the result of length normalization method;

By described length normalization method directional wheel apart from vector Obtain the small echo boundary descriptor B of described object video _k={ b ₀, b ₁... b _N-1; b ₀... b _N-1Expression L ₀, L ₁... L _M-1The small echo boundary descriptor;

Wherein, N _pFor forming described contour vector

The point number, M is that the length normalization method directional wheel is apart from vector

Length factor; N represents the wavelet transformation result's that intercepts coefficient length.

12. the method that object video is followed the tracks of according to claim 11 is characterized in that, the described contour vector that is obtained described object video by described point

Process be specially:

The coboundary of the scope of being drawn a circle to approve from the object video index, first point that searches with horizontal direction is the first point P ₀

With the first point P ₀Be the center, utilize search pattern, in the scope that described search pattern is determined, search for, obtain the second point P according to counterclockwise direction ₁

According to obtain the identical step of second point by first point, up to finding N _pIndividual point

According to (the first point P ₀..., N _pIndividual point ) rank order, obtain the contour vector of described object video $P_k^i=(P_0,...,P_{N_p-1}).$

13. the method that object video is followed the tracks of according to claim 11 is characterized in that, by described contour vector

Calculate the normalization wheelspan vector of described object video

Process be specially:

By described contour vector

Calculate the center-of-mass coordinate of described object video profile Described center-of-mass coordinate

Computing formula be:

${\mathrm{TX}}_{k+1}^i=\frac{1}{N_p}\underset{n=0}{\overset{N_p-1}{\mathrm{\Σ}}}{x}_n,$ ${\mathrm{TY}}_{k+1}^j=\frac{1}{N_p}\underset{n=0}{\overset{N_p-1}{\mathrm{\Σ}}}{y}_n,$

Wherein, (x _n, y _n) be the coordinate of each point, n=0,1......N _p-1;

According to described center-of-mass coordinate

Calculate normalization wheelspan vector

$r_n=\sqrt{{(x_n-{\mathrm{TX}}_k^j)}^2+{(y_n-{\mathrm{TY}}_k^j)}^2}$

r _max＝Max(r ₀，r ₁，...r _N-1)

U _n＝r _n/r _max；

Wherein, r _nBe the distance of each point to described barycenter, r _MaxBe the maximal value of each point in the described centroid distance, n=0,1......N _p-1.

14. the method that object video is followed the tracks of according to claim 11 is characterized in that, by described normalization wheelspan vector

The directional wheel that calculates described object video is apart from vector $Q_k^i=(q_0,q_1\·\·\·\·\·\·q_{N_p-1})$ Process be specially:

At described normalization wheelspan vector

In, find maximal value and minimum value, form " maximal value-minimum value to ";

From described " maximal value-minimum value to ", find out maximum at interval a pair of of described maximal value and minimum value subscript;

According to the order of " minimum value-maximal value ", " maximal value-minimum value to " ordering that described subscript is maximum at interval obtains described directional wheel apart from vector $Q_k^i=(q_0,q_1\·\·\·\·\·\·q_{N_p-1});$

The order of wherein said " minimum value-maximal value " is meant that with described minimum value be initial, guarantees that maximal value is at preceding N _pOrdering in/2.

15. the method that object video is followed the tracks of according to claim 11 is characterized in that, to described directional wheel apart from vector Carry out length normalization method, obtain the length normalization method directional wheel apart from vector $L_k^i=(L_0,L_1\·\·\·\·\·\·L_{M-1})$ Process be specially:

$a=\left[\frac{i}{M}{N}_p\right];$

b＝[a+1]；

$c=\frac{i}{M}{N}_p-a;$

L _i＝(1-c)×q _a+c×q _b，(i＝0，1，......M-1)；

Wherein, a, b, c are constant.

16. the method that object video is followed the tracks of according to claim 11 is characterized in that, by described length normalization method directional wheel apart from vector

Obtain the small echo boundary descriptor B of described object video _k={ b ₀, b ₁... b _N-1Process be specially:

To described length normalization method directional wheel apart from vector

Carry out wavelet transformation, obtain transformation results

w ₀, w ₁... w _M-1Be expressed as L ₀, L ₁... L _M-1Carry out the result behind the wavelet transformation;

According to the difference of image resolution ratio, intercept described transformation results

Coefficient, obtain B _k={ b ₀, b ₁... b _N-1}=(w ₀, w ₁... w _N-1), the coefficient number that is intercepted is identical with the numerical value of described resolution.

17. the method that object video is followed the tracks of according to claim 2, it is characterized in that, when described contour feature value was the small echo boundary descriptor, the process that the contour feature value of the contour feature value of described candidate's profile and the object video in the current frame image is compared was specially:

With the small echo boundary descriptor of each candidate's profile of object video in the next frame image, compare with the similarity of the small echo boundary descriptor of object video in the current frame image;

If the similarity value of the small echo boundary descriptor of front and back two frames surpasses the similarity threshold value, then with the candidate target profile in the next frame image, as the profile of the object video in the current frame image that traces into.

18. the method that object video according to claim 17 is followed the tracks of is characterized in that the computing method of described similarity value are:

$\mathrm{Similarity}\left(i\right)=1-\frac{1}{N}\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}{(b_i^{k-1}-b_i^k)}^2,$

Wherein,

The small echo boundary descriptor of object video described in the expression current frame image,

The small echo boundary descriptor of object video described in the expression next frame image, N are represented the wavelet transformation result's that intercepts coefficient length.

19. the device that object video is followed the tracks of is characterized in that, described device comprises:

First positioning unit is used to obtain the unique point of the profile of object video described in the current frame image;

Second positioning unit is used for finding matching characteristic point with described Feature Points Matching at the next frame image;

The profile detecting unit, be used for according to described matching characteristic point, in the next frame image, detect at least one candidate's profile of described object video, comprise: the regional prediction module, being used for described matching characteristic point is the center, obtains the appearance zone of described object video profile in the next frame image by linear transformation; Profile is chosen module, is used for detecting in described appearance zone at least one candidate's profile of described object video;

First computing unit is used for calculating the contour feature value of the object video of current frame image;

Second computing unit is used to calculate the contour feature value of described candidate's profile;

The outline unit is used for the contour feature value of the object video of the contour feature value of described candidate's profile and current frame image is compared, if the two coupling, then described candidate's profile is the profile of described object video in the next frame image.

20. the device that object video is followed the tracks of according to claim 19 is characterized in that, described first computing unit comprises:

First profile detection module is used at current frame image, described object video is carried out profile detect, and obtains the point of described object video;

The first normalization wheelspan vector computing module is used for being obtained by described point the normalization wheelspan vector of described object video;

The first directed profile vector meter is calculated module, is used for being obtained by the normalization wheelspan vector that described normalization wheelspan vector computing module calculates the directed contour vector of described object video;

The directed profile vector meter of first length normalization method is calculated module, is used for described directed contour vector is carried out length normalization method, obtains the directed contour vector of length normalization method;

The first profile characteristic value calculating module is used for being obtained by the directed contour vector of described length normalization method the small echo boundary descriptor of described object video.

21. the device that object video is followed the tracks of according to claim 19 is characterized in that, described second computing unit comprises:

The second normalization wheelspan vector computing module is used for being obtained by point the normalization wheelspan vector of described object video;

The second directed profile vector meter is calculated module, is used for being obtained by the normalization wheelspan vector that described normalization wheelspan vector computing module calculates the directed contour vector of described object video;

The directed profile vector meter of second length normalization method is calculated module, is used for described directed contour vector is carried out length normalization method, obtains the directed contour vector of length normalization method;

The second contour feature value computing module is used for being obtained by the directed contour vector of described length normalization method the small echo boundary descriptor of described object video.

Images