CN105096345A - Target tracking method based on dynamic measurement matrix and target tracking system based on dynamic measurement matrix - Google Patents

Abstract

The invention discloses a target tracking method based on a dynamic measurement matrix and a target tracking system based on the dynamic measurement matrix. The method comprises the following steps of: 1, compressing high-dimension features of samples into low-dimension features, and initializing the dynamic measurement matrix; 2, collecting a plurality of positive sample sets and negative sample sets around a target position to perform classifier updating learning; 3, determining the position of a current frame target; and 4, updating the dynamic measurement matrix, and returning to the second step until the tracking is completed. The dynamic measurement matrix is used for extracting compression features of the target in the tracking process, i.e., in the tracking process, the measurement matrix is updated by utilizing the features of the tracked target and the features of a Naive Bayes classifier; the fixed form of the measurement matrix is improved; and the adaptability of the tracking method is high. Experimental results show that the tracking method has good robustness.

Description

A kind of method for tracking target based on kinetic measurement matrix and system

Technical field

The present invention relates to computer vision field, particularly relate to a kind of method for tracking target based on kinetic measurement matrix and system.

Background technology

Target following based on video or image sequence is one of hot issue of computer vision, there is contact closely with target detection, also have benefited from the lifting of efficiency and accuracy on test problems based on the tracking (tracking-by-detection) detected.Target Tracking Problem needs the difficulty of reply a lot, and such as illumination variation, target shape or attitudes vibration and complex scene etc. all have a great impact the effect of following the tracks of.

Tracking problem roughly can be divided into two classes.One class is the tracking based on generation model, the display model of this kind of method learning objective, when target location searched by rope, and searching and the region that modal distance is nearest or reconstructed error is minimum; Another kind of is tracking based on discrimination model, and these class methods, using the classification problem of tracking problem as two classes, using target and background as positive and negative two classes, learn the classification boundaries in two class sample characteristics spaces, take full advantage of contextual information.

What MIT tracking target adopted is in subrange, detect order calibration method, is a kind of typically based on the tracking detected.MIT algorithm uses onlineboosting method to select Haar-like to select feature interpretation target and background, trains a Bayes classifier to realize following the tracks of.But the fixed relationship of Haar-like feature on space structure also limit the descriptive power to target, although there is onlineboosting method to go to select feature, training process is loaded down with trivial details.CT track algorithm improves space constraint and the training process of Haar-like feature, namely adopts broad sense Haar-like feature interpretation target and simplifies training process greatly.But calculation matrix just no longer changes after initialization in CT algorithm, result in the solidification of feature mode, cause to follow the tracks of in some scenarios and drift about.Application number is CN201410660331.8, name is called that the patent of invention of " a kind of video target tracking method based on compressed sensing " just have employed CT algorithm, but this application whole tracing process all the time according to this matrix come realization character compression, that is in whole tracing process, keep sparse matrix constant, namely result in the solidification of feature mode, cause to follow the tracks of in some scenarios and drift about.

Summary of the invention

The object of the invention is to overcome the deficiencies in the prior art, a kind of method for tracking target based on kinetic measurement matrix and system are provided, in tracing process, improve the fixed form of calculation matrix, namely upgrade calculation matrix according to the compressive features of the target traced into and the characteristic of Naive Bayes Classifier, thus the method for tracking target that a kind of adaptivity is stronger is provided.

The object of the invention is to be achieved through the following technical solutions: a kind of method for tracking target based on kinetic measurement matrix, it comprises the following steps:

S1: by the low dimensional feature of high dimensional feature boil down to of sample, and initialization kinetic measurement matrix;

S2: gather multiple positive sample set and negative sample set around target location, carry out sorter renewal learning;

S3: determine the position in present frame target;

S4: upgrade kinetic measurement matrix, and return step S2, until followed the tracks of.

Step S1 comprises following sub-step:

S11: by sample high dimensional feature the low dimensional feature of boil down to that is:

v＝R(t)x；

In formula, wherein m=(wh) ², n < < m, it is kinetic measurement matrix;

S12: initialization kinetic measurement matrix R (t), namely generate R (0), the element in formula is:

In formula, s=m/4, the sparse degree of representing matrix; I, j are respectively the ranks mark of matrix R (0), r _i,j(0) element of the capable j row of i is represented.

Step S2 comprises following sub-step:

S21: gather multiple positive sample respectively and multiple negative sample forms positive sample set Z around target ^αwith negative sample set Z ^β, wherein, α and β is the radius choosing positive and negative samples, 0 < α < β; And obtain positive sample set Z ^αwith negative sample set Z ^βcompressive features collection V ^αand V ^β;

S22: utilize Naive Bayes Classifier H (v), sets up probability model respectively to each dimension of compressive features v:

$H\left(v\right)=\log \left(\frac{{\&Pi;}_{i=1}^{n}p\left(v_i|y=1\right)p\left(y=1\right)}{{\&Pi;}_{i=1}^{n}p\left(v_i|y=0\right)p\left(y=0\right)}\right)=\underset{i=1}{\overset{n}{\&Sigma;}}\log \left(\frac{p\left(v_i|y=1\right)}{p\left(v_i|y=0\right)}\right);$

In formula, compressive features v ∈ V ^α∪ V ^β, each element of compressive features v is independent distribution, and positive sample set Z ^αwith negative sample set Z ^βobedience is uniformly distributed i.e. p (y=1)=p (y=0); Y ∈ 0,1} representative sample label, y=0 represents negative sample, and y=1 represents positive sample;

S23: respectively Gauss model is set up to each dimension of compressive features v:

$p\left(v_i\right|y=0)~N({\mathrm{\&mu;}}_i^0,{\mathrm{\&sigma;}}_i^0);$

$p\left(v_i\right|y=1)~N({\mathrm{\&mu;}}_i^1,{\mathrm{\&sigma;}}_i^1);$

In formula, average and the variance of positive sample pattern i-th dimension, average and the variance of negative sample model i-th dimension; Target and sample pattern are by parameter represent;

S24: sorter learning process: calculate characteristic statistics value, iteration more the first month of the lunar year sample pattern:

$\begin{array}{c}{\mathrm{\&mu;}}_i^1\&LeftArrow;{\mathrm{\&lambda;\&mu;}}_i^1+\left(1-\mathrm{\&lambda;}\right){\mathrm{\&mu;}}_i^1;\\ {\mathrm{\&sigma;}}_i^1\&LeftArrow;\sqrt{\mathrm{\&lambda;}{\left({\mathrm{\&sigma;}}_i^1\right)}^2+(1-\mathrm{\&lambda;}){\left({\mathrm{\&sigma;}}^1\right)}^2+\mathrm{\&lambda;}(1-\mathrm{\&lambda;}){({\mathrm{\&mu;}}_i^1-{\mathrm{\&mu;}}^1)}^2};\end{array}$

In formula, λ is learning rate, with add up according to the positive sample pattern of step S23 the average and variance that obtain.

Step S3 comprises following sub-step:

S31: collecting sample composition candidate samples set Z around previous frame target ^γ, by Z ^γmiddle sample does feature extraction and compression, obtains compressive features v';

S32: compressive features v' is input to Naive Bayes Classifier H (v), obtains the compressive features v that score is the highest ^*:

$v^{*}=\arg \underset{v\&Element;V^{\mathrm{\&gamma;}}}{max}H\left(v\right);$

V ^*be categorized as the highest feature of positive sample degree of confidence, then this compressive features v ^*corresponding sample is target sample, then obtains the reposition of target.

Step S4 comprises following sub-step:

S41: use Naive Bayes Classifier H (v) to compressive features v ^*each dimensionality analysis:

$h_i\left(v^{*}\right)=\log \left(\frac{p\left({v^{*}}_i|y=1\right)}{p\left({v^{*}}_i|y=0\right)}\right);$

S42: definition error function:

error(v ^* _i)＝-h _i( _v ^*)；

And adopt idx to represent the row vector needing the correspondence upgraded in kinetic measurement matrix, i.e. the i-th dx dimensional vector:

$idx=\arg \underset{1\&le;i\&le;n}{max}error\left({v^{*}}_i\right);$

S43: adopt random mode to produce new row vector row vector R ^*element be:

In formula, r _{1, j} ^*row vector R ^*a jth element; S=m/4, the sparse degree of representing matrix;

S44: the new row vector R obtained with step S43 ^*replace the i-th dx dimensional vector in kinetic measurement matrix, formula is as follows:

$r_{i,j}\left(t\right)=\left\{\begin{array}{cc}{r}_{i,j}(t-1),& i\&NotEqual;idx\\ r_{1,j}^{*},& i=idx\end{array}\right.,$

In formula, t represents current frame number, namely at t frame;

S45: if target following completes, terminate, otherwise return step S2.

Based on a Target Tracking System for kinetic measurement matrix, it comprises kinetic measurement matrix initialisation module, training classifier module, search object module and kinetic measurement matrix update module; Described kinetic measurement matrix initialisation module is used for initialization kinetic measurement matrix; Described training classifier module is used for upgrading sorter according to the compressive features of positive negative sample; Described search object module is used for gathering candidate samples around the target of previous frame, and by the compressive features of candidate samples input sorter, sample corresponding to the compressive features that score is the highest is just regarded as the target sample of present frame; Described kinetic measurement matrix update module is utilized the characteristic of sorter and is upgraded kinetic measurement matrix by the target sample that search object module obtains.

Described sorter is Naive Bayes Classifier.

The invention has the beneficial effects as follows: the present invention proposes a kind of method for tracking target based on kinetic measurement matrix.The method adopts kinetic measurement matrix to extract the compressive features of target in tracing process, namely in tracing process, utilize the characteristic of clarification of objective and the Naive Bayes Classifier traced into upgrade calculation matrix, improve the fixed form of calculation matrix, the adaptability of this tracking is stronger.The results show, this tracking has good robustness.

Accompanying drawing explanation

Fig. 1 is the inventive method process flow diagram;

Fig. 2 is feature extraction of the present invention and compression process schematic diagram;

The renewal process schematic diagram of Fig. 3 position kinetic measurement matrix of the present invention.

Embodiment

Below in conjunction with accompanying drawing, technical scheme of the present invention is described in further detail: as shown in Figure 1, a kind of method for tracking target based on kinetic measurement matrix, it comprises the following steps:

S1: by the low dimensional feature of high dimensional feature boil down to of sample, and initialization kinetic measurement matrix;

S2: gather multiple positive sample set and negative sample set around target location, carry out sorter renewal learning;

S3: determine the position in present frame target;

S4: upgrade kinetic measurement matrix, and return step S2, until followed the tracks of.

Step S1 comprises following sub-step:

S11: by sample high dimensional feature the low dimensional feature of boil down to that is:

v＝R(t)x；

In formula, wherein m=(wh) ², n < < m, it is kinetic measurement matrix;

S12: initialization kinetic measurement matrix R (t), namely generate R (0), the element in formula is:

In formula, s=m/4, the sparse degree of representing matrix; I, j are respectively the ranks mark of matrix R (0), r _i,j(0) element of the capable j row of i is represented.

Step S2 comprises following sub-step:

S21: as shown in Figure 2, gathers multiple positive sample respectively according to the formula in step S12 around target and multiple negative sample forms positive sample set Z ^αwith negative sample set Z ^β, wherein, α and β is the radius choosing positive and negative samples, 0 < α < β; And obtain positive sample set Z ^αwith negative sample set Z ^βcompressive features collection V ^αand V ^β;

S22: utilize Naive Bayes Classifier H (v), sets up probability model respectively to each dimension of compressive features v:

$H\left(v\right)=\log \left(\frac{{\&Pi;}_{i=1}^{n}p\left(v_i|y=1\right)p\left(y=1\right)}{{\&Pi;}_{i=1}^{n}p\left(v_i|y=0\right)p\left(y=0\right)}\right)=\underset{i=1}{\overset{n}{\&Sigma;}}\log \left(\frac{p\left(v_i|y=1\right)}{p\left(v_i|y=0\right)}\right);$

In formula, compressive features v ∈ V ^α∪ V ^β, each element of compressive features v is independent distribution, and positive sample set Z ^αwith negative sample set Z ^βobedience is uniformly distributed i.e. p (y=1)=p (y=0); Y ∈ 0,1} representative sample label, y=0 represents negative sample, and y=1 represents positive sample;

S23: respectively Gauss model is set up to each dimension of compressive features v:

$p\left(v_i\right|y=0)~N({\mathrm{\&mu;}}_i^0,{\mathrm{\&sigma;}}_i^0);$

$p\left(v_i\right|y=1)~N({\mathrm{\&mu;}}_i^1,{\mathrm{\&sigma;}}_i^1);$

In formula, average and the variance of positive sample pattern i-th dimension, average and the variance of negative sample model i-th dimension; Target and sample pattern are by parameter represent;

S24: sorter learning process: calculate characteristic statistics value, iteration more the first month of the lunar year sample pattern:

$\begin{array}{c}{\mathrm{\&mu;}}_i^1\&LeftArrow;{\mathrm{\&lambda;\&mu;}}_i^1+\left(1-\mathrm{\&lambda;}\right){\mathrm{\&mu;}}_i^1;\\ {\mathrm{\&sigma;}}_i^1\&LeftArrow;\sqrt{\mathrm{\&lambda;}{\left({\mathrm{\&sigma;}}_i^1\right)}^2+(1-\mathrm{\&lambda;}){\left({\mathrm{\&sigma;}}^1\right)}^2+\mathrm{\&lambda;}(1-\mathrm{\&lambda;}){({\mathrm{\&mu;}}_i^1-{\mathrm{\&mu;}}^1)}^2};\end{array}$

In formula, λ is learning rate, with add up according to the positive sample pattern of step S23 the average and variance that obtain.

Step S3 comprises following sub-step:

S31: collecting sample composition candidate samples set Z around previous frame target ^γ, by Z ^γmiddle sample does feature extraction and compression, obtains compressive features v', as shown in the formula in step S12;

S32: compressive features v' is input to Naive Bayes Classifier H (v), obtains the compressive features v that score is the highest ^*:

$v^{*}=\arg \underset{v\&Element;V^{\mathrm{\&gamma;}}}{max}H\left(v\right);$

V ^*be categorized as the highest feature of positive sample degree of confidence, then this compressive features v ^*corresponding sample is target sample, then obtains the reposition of target.

Step S4 comprises following sub-step:

S41: use Naive Bayes Classifier H (v) to compressive features v ^*each dimensionality analysis:

$h_i\left(v^{*}\right)=\log \left(\frac{p\left({v^{*}}_i|y=1\right)}{p\left({v^{*}}_i|y=0\right)}\right);$

S42:v ^*in be classified as positive sample may a minimum dimension be that we are concerned about " bad feature ", it represents that this dimensional feature can not describe positive negative sample preferably, and aligning sample classification is that positive sample has the opposite effect; Still define error function:

error(v ^* _i)＝-h _i(v ^*)；

The dimension of the row vector that above-mentioned " bad feature " is corresponding in kinetic measurement matrix is represented, that is: with idx

$idx=\arg \underset{1\&le;i\&le;n}{max}error\left({v^{*}}_i\right);$

Like this, determine the row vector needing to change in kinetic measurement matrix, be the i-th dx dimensional vector;

S43: adopt random mode to produce new row vector row vector R ^*element be:

In formula, r _{1, j} ^*row vector R ^*a jth element; S=m/4, the sparse degree of representing matrix;

S44: the renewal process of kinetic measurement matrix as shown in Figure 3, the new row vector R obtained with step S43 ^*replace the i-th dx dimensional vector in kinetic measurement matrix, formula is as follows:

$r_{i,j}\left(t\right)=\left\{\begin{array}{cc}{r}_{i,j}(t-1),& i\&NotEqual;idx\\ r_{1,j}^{*},& i=idx\end{array}\right.,$

In formula, t represents current frame number, namely at t frame;

S45: if target following completes, terminate, otherwise return step S2.

Based on a Target Tracking System for kinetic measurement matrix, it comprises kinetic measurement matrix initialisation module, training classifier module, search object module and kinetic measurement matrix update module; Described kinetic measurement matrix initialisation module is used for initialization kinetic measurement matrix; Described training classifier module is used for upgrading sorter according to the compressive features of positive negative sample; Described search object module is used for gathering candidate samples around the target of previous frame, and by the compressive features of candidate samples input sorter, sample corresponding to the compressive features that score is the highest is just regarded as the target sample of present frame; Described kinetic measurement matrix update module is utilized the characteristic of sorter and is upgraded kinetic measurement matrix by the target sample that search object module obtains.

Described sorter is Naive Bayes Classifier.

Claims (7)

1. based on a method for tracking target for kinetic measurement matrix, it is characterized in that: it comprises the following steps:

S1: by the low dimensional feature of high dimensional feature boil down to of sample, and initialization kinetic measurement matrix;

S2: gather multiple positive sample set and negative sample set around target location, carry out sorter renewal learning;

S3: determine the position in present frame target;

S4: upgrade kinetic measurement matrix, and return step S2, until followed the tracks of.

2. a kind of method for tracking target based on kinetic measurement matrix according to claim 1, is characterized in that: step S1 comprises following sub-step:

S11: by sample high dimensional feature the low dimensional feature of boil down to that is:

v＝R(t)x；

In formula, wherein m=(wh) ², n < < m, it is kinetic measurement matrix;

S12: initialization kinetic measurement matrix R (t), namely generate R (0), the element in formula is:

In formula, s=m/4, the sparse degree of representing matrix; I, j are respectively the ranks mark of matrix R (0), r _i,j(0) element of the capable j row of i is represented.

3. a kind of method for tracking target based on kinetic measurement matrix according to claim 1, is characterized in that: step S2 comprises following sub-step:

S21: gather multiple positive sample respectively and multiple negative sample forms positive sample set Z around target ^αwith negative sample set Z ^β, wherein, α and β is the radius choosing positive and negative samples, 0 < α < β; And obtain positive sample set Z ^αwith negative sample set Z ^βcompressive features collection V ^αand V ^β;

S22: utilize Naive Bayes Classifier H (v), sets up probability model respectively to each dimension of compressive features v:

$H\left(v\right)=log\left(\frac{{\&Pi;}_{i=1}^{n}p\left(v_i\right|y=1)p(y=1)}{{\&Pi;}_{i=1}^{n}p\left(v_i\right|y=0)p(y=0)}\right)=\underset{i=1}{\overset{n}{\&Sigma;}}log\left(\frac{p\left(v_i\right|y=1)}{p\left(v_i\right|y=0)}\right);$

In formula, compressive features v ∈ V ^α∪ V ^β, each element of compressive features v is independent distribution, and positive sample set Z ^αwith negative sample set Z ^βobedience is uniformly distributed i.e. p (y=1)=p (y=0); Y ∈ 0,1} representative sample label, y=0 represents negative sample, and y=1 represents positive sample;

S23: respectively Gauss model is set up to each dimension of compressive features v:

$p\left(v_i\right|y=0)~N({\mathrm{\&mu;}}_i^0,{\mathrm{\&sigma;}}_i^0);$

$p\left(v_i\right|y=1)~N({\mathrm{\&mu;}}_i^1,{\mathrm{\&sigma;}}_i^1);$

In formula, average and the variance of positive sample pattern i-th dimension, average and the variance of negative sample model i-th dimension; Target and sample pattern are by parameter represent;

S24: sorter learning process: calculate characteristic statistics value, iteration more the first month of the lunar year sample pattern:

${\mathrm{\&mu;}}_i^1\&LeftArrow;{\mathrm{\&lambda;\&mu;}}_i^1+(1-\mathrm{\&lambda;}){\mathrm{\&mu;}}^1;$

${\mathrm{\&sigma;}}_i^1\&LeftArrow;\sqrt{\mathrm{\&lambda;}{\left({\mathrm{\&sigma;}}_i^1\right)}^2+(1-\mathrm{\&lambda;}){\left({\mathrm{\&sigma;}}^1\right)}^2+\mathrm{\&lambda;}(1-\mathrm{\&lambda;}){({\mathrm{\&mu;}}_1^i-{\mathrm{\&mu;}}^1)}^2};$

In formula, λ is learning rate, ${\mathrm{\&mu;}}^1=\frac{1}{n}{\mathrm{\&Sigma;}}_{k|y=1}{v}_i\left(k\right)$ With ${\mathrm{\&sigma;}}^1=\sqrt{\frac{1}{n}{\mathrm{\&Sigma;}}_{k|y=1}{\left(v_i\right(k)-{\mathrm{\&mu;}}^1)}^2}$ Add up according to the positive sample pattern of step S23 the average and variance that obtain.

4. a kind of method for tracking target based on kinetic measurement matrix according to claim 1, is characterized in that: step S3 comprises following sub-step:

S31: collecting sample composition candidate samples set Z around previous frame target ^γ, by Z ^γmiddle sample does feature extraction and compression, obtains compressive features v';

S32: compressive features v' is input to Naive Bayes Classifier H (v), obtains the compressive features v that score is the highest ^*:

$v^{*}=\arg \underset{v\&Element;V^{\mathrm{\&gamma;}}}{max}H\left(v\right);$

V ^*be categorized as the highest feature of positive sample degree of confidence, then this compressive features v ^*corresponding sample is target sample, then obtains the reposition of target.

5. a kind of method for tracking target based on kinetic measurement matrix according to claim 4, is characterized in that: step S4 comprises following sub-step:

S41: use Naive Bayes Classifier H (v) to compressive features v ^*each dimensionality analysis:

$h_i\left(v^{*}\right)=log\left(\frac{p\left({v^{*}}_i\right|y=1)}{p\left({v^{*}}_i\right|y=0)}\right);$

S42: definition error function:

error(v ^* _i)＝-h _i(v ^*)；

And adopt idx to represent the row vector needing the correspondence upgraded in kinetic measurement matrix, i.e. the i-th dx dimensional vector:

$idx=\arg \underset{1\&le;i\&le;n}{max}error\left({v^{*}}_i\right);$

S43: adopt random mode to produce new row vector row vector R ^*element be:

In formula, r _{1, j} ^*row vector R ^*a jth element; S=m/4, the sparse degree of representing matrix;

S44: the new row vector R obtained with step S43 ^*replace the i-th dx dimensional vector in kinetic measurement matrix, formula is as follows:

$r_{i,j}\left(t\right)=\left\{\begin{array}{cc}{r}_{i,j}(t-1),& i\&NotEqual;idx\\ r_{1,j}^{*},& i=idx\end{array}\right.,$

In formula, t represents current frame number, namely at t frame;

S45: if target following completes, terminate, otherwise return step S2.

6. based on a Target Tracking System for kinetic measurement matrix, it is characterized in that: it comprises kinetic measurement matrix initialisation module, training classifier module, search object module and kinetic measurement matrix update module; Described kinetic measurement matrix initialisation module is used for initialization kinetic measurement matrix; Described training classifier module is used for upgrading sorter according to the compressive features of positive negative sample; Described search object module is used for gathering candidate samples around the target of previous frame, and by the compressive features of candidate samples input sorter, sample corresponding to the compressive features that score is the highest is just regarded as the target sample of present frame; Described kinetic measurement matrix update module is utilized the characteristic of sorter and is upgraded kinetic measurement matrix by the target sample that search object module obtains.

7. a kind of Target Tracking System based on kinetic measurement matrix as claimed in claim 6, is characterized in that: described sorter is Naive Bayes Classifier.