CN103839273A - Real-time detection tracking frame and tracking method based on compressed sensing feature selection - Google Patents

Abstract

The invention provides a real-time detection tracking frame and tracking method based on compressed sensing feature selection. According to the tracking frame and tracking method, compressed features can be selected, and only the sample characteristic with a high distinction degree is used for classification. Real-time tracking can be achieved through the frame and method, the tracking failure phenomenon caused by selection of wrong features is avoided, the influence of bad features on the tracking result is effectively restrained, the tracking speed is obviously increased, and the tracking precision is obviously improved.

Description

Framework and tracking are followed the tracks of in real-time detection based on compressed sensing feature selecting

Technical field

The present invention relates to target tracking domain, relate in particular to a kind of real-time detection based on compressed sensing feature selecting and follow the tracks of framework and tracking.

Background technology

Known, target following is a key areas in computer vision technique, and in military affairs, medical treatment, has important application in monitoring and man-machine interaction.There are many methods recent years for solving the problem of target following, but due to the deformation of target, the variation of illumination, and the target reason such as be blocked, target following remains a difficult point.

The method for real time tracking of main flow all has adaptivity at present.In general tracking can be divided into two classes: generation method and method of discrimination.Generation method can learning objective characteristic model, then search for the region that target may place, the region that uses the model of having learnt to rebuild with least error is target position.In order to solve target deformation problems, WSL and IVT method are successively suggested.Recently, rarefaction representation method is with solving target by the problem of partial occlusion.But these generation models all do not utilize target background information around, these background informations can be in the time detecting target better by target and background separation out.

Discrimination model is regarded target following as a kind of by target and background separation test problems out.In discrimination model, use the good feature of separating capacity can effectively improve tracking accuracy.Use the boosting method of multiple Weak Classifier composition strong classifiers to be widely used at present.But many boosting methods have only been utilized the information of target itself, do not utilize the information of target background, so after target is not accurately detected, the precision of target following after will affecting, finally causes following the tracks of unsuccessfully.Recently, MIL method shows to utilize background information to use high dimensional feature Random Maps to become low dimensional feature to classify, and can effectively target and background area be separated.But because the feature of Random Maps is not likely suitable for classification, this produces adverse influence by causing using the bad feature of discrimination to carry out classification meeting to the effect of target following.

Summary of the invention

The present invention, in order to solve above-mentioned technical matters, has proposed a kind of real-time detection based on compressed sensing feature selecting and has followed the tracks of framework and tracking.

Technical scheme of the present invention is: framework and tracking are followed the tracks of in a kind of real-time detection based on compressed sensing feature selecting, comprise the steps:

Step 1, initiation parameter, determines the target of t frame

described target location is a rectangle frame, is the target that needs tracking in frame;

comprise four parameters: the row-coordinate of target in this frame

row coordinate

width width and height height.

Step 2, makes l (x)={ row (x), col (x) } represent the position of sample x, comprises row-coordinate and the row coordinate of center of a sample,

in radius s pixel, gather positive sample set around $X_t^s=\left\{x_t\right|\left|\right|l\left(x_t\right)-l\left(x_t^{*}\right)\left|\right|\≤s\};$ ?

radius r around, gathers negative sample collection between β pixel $X_t^{r,\mathrm{\β}}=\left\{x_t\right|r\≤\left|\right|l\left(x_t^{*}\right)-l\left(x_t\right)\left|\right|\≤\mathrm{\β}\};$ Wherein s, r, β is empirical parameter, unit is pixel.

Step 3, aligns negative sample and concentrates each sample

extract n class Lis Hartel and levy f (x _t)={ f ₁(x _t), f ₂(x _t) ..., f _n(x _t).

Step 4, is used the features training Weak Classifier of positive sample and negative sample.

Step 5, defining classification interval is assessed each Weak Classifier, and selects Weak Classifier to be used for classification.

Step 6, is made up of the strong classifier of t frame the Weak Classifier choosing

Step 7, at t+1 frame

collecting test sample in radius λ around

use H _t(x) test sample book is classified.

Step 8, classification results is the position at t+1 frame as target

Step 9, if t+1 frame is not last frame, makes t=t+1, returns to step 1.

And, s=4 in described step 2, the span that the span of r is 6～10, β is 15～30.

And, further comprising the steps of in described step 3:

3.1 are provided with template set

wherein element definition is:

Being w for width, is highly the sample of h, carries out convolution with template set T, altogether can generate m=(wh) ²individual feature;

3.2 use stochastic matrix R ^{n × m}the m of a sample feature is compressed, and n represents the vectorial dimension after compression.Element r in R _ijmeet the following conditions:

Use this matrix to compress the high dimensional feature of sample x.After compression, be characterized as f (x _t)={ f ₁(x _t), f ₂(x _t) ..., f _n(x _t)=RT (x _t), wherein T (x _t) represent that practical template set T is to sample x _tcarry out the vector of m feature of convolution extraction; Wherein, empirical parameter ζ represents degree of rarefication.

And, in described step 3.2, represent degree of rarefication empirical parameter ζ=4.

And described step 4 is further comprising the steps of:

The feature of the 4.1 positive samples of hypothesis and negative sample meets normal distribution, and positive sample obedience sample characteristics average is μ ₁, sample characteristics standard deviation is σ ₁normal distribution N (μ ₁, σ ₁), be designated as p (f (x) | y=1)～N (μ ₁, σ ₁), it is μ that negative sample is obeyed sample characteristics average ₀, sample characteristics standard deviation is σ ₀normal distribution N (μ ₀, σ ₀), be designated as p (f (x) | y=0)～N (μ ₀, σ ₀); Utilize positive negative sample to obtain parameter μ ₁, σ ₁, μ ₀and σ ₀;

The parameter update mode of positive sample is: ${\mathrm{\μ}}_t^1=\mathrm{\λ}{\mathrm{\μ}}_{t-1}^1+(1-\mathrm{\λ}){\widetilde{\mathrm{\μ}}}_t^1$

${\mathrm{\σ}}_t^1=\sqrt{\mathrm{\λ}{\left({\mathrm{\σ}}_{t-1}^1\right)}^2+(1-\mathrm{\λ}){\left({\widetilde{\mathrm{\σ}}}_t^1\right)}^2+\mathrm{\λ}(1-\mathrm{\λ}){({\mathrm{\μ}}_{t-1}^1-{\mathrm{\μ}}_t^1)}^2}$

Wherein, λ is learning rate,

represent the positive sample characteristics average of t moment of directly obtaining,

represent the study positive sample characteristics average in rear t moment, represent the study positive sample characteristics average in rear t-1 moment,

represent the standard deviation of the positive sample characteristics in rear t moment of study,

the standard deviation of the positive sample characteristics that expression Direct Sampling is obtained,

represent the standard deviation of the positive sample in t-1 moment;

The parameter update mode of negative sample is: ${\mathrm{\μ}}_t^0=\mathrm{\λ}{\mathrm{\μ}}_{t-1}^0+(1-\mathrm{\λ}){\widetilde{\mathrm{\μ}}}_t^0$

${\mathrm{\σ}}_t^0=\sqrt{\mathrm{\λ}{\left({\mathrm{\σ}}_{t-1}^0\right)}^2+(1-\mathrm{\λ}){\left({\widetilde{\mathrm{\σ}}}_t^0\right)}^2+\mathrm{\λ}(1-\mathrm{\λ}){({\mathrm{\μ}}_{t-1}^0-{\mathrm{\μ}}_t^0)}^2}$

Wherein, λ is learning rate,

represent the t moment negative sample characteristic mean of directly obtaining,

represent the study negative sample characteristic mean in rear t moment, represent the study negative sample characteristic mean in rear t-1 moment,

represent the standard deviation of the negative sample feature in rear t moment of study,

the standard deviation of the negative sample feature that expression Direct Sampling is obtained,

represent the standard deviation of the negative sample in t-1 moment;

4.2 Weak Classifier

h _k(x) represent sample x to use k tagsort.Wherein

f _k(x) k the feature of expression sample x.

And the learning rate λ in described step 4.1 and 4.2 gets 0.85.

And, in described step 5, comprise the following steps:

5.1 definition $m\arg \mathrm{in}\left(h_k\right)=\frac{1}{n_p}\underset{i=1}{\overset{n_p}{\mathrm{\Σ}}}{h}_k\left(x_i\right)-\frac{1}{n_n}\underset{j=1}{\overset{n_n}{\mathrm{\Σ}}}{h}_k\left(x_j\right),$ Wherein n _prepresent the number of positive sample, n _nrepresent the number of negative sample, $x_i\∈X_t^s,x_j\∈X_t^{r,\mathrm{\β}};$

5.2 select front K maximum feature of margin value

composition strong classifier, wherein i _krepresent to select the feature for classifying;

The strong classifier of 5.3 t frames is defined as

And, in described step 7, comprise the following steps:

7.1 at t+1 frame

collecting test sample in radius λ pixel around

7.2 obtain the classification results of each sample

7.3 targets are

the position of target in t+1 frame is

As preferably, in described step 2, the quantity of the negative sample of getting is 1.5 times of positive sample collection quantity.

The invention has the beneficial effects as follows: the present invention (CFS) contrasts with the result of current other popular trackings, is better than other the whole bag of tricks from precision.The present invention can reach real-time follow-up, and speed is obviously better than most tracking.Secondly, the present invention is also obviously better than most of trackings in tracking accuracy.Finally, the present invention has avoided due to the tracking failure phenomenon of having selected wrong feature to cause, has effectively suppressed the impact on tracking results of bad feature.

Brief description of the drawings

Fig. 1 is process flow diagram of the present invention;

Fig. 2 is Margin schematic diagram of the present invention;

Fig. 3 is test video sequence david result comparison diagram.

Embodiment

In order to overcome above-mentioned the deficiencies in the prior art, the invention provides a kind of real-time detection based on compressed sensing feature selecting and follow the tracks of framework and tracking.The present invention can select the feature after compression, only uses the sample characteristics that discrimination is high to classify, and can obviously improve tracking velocity and tracking accuracy.

Below in conjunction with embodiment and accompanying drawing, the present invention is described in further detail, but embodiments of the present invention are not limited to this.

Framework and a tracking are followed the tracks of in real-time detection based on compressed sensing feature selecting, comprise the steps:

Step 1, initiation parameter, determines the target of t frame

described target location is a rectangle frame, is the target that needs tracking in frame;

comprise four parameters: the row-coordinate of target in this frame

row coordinate

width width and height height.

Step 2, makes l (x)={ row (x), col (x) } represent the position of sample x, comprises row-coordinate and the row coordinate of center of a sample,

in radius s pixel, gather positive sample set around $X_t^s=\left\{x_t\right|\left|\right|l\left(x_t\right)-l\left(x_t^{*}\right)\left|\right|\≤s\};$ ?

radius r around, gathers negative sample collection between β pixel $X_t^{r,\mathrm{\β}}=\left\{x_t\right|r\≤\left|\right|l\left(x_t^{*}\right)-l\left(x_t\right)\left|\right|\≤\mathrm{\β}\};$ Wherein s, r, β is empirical parameter, unit is pixel, s=4 in the present invention, the span that the span of r is 6～10, β is 15～30, and the collection quantity of the negative sample of getting is generally 1.5 times of positive sample collection quantity, and this test gathers 70 negative samples.

Step 3, aligns negative sample and concentrates each sample extract n class Lis Hartel and levy f (x _t)={ f ₁(x _t), f ₂(x _t) ..., f _n(x _t), step 3 is further comprising the steps of:

A stochastic matrix R ∈ R ^{n × m}can be by a high dimension vector x ∈ R ^mbe mapped to the vector space v ∈ R of low-dimensional ⁿ, v=Rx, here n < < m.If stochastic matrix R meets the limited isometry of RIP() condition, the position relationship of former vector in vector space that can preserve value of the vector after shining upon.Gauss's matrix meets the limited isometry of RIP() condition, but Gauss's matrix sparse not (in matrix neutral element number be directly proportional to the sparse degree of matrix), if can increase calculated amount for compressing.

3.1 are provided with template set T={h _ij} _{i=1:w; J=1:h}, wherein element definition is:

Being w for width, is highly the sample of h, carries out convolution with template set T, altogether can generate m=(wh) ²individual feature;

3.2 use stochastic matrix R ^{n × m}the m of a sample feature is compressed, and n represents the vectorial dimension after compression.Element r in R _ijmeet the following conditions:

Use this matrix to compress the high dimensional feature of sample x.After compression, be characterized as f (x _t)={ f ₁(x _t), f ₂(x _t) ..., f _n(x _t)=RT (x _t), wherein T (x _t) represent that practical template set T is to sample x _tcarry out the vector of m feature of convolution extraction; Wherein, empirical parameter ζ represents degree of rarefication, empirical parameter ζ=4 in the present invention.

Step 4, is used the features training Weak Classifier of positive sample and negative sample, and step 4 is further comprising the steps of:

The feature of the 4.1 positive samples of hypothesis and negative sample meets normal distribution, and positive sample obedience sample characteristics average is μ ₁, sample characteristics standard deviation is σ ₁normal distribution N (μ ₁, σ ₁), be designated as p (f (x) | y=1)～N (μ ₁, σ ₁), it is μ that negative sample is obeyed sample characteristics average ₀, sample characteristics standard deviation is σ ₀normal distribution N (μ ₀, σ ₀), be designated as p (f (x) | y=0)～N (μ ₀, σ ₀); Utilize positive negative sample to obtain parameter μ ₁, σ ₁, μ ₀and σ ₀;

The parameter update mode of positive sample is: ${\mathrm{\&mu;}}_t^1=\mathrm{\&lambda;}{\mathrm{\&mu;}}_{t-1}^1+(1-\mathrm{\&lambda;}){\widetilde{\mathrm{\&mu;}}}_t^1$

${\mathrm{\&sigma;}}_t^1=\sqrt{\mathrm{\&lambda;}{\left({\mathrm{\&sigma;}}_{t-1}^1\right)}^2+(1-\mathrm{\&lambda;}){\left({\widetilde{\mathrm{\&sigma;}}}_t^1\right)}^2+\mathrm{\&lambda;}(1-\mathrm{\&lambda;}){({\mathrm{\&mu;}}_{t-1}^1-{\mathrm{\&mu;}}_t^1)}^2}$

Wherein, λ is learning rate,

represent the positive sample characteristics average of t moment of directly obtaining, represent the study positive sample characteristics average in rear t moment,

represent the study positive sample characteristics average in rear t-1 moment,

represent the standard deviation of the positive sample characteristics in rear t moment of study,

the standard deviation of the positive sample characteristics that expression Direct Sampling is obtained,

represent the standard deviation of the positive sample in t-1 moment;

The parameter update mode of negative sample is: ${\mathrm{\&mu;}}_t^0=\mathrm{\&lambda;}{\mathrm{\&mu;}}_{t-1}^0+(1-\mathrm{\&lambda;}){\widetilde{\mathrm{\&mu;}}}_t^0$

${\mathrm{\&sigma;}}_t^0=\sqrt{\mathrm{\&lambda;}{\left({\mathrm{\&sigma;}}_{t-1}^0\right)}^2+(1-\mathrm{\&lambda;}){\left({\widetilde{\mathrm{\&sigma;}}}_t^0\right)}^2+\mathrm{\&lambda;}(1-\mathrm{\&lambda;}){({\mathrm{\&mu;}}_{t-1}^0-{\mathrm{\&mu;}}_t^0)}^2}$

Wherein, λ is learning rate,

represent the t moment negative sample characteristic mean of directly obtaining,

represent the study negative sample characteristic mean in rear t moment,

represent the study negative sample characteristic mean in rear t-1 moment,

represent the standard deviation of the negative sample feature in rear t moment of study,

the standard deviation of the negative sample feature that expression Direct Sampling is obtained,

represent the standard deviation of the negative sample in t-1 moment;

4.2 Weak Classifier

h _k(x) represent sample x to use k tagsort.Wherein

f _k(x) k the feature of expression sample x;

Wherein said learning rate λ gets 0.85.

Step 5, defining classification interval is assessed each Weak Classifier, and selects Weak Classifier to be used for classification, and step 5 comprises the following steps:

5.1 definition $m\arg \mathrm{in}\left(h_k\right)=\frac{1}{n_p}\underset{i=1}{\overset{n_p}{\mathrm{\&Sigma;}}}{h}_k\left(x_i\right)-\frac{1}{n_n}\underset{j=1}{\overset{n_n}{\mathrm{\&Sigma;}}}{h}_k\left(x_j\right),$ Wherein n _prepresent the number of positive sample, n _nrepresent the number of negative sample, $x_i\&Element;X_t^s,x_j\&Element;X_t^{r,\mathrm{\&beta;}};$

As shown in Figure 2, circle represents positive sample, and triangle represents negative sample.Horizontal ordinate represents first feature of sample, and ordinate represents second feature of sample.In the time using sample Second Characteristic to classify, can better positive and negative sample area be separated.The Weak Classifier that margin is larger, shows that the separating capacity of this sorter is stronger.

5.2 select front K maximum feature of margin value

composition strong classifier, wherein i _krepresent to select the feature for classifying;

The strong classifier of 5.3 t frames is defined as

Step 6, is made up of the strong classifier of t frame the Weak Classifier choosing

Step 7, at t+1 frame

collecting test sample in radius λ around

use H _t(x) test sample book is classified, in step 7, comprises the following steps:

7.1 at t+1 frame

collecting test sample in radius λ pixel around

7.2 obtain the classification results of each sample

7.3 targets are the position of target in t+1 frame is

Step 8, classification results is the position at t+1 frame as target

Step 9, if t+1 frame is not last frame, makes t=t+1, returns to step 1.

As shown in table 1, framework and tracking (CFS) are followed the tracks of in a kind of real-time detection based on compressed sensing feature selecting, and on each tracking and testing collection, to carry out test chart now good, average accuracy (average) is higher than additive method, wherein (CT) is compressed sensing tracking, is (FRAG) piecemeal tracking.Illustrate that the present invention (CFS) is better than additive method in tracking accuracy.

Table 1 is followed the tracks of accuracy

Cycle tests	CT	FRAG	CFS
				bolt	0.0067	0.1233	0.8523
david	0.3843	0.0955	0.7622
				Tiger	0.3894	0.3037	0.5473
FaceOcc1	0.8789	0.9899	0.9092
				FaceOcc2	0.6995	0.7217	0.6897
Fish	0.8739	0.5294	0.6933
				Jumping	0.0064	0.8371	0.8037
Tigger2	0.4188	0.1406	0.6423
				Shaking	0.0411	0.0877	0.5863
average	0.411	0.425423	0.7207

Claims (9)

1. framework and a tracking are followed the tracks of in the real-time detection based on compressed sensing feature selecting, it is characterized in that: comprise the steps:

Step 1, initiation parameter, determines the target of t frame

described target location is a rectangle frame, is the target that needs tracking in frame;

comprise four parameters: the row-coordinate of target in this frame row coordinate

width width and height height;

Step 2, makes l (x)={ row (x), col (x) } represent the position of sample x, comprises row-coordinate and the row coordinate of center of a sample, in radius s pixel, gather positive sample set around $X_t^s=\left\{x_t\right|\left|\right|l\left(x_t\right)-l\left(x_t^{*}\right)\left|\right|\&le;s\};$ ?

radius r around, gathers negative sample collection between β pixel $X_t^{r,\mathrm{\&beta;}}=\left\{x_t\right|r\&le;\left|\right|l\left(x_t^{*}\right)-l\left(x_t\right)\left|\right|\&le;\mathrm{\&beta;}\};$ Wherein s, r, β is empirical parameter, unit is pixel;

Step 3, aligns negative sample and concentrates each sample

extract n class Lis Hartel and levy f (x _t)={ f ₁(x _t), f ₂(x _t) ..., f _n(x _t);

Step 4, is used the features training Weak Classifier of positive sample and negative sample;

Step 5, defining classification interval is assessed each Weak Classifier, and selects Weak Classifier to be used for classification;

Step 6, is made up of the strong classifier of t frame the Weak Classifier choosing

Step 7, at t+1 frame

collecting test sample in radius λ around

use H _t(x) test sample book is classified;

Step 8, classification results is the position at t+1 frame as target

Step 9, if t+1 frame is not last frame, makes t=t+1, returns to step 1.

2. framework and tracking are followed the tracks of in a kind of real-time detection based on compressed sensing feature selecting according to claim 1, it is characterized in that: s=4 in described step 2, the span that the span of r is 6～10, β is 15～30.

3. framework and tracking are followed the tracks of in a kind of real-time detection based on compressed sensing feature selecting according to claim 1, it is characterized in that: further comprising the steps of in described step 3:

3.1 are provided with template set T={h _ij} _{i=1:w; J=1:h}, wherein element definition is:

Being w for width, is highly the sample of h, carries out convolution with template set T, altogether can generate m=(wh) ²individual feature;

3.2 use stochastic matrix R ^{n × m}the m of a sample feature is compressed, and n represents the vectorial dimension after compression, the element r in R _ijmeet the following conditions:

Use this matrix to compress the high dimensional feature of sample x, after compression, be characterized as f (x _t)={ f ₁(x _t), f ₂(x _t) ..., f _n(x _t)=RT (x _t), wherein T (x _t) represent that practical template set T is to sample x _tcarry out the vector of m feature of convolution extraction; Wherein, empirical parameter ζ represents degree of rarefication.

4. framework and tracking are followed the tracks of in a kind of real-time detection based on compressed sensing feature selecting according to claim 3, it is characterized in that: in described step 3.2, represent degree of rarefication empirical parameter ζ=4.

5. framework and tracking are followed the tracks of in a kind of real-time detection based on compressed sensing feature selecting according to claim 1, it is characterized in that: described step 4 is further comprising the steps of:

The feature of the 4.1 positive samples of hypothesis and negative sample meets normal distribution, and positive sample obedience sample characteristics average is μ ₁, sample characteristics standard deviation is σ ₁normal distribution N (μ ₁, σ ₁), be designated as p (f (x) | y=1)～N (μ ₁, σ ₁), it is μ that negative sample is obeyed sample characteristics average ₀, sample characteristics standard deviation is σ ₀normal distribution N (μ ₀, σ ₀), be designated as p (f (x) | y=0)～N (μ ₀, σ ₀); Utilize positive negative sample to obtain parameter μ ₁, σ ₁, μ ₀and σ ₀;

The parameter update mode of positive sample is: ${\mathrm{\&mu;}}_t^1=\mathrm{\&lambda;}{\mathrm{\&mu;}}_{t-1}^1+(1-\mathrm{\&lambda;}){\widetilde{\mathrm{\&mu;}}}_t^1$

${\mathrm{\&sigma;}}_t^1=\sqrt{\mathrm{\&lambda;}{\left({\mathrm{\&sigma;}}_{t-1}^1\right)}^2+(1-\mathrm{\&lambda;}){\left({\widetilde{\mathrm{\&sigma;}}}_t^1\right)}^2+\mathrm{\&lambda;}(1-\mathrm{\&lambda;}){({\mathrm{\&mu;}}_{t-1}^1-{\mathrm{\&mu;}}_t^1)}^2}$

Wherein, λ is learning rate,

represent the positive sample characteristics average of t moment of directly obtaining,

represent the study positive sample characteristics average in rear t moment,

represent the study positive sample characteristics average in rear t-1 moment,

represent the standard deviation of the positive sample characteristics in rear t moment of study,

the standard deviation of the positive sample characteristics that expression Direct Sampling is obtained,

represent the standard deviation of the positive sample in t-1 moment;

The parameter update mode of negative sample is: ${\mathrm{\&mu;}}_t^0=\mathrm{\&lambda;}{\mathrm{\&mu;}}_{t-1}^0+(1-\mathrm{\&lambda;}){\widetilde{\mathrm{\&mu;}}}_t^0$

${\mathrm{\&sigma;}}_t^0=\sqrt{\mathrm{\&lambda;}{\left({\mathrm{\&sigma;}}_{t-1}^0\right)}^2+(1-\mathrm{\&lambda;}){\left({\widetilde{\mathrm{\&sigma;}}}_t^0\right)}^2+\mathrm{\&lambda;}(1-\mathrm{\&lambda;}){({\mathrm{\&mu;}}_{t-1}^0-{\mathrm{\&mu;}}_t^0)}^2}$

Wherein, λ is learning rate, represent the t moment negative sample characteristic mean of directly obtaining,

represent the study negative sample characteristic mean in rear t moment,

represent the study negative sample characteristic mean in rear t-1 moment,

represent the standard deviation of the negative sample feature in rear t moment of study,

the standard deviation of the negative sample feature that expression Direct Sampling is obtained,

represent the standard deviation of the negative sample in t-1 moment;

4.2 Weak Classifier

h _k(x) represent sample x to use k tagsort, wherein

f _k(x) k the feature of expression sample x.

6. framework and tracking are followed the tracks of in a kind of real-time detection based on compressed sensing feature selecting according to claim 5, it is characterized in that: the learning rate λ in described step 4.1 and 4.2 gets 0.85.

7. framework and tracking are followed the tracks of in a kind of real-time detection based on compressed sensing feature selecting according to claim 1, it is characterized in that: in described step 5, comprise the following steps:

5.1 definition $m\arg \mathrm{in}\left(h_k\right)=\frac{1}{n_p}\underset{i=1}{\overset{n_p}{\mathrm{\&Sigma;}}}{h}_k\left(x_i\right)-\frac{1}{n_n}\underset{j=1}{\overset{n_n}{\mathrm{\&Sigma;}}}{h}_k\left(x_j\right),$ Wherein n _prepresent the number of positive sample, n _nrepresent the number of negative sample,

5.2 select front K maximum feature of margin value composition strong classifier, wherein ik represents to select the feature for classifying;

The strong classifier of 5.3 t frames is defined as

8. framework and tracking are followed the tracks of in a kind of real-time detection based on compressed sensing feature selecting according to claim 1, it is characterized in that: in described step 7, comprise the following steps:

7.1 at t+1 frame

collecting test sample in radius λ pixel around

7.2 obtain the classification results of each sample

7.3 targets are

the position of target in t+1 frame is

9. framework and tracking are followed the tracks of in a kind of real-time detection based on compressed sensing feature selecting according to claim 1, it is characterized in that: in described step 2, the quantity of the negative sample of getting is 1.5 times of positive sample collection quantity.

Images