CN103324953B - Video monitoring multi-target detection and tracking - Google Patents

Abstract

The invention discloses a kind of video monitoring multi-target detection and tracking, comprising: by Adaboost learning algorithm, the target sample of input is trained, generate Adaboost cascade classifier; By described Adaboost cascade classifier, the view data of input is carried out to target detection, the confidence level of localizing objects export target; The target of locating using described Adaboost cascade classifier, as observation, is carried out target following by many Hypothesis Tracking Algorithms. Application technical solution of the present invention, can improve the accuracy rate of target detection, and loss and rate of false alarm while reducing multiple target tracking.

Description

Video monitoring multi-target detection and tracking

Technical field

The present invention relates to technical field of video monitoring, particularly relate to a kind of video monitoring multi-target detection and track sideMethod.

Background technology

Multi-target detection and tracking technique are the key technologies of video monitoring, detect in monitoring scene interested manyIndividual target, need to follow the tracks of each target, determines the movement locus of target, for subsequent analysis.

In the prior art, for target detection, be generally to adopt background modeling algorithm, background modeling algorithm is video figureA kind of method of moving object detection in picture, its basic thought is that the background of image is carried out to modeling, once Background Modeling,Current image and background model are compared, determine foreground target according to comparative result. Background modeling algorithm comprises colorBackground model and grain background model.

Prior art, detecting after target, comprises data correlation and dbjective state (position, size, speed to multiple target trackingDegree etc.) prediction. Data correlation is for determining the corresponding relation between target and observation (target detection result), and common algorithm comprisesField association recently, joint probability filtering etc. Dbjective state prediction is the state at next frame image for estimating target, commonAlgorithm has linear prediction, particle filter etc.

Inventor finds that prior art at least exists following technical problem under study for action: prior art is carried out target detection and adoptedWith the illumination condition of background modeling method to environment and the motion of background more responsive, the shade of moving object also can be to relativelyResult exerts an influence. Therefore background modeling method is not suitable for illumination condition variation and boisterous scene, and ratio of precision is poor. ThisOuter existing multiple target tracking algorithm target loss, rate of false alarm are higher.

Summary of the invention

Based on this, be necessary for above-mentioned technical problem, provide a kind of video monitoring multi-target detection and tracking, energyEnough improve the accuracy rate of target detection, and loss and rate of false alarm while reducing multiple target tracking.

A kind of video monitoring multi-target detection and tracking, comprising:

By Adaboost learning algorithm, the target sample of input is trained, generate Adaboost cascade classifier;

By described Adaboost cascade classifier, the view data of input is carried out to target detection, localizing objects outputThe confidence level of target;

The target of locating using described Adaboost cascade classifier, as observation, is carried out target by many Hypothesis Tracking AlgorithmsFollow the tracks of.

Above-mentioned video monitoring multi-target detection and tracking, the target sample by Adaboost learning algorithm to inputTrain, generate Adaboost cascade classifier, then carry out target detection by Adaboost cascade classifier, with respect to existingThere is the background model modeling of technology to improve the accuracy rate of target detection. In addition carrying out target by many Hypothesis Tracking Algorithms followsTrack, has reduced target loss, rate of false alarm.

Brief description of the drawings

The video monitoring multi-target detection that Fig. 1 provides for an embodiment and the schematic flow sheet of tracking;

The schematic flow sheet of the Adaboost learning algorithm that Fig. 2 provides for an embodiment;

The schematic flow sheet of many Hypothesis Tracking Algorithms that Fig. 3 provides for an embodiment.

Detailed description of the invention

In order to make object of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, rightThe present invention is further elaborated. Should be appreciated that specific embodiment described herein is only in order to explain the present invention, andBe not used in restriction the present invention.

Referring to Fig. 1, in one embodiment, provide a kind of video monitoring multi-target detection and tracking, comprising:

Step S102, trains the target sample of input by Adaboost learning algorithm, generates Adaboost levelConnection grader.

In image object recognition technology field, usually need, according to any given image, it to be entered by certain strategyLine search to be to determine whether to contain specific objective, such as car, ship, face etc. Adaboost is as a kind of iterative algorithm, and it is basicThought is exactly: for a concrete identification problem, by certain algorithm, one group of Weak Classifier is promoted to strong classifier. ThisIf in grader just refer to the grader that the identification of target is only better than to random conjecture, strong classifier can think that it is logicalThe study of crossing certain sample has reached desirable object recognition rate. In this step, by Adaboost learning algorithm to manuallyThe sample training gathering, obtains Adaboost cascade classifier, for carrying out follow-up target detection.

Step S104, carries out target detection, localizing objects by Adaboost cascade classifier to the view data of inputAnd the confidence level of export target.

The view data of input can be video sequence or static image, adopts Adaboost cascade classifier to imageData are carried out target detection, and it is really target that the target detecting can not be defined as completely, need the confidence level of export target to doDetect for determining the probability that target is real goal.

Step S106, the target of locating using Adaboost cascade classifier, as observation, is entered by many Hypothesis Tracking AlgorithmsRow target following.

Video monitoring multi-target detection and tracking that the present embodiment provides, by Adaboost learning algorithm to inputTarget sample train, generate Adaboost cascade classifier, then carry out target inspection by Adaboost cascade classifierSurvey, improved the accuracy rate of target detection with respect to the background model modeling of prior art. In addition follow the tracks of and calculate by many hypothesisMethod is carried out target following, has reduced target loss, rate of false alarm.

Referring to Fig. 2, in one embodiment, Adaboost learning algorithm flow process comprises

Step S202, Weak Classifier training.

Weak Classifier training algorithm, trains and obtains Weak Classifier according to the target sample of input.

Step S204, Weak Classifier is demarcated.

Weak Classifier calibration algorithm, the Weak Classifier that described training is obtained reorders.

In one embodiment, Weak Classifier training algorithm comprises:

(1) the initial number a of input negative sample, the initial number b of positive sample and the number T of Weak Classifier, the negative sample of inputThis and positive sample (x₁,y₁),...,(x_a+b,y_a+b), wherein y_i=0 represents negative sample, y_i=1 represents positive sample.

The positive sample of input and negative sample are by manually gathering. Positive sample is for comprising the image that will detect target, negative sampleFor not comprising the image that will detect target.

(2) initialization sample weight: if y_i=0,If y_i=1,

When arranging, initialization thinks that the probability of positive sample and negative sample is for being evenly distributed.

(3) for y_i=1, complete following circulation:

A. the grader h of each characteristics of image training to sample_j, the error of grader is ${\mathrm{\ξ}}_i={\mathrm{\Σ}}_i{w}_i|h_j\left(x_i\right)-y_i|.$

In the present embodiment, characteristics of image can include but not limited to it is to comprise expansion histogram of gradients feature or Haar spyLevy

B. select to have minimal error ξ_tGrader h_t, establishα_t=-logβ_t，c_t=α_th_t。

C. the sample weights of existing negative sample is carried out to convergent-divergent: rightIf, y_i=0, by w_t-1,iBe updated to

D. to k=1 ..., K, increases sample (x_N+k, 0), weight isAnd

K is setup parameter, and value is generally several thousand, the weight of the negative sample of increase and the negative sample of convergent-divergent and be normalNumber. The output of all Weak Classifiers of current strong classifier is cumulativeBe not less than cumulative two points of the weight of Weak ClassifierOne of, show that sample is correctly classified.

E. the number a of negative sample is updated to a+K,

F. dwindling can be by h_jThe weight of the sample of correct classification:By w_t-1,iBe updated to β_tw_t-1,i。

G. the weight of all samples is normalized:

H. export Weak Classifier c_t。

In one embodiment, Weak Classifier calibration algorithm comprises:

(1) sample set X={ (x is demarcated in input₁,y₁),...,(x_N,y_N), wherein y_i=0,1 is respectively negative sample and positive sampleThis, a=∑ (1-y_i),b=∑y_iFor the quantity of negative sample and positive sample.

(2) input v₁,...,v_T, be the Weak Classifier c of described Weak Classifier training algorithm output₁,...,c_TFail to reportRate.

(3) input the Weak Classifier c that described Weak Classifier training algorithm is exported_tSet { C}.

(4) initialization sample response d_0,i=0, positive sample reject rate p=0, current negative sample number A ← a.

(5) to t=1 ..., T, carries out following circulation:

A. more the first month of the lunar year, sample reject rate p was p+v_t, the number of upgrading negative sample and positive sample is respectively $a_t={\mathrm{\Σ}}_{(x_i,y_i)\∈X}(1-y_i),$ $b_t={\mathrm{\Σ}}_{(x_i,y_i)\∈X}{y}_i,$

B. { C}, select Weak Classifier from Weak Classifier set $q\left(t\right)=\arg {\max }_j\underset{i}{\mathrm{\Σ}}{f}_{t,i,j}{y}_i/b_t-\underset{i}{\mathrm{\Σ}}{f}_{t,i,j}(1-y_i)/a_t,$ Wherein, f_t,i,j=d_t,i-1+c_j，

C. upgrade sample responses d_t,i=d_t-1,i+c_q(t)(x_i)，

D. determine threshold values r_t, the r that all predicted values are no more than_tThe strong classifier output of sample and taking advantage of of sample labelLong-pending cumulative sum is less than a designated value pb: ∑_ipred(d_t,i≤r_t)y_i≤ pb, described strong classifier is current Weak ClassifierIt is cumulative,

E. upgradeX←X-{(x_i,y_i)|d_t,i<r_t}，C←C-{c_q(t)}，

(6) choose at random A_tIndividual negative sample, until determine k=1 ..., K, adds sample (x_N+k, 0), it is output as $d_{t,N+k}=\underset{j=1}{\overset{t}{\mathrm{\&Sigma;}}}{c}_q\left(j\right)\left(x_{N+k}\right),$ And meet $\&ForAll;j\&Element;\{1,...,t\},$ $\underset{m=1}{\overset{j}{\mathrm{\&Sigma;}}}{c}_{q\left(m\right)}\left(x_{N+k}\right)\&GreaterEqual;r_j,$

Upgrade and demarcate sample set element number N ← N+K, upgrade current negative sample number A ← A+A_t。

(7) deferent segment output function c_qAnd r (t)_t。

In one embodiment, by described Adaboost cascade classifier, the view data of input is carried out to target inspectionSurvey, the confidence level of localizing objects export target comprises:

(1) input image data is to Adaboost cascade classifier.

View data is video sequence or still image.

(2) the output s=0 of initialization Adaboost cascade classifier.

(3) to t=1 ..., T circulates:

Calculate the calibrated Weak Classifier output of Weak Classifier c_t(x)，

s=s+c_t(x)

If s < ξ_t, the view data of returning to current input is negative sample.

(4) view data of returning to current input is positive sample and s.

(5) calculate confidence level $P0=\frac{e^s}{e^{-s}+e^s}.$

Referring to Fig. 3, in one embodiment, many Hypothesis Tracking Algorithms comprise:

Step S302, carries out data to target and observation and clusters.

Data cluster according to being correlation threshold, if target and observation between distance drop on outside correlation threshold,Do not belong to same bunch, the calculating of distance adopts Euclidean distance.

Step S304, determines and the associated cost of each target and observation forms incidence matrix.

By determining incidence matrix, provide the similarity between target and observation.

Step S306, utilizes Kalman filter to predict existing target.

Step S308, utilizes optimization algorithm to calculate optimum M allocative decision of this incidence matrix, forms M associated vacationAnd if the probability of each relevance assumption.

Do not enumerating under all situations, M optimum allocative decision enumerated out. For example first structure likely dividesQueue, adopt Hungary's optimal algorithm to find optimum linearity to distribute at every turn, in distribution queue, delete this optimal sortingJoin, then from remaining distribution queue, find an optimum allocation, as suboptimum allocative decision. So circulation just can be found M M timeThe allocative decision of individual optimum.

Step S310, carries out beta pruning to hypothesis.

Hypothesis is carried out to beta pruning, remove the solution of suboptimum, only retain optimum solution, accelerated the computational speed of algorithm.

Step S312, the highest hypothesis of output confidence level, as tracking results.

The above embodiment has only expressed several embodiment of the present invention, and it describes comparatively concrete and detailed, but alsoCan not therefore be interpreted as the restriction to the scope of the claims of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the inventive concept of the premise, can also make some distortion and improvement, these all belong to guarantor of the present inventionProtect scope. Therefore, the protection domain of patent of the present invention should be as the criterion with claims.

Claims (6)

1. video monitoring multi-target detection and a tracking, comprising:

By Adaboost learning algorithm, the target sample of input is trained, generate Adaboost cascade classifier;

By described Adaboost cascade classifier, the view data of input is carried out to target detection, localizing objects export targetConfidence level;

The target of locating using described Adaboost cascade classifier, as observation, is carried out target by many Hypothesis Tracking Algorithms and is followedTrack;

Described Adaboost learning algorithm comprises Weak Classifier training algorithm and Weak Classifier calibration algorithm;

Described Weak Classifier training algorithm, trains and obtains Weak Classifier according to the target sample of input, comprising: the negative sample of inputThis initial number a, the initial number b of positive sample and the number T of Weak Classifier, input negative sample and positive sample (x₁,y₁),...,(x_a+b,y_a+b), wherein y_i=0 represents negative sample, y_i=1 represents positive sample, wherein, and 1≤i≤a+b;

Initialization sample weight: if y_i=0,If y_i=1,

For y_i=1, complete following circulation:

To a grader h of each characteristics of image training of sample_j, the error of grader is ξ_i＝∑_iw_i|h_j(x_i)-y_i|，

Selection has minimal error ξ_tGrader h_t, establishα_t＝-logβ_t，c_t＝α_th_t, wherein, 1≤t≤T;

The sample weights of existing negative sample is carried out to convergent-divergent:By w_t-1,iBe updated to

To k=1 ..., K, increases sample (x_N+k, 0), weight isAndWherein, α_jFor the weight of Weak Classifier, N=a+b;

The number a of negative sample is updated to a+K,

Dwindling can be by h_jThe weight of the sample of correct classification:By w_t-1,iBe updated to β_tw_t-1,i，

The weight of all samples is normalized:

Output Weak Classifier c_t；

Described Weak Classifier calibration algorithm, the Weak Classifier that described training is obtained reorders.

2. method according to claim 1, is characterized in that, described characteristics of image comprise expansion histogram of gradients feature orHaar feature.

3. method according to claim 1, is characterized in that, described Weak Classifier calibration algorithm comprises:

Sample set X={ (x is demarcated in input₁,y₁),...,(x_N,y_N), wherein y_i=0,1 is respectively negative sample and positive sample, a=∑(1-y_i),b＝∑y_iFor the number of original negative sample and positive sample;

Input v₁,...,v_T, be the Weak Classifier c of described Weak Classifier training algorithm output₁,...,c_TRate of failing to report;

Input the Weak Classifier c of described Weak Classifier training algorithm output_tSet { C};

Initialization sample response d_0,i=0, positive sample reject rate p=0, current negative sample number A ← a;

To t=1 ..., T, carries out following circulation:

More the first month of the lunar year, sample reject rate p was p+v_t, the number of upgrading negative sample and positive sample is respectively $a_t={\mathrm{\&Sigma;}}_{(x_i,y_i)\&Element;X}(1-y_i),b_t={\mathrm{\&Sigma;}}_{(x_i,y_i)\&Element;X}{y}_i,$

{ C}, select Weak Classifier from Weak Classifier set $q\left(t\right)={\mathrm{argmax}}_j\underset{i}{\&Sigma;}{f}_{t,i,j}{y}_i/b_t-\underset{i}{\&Sigma;}{f}_{t,i,j}(1-y_i)/a_t,$ ItsIn, f_t,i,j＝d_t,i-1+c_j，

Upgrade sample responses d_t,i＝d_t-1,i+c_q(t)(x_i)，

Determine threshold values r_t, the r that all predicted values are no more than_tThe strong classifier output of sample and the tiring out of the product of sample labelAdd and be less than a designated value pb: ∑_ipred(d_t,i≤r_t)y_i< pb, what described strong classifier was current Weak Classifier is cumulative,

UpgradeX←X-{(x_i,y_i)|d_t,i<r_t}，C←C-{c_q(t)}，

Choose at random A_tIndividual negative sample, until determine k=1 ..., K, adds sample (x_N+k, 0), it is output as $d_{t,N+k}=\underset{j=1}{\overset{t}{\&Sigma;}}{c}_q\left(j\right)\left(x_{N+k}\right),$ And meet $\&ForAll;j\&Element;\{1,\mathrm{...},t\},\underset{m=1}{\overset{j}{\&Sigma;}}{c}_{q\left(m\right)}\left(x_{N+k}\right)\&GreaterEqual;r_j,$

Upgrade and demarcate sample set element number N ← N+K, upgrade current negative sample number A ← A+A_t，

Deferent segment output function c_qAnd r (t)_t。

4. method according to claim 3, is characterized in that, described by described Adaboost cascade classifier to inputView data carry out target detection, the step of the confidence level of localizing objects export target comprises algorithm:

Input image data is to Adaboost cascade classifier;

Initialize the output s=0 of Adaboost cascade classifier;

To t=1 ..., T circulates:

Calculate the calibrated Weak Classifier output of Weak Classifier c_t(x)，

s＝s+c_t(x)；

If s < ξ_t, the view data of returning to current input is negative sample;

The view data of returning to current input is positive sample and s;

Confidence level $P0=\frac{e^s}{e^{-s}+e^s}.$

5. method according to claim 3, is characterized in that, described view data is video sequence or still image.

6. method according to claim 1, is characterized in that, described many Hypothesis Tracking Algorithms comprise:

Target and observation are carried out to data to cluster;

Determine the associated cost of each target and observation, form incidence matrix;

Utilize Kalman filter to predict existing target;

Utilize optimization algorithm to calculate optimum M allocative decision of this incidence matrix, form M relevance assumption associated with eachThe probability of supposing;

Hypothesis is carried out to beta pruning;

The highest hypothesis of output confidence level, as tracking results.