CN106156808A - The on-line study method and apparatus of object classifier and detection method and equipment - Google Patents

Abstract

Present invention provide for the on-line study method and apparatus of a group objects grader and for detecting the method and apparatus of object in image sequence.This on-line study method (a wherein said group objects grader comprises at least one one-class classifier and at least one binary classifier) being used for a group objects grader includes: extract characteristic vector from newly added sample；Update at least one at least one one-class classifier described based on the support vector of the characteristic vector extracted and at least one binary classifier described；The support vector of at least one one-class classifier described in based on the characteristic vector extracted and being partly or wholly updated, updates at least one at least one binary classifier described.Utilize the present invention, high-caliber false alarm rate can be reduced to extremely low level being maintained at recall rate simultaneously.

Description

The on-line study method and apparatus of object classifier and detection method and equipment

Technical field

The present invention relates to image procossing, computer vision (computer vision) and pattern to know Not, and be particularly directed to a group objects grader on-line study method and apparatus and For detecting the method and apparatus of object in image sequence.

Background technology

Nearest decades are at detection special object or species (face, people, automobile etc.) aspect Have been achieved with very big progress.Major part detector must be by off-line training, and it often comprises profit With the intensive study stage of the larger set of the training example of hand labeled.Currently, get more and more User tend to shoot special object.This means that user needs be not traditional as face/ Detector as dog/cat detector, but can learn user oneself object (for example, he Pet) the detector of outward appearance (appearance).For example, user is frequently desired to clapping Their pet is automatically focused on when taking the photograph.

Having been developed for the technology of a kind of novelty, referred to as " user registers object detection (User Registered Object Detection) ", it can provide as positive sample based on by user One or several object images detect user and register object.But it is typically difficult to only lead to Cross and use the positive sample of very small set to detect object from sequence of frames of video exactly.

Therefore it has already been proposed that a solution, it uses based on being used for detecting user's registration The one-class classifier (one-class classifier) of the support vector of object.This method can Based on the positive sample (that is, the object images registered by user) of minority, even only one just Sample carries out on-line study to one-class classifier.Owing to not using any negative sample, therefore should One-class classifier can be suitable for any scene.This method can be by adjustable parameter by list class The false alarm rate (false alarm rate) of grader is maintained at stable level.

However it remains to the need reducing false alarm rate in some special applications further Want.For example, in the supervision (surveillance) user being registered to object is applied, classification Device needs to work long hours in limited scene.Extremely low false alarm rate for various lighting conditions It is important requirement with high robust.

Content of the invention

In view of foregoing, the present inventor proposes by entering a group objects grader The new solution registering object detection for user of row on-line study, it can will be reported by mistake Alert rate is maintained at extremely low level.This classifiers include one or more one-class classifier and One or more binary classifiers (binary classifier), and the two type point Class device is all based on supporting vector and will share support when they are by on-line study in-between Vector, thus improve on-line study effect.Furthermore, it is possible to utilize in-between share support to Such one group of on-line study grader of amount detects user from sequence of frames of video more accurately Registration object.

Note that in the prior art, binary classifier is to be based only on huge number of positive sample Originally with negative sample off-line learning.Binary classifier cannot be used for simply using small set sample User registers object detection, because its false alarm rate will be very high.This is because study two Class grader needs many enough negative samples, but during user registers object detection, User is not provided that enough negative samples to cover all scenario in background.When to be detected When there is the negative sample of the unknown in scene, false alarm rate will be higher, say, that many is not It is that user registers the object of object and will be considered object.Additionally, by use directly feed back right Detector is made to tend to drift about (drift) in the on-line study of binary classifier, and mistake Ground updates and detector will be made to be in insecure state.But, by further investigation, this Bright it was found by the inventors that by sharing in-between during user registers object detection Hold vector to be used in combination with binary classifier with one-class classifier, false alarm rate can be improved.

One aspect of the present invention provides a kind of on-line study for a group objects grader Method, a wherein said group objects grader comprise at least one one-class classifier and at least one Binary classifier, described method includes: characteristic vector pickup step, from newly added sample Extract characteristic vector；One-class classifier updates step, based on the characteristic vector extracted and institute The support vector stating at least one binary classifier updates at least one one-class classifier described In at least one；Binary classifier update step, based on the characteristic vector extracted and Described one-class classifier update in step be partly or wholly updated described at least The support vector of one one-class classifier, updates at least one binary classifier described extremely Few one.

Another aspect of the present invention provides a kind of right for being had been determined as by use The positive sample of at least one of elephant detects the method for described object, described method bag in image sequence Include: one-class classifier produces step, according at least one positive sample described, produce based on being used for At least one one-class classifier of the support vector that described object is classified；Binary classifier Produce step, produce use based on first image and described one-class classifier of described image sequence In at least one binary classifier that described object is classified；Object detection step, passes through Use the group objects grader including described one-class classifier and described binary classifier in institute State in the second image in sequence and detect described object；Grader updates step, on using The on-line study method stated, updates a described group objects grader according to described testing result；With And follow-up detection and update step, for each the follow-up image in described sequence, logical Cross in the image using the group objects grader being updated over to continue in the rear and detect described object And update a described group objects grader subsequently.

An additional aspect of the present invention provides a kind of online for a group objects grader Habit equipment, a wherein said group objects grader comprises at least one one-class classifier and at least one Individual binary classifier, described equipment includes: characteristic vector pickup unit, is configured to add from newly The sample adding extracts characteristic vector；One-class classifier updating block, is configured to based on being carried The characteristic vector taking and the support vector of at least one binary classifier described update described At least one at least one one-class classifier；Binary classifier updating block, is configured to Based on the characteristic vector extracted and by described one-class classifier updating block partly Or at least one one-class classifier described updating fully support vector, update described in extremely At least one in a few binary classifier.

The still another aspect of the present invention provides a kind of right for being had been determined as by use The positive sample of at least one of elephant detects the equipment of described object, described equipment bag in image sequence Include: one-class classifier generation unit, be configured to, according at least one positive sample described, produce At least one one-class classifier based on the support vector for described object is classified；Two Class grader generation unit, is configured to based on the first image of described image sequence and described One-class classifier produces at least one binary classifier for classifying described object；Right It as detector unit, is configured to use and includes described one-class classifier and described two class classification In one group objects grader of device each follow-up image in described sequence, detection is described Object；And as above on-line study equipment, it is configured to according to by described object detection The testing result that unit obtains updates a described group objects grader.

In aforementioned aspect of the present invention, divide sharing from single class during the on-line study stage The support vector of class device and binary classifier.Therefore, will be tight by shared support vector Thickly combine the grader of the two type so that by the classification utilizing this classifiers to carry out More accurate.Additionally, more characteristic vector will be used for the on-line study of this classifiers, because of This shares and supports that vector is very suitable for on-line study application.

In addition, use this classifiers that single grader will be made to avoid drifting problem and therefore False alarm rate is maintained at extremely low level.Additionally, this classifiers can not only be by object and class As scene distinguish, and it can be avoided that the unknown negative sample also not found from us Noise.

In view of foregoing, utilize on-line study and detection method according to the present invention and set Standby, high-caliber false alarm rate can be reduced to extremely low water being maintained at recall rate simultaneously Flat, particularly with continuing and limited scene is such.Therefore, it is especially applicable to monitor Application etc..

According to following description referring to the drawings, other property features of the present invention and advantage will become Clearly.

Brief description

It is incorporated in specification and constitutes a part of accompanying drawing of specification show the present invention's Embodiment, and together with the description for the principle of the present invention is described.

Fig. 1 is the hardware configuration illustrating the ability to implement the computer system of embodiments of the invention Block diagram.

Fig. 2 show according to the first embodiment of the present invention for a group objects grader The flow chart of on-line study method.

Fig. 3 be exemplarily illustrated for determine for update the candidate feature of one-class classifier to The judgement hypersphere (decision hypersphere) of the one-class classifier of amount.

Fig. 4 show according to a first embodiment of the present invention for a group objects grader The block diagram of on-line study equipment 400.

Fig. 5 is illustrate according to a second embodiment of the present invention right for detecting in image sequence The flow chart of the method for elephant.

Fig. 6 is illustrate according to a second embodiment of the present invention right for detecting in image sequence The block diagram of the equipment 600 of elephant.

Fig. 7 illustrates an example of image pick up equipment according to other embodiments of the present invention.

Fig. 8 illustrate by the detection method of a pair grader of use according to embodiments of the present invention with Simply use the prior art detection method of one-class classifier compare obtained from compare knot Really.

Detailed description of the invention

Describe a preferred embodiment of the present invention below with reference to the accompanying drawings in detail.

Note that the similar project that similar reference numeral refers in figure with letter, thus Once a project defined in a width figure, avoids the need for discussing in figure afterwards.

In the disclosure, term " first ", " second " etc. are only used only in element or step Make a distinction between Zhou, and be not intended to mean that time sequencing, priority or importance.

(hardware configuration of computer system)

Fig. 1 is the hardware illustrating the ability to implement the computer system 1000 of embodiments of the invention The block diagram of configuration.

As shown in fig. 1, computer system includes computer 1110.Computer 1110 includes Processing unit the 1120th, the system storage connecting via system bus 1121 is the 1130th, fixing non- The 1140th, removable non-volatile memory interface the 1150th, user is defeated for volatile memory interface Incoming interface the 1160th, network interface the 1170th, video interface 1190 and output peripheral interface 1195.

System storage 1130 includes that ROM (read-only storage) 1131 and RAM is (random Access memory) 1132.BIOS (basic input output system) 1133 resides in ROM 1131 In.Operating system the 1134th, application program the 1135th, other program modules 1136 and some program Data 1137 reside in RAM 1132.

The fixed non-volatile memory 1141 of such as hard disk etc is connected to fixing non-volatile Memory interface 1140.Fixed non-volatile memory 1141 for example can store operating system 1144th, application program the 1145th, other program modules 1146 and some routine data 1147.

Such as floppy disk 1151 and CD-ROM drive 1155 etc removable non- Volatile memory is connected to removable non-volatile memory interface 1150.For example, floppy disk 1152 can be inserted in floppy disk 1151, and CD (CD) 1156 is permissible It is inserted in CD-ROM drive 1155.

The such as input equipment of microphone 1161 and keyboard 1162 etc is connected to user's input Interface 1160.

Computer 1110 can be connected to remote computer 1180 by network interface 1170. For example, network interface 1170 can be connected to remote computer 1180 via LAN 1171. Or, network interface 1170 may be coupled to modem (modulator-demodulator) 1172, And modem 1172 is connected to remote computer 1180 via wide area network 1173.

Remote computer 1180 can include the memory 1181 of such as hard disk etc, its storage Remote application 1185.

Video interface 1190 is connected to monitor 1191.

Output peripheral interface 1195 is connected to printer 1196 and loudspeaker 1197.

Computer system shown in Fig. 1 be merely illustrative and be never intended to the present invention, Its application or purposes carry out any restriction.

Computer system shown in Fig. 1 can be incorporated in any embodiment, can be as independent meter Calculation machine, or also can as the processing system in equipment, can remove one or more need not The assembly wanted, it is also possible to be added to one or more additional assembly.

(first embodiment)

Fig. 2 show according to the first embodiment of the present invention for a group objects grader The flow chart of on-line study method, it is single that a wherein said group objects grader comprises at least one Class grader and at least one binary classifier.

As in figure 2 it is shown, in characteristic vector pickup step 210, carry from newly added sample Take characteristic vector.Now, this group objects grader has been based on user and registers object images ( Individual or multiple) and the image of previously detection alternatively and be generated, and will be based on conduct The new images (for example, the frame of video of capture) of newly added sample comes by online updating.Obviously, The type of the characteristic vector extracted and the support vector of one-class classifier and binary classifier Type is identical so that these characteristic vectors can between one-class classifier and binary classifier quilt Share.

Then, update in step 220 at one-class classifier, based on the characteristic vector extracted with And the support vector of described at least one binary classifier updates at least one single class described and divides At least one in class device.

In one implementation, described one-class classifier update step 220 may include that right In one of at least one one-class classifier described one-class classifier, calculate and divide from this single class The cluster centre (clustering centre) of all support vectors of class device is to the feature extracted Each each in the described support vector of at least one binary classifier described in vector sum Distance；Set threshold value based on the support vector of described one-class classifier；From the feature extracted At least one binary classifier described in vector sum described support vector in select candidate feature to Amount, the calculating of each distance in wherein said candidate feature vector is less than described threshold value； And update the support vector of described one-class classifier by using described candidate feature vector.

An example of such one-class classifier update method is described more fully below.This area Artisans will appreciate that, the invention is not restricted to this specific example, but can use any Suitable mode updates one-class classifier.

In this example, one-class classifier is the one-class classifier based on SVDD, and it passes through Simply use positive sample training, and it is intended to obtain the spherical border with minimum volume Make this ball can surround positive sample as much as possible.The set of support vectors of one-class classifier is such as Under: SVs={ (x_i,α_i) | i=1,2 ... N_p, wherein x_iIt is to support vector and α_iIt is corresponding weight.Right Described support vector in the characteristic vector extracted and at least one binary classifier described In any one characteristic vector z, can calculate this one-class classifier all support to The distance of the cluster centre of amount.For from the characteristic vector extracted and described at least one two The described support vector of class grader select the decision function of candidate feature vector may is that

$f_{\mathrm{one}}\left(z\right)={R_z}^2=K(z,z)-2\underset{x_i\∈\mathrm{SVs}}{\mathrm{\Σ}}{a}_{i}K(x_i,z)+\underset{x_i\∈\mathrm{SVs}}{\mathrm{\Σ}}{a}_i{a}_{j}K(x_i,x_j)\≤R_T^2---\left(1\right)$

Wherein, R_zIt is the distance from characteristic vector z to cluster centre, R_TIt is based on described list The threshold value that the support vector of class grader sets, and K () is kernel function, for example, formulation For following histogram intersection core (Histogram Intersection Kernel, HIK):

$K(T,Q)=\underset{i=1}{\overset{d}{\mathrm{\Σ}}}\min (T_i,Q_i)---\left(2\right)$

Wherein, d is characteristic vector dimension.

Here, we are not intended to find the described cluster centre supporting vector in feature space Position, we only wish to determine from any characteristic vector z to support vector described cluster The distance function at center.

In this step, divide calculating from described center to the characteristic vector extracted and two classes All distances of the support vector of class device.

Then, the method will set threshold by using the support vector of described one-class classifier Value, i.e. above-mentioned R_T。

Hereinafter it is assumed that parameter P, i.e. point out outside the judgement hypersphere at one-class classifier The predetermined probabilities of the ratio of positive sample, by using R_minAnd R_max, produce judgement hyperspherical half Footpath R_T, i.e. above-mentioned threshold value.Parameter P is defined as foloows:

P=(f_T(z)-f_min)/(f_max-f_min) (3)。

According to formula (3), adjudicate hyperspherical radius R_TCan be determined that:

$\begin{array}{c}{f}_T\left(z\right)={\mathrm{Pf}}_{\max }+(1-P)f_{\min }\\ \⇒\mathrm{\π}{R}_T^2=\mathrm{P\π}{R}_{\max }^2+(1-P)\mathrm{\π}{P}_{\min }^2\\ \⇒R_T^2=P{R}_{\max }^2+(1-P)R_{\min }^2\end{array}---\left(4\right)$

Wherein R_minAnd R_maxIt is the least radius and maximum radius that can obtain based on formula (1).

According to formula (4), for characteristic vector z, estimate suitable threshold value, i.e. f_T(z).? Eventually, vector X is supported for all_i∈ SVs estimates f_T:

$f_T={\mathrm{\Σ}}_{i=1}^M{\mathrm{\α}}_i{f}_T\left(X_i\right)---\left(5\right)$

f_TIt is threshold value R_TSquare.

As for the f according to formula (1)_oneZ (), (z will be z) fixing for any characteristic vector z to K , because K () is HIK core and data z have been normalized (normalize).

$f_{\max }\left(z\right)=R_{\max }^2=K(z,z)+\underset{x_i\∈\mathrm{SVs}}{\mathrm{\Σ}}\underset{x_j\∈\mathrm{SVs}}{\mathrm{\Σ}}{\mathrm{\α}}_i{\mathrm{\α}}_{j}K(x_i,x_j)$

$f_{\min }\left(z\right)=R_{\min }^2=f_{\max }-2\max \left(\underset{x_i\∈\mathrm{SVs}}{\mathrm{\Σ}}{\mathrm{\α}}_{i}K(x_i,z)\right)---\left(6\right)$

In conjunction with formula (4)-(6), can be by setting fixation probability P_oneCalculate threshold value R_T。

Then, for the characteristic vector extracted and described binary classifier described support to Each in amount, compares calculating distance with the threshold value of setting, and works asWhen, Characteristic vector z is chosen as candidate feature vector, and described candidate feature vector is considered as to be positioned at it Radius is R_TJudgement hypersphere in positive sample, as shown in Figure 3.

It follows that by use described candidate feature vector update the support of one-class classifier to Amount.Here any update method for having the one-class classifier supporting vector can be used. Following methods is it is preferable that the sequence being realized by LIBSVM is minimum to be optimized by using (Sequential Minimal Optimization, SMO) method, based on previous support One-class classifier is learnt by candidate feature vector after vector sum updates again.SMO method Bigger quadratic programming (quadratic programming, QP) optimization problem is resolved into The possible QP problem of a series of minimum.Analytically solve these little QP problems, this Avoid and use time-consuming numerical value QP to optimize as interior circulation.LIBSVM is for SVM Storehouse, it has since two thousand been developed.It is that currently used widest SVM is soft Part.

It follows that update in step 230 at binary classifier, based on the characteristic vector extracted And at least one is single described in being partly or wholly updated in described step 220 The support vector of class grader, updates at least at least one binary classifier described Individual.

In one implementation, described binary classifier update step 230 may include that right In one of at least one binary classifier described binary classifier, determine extracted feature Vector is positive characteristic vector or negative characteristic vector；And by using all one-class classifiers Support that vector as positive characteristic vector and uses determination result to update described binary classifier Support vector.It will be appreciated by those skilled in the art that and the invention is not restricted to this implementation, But can make to update binary classifier in any suitable manner.

Specifically, the support vector of one-class classifier is obtained as positive characteristic vector.And institute The characteristic vector extracted can be by a described classifiers or only one-class classifier or only Only binary classifier is defined as positive characteristic vector or negative characteristic vector.

Then, binary classifier is updated by using all positive characteristic vectors and negative characteristic vector Support vector.

Based on positive characteristic vector and negative characteristic vector, can be handed over based on histogram by using The SVM of fork core trains binary classifier.For undetermined sample z, the judgement of SVM Function is shown as:

$f_{\mathrm{binary}}\left(z\right)=\underset{x_i\∈\mathrm{SVs}}{\mathrm{\Σ}}{\mathrm{\α}}_{i}K(x_i,z)+b>0---\left(7\right)$

Wherein x_iIt is to support vector, α_iBeing corresponding weight factor, K () is kernel function, and b It is judgement biasing (bias).Work as f_binaryZ, during () ＞ 0, sample z is classified as positive.

Additionally, the decision function of the SVM as shown in formula (7) can be normalized as generally Rate exports SVM.Probability output for two classes SVM is formulated as:

$f_{\mathrm{binary}}=1.0-\frac{1}{1+\exp (\mathrm{Af}\left(z\right)+B)}---\left(8\right)$

Wherein use 5 folding cross validations (5-fold cross-validation) in the training process Estimate A and B.

Similar with one-class classifier update method, the 5 folding cross validation phases in the training process Between, sequence minimum optimization (SMO) method and LIBSVM can be used renewal two classes and divide Class device.Those skilled in the art would appreciate that, any of or new suitable method Can be used renewal binary classifier.

Note that in some instances, comprise described the one of one-class classifier and binary classifier Classifiers can be cascade structure (cascade structure), and it passes through the two type Grader classify by proper order.But a described classifiers can also have weighting Accumulation structure, the fraction of its weighted accumulation by all graders is classified.Described In the case that one classifiers has weighted accumulation structure, according to the on-line study of first embodiment Method can also include updating single class after described binary classifier updates step 230 alternatively The weight of the weight of grader and binary classifier updates step.

In weighted accumulation structure, two kinds of grader has respective normalized probability Output, and final classification can be calculated by being weighted summation as follows to all outputs Fraction.

f_win=w1_one*f1_one+w2_one*f2_one+…+wm_one*fm_one+w1_binary*f1_binary+w2_binary*f2_binary

+…+wn_binary*fn_binary (9)

Wherein w1_one,…,wm_oneAnd w1_binary,…,wn_binaryIt is for one-class classifier and binary classifier Respective weight, it can be assessed based on Consumer's Experience or nonparametric or training comes Arrive.

Many methods can be used for updating weight w1_one,…,wm_oneAnd w1_binary,…,wn_binary.Below It is preferably effective method:

1) collection includes extracted characteristic vector and described one-class classifier and the classification of two classes The all positive characteristic vector of the support vector of device and negative characteristic vector conduct

F={X_pos,X_neg, and X_pos={ (x_i,f_i) | i=1,2 ... N_p, X_neg={ (x_i,f_i) | i=1,2 ... N_n}

Wherein X_posIt is positive characteristic vector set, and X_negIt is negative characteristic vector set.

2) weight w1 is regulated_one,…,wm_oneAnd w1_binary,…,wn_binary, so that it is guaranteed that positive characteristic vector and Difference between the average mark of negative characteristic vector is maximum；

$\left\{\begin{array}{c}{w1}_{\mathrm{one}}+...+{\mathrm{wm}}_{\mathrm{one}}+{w1}_{\mathrm{binary}}+...{\mathrm{wn}}_{\mathrm{binary}}=1\\ \mathrm{diffrence}=f_{\mathrm{positive}}^{\mathrm{average}}-f_{\mathrm{negative}}^{\mathrm{average}}\end{array}\right..$

Fig. 4 show according to a first embodiment of the present invention for a group objects grader The block diagram of on-line study equipment 400, it is single that a wherein said group objects grader comprises at least one Class grader and at least one binary classifier.

As shown in Figure 4, the on-line study equipment 400 for a group objects grader includes: special Levy vector extraction unit the 410th, one-class classifier updating block 420 and binary classifier updates single Unit 430.

More specifically, characteristic vector pickup unit 410 is configured to carry from newly added sample Take characteristic vector.

One-class classifier updating block 420 be configured to based on the characteristic vector extracted and The support vector of at least one binary classifier described updates at least one single class classification described At least one in device.

Binary classifier updating block 430 be configured to based on the characteristic vector extracted and Described in described one-class classifier updating block 420 has partly or wholly updated at least The support vector of one one-class classifier, updates at least one binary classifier described extremely Few one.

Unit in equipment 400 can be configured to perform shown in the flow chart in Fig. 2 Each step.

At the above-mentioned group objects grader for comprising one-class classifier and binary classifier On-line study method and apparatus in, between one-class classifier and binary classifier share support Vector, and therefore, it is possible to high-caliber simultaneously by false alarm rate reduction being maintained at recall rate To extremely low level.

(the second embodiment)

The present embodiment relates to being applied to the on-line study method and apparatus according to first embodiment Detect user in image sequence, register object.

Fig. 5 be illustrate according to a second embodiment of the present invention by use have determined as object At least one positive sample in image sequence, detect the flow chart of method of object.

As it is shown in figure 5, produce in step 510 at one-class classifier, according to described at least one Positive sample, produces at least one based on the support vector for classifying described object single Class grader.Those skilled in the art would appreciate that, any of or new is suitable Method can be used generation one-class classifier.

It follows that produce in step 520 at binary classifier, based on the of described image sequence One image and described one-class classifier produce be used for classifying described object at least one Individual binary classifier.

In one implementation, described binary classifier produce step 520 may include that by Described first image is divided into multiple region and determines institute by using described one-class classifier Whether each in the region dividing is described object；Collect the support of described one-class classifier Vector and the characteristic vector extracted from the region having been determined as described object, as just Characteristic vector；Collect the characteristic vector extracted from the region having been determined as non-object, make For negative characteristic vector；And by using collected positive characteristic vector and the training of negative characteristic vector At least one binary classifier.

An example of such binary classifier production method is described more fully below.This area Artisans will appreciate that, the invention is not restricted to this specific example, but can use any Suitable mode produces binary classifier.

In this example, first, by sliding window, the first image is divided into multiple window image, Then for each window, characteristic vector is extracted.Then, single class object grader is based on being extracted Characteristic vector determine that the image in sliding window is positive or negative, i.e. determine this image It whether is object to be detected.

Based on classification results, it is fenestrate that collection is classified as positive institute, uses non-maximum suppression Positive window is carried out clustering and merging with facility by (non-maximum suppression) With bounding box (bounding box) w_objPositioning object.

Then, all described bounding box w_objCharacteristic vector be extracted, and be considered positive spy Levy vector.

Additionally, all support vectors of one-class classifier are also collected as positive characteristic vector.

After positioning subject area in the picture, the registration (overlap with object bounding box Ratio) it is considered negative less than the window of specific threshold value thr of user, it may be assumed that

IfThen window w_sIt is considered negative.

All of negative window is collected, and arranges by ascending order based on their classification fraction, and And only head N_nIndividual window selected difficult (hard) negative sample, wherein N_n=N_p, and N_pIt is general One-class classifier just support vector quantity.Extract characteristic vector simultaneously from those negative samples And it is added to negative characteristic vector set X_negIn.

Then, training is based on the binary classifier supporting vector, and it can be by object and current field Scenic spot separates.Sequence minimum optimization (SMO) method and LIBSVM also can be used for this Plant training.

Based on positive characteristic vector and negative characteristic vector (entitled X_posAnd X_neg) both, by adopting Training binary classifier with the SVM based on histogram intersection core, object is divided by it with background From.For undetermined sample z, the decision function of SVM is shown as:

$f_{\mathrm{binary}}\left(z\right)=\underset{x_i\∈\mathrm{SVs}}{\mathrm{\Σ}}{\mathrm{\α}}_{i}K(x_i,z)+b>0---\left(10\right)$

Wherein x_iIt is to support vector, α_iBeing corresponding weight factor, K () is kernel function, and b It is judgement biasing (bias).Work as f_binaryZ, during () ＞ 0, sample z is classified as positive.

Additionally, the decision function of the SVM as shown in formula (10) can be normalized as generally Rate exports SVM.Probability output for two classes SVM is formulated as:

$f_{\mathrm{binary}}=1.0-\frac{1}{1+\exp (\mathrm{Af}\left(z\right)+B)}---\left(11\right)$

Wherein use 5 folding cross validations to estimate A and B in the training process.

Then, in object detection step 530, by use include described one-class classifier and One group objects grader of described binary classifier in described sequence follow the first image it After the second image in detect described object.After this detection, the method performs grader more New step 540, by using the on-line study side according to first embodiment in this step 540 Method, updates a described group objects grader according to described testing result.Then, the method is returned Return to object detection step 530, for by using the object classifier after updating in this sequence Next image in row detects described object.Object detection step 530 and grader update Step 540 is repeated, until all images in this sequence have all been detected or have detected Terminate.Grader updates step 540 and uses the on-line study method according to first embodiment, its It has been discussed in detail and here will not repeated description before.Therefore, we below will be in detail Object detection step 530 is discussed.

In a kind of implementation of method for checking object, detect described object may include that by Described image is divided into multiple region；By using a described group objects grader to being divided Region in each carry out classifying and obtain for being divided from a described classifiers Region in the classification fraction of each；And according to described classification fraction and predefined threshold Value determines that each in divided region is described object or non-object.

Alternately, in another kind of implementation, detect described object and may include that institute State image and be divided into multiple region；By using described one-class classifier to the region being divided In each carry out classifying and obtain for the region being divided from described one-class classifier In each first classification fraction；Come to by described list by using binary classifier Class grader is categorized as positive region and classifies, and obtains for these regions from binary classifier Second classification fraction, and by first classification fraction and second classification fraction be weighted Summation calculates final classification fraction；And collect its fraction of finally classifying more than predefined threshold The region of value as subject area, and collect its finally classification fraction be not more than described predefined The region of threshold value as non-object region.

An example of such detection method is described more fully below.Those skilled in the art's meeting Understand, the invention is not restricted to this specific example, but can make in any suitable manner Detect object.

In this example, including this classifiers of one-class classifier and binary classifier has level Connection structure.The first order of this structure is one-class classifier and the second level is binary classifier.

Detect object in the picture by this classifiers and collect positive characteristic vector and negative spy Levy vector for ensuing renewal (on-line study).

Concrete detection process is as follows.

Divide an image into multiple sample window image by sliding window, then for each window w_s, extract characteristic vector x_s.Then, sample image is classified by this classifiers as follows:

1) by using one-class classifier to sampling feature vectors x_sCarry out classifying and being divided Class fraction

2) if by one-class classifier by this sampling feature vectors x_sIt is categorized as positive, then pass through Binary classifier is used to carry out to it classifying and obtain fraction of classifying

3) when being subject area by above-mentioned two grader by sample classification, by using FormulaFinally classified fraction f_s。

During detecting, collect positive window and negative window: the W of candidate_pos={ (w_i,f_i) | i=1 ..., N_p, W_neg={ (w_i,f_i) | i=1 ..., N_n, each of which associates classification fraction, wherein a W_posIt just is Window set, and W_negIt is negative window set.Positive candidate W_posIt is by this component class Device is categorized as those positive samples, i.e. its fraction of finally classifying is more than that of predefined threshold value A little samples, and negative candidate W_negBe by one-class classifier be categorized as positive but by two Class grader is categorized as those samples born, i.e. its fraction of finally classifying is no more than predefined Those samples of threshold value.

Fig. 6 be illustrate according to a second embodiment of the present invention for by use have determined as The positive sample of at least one of object detects the block diagram of the equipment 600 of object in image sequence.

As shown in Figure 6, this equipment 600 includes: one-class classifier generation unit the 610th, two class Grader generation unit the 620th, subject detecting unit 630 and according to first embodiment Line facility for study 400.

More specifically, one-class classifier generation unit 610 be configured to according to described at least one Positive sample, produces at least one single class based on the support vector for classifying described object Grader.

Binary classifier generation unit 620 is configured to the first figure based on described image sequence As and described one-class classifier produce for described object is classified at least one two Class grader.

Subject detecting unit 630 is configured to use and includes described one-class classifier and institute State a group objects grader of binary classifier each follow-up image in described sequence The described object of middle detection.

On-line study equipment 400 according to first embodiment is configured to according to by described object The testing result that detector unit obtains updates a described group objects grader.

Unit in equipment 600 can be configured to perform shown in the flow chart in Fig. 5 Each step.

In above-mentioned detection method and equipment, between one-class classifier and binary classifier altogether Enjoy support vector, and therefore, it is possible to high-caliber simultaneously by false alarm being maintained at recall rate Rate is reduced to extremely low level.

(other embodiments)

The above-mentioned on-line study method according to first embodiment or according to the second embodiment Method for detection object in image sequence can be used in many application.One of application is Register object for automatically focusing on user in image pick up equipment.

As it is shown in fig. 7, image pick up equipment 700 may include that optical system 710, joined It is set to captured image or video；And according to the on-line study equipment 400 of first embodiment or Person is according to the equipment 600 for detecting object in the sequence of frame of video of the second embodiment.

Additionally, image pick up equipment 700 can be included in image processing equipment, this figure As processing equipment e.g. camera, mobile phone, desktop computer, tablet PC or Laptop computer etc..

(by the assessment of the technique effect that the present invention realizes)

Show the principle according to the present invention for detect the method for object with in prior art The example that compares of other method testing.Noting, shown here example is merely exemplary , it is used for illustrating the Advantageous techniques effect of the present invention, rather than restricted.

It is used in user according to the method for checking object of the present invention and register object detection (UROD) In system.By using single class for special object for little sample training of user's offer Grader；And the first frame of video sequence is used to construct binary classifier.The two is classified Device is combined to form a classifiers, and based on testing result by online updating.

We by the performance of the detection method of the present invention with simply use the existing of one-class classifier The method of technology compares.It is color and local three value moulds for the feature that image-region describes Formula (Local ternary pattern, LTP).

Assessment level is PASCAL assessment level, whenWhen, detection Region is considered correct detection region.Here, T is 0.5.

Recall rate (recall rate), i.e. recall rate, be

Empty inspection rate (FPPI) isWherein the frame of mistake detection includes Not there is object but be detected as the frame of object.

Table 1 illustrates the hardware and software configuration in test.

Table 1

Assessment example 1

Fig. 8 illustrate by the detection method of a pair grader of use according to embodiments of the present invention with Only use one-class classifier prior art detection method compare obtained from comparative result, That is, ROC curve.The data set using is all to be regarded by 20 that handheld camera shoots Frequently.To be detected to as if many plant animal, such as cat, dog etc..Each animal includes not Same outward appearance and background change at leisure.

As shown in Figure 8, compared with the method only with one-class classifier, according to the present embodiment Method has better performance (higher recall rate and lower FPPI).

Assessment example 2

In this example, equally, the inspection by a pair grader of use according to embodiments of the present invention Survey method compares with only using the prior art detection method of one-class classifier.

The data set using is 144 videos all being shot by handheld camera.To be detected To as if 28 kinds of animals.

The comparison of FPPI is performed in the case that recall rate is maintained at 92.1%.

FPPI is illustrated in table 2:

Table 2

As shown in the above Table, when the recall rate of both approaches is identical, according to the side of the present invention The FPPI of method is only about the 30% of the FPPI of one-class classifier method.

From comparative result above it can be seen that can subtract according to the method for checking object of the present invention Few FPPI leads (false alarm).It is to say, the combination of one-class classifier and binary classifier More effectively the scene of object and surrounding can be distinguished.Additionally, according to the object inspection of the present invention Survey method can improve recall rate.It is to say, will be made this according to the on-line study of the present invention Classifiers can adapt to the cosmetic variation in object.

The method and apparatus that can implement the present invention in many ways.For example, it is possible to pass through The method and apparatus that the present invention is implemented in software, hardware, firmware or its any combination.Above-mentioned The order of method step be merely illustrative, the method step of the present invention be not limited to above specifically The order describing, clearly states unless otherwise.Additionally, in certain embodiments, this Invention can also be implemented as recording program in the recording medium, and it includes for realizing basis The machine readable instructions of the method for the present invention.Thus, the present invention also covers storage and is used for realizing root The record medium of program according to the method for the present invention.

Although passed through example detail display some specific embodiments of the present invention, but ability It is illustrative and do not limit the present invention that field technique personnel it should be appreciated that above-mentioned example is intended merely to Scope.It should be appreciated by those skilled in the art that above-described embodiment can be without departing from the present invention's It is modified in the case of scope and spirit.The scope of the present invention is by appended claim limit Fixed.

Claims (20)

1. the on-line study method for a group objects grader, wherein said one group pair As grader comprises at least one one-class classifier and at least one binary classifier, described method Including:

Characteristic vector pickup step, extracts characteristic vector from newly added sample；

One-class classifier update step, based on the characteristic vector extracted and described at least one The support vector of binary classifier updates at least at least one one-class classifier described Individual；

Binary classifier updates step, divides based on the characteristic vector extracted and in described single class Class device updates at least one the single class described being partly or wholly updated in step and divides The support vector of class device, updates at least one at least one binary classifier described.

2. on-line study method according to claim 1, a wherein said group objects divides Appliances has weighted accumulation structure, and described on-line study method is at described binary classifier more The step updating the weight of one-class classifier and binary classifier is also included after new step.

3. on-line study method according to claim 1, wherein said one-class classifier Update step to include:

For one of at least one one-class classifier described one-class classifier,

Calculate the cluster centre of all support vectors from this one-class classifier to being extracted Each in the described support vector of characteristic vector and at least one binary classifier described Each distance；

Set threshold value based on the support vector of described one-class classifier；

From the described support of the characteristic vector extracted and at least one binary classifier described to Amount selects the candidate feature vector with the calculating distance less than described threshold value；And

Update the support vector of described one-class classifier by using described candidate feature vector.

4. on-line study method according to claim 1, wherein said binary classifier Update step to include:

For one of at least one binary classifier described binary classifier,

Determine that extracted characteristic vector is positive characteristic vector or negative characteristic vector；And

By using the support vector of all one-class classifiers as positive characteristic vector and use Determine result to update the support vector of described binary classifier.

5. scheming at least one the positive sample being had been determined as object by use for one kind As detecting the method for described object in sequence, described method includes:

One-class classifier produces step, according at least one positive sample described, produces based on being used for At least one one-class classifier of the support vector that described object is classified；

Binary classifier produces step, based on the first image and the described list of described image sequence Class grader produces at least one binary classifier for classifying described object；

Object detection step, includes described one-class classifier and described binary classifier by using The second image in described sequence for the group objects grader in detect described object；

Grader updates step, by using on-line study side according to claim 1 Method, updates a described group objects grader according to described testing result；And

Follow-up detection and renewal step, for each the follow-up image in described sequence, It is described right to detect in the image that the group objects grader being updated over by use is continued in the rear As and update a described group objects grader subsequently.

6. method according to claim 5, wherein said binary classifier produces step Including:

Described first image is divided into multiple region and by using described one-class classifier Determine whether each in divided region is described object；

Collect the support vector of described one-class classifier and from having been determined as described object Region in extract characteristic vector, as positive characteristic vector；

Collect the characteristic vector extracted from the region having been determined as non-object, as negative spy Levy vector；And

Train at least one two class by using collected positive characteristic vector and negative characteristic vector Grader.

7. method according to claim 5, wherein detects described object and includes:

Divide the image into multiple region；

By using a described group objects grader, each in the region being divided is entered Row is classified and obtains for each region being divided from a described classifiers Classification fraction；And

According to described classification fraction and predefined threshold value determine in divided region each Individual is described object or non-object.

8. method according to claim 5, wherein detects described object and includes:

Divide the image into multiple region；

Each in the region being divided is carried out point by using described one-class classifier Class and obtaining for each first region being divided from described one-class classifier Classification fraction；

By using binary classifier to being categorized as positive district by described one-class classifier Classifying in territory, obtains the second classification fraction for these regions from binary classifier, and Calculate final classification by summation is weighted to the first classification fraction and the second classification fraction Fraction；And

Collect its finally classification fraction more than region of predefined threshold value as subject area, and And collect its finally classification fraction be not more than the region of described predefined threshold value as non-object Region.

9. the on-line study equipment for a group objects grader, wherein said one group pair As grader comprises at least one one-class classifier and at least one binary classifier, described equipment Including:

Characteristic vector pickup unit, is configured to from newly added sample extract characteristic vector；

One-class classifier updating block, is configured to based on the characteristic vector extracted and described The support vector of at least one binary classifier updates at least one one-class classifier described At least one；

Binary classifier updating block, is configured to based on the characteristic vector extracted and by institute State at least one list described that one-class classifier updating block has partly or wholly updated The support vector of class grader, updates at least at least one binary classifier described Individual.

10. on-line study equipment according to claim 9, a wherein said group objects divides Appliances has weighted accumulation structure, and described on-line study equipment also includes being configured to more The unit of the weight of renewal one-class classifier and binary classifier after new binary classifier.

11. on-line study equipment according to claim 9, wherein said one-class classifier Updating block includes:

It is configured to for one of at least one one-class classifier described one-class classifier, meter Calculate from this one-class classifier all support vector cluster centre to the feature extracted to Amount and at least one binary classifier described described support vector in each each away from From unit；

It is configured to for one of at least one one-class classifier described one-class classifier, base Set the unit of threshold value in the support vector of described one-class classifier；

It is configured to the institute from the characteristic vector extracted and at least one binary classifier described State the candidate feature vector supporting to select to have the calculating distance less than described threshold value in vector Unit；And

It is configured to use described candidate feature vector to update described one-class classifier Support the unit of vector.

12. on-line study equipment according to claim 9, wherein said binary classifier Updating block includes:

It is configured to for one of at least one binary classifier described binary classifier, really Fixed extracted characteristic vector is positive characteristic vector or the unit of negative characteristic vector；And

It is configured to for one of at least one binary classifier described binary classifier, logical Cross the support vector using all one-class classifiers and as positive characteristic vector and use determination knot Fruit updates the unit of the support vector of described binary classifier.

13. 1 kinds for by using at least one the positive sample having been determined as object at figure As detecting the equipment of described object in sequence, described equipment includes:

One-class classifier generation unit, is configured to, according at least one positive sample described, produce At least one one-class classifier based on the support vector for described object is classified；

Binary classifier generation unit, be configured to based on described image sequence the first image with And described one-class classifier produces and divides at least one two class classifying described object Class device；

Subject detecting unit, is configured to use and includes described one-class classifier and described two One group objects grader of class grader each follow-up image in described sequence is examined Survey described object；And

On-line study equipment according to claim 9, is configured to according to by described object The testing result that detector unit obtains updates a described group objects grader.

14. equipment according to claim 13, wherein said binary classifier produces single Unit includes:

It is configured to described first image is divided into multiple region and by using described list Class grader determines the unit whether each in divided region is described object；

It is configured to collect the support vector of described one-class classifier and from having been determined as The characteristic vector extracted in the region of described object, as the unit of positive characteristic vector；

It is configured to collect the characteristic vector extracted from the region having been determined as non-object, Unit as negative characteristic vector；And

It is configured to use collected positive characteristic vector and negative characteristic vector to train at least The unit of one binary classifier.

15. equipment according to claim 13, wherein said subject detecting unit includes:

It is configured to divide the image into the unit in multiple region；

It is configured to use a described group objects grader in the region being divided Each carries out classifying and obtains for the region being divided from a described classifiers The unit of the classification fraction of each；And

It is configured to determine divided region according to described classification fraction and predefined threshold value In each be the unit of described object or non-object.

16. equipment according to claim 13, wherein said subject detecting unit includes:

It is configured to divide the image into the unit in multiple region；

It is configured to use each in the region being divided of described one-class classifier Individual carry out classifying and from described one-class classifier obtain for the region being divided each The unit of the first individual classification fraction；

It is configured to use binary classifier to being classified by described one-class classifier Classify for positive region, obtain the second classification point for these regions from binary classifier Number, and by the first classification fraction and the second classification fraction are weighted suing for peace and calculate The unit of classification fraction eventually；And

Be configured to collect its finally classification fraction more than the region of predefined threshold value as right As region, and collect the region work that its fraction of finally classifying is not more than described predefined threshold value Unit for non-object region.

17. on-line study methods according to any one in claim 1-4 or according to Existing for the method detecting object in image sequence described in any one in claim 5-8 Register the use in the image pick up equipment of object for autofocusing on user.

18. 1 kinds of image pick up equipment, comprising:

Optical system, is configured to captured image or video；And

On-line study equipment according to any one in claim 9-12 or according to power Profit require described in any one in 13-16 for detection the setting of object in the sequence in frame of video Standby.

19. 1 kinds of image processing equipments, including image pickup according to claim 18 Equipment.

20. image processing equipments according to claim 19, wherein said image procossing Equipment is camera, mobile phone, desktop computer, tablet PC or calculating on knee Machine.