CN103593672A - Adaboost classifier on-line learning method and Adaboost classifier on-line learning system - Google Patents

Abstract

The invention relates to an Adaboost classifier on-line learning method and an Adaboost classifier on-line learning system. The Adaboost classifier on-line learning method comprises steps that: object detection is carried out by a strong classifier acquired through employing off-line training, and an object detection result is acquired; object detection is acquired by employing a background model to acquire a motion object; the object detection result acquired by the strong classifier is compared with the motion object acquired through detection by employing the background model to acquire an error classifier object; the error classifier object is taken as an on-line training sample to carry out on-line training to acquire a strong classifier after updating. According to the Adaboost classifier on-line learning method and the Adaboost classifier on-line learning system, the object detection result acquired by the off-line classifier is compared with the motion object acquired through detection by employing the background model to acquire the error classifier object, the error classifier object is taken as the on-line training sample to acquire the strong classifier after updating. Generalization performance of the object detection classifier is effectively improved, so the object detection classifier can adapt to a monitoring scene during operation, a detection ratio is improved, and a rate of false alarm is reduced.

Description

Adaboost sorter on-line study method and system

Technical field

The present invention relates to sorter field, particularly relate to a kind of Adaboost sorter on-line study method and system.

Background technology

The target detection techniques such as the detection of people's face, pedestrian detection, vehicle detection are one of core technologies of intelligent video monitoring.At present, target detection has the method for two kinds of main flows: based on motion detection with based on detection of classifier.Based on motion detection, be to cut apart the moving target (prospect) in appearance scape by technology such as background modelings, speed is fast, but very sensitive to illumination variation, inclement weather, chaff interference etc., when target is intensive as inter-adhesive, also cannot be partitioned into accurately each target, in addition, also cannot distinguish accurately the type of each target.Based on detection of classifier, be the method for using machine learning, the sorter of a specific objective of precondition (as face classification device), during operation, scans whole frame of video, detects wherein all targets.

In target detection sorter, what be widely used is Adaboost sorter, and the method accuracy rate is high, speed is fast.Adaboost sorter is generally that off-line training generates, and, by the training sample of collecting is trained, obtains sorter model, then for actual target detection task.Because off-line training sample and monitoring scene there are differences, may cause the Generalization Capability of Adaboost sorter poor, verification and measurement ratio is low, and rate of false alarm is high.For example, if the pedestrian's sample in training sample under normal illumination condition, take, so, when monitoring scene light is very weak, cannot be correct detect the pedestrian in scene.

Summary of the invention

Based on this, be necessary to adopt off-line training to generate for traditional Adaboost sorter, detecting goal task exists verification and measurement ratio low, the problem that rate of false alarm is high, a kind of Adaboost sorter on-line study method is provided, while adopting this Adaboost detection of classifier goal task, can improves verification and measurement ratio and rate of false alarm is low.

In addition, be also necessary to provide a kind of Adaboost sorter on-line study system, while adopting this Adaboost detection of classifier goal task, can improve verification and measurement ratio and rate of false alarm is low.

A sorter on-line study method, comprising:

The strong classifier that adopts off-line training to obtain carries out target detection, and obtains target detection result;

Adopt background model to carry out target detection, obtain moving target;

The target detection result that strong classifier is obtained and background model detection obtain moving target and compare, and obtain the target of misclassification device;

Using the target of described misclassification device as online training sample, train online the strong classifier after being upgraded.

A sorter on-line study system, comprising:

First detection module, for adopting the strong classifier that off-line training obtains to carry out target detection, and obtains target detection result;

The second detection module, for adopting background model to carry out target detection, obtains moving target;

Comparison module, detects and obtains moving target and compare for target detection result that strong classifier is obtained and background model, obtains the target of misclassification device;

Online training module, for using the target of described misclassification device as online training sample, trains the strong classifier after being upgraded online.

Above-mentioned Adaboost sorter on-line study method and system, by the moving target of the objective result of off-line detection of classifier and background model detection is compared, obtain the target of misclassification device, using the target of misclassification device as online training sample, carry out on-line study, strong classifier after being upgraded, by the strong classifier after upgrading, detect target again, effectively improved the Generalization Capability of target detection sorter, monitoring scene in the time of can adapting to operation, improve verification and measurement ratio, reduced rate of false alarm.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of Adaboost sorter on-line study method in an embodiment;

Fig. 2 adopts background model to carry out target detection in an embodiment, obtains the process flow diagram of movement destination step;

Fig. 3 adopts background model to carry out target detection in another embodiment, obtains the process flow diagram of movement destination step;

Fig. 4 adopts background model to carry out target detection in another embodiment, obtains the process flow diagram of movement destination step;

Fig. 5 is using the target of this misclassification device as online training sample, trains online the process flow diagram of the strong classifier step after being upgraded;

Fig. 6 is the process flow diagram of the strong classifier that in an embodiment, off-line training obtains;

Fig. 7 is according to this N off-line training sample and a plurality of Weak Classifiers of weight circuit training, and obtains the particular flow sheet of the weight step of each Weak Classifier;

Fig. 8 is the structured flowchart of Adaboost sorter on-line study system in an embodiment;

Fig. 9 is the inner structure block diagram of the second detection module in Fig. 8;

Figure 10 is the inner structure block diagram of online training module in Fig. 8;

Figure 11 is the structured flowchart of Adaboost sorter on-line study system in another embodiment;

Figure 12 is the inner structure block diagram of off-line training module in Figure 11;

Figure 13 is the inner structure block diagram of off-line training unit in Figure 12.

Embodiment

Below in conjunction with specific embodiment and accompanying drawing, the technical scheme of Adaboost sorter on-line study method and system is described in detail, so that it is clearer.

As shown in Figure 1, be the process flow diagram of Adaboost sorter on-line study method in an embodiment.This Adaboost sorter on-line study method, comprising:

Step S102, the strong classifier that adopts off-line training to obtain carries out target detection, and obtains target detection result.

Step S104, adopts background model to carry out target detection, obtains moving target.

Concrete, this background model is gauss hybrid models, background subtraction point-score model or frame average background model.

First, when this background model is gauss hybrid models, as shown in Figure 2, adopt background model to carry out target detection, the step that obtains moving target comprises:

Step S202, if present frame is the first frame, the gaussian component in initialization gauss hybrid models.

Concrete, the weight of each gaussian component in initialization gauss hybrid models, average and variance.

Step S204, if present frame is not or not first frame, utilizes each pixel of present frame to upgrade the gaussian component in gauss hybrid models, and according to the comparison of present frame and each gaussian component, obtains foreground pixel, forms foreground picture.

Concrete, utilize each pixel of present frame to upgrade weight, variance and the average of each gaussian component in gauss hybrid models, and can replace the gaussian component of weight minimum.

Step S206, processes this foreground picture, extracts UNICOM's component, obtains moving target.

Concrete, foreground picture is processed, comprise and carry out morphological dilations and corrosion.Morphological dilations, the process that for example image A is expanded by structural element B is: (1) uses structural element B, each pixel of scan image A; (2) with the bianry image of structural element B and its covering, do AND-operation; (3) if be all 0, this pixel of image is 0, otherwise is 1.

Morphological erosion, for example, structural element B to the process of image A corrosion is: (1) uses structural element B, each pixel of scan image A; (2) with the bianry image of structural element B and its covering, do AND-operation; (3) if be all 1, this pixel of image is 1, otherwise is 0.

UNICOM's component refers to the region that becomes piece in foreground picture, and these regions are not connected with other regions.

Secondly, when this background model is background subtraction point-score model, as shown in Figure 3, adopt background model to carry out target detection, obtain the step of moving target, comprising:

Step S302, adopts present frame and former frame to subtract each other and obtains difference diagram, carries out thresholding extract foreground picture according to difference diagram.

Concrete, thresholding can pass through to specify fixed threshold, as sets a certain fixed threshold, then filters out the pixel that is greater than this fixed threshold, extracts foreground picture.Thresholding also can pass through adaptive threshold, be that threshold value corresponding to each pixel may be different, as the threshold value of each pixel is determined by the field window centered by self, this is usingd to the intermediate value of the field window centered by certain pixel or average or the Gaussian convolution threshold value as this pixel.

Step S304, processes this foreground picture, extracts UNICOM's component, obtains moving target.

Concrete, foreground picture is processed, comprise and carry out morphological dilations and corrosion.

Again, when this background model is frame average background model, as shown in Figure 4, adopt background model to carry out target detection, obtain the step of moving target, comprising:

Step S402, by the weighted mean of all frames before present frame as a setting, then obtains difference diagram with present frame and this background subtracting, carries out thresholding extract foreground picture according to difference diagram.

Concrete, thresholding can pass through to specify fixed threshold, as sets a certain fixed threshold, then filters out the pixel that is greater than this fixed threshold, extracts foreground picture.Thresholding also can pass through adaptive threshold, be that threshold value corresponding to each pixel may be different, as the threshold value of each pixel is determined by the field window centered by self, this is usingd to the intermediate value of the field window centered by certain pixel or average or the Gaussian convolution threshold value as this pixel.

Step S404, processes this foreground picture, extracts UNICOM's component, obtains moving target.

Concrete, foreground picture is processed, comprise and carry out morphological dilations and corrosion.

Step S106, the target detection result that strong classifier is obtained and background model detection obtain moving target and compare, and obtain the target of misclassification device.

Concrete, the target of misclassification device comprises the target of wrong report and the target of failing to report.Strong classifier detects target at Background, thinks the target of wrong report; If do not detect target in foreground picture, think the target of failing to report.Target for wrong report can, using the detected decoy of strong classifier as negative training sample, for the target of failing to report, can obtain target, then using it as positive training sample from foreground picture.

In addition,, after relatively, if there is no misclassification device target, process next frame figure.

Step S108, using the target of this misclassification device as online training sample, trains online, the strong classifier after being upgraded.

In one embodiment, as shown in Figure 5, step S108 comprises:

Step S502, using the target of this misclassification device as online training sample, the weight of the online training sample of initialization.

Concrete, Input Online training sample (x, y), online Weak Classifier learning algorithm L ₀with strong classifier h (x).Weight λ=1 of the online training sample of initialization.

Step S504, arranges the number of times of training according to take Poisson distribution that weight is parameter.

Concrete, the Poisson distribution Poisson (λ) that the λ of take is parameter arranges the number of times k of training.

Step S506, trains according to this online training sample Weak Classifier according to the number of times arranging.

Concrete, by Weak Classifier h _maccording to training sample (x, y), train k time:

h _m←L _o(h _m，(x，y)) (1)

In formula (1), h _mrepresent m Weak Classifier, total M Weak Classifier, M is natural number.

Step S508, if online training sample is correctly classified, upgrades the parameter of Poisson distribution, calculates the number of times accumulated value of correct classification, the error rate of calculating Weak Classifier.

Concrete, computing formula is as follows:

${\mathrm{\λ}}_m^{\mathrm{sc}}\←{\mathrm{\λ}}_m^{\mathrm{sc}}+\mathrm{\λ},e_m\←\frac{{\mathrm{\λ}}_m^{\mathrm{sw}}}{{\mathrm{\λ}}_m^{\mathrm{sc}}+{\mathrm{\λ}}_m^{\mathrm{sw}}},\mathrm{\λ}\←\mathrm{\λ}\×\frac{1}{2(1-e_m)}---\left(2\right)$

In formula (2),

for current online sample is by the number of times accumulated value of the correct classification of Weak Classifier,

for current online sample is by the accumulative frequency of Weak Classifier mis-classification, e _mfor the error rate of Weak Classifier, the parameter that λ is Poisson distribution.

with

initial value be all 1.

Step S510, if online training sample is by mis-classification, upgrades the parameter of Poisson distribution, the error rate of the number of times accumulated value of miscount classification, calculating sorter.

Concrete, computing formula is as follows:

${\mathrm{\λ}}_m^{\mathrm{sw}}\←{\mathrm{\λ}}_m^{\mathrm{sw}}+\mathrm{\λ},e_m\←\frac{{\mathrm{\λ}}_m^{\mathrm{sw}}}{{\mathrm{\λ}}_m^{\mathrm{sc}}+{\mathrm{\λ}}_m^{\mathrm{sw}}},\mathrm{\λ}\←\mathrm{\λ}\×\frac{1}{2e}---\left(3\right)$

In formula (3), for current online sample is by the number of times accumulated value of the correct classification of Weak Classifier,

for current online sample is by the accumulative frequency of Weak Classifier mis-classification, e _mfor the error rate of Weak Classifier, the parameter that λ is Poisson distribution, e is natural logarithm.

with

initial value be all 1.

Step S512, the strong classifier after being upgraded.

In formula (4), M is the number of Weak Classifier in strong classifier,

for current sample is by the number of times accumulated value of the correct classification of Weak Classifier,

for current sample is by the accumulative frequency of Weak Classifier mis-classification, e _merror rate for Weak Classifier.

with initial value be all 1.

In formula (4)

represent to calculate the output to current online sample to be sorted with each Weak Classifier, then that the output that is judged as each Weak Classifier is cumulative, find that maximum class of cumulative system as the classification of sample to be sorted.

Then, adopt this strong classifier to carry out target detection, process next frame figure.

In addition, the Weak Classifier in foregoing description is Adaboost Weak Classifier, and strong classifier is Adaboost strong classifier.

Above-mentioned Adaboost sorter on-line study method, by the moving target of the objective result of off-line detection of classifier and background model detection is compared, obtain the target of misclassification device, using the target of misclassification device as online training sample, carry out on-line study, strong classifier after being upgraded, by the strong classifier after upgrading, detect target again, effectively improved the Generalization Capability of target detection sorter, monitoring scene in the time of can adapting to operation, improve verification and measurement ratio, reduced rate of false alarm.

Further, in one embodiment, as shown in Figure 6, above-mentioned Adaboost sorter on-line study method, also comprises the step of the strong classifier that off-line training obtains.The strong classifier that this off-line training obtains, comprising:

Step S602, given N off-line training sample, and the weight of initialization off-line training sample.

Concrete, given N off-line training sample { (x _i, y _i), i=1 ..., N, x _i∈ R ^k, y _i∈ 1, and+1}, each off-line training sample comprises proper vector and class label.Proper vector x is a n-tuple, represents the feature of sample; Class label y is the classification under sample, and+1 is positive sample, and-1 is negative sample.Initialization off-line training sample weights w _i=1/N, i=1 ..., N, all samples have equal weight.

Step S604, according to this N off-line training sample and a plurality of Weak Classifiers of weight circuit training, and obtains the weight of each Weak Classifier.

Further, in one embodiment, as shown in Figure 7, according to this N off-line training sample and a plurality of Weak Classifiers of weight circuit training, and the step that obtains the weight of each Weak Classifier comprises:

Step S702, according to this N off-line training sample and a plurality of Weak Classifiers of weight training.

Concrete, use weight w _iwith a Weak Classifier f of training sample set training _m(x) ∈ 1 ,+1}.

Step S704, calculates the total error of described N off-line training sample on each Weak Classifier, and according to the total error of described each Weak Classifier, obtains the weight of each Weak Classifier.

Concrete, the total error of calculation training sample on this Weak Classifier

${\mathrm{err}}_m=E_w\left[1_{y\≠f_m\left(x\right)}\right],c_m=\log \left(\frac{1-{\mathrm{err}}_m}{{\mathrm{err}}_m}\right)$

Wherein, err _mtotal error, c _mit is the weight of m Weak Classifier.1 _{y ≠ fm (x)}represent that training sample x is correctly classified, i.e. y ≠ f (x), this value is 1, otherwise is 0, E _wrepresent weighting mathematical expectation.

Step S706, upgrades the weight of this off-line training sample, and to this weight normalization.

Concrete, the weight of renewal off-line training sample

to weight normalization refer to weight after the renewal of calculating all off-line training samples and, and then by the weight of each off-line training sample divided by this weight and.

Step S606, obtains strong classifier according to the plurality of Weak Classifier and corresponding weight.

Concrete, obtain strong classifier $h\left(x\right)=\mathrm{sgn}\left(\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{c}_m{f}_m\left(x\right)\right).$

As shown in Figure 8, be the structured flowchart of Adaboost sorter on-line study system in an embodiment.This Adaboost sorter on-line study system, comprises first detection module 820, the second detection module 840, comparison module 860 and online training module 880.Wherein:

First detection module 820 carries out target detection for the strong classifier that adopts off-line training and obtain, and obtains target detection result.

The second detection module 840, for adopting background model to carry out target detection, obtains moving target.

Concrete, this background model is gauss hybrid models, background subtraction point-score model or frame average background model.

First, when this background model is gauss hybrid models, in one embodiment, as shown in Figure 9, the second detection module 840 comprises component initialization unit 842, extraction unit 844 and acquiring unit 846.Wherein:

If component initialization unit 842 is the first frame for present frame, the gaussian component in initialization gauss hybrid models.

Concrete, the weight of each gaussian component in initialization gauss hybrid models, average and variance.

If extraction unit 844 is not the first frame for present frame, utilize each pixel of present frame to upgrade the gaussian component in gauss hybrid models, and according to the comparison of present frame and each gaussian component, obtain foreground pixel, form foreground picture.

Concrete, utilize each pixel of present frame to upgrade weight, variance and the average of each gaussian component in gauss hybrid models, and can replace the gaussian component of weight minimum.

Acquiring unit 846, for this foreground picture is processed, extracts UNICOM's component, obtains moving target.

Concrete, foreground picture is processed, comprise and carry out morphological dilations and corrosion.UNICOM's component refers to the region that becomes piece in foreground picture, and these regions are not connected with other regions.

Secondly, when this background model is background subtraction point-score model, the second detection module 840 comprises extraction unit 844 and acquiring unit 846.Wherein:

Extraction unit 844 obtains difference diagram for adopting present frame and former frame to subtract each other, and carries out thresholding extract foreground picture according to difference diagram.

Concrete, thresholding can pass through to specify fixed threshold, as sets a certain fixed threshold, then filters out the pixel that is greater than this fixed threshold, extracts foreground picture.Thresholding also can pass through adaptive threshold, be that threshold value corresponding to each pixel may be different, as the threshold value of each pixel is determined by the field window centered by self, this is usingd to the intermediate value of the field window centered by certain pixel or average or the Gaussian convolution threshold value as this pixel.

Acquiring unit 846, for this foreground picture is processed, extracts UNICOM's component, obtains moving target.

Concrete, foreground picture is processed, comprise and carry out morphological dilations and corrosion.

When this background model is frame average background model, the second detection module 840 comprises extraction unit 844 and acquiring unit 846.Wherein:

Extraction unit 844 is for by the weighted mean of all frames before present frame as a setting, then obtains difference diagram with present frame and this background subtracting, according to difference diagram, carries out thresholding extraction foreground picture.

Acquiring unit 846, for this foreground picture is processed, extracts UNICOM's component, obtains moving target.

Comparison module 860 detects and obtains moving target and compare for target detection result that strong classifier is obtained and background model, obtains the target of misclassification device.

Concrete, the target of misclassification device comprises the target of wrong report and the target of failing to report.Strong classifier detects target at Background, thinks the target of wrong report; If do not detect target in foreground picture, think the target of failing to report.Target for wrong report can, using the detected decoy of strong classifier as negative training sample, for the target of failing to report, can obtain target, then using it as positive training sample from foreground picture.

In addition,, after relatively, if there is no misclassification device target, process next frame figure.

Online training module 880, for using the target of this misclassification device as online training sample, is trained the strong classifier after being upgraded online.

In one embodiment, as shown in figure 10, online training module 880 comprises online initialization unit 882, setting unit 884, online training unit 886 and updating block 888.Wherein:

Online initialization unit 882 is for using the target of described misclassification device as online training sample, the weight of initialization training sample.

Concrete, Input Online training sample (x, y), online Weak Classifier learning algorithm L ₀with strong classifier h (x).Weight λ=1 of the online training sample of initialization.

Setting unit 884 is for arranging the number of times of training according to take Poisson distribution that weight is parameter.

Concrete, the Poisson distribution Poisson (λ) that the λ of take is parameter arranges the number of times k of training.

Online training unit 886 is for training according to described online training sample Weak Classifier according to the number of times of this setting.

Concrete, by Weak Classifier h _maccording to training sample (x, y), train k time:

h _m←L _o(h _m,(x,y)) （1）

In formula (1), h _mrepresent m Weak Classifier, total M Weak Classifier, M is natural number.

Updating block 888, for when online training sample is correctly classified, upgrades the parameter of Poisson distribution, calculates the number of times accumulated value of correct classification, the error rate of calculating sorter; And when online training sample is during by mis-classification, upgrade the parameter of Poisson distribution, the number of times accumulated value of miscount classification, calculate the error rate of sorter the strong classifier after being upgraded.

Concrete, computing formula is as follows:

${\mathrm{\λ}}_m^{\mathrm{sc}}\←{\mathrm{\λ}}_m^{\mathrm{sc}}+\mathrm{\λ},e_m\←\frac{{\mathrm{\λ}}_m^{\mathrm{sw}}}{{\mathrm{\λ}}_m^{\mathrm{sc}}+{\mathrm{\λ}}_m^{\mathrm{sw}}},\mathrm{\λ}\←\mathrm{\λ}\×\frac{1}{2(1-e_m)}---\left(2\right)$

In formula (2),

for current online sample is by the number of times accumulated value of the correct classification of Weak Classifier,

for current online sample is by the accumulative frequency of Weak Classifier mis-classification, e _mfor the error rate of Weak Classifier, the parameter that λ is Poisson distribution.

with

initial value be all 1.

Computing formula is as follows:

${\mathrm{\λ}}_m^{\mathrm{sw}}\←{\mathrm{\λ}}_m^{\mathrm{sw}}+\mathrm{\λ},e_m\←\frac{{\mathrm{\λ}}_m^{\mathrm{sw}}}{{\mathrm{\λ}}_m^{\mathrm{sc}}+{\mathrm{\λ}}_m^{\mathrm{sw}}},\mathrm{\λ}\←\mathrm{\λ}\×\frac{1}{2e}---\left(3\right)$

In formula (3),

for current online sample is by the number of times accumulated value of the correct classification of Weak Classifier,

for current online sample is by the accumulative frequency of Weak Classifier mis-classification, e _mfor the error rate of Weak Classifier, the parameter that λ is Poisson distribution, e is natural logarithm. with

initial value be all 1.

In formula (4), M is the number of Weak Classifier in strong classifier,

for current sample is by the number of times accumulated value of the correct classification of Weak Classifier,

for current sample is by the accumulative frequency of Weak Classifier mis-classification, e _merror rate for Weak Classifier. with

initial value be all 1.

In formula (4)

represent to calculate the output to current online sample to be sorted with each Weak Classifier, then that the output that is judged as each Weak Classifier is cumulative, find that maximum class of cumulative system as the classification of sample to be sorted.

Then, adopt this strong classifier to carry out target detection, process next frame figure.

In addition, the Weak Classifier in foregoing description is Adaboost Weak Classifier, and strong classifier is Adaboost strong classifier.

Above-mentioned Adaboost sorter on-line study system, by the moving target of the objective result of off-line detection of classifier and background model detection is compared, obtain the target of misclassification device, using the target of misclassification device as online training sample, carry out on-line study, strong classifier after being upgraded, by the strong classifier after upgrading, detect target again, effectively improved the Generalization Capability of target detection sorter, monitoring scene in the time of can adapting to operation, improve verification and measurement ratio, reduced rate of false alarm.

As shown in figure 11, be the structured flowchart of Adaboost sorter on-line study system in another embodiment.This Adaboost sorter on-line study system, except comprising first detection module 820, the second detection module 840, comparison module 860 and online training module 880, also comprises off-line training module 890.Wherein:

The strong classifier that off-line training module 890 obtains for off-line training.As shown in figure 12, off-line training module 890 comprises off-line initialization unit 892, off-line initialization unit 894 and off-line output unit 896.Wherein:

Off-line initialization unit 892 is for given N off-line training sample, and the weight of initialization off-line training sample.

Concrete, given N off-line training sample { (x _i, y _i), i=1 ..., N, x _i∈ R ^k, y _i∈ 1, and+1}, each off-line training sample comprises proper vector and class label.Proper vector x is a n-tuple, represents the feature of sample; Class label y is the classification under sample, and+1 is positive sample, and-1 is negative sample.Initialization off-line training sample weights w _i=1/N, i=1 ..., N, all samples have equal weight.

Off-line training unit 894 is for according to this N off-line training sample and a plurality of Weak Classifiers of weight circuit training, and obtains the weight of each Weak Classifier.

As shown in figure 13, off-line training unit 894 comprises Weak Classifier training subelement 894a, Weak Classifier weight calculation subelement 894b and training sample weight renewal subelement 894c.Wherein:

Weak Classifier training subelement 894a is used for according to described N off-line training sample and a plurality of Weak Classifiers of weight training.

Concrete, use weight w _iwith a Weak Classifier f of training sample set training _m(x) ∈ 1 ,+1}.

Weak Classifier weight calculation subelement 894b is used for calculating the total error of described N off-line training sample on each Weak Classifier, and according to the total error of described each Weak Classifier, obtains the weight of each Weak Classifier.

Concrete, the total error of calculation training sample on this Weak Classifier:

${\mathrm{err}}_m=E_w\left[1_{y\≠f_m\left(x\right)}\right],c_m=\log \left(\frac{1-{\mathrm{err}}_m}{{\mathrm{err}}_m}\right)$

Wherein, err _mtotal error, c _mit is the weight of m Weak Classifier.1 _{y ≠ fm (x)}represent that training sample x is correctly classified, i.e. y ≠ f (x), this value is 1, otherwise is 0, E _wrepresent weighting mathematical expectation.

Training sample weight is upgraded subelement 894c for upgrading the weight of this off-line training sample, and to this weight normalization.

Concrete, the weight of renewal off-line training sample

to weight normalization refer to weight after the renewal of calculating all off-line training samples and, and then by the weight of each off-line training sample divided by this weight and.

Off-line output unit 896 is for obtaining strong classifier according to the plurality of Weak Classifier and corresponding weight.

Concrete, obtain strong classifier $h\left(x\right)=\mathrm{sgn}\left(\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{c}_m{f}_m\left(x\right)\right).$

One of ordinary skill in the art will appreciate that all or part of flow process realizing in above-described embodiment method, to come the hardware that instruction is relevant to complete by computer program, described program can be stored in a computer read/write memory medium, this program, when carrying out, can comprise as the flow process of the embodiment of above-mentioned each side method.Wherein, described storage medium can be magnetic disc, CD, read-only store-memory body (Read-Only Memory, ROM) or random store-memory body (Random Access Memory, RAM) etc.

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

Claims (10)

1. an Adaboost sorter on-line study method, comprising:

The strong classifier that adopts off-line training to obtain carries out target detection, and obtains target detection result;

Adopt background model to carry out target detection, obtain moving target;

The target detection result that strong classifier is obtained and background model detection obtain moving target and compare, and obtain the target of misclassification device;

Using the target of described misclassification device as online training sample, train online the strong classifier after being upgraded.

2. Adaboost sorter on-line study method according to claim 1, is characterized in that, described method also comprises:

The strong classifier that off-line training obtains, comprising:

Given N off-line training sample, and the weight of initialization off-line training sample;

According to described N off-line training sample and a plurality of Weak Classifiers of weight circuit training, and obtain the weight of each Weak Classifier;

According to described a plurality of Weak Classifiers and corresponding weight, obtain strong classifier.

3. Adaboost sorter on-line study method according to claim 1, is characterized in that, described according to described N off-line training sample and a plurality of Weak Classifiers of weight circuit training, and the step that obtains the weight of each Weak Classifier comprises:

According to described N off-line training sample and a plurality of Weak Classifiers of weight training;

Calculate the total error of described N off-line training sample on each Weak Classifier, and according to the total error of described each Weak Classifier, obtain the weight of each Weak Classifier;

Upgrade the weight of described off-line training sample, and to described weight normalization.

4. Adaboost sorter on-line study method according to claim 1, is characterized in that, described background model is gauss hybrid models, background subtraction point-score model or frame average background model;

When described background model is gauss hybrid models, described employing background model is carried out target detection, and the step that obtains moving target comprises:

If present frame is the first frame, the gaussian component in initialization gauss hybrid models;

If present frame is not the first frame, utilize each pixel of present frame to upgrade the gaussian component in gauss hybrid models, and according to the comparison of present frame and each gaussian component, obtain foreground pixel, form foreground picture;

Described foreground picture is processed, extracted UNICOM's component, obtain moving target;

When described background model is background subtraction point-score model, described employing background model is carried out target detection, and the step that obtains moving target comprises:

Adopt present frame and former frame to subtract each other and obtain difference diagram, according to difference diagram, carry out thresholding and extract foreground picture;

Described foreground picture is processed, extracted UNICOM's component, obtain moving target;

When described background model is frame average background model, described employing background model is carried out target detection, and the step that obtains moving target comprises:

By the weighted mean of all frames before present frame as a setting, then obtain difference diagram with present frame and described background subtracting, according to difference diagram, carry out thresholding and extract foreground picture;

Described foreground picture is processed, extracted UNICOM's component, obtain moving target.

5. according to the Adaboost sorter on-line study method described in any one in claim 1 to 4, it is characterized in that, describedly using the target of described misclassification device as training sample, train online, the step of the strong classifier after being upgraded comprises:

Using the target of described misclassification device as online training sample, the weight of initialization training sample;

According to take Poisson distribution that weight is parameter, the number of times of training is set;

Weak Classifier is trained according to the number of times of described setting according to described online training sample;

If online training sample is correctly classified, upgrade the parameter of Poisson distribution, calculate the number of times accumulated value of correct classification, the error rate of calculating sorter;

If online training sample, by mis-classification, upgrades the parameter of Poisson distribution, the error rate of the number of times accumulated value of miscount classification, calculating sorter;

Strong classifier after being upgraded.

6. an Adaboost sorter on-line study system, is characterized in that, comprising:

First detection module, for adopting the strong classifier that off-line training obtains to carry out target detection, and obtains target detection result;

The second detection module, for adopting background model to carry out target detection, obtains moving target;

Comparison module, detects and obtains moving target and compare for target detection result that strong classifier is obtained and background model, obtains the target of misclassification device;

Online training module, for using the target of described misclassification device as online training sample, trains the strong classifier after being upgraded online.

7. Adaboost sorter on-line study system according to claim 6, is characterized in that, described system also comprises:

Off-line training module, the strong classifier obtaining for off-line training, comprising:

Off-line initialization unit, for given N off-line training sample, and the weight of initialization off-line training sample;

Off-line training unit, for according to described N off-line training sample and a plurality of Weak Classifiers of weight circuit training, and obtains the weight of each Weak Classifier;

Off-line output unit, for obtaining strong classifier according to described a plurality of Weak Classifiers and corresponding weight.

8. Adaboost sorter on-line study system according to claim 6, is characterized in that, described off-line training unit comprises:

Weak Classifier training subelement, for according to described N off-line training sample and a plurality of Weak Classifiers of weight training;

Weak Classifier weight calculation subelement, for calculating the total error of described N off-line training sample on each Weak Classifier, and obtains the weight of each Weak Classifier according to the total error of described each Weak Classifier;

Training sample weight is upgraded subelement, for upgrading the weight of described off-line training sample, and to described weight normalization.

9. Adaboost sorter on-line study system according to claim 6, is characterized in that, described background model is gauss hybrid models, background subtraction point-score model or frame average background model;

When described background model is gauss hybrid models, described the second detection module comprises:

Component initialization unit, if be the first frame for present frame, the gaussian component in initialization gauss hybrid models;

Extraction unit, if be not the first frame for present frame, utilize each pixel of present frame to upgrade the gaussian component in gauss hybrid models, and according to the comparison of present frame and each gaussian component, obtains foreground pixel, forms foreground picture;

Acquiring unit, for described foreground picture is processed, extracts UNICOM's component, obtains moving target;

When described background model is background subtraction point-score model, described the second detection module comprises:

Extraction unit, obtains difference diagram for adopting present frame and former frame to subtract each other, and carries out thresholding extract foreground picture according to difference diagram;

Acquiring unit, for described foreground picture is processed, extracts UNICOM's component, obtains moving target;

When described background model is frame average background model, described the second detection module comprises:

Extraction unit, for by the weighted mean of all frames before present frame as a setting, then obtains difference diagram with present frame and described background subtracting, carries out thresholding extract foreground picture according to difference diagram;

Acquiring unit, for described foreground picture is processed, extracts UNICOM's component, obtains moving target.

10. according to the Adaboost sorter on-line study method described in any one in claim 6 to 9, it is characterized in that, described online training module comprises:

Online initialization unit, for using the target of described misclassification device as online training sample, the weight of initialization training sample;

Setting unit, for arranging the number of times of training according to take Poisson distribution that weight is parameter;

Online training unit, for training according to described online training sample Weak Classifier according to the number of times of described setting;

Updating block, for when online training sample is correctly classified, upgrades the parameter of Poisson distribution, calculates the number of times accumulated value of correct classification, the error rate of calculating sorter; And when online training sample is during by mis-classification, upgrade the parameter of Poisson distribution, the number of times accumulated value of miscount classification, calculate the error rate of sorter, the strong classifier after then being upgraded.

Images