CN103902968A - Pedestrian detection model training method based on AdaBoost classifier - Google Patents

Abstract

The invention discloses a pedestrian detection model training method based on an AdaBoost classifier. The pedestrian detection model training method comprises the steps of firstly, conducting real-time statistics on the sum of sample weight values in the AdaBoost training process, when degeneration is carried out to a certain extent, using a currently-trained weak classifier set for scanning a non-pedestrian image for a false detection window, using the false detection window as a difficult sample to be added in negative sample training sets, and decreasing a degeneration degree threshold value so as to reduce sample update efficiency; finally, removing a part of negative samples through random sampling, and reducing the number of the negative sample training sets so as to reduce the calculated amount of the training process. According to the pedestrian detection model training method based on the AdaBoost classifier, on the premise that a feature extraction method is not changed, the training effect of the classifier can be improved to the maximum extent, and the final detection precision is improved.

Description

A kind of pedestrian detection model training method based on AdaBoost sorter

Technical field:

The present invention is mainly concerned with the pedestrian detection field based on still image, refers in particular to a kind of pedestrian detection model training method based on AdaBoost sorter.

Background technology:

Traffic hazard is to cause casualties peacetime and the major reason of property loss, wherein, there is the traffic hazard that pedestrian participates in to account for 14% of traffic hazard total amount, extremely whole society's extensive concern of the safety problem of pedestrian in road traffic, researchist starts to be devoted to vehicle assistant drive technology and reduces the accident rate of pedestrian in road traffic, and it is particularly important that pedestrian detection technology seems therein.Because pedestrian has the characteristic of rigidity and flexible article concurrently, the impacts [4] such as outward appearance is subject to dress, yardstick, blocks, attitude and visual angle, make pedestrian detection become Research Challenges and the focus of computer vision.

In the last few years, the method for machine learning was introduced in the research of pedestrian detection, and Detection accuracy is improved, and the research of pedestrian detection at present mainly concentrates on the selection of feature extraction and sorter.The method that the people such as Pepageorgios have proposed sliding window for the first time for detection of, the mode that they have adopted SVM and Multi-Scale Haar Wavelet to cross complete base combination detects, the people such as Viola and Jones, based on this thinking, has introduced integrogram thought and AdaBoost Fast Classification device (VJ).Above-mentioned two kinds of thinkings have formed the nowadays foundation stone of pedestrian detection operator.Then the people such as Dalal, based on pedestrian's overall profile feature, proposes histograms of oriented gradients feature (HOG), and HOG feature is in conjunction with svm classifier device, and Detection accuracy has been obtained revolutionary progress.The people such as Dollar are merged the thought of Viola and Dalal, propose integration channel characteristics (ChnFtr), in conjunction with AdaBoost sorter, have promoted detection speed and accuracy in detection.

AdaBoost, is the abbreviation of English " Adaptive Boosting " (self-adaptation enhancing), is a kind of machine learning method, is proposed by Yoav Freund and Robert Schapire ^[1].AdaBoost method is a kind of iterative algorithm, adds a new Weak Classifier in each is taken turns, until reach certain predetermined enough little error rate or reach predetermined Weak Classifier number.The self-adaptation of AdaBoost method is: the sample of previous sorter misclassification can be used to train next sorter.AdaBoost method is very sensitive for noise data and abnormal data.But in some problems, than most of learning algorithms, AdaBoost method, for other learning algorithm of great majority, can not be easy to occur over-fitting phenomenon.The sorter using in AdaBoost method may very weak (such as occurring very serious mistake rate), but as long as its classifying quality better than at random (such as two class Question Classification error rates are slightly less than 0.5) just can improve the model finally obtaining.And error rate is also useful higher than the Weak Classifier of random assortment device, because in the linear combination of the multiple sorters that finally obtain, can give negative coefficient to them, equally also can promote classifying quality.

No matter based on which kind of method, all need great amount of samples train sorter that accuracy rate is high for detection of.In order better to contrast the effect of algorithms of different, researchers have proposed the research of series of standards data set for pedestrian detection algorithm, conventional have MIT data set, INRIA data set, ETH data set and a Caltech data set, these data sets are all made up of two parts sample set, be respectively used to training and testing, every partitioned data set (PDS) all comprises pedestrian's image collection and non-pedestrian's image collection, is positive sample image and negative sample image.Wherein MIT and INRIA are made up of still image, and ETH and Caltech are made up of sequence of frames of video.Obtaining in training set process, positive sample need to be according to comment file, from positive sample image, extract pedestrian's image of fixed resolution, general all for the training of sorter, negative sample can obtain from negative sample picture frame block diagram non-pedestrian's image of magnanimity, researcher can carry out cutting and screening according to own actual demand, and the image resolution ratio of the positive and negative sample training collection obtaining is 64*128.In the last few years, in the research in pedestrian detection field, researchers mostly only pay close attention to characteristic Design and sorter is selected, and had ignored the reasonable utilization of training sample information, can not fully excavate the effect of institute's using method.For the negative sample information of magnanimity, how it is carried out to adequately and reasonably cutting and screening, the false drop rate, the lifting sorter training effectiveness that reduce sorter are played to vital effect.

Summary of the invention:

The technical problem to be solved in the present invention is: current most of pedestrian detection method is mainly paid close attention in the selection and optimization of feature and sorter, and ignore the reasonable application to magnanimity negative sample in training process, the present invention aims to provide a kind of pedestrian detection model training method based on AdaBoost sorter, in conjunction with AdaBoost Weak Classifier training process, according to the degree of degeneration of sample training centralization of state power value, negative sample training set is carried out reasonably dynamically adjusting, keep the diversity of negative sample training set, promote the training effectiveness of Weak Classifier, can effectively promote the accuracy of final disaggregated model, optimize the efficiency of training process, reduce the training time.

In order to solve the problems of the technologies described above, the technical scheme that the present invention proposes is: a kind of pedestrian detection model training method based on AdaBoost sorter, it is characterized in that: the weight of first determining initial positive and negative sample training collection and positive negative sample by known standard pedestrian's data set, then in AdaBoost training process in real-time statistics training set sample weights value and, in the time there is obvious weight degradation in training set, the non-pedestrian's image that goes sliding window scanning detection raw data to concentrate by the current Weak Classifier group having trained, and add in negative sample training set flase drop video in window as difficult sample, then reject part negative sample by random sampling, finally reduce degree of degeneration threshold value, reduce training set renewal frequency, continue AdaBoost training process, until finally obtain pedestrian detection model.

The specific implementation step of the above-mentioned pedestrian detection model training method based on AdaBoost sorter is:

1) determine initial positive and negative sample training collection:

First, select a pedestrian detection standard data set, comprising the positive sample image that comprises pedestrian with do not comprise pedestrian's negative sample image.According to the comment file of data set, from obtaining pedestrian's image of 64*128 and carry out mirror process, positive sample image Block Diagrams extracts image integration channel characteristics, form positive sample training collection

wherein N=2416, negative sample image Block Diagrams obtains non-pedestrian's image of countless 64*128, is regarded as magnanimity negative sample raw data set $S_{\∞}^{-}=\{(x_1,-1),\·\·\·,(x_{N_{\mathrm{\ω}}^{-}},-1)\},(N_{\∞}^{-}\→\∞),$ Again by get non-pedestrian's image of fixed qty with machine frame on every width negative sample image, from S ^-in obtain original negative sample training collection wherein N ^-=20,000.Therefore, initial training collection is $S=S^{+}\∪S_0^{-}=\{(x_1,y_1),\·\·\·,(x_N,y_N)\},$ Wherein N=N ⁺+ N ^-;

2) weight of the positive and negative sample training collection of initialization is:

$\mathrm{\ω}=\{{\mathrm{\ω}}_1,{\mathrm{\ω}}_2,\·\·\·,{\mathrm{\ω}}_N\}=\left\{\begin{array}{c}{\mathrm{\ω}}^{+}=\left\{{\mathrm{\ω}}_i^{+}\right|{\mathrm{\ω}}_i^{+}=1/2N^{+},i=\mathrm{1,2},\·\·\·,N^{+}\}\\ {\mathrm{\ω}}^{-}=\left\{{\mathrm{\ω}}_i^{-}\right|{\mathrm{\ω}}_i^{-}=1/2N^{-},i=\mathrm{1,2},\·\·\·,N^{-}\}\end{array}\right.$

Wherein with

represent respectively the weight of positive sample and negative sample;

3) initialization difficulty sample adds the upper limit

initial value τ=0.1 of weights degree of degeneration threshold value is set;

4) establishing AdaBoost Weak Classifier number is T=2048, initialization t=0;

5) t=t+1 forwards step 13 in the time of t>T);

6) according to known Weak Classifier training method (referring to specific implementation method), based on step 1) and step 2) the training set S that obtains and its weight sets ω, training obtains t Weak Classifier ht (x) ∈ { 1,1};

7) the classification error rate of statistics Weak Classifier in sample

this Weak Classifier h _t(x) weight is α _t=log ((1-err _t) err _t);

8) upgrade sample weights ω _n=ω _nexp (α _th _t(x _n)), and calculate negative sample weight and ${\mathrm{loss}}_t=\underset{y_n=-1}{\underset{{\mathrm{\ω}}_n\∈\mathrm{\ω}}{\mathrm{\Σ}}}{\mathrm{\ω}}_n;$

9) if loss _t> τ, training set S remains unchanged, and jumps to step 5); Otherwise, make τ=τ/10, continue following steps;

10) using the t a having trained Weak Classifier as current sorter

on original minus sample image, use current sorter H _t(x) scan N ^fPthe sliding window FP(False of individual flase drop Positives), after extraction integration channel characteristics, add negative sample training set S to ^-in, wherein

wherein

for difficult sample adds the upper limit;

11) from negative sample training set S ^-in random reject a part of negative sample, make final negative sample number be

12) upgrade negative sample training set S ^-in the weights omega of the new difficult sample adding _n=exp (H _t(x _n)) 2N ^-, wherein (x _n, y _n) ∈ FP, FP refers to difficult sample, then jumps to step 5);

13) obtain final disaggregated model $H\left(x\right)=\mathrm{sign}\left(\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\α}}_t{h}_t\left(x\right)\right).$

In sum, of the present inventionly reduce the false drop rate of final mask based on taking full advantage of negative sample training set information, and by dynamically controlling training set sample size, promoted training effectiveness.On the basis based on same characteristic features extracting method, the present invention trains gained disaggregated model to have higher detection precision.

List of references:

[1]Freund Y,Schapire R E.A desicion-theoretic generalization of on-line learning and an application to boosting[C]//Computational learning theory.Springer Berlin Heidelberg,1995:23-37.

[2]Dollár P,Tu Z,Perona P,et al.Integral Channel Features[C]//BMVC.2009,2(4):5.

[3]Dollár P,Belongie S,Perona P.The Fastest Pedestrian Detector in the West[C]//BMVC.2010,2(3):7.

[4]Dollar P,Wojek C,Schiele B,et al.Pedestrian detection:An evaluation of the state of the art[J].Pattern Analysis and Machine Intelligence,IEEE Transactions on,2012,34(4):743-761.

Accompanying drawing explanation:

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

It is the obtain manner of positive and negative sample training collection that Fig. 2 (a), Fig. 2 (b) distribute, positive sample training collection is pedestrian's image of getting 64*128 from positive sample image center, and negative sample training set is non-pedestrian's image of getting corresponding 64*128 from background image according to certain tactful frame;

The integrogram of 10 passages of Fig. 3 mono-secondary pedestrian image;

Fig. 4 is the training effect comparison diagram based under different threshold values, this figure is the ROC curve of gained after sorter that application training is good detects on test set, wherein horizontal ordinate fppi (false positive per-image) represents the average false drop rate of every width image, and ordinate miss rate represents loss;

Fig. 5 is the training effect comparison diagram under different training set update modes, by contrasting negative sample at t=[16642561024] moment renewal and the even training effect upgrading, give prominence to the importance of training process in early stage, verify the outstanding immanent cause of training effect under dynamic threshold, wherein horizontal ordinate fppi represents the average false drop rate of every width image, and ordinate miss rate represents loss;

Fig. 6 (a) is the disaggregated model effect contrast figure who randomly draws under varying number negative sample training set, (b) be to add difficult sample by the Bootstrapping method of different rounds, the modelling effect comparison diagram training under different sample diversity conditions, wherein as above figure of transverse and longitudinal coordinate definition, can find out by two width figure, the diversity of training sample is more remarkable on the impact of final disaggregated model;

Fig. 7 is under based on integration channel characteristics method, the comparison diagram of the final disaggregated model that the original training method of training method of the present invention and Dollar trains.

Embodiment:

As shown in Figure 1, the present embodiment provides a kind of pedestrian detection model training method based on AdaBoost sorter:

1. determining of positive negative sample initial training collection:

First, select a suitable pedestrian detection standard data set, for example INRIA data set, comprising the positive sample image that comprises pedestrian with do not comprise pedestrian's negative sample image.According to the comment file of data set, from positive sample image Block Diagrams obtains pedestrian's image of 64*128 and carries out Mirror Symmetry processing and copies, the extraction image integration channel characteristics [2] (the 4th of aftermentioned has a detailed description) that adopts document 3 to propose, namely all positive sample pedestrian images are carried out after reverse symmetry copies, joining positive sample training and concentrating.Form by the way positive sample training collection

wherein N=2416, can block diagram in negative sample image obtains non-pedestrian's image of countless 64*128, and we are regarded as magnanimity negative sample raw data set $S_{\∞}^{-}=\{(x_1,-1),\·\·\·,(x_{N_{\mathrm{\ω}}^{-}},-1)\},(N_{\∞}^{-}\→\∞),$ Can be by get non-pedestrian's image of fixed qty with machine frame on every width negative sample image, from S ^-in obtain original negative sample training collection

wherein N ^-=20,000.Therefore, initial training collection is $S=S^{+}\∪S_0^{-}=\{(x_1,y_1),\·\·\·,(x_N,y_N)\},$ Wherein N=N ⁺+ N ^-;

And the weight of these positive and negative sample training collection of initialization:

$\mathrm{\ω}=\{{\mathrm{\ω}}_1,{\mathrm{\ω}}_2,\·\·\·,{\mathrm{\ω}}_N\}=\left\{\begin{array}{c}{\mathrm{\ω}}^{+}=\left\{{\mathrm{\ω}}_i^{+}\right|{\mathrm{\ω}}_i^{+}=1/2N^{+},i=\mathrm{1,2},\·\·\·,N^{+}\}\\ {\mathrm{\ω}}^{-}=\left\{{\mathrm{\ω}}_i^{-}\right|{\mathrm{\ω}}_i^{-}=1/2N^{-},i=\mathrm{1,2},\·\·\·,N^{-}\}\end{array}\right.$

Wherein

with

represent respectively the weight of positive sample and negative sample.

This weight sets represents the significance level of training sample set in training Weak Classifier process, and by the more samples of Weak Classifier classification error number of times, weight is larger, often completes the training of a sorter later, and weight is according to following Policy Updates:

ω _n=ω _nexp(α _th _t(x _n)) (2)

: ω _n=exp (H _t(x _n)) 2N ^-, (3) wherein $H_t\left(x\right)=\underset{k=1}{\overset{t}{\mathrm{\Σ}}}{\mathrm{\α}}_k{h}_k\left(x\right),$ (x _n,y _n)∈FP；

Finally, according to weight sets training Weak Classifier h _t(x) { 1 ,+1}, adds up this Weak Classifier error rate err to ∈ _t, and calculate these Weak Classifier weights α _t, computation process is as follows:

${\mathrm{err}}_t=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\ω}}_i\·1[y_i\≠h_i\left(x_i\right)]---\left(4\right)$

α _t=log((1-err _t)err _t) (5)

Use T final Weak Classifier to form final strong classifier

Wherein Weak Classifier training method is the published content of document 3, specifically sees the 5th description of this part.

2. degenerate and control the renewal of negative sample training set according to weights:

Weights and the weights degree of degeneration that represents negative sample training set with negative sample:

${\mathrm{loss}}_t=\underset{y_n=-1}{\underset{{\mathrm{\ω}}_n\∈\mathrm{\ω}}{\mathrm{\Σ}}}{\mathrm{\ω}}_n---\left(6\right)$

Wherein weights omega _n=ω _nexp (α _th _t(x _n))

Regard front t Weak Classifier group as current sorter H _t(x), weights degree of degeneration found out by formula (6), as weights degree of degeneration loss _tmuch smaller than 1/2 o'clock, illustrate that most of negative sample meets H _t(x _i) <0, i.e. current sorter H _t(x) negative sample training set has been possessed to good separating capacity, there is serious degradation in current negative sample training set, ensuing sorter training process affected to conspicuousness variation, need to from original background image, again obtain new difficult sample, weights degree of degeneration loss is set _tinitial value τ=0.1 of threshold value, in the time being less than threshold values, upgrades negative sample training set; And upgrade weight degradation threshold tau=τ/10, can progressively reduce like this negative sample renewal frequency, by computational resource more for the training process in early stage, guaranteeing in training effect, the less training time to greatest extent.

In the AdaBoost sorter based on softcascade, often most of detection window only needs front several Weak Classifier just can be defined as non-pedestrian, so in actual testing process, sorter frequency of utilization more forward in AdaBoost Weak Classifier group is higher, its significance level is also just more remarkable, should more pay close attention to so the training of training Weak Classifier in early stage in training process.

3. dynamically control negative sample training set quantity:

Compare sample size, sample diversity is more obvious for the impact of training effect, is the shortening training time in training process, dynamically control negative sample quantity, complete after the renewal of negative sample training set at every turn, weed out at random part negative sample, the number of new negative sample training set is met:

like this, be greater than the negative sample quantity in training later stage in the quantity of training negative sample in early stage, reduced the average computation amount of Weak Classifier, realized the acceleration of training process.

4. integration channel characteristics:

The present invention has used published integration channel characteristics extracting method in document 3, its core concept is carried out multiple linearity or nonlinear conversion to input picture exactly, and with the form storage of integrogram, to be similar to the rectangular area of Harr feature and as feature, can to pass through the quick calculating of integrogram realization character.

Suppose that input picture is I, Ω represents the mapping relations of certain channel image and source images I, can obtain thus respective channel mapping graph C=Ω (I), and C is converted to the form of integrogram f, wherein

dollar has used in test ten class integration channel C _k(k=1,2 ..., 10), be respectively three LUV Color Channels, gradient assignment passage, six gradient direction passages (as shown in Figure 3).

(1) LUV Color Channel: input original image I is RGB color space, is translated into LUV color space and represents, three the channel information composition LUV Color Channel figures of image I on LUV color space,

k=1,2,3.

(2) gradient assignment passage: gradient assignment channel C ₄the gray-scale map I of calculating based on image I ^gray, computing method are as follows:

$C_4(x,y)=\sqrt{{\left(\frac{{\&PartialD;I}^{\mathrm{gray}}(x,y)}{\&PartialD;x}\right)}^2+{\left(\frac{{\&PartialD;I}^{\mathrm{gray}}(x,y)}{\&PartialD;y}\right)}^2}.$

(3) gradient direction passage: first will be based on gray-scale map I ^graycalculate the gradient direction of each location of pixels:

$\mathrm{\&Theta;}(x,y)={\tan }^{-1}\left(\frac{\frac{{\&PartialD;I}^{\mathrm{gray}}}{\&PartialD;x}(x,y)}{\frac{{\&PartialD;I}^{\mathrm{gray}}}{\&PartialD;y}(x,y)}\right),$

Then

discrete is six gradient directions obtain the gradient assignment passage figure on six direction: $C_k(x,y)=C_4(x,y)\&CenterDot;1[\mathrm{\&Theta;}(x,y)={\mathrm{\&theta;}}_{k-5}],$ Wherein k=5,6 ..., 10.

For the quick calculating of realization character, Dollar by characterizing definition be random rectangular area on passage figure pixel value and, can under constant complexity, complete calculating by passage integrogram, that is:

$\underset{{\mathrm{<figure-callout\; id="1"\; label="y"\; filenames="DEST\_PATH\_HDA0000502237640000011.png,DEST\_PATH\_HDA0000502237640000031.png"\; state="\{\{state\}\}">y</figure-callout>}}_1<y<y_2}{\underset{x_1<x<x_2}{\mathrm{\&Sigma;}}}{C}_k(x,y)=f_k({\mathrm{<figure-callout\; id="2"\; label="x"\; filenames="DEST\_PATH\_HDA0000502237640000051.png"\; state="\{\{state\}\}">x</figure-callout>}}_2,y_2)-f_k({\mathrm{<figure-callout\; id="1"\; label="x"\; filenames="DEST\_PATH\_HDA0000502237640000011.png,DEST\_PATH\_HDA0000502237640000031.png"\; state="\{\{state\}\}">x</figure-callout>}}_1,y_2)-f_k({\mathrm{<figure-callout\; id="2"\; label="x"\; filenames="DEST\_PATH\_HDA0000502237640000051.png"\; state="\{\{state\}\}">x</figure-callout>}}_2,y_1)+f_k({\mathrm{<figure-callout\; id="1"\; label="x"\; filenames="DEST\_PATH\_HDA0000502237640000011.png,DEST\_PATH\_HDA0000502237640000031.png"\; state="\{\{state\}\}">x</figure-callout>}}_1,y_1).$

Integration passage belongs to a kind of feature of many Fusion Features, but it has solved the slow-footed shortcoming of many Fusion Features by the method for integrogram, and in testing process, space orientation accuracy is higher, there is less parameter setting, be well suited for, with cascade classifier combination, reaching detection speed faster.

5. the present invention has adopted training and the screening of document 3 disclosed Weak Classifiers:

For each sample image, first we ask for the integrogram of 10 passages, then on passage integrogram random 10, the rectangle of 000 4*4, using the pixel value of the passage figure in rectangle and as eigenwert, this value can, by the form of integrogram, be obtained fast under constant complexity.

In the time of t Weak Classifier of training, we are for 10, the feature that had not participated in final Weak Classifier training in 000 dimensional feature uses respectively the mode of AdaBoost to train multiple Weak Classifiers, and the error rate of adding up them, we are using Weak Classifier minimum classification error rate as t final Weak Classifier, complete the screening to Weak Classifier, and mark respective dimension feature is no longer applied to ensuing training process, the Weak Classifier remaining finally participates in forming AdaBoost strong classifier.

The concrete methods of realizing process description of technique scheme is as follows:

Input:

● pedestrian detection training set (comprising the non-pedestrian's image of positive sample pedestrian image and negative sample);

● AdaBoost Weak Classifier is counted T;

● difficult sample (FP) adds the upper limit

with negative sample training set capacity N ^-.

Initialization:

● using positive sample pedestrian image as positive sample training collection S ⁺, original negative sample training collection S ^-from non-pedestrian's image, random sliding window frame is got and is obtained, and sample number is N ^-, form AdaBoost training set S=S ⁺∪ S ^-.

● the positive sample weights of initialization

with negative sample weights ${\mathrm{\&omega;}}^{-}=\left\{{\mathrm{\&omega;}}_n^{-}\right|{\mathrm{\&omega;}}_n^{-}=1/2N^{-}\},$ Form weight sets ω=ω ⁺∪ ω ^-;

● weights degree of degeneration threshold tau=0.1 is set.

Flow process:

For t=1,…,T

1. based on training set (S, ω), train t Weak Classifier ht (x) ∈ { 1,1};

2. miscount rate α _t=log ((1-err _t) err _t);

3. upgrade sample weights ω={ ω _n| ω _n=ω _nexp (α _th _t(x _n)), calculate negative sample weight and: ${\mathrm{loss}}_t=\underset{{\mathrm{\&omega;}}_n^{-}\&Element;{\mathrm{\&omega;}}^{-}}{\mathrm{\&Sigma;}}{\mathrm{\&omega;}}_n^{-};$

4.If loss _t<τ

4.1 on original non-pedestrian's image, by current sorter

scanning is less than flase drop window add negative sample training set S to ^-in;

4.2 from S ^-in randomly draw individual negative sample, forms new negative sample training set S ^-;

4.3 upgrade S ^-in the weights omega of the new difficult sample adding _n=exp (H _t(x _n)) 2N ^-, (x _n, y _n) ∈ FPs;

5. obtain final disaggregated model $H\left(x\right)=\mathrm{sign}\left(\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_t{h}_t\left(x\right)\right)$

For verifying effect of the present invention, the integration channel characteristics that the present embodiment proposes based on people such as Dollar, improve according to said method, having carried out programming realizes, train the sorter that final sorter and Dollar original method train and contrast, as shown in Figure 7, under same characterization method, the final disaggregated model effect training by the present embodiment training method is better than original method, has higher accuracy of detection.

Claims (2)

1. the pedestrian detection model training method based on AdaBoost sorter, it is characterized in that: the weight of first determining initial positive and negative sample training collection and positive negative sample by known standard pedestrian's data set, then in AdaBoost training process in real-time statistics training set sample weights value and, in the time there is obvious weight degradation in training set, the non-pedestrian's image that goes sliding window scanning detection raw data to concentrate by the current Weak Classifier group having trained, and add in negative sample training set flase drop video in window as difficult sample, then reject part negative sample by random sampling, finally reduce degree of degeneration threshold value, reduce training set renewal frequency, continue AdaBoost training process, until finally obtain pedestrian detection model.

2. the pedestrian detection model training method based on AdaBoost sorter according to claim 1, is characterized in that, specifically comprises the steps:

1) determine initial positive and negative sample training collection:

First, select a pedestrian detection standard data set, comprising the positive sample image that comprises pedestrian with do not comprise pedestrian's negative sample image; According to the comment file of described data set, from obtaining pedestrian's image of 64*128 and carry out mirror process, positive sample image Block Diagrams extracts image integration channel characteristics, form positive sample training collection

wherein N=2416, negative sample image Block Diagrams obtains non-pedestrian's image of countless 64*128, and is regarded as magnanimity negative sample raw data set $S_{\&infin;}^{-}=\{(x_1,-1),\&CenterDot;\&CenterDot;\&CenterDot;,(x_{N_{\mathrm{\&omega;}}^{-}},-1)\},(N_{\&infin;}^{-}\&RightArrow;\&infin;),$ Again by get non-pedestrian's image of fixed qty with machine frame on every width negative sample image, from S ^-in obtain original negative sample training collection

wherein N ^-=20,000; Therefore, initial training collection is $S=S^{+}\&cup;S_0^{-}=\{(x_1,y_1),\&CenterDot;\&CenterDot;\&CenterDot;,(x_N,y_N)\},$ Wherein N=N ⁺+ N ^-;

2) weight of the positive and negative sample training collection of initialization is:

$\mathrm{\&omega;}=\{{\mathrm{\&omega;}}_1,{\mathrm{\&omega;}}_2,\&CenterDot;\&CenterDot;\&CenterDot;,{\mathrm{\&omega;}}_N\}=\left\{\begin{array}{c}{\mathrm{\&omega;}}^{+}=\left\{{\mathrm{\&omega;}}_i^{+}\right|{\mathrm{\&omega;}}_i^{+}=1/2N^{+},i=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,N^{+}\}\\ {\mathrm{\&omega;}}^{-}=\left\{{\mathrm{\&omega;}}_i^{-}\right|{\mathrm{\&omega;}}_i^{-}=1/2N^{-},i=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,N^{-}\}\end{array}\right.$

Wherein

with

represent respectively the weight of positive sample and negative sample;

3) initialization difficulty sample adds the upper limit initial value τ=0.1 of weights degree of degeneration threshold value is set;

4) establishing AdaBoost Weak Classifier number is T=2048, initialization t=0;

5) t=t+1 forwards step 13 in the time of t>T);

6) according to Weak Classifier training method, based on step 1) and step 2) the training set S that obtains and its weight sets ω, training obtains t Weak Classifier h _t(x) ∈ { 1,1};

7) the classification error rate of statistics Weak Classifier in sample this Weak Classifier h _t(x) weight is α _t=log ((1-err _t) err _t);

8) upgrade sample weights ω _n=ω _nexp (α _th _t(x _n)), and calculate negative sample weight and ${\mathrm{loss}}_t=\underset{y_n=-1}{\underset{{\mathrm{\&omega;}}_n\&Element;\mathrm{\&omega;}}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_n;$

9) if loss _t> τ, training set S remains unchanged, and jumps to step 5); Otherwise, make τ=τ/10, continue following steps;

10) using the t a having trained Weak Classifier as current sorter

on original minus sample image, use current sorter H _t(x) scan N ^fPthe sliding window FP(False of individual flase drop Positives), after extraction integration channel characteristics, add negative sample training set S to ^-in, wherein

wherein

for difficult sample adds the upper limit;

11) from negative sample training set S ^-in random reject a part of negative sample, make final negative sample number be

12) upgrade negative sample training set S ^-in the weights omega of the new difficult sample adding _n=exp (H _t(x _n)) 2N ^-, wherein (x _n, y _n) ∈ FP, then jump to step 5);

13) obtain final disaggregated model $H\left(x\right)=\mathrm{sign}\left(\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_t{h}_t\left(x\right)\right).$

Images