CN104732248A - Human body target detection method based on Omega shape features - Google Patents

Abstract

The invention discloses a human body target detection method based on Omega shape features. The method includes the steps that first, the HOG features of all samples in training samples are extracted, the feature column vectors of the positive samples and the negative samples are combined to obtain two feature matrixes, the feature matrix of the positive samples is subjected to orthogonal non-negative matrix factorization, an obtained orthogonal basis matrix is used as the initial value of a feature projection matrix, and a final feature projection matrix is obtained through an LMNN algorithm according to the feature column vectors of all the samples; then the two feature matrixes are mapped according to the feature projection matrix, and each feature column vector in an obtained new feature matrix and a corresponding class tag message are input to an SVM classifier to be trained; afterwards, the obtained SVM classifier is used for conducting human body target detection. According to the human body target detection method, a learning method based on orthogonal non-negative matrix factorization and large margin nearest neighbor distance measure is used for lowering data redundancy and improving detection precision on the basis of the HOG features.

Description

Based on the human body target detection method of Omega shape facility

Technical field

The invention belongs to technical field of computer vision, more specifically say, relate to a kind of human body target detection method based on Omega shape facility.

Background technology

Human detection is under the jurisdiction of the important artificial intelligence technology of of computer vision field.It gives the function of machine type like human visual, take human body as searched targets, obtain ambient image by image collecting device and convert thereof into digital signal, and utilize different image processing techniquess to carry out statistics and analysis to image, and then human body is identified and locates, be follow-up application demand, such as human body tracking, behavioural analysis, scene cut etc., establish certain identification basis.In recent years, along with the constantly perfect of detection algorithm and the significantly increase of computer computation ability, Human Detection is widely used in intelligent monitoring, Smart Home, intelligent transportation, robot visual guidance, the aspects such as Entertainment.Requirement for human body target detection algorithm is how robustly to carry out human body target detection, reduces false drop rate, improves verification and measurement ratio.

At present, there is a lot of researchist to detect association area to human body target and be studied work.Prior art adopts the detection method of feature based and sorter to detect human body target mostly.Specific algorithm can list of references Viola P, Jones M J, Snow D.Detecting pedestrians using patterns of motionand appearance [C] .Computer Vision, 2003.Proceedings.Ninth IEEE InternationalConference on.IEEE, 2003:734-741. document Dalal N, Triggs B.Histograms of orientedgradients for human detection [C] .Computer Vision and Pattern Recognition, 2005.CVPR 2005.IEEE Computer Society Conference on.IEEE, 2005, 1:886-893. document Sermanet P, Kavukcuoglu K, Chintala S, et al.Pedestrian detection with unsupervisedmulti-stage feature learning [C] .Computer Vision and Pattern Recognition (CVPR), 2013 IEEE Conference on.IEEE, 2013:3626-3633. and document Doll á r P, Tu Z, PeronaP, et al.Integral Channel Features [C] .BMVC.2009, 2 (3): 5.

Said method is all detect human body target based on human body global characteristics, and these methods can realize the detection of human body target to a certain extent.But in actual scene, the detection of human body target also exists many difficult points.First, because human body is non-rigid object, along with the difference of person action, can there is significant change in the outer shape profile of human body, these non-rigid characteristic bring certain difficulty to the detection of human body target; Secondly, the human body target in motion to be easy to block by various barrier, also comparatively serious blocking can occur between men, the method now based on human body global characteristics is often greatly affected because of blocking, and causes serious undetected; Again, complicated background also can cause difficulty in many aspects to pedestrian detection, reduces the precision detected; Finally, the usual dimension of human body target feature obtained by existing method is higher, the time which increasing training and detect, and make sorter model solve difficulty, complex structure, and traditional dimension reduction method, such as principal component analysis (PCA) etc., although reduce characteristic dimension, cannot ensure the nonnegativity of decomposition result, and negative value result is often nonsensical in practical problems, cannot explain.And traditional dimension reduction method to the expression of data be generally based on entirety instead of based on local, and in practical problems, by portraying the local of data, the feature of essence, dimensionality reduction is carried out to data, for reduction data redundancy, improve accuracy of detection significant.

In actual applications, monitoring camera installation site as image collecting device is generally positioned at high-altitude, in the human body image that its visual angle, place obtains, the human body lower part of the body is often easily by tables and chairs, the barriers such as roadblock stopped, and human body head-take on is because of upper half of human body residing for it, the probability be blocked is also less.Meanwhile, compared to the whole detection of human body target, the head of upper half of human body-take on similar Omega shape facility is smaller by attitude, angle effects.Fig. 1 is upper half of human body head shoulder Omega example of shape figure.Academia has certain research to the method for being carried out human body target detection by human body head-shoulder model, as document Li M, Zhang Z, HuangK, et al.Rapid and robust human detection and tracking based on omega-shapefeatures [C] .Image Processing (ICIP), 200916th IEEE International Conference on.IEEE, 2009:2545-2548. existing algorithm mainly adopts HOG (Histogram of Oriented Gradient, gradient orientation histogram) feature statement is carried out to target, and in conjunction with the method for SVM linear classifier, human body head-shoulder target is detected, this mode improves the verification and measurement ratio of human body target, but cannot solve characteristic dimension higher with complex background on detecting the impact caused.

Summary of the invention

The object of the invention is to overcome the deficiencies in the prior art, a kind of human body target detection method based on Omega shape facility is provided, the HOG feature takeing on similar Omega shape by extracting upper half of human body head solves the problem of human body non-rigid shape influence of crust deformation testing result, and apply on the basis of HOG feature and estimate learning method based on orthogonal Non-negative Matrix Factorization and large-spacing nearest neighbor distance and reduce data redundancy, improve accuracy of detection.

For achieving the above object, the human body target detection method that the present invention is based on Omega shape facility comprises the following steps:

S1: select the image including upper half of human body head-shoulder Omega shape as positive sample in advance, selects the Background not containing human body target as negative sample, extracts HOG feature, obtain d dimensional feature column vector to the gray-scale map of each sample; By the d dimensional feature column vector that n in positive Sample Storehouse positive sample obtains be combined into eigenmatrix V _pos, the d dimensional feature column vector that the negative sample of m in negative example base is obtained be combined into eigenmatrix V _neg,

S2: to eigenmatrix V _poscarry out orthogonal Non-negative Matrix Factorization, obtain orthogonal basis matrix W _f;

S3: the sample characteristics column vector obtained according to step S1, adopts LMNN algorithm to obtain Projection Character matrix L, by orthogonal basis matrix W in LMNN Algorithm for Solving process _fas the initial value of Projection Character matrix L;

S4: the Projection Character matrix L obtained according to step S3, maps the eigenmatrix of positive and negative samples, obtains new eigenmatrix V _pos', V _neg', mapping equation is:

[V _pos',V _neg']＝L*[V _pos,V _neg]

By eigenmatrix V _pos' and V _neg' in each characteristic series vector input in SVM classifier with corresponding class label information and train, obtain SVM classifier;

S5: the SVM classifier adopting step S4 to obtain is treated detected image and carried out human body target detection, concrete grammar is: adjusted according to each zoom scale preset by image to be detected, travel through the region of search of the rear image of each adjustment with the search box of default size, extract the HOG characteristic series vector V of image in search box _o, calculate corresponding decision value f; Computing formula is:

f＝sv_coef*SVs*V _o'

V _o'＝(L*V _o)

All search boxes are judged one by one, if its decision value f is greater than predetermined threshold value F, then contains human body target in this search box, search box positional information is stored as Search Results; Finally the search box in Search Results is merged, obtain final testing result.

The present invention is based on the human body target detection method of Omega shape facility, first the HOG feature of each sample in training sample is extracted, respectively the characteristic series vector combination of positive sample and negative sample is obtained two eigenmatrixes, the eigenmatrix aligning sample carries out orthogonal Non-negative Matrix Factorization, using the orthogonal basis matrix that the obtains initial value as Projection Character matrix, characteristic series vector according to all samples adopts LMNN algorithm to obtain final Projection Character matrix, then according to Projection Character matrix, two eigenmatrixes are mapped, each characteristic series vector in the new feature matrix obtained is inputted SVM classifier with corresponding class label information train, then human body target detection is carried out by the SVM classifier obtained, when carrying out human body target and detecting, also the HOG characteristic series vector using Projection Character matrix to treat detected image is needed to map.

The present invention has following beneficial effect:

(1) based on upper half of human body head-shoulder Omega shape framework, carry out human body target detection, the probability decreasing detected object generation non-rigid shape deformations Yu be blocked, reduce the loss that human body target detects;

(2) have employed HOG to extract object edge characteristic information, this feature can characterize upper half of human body Omega shape better;

(3) adopt orthogonal NMF method to be further analyzed feature, extract the local feature that data are the most essential, effectively suppress to block the interference caused detection with complex background, lower characteristic dimension simultaneously, reduce sorter and train complexity;

(4) adopt LMNN distance measure learning method, set up multi-modal discriminant apparent model, make sorter have higher identification.

Accompanying drawing explanation

Fig. 1 is upper half of human body head shoulder Omega example of shape figure;

Fig. 2 is the embodiment process flow diagram of the human body target detection method that the present invention is based on Omega shape facility;

Fig. 3 is the concrete implementing procedure figure of orthogonal Non-negative Matrix Factorization

Fig. 4 is the embodiment process flow diagram that human body target detects

Fig. 5 is the FPPI-MR comparison diagram of the test result of Lung biopsy

Fig. 6 is partial detection exemplary plot of the present invention.

Embodiment

Below in conjunction with accompanying drawing, the specific embodiment of the present invention is described, so that those skilled in the art understands the present invention better.Requiring particular attention is that, in the following description, when perhaps the detailed description of known function and design can desalinate main contents of the present invention, these are described in and will be left in the basket here.

Embodiment

Fig. 2 is the embodiment process flow diagram of the human body target detection method that the present invention is based on Omega shape facility.As shown in Figure 2, the concrete steps that the present invention is based on the human body target detection method of Omega shape facility comprise:

S201: obtain sample HOG eigenmatrix:

In the present invention, the positive sample for training classifier is the image including upper half of human body head-shoulder Omega shape, and negative sample is not then containing the Background of human body target.If sample is RGB figure, be first translated into gray-scale map, then HOG feature extracted to the gray-scale map of each sample, obtain d dimensional feature column vector.By the d dimensional feature column vector that n in positive Sample Storehouse positive sample obtains be combined into eigenmatrix V _pos, similarly, the d dimensional feature column vector negative sample of m in negative example base obtained be combined into eigenmatrix V _neg,

S202: align sample characteristics matrix and carry out orthogonal Non-negative Matrix Factorization:

Orthogonal Non-negative Matrix Factorization (Nonnegative Matrix Factorization NMF) is a kind of dimension reduction method of Non-negative Matrix Factorization, a Non-negative Matrix Factorization is the product form of the nonnegative matrix (basis matrix W and matrix of coefficients H) of two low-ranks by it, thus obtains the potential feature of data.Only comprise non-negative element owing to decomposing in the matrix of front and back, primitive character matrix V can be interpreted as the weighted sum of all essential local features in basis matrix W, and weight coefficient is matrix of coefficients H.This resolution characteristic based on local feature can solve partial occlusion because the factors such as camera visual angle cause and complex background to a certain extent on detecting the impact caused, and makes characteristic present more effective.Simultaneously, list of references Mauthner T of the present invention, Kluckner S, Roth P M, etal.Efficient object detection using orthogonal NMF descriptor hierarchies [M] .PatternRecognition.Springer Berlin Heidelberg, 2010:212-221. adds orthogonality constraint W in the condition of Non-negative Matrix Factorization ^tw=I, I representation unit matrix, compared with general Algorithms of Non-Negative Matrix Factorization, iteration speed is faster, and after iteration, the redundance of basis matrix is lower, and feature representation ability is stronger.

The present invention, needs the eigenmatrix V aligning sample _poscarry out orthogonal Non-negative Matrix Factorization.Fig. 3 is the concrete implementing procedure figure of orthogonal Non-negative Matrix Factorization.As shown in Figure 3, the concrete steps of the orthogonal Non-negative Matrix Factorization of the present invention's employing comprise:

S301: make hierarchical alterative sequence number k=0;

S302: the initial basis matrix of stochastic generation and matrix of coefficients:

Initialization base space dimension angle value r ₀, in the present embodiment, make r ₀=2, stochastic generation non-negative basis matrix W ₀with a no negative coefficient matrix H ₀, wherein W ₀for d × r ₀dimension matrix, H is r ₀× n ties up matrix.

S303: iteration obtains basis matrix:

By non-negative basis matrix W _kwith no negative coefficient matrix H _kiteration is carried out by following formula:

$H_k^{t+1}=H_k^t*\frac{\left[{\left(W_k^t\right)}^T{V}_{\mathrm{pos}}\right]}{\left[{\left(W_k^t\right)}^T{W}_k^t{H}_k^t\right]}---\left(1\right)$

$W_k^{t+1}=W_k^t*\frac{\left[V{\left(H_k^t\right)}^T\right]}{\left[W_k^t{H}_k^t{V_{\mathrm{pos}}}^T{W}_k^t\right]}---\left(\text{2}\right)$

Wherein, subscript T representative carries out transposition to matrix, and t represents iterations.

Often complete an iteration namely according to error formula || V-WH|| ²calculate first-order error, || || ²represent two norms, iteration is until error formula always || V-WH|| ²reach minimum value, namely when the error that the t time iteration result obtains || V-WH|| ² _tbe less than the error that the t+1 time iteration result obtains || V-WH|| ² _t+1, then the basis matrix in the t time iteration result is required basis matrix W _k.

S304: make base Spatial Dimension r _k+1=2r _k.

S305: judge whether r _k+1> R, R are default base Spatial Dimension threshold values, if so, and W _kbe required orthogonal basis matrix W _f, algorithm terminates, otherwise enters step S305.

S306: generate new basis matrix:

The basis matrix obtained by step S303 is combined into new basis matrix, even W _k+1=[W _k, W _k], stochastic generation r _k+1the no negative coefficient matrix H of × n _k+1, make k=k+1, return step S303.

The orthogonal basis matrix W that orthogonal Non-negative Matrix Factorization gained is final _f, corresponding matrix of coefficients H _kbe final no negative coefficient matrix H _f.Due to V _pos≈ W _fh _f, orthogonal basis matrix W _feach row represent former eigenmatrix V _posin make reconstructed image error minimum certain local feature, matrix of coefficients H _fevery a line represent according to Ji Tezheng W _freconstruct former eigenmatrix V _posthe coefficient of character pair.Owing to the local feature of last layer being carried out Breaking Recurrently as the lower one deck of initial value input in orthogonal Algorithms of Non-Negative Matrix Factorization, local feature by last layer decomposes further, final local feature is made to have more representativeness, can the global feature of better token image.Due to W _fmatrix is orthogonal nonnegative matrix, W _f ^tw _f=I, so for the characteristic series vector V detecting sample _d, by formula H _d=W _f ^tv _dnamely the new feature matrix of former Projection Character to base space can be obtained.Now, the characteristic series vector V detected is needed _dby representative local feature matrix W _fthrough matrix of coefficients H _dweighting forms, and this characteristic present mode reduces the dimension of feature, simultaneously after signature analysis, to the statement of detection clarification of objective by characteristic series vector V _dbe converted to local feature W _fcoefficient value H _d, this process makes proper vector V _din certain more essential feature become more clear.So just can solving partial occlusion because the factors such as camera visual angle cause and complex background to a certain extent to detecting the impact caused, making characteristic present more effectively, improve the Detection results of sorter.

S203: adopt LMNN algorithm to obtain Projection Character matrix:

Large-spacing arest neighbors sorting technique is statistical a kind of machine learning algorithm having supervision, and this kind of algorithm makes similar data point concentrate as far as possible, inhomogeneous data point as far as possible away from.Different from other distance measure learning methods, LMNN (Large margin nearest neighbor, large-spacing nearest neighbor distance is estimated) algorithm do not carry out the Optimization Learning of distance measure for overall sample, and be only optimized with invasion neighbour for K target neighbor of sample, make in new sample space, k nearest neighbor for certain input amendment all belongs to same classification, and different classes of sample then and the large interval of its maintenance one.This characteristic is conducive to setting up multi-modal discriminant apparent model, sorter is made to have higher identification, specific algorithm can list of references Weinberger K Q, Blitzer J, Saul L K.Distance metric learning for largemargin nearest neighbor classification [C] .Advances in neural information processingsystems.2005:1473-1480.The present invention applies LMNN and carries out distance measure study on the basis of orthogonal Non-negative Matrix Factorization, can reduce the error rate of SVM classifier further, improves classifying quality.

The concrete grammar of large-spacing nearest neighbor distance Measure Algorithm used in the present invention is the orthogonal basis matrix W obtained through orthogonal Non-negative Matrix Factorization by step S202 _fas the initial value of Projection Character matrix L, i.e. L ₀=W _f.Stochastic choice target sample, the HOG characteristic series vector of note target sample is x _i, class label information is c _i, namely class label information is positive sample or negative sample for identifying target sample.Definition x _lfor the invasion neighbour of target sample, namely with K sample of target sample arest neighbors, class label c _l≠ c _isample.Definition x _jfor the target neighbor of target sample, namely with K sample of target sample arest neighbors, class label c _j=c _isample.If distance function is Euclidean distance function D _l(x, y)=|| L (x-y) || ², then can define penalty:

${\mathrm{\ϵ}}_{\mathrm{push}}=\underset{i,j}{\mathrm{\Σ}}\underset{l}{\mathrm{\Σ}}[1+D_L(x_i,x_j)-D_L(x_i,x_l){]}_{+}---\left(3\right)$

${\mathrm{\ϵ}}_{\mathrm{pull}}=\underset{i,j}{\mathrm{\Σ}}{D}_L(x_i,x_j)---\left(4\right)$

Wherein, [Z] ₊=max (Z, 0), D _l(x _i, x _j) represent the Euclidean distance metric function of target neighbor and target sample, D _l(x _i, x _l) represent the Euclidean distance metric function of invading neighbour and target sample.Formula (3) only punishes x _iinvasion neighbour, optimization aim makes target sample x _ito its target neighbor x _jdistance and it is to invading neighbour x _ldistance at least keep 1 large-spacing; Formula (4) only punishes x _itarget neighbor, optimization aim makes target sample x _ito its target neighbor x _jdistance minimization.In conjunction with two formula, final loss function can be obtained:

ε(L)＝(1-μ)ε _push(L)+με _pull(L) (5)

Wherein, μ is the weights of two penalty, is generally taken as 0.5.

Above-mentioned formula is converted to convex programming problem solve, introduces slack variable ξ, construct following semi definite programming problem:

$\min (1-\mathrm{\μ})\underset{i,j}{\mathrm{\Σ}}{(x_i-x_j)}^{T}M(x_i-x_j)+\mathrm{\μ}\underset{i,j}{\mathrm{\Σ}}\underset{l}{\mathrm{\Σ}}\mathrm{\ξ}$

$\begin{array}{cc}s.t.& {(x_i-x_l)}^{T}M(x_i-x_l)-{(x_i-x_j)}^{T}M(x_i-x_j)\≥1-\mathrm{\ξ}---\left(6\right)\end{array}$

ξ＞0

M≥0

Wherein, distance matrix metric M=L ^tl, the implication of all the other symbols is identical with formula (3) (4).This optimization problem can pass through subgradient descent method (Sub-Gradient Descent) and solve.Finally obtain distance matrix metric L.

S204: sorter is trained:

Obtain Projection Character matrix L according to step S203, by following formula, the HOG eigenmatrix of positive negative sample is mapped to new feature space, obtain the eigenmatrix V after mapping _pos', V _neg':

[V _pos',V _neg']＝L*[V _pos,V _neg] (7)

By eigenmatrix V _pos' and V _neg' in each characteristic series vector input in SVM (Support Vector Machine, support vector machine) sorter with corresponding class label information and train, obtain final SVM classifier.

S205: human body target detects:

Adopt the sorter that obtains of step S204, namely availability moment matrix L and SVM classifier can calculate decision function value, treat detected image and carry out human body target detection.Fig. 4 is the embodiment process flow diagram that human body target detects.As shown in Figure 4, the concrete steps of human body target detection are as follows:

S401: make zoom scale sequence number g=1.

S402: adjustment image, calculates the decision value of search box:

Image to be detected is adjusted according to g zoom scale, with the region of search of image after the adjustment of the search box of default size traversal, extracts the HOG characteristic series vector V of image in search box _o, calculate corresponding decision value f.Computing formula is:

f＝sv_coef*SVs*V _o' (8)

V _o'＝(L*V _o) (9)

Wherein, sv_coef is support vector matrix of coefficients, and SVs is support vector matrix, and these two coefficients obtain when SVM classifier is trained.

Preserve the positional information of search box and the decision value of correspondence.

S403: judge that whether g=G, G represent default zoom scale quantity to zoom scale sequence number, if so, enter step S405, otherwise enter step S404.

S404: make g=g+1, returns step S402.

S405: judge one by one all search boxes, if its decision value f is greater than predetermined threshold value F, then contains human body target in this search box, search box positional information is stored as Search Results.

S406: the search box in the Search Results obtain step S405 merges, and obtains final testing result.In the present embodiment, when two search box overlapping areas are greater than predetermined threshold value, to permeate frame by two search boxes.

In order to verify beneficial effect of the present invention, wherein two sequences of common data sets CAVIAR carry out proof of algorithm test.In order to intuitive and convenient carry out the comparison of algorithm performance, the test index of employing is FPPI-MR figure.Wherein, FPPI (False Positive Per Image) represents the quantity of false target in each frame picture, and MR (Miss Rate) to represent in each frame picture correctly but the target be not detected accounts for the quantity of general objective.

Test identical with general testing process, first test pattern is traveled through, search box is extracted according to certain step-length and zoom scale, then the HOG feature of search box is extracted, according to formula (8) by original HOG Feature Mapping to new feature space, and input SVM classifier and carry out discriminant detection, stored search frame position, yardstick and corresponding decision function value, finally merge obtain Search Results according to decision function threshold value screening search box.And calculate FPPI and MR under specific decision function threshold value with this, according to the difference of decision function threshold value, FPPI and MR also changes thereupon.Under same FPPI, MR value is lower, and sorter effect is better.

Have employed Lung biopsy herein to compare, the first Omega human body target detection method based on NMF and LMNN proposed for the present invention, represent with NMF+LMNN, the second is the human body target detection method based on HOG primitive character, represent with HOG, the third is the human body target detection method based on NMF, represent with NMF, 4th kind is Based PC A (Principal components analysis, principal component analysis (PCA)) human body target detection method, represent with PCA, 5th kind is the human body target detection method of Based PC A and LMNN, represent with PCA+LMNN.The sorter adopted is SVM linear classifier.Fig. 5 is the FPPI-MR comparison diagram of the test result of Lung biopsy.As shown in Figure 5, method proposed by the invention is under the condition of identical FPPI, and accuracy rate is all better than all the other four kinds of methods.Fig. 6 is partial detection exemplary plot of the present invention.As shown in Figure 6, the present invention is adopted accurately can to detect human body target.As can be seen from Fig. 5 and Fig. 6, method proposed by the invention can effectively improve sorter accuracy, realizes stable, human body target detection efficiently.

Although be described the illustrative embodiment of the present invention above; so that those skilled in the art understand the present invention; but should be clear; the invention is not restricted to the scope of embodiment; to those skilled in the art; as long as various change to limit and in the spirit and scope of the present invention determined, these changes are apparent, and all innovation and creation utilizing the present invention to conceive are all at the row of protection in appended claim.

Claims (3)

1., based on a human body target detection method for Omega shape facility, it is characterized in that, comprise the following steps:

S1: select the image including upper half of human body head-shoulder Omega shape as positive sample in advance, selects the Background not containing human body target as negative sample, extracts HOG feature, obtain d dimensional feature column vector to the gray-scale map of each sample; By the d dimensional feature column vector that n in positive Sample Storehouse positive sample obtains be combined into eigenmatrix V _pos, the d dimensional feature column vector that the negative sample of m in negative example base is obtained be combined into eigenmatrix V _pos, ${\left[V_{\mathrm{pos}}\right]}_{d\×m}=[V_1^{-}...V_m^{-}];$

S2: to eigenmatrix V _poscarry out orthogonal Non-negative Matrix Factorization, obtain orthogonal basis matrix W _f;

S3: all sample characteristics column vectors obtained according to step S1, adopts LMNN algorithm to obtain Projection Character matrix L, by orthogonal basis matrix W in LMNN Algorithm for Solving process _fas the initial value of Projection Character matrix L;

S4: the Projection Character matrix L obtained according to step S3, maps the eigenmatrix of positive and negative samples, obtains new eigenmatrix V _pos', V _neg', mapping equation is:

[V _pos',V _neg']＝L*[V _pos,V _neg]

By eigenmatrix V _pos' and V _neg' in each characteristic series vector input in SVM classifier with corresponding class label information and train, obtain SVM classifier;

S5: the SVM classifier adopting step S4 to obtain is treated detected image and carried out human body target detection, concrete grammar is: adjusted according to each zoom scale preset by image to be detected, travel through the region of search of the rear image of each adjustment with the search box of default size, extract the HOG characteristic series vector V of image in search box _o, calculate corresponding decision value f; Computing formula is:

f＝sv_coef*SVs*V _o'

V _o'＝(L*V _o)

All search boxes are judged one by one, if its decision value f is greater than predetermined threshold value F, then contains human body target in this search box, search box positional information is stored as Search Results; Finally the search box in Search Results is merged, obtain final testing result.

2. human body target detection method according to claim 1, is characterized in that, in described step S2, orthogonal Non-negative Matrix Factorization comprises the following steps:

S2.1: make hierarchical alterative sequence number k=0; Initialization base space dimension angle value r ₀, stochastic generation non-negative basis matrix W ₀with a no negative coefficient matrix H ₀, wherein W ₀for d × r ₀dimension matrix, H is r ₀× n ties up matrix;

S2.2: by non-negative basis matrix W _kwith no negative coefficient matrix H _kiteration is carried out by following formula:

$H_k^{t+1}=H_k^t*\frac{\left[{\left(W_k^t\right)}^T{V}_{\mathrm{pos}}\right]}{\left[{\left(W_k^t\right)}^T{W}_k^t{H}_k^t\right]}$

$W_k^{t+1}=W_k^t*\frac{\left[V{\left(H_k^t\right)}^T\right]}{\left[W_k^t{H}_k^t{V_{\mathrm{pos}}}^T{W}_k^t\right]}$

Wherein, subscript T representative carries out transposition to matrix, and t represents iterations; Often complete an iteration according to error formula || V-WH|| ²calculate first-order error, || || ²represent two norms, when || V-WH|| ²reach minimum value, the basis matrix of its correspondence is required basis matrix W _k;

S2.3: make base Spatial Dimension r _k+1=2r _kif, r _k+1> R, R are default base Spatial Dimension threshold values, W _kbe required orthogonal basis matrix W _f, otherwise enter step S2.4;

S2.4: make W _k+1=[W _k, W _k], random generation r _k+1the no negative coefficient matrix H of × n _k+1, make k=k+1, return step S2.2.

3. human body target detection method according to claim 1, is characterized in that, in described step S3, the concrete grammar of LMNN algorithm is: solve following semi definite programming problem:

$\min (1-\mathrm{\μ})\underset{i,j}{\mathrm{\Σ}}{(x_i-x_j)}^{T}M(x_i-x_j)+\mathrm{\μ}\underset{i,j}{\mathrm{\Σ}}\underset{l}{\mathrm{\Σ}}\mathrm{\ξ}$

s.t. (x _i-x _l) ^TM(x _i-x _l)-(x _i-x _j) ^TM(x _i-x _j)≥1-ξ

ξ＞0

M≥0

Wherein, distance matrix metric M=L ^tl, x _ifor the HOG characteristic series vector of target sample, x _jfor the target neighbor of target sample, x _lrepresent the invasion neighbour of target sample, μ is default weights, and ξ is slack variable.