CN102426653A - Static human body detection method based on second generation Bandelet transformation and star type model - Google Patents

Abstract

The invention provides a human body detection method based on a second generation Bandelet transformation and a star type model. The current method can not process an occlusion problem well. The method of the invention can solve the above problem. The method comprises the following steps: (1) acquiring a large amount of negative samples through an INRIN database, forming a training sample set FB of a whole human body with other positive samples in the database, cutting the FB so as to obtain the training sample sets of each part; (2) using the Bandelet transformation, calculating a Bandelet combination characteristic of the each training sample set and forming six training sample characteristic sets; (3) using an Adaboost algorithm to perform classification training to the sample characteristic sets so as to obtain a corresponding part classifier and acquiring the human body star type model through learning the each part; (4) calculating a Bandelet coefficient matrix of an image to be detected, carrying out human body position detection through the star type model, using a main window merging method to combine scanning windows of all human bodies and parts so as to obtain a final human body detection result. By using the method of the invention, the human body can be accurately detected. The method can be used in a video processing field, such as video monitoring, object identification and the like.

Description

Static human detection method based on second generation Bandelet conversion and star-like model

Technical field

The invention belongs to technical field of image processing, relate to static human detection method, can be used for intelligent monitoring, driver assistance system, human body motion capture, porny filtration, virtual video etc.

Background technology

Human detection in the computer vision is the very wide technology of an application prospect; Human detection all has application promise in clinical practice in a plurality of fields in fact; But because the diversity of human body attitude; Noisy and the clothes texture of background, illumination condition many-sided factor such as self blocks and causes human detection to become a very problem of difficulty.

At present, in the still image method of human detection mainly contain method based on manikin, based on the method for template matches with based on the method for statistical classification.Wherein:

Based on the method for manikin, clear and definite manikin arranged, carry out human body identification according to each position and the relation between the human body of model construction then.This method can be handled occlusion issue, and can infer the attitude of human body.But the deficiency of this method is that the structure of model is difficult, finds the solution also more complicated.

Based on the method for template matches, to template of human body target structure, carried out the identification of human body then according to template matching algorithm before this.This method is calculated fairly simplely with comparing based on the method for model, and shortcoming is because the complicacy of human body attitude is difficult to construct enough templates to handle different attitudes.

Based on the method for statistical classification, from a series of training data middle school acquistion to a sorter, represent human body through machine learning with this sorter, utilize this sorter input window is classified and to discern then.Comparing based on the method for the method of the advantage of the method for statistical classification and manikin and template matches is the comparison robust, and testing result is more excellent, but its shortcoming is to need a lot of training datas, and is difficult to solve attitude and the problem of blocking.

Summary of the invention

The present invention seeks to deficiency to above-mentioned prior art; A kind of human body detecting method that conversion combines with hub-and-spoke configuration based on second generation Bandelet that proposes; From the geometry flow characteristic of image and the geometry of human body; Effectively handle occlusion issue, reduce the empty scape rate of human detection, improve correct rate of human body detection.

For realizing above-mentioned purpose, technical scheme of the present invention comprises as follows:

(1) from the INRIA database, obtain a large amount of negative samples, and other positive sample constitutes whole human body training sample set FB in database, and FB is manually cut through the bootstrapping operation; Obtain human body head, left side shoulder, right shoulder; Lower limb, the training sample set of five human bodies of step;

(2) extract the Bandelet coefficient of FB and five position training sample sets and the statistical value of this coefficient; It is united characteristic as Bandelet; Obtain the respective sample feature set; And utilize the Adaboost algorithm that these sample characteristics collection are carried out classification based training, obtain whole human body sorter and five position sorters;

(3) study based on star-like model is carried out at each position of human body, obtained the star-like model of whole human body, the output valve of the whole human body sorter that obtains in the step (2) and five position sorters input value as corresponding site wave filter in the star-like model of human body;

(4) import the image to be checked of size arbitrarily, utilize the Bandelet matrix of coefficients of Bandelet transformation calculations image to be checked;

(5) according to the Bandelet matrix of coefficients of image to be checked; Calculate the Bandelet associating characteristic of all scanning windows in the image to be checked; Gained whole human body sorter and five position sorters in the input step (2); From left to right 9 * 9 neighborhood scanning is carried out in human body candidate region in the preliminary judgement image to be checked in the human body candidate region;

(6) based on the manikin of gained in the step (3); Calculate in the image to be checked each scanning window in the output valve of star-like model median filter; And whether there is human body in the output valve judgement scanning window based on wave filter; There are human body and human body if filter output value is thought greater than zero in this scanning window, otherwise then think and do not have human body and human body;

(7) utilizing the main window act of union to carry out window to the window that has human body and human body merges, obtains final human detection result.

The present invention has the following advantages compared with prior art:

1, the present invention is the human body detecting method that combines Bandelet characteristic and star-like model, can get the present invention through experimental result and effectively improve correct rate of human body detection, has reduced the empty scape rate of human detection.

2, the present invention utilizes star-like model to represent human body, has made full use of the geometry of human body, effectively the into treatment sites occlusion issue.

3, the present invention utilizes Bandelet geometry flow characteristic; Unite the geometrical property that characteristic is portrayed human body image through Bandelet; Combining with star-like model makes each position of human body to be come out by abundant sign again; Compare more with the HOG method and can embody human geometry's architectural characteristic, obtain more characteristics of human body's quantity of information.

Description of drawings

Fig. 1 is a schematic flow sheet of the present invention;

Fig. 2 is the whole human body used among the present invention and the positive sample image at each position;

Fig. 3 is the whole human body used among the present invention and the negative sample image at each position;

Fig. 4 is the performance comparison diagram at each position;

Fig. 5 is the performance comparison diagram of the present invention and HOG method;

Fig. 6 is that the present invention is used for the figure as a result that human body image detects.

Embodiment

The present invention is from the human detection based on the position; A kind of human body detecting method based on second generation Bandelet conversion and star-like model has been proposed; Utilize Bandelet coefficient and statistical nature thereof in the Bandelet conversion as the final human characteristic; Utilize study and the matching detection of the star-like model human body in the picture of publishing picture then, and experimental result and HOG experimental result are compared.

With reference to Fig. 1, concrete performing step of the present invention is following:

Step 1 is obtained the training sample set of whole human body training sample set and each human body.

Test required training sample set in order to obtain the present invention; Be the basis with the INRIA database; Operate through the bootstrapping of negative sample and to obtain a large amount of negative samples; And other positive sample constitutes the training sample set of whole human body in database, utilizes the method for manually cutting to obtain the training sample set of each human body again, and its concrete performing step is following:

(1.1) required training sample set is from INRIA somatic data storehouse, and download address does Http:// pascal.inrialpes.fr/data/human/, because this database does not provide enough negative samples,, to operate through the bootstrapping of negative sample and obtain a large amount of negative samples so need be the basis with this database, its detailed process is following:

(1.1a) first from the INRIA database, appointing, got positive sample of sub-fraction and negative sample, carries out feature extraction, training preliminary classification device;

(1.1b) use this preliminary classification device; Detect all the other the non-human body images in the database; A random choose part is formed new negative sample collection with current negative sample from the scanning window image that mistake is divided, and uses random choose can avoid sneaking into the sample image of a large amount of feature similarities;

(1.1c) repeated characteristic extraction, training classifier, the non-human body image of detection and form new this process of negative sample collection until collect with the INRIA database in the positive identical final negative sample of number of samples;

(1.2) in order to make training sample have more band table property; Use respectively based on gradient orientation histogram HOG characteristic and two kinds of characteristics of Bandelet associating characteristic and carry out feature extraction; Carry out the bootstrapping operation of negative sample like this, can obtain two negative sample collection, get it then and occur simultaneously as final negative sample collection; At last add the positive sample set in the INRIA database, obtain whole human body sample set FB by this negative sample collection;

(1.3) all positive negative samples among the whole human body sample set FB are carried out the sample image position cutting of head and foot's 40 * 40 pixels; The sample image position cutting of left side shoulder and right shoulder 32 * 48 pixels, the sample image position cutting of lower limb 40 * 32 pixels, deletion departs from the bigger cutting result of human body and can obtain head sample set H; Left side shoulder sample set LS; Right shoulder sample set RS, lower limb sample set L, five position sample sets of the sample set F of foot human body;

(1.4) five position sample sets with whole human body sample set and human body constitute final training sample set; Wherein the whole human body sample set has 2416 positive samples and 10000 negative samples as training sample set; Size is 128 * 64 pixels, and five position sample sets of human body respectively have 2074 positive samples and 10000 negative samples, and head and foot's sample size are 40 * 40 pixels; Left side shoulder is 48 * 32 pixels with right shoulder, and lower limb are 32 * 40 pixels.Fig. 2 has provided part sample image wherein, and wherein Fig. 2 (a) is the positive sample of part, and Fig. 2 (b) is the part negative sample.

Step 2 is utilized the Bandelet conversion, calculates the Bandelet associating characteristic of all training samples and forms six training sample feature sets, and it comprises five position sample characteristics collection of a whole human body sample characteristics collection and human body, and its concrete performing step is following:

(2.1) each sample image is done following two-dimensional discrete orthogonal wavelet transformation:

(2.1a) every row of sample image is done one-dimensional wavelet transform:

$X_i=(X_{i1}+X_{i2})/\sqrt{2}$ i＝1，2，K?nx/2?i1＝1，3，5K i2＝2，4，6K

$\mathrm{Xu}=(X_{u1}-X_{u2})/\sqrt{2}$ u＝nx/2，nx/2+1，K?nx?u1＝1，3，5K u2＝2，4，6K

Wherein, X _i, Xu, X _I1, X _I2, X _U1, X _U2Represent the last i of each row respectively, u, i1, i2, u1, u2 gray values of pixel points, the line number of nx presentation video gray matrix;

(2.1b) every row of sample image are done one-dimensional wavelet transform:

$X_j=(X_{\mathrm{<figure-callout\; id="1"\; label="j"\; filenames="HDA0000103279350000041.png,HDA0000103279350000042.png"\; state="\{\{state\}\}">j</figure-callout>}1}+X_{j2})/\sqrt{2}$ j＝1，2，K?ny/2?j1＝1，3，5K?j2＝2，4，6K

$\mathrm{Xv}=(X_{v1}-X_{v2})/\sqrt{2}$ v＝ny/2，ny/2+1，K?ny?v1＝1，3，5K?v2＝2，4，6K

Wherein, X _j, Xv, X _J1, X _J2, X _V1, X _V2Represent that respectively each lists j, v, j1; J2, v1, v2 gray values of pixel points; The columns of ny presentation video gray matrix can get sample image two-dimensional discrete orthogonal wavelet transformation coefficient, and the present invention draws through the contrast experiment; The number of plies of doing two-dimensional wavelet transformation is 1 o'clock, and the sign performance of characteristic is best.

(2.2) combine quaternary tree to cut apart the optimal quadtree decomposition that optimized Algorithm is set up multiple dimensioned each subband of figure the sample image behind the two-dimensional discrete orthogonal wavelet transformation, obtain the Bandelet piece of L * L, L representes the width of Bandelet piece.Described quaternary tree is cut apart optimized Algorithm can be referring to document G.Peyr é and S.Mallat; " Surface compression with geometric bandelets ", ACM Transactions on Graphics (TOG) Vol24-3, ACM New York; NY; USA, July 2005, pp.601-608;

(2.3) each Bandelet piece is asked for a best geometry flow side according to the Lagrange penalty function method

(2.3a) for the sub-piece S that is of a size of L * L, [0, π) equal angles is discrete is L with angle of circumference ²-1 sampling angle α;

$\mathrm{\&alpha;}=\frac{\mathrm{r\&pi;}}{L^2-1},r=\mathrm{0,1,2}\mathrm{\&Lambda;}{L}^2-1$

For the situation of no geometry flow, mark sampling angle α=inf, inf representes infinity, Bandeletization is not implemented in expression, so the total L of the value of sampling angle α ²Individual;

(2.3b) for each sampling angle α, α ≠ inf constructs the big or small net point of a L * L; Calculate the net point of each coordinate, h for [x (h), y (l)]; L be each pixel respectively in the position at x axle and y axle place, the rectangular projection error t of grid points on sampling angle α:

t＝-sin(α)·x(h)+cos(α)·y(l)

Net point is sorted by error amount t from small to large, obtain a L ²* 1 ranking index reorders the coefficient behind the two-dimensional quadrature wavelet transform of this sub-block S then by ranking index, obtain an one-dimensional signal f _s

(2.3c) to one-dimensional signal f _sCarry out one-dimensional wavelet transform and obtain one-dimensional signal f _α, calculate the quantized value of this signal simultaneously

Quantitative formula is:

${\tilde{f}}_{\mathrm{\&alpha;}}=\left\{\begin{array}{cc}0& \left|f_{\mathrm{\&alpha;}}\right|\&le;T\\ \mathrm{sign}\left(f_{\mathrm{\&alpha;}}\right)\&CenterDot;(q+0.5)\&CenterDot;T& \mathrm{qT}\&le;\left|f_{\mathrm{\&alpha;}}\right|\&le;(q+1)T\end{array}\right.$

Wherein, T is a quantization threshold, and q is an integer, and sign is a sign function;

(2.3d) according to f _αWith

Value calculate the best geometry flow direction in the sub-piece S, this direction makes the Lagrange function exactly:

Minimum geometry flow direction, wherein, R _gThe required bit number of presentation code geometry flow, R _bThe required bit number of Bandelet coefficient after presentation code quantizes, λ is that the Lagrange multiplier gets 3/28;

(2.3e) get the corresponding one-dimensional wavelet transform coefficient f of minimum Lagrange functional value _θ, and by the Mallet rule this coefficient is reconfigured and to be two dimensional form, as the Bandelet in the sub-piece S be;

(2.4) extract the Bandelet coefficient of sample image and the statistic of Bandelet coefficient by every sub-block S, the statistic that the present invention preferentially chooses comprises as follows: energy:

Entropy:

Average:

And maximal value: max (c _{M, n}), wherein, c _{M, n}For (m n) is Bandelet coefficient on the coordinate, and N is the number of contained element among each sub-block S among each sub-block S.

With the statistic of Bandelet coefficient and Bandelet coefficient as final characteristics of image; Five position sample sets to whole human body sample set and human body extract characteristics of image, obtain five position training sample feature sets of corresponding whole human body training sample feature set and human body.

Step 3 is utilized the AdaBoost algorithm, each training sample feature set is carried out classification based training obtain corresponding AdaBoost sorter, and its concrete steps are following:

(3.1) five position training sample feature sets of input whole human body training sample feature set and human body can be set up corresponding a plurality of Weak Classifier;

(3.2) through study to Weak Classifier; Obtain the weighted value of each sample characteristics; Make the weighted value of each sample characteristics represent the size that this sample characteristics is divided by mistake, thereby constructed five position sorters of corresponding whole human body sorter and human body by the weighted value of this sample characteristics of size renewal of mistake branch according to this sample characteristics.

Step 4 is carried out the study based on star-like model to each position of human body, obtains the star-like model of whole human body, and its concrete steps are following:

(4.1) through connecting parameters C _kSet up each position of human body and connect with the elasticity of corresponding human body root interdigit, tentatively confirm the star-like model of whole human body, this model meets Gaussian distribution, is expressed as:

P(l _k，l ₀/C _k)＝N(l _k-l ₀，V _k，∑ _k)

In the formula, N representes Gaussian distribution, l _kExpression position, k position of human body, k=1 ..., 5, represent head successively, left shoulder, right shoulder, lower limb portion, foot position; l ₀Expression position, human body root position, this root position is meant the center of whole human body; V _kBeing a two-dimensional vector, is the ideal position of k position with respect to its root position, with the output valve of five position sorters of the human body input value as corresponding human body region wave filter in the star-like model of whole human body, V _kConfirm by k the maximum output valve of position wave filter; Connect parameters C _k=(V _k, ∑ _k), ∑ _kBe k position with corresponding root position between the full covariance matrix of elasticity coupling stiffness;

(4.2) to full covariance matrix ∑ _kCarry out svd, be expressed as:

In the formula, U _kIt is ∑ _kSingular vector, U _k=∑ _k∑ _k ^T,

Be T another one singular vector,

D _kBe a diagonal matrix, represent ∑ _kSingular value, the transposition of subscript T representing matrix; By ∑ _kSvd obtain the transform T of k human body and this human body root position _K0(l _k) and the transform T of this human body root position itself _0k(l ₀):

$T_{k0}\left(l_k\right)=U_k^T(l_k-V_k),$ $T_{0k}\left(l_0\right)=U_k^T\left(l_0\right);$

The star-like model that obtains whole human body according to above-mentioned transform is:

P(l _k，l ₀/C _k)＝N[T _k0(l _k)-T _0k(l ₀)，0，D _k]

In the formula, N representes Gaussian distribution, T _K0(l _k)-T _0k(l ₀) represent the stochastic variable of this Gaussian distribution, the average of 0 this Gaussian distribution of expression, D _kThe covariance of representing this Gaussian distribution;

(4.3) with k position of human body with respect to its ideal position V _kThe deformation cost

Be expressed as:

$d_{k,l_k}={[T_{k0}\left(l_k\right)-T_{0k}\left(l_0\right)]}^T{M}_k^{-1}[T_{k0}\left(l_k\right)-T_{0k}\left(l_0\right)]$ M in the formula _k=2 * D _k, the transposition of subscript T representing matrix;

Can the output valve

of everyone body region wave filter be expressed as thus:

$R_{k,l_k}=\underset{l_k}{\max }(f_{k,l_k}-d_{k,l_k}+b)$

In the formula; expression k position is in the input value of lk position filtering device, and b is the side-play amount of k position with respect to its root position;

(4.4) obtain whole human body wave filter F0 output valve according to each human body filter output value by following formula:

$\mathrm{Score}\left(F_0\right)={\mathrm{<figure-callout\; id="0"\; label="f"\; filenames="HDA0000103279350000042.png"\; state="\{\{state\}\}">f</figure-callout>}}_0+\underset{k=1}{\overset{5}{\mathrm{\&Sigma;}}}{R}_{k,l_k}({\mathrm{<figure-callout\; id="0"\; label="l"\; filenames="HDA0000103279350000042.png"\; state="\{\{state\}\}">l</figure-callout>}}_0+V_k),$

In the formula, f ₀Be the input value of whole human body wave filter, l ₀Expression position, human body root position, V _kBe the ideal position of k position, with whole human body wave filter F with respect to its root position ₀Output valve as the final output valve of the star-like model of whole human body.

Step 5, input is the image to be checked of size arbitrarily, calculates the characteristics of image that its Bandelet matrix of coefficients obtains image to be checked, and its concrete steps are following:

(5.1) any big or small image to be checked of input, with its convergent-divergent according to a certain percentage, pantograph ratio is got [0.5,0.6,0.7,0.8,0.9], clips the right less than 8 pixels and following redundance less than 16 pixels by following formula then:

X＝|X/8|×8 Y＝|Y/16|×16

Wherein, X is an image to be checked pixel value in the horizontal direction, and Y is the pixel value of image to be checked in vertical direction, || the expression truncation rounds;

Image to be checked is carried out the two-dimensional discrete orthogonal wavelet transformation, and the number of plies of wavelet transformation is 1;

(5.2) image behind the two-dimensional discrete orthogonal wavelet transformation is carried out two of 4 * 4 pixels and advances subdivision, and with the fritter of each 4 * 4 pixel as a Bandelet piece;

(5.3) calculate the projection error of each Bandelet piece on all directions, and, ask for the projection error ranking index of an optimum by minimum Lagrange function method;

(5.4) according to the optimal sequencing index of each Bandelet piece, the two-dimensional discrete wavelet conversion coefficient on it is reordered, obtain corresponding one-dimensional signal;

(5.5) each one-dimensional signal is obtained the one-dimensional wavelet transform coefficient as one-dimensional wavelet transform; By the Mallet rule this coefficient is reconfigured and to be two dimensional form; Bandelet coefficient as corresponding Bandelet piece; With the Bandelet coefficient of all Bandelet pieces, constitute the Bandelet matrix of coefficients of the image of handling through step (5.1) to be checked, thereby obtain the characteristics of image of image to be checked.

Step 6 with characteristics of image input whole human body sorter and each position sorter of human body of image to be checked, obtains the output valve of whole human body wave filter, and its concrete steps are following:

(6.1) tentatively confirm the candidate region of human body according to the output valve of whole human body sorter, from left to right scanning in 9 * 9 neighborhoods of candidate region then scans the head of human body, left side shoulder, right shoulder, lower limb and foot successively;

(6.2) human body of every scanning calculates the deformation cost of this human body and the output valve of this human body wave filter, finally confirms the output valve of whole human body wave filter.

Step 7 according to the output valve of whole human body wave filter, is utilized the main window act of union, and all scanning windows that are divided into human body are made up, and forms final human detection result, and concrete steps are following:

(7.1) judge by human body candidate region in the altimetric image based on the output window of whole human body wave filter; If the scanning window of wave filter output does not contain the human body window; Do not contain human body in the image then to be checked, otherwise, find out the maximum human body window of whole human body filter output value as the human body main window;

(7.2) in the candidate region, carrying out x direction step-length is 9, and y direction step-length is 9 dense scanning, obtains the output valve of each human body window filter, finds out the maximum human body window of position filter output value as the human body main window;

(7.3) human body main window and human body main window and other human body windows or human body window are made up judgement; When other human body windows are in around the human body main window and overlapping less than 3/5 the time; Be judged to and do not make up; In like manner, when other people body region window is in around the human body main window and overlapping yet being judged to less than 3/5 time do not made up, otherwise make up;

(7.4) with the border average of the human body window of human body main window and human body main window and required combination or human body window as a testing result;

(7.5) the human body window and the human body window of deletion main window and position main window and all participation combinations;

(7.6) if also have remaining window, then find out again sorter mark wherein the highest as main window, and the operation of repeating step (7.2)-(7.5);

(7.7) on human body image to be checked, mark all testing results, by the final detection result of altimetric image, this final detection result can adopt rectangle frame to represent as this, and the human body that is detected is in the rectangle frame.

Effect of the present invention can further be explained through following experimental result and analysis:

1, experiment condition:

Test required positive sample and negative sample and all be taken from the INRIA database; Can obtain 2416 positive samples and 10000 negative samples thus as training set; 1132 positive samples and 4050 negative samples are as test set; Size is 128 * 64 pixels; Cutting obtains each position sample set on the basis of above sample set, and then manually removes some and depart from bigger sample and obtain the position collection at last and all comprise 2074 positive samples and 1000 negative samples as training set, 1132 positive samples in 4050 negative samples as test set.Wherein H head size is 40 * 40 pixels, and LS left side shoulder and the right shoulder of RS sample size are 32 * 48 pixels, and L shank sample size is 40 * 32 pixels, and F foot sample size is 40 * 40 pixels; Fig. 2 has provided wherein part sample image, and wherein Fig. 2 (a) is the positive sample of part whole human body, and 2 (b) are the positive sample of part head, and 2 (c) are the positive sample of part left side shoulder; 2 (d) bear sample for the part right side, and 2 (e) are the positive samples of part lower limb, and 2 (f) are the positive sample of part foot; Fig. 3 (a), 3 (b), 3 (c); 3 (d), 3 (e), 3 (f) are the part negative sample corresponding with positive sample.

Hardware platform is: Intel Core2 Duo CPU E6550 2.33GHZ, 2GB RAM.Software platform is MATLAB R2009a.

2, method of contrast

The method of contrast that the present invention uses is the human body detecting method based on gradient orientation histogram HOG characteristic that N.Dalal and B.Triggs propose in article " N.Dalal and B.Triggs; " Historgram of oriented gradient for human detection "; In CVPR Vol.1, San Diego, California; June 2005, pp.886-893. ".At first extract the HOG characteristic of human body, utilize the svm classifier device that characteristic is carried out classification based training again, import image to be checked, obtain final testing result according to the output valve of sorter.

3, experiment content and interpretation

Experiment 1: the same width of cloth is carried out human detection from the image to be checked of MIT database with the inventive method and method of contrast; Testing result such as Fig. 6, wherein image 6 (a) expression the inventive method is to the result of the Classification and Identification at each position, and image 6 (b) is the final detection result of the inventive method; Fig. 6 (c) adopts the resulting testing result of method of contrast; As can be seen from Figure 6, use method of the present invention to have higher classification accuracy rate, make it have higher human detection accuracy; Be very suitable for the human detection of still image; Method of the present invention can reduce the empty scape rate of detection greatly simultaneously, compares with the HOG method, under the situation that detection has the position to block, can show stronger robustness.

Experiment 2: classification performance is carried out relatively in each position of human body with the inventive method; Its classification performance comparative result such as Fig. 4; As can beappreciated from fig. 4 the classifying quality of head and step is better, and this is that left side shoulder is relatively poor with the classifying quality of right shoulder because the deformation in walking of head and step is minimum; This is bigger because of the amplitude of when walking, getting rid of arm, and the deformation of generation is also bigger.

Experiment 3: the detection with the inventive method and method of contrast can compare, and it detects performance comparison result such as Fig. 5, and as can beappreciated from fig. 5 the detection performance of the inventive method is higher than the detection performance of additive method.

Claims (5)

1. the static human detection method based on second generation Bandelet conversion and star-like model comprises the steps:

(1) from the INRIA database, obtain a large amount of negative samples, and other positive sample constitutes whole human body training sample set FB in database, and FB is manually cut through the bootstrapping operation; Obtain human body head, left side shoulder, right shoulder; Lower limb, the training sample set of five human bodies of step;

(2) extract the Bandelet coefficient of FB and five position training sample sets and the statistical value of this coefficient; It is united characteristic as Bandelet; Obtain the respective sample feature set; And utilize the Adaboost algorithm that these sample characteristics collection are carried out classification based training, obtain whole human body sorter and five position sorters;

(3) study based on star-like model is carried out at each position of human body, obtained the star-like model of whole human body, the output valve of the whole human body sorter that obtains in the step (2) and five position sorters input value as corresponding site wave filter in the star-like model of human body;

(4) import the image to be checked of size arbitrarily, utilize the Bandelet matrix of coefficients of Bandelet transformation calculations image to be checked;

(5) according to the Bandelet matrix of coefficients of image to be checked; Calculate the Bandelet associating characteristic of all scanning windows in the image to be checked; Gained whole human body sorter and five position sorters in the input step (2); From left to right 9 * 9 neighborhood scanning is carried out in human body candidate region in the preliminary judgement image to be checked in the human body candidate region;

(6) based on the manikin of gained in the step (3); Calculate in the image to be checked each scanning window in the output valve of star-like model median filter; And whether there is human body in the output valve judgement scanning window based on wave filter; There are human body and human body if filter output value is thought greater than zero in this scanning window, otherwise then think and do not have human body and human body;

(7) utilizing the main window act of union to carry out window to the window that has human body and human body merges, obtains final human detection result.

2. human body detecting method according to claim 1, wherein step (3) is described carries out the study based on star-like model to each position of human body, obtains the star-like model of whole human body, carries out as follows:

(3a) through connecting parameters C _kSet up each position of human body and connect with the elasticity of corresponding human body root interdigit, the preliminary star-like model of confirming based on human body, this model meets Gaussian distribution, is expressed as:

P(l _k，l ₀/C _k)＝N(l _k-l ₀，V _k，∑ _k)

In the formula, l _kExpression position, k position of human body, k=1 ..., 5, represent head successively, left shoulder, right shoulder, lower limb portion, foot position; l ₀Expression position, human body root position, this root position is meant the center of whole human body; V _kBeing a two-dimensional vector, is the ideal position of k position with respect to its root position, is confirmed by k the maximum output valve of position wave filter; Connect parameters C _k=(V _k, ∑ _k), ∑ _kBe k position with corresponding root position between the full covariance matrix of elasticity coupling stiffness;

(3b) to ∑ _kCarry out svd, be expressed as:

In the formula, U _kIt is ∑ _kSingular vector, U _k=∑ _k∑ _k ^T,

It is ∑ _kThe another one singular vector,

D _kBe a diagonal matrix, represent ∑ _kSingular value, by ∑ _kSvd, obtain the transform T of k position and root position _K0(l _k) and the transform T of root position itself _0k(l ₀):

$T_{k0}\left(l_k\right)=U_k^T(l_k-V_k),$ $T_{0k}\left(l_0\right)=U_k^T\left(l_0\right);$

Obtain based on the star-like model of human body according to above-mentioned transform be:

P(l _k，l ₀/C _k)＝N[T _k0(l _k)-T _0k(l ₀)，0，D _k]

In the formula, N representes Gaussian distribution, T _K0(l _k)-T _0k(l ₀) represent the stochastic variable of this Gaussian distribution, the average of 0 this Gaussian distribution of expression, D _kThe covariance of representing this Gaussian distribution;

(3c) can be with l according to step (3b) _kWith respect to V _kThe deformation cost

Be expressed as:

$d_{k,l_k}={[T_{k0}\left(l_k\right)-T_{0k}\left(l_0\right)]}^T{M}_k^{-1}[T_{k0}\left(l_k\right)-T_{0k}\left(l_0\right)]$ M in the formula _k=2 * D _k

Can the output valve

of everyone body region wave filter be expressed as thus:

$R_{k,l_k}=\underset{l_k}{\max }(f_{k,l_k}-d_{k,l_k}+b)$

In the formula,

Represent that the k position is at l _kThe input value of position filtering device, b is the side-play amount of k position with respect to its root position;

(3d) according to each human body filter output value Obtain whole human body wave filter F by following formula ₀Output valve:

$\mathrm{Score}\left(F_0\right)=f_0+\underset{k=1}{\overset{5}{\mathrm{\&Sigma;}}}{R}_{k,l_k}(l_0+V_k),$

In the formula, f ₀Be the input value of whole human body wave filter, l ₀Expression position, human body root position, V _kBe the ideal position of k position with respect to its root position.

3. human body detecting method according to claim 1, any image to be checked of size of the said input of step (4) wherein utilizes the Bandelet matrix of coefficients of Bandelet transformation calculations image to be checked, carries out as follows:

(4a) any image to be checked of size of input, earlier with its convergent-divergent by a certain percentage, clip the right less than 8 pixels and following redundance by following formula then less than 16 pixels:

X＝|X/8|×8 Y＝|Y/16|×16

Wherein, X is an image to be checked pixel value in the horizontal direction, and Y is the pixel value of image to be checked in vertical direction, || the expression truncation rounds;

(4b) carry out the two-dimensional discrete orthogonal wavelet transformation to clipping the image of handling to be checked, the number of plies of wavelet transformation is 1;

(4c) image to be checked behind the two-dimensional discrete orthogonal wavelet transformation is carried out two of 4 * 4 pixels and advances subdivision, and with the fritter of each 4 * 4 pixel as a Bandelet piece;

(4d) calculate the projection error of each Bandelet piece on all directions, and, ask for the projection error ranking index of an optimum by minimum Lagrange function method;

(4e), the two-dimensional discrete wavelet conversion coefficient on it is reordered, obtain corresponding one-dimensional signal according to the optimal sequencing index of each Bandelet piece;

(4f) each one-dimensional signal is made one-dimensional wavelet transform; And with the one-dimensional wavelet transform coefficient of correspondence; Reconfigure by the Mallet rule and to be two dimensional form; As the Bandelet coefficient of corresponding Bandelet piece, the Bandelet coefficient of all Bandelet pieces is constituted the Bandelet matrix of coefficients of this image to be checked.

4. human body detecting method according to claim 1, wherein the Bandelet of all scanning windows unites characteristic in the said calculating of step (5) image to be checked, carries out as follows:

(5a) with the zone of the sample size in the image to be detected upper left corner as first scanning window, every to 8 pixels of right translation or downwards 16 pixels of translation obtain one group of scanning window as a new scanning window;

(5b) calculate these several types of statistical values of energy, entropy, average and maximal value of Bandelet coefficient on each scanning window, and with these statistical values and Bandelet coefficient together as the Bandelet associating characteristic of corresponding scanning window.

5. human body detecting method according to claim 1, wherein each scanning window carries out in the output valve of star-like model median filter as follows in the described calculating of step (6) image to be checked:

(6a) tentatively confirm the candidate region of human body according to the output valve of each scanning window wave filter, in 9 * 9 neighborhoods of candidate region, from left to right scan then;

(6b) scan the head of human body successively, left side shoulder, right shoulder, lower limb and foot, human body of every scanning calculates the deformation cost of this human body and the output valve of this human body wave filter, finally confirms the output valve of whole human body wave filter.

Images