CN101630369B - Pedestrian detection method based on wavelet fractal characteristic - Google Patents

Abstract

The invention discloses a pedestrian detection method based on wavelet fractal characteristic, which relates to the technology of intelligent vehicles in an intelligent traffic system. The pedestrian detection method comprises the following steps: (1) reading a training sample set and standardizing all sample images to the pixel size of 48*96; (2) carrying three times of two-dimensional wavelet transformation on the sample images, and getting six wavelet sub-graphs in the second layer and the third layer; (3) getting absolute values of the wavelet sub-graphs obtained in the step (2), and carrying out stretching and scaling so that the numeric area is uniformly mapped to 0 to 255; (4) respectively getting fractal dimension vectors of the wavelet sub-graphs treated in the step (3); (5) standardizing the fractal vectors corresponding to each sub-graph obtained in the step (4); (6) combining the fractal vectors standardized in the step (5), and obtaining the wavelet fractal characteristic vector of 6*(n-1) dimension; and (7) using the wavelet fractal characteristic obtained in the step (6) for training a soft-supporter vector machine, and using the obtained discriminant function for realizing pedestrian detection. The invention has the characteristics of concise characteristic expression form, strong characteristic distinguishing capability and higher pedestrian detection efficiency.

Description

A kind of pedestrian detection method based on wavelet fractal characteristic

Affiliated technical field

The present invention relates to the The intelligent vehicles technology in the intelligent transportation system, especially a kind of pedestrian detection method based on wavelet fractal characteristic.

Background technology

Target detection is the important content of machine vision research, and at video monitoring, the mobile robot is widely used in the fields such as intelligent vehicle.Because pedestrian's appearance, the combination of posture and intensity of illumination is ever-changing, cause pedestrian detection become sensation target detect in one of complicated problems the most.

Existing pedestrian detection method can be summed up as two main treatment steps, i.e. feature extraction and target detection on process.The main task of target detection step is the target signature that characteristic extraction step is constructed, and is used for the sorter training, thereby new test sample book is made differentiation.Therefore, feature extracting method is to the effect important influence of pedestrian detection.Feature extraction can be divided into two types of sparse expression and dense expression again according to the mode of feature representation.The sparse expression method extracts the point of interest in the target area through certain key feature extraction algorithm, forms the sparse features expression pattern of target.For example yardstick invariant features conversion (Scale-invariant feature transform, SIFT) and principal component analysis (PCA) (Principal Component Analysis, PCA).The target signature of using these feature representation modes to extract has the advantage that is concise in expression, but when describing complex patterns, the information that its proper vector comprised is unfavorable for the sorter training, thereby not good at the effect in pedestrian detection field.And dense expression is that point by point scanning is carried out in the given area, extracts the dense characteristic expression pattern of target area.Typically have Lis Hartel seek peace gradient orientation histogram (Histograms of Oriented Gradient, HOG) etc.What the target signature that these method for distilling extract comprised contains much information, and help the sorter training, but the proper vector dimension is too high.With the HOG method is example, and the image of 64 * 128 sizes will produce the proper vector of at least 4096 dimensions, if adopt Lis Hartel to levy obtaining more intrinsic dimensionality, causes the characteristic amount of redundancy big, and counting yield is low, generally also needs the cooperation of feature selection approach.

Summary of the invention

The object of the invention is to provide a kind of feature representation succinct, and information redundancy is little, the pedestrian detection method based on wavelet fractal characteristic that detection efficiency is high.

The technical scheme that the present invention adopted is:

(1), read in training sample set, sample set comprises complete pedestrian's image of sorter training and does not comprise pedestrian's natural scene image that all sample images are normalized to 48 * 96 pixel sizes, and wherein training sample image is a gray level image;

(2), (x y) carries out two-dimensional wavelet transformation 3 times, obtains 12 sub-graphs after 3 layers of wavelet decomposition, and the yardstick of each layer correspondence is respectively 2 * 2,4 * 4,8 * 8, gets its mesoscale and be 4 * 4,8 * 86 small echo subgraphs in two-layer, promptly to sample image f

$\{f_2^H,f_2^V,f_2^D,f_3^H,f_3^V,f_3^D\}$

Wherein, f (x y) is the gradation of image value, (and x y) is pixel coordinate, and the subscript A of f, V, D, H are each subgraph that the 2-d wavelet branch solution subspace comprises, the corresponding LL subgraph of difference, LH subgraph, HL subgraph and HH subgraph, and the number of plies is decomposed in the subscript representative of f;

(3), step (2) is obtained carrying out suitable stretching and scale transformation again after each small echo subgraph takes absolute value respectively, make its span unification be mapped to 0～255;

(4), the small echo subgraph after step (3) handled asks for the fractal dimension vector respectively, makes yardstick ε=1,2,3 ..., n, n get the positive integer in 2～100, try to achieve D successively according to following formula ₁, D ₂..., D _N-1, promptly fractal dimension is vectorial,

$D_k=2-\frac{\ln (V_{{\mathrm{\ϵ}}_k}-V_{{\mathrm{\ϵ}}_{k-1}})-\ln (V_{{\mathrm{\ϵ}}_{k+1}}-V_{{\mathrm{\ϵ}}_k})}{\ln {\mathrm{\ϵ}}_k-{\ln \mathrm{\ϵ}}_{k+1}}$

$V_{\mathrm{\ϵ}}=\underset{i,j}{\mathrm{\Σ}}(u_{\mathrm{\ϵ}}(i,j)-b_{\mathrm{\ϵ}}(i,j))$

$u_{\mathrm{\ϵ}}(i,j)=\max \{u_{\mathrm{\ϵ}-1}(i,j),\underset{|(m,n)-(i,j)\≤1|}{\max }{u}_{\mathrm{\ϵ}-1}(m,n)\}$

$b_{\mathrm{\ϵ}}(i,j)=\max \{b_{\mathrm{\ϵ}-1}(i,j),\underset{|(m,n)-(i,j)\≤1|}{\max }{b}_{\mathrm{\ϵ}-1}(m,n)\}$

Wherein: D represents fractal dimension; K representes the subscript to dependent variable; ε _kRepresent a value in the yardstick span; u _εThe upper surface area of the fractal target when the expression yardstick is ε; b _εThe lower surface area of the fractal target when the expression yardstick is ε; V _εWhen the expression yardstick is ε, the volume that the upper and lower surface area of fractal target surrounds;

(5), the pairing fractal vector of each subgraph that step (4) is obtained is done standardization processing respectively;

(6), the fractal vector after the standardization that all subgraphs after step (5) standardization are corresponding makes up, and obtains one 6 * (n-1) dimensional vector, i.e. the wavelet fractal characteristic vector;

(7), wavelet fractal characteristic that step (6) is extracted is used to train soft support vector sorter, realizes the detection to unknown sample with the discriminant function after the training.

The present invention combines small echo and fractal theory; A kind of new pedestrian's target's feature-extraction method is proposed, be called wavelet fractal characteristic (Wavelet Fractal Signature, WFS); Be used to train the svm classifier device; This characteristic has succinct feature representation form, more excellent characteristic resolution characteristic, thus can improve the pedestrian detection effect.

The WFS characteristic comprises dense sampling and abstract two links:

(1) input picture is decomposed the small echo subpattern of nonautocorrelation under the different resolution at dense sampling element; Select wherein to help most portraying 6 sub-modes of pedestrian's characteristic; Comprised a large amount of luminance difference information; Directional information, thus the pedestrian detection problem is converted into the pattern classification problem of 6 dimension spaces, improved classifying quality.

(2) in abstract link; Through calculating the FRACTAL DIMENSION vector; Extract the characteristic difference of each subpattern under different scale; Can under considerably less intrinsic dimensionality (the maximum occurrences n of yardstick ε is, the proper vector of generation only has 48 dimensions), take out the characteristic of target under different resolution at 9 o'clock, have the advantage that is concise in expression.

(3) according to general knowledge, 2-d wavelet involved in the present invention decomposes, and link counting yielies such as FRACTAL DIMENSION calculating, L1 standardization are all very high, and the computing cost of SVM depends primarily on intrinsic dimensionality and support vector number.On the one hand, the WFS characteristic that provides only has 48 dimensions, is merely 1% of HOG characteristic; On the other hand, adopt the WFS characteristic when considerably less training sample, just can obtain higher detection efficient, significantly reduced the number of support vector, thereby reduced computing cost.

Therefore, the present invention had both had the big advantage of dense sample information amount, possessed the advantage of sparse sampling terseness again.When reducing counting yield, improved verification and measurement ratio.Another beneficial effect of the present invention is to facts have proved that when training sample was considerably less, the present invention program still can have very high verification and measurement ratio, can be applicable to the occasion that training sample is difficult to collect.

Description of drawings

Fig. 1 is a method flow diagram of the present invention.

Embodiment

Below in conjunction with accompanying drawing working of an invention is made and to be further specified.Fig. 1 is method flow diagram of the present invention, and is as shown in Figure 1, and this method comprises following 7 steps.

Step 101: read in training sample set, sample set comprises complete pedestrian's image of sorter training and does not comprise pedestrian's natural scene image that all sample images are normalized to 48 * 96 pixel sizes, and wherein training sample image is a gray level image;

Step 102, (x y) carries out two-dimensional wavelet transformation 3 times, obtains 3 layers of subgraph after the wavelet decomposition, and the yardstick of each layer correspondence is respectively 2 * 2,4 * 4,8 * 8 to sample image f.Remove yardstick and be all subgraphs of 2 * 2, and the f in other two kinds of yardsticks ^ASubgraph, remaining subgraph does

$\{f_2^H,f_2^V,f_2^D,f_3^H,f_3^V,f_3^D\}---\left(1\right)$

Wherein, f (x y) is the gradation of image value, (and x y) is pixel coordinate, and the subscript A of f, V, D, H are each subgraph that the 2-d wavelet branch solution subspace comprises, the corresponding LL subgraph of difference, LH subgraph, HL subgraph and HH subgraph, and the number of plies is decomposed in the subscript representative of f;

Step 103: carry out suitable stretching and scale transformation again after each the small echo subgraph in the formula (1) taken absolute value, make its span unification be mapped to 0～255;

Step 104: the small echo subgraph to after step 103 processing is asked for the fractal dimension vector respectively, is specially, and makes yardstick ε=1,2,3 ..., n tries to achieve D successively according to formula (2) ₁, D ₂..., D _N-1, promptly fractal dimension is vectorial, and n gets the integer between 2～100.The value of n is directly proportional with the intrinsic dimensionality of generation, but its value is big more, and the characteristic information amount that comprises is big more, otherwise then more little.Obtain when n gets 9 to optimum through a large amount of experiments;

$D_k=2-\frac{\ln (V_{{\mathrm{\ϵ}}_k}-V_{{\mathrm{\ϵ}}_{k-1}})-\ln (V_{{\mathrm{\ϵ}}_{k+1}}-V_{{\mathrm{\ϵ}}_k})}{\ln {\mathrm{\ϵ}}_k-{\ln \mathrm{\ϵ}}_{k+1}}$

$V_{\mathrm{\ϵ}}=\underset{i,j}{\mathrm{\Σ}}(u_{\mathrm{\ϵ}}(i,j)-b_{\mathrm{\ϵ}}(i,j))$ $\left(2\right)$

$u_{\mathrm{\ϵ}}(i,j)=\max \{u_{\mathrm{\ϵ}-1}(i,j),\underset{|(m,n)-(i,j)\≤1|}{\max }{u}_{\mathrm{\ϵ}-1}(m,n)\}$

$b_{\mathrm{\ϵ}}(i,j)=\max \{b_{\mathrm{\ϵ}-1}(i,j),\underset{|(m,n)-(i,j)\≤1|}{\max }{b}_{\mathrm{\ϵ}-1}(m,n)\}$

Step 105: the pairing fractal vector of each subgraph to step 104 obtains is done standardization respectively, promptly uses formula (3) the corresponding FRACTAL DIMENSION vector of each subgraph that standardizes,

${\mathrm{\β}}_i=\frac{{\mathrm{\α}}_i}{{\left|\right|{\mathrm{\α}}_i\left|\right|}_1+\mathrm{\λ}}---\left(3\right)$

Wherein, || || ₁Be 1 norm, α _iSubgraph FRACTAL DIMENSION vector before the expression standardization, β _iStandardized vector after the expression standardization, λ is a very little positive number, for assurance formula (3) effective;

This step can also be used alternate manners such as the L2 norm fractal vector that standardizes, but the L1 norm has the highest counting yield.

Step 106: the fractal vector after the standardization that all subgraphs after step 105 standardization are corresponding makes up, and obtains one 6 * (n-1) dimensional vector, promptly final wavelet fractal characteristic vector;

Step 107: the wavelet fractal characteristic that step 106 is extracted is used to train soft support vector sorter.Specific algorithm is following:

(A) known training sample set

T＝{(x ₁，y ₁)，…，(x _l，y _l)}∈(R ⁿ×{-1，1})

(B) select kernel function and punishment parameters C, construct and find the solution optimization problem

$\underset{\mathrm{\α}}{\min }L\left(\mathrm{\α}\right)=\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i-\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}\underset{j=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_{j}K(x_i,x_j)$ $\left(4\right)$

s.t.?0≤α _i?≤C，

i＝1，2.…，l.

Wherein,

is kernel function.

(C) solve the optimum solution of step (B) Chinese style (4), note is made α ^*, obtain optimal classification lineoid (w with formula (5) ^*, b ^*),

$b^{*}=y_j-\underset{i\∈\mathrm{SV}}{\mathrm{\Σ}}{y}_i{\mathrm{\α}}_i^{*}K(x_i,x_j)$

(D) thus obtain the discriminant function f (x) that is used to classify,

$f\left(x\right)=\mathrm{sgn}((w^{*}\·x)+b^{*})$

$=\mathrm{sgn}(\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i^{*}K(x_i,x)+y_i-\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{y}_i{\mathrm{\α}}_i^{*}K(x_i,x))---\left(6\right)$

Wherein, the penalty factor C of SVM is the constant greater than zero, and its optimal value need be confirmed through test, get 2000 among the present invention.Kernel function can use polynomial kernel, two layers of neural network to examine and gaussian kernel, the first-selected gaussian kernel of the present invention, and σ gets 1.5, and other nuclear takes second place.

Utilize a window in by altimetric image, to slide with the same size of training sample image; Through obtaining discriminant function f (x) target in the window is classified; Judge whether this pattern is the pedestrian; Thereby the realization pedestrian detection, the WFS characteristic of x representative image wherein, pedestrian or non-pedestrian are represented in the output of discriminant function f (x).

Claims (3)

1. the pedestrian detection method based on wavelet fractal characteristic is characterized in that, this method may further comprise the steps:

(1), read in training sample set, sample set comprises complete pedestrian's image of sorter training and does not comprise pedestrian's natural scene image that all sample images are normalized to 48 * 96 pixel sizes, and wherein training sample image is a gray level image;

(2), (x y) carries out two-dimensional wavelet transformation 3 times, obtains 12 sub-graphs after 3 layers of wavelet decomposition, and the yardstick of each layer correspondence is respectively 2 * 2,4 * 4,8 * 8, gets its mesoscale and be 4 * 4,8 * 86 small echo subgraphs in two-layer, promptly to sample image f

$\{f_2^H,f_2^V,f_2^D,f_3^H,f_3^V,f_3^D\}$

Wherein, f (x y) is the gradation of image value, (and x y) is pixel coordinate, and the subscript A of f, V, D, H are each subgraph that the 2-d wavelet branch solution subspace comprises, the corresponding LL subgraph of difference, LH subgraph, HL subgraph and HH subgraph, and the number of plies is decomposed in the subscript representative of f;

(3), step (2) is obtained carrying out suitable stretching and scale transformation again after each small echo subgraph takes absolute value respectively, make its span unification be mapped to 0～255;

(4), the small echo subgraph after step (3) handled asks for the fractal dimension vector respectively, makes yardstick ε=1,2,3 ..., n, n get the positive integer in 2～100, try to achieve D successively according to following formula ₁, D ₂..., D _N-1, promptly fractal dimension is vectorial,

$D_k=2-\frac{\ln (V_{{\mathrm{\ϵ}}_k}-V_{{\mathrm{\ϵ}}_{k-1}})-\ln (V_{{\mathrm{\ϵ}}_{k+1}}-V_{{\mathrm{\ϵ}}_k})}{\ln {\mathrm{\ϵ}}_k-{\ln \mathrm{\ϵ}}_{k+1}}$

$V_{\mathrm{\ϵ}}=\underset{i,j}{\mathrm{\Σ}}(u_{\mathrm{\ϵ}}(i,j)-b_{\mathrm{\ϵ}}(i,j))$

$u_{\mathrm{\ϵ}}(i,j)=\max \{u_{\mathrm{\ϵ}-1}(i,j),\underset{|(m,n)-(i,j)\≤1|}{\max }{u}_{\mathrm{\ϵ}-1}(m,n)\}$

$b_{\mathrm{\ϵ}}(i,j)=\max \{b_{\mathrm{\ϵ}-1}(i,j),\underset{|(m,n)-(i,j)\≤1|}{\max }{b}_{\mathrm{\ϵ}-1}(m,n)\}$

Wherein: D represents fractal dimension; K representes the subscript to dependent variable; ε _kRepresent a value in the yardstick span; u _εThe upper surface area of the fractal target when the expression yardstick is ε; b _εThe lower surface area of the fractal target when the expression yardstick is ε; V _εWhen the expression yardstick is ε, the volume that the upper and lower surface area of fractal target surrounds;

(5), the pairing fractal vector of each subgraph that step (4) is obtained is done standardization processing respectively;

(6), the fractal vector after the standardization that all subgraphs after step (5) standardization are corresponding makes up, and obtains one 6 * (n-1) dimensional vector, i.e. the wavelet fractal characteristic vector;

(7), the wavelet fractal characteristic that step (6) is extracted is used to train soft support vector sorter; The discriminant function that acquisition is used to classify; Utilize a window in by altimetric image, to slide again with the same size of training sample image; Discriminant function with obtaining after the training is classified to the target in the moving window, judges whether this pattern is the pedestrian, thereby realizes pedestrian detection.

2. a kind of pedestrian detection method based on wavelet fractal characteristic according to claim 1 is characterized in that: the maximum occurrences n of said step (4) mesoscale ε is 9.

3. a kind of pedestrian detection method based on wavelet fractal characteristic according to claim 1 is characterized in that: the employed normalized method of said step (5) is: with the corresponding FRACTAL DIMENSION vector of each subgraph of L1 norm standardization, promptly

${\mathrm{\β}}_i=\frac{{\mathrm{\α}}_i}{{\left|\right|{\mathrm{\α}}_i\left|\right|}_1+\mathrm{\λ}}$

Wherein, || || ₁Be 1 norm, α _iSubgraph FRACTAL DIMENSION vector before the expression standardization, β _iStandardized vector after the expression standardization, λ is a very little positive number, and is effective in order to guarantee division.

Images