CN104091157A - Pedestrian detection method based on feature fusion - Google Patents

Abstract

The invention discloses a pedestrian detection method based on feature fusion. Pedestrians are detected through fusion of pedestrian features in pictures, limitation of a single feature can be overcome, and therefore the pedestrian detection rate is improved; meanwhile, dimensions are reduced and the detection speed is increased. The HOG and textural feature LBP fusion pedestrian detection method achieving dimension reduction through PCA is provided, dimension reduction is carried out on the HOG through the PCA, and the feature dimension is decreased to 300 from 3780. By the adoption of the multi-feature fusion method, LBP textural features are fused, an SVM classifier is combined, the training and detecting time is shortened while the HOG-LBP recognition rate is increased, and the problem of pedestrian blocking can be well solved.

Description

A kind of pedestrian detection method based on Fusion Features

Technical field

The invention belongs to pedestrian detection field in computer vision, particularly a kind of pedestrian detection method based on Fusion Features.

Background technology

Pedestrian detecting system relates to subjects and the computer technology problems such as pattern-recognition, computer vision, be a comparatively complicated and huge engineering, in the fields such as intelligent automobile, intelligent transportation, video monitoring, robot and senior man-machine interaction, have application prospect very widely.Although there has been preliminary achievement in research now, but also there are a lot of difficult points, as features such as non-rigid, the attitude of human body self, the varietys of appearance kimonos, be easy to be subject to illumination, weather and external environment condition complicated and changeable etc. and permitted multifactorial interference, cause human detection to become a quite complicated problem and face lot of challenges.

The technical method that has at present many pedestrian detection, wherein great majority are the detection methods based on machine learning, mainly comprise two importances, and one is that feature is described operator, and another is learning algorithm.Feature is described operator haar-like, HOG (gradient orientation histogram), LBP (local binary patterns) and edgelet (edge feature) etc.Learning algorithm has support vector machine (SVM) and cascade classifier Adaboost.Existing pedestrian detection technology has the shortcoming that detection speed is slow, accuracy rate is not high enough.

Summary of the invention

Goal of the invention: the problem existing for prior art, the invention provides the pedestrian detection method based on Fusion Features that a kind of detection speed is fast, accuracy rate is high.

Summary of the invention: the invention provides a kind of pedestrian detection method based on Fusion Features, comprise the following steps:

Step 1: gather image;

Step 2: the gamma of the image that step 1 is collected (gamma) space and color space carry out standardization;

Step 3: the image obtaining in step 2 is carried out to the calculating of pixel gradient;

Step 4: the image that step 3 is obtained carries out the histogrammic statistics of unit inside gradient;

Step 5: the unit inside gradient Nogata segment normalized that step 4 statistics is obtained obtains piece normalization histogram;

Step 6: step 5 is obtained to piece normalization histogram and extract feature, obtaining dimension is the HOG proper vector of 3780*3780 dimension;

Step 7: the image of step 6, by the HOG feature of training sample, is obtained to sample average, computation of characteristic values, proper vector and covariance matrix U, wherein the matrix size of U is 3780*3780;

Step 8: for each the HOG eigenwert in covariance matrix, carry out dimensionality reduction by PCA dimensionality reduction formula, obtain having the p dimension matrix of HOG-PCA feature;

Step 9: the image gathering for step 1 carries out image to be cut apart;

Step 10: every LBP feature histogram in the image after asking for step 9 and cutting apart;

Step 11: the LBP feature histogram obtaining for step 10 is normalized histogram and processes;

Step 12: the histogram obtaining after normalized for step 11 extracts texture LBP feature;

Step 13: the LBP eigenmatrix level that the HOG-PCA eigenmatrix that step 8 is obtained and step 12 obtain is linked togather, and obtains HOG-PCA-LBP eigenmatrix;

Step 14, treats all pedestrian's images Output rusults that is disposed.

Further, in described step 9, adopting 16 * 16 blocks of images of cutting apart gathering in step 1 to carry out image cuts apart.Help like this when the image to after cutting apart carries out texture feature extraction, can not increase the complexity of calculating, guarantee to extract most texture information simultaneously.

Beneficial effect: compared with prior art, the present invention is by the fusion of pedestrian's feature in picture is detected to pedestrian, thus the limitation that can make up single features improves pedestrian detection rate, reduces dimension simultaneously and has improved detection speed.A kind of HOG of the PCA of utilization dimensionality reduction and the pedestrian detection method that textural characteristics LBP merges have been proposed.Utilize PCA to carry out dimensionality reduction to HOG, intrinsic dimensionality is reduced to 300 dimensions from 3780 dimensions.Adopt the method for many Fusion Features, merge LBP textural characteristics, combine with svm classifier device, when improving HOG-LBP discrimination, the time of having reduced training and having detected, and can fine processing pedestrian occlusion issue.

Accompanying drawing explanation

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

Embodiment

Below in conjunction with accompanying drawing, the present invention is described in detail.

As shown in Figure 1, the pedestrian detection method based on Fusion Features of the present invention, its step is as follows

Step 1: gather image.

Step 2: the gamma of the image that step 1 is collected (gamma) space and color space carry out standardization.

Square root gamma standardization is the impact of removal of images global illumination and contrast well.In the present embodiment, adopt square root and log compression to carry out denoising to the image collecting, and zoom to same size.On each passage of color space, to use these two kinds of methods to calculate.

Step 3: the image obtaining in step 2 is carried out to the calculating of pixel gradient;

The First-order Gradient of computed image.Calculate derivative and not only can obtain human body contour outline and texture information, can also further weaken the impact of illumination.Because the computing of HOG feature is to template operator sensitivity very, through Experimental Comparison, find, that the simplest one-dimensional discrete differential template (1,0 ,+1) and transposition thereof carry out gradient calculation to each pixel of image and can obtain best detection effect in horizontal and vertical directions on the contrary.Can distinguish by formula gradient-norm value and the deflection of calculating pixel point (x, y):

$G(x,y)=\sqrt{{(H(x+1,y)-H(x-1,y))}^2+{(H(x,y+1)-H(x,y-1))}^2}$

$\mathrm{\α}(x,y)=\mathrm{ta}{n}^{-1}\left(\frac{H(x,y+1)-H(x,y-1)}{H(x+1,y)-H(x-1,y)}\right)$

Wherein, G (x, y), α (x, y), H (x, y) represents respectively the gray-scale value of gradient magnitude, gradient direction and the pixel of pixel.For coloured image, can calculate respectively the gradient of their each Color Channels, select the value of amplitude maximum wherein as the gradient of this pixel.

Step 4: the image that step 3 is obtained carries out the histogrammic statistics of unit inside gradient;

The essence of this step is to local image-region coding, and can keep the hyposensitiveness perception to the outward appearance of human object in image and posture.We are divided into several zonules image window, and these zonules are called as " cell ", i.e. cell.Image averaging is divided into some foursquare cells (cell), and each cell comprises 8 * 8 pixels, handle in each cell gradient direction be divided into 9 intervals (bin), then in each cell, the Grad of all pixels carries out statistics with histogram in each bin interval respectively again, such cell obtain one 9 dimension proper vector.

Step 5: the unit inside gradient Nogata segment normalized that step 4 statistics is obtained obtains piece normalization histogram;

The main cause that the variation range of Grad is very wide is the variation due to the contrast of exposure rate local in image and foreground-background.Thereby the detection effect that will obtain, must carry out local contrast standardization effectively.Standardized method has a lot, and general method is all that one group of cell is put in a piece, then distinguishes each piece of standardization.2 * 2 cell form a piece, and such piece just forms the proper vector of 36 dimensions, and recycling L2-norm is normalized whole, obtains final proper vector.

Step 6: step 5 is obtained to piece normalization histogram and extract feature, obtaining dimension is the HOG proper vector of 3780*3780 dimension;

The image gathering in embodiment is that 64 * 128, cell is 8 * 8, and piece is 16 * 16, and piece image just comprises 105 pieces so, and each piece is 36 dimensional vectors, so the HOG proper vector of the image of 64 * 128 sizes is 36 * 105=3780 dimension.

Step 7: the image of step 6, by the HOG feature of training sample, is obtained to sample average, and according to formula computation of characteristic values, proper vector and covariance matrix U, wherein the matrix size of U is 3780 * 3780;

By the HOG feature of training sample, N, x _i, represent respectively sample size, sample value and sample mean.According to following formula computation of characteristic values, proper vector and covariance matrix U, wherein the matrix size of U is 3780*3780.

$U^T=\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}(x_i-\stackrel{\‾}{x}){(x_i-\stackrel{\‾}{x})}^T$

Step 8: for each the HOG eigenwert in covariance matrix, carry out dimensionality reduction by PCA dimensionality reduction formula, obtain having the p dimension matrix of HOG-PCA feature;

The present invention adopts PCA (principal component analysis (PCA)) to carry out dimensionality reduction to proper vector, experimental results show that PCA dimensionality reduction effect is relatively good.PCA main thought: carry out spatial alternation by the sample space to original, make on the lower and mutually orthogonal space of original coordinate projection to new dimension.

Get front p the major component of covariance matrix, to each the HOG eigenwert in training sample, by following formula, carry out Feature Dimension Reduction, obtain HOG-PCA feature, vectorial dimension is p dimension, and it is 300 that the numerical value of p will be determined by experiment.

$y=U^T(x_i-\stackrel{\‾}{x})$

Step 9: the image gathering for step 1 carries out image to be cut apart;

In the present embodiment, the picture of 64 * 128 original sizes being cut apart according to 16 * 16, is mainly that 8 * 8 too small, can increase computation complexity because when texture feature extraction, and 32 * 32 is too large, can lost part texture information.But based on gray level image during the required LBP feature of real process, so first coloured image should be converted into gray level image.

Step 10: every LBP feature histogram in the image after asking for step 9 and cutting apart;

For the subimage of every 16 * 16, according to LBP82 operator, ask for the textural characteristics of image, obtain 256 dimensional feature vectors, then 256 dimensional feature vectors are converted into 59 dimensional feature vectors.

Step 11: the LBP feature histogram obtaining for step 10 is normalized histogram and processes;

In order to improve the robustness of proper vector, overcome the interference of some noises, with HOG feature class seemingly, need to be normalized operation to 59 dimensional feature vectors that extract.Consider different normalized factors, the anti-interference of subtend measure feature has larger impact, adopts L2-norm normalized factor in the present embodiment, and that obtains is effective.

Step 12: the histogram obtaining after normalized for step 11 extracts texture LBP feature;

In the present embodiment, image is 64 * 128, and detection window is divided into 32 cell, and the proper vector of each cell is 59 dimensions, so finally obtain the LBP proper vector of 1888 dimensions.

Step 13: the LBP eigenmatrix level that the HOG-PCA eigenmatrix that step 8 is obtained and step 12 obtain is linked togather, and obtains HOG-PCA-LBP eigenmatrix.

Step 14, treats all pedestrian's images Output rusults that is disposed.

Experiment adopts the method for HOG feature, HOG+LBP feature and HOG-PCA+LBP feature in this paper to verify detection effect, and result is as shown in table 1.Experimental data shows that method in this paper has advantage than the method for Dalal and Wang, and on the training time, this paper method time used is approximately half of HOG, is about 1/3 of HOG+LBP.On detection time, be also like this.Aspect discrimination, the discrimination of HOG+LBP than HOG, improve 5.7% and can well process occlusion issue.This paper method HOG+LBP that compares, although that discrimination improves is few, suitable with the former when processing occlusion issue, still the time used but greatly reduces approximately 2/3, so this paper method is effective.Table 1

Feature kind	Vector dimension/dimension	Training time/s	Detection time/ms	Discrimination
					HOG	3780	223	91	89％
HOG+LBP	5668	334	136	94.7％
					HOG-PCA+LBP	2188	128	47	96.37％

Claims (2)

1. the pedestrian detection method based on Fusion Features, is characterized in that: comprise the following steps:

Step 1: gather image;

Step 2: the gamma space of the image that step 1 is collected and color space carry out standardization;

Step 3: the image obtaining in step 2 is carried out to the calculating of pixel gradient;

Step 4: the image that step 3 is obtained carries out the histogrammic statistics of unit inside gradient;

Step 5: the unit inside gradient Nogata segment normalized that step 4 statistics is obtained obtains piece normalization histogram;

Step 6: step 5 is obtained to piece normalization histogram and extract feature, obtaining dimension is the HOG proper vector of 3780*3780 dimension;

Step 7: the image of step 6, by the HOG feature of training sample, is obtained to sample average, computation of characteristic values, proper vector and covariance matrix U, wherein the matrix size of U is 3780*3780;

Step 8: for each the HOG eigenwert in covariance matrix, carry out dimensionality reduction by PCA dimensionality reduction formula, obtain having the p dimension matrix of HOG-PCA feature;

Step 9: the image gathering for step 1 carries out image to be cut apart;

Step 10: every LBP feature histogram in the image after asking for step 9 and cutting apart;

Step 11: the LBP feature histogram obtaining for step 10 is normalized histogram and processes;

Step 12: the histogram obtaining after normalized for step 11 extracts texture LBP feature;

Step 13: the LBP eigenmatrix level that the HOG-PCA eigenmatrix that step 8 is obtained and step 12 obtain is linked togather, and obtains HOG-PCA-LBP eigenmatrix.

Step 14, treats all pedestrian's images Output rusults that is disposed.

2. the pedestrian detection method based on Fusion Features according to claim 1, is characterized in that: in described step 9, adopt 16 * 16 blocks of images of cutting apart gathering in step 1 to carry out image and cut apart.