CN103455826A - Efficient matching kernel body detection method based on rapid robustness characteristics - Google Patents

Abstract

The present invention proposes an efficient matching kernel human body detection method based on fast and robust features, which mainly solves the problem that traditional methods cannot better deal with mixed image backgrounds or uneven illumination. The implementation steps are: (1) Select the image of the training sample set; (2) Extract the SURF feature points of the image; (3) Construct the initial vector base of each layer; (4) Obtain the maximum kernel function feature of the sampling layer; (5) Obtain Efficient image matching of nuclear features; (6) classification training; (7) input image for scanning; (8) detection of scanning window; (9) output of detection results. The invention extracts the local information of the image layer by layer, performs feature learning to map the features to a low-dimensional space, assembles them into a feature set, and then uses a linear classifier to train the features to obtain a human body detection classifier. The invention can accurately detect human body information in natural images in the field of image processing.

Description

An Efficient Matching Kernel Human Detection Method Based on Fast and Robust Features

technical field

The invention belongs to the technical field of image processing, and further relates to a fast and robust feature-based Efficient Match Kernel (Efficient Match Kernel EMK) human detection method in the technical field of static human detection. The invention can be used to detect human body information from static images to achieve the purpose of identifying human body targets.

Background technique

Human body detection is the process of judging the location of human body information from natural images. In recent years, due to its application value in the fields of intelligent monitoring, driver assistance systems, human motion capture, and pornographic image filtering, it has become a computer vision field. key technology. However, human body detection has become a very difficult problem due to the diversity of human body poses, background clutter, clothing textures, lighting conditions, and self-occlusion. At present, the methods of human detection in static images are mainly divided into two categories: human body detection methods based on human body models and human body detection methods based on learning.

The first is a human detection method based on a human body model. This method does not need to learn the database, has a clear human body model, and then performs human body recognition according to the relationship between each part of the model construction and the human body.

Beijing Jiaotong University disclosed a human body model-based detection method in its patent application "a human body detection method" (patent application number CN201010218630.8, publication number CN101908150A). In this method, human body detection templates with a certain degree of ambiguity are established through human body samples of various body shapes and postures to determine human body candidate areas. This method can better deal with the occlusion problem, can calculate the posture of the human body, and improve the efficiency and accuracy of human body detection. It is difficult to achieve good detection results under the circumstances.

The second is a learning-based human detection method. The method learns a classifier from a series of training data through machine learning, and then uses the classifier to classify and identify input windows.

Harbin Engineering University disclosed a method combining multi-scale gradient histogram HOG features and head Human detection method based on color histogram features. This method extracts the HOG feature of the gradient histogram and combines the feature template at the same time, which increases the function of head feature discrimination, improves the detection rate compared with the traditional method, and has a good feature recognition effect especially for the image space with little background change. , however, this method still has the disadvantage that the detection results will be disturbed in the case of mixed background or uneven illumination.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies of the above-mentioned prior art, propose a kind of high-efficiency matching kernel human body detection method based on fast and robust features, adopt the human body detection method based on learning, extract the local information of the image by layering, and then perform dictionary Learn to map features to a low-dimensional space, assemble them into a feature set, use a linear classifier to train the feature set, get a classifier for human detection, and then use this classifier to perform human detection on the image to be detected.

In order to achieve the above object, the present invention includes two processes of obtaining a detection classifier and utilizing the obtained classifier to detect images, and the specific implementation steps are as follows:

In the first process, the specific steps to obtain the detection classifier are as follows:

(1) Select training sample set images:

1a) Using a bootstrap operation, obtain enough negative samples from the non-human natural images of the INRIA database;

1b) Combining the obtained negative sample image with the negative sample set in the INRIA database to form a new negative sample set;

1c) The obtained new negative sample set image and the positive sample set in the INRIA database constitute the human training sample set.

(2) Extract image SURF feature points:

2a) Divide each image in the human training sample set into a grid of 8*8 pixels, and sample each grid at a scale of 16, 25, and 36 pixels, and each scale sampling forms a sampling layer of the grid;

2b) For each 8*8 pixel grid, calculate the sum of the squares of the horizontal gradient and the vertical gradient of the sampling points in the grid after each layer of sampling, and use the sampling point corresponding to the maximum value of the gradient square sum as the pixel grid in the sampling layer The fast and robust feature SURF feature points;

2c) For each image in the human training sample set, randomly select 15 feature points from the fast robustness feature SURF feature points of all pixel grids on each sampling layer, and use them as images of the human training sample set on the sampling layer The fast robustness features of SURF feature points.

(3) Construct the initial vector base of each layer:

Use the k-means clustering method to cluster the fast and robust feature SURF feature points of all images in the human training sample set on each sampling layer, define 450 cluster centers, and obtain the entire human training sample set on the sampling layer The 450-dimensional visual vocabulary constitutes the initial basis vector of the sampling layer.

(4) Obtain the maximum kernel function feature of the sampling layer:

For the initial basis vector of each sampling layer, the dictionary learning is carried out by using the kernel singular value decomposition CKSVD with constraints respectively, and the maximum kernel function feature of the sampling layer is obtained.

(5) Obtain image efficient matching kernel features:

5a) For each sampling layer, arrange the element values of the maximum kernel function feature of the sampling layer in descending order, and judge whether the number of elements of the maximum element value is 1, and if so, use the maximum kernel function feature of the sampling layer as the feature of the sampling layer Vector output, otherwise, the element value equal to the maximum value in the maximum kernel function feature of the sampling layer is set to zero, and the maximum kernel function feature of the sampling layer after zeroing is output as the feature vector of the sampling layer;

5b) Perform weighted summation of the feature vectors of all sampling layers to obtain all scale features and store all scale features;

5c) Calculate the mean value of the elements of each row of all scale feature vectors, and accumulate the number of mean values at the corresponding mean points on the horizontal axis of the coordinates to obtain the distribution of the mean values of the elements of all rows of all scale feature vectors, and select the distribution of the mean values of all rows Gaussian-like distribution of features as efficient matching kernel features for fast and robust properties of the final image.

(6) Classification training:

Using the support vector machine (SVM) classifier to classify and train the extracted high-efficiency matching kernel features of fast and robust features, a detection classifier is obtained.

In the second process, the specific steps of using the obtained detection classifier to detect the image are as follows:

(7) Input image to scan:

Input a detected image, scan the entire detected image by window scanning method to obtain a group of scanning window images, and input the group of scanning window images into the detection classifier.

(8) Detection scan window:

8a) Use the detection classifier to judge whether the input scanning window image contains human body information, if there is no human body information, then locate the detected image as a non-human natural image, otherwise, from all the judged human body information In the scan window image, find the scan window image with the highest score of the detection classifier as the main window image;

8b) From the remaining scanning window images with human body information other than the main window image, the scanning window image overlapping with the main window image by more than 50% is combined with the main window image, and the window obtained by window combination is used as a detection result Save, delete all images participating in the window composition;

8c) Determine whether there are any scan window images with human body information left, if so, find the image with the highest detection classifier score among the remaining scan window images as the main window image, and perform step 8b), otherwise, perform step (9) .

(9) Output detection results:

All the windows obtained by combining the windows are marked on the detected image, and the marked image is output as the human body detection result of the detected image.

Compared with the prior art, the present invention has the following advantages:

First, the present invention uses fast robust features in the feature extraction process of human body detection. The fast robust features perform statistics on regional images by calculating local image gradient changes to form a statistical image of the entire image. feature, which can avoid the problem of fuzzy representation produced by edge-based and contour-based image representation methods in the prior art, so that the present invention can obtain better detection results when dealing with mixed background and unevenly illuminated images.

Second, in the feature extraction process of human body detection, the present invention extracts features layer by layer from images, effectively utilizes feature point information on different scales, and avoids local matching errors caused by too small scales in the prior art, so that the present invention can Get better test results.

Third, the present invention uses the dictionary learning method to map the extracted image features to a low-dimensional space, and assembles them into a feature set. Compared with the prior art, the dimensionality of image features is reduced, and the calculation time of image features is effectively reduced. and data calculations.

Description of drawings

Fig. 1 is a flow chart of the present invention;

Fig. 2 is a sample image used in the present invention;

Fig. 3 is the comparison diagram of the classifier classification performance of the present invention and the human body detection method based on gradient histogram HOG feature;

Fig. 4 is the simulation diagram of the method of the present invention and the human body detection method based on the gradient histogram HOG feature to the uneven illumination image;

FIG. 5 is a simulation diagram of the method of the present invention and the human body detection method based on the gradient histogram HOG feature for human body detection on complex background images.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings.

With reference to accompanying drawing 1, concrete steps of the present invention are as follows:

Step 1. Select training sample set images.

Using a bootstrap operation, enough negative images are obtained from the non-human natural images of the INRIA database.

The specific steps of the bootstrap operation are as follows:

In the first step, m positive sample images and n negative sample images are randomly selected from the INRIA database, among which 100≤m≤500, 100≤n≤800, and n≤m≤3n, using the gradient orientation histogram HOG feature extraction method, Feature extraction is performed on all the selected positive and negative sample images, and the SVM classifier is used to perform classification training on the extracted features to obtain the initial classifier.

The second step is to continuously randomly select non-human natural images in the INRIA database, use the scanning window of the size of the sample image, and use 8 pixels as the moving unit from left to right, and 16 pixels as the moving unit from top to bottom to scan the whole body. A detected non-human natural image; input all the images in the scanning window to the initial classifier for detection, save the misclassified scanning window images until the number of misclassified scanning window images reaches a, 200≤a ≤500, stop selecting non-human natural images; randomly select b images from misclassified scan window images, 1/5a≤b≤1/3a, and form a new negative sample set with the current negative sample image.

The third step is to extract m positive sample images and a new negative sample set at random, perform gradient direction histogram HOG feature extraction, train a classifier, detect non-human natural images and update the negative sample set.

The fourth step is to repeat the operation of the third step until the updated final training sample set consists of 2416 positive sample images and 13500 negative sample images, and the size of the samples is 128×64 pixels.

The obtained negative sample image and the negative sample set in the INRIA database are combined to form a new negative sample set.

The obtained new negative sample set image and the positive sample set in the INRIA database constitute the human training sample set.

In the embodiment of the present invention, in the final human training sample set, the training sample set consists of 2416 positive samples and 13500 negative samples, the test sample set consists of 1132 positive samples and 4050 negative samples, and the size of the sample images is 128×64 pixels.

Figure 2 is a partial sample image used in the present invention, wherein Figure 2(a) is a partial positive sample image used in the present invention, and Figure 2(b) is a partial negative sample image used in the present invention.

Step 2, extract the SURF feature points of the image.

Each image in the human training sample set is divided into 8*8 pixel grids, and each grid is sampled at a scale of 16, 25, and 36 pixels, and each scale sampling forms a sampling layer of the grid.

For each 8*8 pixel grid, calculate the sum of the squares of the horizontal gradient and the vertical gradient of the sampling points in the grid after each layer of sampling, and use the sampling point corresponding to the maximum value of the gradient square sum as the fast speed of the pixel grid in the sampling layer Robustness features SURF feature points.

For each image in the human body training sample set, 15 feature points are randomly selected from the fast robust feature SURF feature points of all pixel grids on each sampling layer, and used as the fast Robustness features SURF feature points.

Step 3, construct the initial vector base of each layer.

Using the k-means clustering method, cluster all the fast and robust feature SURF feature points of all training sample images on each sampling layer, define 450 cluster centers, and obtain the 450-dimensional visual vocabulary of the entire training sample at the sampling layer , constituting the initial vector basis of the sampling layer.

The specific steps of the k-means clustering method are as follows:

In the first step, for each sampling layer, randomly select 450 fast robustness feature SURF feature points from all sample images of the human training sample set in the sampling layer as the fast robustness feature SURF feature points, as the sampling layer For the initial clustering centers, the data values of 450 initial clustering centers are used as the clustering center values of the initial clustering centers.

The second step is to calculate the Euclidean distance from all the fast and robust SURF feature points of the human training sample set at the sampling layer to each cluster center.

The third step is to classify each fast robust feature SURF feature point of the sampling layer into the category of the cluster center closest to itself.

The fourth step is to judge whether the data average value of the fast robust feature SURF feature points of each class after classification is equal to the cluster center value, if yes, perform the fifth step, otherwise, calculate the feature points of each class The average value of the data is used as the new cluster center value, and returns to the second step.

The fifth step is to save 450 cluster center values, use the 450 cluster center values to form a column vector, and output the column vector as the initial basis vector of the entire human training sample set on the sampling layer.

Step 4, obtain the maximum kernel function feature of the sampling layer.

For the initial basis vector of each sampling layer, the dictionary learning is carried out by using the kernel singular value decomposition CKSVD with constraints respectively, and the maximum kernel function feature of the sampling layer is obtained.

The steps for bootstrap operation are as follows:

In the first step, for each sampling layer, the initial basis vector of the sampling layer is projected onto a 450-dimensional space, and the projection vector of the initial basis vector of the sampling layer is obtained by the following formula:

R=R ₁ ×[v ₁ ,...v _j ...,v _N ]

Among them, R represents the projection vector of the initial basis vector of the sampling layer, _R1 represents the initial basis vector of the sampling layer, and _vj represents the projection coefficient of the jth feature point extracted on the sampling layer from all sample images of the human training sample set Vector, v _j =[v _1j ,v _2j ,...v _ij ...,v _Mj ] ^T , v _ij represents the jth feature point extracted from the i-th sample image in the human training sample set on the sampling layer Projection coefficient, M represents the number of all sample images in the human training sample set, j=1, 2,..., N, N represents the number of feature points randomly selected on the sampling layer for each sample image in the human training sample set number.

In the second step, an approximation function is constructed according to the following formula to fit the projection of the initial basis vector on the approximation sampling layer in the projection space:

f(r)=argmin||r-R||

Among them, r represents the maximum kernel function feature on the sampling layer, R represents the projection of the initial basis vector R ₁ of the sampling layer, ||·|| represents the 2-norm, and argmin||·|| represents the minimum value.

Substitute R=R ₁ ×[v ₁ ,...v _j ...,v _N ] into the above formula, and expand r by r=[r ₁ ,...r _j ...,r _N ], The quadratic approximation function f(v,r) of the maximum kernel function feature r to the initial basis vector R ₁ is obtained by the following formula:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; v</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

Among them, v represents the low-dimensional projection coefficient vector of all feature points extracted from all sample images of the human training sample set, and v=[v ₁ ,...v _j ...,v _N ], v _j represents the human training sample The vector of the projection coefficients of the jth feature point extracted on the sampling layer from all sample images in the set, N represents the number of randomly selected feature points on the sampling layer for each sample image in the human training sample set, r _j represents The largest kernel feature vector of the jth feature point extracted from all sample images of the human training sample set, R ₁ represents the initial basis vector of the sampling layer.

The third step is to use the stochastic gradient descent method to solve the approximation function, and obtain the following formula to iteratively update the maximum kernel function feature on the sampling layer to form a low-dimensional image feature representation:

$\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; \&eta;</span>}}{\mathrm{<span\; class="notranslate">\; k</span>}}\frac{<span\; class="notranslate">\; \&PartialD;</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; R</span>}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; )</span>}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; r</span>}}$

表示计算括号里的式子对r的导数，r_j表示人体训练样本集中的所有样本图像在采样层上提取的第j个特征点的最大核特征向量，R₁表示采样层上的初始基向量，R₁ ^T表示采样层上的初始基向量R₁的转置向量，设定迭代次数为1000次，迭代完成后得到的r(1000)作为采样层上的最终的最大核函数特征，j=1,2,...,N，N表示人体训练样本集中的每一幅样本图像在采样层上的随机选取的特征点的数量。Among them, r(k+1) represents the maximum kernel function feature on the sampling layer obtained by iterating k+1 times, k represents the number of iterations, r(k) represents the maximum kernel function feature on the sampling layer obtained by iterating k times, η Indicates the learning rate, which is a constant,

Indicates the calculation of the derivative of the formula in parentheses to r, r _j represents the maximum kernel feature vector of the jth feature point extracted from all sample images in the human training sample set on the sampling layer, and R ₁ represents the initial basis vector on the sampling layer , R ₁ ^T represents the transpose vector of the initial basis vector R ₁ on the sampling layer, set the number of iterations to 1000, and the r(1000) obtained after the iteration is used as the final maximum kernel function feature on the sampling layer, j= 1, 2,..., N, N represents the number of randomly selected feature points on the sampling layer for each sample image in the human training sample set.

Step 5, obtain image high-efficiency matching kernel features.

For each sampling layer, arrange the element values of the maximum kernel function feature of the sampling layer in descending order, and judge whether the number of elements of the maximum element value is 1, and if so, output the maximum kernel function feature of the sampling layer as the feature vector of the sampling layer , otherwise, the element whose value is equal to the maximum value in the maximum kernel function feature of the sampling layer is set to zero, and the maximum kernel function feature of the sampling layer after zeroing is output as the feature vector of the sampling layer.

Weighted and summed the feature vectors of all sampling layers to obtain all scale features and store all scale features.

The method of weighted summation is as follows:

G*=G _i ×A _i

w_i=1/p_i，p_i表示每个采样层的采样尺度的像素大小，p_i={16,25,36}。Among them, G* represents the characteristics of all sampling layers, G _i represents the feature vector of each sampling layer, i=1,2,3, A _i represents the weight corresponding to each sampling layer,

w _i =1/p _i , p _i represents the pixel size of the sampling scale of each sampling layer, p _i ={16,25,36}.

Calculate the mean value of each row element of all scale feature vectors, and accumulate the number of mean values at the corresponding mean points on the horizontal axis of the coordinates to obtain the distribution of the element mean values of all rows of all scale feature vectors, and select the element mean value distribution of all rows similar to Gaussian-distributed features as efficient matching kernel features for fast robustness properties of the final image.

Step 6, classification training.

Using the support vector machine (SVM) classifier to classify and train the extracted high-efficiency matching kernel features of fast and robust features, a detection classifier is obtained.

Step 7, input image to scan.

Input a detected image, scan the entire detected image by window scanning method to obtain a group of scanning window images, and input the group of scanning window images into the detection classifier.

The specific steps of window scanning are as follows:

In the first step, the area of the sample image size in a human training sample set in the upper left corner of the input detected image is used as the first scanning window, and the scanning window is used as the current scanning window, and the current scanning window image is saved.

In the second step, the current scanning window is shifted to the right by 8 pixels or down by 16 pixels on the detected image to obtain a new scanning window, and the current scanning window is replaced with the new scanning window, and the image of the current scanning window is saved.

The third step is to move the current scanning window according to the above method, and replace the current scanning window with the moved scanning window until the complete image to be detected is scanned, and all the scanning window images are saved.

Step 8, detecting the scan window.

8a) Use the detection classifier to judge whether the input scanning window image contains human body information, if there is no human body information, then locate the detected image as a non-human natural image, otherwise, from all the judged human body information Among the scan window images, find the scan window image with the highest detection classifier score as the main window image.

8b) From the remaining scanning window images with human body information other than the main window image, the scanning window image overlapping with the main window image by more than 50% is combined with the main window image, and the window obtained by window combination is used as a detection result Save, delete all images participating in the window composition.

The specific steps of window composition are as follows:

The first step is to sequentially number all images that require window combinations starting from 1.

In the second step, the proportion of the classifier score of each image requiring window combination in the sum of the classifier scores of all images requiring window combination is used as the weight of image boundary weighting.

In the third step, use the following formula to weight each boundary of the image that needs window combination:

$\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; \&times;</span>\frac{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}{\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; \&times;</span>\frac{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; \&times;</span>\frac{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}}}{\mathrm{<span\; class="notranslate">\; A</span>}}$

N表示参与窗口组合的图像个数，i表示窗口组合图像的编号，m_i表示第i幅参与窗口组合的图像的分类器分数。Among them, X represents the pixel value of the row or column of the window boundary on the detected image obtained after weighting, and x ₁ , x ₂ ,...x _N respectively represent the image boundaries participating in the window combination in the detected image The pixel value of the row or the pixel value of the column on the image, m ₁ , m ₂ ,...m _N respectively represent the classifier scores corresponding to the images participating in the window combination, N represents the number of images participating in the window combination, A Denotes the sum of the image classifier scores participating in the window combination,

N represents the number of images participating in the window combination, i represents the number of the window combination image, and m _i represents the classifier score of the i-th image participating in the window combination.

In the fourth step, the weighted boundaries are combined into a window.

8c) Determine whether there are any scan window images with human body information left, if so, find the image with the highest detection classifier score among the remaining scan window images as the main window image, and perform step 8b), otherwise, perform step 9.

Step 9, output the detection result.

All the windows obtained by combining the windows are marked on the detected image, and the marked image is output as the human body detection result of the detected image.

Effect of the present invention can be further illustrated by following simulation:

1. Simulation experiment condition setting:

The simulation experiment of the present invention is compiled on Matlab2009a, and the simulation environment is an HP workstation under the Windows framework. The positive sample images and negative sample images required for the experiment are taken from the INRIA database. The training samples include 2416 positive samples and 13500 negative samples. The test sample set includes 1132 positive samples and 4050 negative samples. The positive samples and negative samples The size of the image is 128×64 pixels, and Fig. 2 is a part of the sample image used in the present invention, wherein Fig. 2 (a) is a part of the positive sample image used in the present invention, and Fig. 2 (b) is a part of the sample image used in the present invention Partial negative sample images.

2. Simulation content and result analysis:

Simulation 1:

The accuracy rate obtained after the classifier training is completed is one of the important indicators for judging the performance of the classifier. In order to obtain a classifier with better performance, we have done a lot of experiments on the selection of two parameters, the number of sampling layers of the sample image and the projection dimension of the low-dimensional projection of the initial basis vector, when extracting the features of the sample image. The number of sampling layers samples the sample image, and different projection dimensions project the initial basis vector to the accuracy rate of the training classifier. The comparison results are shown in Table 1.

It can be seen from Table 1 that for the same projection dimension, the accuracy rate of the classifier when the sample image is sampled with 3 layers is higher than that when the sample image is sampled with 2 layers; while for the same number of sampling layers , it does not necessarily mean that the higher the projection dimension, the higher the accuracy of the classifier. It can be seen from the data in the figure that the accuracy of the classifier obtained when the sample image is sampled at three layers and the initial basis vector is projected to 450 dimensions is the highest, and the classification performance obtained is the best.

Simulation 2:

Using the present invention and the human body detection method based on the gradient histogram HOG feature to extract the features of the human body training sample set, train the classifier, and compare the performance of the obtained classifiers. The performance comparison diagram of the classifier is shown in Figure 3. In Figure 3, the ROC (Receiver Operating Characteristic) curve of the relationship between the true positive rate TPR (True Positive Rate) and the false positive rate FPR (False Positives Rates) is selected to evaluate the classification device performance. The higher the ROC curve is towards the top left corner, the better the corresponding classifier.

The axis of abscissa in Figure 3 represents the false positive rate FPR (False Positives Rates), and the axis of ordinate represents the true positive rate TPR (True Positive Rate). In accompanying drawing 3, represent the ROC curve of classifier true positive rate and the false positive rate relation of classifier true positive rate and false positive rate with the curve marked with box, represent the true positive rate and false positive rate of classifier based on gradient histogram HOG feature human body detection method with the curve marked with cross The ROC curve of the rate relationship. It can be seen from Figure 3 that the ROC curve obtained by the present invention is more inclined to the left top corner than the ROC curve obtained by the human body detection method based on the gradient histogram HOG feature, indicating that the classification performance of the present invention is better than that based on the gradient histogram HOG feature Classification performance of human detection methods.

Simulation 3:

Using the present invention and the human body detection method based on the gradient histogram HOG feature to detect the human body from the natural image of the INRIA database, the detection results are shown in Figure 4 and Figure 5 .

Fig. 4 is an image of uneven illumination, and image 4 (a) represents the human body detection result of the present invention, and the white box in the figure represents the result of window merging after the human body information in the detection classifier detection image of the present invention. Image 4(b) shows the result of human detection based on the gradient histogram HOG feature human detection method. The white box in the figure represents the result of window merging after the detection classifier of this method detects human information in the image. It can be seen from Fig. 4 that in the case of uneven illumination, the method of the present invention can greatly reduce the false alarm rate and more accurately detect the human body detection method based on the gradient histogram HOG feature. all human body information.

Fig. 5 is an image with complex background and people occlusion, image 5 (a) represents the human body detection result of the present invention, and the white box in the figure represents that the detection classifier of the present invention detects the human body information in the image and the window merges the result of. Image 5(b) shows the human detection result based on the gradient histogram HOG feature human detection method, and the white box in the figure represents the result of window merging after the detection classifier of this method detects the human body information in the image. As can be seen from Figure 5, in the case of complex backgrounds and people occluded, using the method of the present invention can more accurately mark human body information, and the window size obtained after window merging is larger than that of the human body detection method based on the gradient histogram HOG feature It is more suitable and has a higher accuracy rate of human detection.

To sum up, the method of the present invention can detect the human body under the conditions of uneven illumination, complex background and partial occlusion. It shows that this method is very suitable for human detection in natural images.

Claims (7)

1. the efficient coupling core human body detecting method based on the fast robust feature, comprise and obtain detecting sorter and utilize the sorter obtained to be detected two processes to image, and the specific implementation step is as follows:

First process, the concrete steps that obtain detecting sorter are as follows:

(1) select the training sample set image:

1a) utilize the bootstrapping operation, from the non-human body natural image of INRIA database, obtain enough negative sample images;

1b) the negative sample image of acquisition and the negative sample collection in the INRIA database are formed to new negative sample collection;

Positive sample set in the new negative sample collection image that 1c) will obtain and INRIA database forms the human body training sample set;

(2) extract image SURF unique point:

Every width image of 2a) the human body training sample being concentrated is divided into the grid of 8*8 pixel, to each grid, by the yardstick of 16,25,36 pixel sizes, samples respectively, and the trellised sample level of each yardstick sampling shape;

2b) to each 8*8 pixel grid, calculate after every layer of sampling the horizontal direction gradient of sampled point and the quadratic sum of vertical gradient in grid, by sampled point corresponding to gradient quadratic sum maximal value, as this pixel grid in the fast robust feature SURF of sample level unique point;

Every width image of 2c) the human body training sample being concentrated, each sample level in the fast robust feature SURF unique point of all pixel grid, choose at random 15 unique points, as the fast robust feature SURF unique point of image on sample level of human body training sample set;

(3) construct every layer of initial vector base:

Utilize the k means clustering method, to the human body training sample, concentrate the fast robust feature SURF unique point of all images on each sample level to carry out cluster, define 450 cluster centres, obtain the 450 dimension visual vocabularies of whole human body training sample set on sample level, form the initial base vector of sample level;

(4) obtain the maximum kernel Function feature of sample level:

For the initial base vector of each sample level, utilize respectively the core svd CKSVD of belt restraining to carry out dictionary learning, obtain the maximum kernel Function feature of sample level;

(5) obtain image and efficiently mate the core feature:

5a) to each sample level, press the element value of the maximum kernel Function feature of descending sort sample level, whether the element number that judges the greatest member value is 1, if, proper vector output using the maximum kernel Function feature of sample level as sample level, otherwise, the element that in the maximum kernel Function feature of sample level, element value equates with maximal value is set to zero, the proper vector output using the maximum kernel Function feature of the sample level after zero setting as sample level;

5b) proper vector of all sample level is weighted to summation, obtains all scale features, store all scale features;

5c) every row element of all scale feature vectors is averaged, on the coordinate transverse axis, corresponding average point carries out the cumulative of average number, obtain the distribution of the element average of all row of all scale feature vectors, select the feature of similar Gaussian distribution in the element distribution of mean value of all row, as the efficient coupling core feature of the fast robust characteristic of final image;

(6) classification based training:

Use the support vector machines sorter to carry out classification based training to the efficient coupling core feature of the fast robust characteristic extracted, obtain detecting sorter;

Second process, the concrete steps that the detection sorter that utilization obtains is detected image are as follows:

(7) input picture is scanned:

Input the detected image of a width, with the detected image of window scanning method scanning view picture, obtain one group of scanning window image, this group scanning window image is input to the detection sorter;

(8) detect scanning window:

8a) with detecting sorter, judge in the scanning window image of inputting whether include human body information, if there is not human body information, should be detected framing is non-human body natural image, otherwise, from all scanning window images that human body information arranged of judging, find out and detect scanning window image that the sorter mark is the highest as the main window image;

8b) from the remaining scanning window image that human body information arranged beyond the main window image, to be greater than 50% scanning window image and main window image with the main window doubling of the image and carry out the window combination operation, the window that window combination is obtained is preserved as a testing result, deletes the image of all participation window combination;

8c) judgement has the scanning window image of human body information whether to also have residue, if having, finds out in remaining scanning window image and detects image that the sorter mark is the highest as the main window image, execution step 8b), otherwise, execution step (9);

(9) output detections result:

All windows that window combination is obtained mark on detected image, and the image after output marks, as the human detection result of detected image.

2. the efficient coupling core human body detecting method based on the fast robust feature according to claim 1 is characterized in that: step 1a) concrete steps of described bootstrapping operation are as follows:

The first step, choose at random m positive sample image and n negative sample image from the INRIA database, 100≤m≤500 wherein, 100≤n≤800, and n≤m≤3n, used gradient orientation histogram HOG feature extracting method, and selected all positive and negative sample image is carried out to feature extraction, utilize the support vector machines sorter to carry out classification based training to the feature of extracting, obtain the preliminary classification device;

Second step, choose at random continuously the non-human body natural image in the INRIA database, adopt the scanning window of sample image size, 8 pixels of take from left to right are Moving Unit, 16 pixels of take from top to bottom are Moving Unit, the detected non-human body natural image of scanning view picture; Image in all scanning windows is input to the preliminary classification device and is detected, preserve the sorter scanning window image of wrong minute, open until the scanning window amount of images of wrong minute reaches a, 200≤a≤500, stop choosing non-human body natural image; The scanning window image divided from mistake, random choose b opens image, and 1/5a≤b≤1/3a forms new negative sample collection with current negative sample image;

The 3rd step, to m positive sample image and the new negative sample collection of choosing at random, carry out gradient orientation histogram HOG feature extraction, training classifier, the non-human body natural image of detection and upgrade the negative sample collection;

The 4th step, repeat the 3rd step operation, until the final training sample set after upgrading is comprised of 2416 positive sample images and 13500 negative sample images, size is 128 * 64 pixels.

3. the efficient coupling core human body detecting method based on the fast robust feature according to claim 1, it is characterized in that: the concrete steps of the described k means clustering method of step (3) are as follows:

The first step, to each sample level, random all sample images from the human body training sample set are the fast robust feature SURF of sample level unique point, choose 450 fast robust feature SURF unique points, initial cluster center as sample level, respectively by the data value of 450 initial cluster centers, as the cluster centre value of place initial cluster center;

Second step, calculate the human body training sample set and arrive the Euclidean distance of each cluster centre in all fast robust feature SURF unique points of sample level;

The 3rd step, be grouped into each fast robust feature SURF unique point of sample level in the classification at the cluster centre place nearest with self;

The 4th step, whether the data mean value that the fast robust feature SURF unique point of rear each class is sorted out in judgement equals the cluster centre value, if, carry out the 5th step, otherwise, using the data mean value of the unique point of required each class as new cluster centre value, return to second step;

The 5th step, preserve 450 cluster centre values, by 450 cluster centre values, forms column vector, using the initial base vector output on sample level as whole human body training sample set of this column vector.

4. the efficient coupling core human body detecting method based on the fast robust feature according to claim 1, it is characterized in that: the concrete steps of the described dictionary learning of step (4) are as follows:

The first step, for each sample level, project to the initial base vector of sample level on the space of one 450 dimension, by following formula, calculates, and obtains the projection vector of the initial base vector of sample level:

R=R ₁×[v ₁,...v _j...,v _N]

Wherein, R means the projection vector of the initial base vector of sample level, R ₁the initial base vector that means sample level, v _jthe vector of the projection coefficient of j the unique point that all sample images of expression human body training sample set extract on sample level, v _j=[v _1j, v _2j... v _ij..., v _mj] ^t, v _ijthe projection coefficient that means j the unique point that the human body training sample concentrates i width sample image to extract on sample level, M means the sample image number of human body training sample set, j=1,2, ..., N, N means the number of the unique point that each concentrated width sample image of human body training sample is chosen at random on sample level;

Second step, press an approximating function of following formula structure, the projection of the initial base vector on projector space gets on the approach sample level:

f(r)=argmin||r-R||

Wherein, r means the maximum kernel Function feature on sample level, and R means the initial base vector R on sample level ₁projection, || || mean 2 norms, argmin|||| means to minimize;

The 3rd step, calculate approximating function, obtains following formula and come iteration to upgrade the maximum kernel Function feature on sample level, forms the image feature representation of low-dimensional:

$r(k+1)=r\left(k\right)-\frac{\mathrm{\&eta;}}{k}\frac{\&PartialD;\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left|\right|r_j-R_1{\left({R_1}^T{R}_1\right)}^{-1}\left({R_1}^T{r}_j\right)\left|\right|}^2}{\&PartialD;r}$

Wherein, the maximum kernel Function feature on the sample level that r(k+1) the expression iteration obtains for k+1 time, k means iterations, the maximum kernel Function feature on the sample level that r(k) the expression iteration obtains for k time, η means learning rate, is a constant,

the derivative of formula in expression calculating bracket to r, r _jthe maximum kernel proper vector that means j the unique point that the concentrated all sample images of human body training sample extract on sample level, R ₁mean the initial base vector on sample level, R ₁ ^tmean the initial base vector R on sample level ₁transposed vector, setting iterations is 1000 times, the r obtained after iteration completes (1000) is as the final maximum kernel Function feature on sample level, j=1,2, ..., N, N means the quantity of the unique point at random chosen of each width sample image on sample level that the human body training sample is concentrated.

5. the efficient coupling core human body detecting method based on the fast robust feature according to claim 1, it is characterized in that: step 5b) mode of described weighted sum is as follows:

G*=G _i×A _i

Wherein, G* means all scale features, G _ithe proper vector that means each sample level, i=1,2,3, A _imean the weight that each sample level is corresponding,

w _i=1/p _i, p _ithe pixel size that means the sampling scale of each sample level, p _i={ 16,25,36}.

6. the efficient coupling core human body detecting method based on the fast robust feature according to claim 1, it is characterized in that: the concrete steps of the described window scanning method of step (7) are as follows:

The first step, using the zone of sample image size in a human body training sample set in the detected image upper left corner of input as first scanning window, using this scanning window as current scanning window, preserve current scanning window image;

Second step, by current scanning window on detected image to 8 pixels of right translation or move down 16 pixels and obtain a new scanning window, remove to replace current scanning window with new scanning window, preserve current scanning window image;

The 3rd step, mobile current scanning window as stated above, go to replace current scanning window with the scanning window after movement until scan the detected image of complete width, preserves all scanning window images.

7. the efficient coupling core human body detecting method based on the fast robust feature according to claim 1 is characterized in that: step 8b) concrete steps of described window combination operation are as follows:

The first step, by all images of window combination that need since 1 serial number;

Second step, need every width the sorter mark of the image of window combination, and the proportion accounted in the detection sorter mark sum of all images that need window combination is as the weight of image boundary weighting;

The 3rd step, utilize following formula, and every border of the image that needs window combination is weighted:

$X=x_1\&times;\frac{m_1}{A}+x_2\&times;\frac{m_2}{A}+...+x_N\&times;\frac{m_N}{A}$

Wherein, X means the pixel value of being expert on detected image of the window edge that obtains after weighting or the pixel value of column, x ₁, x ₂... x _nthe pixel value of being expert at of the image boundary that mean to participate in respectively window combination on detected image or the pixel value of column, m ₁, m ₂... m _nmean to participate in respectively the detection sorter mark corresponding to image of window combination, N means to participate in the image number of window combination, and A means to participate in the image detection sorter mark sum of window combination,

n means to participate in the image number of window combination, and i means the numbering of window combination image, m _imean that the i width participates in the detection sorter mark of the image of window combination;

The 4th step, form a window by the border after weighting.

Images