CN104298955A - Human head detection method and device - Google Patents

Abstract

The application discloses a human head detection method and device. The method comprises the steps that images to be detected are scanned according to windows with the preset dimension; the color self-similarity characteristic of each scanning window is acquired; and human head/non-human head judgment is performed on the scanning window bys utilizing a classifier. The beneficial effects of the method are that a CSS is selected to act as characteristic representation of a human head, and an experiment proves that the CSS characteristic is a robust characteristic and detection performance can be enhanced; and meanwhile, the human heads in the video images can be detected in combination with SVM classifier learning training so that the method can be applied to human body tracking, event detection and human flow statistics, and thus the method has important meaning for realization of automatic monitoring.

Description

A kind of people's head inspecting method and device

Technical field

The application relates to video monitoring and technical field of image processing, particularly relates to a kind of people's head inspecting method and a kind of number of people pick-up unit.

Background technology

In field of intelligent video surveillance, be an important research contents to the Intelligent Measurement of the number of people and statistics.But the gray difference of the number of people of different attitude is comparatively large, such as, from number of people front, side, the back side and top shooting number of people gray difference can be very large, make to realize the number of people and detect more difficult.Therefore, the feature how extracting robust is the importance affecting number of people detection and statistics accuracy rate and real-time.

Summary of the invention

The application provides a kind of people's head inspecting method and corresponding device.

According to the first aspect of the application, the application provides a kind of people's head inspecting method, comprising: treat detected image and scan by the window of preliminary dimension; Obtain the color self-similarity characteristics of each scanning window; Sorter is utilized to carry out the number of people/non-number of people judgement to scanning window.

Further, the acquisition of described color self-similarity characteristics comprises: quantization step, described image to be detected is carried out color space conversion, is converted to HSV space, obtains H, S, V tri-components, carries out uniform quantization to described H, S, V tri-components; Blocking step, divides equally described image to be detected and carries out piecemeal by pre-sizing; Histogram calculation step, calculates the color histogram of each described piecemeal in HSV space; Similarity Measure step, calculates the similarity between any two piecemeals in all described piecemeals, obtains all Similarity value; Normalization step, carries out cascade by described all Similarity value, is normalized cascade result, and result is color self-similarity feature.

Preferably, in described quantization step, described H, S, V tri-components are carried out the uniform quantization of 3*3*3; In described blocking step, described pre-sizing is 8*8 pixel size.

Preferably, in described histogram calculation step, in HSV space, after Tri linear interpolation is carried out to each pixel in each described piecemeal, then calculate color histogram.

Preferably, in described Similarity Measure step, the calculating of described similarity adopts histogram intersection method.

Preferably, in described normalization step, described normalized adopts L2-Norm to realize.

Further, described sorter is that described sample set comprises the positive sample of the number of people and number of people negative sample by extracting color self-similarity characteristics and training in advance obtains based on support vector machine from sample set.

According to the second aspect of the application, the application provides a kind of number of people pick-up unit, comprising: window scan module, scans by the window of preliminary dimension for treating detected image; Feature acquisition module, for obtaining the color self-similarity characteristics of each scanning window; Classification judging module, carries out the number of people/non-number of people judgement for utilizing sorter to scanning window.

Further, described feature acquisition module comprises: quantifying unit, for described image to be detected is carried out color space conversion, is converted to HSV space, obtains H, S, V tri-components, carries out uniform quantization to described H, S, V tri-components; Blocking unit, carries out piecemeal for being divided equally by described image to be detected by pre-sizing; Histogram calculation unit, for calculating the color histogram of each described piecemeal in HSV space; Similarity calculated, for calculating the similarity in all described piecemeals between any two piecemeals, obtains all Similarity value; Normalization unit, for described all Similarity value are carried out cascade, is normalized cascade result, and result is color self-similarity feature.

Preferably, in described histogram calculation unit, in HSV space, after Tri linear interpolation is carried out to each pixel in each described piecemeal, then calculate color histogram; In described similarity calculated, the calculating of described similarity adopts histogram intersection method; In described normalization unit, described normalized adopts L2-Norm to realize.

The beneficial effect of the application is: selected CSS as the character representation of the number of people, and experiment proves, CSS feature is a kind of feature of robust, can improve the performance detected; Meanwhile, in conjunction with SVM classifier learning training, and then the number of people in video image can be detected, may be used for human body tracking, event detection, people flow rate statistical etc., significant for realizing automatic monitoring.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet of people's head inspecting method of a kind of embodiment of the application;

Fig. 2 is the schematic flow sheet of the extraction CSS feature of a kind of embodiment of the application;

Fig. 3 is that the eigenwert of a kind of embodiment of the application and the mapping relations of HSV component represent intention;

Fig. 4 is the structural representation of the number of people pick-up unit of a kind of embodiment of the application.

Embodiment

By reference to the accompanying drawings the present invention is described in further detail below by embodiment.

Because number of people face, shoulder etc. have symmetry, even the number of people of pitching certain angle or the number of people of side, still there is great similarity in the color histogram in image between different piecemeal, therefore, the application utilizes the similarity of this color space, the statistical learning method proposed based on color self similarity (CSS, Color Self-Similarity) feature detects the number of people.In addition, this feature has unchangeability to yardstick, is conducive to multiple scale detecting fast.

Embodiment 1:

As shown in Figure 1, people's head inspecting method of the present embodiment comprises training process and testing process two parts.Wherein, training process comprises step: obtain training sample set S11, extract CSS feature S13 and sorter training S15; Testing process comprises step: scan image S21 to be detected, obtain the CSS feature S23 of each scanning window and carry out the number of people/non-number of people judgement S25 to scanning window.

For acquisition training sample set S21, obtain by video image obtain manner.Training sample set comprises the positive sample of the number of people and number of people negative sample.Wherein, the positive sample of the number of people can comprise from number of people front, side, the back side and top shooting number of people image, namely cover the true number of people image of different attitude, different hair, the different cap of band; And number of people negative sample can comprise the image that such as landscape, animal, word etc. do not comprise arbitrarily the number of people.

For scanning image S23 to be detected, can scan by the window of certain size size the image to be detected of input, first treat detected image and carry out pyramid convergent-divergent, namely the Multiscal process of a coarse resolution to smart resolution is defined, then on the image of each yardstick, carry out intensive scanning according to the candidate's number of people window with training sample set formed objects, each candidate's number of people window is input in sorter, the candidate's number of people window being identified as the number of people can be left.

Carry out the number of people/non-number of people judgement S25 for sorter training S15 with to scanning window, the correlation technique of pattern-recognition can be adopted to realize.The present embodiment adopts support vector machine (SVM, Support Vector Machine) to realize the training study of sorter, and trains the categorised decision obtained to adjudicate according to SVM in testing process.SVM is a kind of conventional machine learning method, and its core concept is, at the inseparable data set of lower dimension Linear, a kind of nonlinear mapping algorithm can be used to be translated into more high-dimensional space, data set can be divided.It needs to seek optimum classifying face, and not only requirement can separate faultless for two class samples, also requires that the class interval of two class samples is maximum.Concrete training method can realize based on existing SVM theory.During specific implementation, can realize training the SVM of CSS feature according to existing SVM knowwhy coding, also directly can call existing SVM storehouse and carry out SVM training, such as use the SVM tool box of Taiwan woods intelligence benevolence (Chih-Jen Lin) professor's exploitation, thus the support vector machine, threshold value etc. that can obtain required for scanning window judgement, namely optimum classifying face, then in testing process, carries out the judgement of the number of people/non-number of people to scanning window.If court verdict is just, then this window comprises the number of people, otherwise is inhuman head region.Treat detected image due to whole testing process and carry out pyramid convergent-divergent, the detection window that a lot of are classified as the number of people can be obtained, mean-shift algorithm is used to merge these windows, by multiple in position with size on can be considered to same destination object window merge become a window, finally testing result is shown.Research shows, the feature of robust can improve the performance of detection.Based on this, the present embodiment propose extraction CSS feature flow process as shown in Figure 2, correspondingly, the CSS characteristic S23 of each scanning window of acquisition that can realize in testing process shown in Fig. 1 according to this flow process.

Usually in classification problem, different sample sizes is selected according to the difference of class, such as, the training picture of 24*24 pixel size is selected in Face datection frequently-used data storehouse, in pedestrian detection, INRIA database selects the training picture of 64*128 pixel size, and in the number of people detects, there is no general database, by the number of people database of camera shooting for training and detect.Here the training sample size detected for the number of people is described for 40*40 pixel, certainly can adopt other pixel size in other embodiment, not be restricted this.CSS feature extraction specifically comprises the steps S131 ~ S139:

Step S131, H, each component uniform quantization of S, V.

For the image to be detected of input, first carry out color space conversion, by image from RGB(Red, Green, Blue) space transforming to HSV(Hue, Saturation, Value) space, obtain H, S, V tri-passages.In HSV space, H, S, V tri-components are carried out uniform quantization.Uniform quantization refers to and quantized interval is divided into some parts.Because the maximin of each passage of H, S, V is determined, after the method for selected uniform quantization, quantized value can be determined.In embodiment, in HSV space, H, S, V tri-components are carried out the uniform quantization of 3*3*3, be divided into 3 parts by each passage.Can call OpenCV storehouse (Open Source Computer Vision Library) when realizing to realize, the concrete quantized value obtained is as follows:

$H=\left\{\begin{array}{c}0,h\∈\left[\mathrm{0,60}\right)\\ 1,h\∈\left[\mathrm{60,120}\right)\\ 2,h\∈\left[\mathrm{0,180}\right]\end{array}\right.$ $S=\left\{\begin{array}{c}0,s\∈\left[\mathrm{0,85}\right)\\ 1,s\∈\left[\mathrm{85,170}\right)\\ 2,s\∈\left[\mathrm{170,225}\right)\end{array}\right.$ $V=\left\{\begin{array}{c}0,v\∈\left[\mathrm{0,85}\right)\\ 1,v\∈\left[\mathrm{85,170}\right)\\ 2,v\∈\left[\mathrm{170,255}\right)\end{array}\right.$

Can find out, each passage of H, S and V of pixel will drop on 0,1 or 2 any one interval, and therefore, the total 3*3*3=27 kind combination of the interval distribution that pixel value may drop on, uses eigenwert k=0,1,2 respectively ..., 26 represent.The mapping of definition k=f (H, S, V), can obtain mapping relations table as shown in Figure 3.Therefore under this quantification manner, the dimension of the color histogram of HSV space has 27 dimensions, represents respectively with H (k).

Step S133, image block.

What consider due to color self-similarity characteristics is similarity between different block in color distribution, as two eye portion, hair portion, shoulders etc. have great color self-similarity, the very large piece of similarity as 20*20 is too coarse, be unfavorable for the characteristic of division extracting robust, very little block is then not enough to characterize the representational part of the number of people as 2*2, is unfavorable for distinguishing with other class.Therefore, the block of the pixel sizes such as 8*8,10*10 is selected in general consideration.For 8*8 in the present embodiment, image uniform is divided into the block of 8*8 pixel, totally 25 blocks.

Step S135, adds up the color histogram H of each piecemeal in HSV space _p(k), wherein p is the sequence of block, p=1,2 ..., 25.

In a kind of embodiment, this step detailed process is: carry out Tri linear interpolation to each pixel in each piecemeal, is equivalent to carry out Tri linear interpolation to the value of H, S, V on three passages respectively, asks for the contribution of each pixel to color histogram.Use h ₀=0, h ₁=60, h ₂=120, h ₃the border of=180 expression H passage quantized intervals, in like manner uses s ₀, s ₁, s ₂, s ₃represent the border of channel S quantized interval, v ₀, v ₁, v ₂, v ₃represent the border of V passage quantized interval; Use h _m0=30, h _m1=90, h _m2the intermediate value of=150 expression H passage quantized intervals, in like manner uses s _m0, s _m1, s _m2represent the intermediate value of channel S quantized interval, v _m0, v _m1, v _m2represent the intermediate value of V passage quantized interval; Use h _lenthe width of=60 expression H passage quantized intervals, in like manner obtains the width s of the quantized interval of channel S and V passage _len=v _len=85; Use h _w[3] represent the weight to its three quantized interval contributions in H channel linear Interpolation Process, in like manner use s _w[3], v _w[3] weight that channel S and V passage are contributed respective quantized interval is represented; Then, for each pixel and h, s, v value thereof, its Tri linear interpolation process can by following pseudo-representation:

if(h>=h ₀&&h<h _m0)

h _w[0]=1.0;

else?if(h>=h _m0&&h<h _m1)

h _w[0]=(h _m1-h)/h _len;

h _w[1]=1-h _w[0];

else?if(h>=h _m1&&h<h _m2)

h _w[1]=(h _m2-h)/h _len;

h _w[2]=1-h _w[1];

else?if(h>=h _m2&&h<=h ₃)

h _w[2]=1.0;

In like manner, s can be obtained _w[3] and v _w[3] value.

Then this pixel to the contribution of each color histogram is:

H[k]=H[f(i,j,m)]=H[k]+h _w[i]×s _w[j]×v _w[m]

Wherein, the span of i, j, m is { 0,1,2}.

Step S137, adds up the similarity between any two blocks.

The calculating of similarity can adopt similarity algorithm conventional in image processing techniques to realize, such as Euclidean distance method, correlation coefficient process and histogram intersection method etc.As adopted Euclidean distance method, its Euclidean distance is less, represents that image is more similar; As adopted correlation coefficient process, its related coefficient is generally arranged between-1 and 1, and-1 represents completely uncorrelated, and 1 represents identical.Here, embodiment employing histogram intersection method calculates the similarity S (x, y) between any two piecemeals, and formula is as follows, and wherein the implication of each parameter is with aforementioned.

$S(x,y)=\frac{\underset{k=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\min (H_x\left(k\right),H_y\left(k\right))}{\underset{k=0}{\overset{M-1}{\mathrm{\&Sigma;}}}{H}_x\left(k\right)}$

Wherein, x and y represents two piecemeals respectively, H _x(k) and H _yk () represents the color histogram of these two piecemeals respectively, combination when M is this three components of uniform quantization H, S, V, and as aforementioned, if carry out the uniform quantization of 3*3*3, now, above formula can be written as follows:

$S(x,y)=\frac{\underset{k=0}{\overset{26}{\mathrm{\&Sigma;}}}\min (H_x\left(k\right),H_y\left(k\right))}{\underset{k=0}{\overset{26}{\mathrm{\&Sigma;}}}{H}_x\left(k\right)}$

All piecemeals are all calculated, can obtain altogether individual Similarity value.

Step S139, by all Similarity value cascades and normalization.

In this step, 300 Similarity value obtained by step S137 carry out cascade (cascade), are connected into the vector of one 300 dimension by 300 one-dimensional vector.Then color self-similarity feature is after being normalized.The mode of normalized can be the normalization mode that L1-Norm, L1-Sqrt and L2-Norm etc. are conventional.L2-Norm normalization mode is selected in embodiment.

People's head inspecting method that the present embodiment proposes, it has selected CSS as the character representation of the number of people, and experiment proves, CSS feature is a kind of feature of robust, can improve the performance detected; Meanwhile, in conjunction with SVM classifier learning training, and then the number of people in video image can be detected, may be used for human body tracking, event detection, significant for realizing automatic monitoring.

Embodiment 2:

As shown in Figure 4, the present embodiment provides a kind of number of people pick-up unit, comprising: for treat window scan module that detected image undertaken scanning by the window of preliminary dimension, for obtain the color self-similarity characteristics of each scanning window feature acquisition module and for utilizing sorter to carry out the classification judging module of the number of people/non-number of people judgement to scanning window.

Wherein, feature acquisition module comprises: quantifying unit, blocking unit, histogram calculation unit, similarity calculated and normalization unit.Quantifying unit is used for image to be detected to carry out color space conversion, is converted to HSV space, obtains H, S, V tri-components, carry out uniform quantization to H, S, V tri-components; Blocking unit is used for image to be detected to divide equally to carry out piecemeal by pre-sizing; Histogram calculation unit is for calculating the color histogram of each piecemeal in HSV space; Similarity calculated, for calculating the similarity in all piecemeals between any two piecemeals, obtains all Similarity value; Normalization unit is used for all Similarity value to carry out cascade, and be normalized cascade result, result is color self-similarity feature.Preferably, in histogram calculation unit, in HSV space, after Tri linear interpolation is carried out to each pixel in each piecemeal, then calculate color histogram; In similarity calculated, the calculating of similarity adopts histogram intersection method; In normalization unit, normalized adopts L2-Norm to realize.

The realization of each module and unit can associated description in reference example 1 above, does not repeat at this.

It will be appreciated by those skilled in the art that, in above-mentioned embodiment, all or part of step of various method can be carried out instruction related hardware by program and completes, this program can be stored in a computer-readable recording medium, and storage medium can comprise: ROM (read-only memory), random access memory, disk or CD etc.

Above content is in conjunction with concrete embodiment further description made for the present invention, can not assert that specific embodiment of the invention is confined to these explanations.For general technical staff of the technical field of the invention, without departing from the inventive concept of the premise, some simple deduction or replace can also be made.

Claims (10)

1. people's head inspecting method, is characterized in that, comprising:

Treat detected image to scan by the window of preliminary dimension;

Obtain the color self-similarity characteristics of each scanning window;

Sorter is utilized to carry out the number of people/non-number of people judgement to scanning window.

2. people's head inspecting method as claimed in claim 1, it is characterized in that, the acquisition of described color self-similarity characteristics comprises:

Quantization step, carries out color space conversion by described image to be detected, is converted to HSV space, obtains H, S, V tri-components, carries out uniform quantization to described H, S, V tri-components;

Blocking step, divides equally described image to be detected and carries out piecemeal by pre-sizing;

Histogram calculation step, calculates the color histogram of each described piecemeal in HSV space;

Similarity Measure step, calculates the similarity between any two piecemeals in all described piecemeals, obtains all Similarity value;

Normalization step, carries out cascade by described all Similarity value, is normalized cascade result, and result is color self-similarity feature.

3. people's head inspecting method as claimed in claim 2, is characterized in that,

In described quantization step, described H, S, V tri-components are carried out the uniform quantization of 3*3*3;

In described blocking step, described pre-sizing is 8*8 pixel size.

4. people's head inspecting method as claimed in claim 2, is characterized in that, in described histogram calculation step, in HSV space, after carrying out Tri linear interpolation, then calculate color histogram to each pixel in each described piecemeal.

5. people's head inspecting method as claimed in claim 2, is characterized in that, in described Similarity Measure step, the calculating of described similarity adopts histogram intersection method, and formula is:

$S(x,y)=\frac{\underset{k=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\min (H_x\left(k\right),H_y\left(k\right))}{\underset{k=0}{\overset{M-1}{\mathrm{\&Sigma;}}}{H}_x\left(k\right)}$

Wherein, S (x, y) represents the similarity between two piecemeal x and y, H _x(k) and H _yk () represents the color histogram of x and y two piecemeals respectively, combination when M is this three components of uniform quantization H, S, V, k is integer.

6. people's head inspecting method as claimed in claim 2, is characterized in that, in described normalization step, described normalized adopts L2-Norm to realize.

7. people's head inspecting method as claimed in claim 1 or 2, it is characterized in that, described sorter is that described sample set comprises the positive sample of the number of people and number of people negative sample by extracting color self-similarity characteristics and training in advance obtains based on support vector machine from sample set.

8. a number of people pick-up unit, is characterized in that, comprising:

Window scan module, scans by the window of preliminary dimension for treating detected image;

Feature acquisition module, for obtaining the color self-similarity characteristics of each scanning window;

Classification judging module, carries out the number of people/non-number of people judgement for utilizing sorter to scanning window.

9. number of people pick-up unit as claimed in claim 8, it is characterized in that, described feature acquisition module comprises:

Quantifying unit, for described image to be detected is carried out color space conversion, is converted to HSV space, obtains H, S, V tri-components, carries out uniform quantization to described H, S, V tri-components;

Blocking unit, carries out piecemeal for being divided equally by described image to be detected by pre-sizing;

Histogram calculation unit, for calculating the color histogram of each described piecemeal in HSV space;

Similarity calculated, for calculating the similarity in all described piecemeals between any two piecemeals, obtains all Similarity value;

Normalization unit, for described all Similarity value are carried out cascade, is normalized cascade result, and result is color self-similarity feature.

10. number of people pick-up unit as claimed in claim 9, is characterized in that,

In described histogram calculation unit, in HSV space, after Tri linear interpolation is carried out to each pixel in each described piecemeal, then calculate color histogram;

In described similarity calculated, the calculating of described similarity adopts histogram intersection method;

In described normalization unit, described normalized adopts L2-Norm to realize.