CN101645135A - Method and system for human detection - Google Patents

Abstract

The invention provides a method and a system for human detection. The method comprises the following steps: extracting all the edge orientation histogram (EOH) features of a first rectangular region with the width thereof being a first width and the height thereof being the first height from a sample image; training according to all the EOH features of the first rectangular region by using a classifier training algorithm, to obtain a human classifier; scaling the original image to be detected to images to be detected on a series of scales according to scaling factors; extracting all the EOH features of a second rectangular region with the width thereof being the first width and the height thereof being the first height from the images to be detected on a series of scales; judging all the EOH features of the second rectangular region by the human classifier obtained from training, so as to detect the candidate region of human positions in the original image to be detected; and carryingout the post-treatment on the candidate region of the human positions in the original image to be detected, so as to detect the region of the human position in the original image to be detected.

Description

A kind of human body detecting method and system

Technical field

The present invention relates to area of pattern recognition, relate in particular to a kind of human body detecting method and system.

Background technology

Human detection is an important subject of computer vision field, as the front end procedure of intelligent video analysis, it has important use and is worth at aspects such as intelligent monitoring, driver assistance system, Content-based Video Retrieval and senior man-machine interfaces.The main target of human detection is exactly in still image or sport video, realizes the automatic detection of human body target.According to the mode of utilizing of human body knowledge, Human Detection is divided into two classes: based on the human detection of feature with based on the human detection of pattern classification.Human Detection based on pattern classification is to utilize the knowledge of method automatic study manikin from sample of statistical learning, human detection is regarded as the problem of a pattern classification.

Existing Human Detection based on pattern classification is with the method best results based on support vector machine (SVM, supportvector machine) and histogram of gradients (HOG, histogram of gradient).Support vector machine is a kind of widely used sorter, and this sorter has preferable generalization ability.But an important defective of this method is that speed is too slow, can't reach the requirement of actual use.The slow-footed most important reason of this method is because in order to detect the human body of different sizes and diverse location, need judge to image zooming and to each position, owing to be in the HOG feature difference of the window of diverse location, thereby need repeatedly calculate the HOG feature, thereby make algorithm speed very slow.

Summary of the invention

Thereby the present invention is directed to and repeatedly to calculate same category feature based on windowhood method in the prior art of human detection and cause the big and high deficiency of computational complexity of operand, a kind of human body detecting method and system are provided.

In a first aspect of the present invention, a kind of human body detecting method is provided, comprise step: extracting in the sample image all wide is first width, and height is the first rectangular area edge orientation histogram feature of first height; According to described all first rectangular area edge orientation histogram features, utilize the sorter training algorithm, training obtains the human body sorter; According to zoom factor is original image zoom to be detected the image to be detected of a series of yardsticks; Extracting in the image to be detected of described a series of yardsticks all wide is first width, and height is the second rectangular area edge orientation histogram feature of first height; The described human body sorter that utilizes training to obtain is judged the edge orientation histogram feature of described all second rectangular areas, detects the position of human body candidate region in the described original image to be detected; Aftertreatment is carried out in described position of human body candidate region, detect the position of human body zone in the described original image to be detected.

Further, in the described extraction sample image all wide be first width, height be first the height the first rectangular area edge orientation histogram feature comprise: with described wide be first width, height is that the rectangular area of first height travels through described sample image in level and vertical direction, division obtains described all first rectangular areas, it is identical that described each first rectangular area is divided into four sizes, wide is second width, height is second height, the 3rd rectangular area that level and vertical direction are adjacent in twos; Adopt the computing method of edge orientation histogram feature, obtain the edge orientation histogram feature of described all the 3rd rectangular areas, the edge orientation histogram feature of four the 3rd rectangular areas in described each first rectangular area head and the tail are linked to each other successively, obtain the proper vector of described each first rectangular area, the proper vector of described each first rectangular area as described each first rectangular area edge orientation histogram feature, is perhaps carried out normalization to described proper vector and obtained described each first rectangular area edge orientation histogram feature.

Further, in the image to be detected of a series of yardsticks of described extraction all wide be first width, height be first the height the second rectangular area edge orientation histogram feature comprise: with described wide be first width, height is that the rectangular area of first height travels through the image to be detected of described a series of yardsticks in level and vertical direction, divides and obtains described all second rectangular areas; It is identical that described each second rectangular area is divided into four sizes, and wide is second width, and height is second height, the 4th rectangular area that level and vertical direction are adjacent in twos; Adopt the computing method of edge orientation histogram feature, obtain the edge orientation histogram feature of described all the 4th rectangular areas, the edge orientation histogram feature of four the 4th rectangular areas in described each second rectangular area head and the tail are linked to each other successively, obtain the proper vector of described each second rectangular area, the proper vector of described each second rectangular area as described each second rectangular area edge orientation histogram feature, is perhaps carried out normalization to described proper vector and obtained described each second rectangular area edge orientation histogram feature.

In a second aspect of the present invention, a kind of detecting system of human body is provided, comprising: image zoom module to be detected, being used for according to zoom factor is original image zoom to be detected the image to be detected of a series of yardsticks; The edge orientation histogram characteristic extracting module, being used for extracting the input sample image wide is first width, height is the edge orientation histogram feature of all first rectangular areas of first height, and the image to be detected of a series of yardsticks that is used for extracting described image zoom module output to be detected wide be first width, height is the edge orientation histogram feature of all second rectangular areas of first height; Human body sorter training module is used for the edge orientation histogram feature according to all first rectangular areas of described edge orientation histogram characteristic extracting module output, utilizes the sorter training algorithm, and training obtains the human body sorter; The human body sort module, be used to utilize the human body sorter of described human body sorter training module output that the edge orientation histogram feature of all second rectangular areas of described edge orientation histogram characteristic extracting module output is judged, detect the position of human body candidate region in the described original image to be detected; Human body classification post-processing module is used for aftertreatment is carried out in the position of human body candidate region of described human body sort module output, detects the position of human body zone in the described original image to be detected.

Further, described edge orientation histogram characteristic extracting module further comprises: module is divided in the rectangular area, be used for described wide be first width, height is that the rectangular area of first height travels through described sample image and a series of image to be detected in level and vertical direction, divide respectively and obtain described all first rectangular areas and second rectangular area, it is identical that described each first rectangular area and each second rectangular area are divided into four sizes respectively, wide is second width, height is second height, the 3rd rectangular area that level and vertical direction are adjacent in twos and the 4th rectangular area; Edge orientation histogram feature calculation module, be used to calculate described rectangular area and divide all the 3rd rectangular areas of module output and the edge orientation histogram feature of all the 4th rectangular areas, and the edge orientation histogram feature of four the 3rd rectangular areas in described each first rectangular area head and the tail are linked to each other successively, obtain the proper vector of described each first rectangular area, the proper vector of described each first rectangular area as described each first rectangular area edge orientation histogram feature, perhaps the proper vector of described each first rectangular area is carried out normalization and obtain described each first rectangular area edge orientation histogram feature, and it is continuous successively the edge orientation histogram feature of four the 4th rectangular areas in described each second rectangular area head and the tail, obtain the proper vector of described each second rectangular area, the proper vector of described each second rectangular area as described each second rectangular area edge orientation histogram feature, is perhaps carried out normalization to the proper vector of described each second rectangular area and obtained described each second rectangular area edge orientation histogram feature.

A kind of human body detecting method provided by the invention and system adopt the HOG feature based on all positions on the thinking computed image of feature, and the HOG feature to obtaining adopts the model that trains to judge then.Owing to only need calculate the HOG feature one time for image, thereby can reduce operand, raising speed.

Description of drawings

Below in conjunction with accompanying drawing the specific embodiment of the present invention is described in further detail, in the accompanying drawing:

Fig. 1 is the process flow diagram of the human body detecting method of one embodiment of the invention;

Fig. 2 is the synoptic diagram of the no symbol discretize edge direction of 6 directions in one embodiment of the invention;

Fig. 3 is the structured flowchart of the detecting system of human body of one embodiment of the invention.

Embodiment

Human body contour outline in the image and background intersection have strong edge, and existing human detection algorithm is verified, and the edge of human body contour outline and background intersection is the important information that carries out human detection; And the gradient of single pixel, very strong expression ability is in addition, relatively inresponsive for the variation of noise and different people human body attitude outward appearance yet.The present invention adopts the normalization edge direction histogram of gradients (HOG, histogram of gradient) of rectangular area to distribute and makes up sorter as feature.

For fear of edge calculation direction histogram feature repeatedly, the present invention only adopts a kind of characterizing definition mode of size area, like this, image is at first calculated the edge orientation histogram feature one time, when then extracting the edge orientation histogram feature of rectangular window of each position, be no longer necessary for each rectangular window and calculate an edge orientation histogram feature separately.Iff the edge orientation histogram feature that adopts a rectangular area, owing to do not utilize the edge orientation histogram feature of adjacent area, the effect that obtains not is fine.The present invention adopts based on four edges of regions direction histogram normalization features as final rectangular area HOG characterizing definition: it is identical that the rectangular area on the image is divided into four sizes, and the less rectangular area that level and vertical direction are adjacent in twos, respectively edge orientation histogram is asked in these four less rectangular areas, the edge orientation histogram head and the tail of then four less rectangular areas being tried to achieve connect successively, obtain a proper vector, this proper vector as rectangular area edge orientation histogram feature, perhaps adopt 2 norms or 1 norm to carry out normalization, obtain rectangular area edge orientation histogram feature this vector.

Fig. 1 is the process flow diagram of the human body detecting method of one embodiment of the invention.As shown in Figure 1, the method for human detection comprises:

101) extracting in the sample image all wide is first width, and height is the first rectangular area edge orientation histogram feature of first height.

This step further comprises step:

1011) with wide be first width, height is that the rectangular area of first height travels through sample image in level and vertical direction, division obtains all first rectangular areas, it is identical that each first rectangular area is divided into four sizes, wide is second width, height is second height, the 3rd rectangular area that level and vertical direction are adjacent in twos.

The height of supposing manikin is H _Sample, width is W _Sample, extract the image zooming that comprises human region to the model size as positive sample, extract do not comprise human region image zooming to the model size as anti-sample, then with wide be 2W _r, height is 2H _rThe rectangular area wide in level and vertical direction traversal be W _Sample, height is H _SamplePositive and negative sample image.Each the wide 2W of being that obtains in the positive and negative sample image _r, height is 2H _rFirst rectangular area, it is identical to be divided into four sizes, wide is W _r, height is H _r, and level three rectangular area adjacent in twos with vertical direction.

Above-mentioned W _rAnd H _rBe to set constant, need satisfy 2W _r≤ W _Sample, and 2H _r≤ H _SampleThis is wide to be W _Sample, height is H _SampleSample image in comprise (H altogether _Sample-2H _r+ 1) * (W _Sample-2W _r+ 1) individual first rectangular area.Claim 2W among the present invention _rBe first width, 2H _rBe first height, claim W _rBe second width, H _rIt is second height.

1012) computing method of employing edge orientation histogram feature, obtain the edge orientation histogram feature of all the 3rd rectangular areas, the edge orientation histogram feature of four the 3rd rectangular areas in each first rectangular area head and the tail are linked to each other successively, obtain the proper vector of each first rectangular area, the proper vector of each first rectangular area as each first rectangular area edge orientation histogram feature, is perhaps carried out normalization to the proper vector of each first rectangular area and obtained each first rectangular area edge orientation histogram feature.

In this step, adopt the computing method of edge orientation histogram feature, calculate the wide W of being _Sample, height is H _SampleSample image in all wide be W _r, height is H _rThe edge orientation histogram feature of the 3rd rectangular area, each wide be 2W _r, height is 2H _rFirst rectangular area in four wide be W _r, height is H _rThe edge orientation histogram feature head and the tail of the 3rd rectangular area link to each other successively, obtain the proper vector of each first rectangular area, the proper vector of each first rectangular area as each first rectangular area edge orientation histogram feature, is perhaps carried out normalization to the proper vector of each first rectangular area and obtained each first rectangular area edge orientation histogram feature.

Further, method for normalizing is to adopt each value of the proper vector of each first rectangular area all divided by 1 norm or 2 norms of this proper vector, thereby obtains each first rectangular area image normalization edge orientation histogram feature.

The step of the computing method of above-mentioned edge orientation histogram feature comprises:

10121) horizontal edge and the vertical edge of each pixel of calculating sample image.

Calculating wide is W _Sample, height is H _SampleThe horizontal edge and the vertical edge of each pixel of sample image.The acquiring method at pixel edge has a variety of, and is commonly used as sobel or prewitt operator:

$S1=\left[\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right]$ $S2=\left[\begin{array}{ccc}-1& -2& -1\\ 0& 0& 0\\ 1& 2& 1\end{array}\right]$

The matrix S 1 in left side is the detection template of sobel horizontal direction, this matrix is used for the horizontal edge of computing center's place's elements A (i.e. the element of the 2nd row the 2nd row), represent with EH (A), the matrix S 2 on right side is the detection template of sobel vertical direction, be used to calculate the vertical edge that A is ordered, with EV (A) expression.

10122), obtain the edge strength and the discretize edge direction of each pixel of sample image according to the horizontal edge and the vertical edge of each pixel of sample image.

According to step 10121) in the horizontal edge EH and the vertical edge EV of each pixel of calculating, further calculate the edge direction of each pixel, represent with ED, and edge strength, represent with EI.Adopt signless edge direction definition, promptly the scope of edge direction is 0～180 degree, and the edge direction that differ 180 degree this moment is same direction.Signless edge direction is carried out discretize, and the scope that is about to 180 degree is divided into N interval.Fig. 2 is the synoptic diagram of the no symbol discretize edge direction of 6 directions in one embodiment of the invention, i.e. the situation of N=6, and at this moment, edge direction belongs to the pixel in the same interval range, and the value of their discretize edge directions is identical, represents with NED.

The account form of edge strength has multiple, for example $\mathrm{EI}=\sqrt{{\mathrm{EH}}^2+{\mathrm{EV}}^2}$ Or EI=|EH|+|EV|, as shown in Figure 2, the no symbol edge direction of pixel $\mathrm{ED}=\mathrm{arccot}\left(\frac{\mathrm{EV}}{\mathrm{EH}}\right)$ , the discretize edge direction of this pixel then , wherein the arccot function is the inverse function of cotangent function.If adopt above-mentioned definition to calculate ED earlier, calculate the value of NED again by ED, owing to have evolution and triangulo operation, computing velocity is slower, can adopt following method to calculate the NED value of each pixel fast:

I) if EH is 0, then setting at this moment, NED is 0; Otherwise, carry out ii) step;

Ii) initialization i=0 calculates

Value;

If iii)

Smaller or equal to

Value, termination process; Otherwise, carry out iv) step;

Iv) i increases by 1, if this moment i＜N-1, then get back to iii) step, otherwise termination process.

The value of the i that obtains is the value of this pixel NED.Obviously, owing to only need carry out the calculating of simple algebraic operation and cotangent function in the said process, so computing velocity is carried out Calculation Method faster than adopting according to the definition of NED greatly.

By said method, traveling through wide is W _Sample, height is H _SampleSample image in all pixels, can obtain the edge strength and the discretize edge direction of each pixel.

10123) according to the edge strength and the discretize edge direction of each pixel of sample image, in the computed image all wide be second width, height is the edge orientation histogram feature of the rectangular area of second height.

According to wide be W _Sample, height is H _SampleThe edge strength and the discretize edge direction of each pixel of sample image, calculating in the sample image all wide is W _r, height is H _rThe edge orientation histogram feature of the 3rd rectangular area.

The definition of i value of the edge orientation histogram feature of rectangular area R is: in this zone all discretize edge direction values be i pixel edge strength accumulation and, be formulated as:

i＝0，1，…N-1

Wherein, (promptly coordinate is (x, pixel y) to P for x, the y) pixel that x is listed as, y is capable in the expression region R.

When the edge orientation histogram feature in compute matrix zone, all the method that adopts the multidimensional integrals image in the prior art, the present invention discloses a kind of computing method, and this method arithmetic speed is also very fast, and simultaneously, this method is saved a lot of internal memories.

Adopting in this method computed image all wide is second width, and height is that the step of edge orientation histogram feature of the 3rd rectangular area of second height is:

101231) calculating left upper apex wide in all of image first row is 1, and height is the edge orientation histogram feature of the rectangular area of second height.

Calculate and preserve that left upper apex is (l, 0) in the sample image by the edge orientation histogram characterizing definition of above-mentioned rectangular area, wide is 1, and height is H _rW _SampleIndividual rectangular area R (l, 0,1, H _r) the edge orientation histogram feature: HOG (R (l, 0,1, H _r)), l=0 wherein, 1 ..., W _Sample-1;

101232) calculating left upper apex other capable all image removes first row wide is 1, and height is the edge orientation histogram feature of the rectangular area of second height.

In this sample image, according to left upper apex be (l, t), wide is 1, height is H _rRectangular area R (l, t, 1, H _r) edge orientation histogram feature HOG (R (l, t, 1, H _r)), iterative computation obtain left upper apex be (l, t+1) wide is 1, height is H _rRectangular area R (l, t+1,1, H _r) the edge orientation histogram feature: HOG (R (l, t+1,1, H _r)), l=0 wherein, 1 ... W _Sample-1, t=0,1 ... H _Sample-H _r-1.Concrete computing method are: HOG (R (l, t, 1, H _r)) and HOG (R (l, t+1,1, H _r)) difference be, pixel that the former is many (l, edge intensity value computing t), and the latter many pixel (l, t+H _r) edge intensity value computing.Suppose pixel (l, t+H _r) the discretize edge direction be i=NED (l, t+H _r), pixel (l, discretize edge direction t) be j=NED (l, t), then with HOG (R (l, t, 1, H _r)) i value add pixel (l, t+H _r) edge intensity value computing EI (l, t+H _r), with HOG (R (l, t, 1, H _r)) j value deduct pixel (l, (l t), and keep HOG (R (l, t, 1, H to edge intensity value computing EI t) _r)) edge intensity value computing on other discretize edge direction is constant, just obtains HOG (R (l, t+1,1, H _r)), l=0 wherein, 1 ... W _Sample-1, t=0,1 ... H _Sample-H _r-1.

101233) calculating left upper apex wide in all of image first row is second width, and height is the edge orientation histogram feature of the rectangular area of second height.

In this sample image, according to left upper apex be (l, t), wide is 1, height is H _rRectangular area R (l, t, 1, H _r) edge orientation histogram feature HOG (R (l, t, 1, H _r)), l=0 wherein, 1 ..., W _r-1, t=0,1 ..., H _Sample-H _r, calculating left upper apex is that (0, t), wide is W _r, height is H _rThe 3rd rectangular area R (0, t, W _r, H _r) the edge orientation histogram feature:

$\mathrm{HOG}\left(R(0,t,W_r,H_r)\right)=\underset{0\≤l\≤W_r-1}{\mathrm{\Σ}}\mathrm{HOG}\left(R(l,t,1,H_r)\right)$

T=0 wherein, 1 ..., H _Sample-H _r

101234) calculating left upper apex all of other row image removes first row wide is second width, and height is second highly the edge orientation histogram feature of rectangular area.

In this sample image, be that (l, t), wide is W according to left upper apex _r, height is H _rThe 3rd rectangular area R (l, t, W _r, H _r) edge orientation histogram feature HOG (R (l, t, W _r, H _r)) and left upper apex be (l, t), wide is 1, height is H _rRectangular area R (l, t, 1, H _r) edge orientation histogram feature HOG (R (l, t, 1, H _r)), it is that (l+1, t), wide is W that iterative computation obtains left upper apex _r, height is H _rThe 3rd matrix area R (l+1, t, W _r, H _r) edge orientation histogram feature HOG (R (l+1, t, W _r, H _r)=HOG (R (l, t, W _r, H _r))+HOG (R (l+W _r, t, 1, H _r))-HOG (R (l, t, 1, H _r)) l=0 wherein, 1 ... W _Sample-W _r-1, t=0,1 ... H _Sample-H _r

The foregoing description provide the 101232nd), 101233) and 101234 the step all be quick calculation method, not necessary in an embodiment entirely, also can adopt the definition of the discretize edge orientation histogram feature of rectangular area to replace wherein any step, thereby calculate the discretize edge orientation histogram feature of rectangular area.Such as adopting the 101231st among the embodiment) step, adopt the definition of the discretize edge orientation histogram feature of rectangular area calculate left upper apex be (l+1, t), wide is 1, height is H _rRectangular area R (l, t+1,1, H _r) edge orientation histogram feature HOG (R (l, t+1,1, H _r)), l=0 wherein, 1 ..., W _r-1, t=0,1 ..., H _Sample-H _r, and then adopt the 101233rd) step and 101234) step, calculate the respectively HOG feature of the 3rd rectangular area.Also can adopt the definition of the discretize edge orientation histogram feature of rectangular area to replace the 101233rd) and 101234) step.

103) obtain all first rectangular area edge orientation histogram features of sample image after, according to all first rectangular area edge orientation histogram features, utilize the sorter training algorithm, training obtains the human body sorter.

In an embodiment of the present invention, the sorter training algorithm is at first to adopt adaboost (adaptive boosting) that all first rectangular area edge orientation histogram features are carried out feature selecting, the criterion of selecting can be defined as in all features, make weighted error and minimum feature, then with all selecteed first rectangular area edge orientation histogram features as feature, adopt svm classifier device training algorithm to train and obtain final sorter.

In another embodiment of the present invention, the sorter training algorithm is the feature of the edge orientation histogram feature of all first rectangular areas as SVM, adopt svm classifier device training algorithm to train, obtain sub-svm classifier device, with the output of this sub-svm classifier device as feature, and adopt level type adaboost algorithm to train, finally obtain the adaboost human body sorter of level type.

105) be original image zoom to be detected the image to be detected of a series of yardsticks according to zoom factor.

According to zoom factor S wide be W _{To be measured}, height is H _{To be measured}Image zoom original to be detected be M wide be W _{To be measured}* S ^m, height is H _{To be measured}* S ^mThe image to be detected of a series of yardsticks, form pyramid image to be detected, m=0 wherein, 1 ... M-1, S are the scaling factors, can be taken as the number less than 1, such as being taken as 0.8.M is a positive integer, and its value satisfies W _{To be measured}* S ^M-1〉=W _SampleAnd H _{To be measured}* S ^M-1〉=H _SampleGet final product.

107) extracting in the image to be detected of a series of yardsticks all wide is first width, and height is the second rectangular area edge orientation histogram feature of first height.

This step further comprises:

1071) with wide be first width, height is that the rectangular area of first height travels through the image to be detected of a series of yardsticks in level and vertical direction, divides and obtains all second rectangular areas; It is identical that each second rectangular area is divided into four sizes, and wide is second width, and height is second height, the 4th rectangular area that level and vertical direction are adjacent in twos.

Wide to all is W _{To be measured}* S ^m, height is H _{To be measured}* S ^mThe image to be detected of a series of yardsticks, wherein S is a zoom factor, m=0,1 ... M-1 adopts the wide 2W of being respectively _r, height is 2H _rThe rectangular area travel through the image to be detected of these a series of yardsticks in level and vertical direction, it is identical that all second rectangular areas that obtain are divided into four sizes, wide is W _r, height is H _r, the 4th rectangular area that level and vertical direction are adjacent in twos.

1072) computing method of employing edge orientation histogram feature, obtain the edge orientation histogram feature of all the 4th rectangular areas, the edge orientation histogram feature of four the 4th rectangular areas in each second rectangular area head and the tail are linked to each other successively, obtain the proper vector of each second rectangular area, the proper vector of each second rectangular area as each second rectangular area edge orientation histogram feature, is perhaps carried out normalization to the proper vector of each second rectangular area and obtained each second rectangular area edge orientation histogram feature.

Adopt the computing method of edge orientation histogram feature, calculate the wide W of being _{To be measured}* S ^m, height is H _{To be measured}* S ^mThe image to be detected of a series of yardsticks in all wide be W _r, height is H _rThe edge orientation histogram feature of the 4th rectangular area, wherein S is a zoom factor, m=0,1, M-1, the edge orientation histogram feature of four the 4th rectangular areas in each second rectangular area head and the tail are linked to each other successively, obtain the proper vector of each second rectangular area, the proper vector of each second rectangular area as each second rectangular area edge orientation histogram feature, is perhaps carried out normalization to this proper vector and obtained each second rectangular area edge orientation histogram feature.

Further, method for normalizing is to adopt each value of the proper vector of each second rectangular area all divided by 1 norm or 2 norms of this proper vector, thereby obtains each second rectangular area image normalization edge orientation histogram feature.

If when calculating the edge orientation histogram feature of first rectangular area, adopted the normalization mode, when then calculate the edge orientation histogram feature of second rectangular area this moment, also adopt the normalization mode.Otherwise, if when calculating the edge orientation histogram feature of first rectangular area, do not adopted the normalization mode, promptly the proper vector of first rectangular area as the first rectangular area edge orientation histogram feature, when then calculate the edge orientation histogram feature of second rectangular area this moment, do not adopt the normalization mode yet, the proper vector of second rectangular area as the second rectangular area edge orientation histogram feature.

Above-mentioned when calculating all the 4th rectangular area edge orientation histogram features, algorithm and the step 10121 of employing), 10122) and 10123) identical: a need is step 10121), 10122) and 10123) in W _SampleReplace with W _{To be measured}* S ^m, H _SampleReplace with H _{To be measured}* S ^m, wherein S is a zoom factor, m=0, and 1 ... M-1.

109) the human body sorter that utilizes above-mentioned training to obtain is judged all second rectangular area edge orientation histogram features, detects the position of human body candidate region in the original image to be detected.

111) aftertreatment is carried out in the position of human body candidate region, detect the position of human body zone in the original image to be detected.

Again a kind of detecting system of human body of one embodiment of the invention is elaborated below.Fig. 3 is the structured flowchart of the detecting system of human body of one embodiment of the invention.As shown in Figure 3, detecting system of human body comprises pantography module 301 to be detected, edge orientation histogram characteristic extracting module 302, human body sorter training module 303, human body sort module 305 and human body classification post-processing module 306.

As shown in Figure 3, image zoom module 301 to be detected is used for according to zoom factor S the wide W of being _{To be measured}, height is H _{To be measured}Image zoom original to be detected be M wide be W _{To be measured}* S ^m, height is H _{To be measured}* S ^mThe image to be detected of a series of yardsticks, form pyramid image to be detected, m=0 wherein, 1 ... M-1.

As shown in Figure 3, edge orientation histogram characteristic extracting module 302 further comprises rectangular area division module 3021 and edge orientation histogram feature calculation module 3022.Shape area dividing module 3021 be used for wide be the first width 2W _r, height is the first height 2H _rThe rectangular area be W the wide of level and vertical direction traversal input _Sample, height is H _SampleSample image, and each wide be 2W _r, height is 2H _rTo be divided into four sizes identical in first rectangular area, wide is second width W _r, height is second height H _r, the 3rd rectangular area that level and vertical direction are adjacent in twos; The rectangular area divide module 3021 also be used for wide be the first width 2W _r, height is the first height 2H _rThe rectangular area be W in level and vertical direction traversal from the wide of image zoom module output to be detected _{To be measured}* S ^m, height is H _{To be measured}* S ^mThe image to be detected of a series of yardsticks, and all wide 2W of being that obtain _r, height is 2H _rTo be divided into four sizes identical in second rectangular area, wide is first width W _r, height is first height H _r, the 4th rectangular area that level and vertical direction are adjacent in twos.Wherein, W _{To be measured}* S ^M-1〉=W _Sample, and H _{To be measured}* S ^M-1〉=H _SampleEdge orientation histogram feature calculation module 3022 is used to calculate the rectangular area and divides all the 3rd rectangular areas of module 3021 outputs and the edge orientation histogram feature of all the 4th rectangular areas, and the edge orientation histogram feature of four the 3rd rectangular areas in each first rectangular area head and the tail are linked to each other successively, obtain the proper vector of each first rectangular area, the proper vector of each first rectangular area as each first rectangular area edge orientation histogram feature, perhaps the proper vector of each first rectangular area is carried out normalization and obtain each first rectangular area normalization edge orientation histogram feature, and it is continuous successively the edge orientation histogram feature of four the 4th rectangular areas in each second rectangular area head and the tail, obtain the proper vector of each second rectangular area, the proper vector of each second rectangular area as each second rectangular area edge orientation histogram feature, is perhaps carried out normalization to the proper vector of each second rectangular area and obtained each second rectangular area normalization edge orientation histogram feature.

Further, above-mentioned method for normalizing is to adopt each value of proper vector all divided by 1 norm or 2 norms of proper vector, thereby obtains normalization edge orientation histogram feature.

If when calculating the edge orientation histogram feature of first rectangular area, adopted the normalization mode, when then calculating the edge orientation histogram feature of second rectangular area, also adopted the normalization mode.Otherwise, if when calculating the edge orientation histogram feature of first rectangular area, do not adopted the normalization mode, promptly the proper vector of first rectangular area as the first rectangular area edge orientation histogram feature, when then calculate the edge orientation histogram feature of second rectangular area this moment, do not adopt the normalization mode yet, the proper vector of second rectangular area as the second rectangular area edge orientation histogram feature.

In an embodiment of the present invention, above-mentioned edge orientation histogram feature calculation module 3022 comprises horizontal edge and vertical edge computing module, edge strength and discretize edge direction computing module and rectangular area edge orientation histogram feature calculation module.Horizontal edge and vertical edge computing module are used to obtain the wide W of being, height is the horizontal edge and the vertical edge of each pixel of the image of H.Edge strength and discretize edge direction computing module are used for horizontal edge and the vertical edge according to each pixel of horizontal edge and vertical edge computing module output, obtain the edge strength of each pixel of image and the edge direction of discretize.Rectangular area edge orientation histogram feature calculation module is used for edge strength and the discretize edge direction according to each pixel of edge strength and discretize edge direction computing module output, and calculating wide is W, height be in the image of H all wide be second width W _r, height is second height H _rThe edge orientation histogram feature of rectangular area.Wherein, wide is W, and height is that the image of H can be represented the wide W of being _Sample, height is H _SampleSample image, also can represent the wide W of being _{To be measured}* S ^m, height is H _{To be measured}* S ^mThe image to be detected of a series of yardsticks.

In an embodiment of the present invention, above-mentioned rectangular area edge orientation histogram feature calculation module further comprises: first computing module, second computing module, the 3rd computing module and the 4th computing module.

It is 1 that first computing module is used to calculate left upper apex wide in all of image first row, and height is the edge orientation histogram feature of the rectangular area of second height, and promptly being used for the computed image left upper apex is (l, 0), and wide is 1, and height is H _rW rectangular area R (l, 0,1, H _r) edge orientation histogram feature HOG (R (l, 0,1, H _r)), l=0 wherein, 1 ..., W-1.

Second computing module is used at this wide W of being, height is in the image of H, according to left upper apex be (l, t), wide is 1, height is H _rRectangular area R (l, t, 1, H _r) edge orientation histogram feature HOG (R (l, t, 1, H _r)), iterative computation obtain left upper apex be (l, t+1), wide is 1, height is H _rRectangular area R (l, t+1,1, H _r) edge orientation histogram feature HOG (R (l, t+1,1, H _r)), l=0 wherein, 1 ... W-1, t=0,1 ... H-H _r-1: as pixel (l, t+H _r) the discretize edge direction be i=NED (l, t+H _r), pixel (l, discretize edge direction t) be j=NED (l, t), then with HOG (R (l, t, 1, H _r)) i value add pixel (l, t+H _r) edge intensity value computing EI (l, t+H _r), with HOG (R (l, t, 1, H _r)) j value deduct pixel (l, (l t), keep HOG (R (l, t, 1, H to edge intensity value computing EI t) _r)) edge intensity value computing on other discretize edge direction is constant, obtains HOG (R (l, t+1,1, H _r)).

The 3rd computing module is used at this wide W of being, height is in the image of H, according to left upper apex be (l, t), wide is 1, height is H _rRectangular area R (l, t, 1, H _r) edge orientation histogram feature HOG (R (l, t, 1, H _r)), calculating left upper apex is that (0, t), wide is W _r, height is H _rRectangular area R (0, t, W _r, H _r) the edge orientation histogram feature, l=0 wherein, 1 ..., W _r-1, t=0,1 ..., H-H _r:

$\mathrm{HOG}\left(R(0,t,W_r,H_r)\right)=\underset{0\≤l\≤W_r-1}{\mathrm{\Σ}}\mathrm{HOG}\left(R(l,t,1,H_r)\right),$ t＝0，1，…H-H _r。

The 4th computing module is used at this wide W of being, height is in the image of H, is that (l, t), wide is W according to left upper apex _r, height is H _rRectangular area R (l, t, W _r, H _r) HOG (R (l, t, the W of edge orientation histogram feature _r, H _r)) and left upper apex be (l, t), wide is 1, height is H _rRectangular area R (l, t, 1, H _r) edge orientation histogram feature HOG (R (l, t, 1, H _r)), it is that (l+1, t), wide is W that iterative computation obtains left upper apex _r, height is H _rMatrix area R (l+1, t, W _r, H _r) edge orientation histogram feature HOG (R (l+1, t, W _r, H _r)=HOG (R (l, t, W _r, H _r))+HOG (R (l+W _r, t, 1, H _r))-HOG (R (l, t, 1, H _r)) l=0 wherein, 1 ..., W-W _r-1, t=0,1 ..., H-H _r

Above-mentioned wide be W, height is that the image of H can be represented the wide W of being _Sample, height is H _SampleSample image, also can represent the wide W of being _{To be measured}* S ^m, height is H _{To be measured}* S ^mThe image to be detected of a series of yardsticks.

As shown in Figure 3, human body sorter training module 303 is used for the edge orientation histogram feature according to all first rectangular areas of edge orientation histogram characteristic extracting module 302 outputs, utilizes the sorter training algorithm, and training obtains the human body sorter.

In an embodiment of the present invention, the sorter training algorithm is at first to adopt the adaboost algorithm that all first rectangular area edge orientation histogram features are carried out feature selecting, the criterion of selecting can be defined as in all features, make weighted error and minimum feature, then with all selecteed first rectangular area edge orientation histogram features as feature, adopt svm classifier device training algorithm to train and obtain final sorter.

In another embodiment of the present invention, the sorter training algorithm is the feature of the edge orientation histogram feature of all first rectangular areas as SVM, adopt svm classifier device training algorithm to train, obtain sub-svm classifier device, with the output of this sub-svm classifier device as feature, and adopt level type adaboost algorithm to train, finally obtain the adaboost human body sorter of level type.

As shown in Figure 3, human body sort module 304 is used to utilize the human body sorter of human body sorter training module 303 outputs that the edge orientation histogram feature of all second rectangular areas of edge orientation histogram characteristic extracting module 302 outputs is judged, detects the position of human body candidate region in the original image to be detected.

As shown in Figure 3, human body classification post-processing module 305 is used for aftertreatment is carried out in the position of human body candidate region of human body sort module 304 outputs, detects the position of human body zone in the original image to be detected.

Obviously, under the prerequisite that does not depart from true spirit of the present invention and scope, the present invention described here can have many variations.Therefore, the change that all it will be apparent to those skilled in the art that all should be included within the scope that these claims contain.

Claims (21)

1, a kind of human body detecting method is characterized in that, comprises step:

Extracting in the sample image all wide is first width, and height is the first rectangular area edge orientation histogram feature of first height;

According to described all first rectangular area edge orientation histogram features, utilize the sorter training algorithm, training obtains the human body sorter;

According to zoom factor is original image zoom to be detected the image to be detected of a series of yardsticks;

Extracting in the image to be detected of described a series of yardsticks all wide is first width, and height is the second rectangular area edge orientation histogram feature of first height;

The described human body sorter that utilizes training to obtain is judged the edge orientation histogram feature of described all second rectangular areas, detects the position of human body candidate region in the described original image to be detected;

Aftertreatment is carried out in described position of human body candidate region, detect the position of human body zone in the described original image to be detected.

2, the method for claim 1 is characterized in that, in the described extraction sample image all wide be first width, height be first the height the first rectangular area edge orientation histogram feature further comprise:

With described wide be first width, height is that the rectangular area of first height travels through described sample image in level and vertical direction, division obtains described all first rectangular areas, it is identical that described each first rectangular area is divided into four sizes, wide is second width, height is second height, the 3rd rectangular area that level and vertical direction are adjacent in twos;

Adopt the computing method of edge orientation histogram feature, obtain the edge orientation histogram feature of described all the 3rd rectangular areas, the edge orientation histogram feature of four the 3rd rectangular areas in described each first rectangular area head and the tail are linked to each other successively, obtain the proper vector of described each first rectangular area, the proper vector of described each first rectangular area is perhaps carried out normalization to described proper vector and is obtained described each first rectangular area edge orientation histogram feature as described each first rectangular area edge orientation histogram feature.

3, the method for claim 1 is characterized in that, in the image to be detected of a series of yardsticks of described extraction all wide be first width, height be first the height the second rectangular area edge orientation histogram feature further comprise:

With described wide be first width, height is that the rectangular area of first height travels through the image to be detected of described a series of yardsticks in level and vertical direction, divides and obtains described all second rectangular areas; It is identical that described each second rectangular area is divided into four sizes, and wide is second width, and height is second height, the 4th rectangular area that level and vertical direction are adjacent in twos;

Adopt the computing method of edge orientation histogram feature, obtain the edge orientation histogram feature of described all the 4th rectangular areas, the edge orientation histogram feature of four the 4th rectangular areas in described each second rectangular area head and the tail are linked to each other successively, obtain the proper vector of described each second rectangular area, the proper vector of described each second rectangular area is perhaps carried out normalization to described proper vector and is obtained described each second rectangular area edge orientation histogram feature as described each second rectangular area edge orientation histogram feature.

4, the method for claim 1, it is characterized in that, described sorter training algorithm is at first to adopt the adaboost algorithm that described all first rectangular area edge orientation histogram features are carried out feature selecting, adopt the first rectangular area edge orientation histogram feature that described selection obtains feature then, and training obtains the final human sorter as the svm classifier device.

5, the method for claim 1, it is characterized in that, described sorter training algorithm is the feature of the edge orientation histogram feature of described all first rectangles as the svm classifier device, training obtains sub-svm classifier device, with the output of described sub-svm classifier device as feature, adopt level type adaboost algorithm to train, obtain level type adaboost human body sorter.

6, as claim 2 or 3 described methods, it is characterized in that the computing method of described edge orientation histogram feature comprise:

The horizontal edge of each pixel of computed image and vertical edge;

According to the horizontal edge and the vertical edge of described each pixel, obtain the edge strength and the discretize edge direction of each pixel of described image;

According to the edge strength and the discretize edge direction of each pixel of described image, calculating in the described image all wide is second width, and height is the edge orientation histogram feature of the rectangular area of second height.

As claim 2 or 3 described methods, it is characterized in that 7, described normalization is to adopt each value of described proper vector 1 norm or 2 norms divided by described proper vector.

8, method as claimed in claim 6 is characterized in that, the horizontal edge and the vertical edge of described each pixel according to image, and the edge strength and the discretize edge direction that obtain each pixel of described image further comprise:

Calculating left upper apex wide in all of described image first row is 1, and height is the edge orientation histogram feature of the rectangular area of second height;

Calculating left upper apex other capable all described image removes first row wide is 1, and height is the edge orientation histogram feature of the rectangular area of second height;

Calculating left upper apex wide in all of described image first row is second width, and height is the edge orientation histogram feature of the rectangular area of second height;

Calculating left upper apex all of other row described image removes first row wide is second width, and height is second highly the edge orientation histogram feature of rectangular area.

9, method as claimed in claim 8 is characterized in that, described calculating left upper apex is wide in image all of other row except that first row to be 1, and height is that second highly the edge orientation histogram feature of rectangular area further comprises:

Is left upper apex 1 the wide of the same row of described image previous row, height is that sequence number is the eigenwert of the same row pixel of previous row discretize edge direction in the edge orientation histogram feature of rectangular area of second height, subtract each other with the edge intensity value computing of the same row pixel of described previous row, obtaining left upper apex is 1 the wide of the same row of described image current line, height be described second the height rectangular area edge orientation histogram feature in sequence number be described eigenwert with delegation's previous column pixel discretize edge direction; Is left upper apex 1 the wide of the same row of described image previous row, height be described second the height rectangular area edge orientation histogram feature in sequence number be previous row add the above second the height, the eigenwert of the pixel discretize edge direction of same row, add the above second height with described previous row, the pixel edge intensity level addition of same row, obtaining left upper apex is 1 the wide of the same row of described image current line, height be described second the height rectangular area edge orientation histogram feature in sequence number be described previous row add the above second the height, the eigenwert of the pixel discretize edge direction of same row; Left upper apex is 1 the wide of the same row of described image current line, height be other value of rectangular area edge orientation histogram feature of described second height to equal left upper apex be 1 the wide of the same row of described image previous row, height is the respective value of the rectangular area edge orientation histogram feature of described second height.

10, method as claimed in claim 8 is characterized in that, described calculating left upper apex is wide in all of image first row to be second width, and height is that the edge orientation histogram feature of the rectangular area of second height further comprises:

Being listed as left upper apex wide with described second width of delegation's second width columns at described image with delegation first successively is 1, height is the edge orientation histogram feature addition of the rectangular area of second height, obtaining left upper apex is described second width at described image with delegation's first the wide of row, and height is the edge orientation histogram feature of the rectangular area of described second height.

11, method as claimed in claim 8 is characterized in that, described calculating left upper apex is wide in image all of other row except that first row to be second width, and height is that second highly the edge orientation histogram feature of rectangular area further comprises:

Is left upper apex second width at described image with the wide of delegation's previous column, height is that the edge orientation histogram feature of the rectangular area of second height adds that left upper apex adds that with described second width of delegation the wide of previous column is 1 at described image, height is the edge orientation histogram feature of the rectangular area of described second height, deducting left upper apex again is 1 at described image with the wide of delegation's previous column, height is the edge orientation histogram feature of the rectangular area of described second height, obtaining left upper apex is described second width the wide of the same trade of described image prostatitis, and height is the edge orientation histogram feature of the rectangular area of described second height.

12, a kind of detecting system of human body is characterized in that, comprising:

Image zoom module to be detected, being used for according to zoom factor is original image zoom to be detected the image to be detected of a series of yardsticks;

The edge orientation histogram characteristic extracting module, being used for extracting the input sample image wide is first width, height is the edge orientation histogram feature of all first rectangular areas of first height, and the image to be detected of a series of yardsticks that is used for extracting described image zoom module output to be detected wide be first width, height is the edge orientation histogram feature of all second rectangular areas of first height;

Human body sorter training module is used for the edge orientation histogram feature according to all first rectangular areas of described edge orientation histogram characteristic extracting module output, utilizes the sorter training algorithm, and training obtains the human body sorter;

The human body sort module, be used to utilize the human body sorter of described human body sorter training module output that the edge orientation histogram feature of all second rectangular areas of described edge orientation histogram characteristic extracting module output is judged, detect the position of human body candidate region in the described original image to be detected;

Human body classification post-processing module is used for aftertreatment is carried out in the position of human body candidate region of described human body sort module output, obtains the position of human body zone in the described original image to be detected.

13, system as claimed in claim 12 is characterized in that, described edge orientation histogram characteristic extracting module further comprises:

Module is divided in the rectangular area, be used for described wide be first width, height is that the rectangular area of first height travels through described sample image and a series of image to be detected in level and vertical direction, divide respectively and obtain described all first rectangular areas and second rectangular area, it is identical that described each first rectangular area and each second rectangular area are divided into four sizes respectively, wide is second width, height is second height, the 3rd rectangular area that level and vertical direction are adjacent in twos and the 4th rectangular area;

Edge orientation histogram feature calculation module, be used to calculate described rectangular area and divide all the 3rd rectangular areas of module output and the edge orientation histogram feature of all the 4th rectangular areas, and the edge orientation histogram feature of four the 3rd rectangular areas in described each first rectangular area head and the tail are linked to each other successively, obtain the proper vector of described each first rectangular area, the proper vector of described each first rectangular area is as described each first rectangular area edge orientation histogram feature, perhaps the proper vector of described each first rectangular area is carried out normalization and obtain described each first rectangular area edge orientation histogram feature, and it is continuous successively the edge orientation histogram feature of four the 4th rectangular areas in described each second rectangular area head and the tail, obtain the proper vector of described each second rectangular area, the proper vector of described each second rectangular area is perhaps carried out normalization to the proper vector of described each second rectangular area and is obtained described each second rectangular area edge orientation histogram feature as described each second rectangular area edge orientation histogram feature.

14, system as claimed in claim 12, it is characterized in that, described sorter training algorithm is at first to adopt the adaboost algorithm that described all first rectangular area edge orientation histogram features are carried out feature selecting, adopt the first rectangular area edge orientation histogram feature that described selection obtains feature then, and training obtains the final human sorter as the svm classifier device.

15, system as claimed in claim 12, it is characterized in that, described sorter training algorithm is at first the feature of the edge orientation histogram feature of described all first rectangles as the svm classifier device, training obtains sub-svm classifier device, with the output of described sub-svm classifier device as feature, adopt level type adaboost algorithm to train, obtain level type adaboost human body sorter.

16, system as claimed in claim 13 is characterized in that, described edge orientation histogram feature calculation module further comprises:

Horizontal edge and vertical edge computing module are used to obtain the horizontal edge and the vertical edge of each pixel of image;

Edge strength and discretize edge direction computing module are used for horizontal edge and vertical edge according to each pixel of described horizontal edge and the output of vertical edge computing module, obtain the edge strength and the discretize edge direction of each pixel of described image;

Rectangular area edge orientation histogram feature calculation module, be used for edge strength and discretize edge direction according to each pixel of described edge strength and the output of discretize edge direction computing module, calculating in the described image all wide is second width, and height is the edge orientation histogram feature of the rectangular area of second height.

17, system as claimed in claim 13 is characterized in that, described normalization is to adopt each value of described proper vector 1 norm or 2 norms divided by described proper vector.

18, system as claimed in claim 16 is characterized in that, described rectangular area edge orientation histogram feature calculation module further comprises:

First computing module, being used to calculate left upper apex wide in all of described image first row is 1, height is the edge orientation histogram feature of the rectangular area of second height;

Second computing module, being used to calculate left upper apex wide in described image all of other row except that first row is 1, height is second highly the edge orientation histogram feature of rectangular area;

The 3rd computing module, being used to calculate left upper apex wide in all of described image first row is second width, height is the edge orientation histogram feature of the rectangular area of second height;

The 4th computing module, being used to calculate left upper apex wide in described image all of other row except that first row is second width, height is second highly the edge orientation histogram feature of rectangular area.

19, system as claimed in claim 18 is characterized in that, described second computing module is further used for:

Is left upper apex 1 the wide of the same row of described image previous row, height is that sequence number is the eigenwert of the same row pixel of previous row discretize edge direction in the edge orientation histogram feature of rectangular area of second height, subtract each other with the edge intensity value computing of the same row pixel of described previous row, obtaining left upper apex is 1 the wide of the same row of described image current line, height be described second the height rectangular area edge orientation histogram feature in sequence number be described eigenwert with delegation's previous column pixel discretize edge direction; Is left upper apex 1 the wide of the same row of described image previous row, height be described second the height rectangular area edge orientation histogram feature in sequence number be previous row add the above second the height, the eigenwert of the pixel discretize edge direction of same row, add the above second height with described previous row, the pixel edge intensity level addition of same row, obtaining left upper apex is 1 the wide of the same row of described image current line, height be described second the height rectangular area edge orientation histogram feature in sequence number be described previous row add the above second the height, the eigenwert of the pixel discretize edge direction of same row; Left upper apex is 1 the wide of the same row of described image current line, height be other value of rectangular area edge orientation histogram feature of described second height to equal left upper apex be 1 the wide of the same row of described image previous row, height is the respective value of the rectangular area edge orientation histogram feature of described second height.

20, system as claimed in claim 18 is characterized in that, described the 3rd computing module is further used for:

Being listed as left upper apex wide with described second width of delegation's second width columns at described image with delegation first successively is 1, height is the edge orientation histogram feature addition of the rectangular area of second height, obtaining left upper apex is described second width at described image with delegation's first the wide of row, and height is the edge orientation histogram feature of the rectangular area of described second height.

21, system as claimed in claim 18 is characterized in that, described the 4th computing module is further used for:

Is left upper apex second width at described image with the wide of delegation's previous column, height is that the edge orientation histogram feature of the rectangular area of second height adds that left upper apex adds that with described second width of delegation the wide of previous column is 1 at described image, height is the edge orientation histogram feature of the rectangular area of described second height, deducting left upper apex again is 1 at described image with the wide of delegation's previous column, height is the edge orientation histogram feature of the rectangular area of described second height, obtaining left upper apex is described second width the wide of the same trade of described image prostatitis, and height is the edge orientation histogram feature of the rectangular area of described second height.

Images