CN102810159A - Human body detecting method based on SURF (Speed Up Robust Feature) efficient matching kernel - Google Patents

Abstract

The invention provides a human body detecting method based on an SURF efficient matching kernel, and mainly solves the problem that image background hybridity can not be better processed in the existing method. The method comprises the steps that a negative sample is obtained through bootstrap in an INRIA (Institute National de Recherce en Informatique et Automatique) database, and a training sample set of the whole human body is formed by the negative sample and a positive sample in the database; SURF descriptor feature points are extracted under different image scales for the training sample; feature points are extracted by random sampling to constitute the initial vector basis of a visual vocabulary; constrained singular value decomposition is utilized for the initial vector basis to obtain the maximum kernel function feature; the maximum kernel function feature in different image scales is weighted to obtain the features under all the image scales; the obtained features are trained in different classes by an SVM (Support Vector Machine) classifier, and a detection classifier is obtained; and the image to be detected is input to the classifier to obtain the final detection result. The method disclosed by the invention can be used for accurately detecting the human body, and can be used for intelligent monitoring, driver auxiliary systems and virtual video.

Description

The human body detecting method that efficiently matees nuclear based on SURF

Technical field

The invention belongs to technical field of image processing, relate to static human detection method, can be used for intelligent monitoring, driver assistance system, human body motion capture, porny filtration and virtual video.

Background technology

Human detection is the very wide technology of an application prospect in computer vision field; Human detection all has application promise in clinical practice in a plurality of fields; But because the diversity of human body attitude; Mixing and the clothes texture of background, illumination condition many-sided factor such as self blocks and causes human detection to become a very problem of difficulty.At present, in the still image method of human detection mainly contain detection method based on kinetic characteristic, based on the method for manikin with based on the method for statistical classification.

Based on the detection method of kinetic characteristic is that attitude when utilizing human body to stablize changes and the symmetry of human body is the cycle and changes this characteristic; Structure self similarity matrix in time domain; Motion change through human cyclin property reflects character different and other motions of matter; And utilize this analytical approach to detect movement human, but this method algorithm complex is big, and higher to the human motion stability requirement.

Based on the method for manikin, clear and definite manikin arranged, carry out human body identification according to each position and the relation between the human body of model construction then.This method can be handled occlusion issue, and can infer the attitude of human body.But the deficiency of this method is the structure difficulty of model, finds the solution complicacy.

Based on the method for statistical classification, from a series of training data middle school acquistion to a sorter, represent human body through machine learning with this sorter, utilize this sorter input is classified and to discern then.Advantage based on the method for statistical classification is that testing result is stable, and effect is better, and shortcoming is to need a lot of training datas, and is difficult to solve the problem that insufficient light and background mix.The human body detecting method that wherein efficiently matees nuclear based on SURF Speed Up Robust Feature; The characteristics of image of its input category device is a kind of based on the part image characterizing method; Traditional background challenge can be avoided, better human detection result can be obtained.

Summary of the invention

The present invention seeks to the deficiency to above-mentioned prior art, a kind of of proposition efficiently mated the human body detecting method of nuclear based on SURF, to reduce the complexity of image characteristics extraction, improves the sign ability of characteristic, effectively improves correct rate of human body detection.

Technical scheme of the present invention realizes through following steps:

(1) from institut national de recherche en infomatique et automatique INRIA database, obtain negative sample, and other positive sample constitutes the whole human body training sample set in database through the bootstrapping operation;

(2) every width of cloth training sample image is divided into 8 * 8 pixel grid, each grid extracts the SURF descriptor unique point F of all training images respectively by the graphical rule sampling of 16 and 25 pixel sizes;

(3) carry out stochastic sampling through SURF descriptor unique point F, obtain the visual vocabulary of whole training sample 350 dimensions, constitute initial base vector R with the 350 dimension visual vocabularies that obtain to all training images;

(4) with initial base vector R, utilize the nuclear svd CKSVD of belt restraining to carry out dictionary study, obtain maximum kernel function characteristic r;

(5) extract the maximum kernel function characteristic r that suppresses similar through the maximization eigenwert; And press descending and extract the kernel function eigenwert, the same element of deletion maximal value obtains proper vector G; Characteristics of image G to each different images yardstick carries out weighted sum, obtains the characteristic G ' of all images yardstick:

G′＝G×A _l，

Wherein, A _lBe the weight of difference figure phase yardstick, l=[1,2],

w _l=1/p _l, p is the pixel size of the graphical rule of the SURF unique point extracted, p={16,25};

(6) store the characteristic G ' of all images yardstick, select the low dimensional feature h of similar Gaussian distribution among the G ', efficiently mate nuclear characteristic X as the SURF of final image;

(7) use support vector machine svm classifier device that resulting SURF is efficiently mated nuclear characteristic X and carry out classification based training, finally be used to the sorter that detects;

(8) import image to be detected, utilize the sorter that has obtained to confirm final testing result.

The present invention has the following advantages compared with prior art:

1, can avoid traditional fuzzy because the SURF that uses efficiently matees the characteristics of image of nuclear among the present invention, can obtain better human detection result based on the expression edge and that produce based on the graphical representation method of profile.

2, the present invention can effectively reduce characteristic time and the data computation amount extracted because the image feature information that extracts is low than traditional image describing method dimension.

3, the present invention when processing mixes background image, can obtain better result owing to be based on the human body detecting method of local visual characteristic information.

Description of drawings

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

Fig. 2 is the positive sample image of part that uses among the present invention;

Fig. 3 is the part negative sample image that uses among the present invention;

Fig. 4 is the detection performance comparison diagram with the present invention and existing method;

Fig. 5 is the figure as a result that human body image is detected with the present invention.

Embodiment

With reference to Fig. 1, practical implementation step of the present invention is following:

Step 1; From institut national de recherche en infomatique et automatique INRIA database, a large amount of negative sample image that obtains through bootstrapping operation, and in database other positive sample image composing training sample set; Wherein the negative sample image is as shown in Figure 2, and positive sample image is as shown in Figure 3.

Step 2, the SURF descriptor unique point F of extraction training sample set.

2a) j width of cloth training image is divided into 8 * 8 pixel grid, each grid obtains the SURF Speed Up Robust Feature descriptor unique point F of i width of cloth training image respectively by the graphical rule sampling of 16 and 25 pixel sizes _j

2b) according to step 2a) extract the SURF descriptor unique point F of all training images, wherein, F={F ₁..., F _j..., F _N, j ∈ [1, M], M is a number of training.

Step 3, the initial base vector R of acquisition visual vocabulary.

3a) to each width of cloth training sample image; On 8 * 8 image grid; According to 16; 25 pixel size yardsticks, 15 SURF unique points that obtained by step (2) of random sampling are designated as

i and represent i width of cloth training image respectively;

3b) repeating step 3a), the SURF unique point of the training sample that random extraction is all is designated as F '; Utilize the k-means clustering method that SURF unique point similar among the F ' is carried out cluster, define 350 cluster centres, obtain the visual vocabulary of whole training image 350 dimensions, constitute the initial base vector R of visual vocabulary.

Step 4 obtains the maximum kernel function proper vector r of initial base vector R.

4a) initial base vector R is used projection coefficient v, projects on the space of one 350 dimension, obtain the projection R ' of R:

R′＝Rv，

v＝[v ₁,...v _i...,v _N]

v _i＝(R ^TR) ^-1(R ^Tr _i)，i∈[1,N]，

Wherein, r _iBe the maximum kernel characteristic of i unique point extracting in the piece image, v _iBe the low dimension projection coefficient of i unique point extracting in the piece image, N is the quantity of the unique point of picked at random in the piece image;

4b) on projector space, maximum kernel function proper vector r approached the projection R ' of initial base vector R, obtains approximating function f (r):

f(r)＝arg?min‖r-R′‖，

With R '=Rv substitution following formula:

f(r)＝arg?min‖r-Rv‖，

Wherein, ‖ ‖ representes 2 norms, and arg min ‖ ‖ representes to minimize;

4c) v among f (r)=arg min ‖ r-Rv ‖ and r are launched, obtain maximum kernel function proper vector r to 2 approximating function f of initial base vector R (v, r):

$f(v,r)=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left|\right|r_i-{\mathrm{Rv}}_i\left|\right|}^2,$

Wherein, r=[r ₁... r _i..., r _N], expression maximum kernel function proper vector;

(v r), obtains maximum kernel function proper vector r 4d) to use at random the gradient descent method to find the solution approximating function f.

Step 5; By the element among the descending sort maximum kernel function proper vector r, the same element of maximal value among the deletion maximum kernel function proper vector r obtains proper vector G; Proper vector G under each different images yardstick carries out weighted sum, obtains the characteristics of image G ' on all images yardstick:

G′＝G×A _l，

Wherein, A _lBe the weight of difference figure phase yardstick, l=[1,2],

w _l=1/p _l, p is the pixel size that extracts the graphical rule of SURF descriptor unique point, p={16,25}.

Step 6 stores all images yardstick characteristics of image G ' down, selects the low dimensional feature h of similar Gaussian distribution among the characteristics of image G ', efficiently matees as the SURF of final image and examines characteristic X.

Step 7 is used support vector machine svm classifier method that the SURF that has obtained is efficiently mated nuclear characteristic X and is carried out classification learning, finally is used for the sorter of human detection.

Step 8 is used the sorter that is used for human detection that has obtained, and confirms final testing result.

(8a) input image to be detected; With the size in the image to be detected upper left corner zone that is 128 * 64 pixels as first scanning window; Every to 8 pixels of right translation or downwards 16 pixels of translation as a new scanning window; Obtain one group of scanning window thus, input step (7) gained sorter obtains the sorter mark of each scanning window;

(8b) judge whether comprised human body in the altimetric image,, then from all scanning windows that contains human body, find out the highest scanning window of sorter mark as main window if the scanning window of sorter output contains human body according to the sorter mark of scanning window;

(8c) main window and other human body windows are made up judgement, when other human body windows are in around the main window and overlapping greater than 1/2 the time, with this window and main window combination, the human body window after obtaining to make up;

(8d) the human body window after the reservation combination, deletion main window and all are participated in the human body window of combination;

(8e) if also have remaining human body window, then find out human body window that wherein the sorter mark is the highest again as main window, and repeating step (8b)-(8d);

(8f) on tested person's volume image, mark all testing results, as by the altimetric image final human detection result, adopt rectangle frame to represent testing result, the human body that is detected is in the rectangle frame.

Effect of the present invention can obtain checking through following emulation experiment:

1) emulation experiment condition setting: emulation experiment of the present invention compiles completion on Matlab 2009a, and execution environment is the HP workstation under the Windows framework.Test required positive sample and negative sample and all be taken from institut national de recherche en infomatique et automatique INRIA database.Use 2416 positive samples and 13500 negative samples as training set; 1132 positive samples and 4050 negative samples are as test set; The size of positive sample and negative sample image is 128 * 64 pixels, and Fig. 2 has provided the wherein positive sample image of part, and Fig. 3 is a part negative sample image.

2) emulation content and interpretation of result

Emulation one: use the present invention and existing method that characteristics of image is classified respectively, classification performance is as shown in Figure 4.Among Fig. 4, top curve is a classification performance curve of the present invention, and following curve is the classification performance curve of existing method, and as can beappreciated from fig. 4, classification performance of the present invention is higher than the classification performance of existing method.

Emulation two: use the inventive method and existing method that the same width of cloth is carried out human detection from the image of Massachusetts science and engineering MIT database respectively, testing result is as shown in Figure 5.Wherein, Fig. 5 (a) is to use existing method to carry out human detection; Carry out the experimental result before window merges, Fig. 5 (b) is the final detection result of existing method, and Fig. 5 (c) expression adopts this method to carry out human detection; Carry out the experimental result before window merges, Fig. 5 (d) is the final detection result of this method.Method as can beappreciated from fig. 5 of the present invention has higher human detection accuracy.

To sum up; The present invention has improved the ability to express of characteristic, thereby has made this method be very suitable for the human detection of still image in the complexity that reduces image characteristics extraction; Compare with existing method simultaneously, this method can greatly reduce the empty scape rate of human detection.

Claims (4)

1. a human body detecting method that efficiently matees nuclear based on SURF comprises the steps:

(1) from institut national de recherche en infomatique et automatique INRIA database, obtain negative sample, and other positive sample constitutes the whole human body training sample set in database through the bootstrapping operation;

(2) every width of cloth training sample image is divided into 8 * 8 pixel grid, each grid extracts the SURF Speed Up Robust Feature descriptor unique point F of all training images respectively by the graphical rule sampling of 16 and 25 pixel sizes;

(3) carry out stochastic sampling through SURF descriptor unique point F, obtain the visual vocabulary of whole training sample 350 dimensions, constitute initial base vector R with the 350 dimension visual vocabularies that obtain to all training images;

(4) with initial base vector R, utilize the nuclear svd CKSVD of belt restraining to carry out dictionary study, obtain maximum kernel function characteristic r;

(5) through maximization eigenwert extraction method; Suppress similar maximum kernel function characteristic r; And press descending and extract the kernel function eigenwert, the same element of deletion maximal value obtains proper vector G; Characteristics of image G to each different images yardstick carries out weighted sum, obtains the characteristic G ' of all images yardstick:

G′＝G×A _l，

Wherein, A _lBe the weight of difference figure phase yardstick, l=[1,2],

w _l=1/p _l, p is the pixel size of the graphical rule of the SURF unique point extracted, p={16,25};

(6) store the characteristic G ' of all images yardstick, select the low dimensional feature h of similar Gaussian distribution among the G ', efficiently mate nuclear characteristic X as the SURF of final image;

(7) use support vector machine svm classifier device that resulting SURF is efficiently mated nuclear characteristic X and carry out classification based training, finally be used to the sorter that detects;

(8) import image to be detected, utilize the sorter that has obtained to confirm final testing result.

2. method according to claim 1, wherein step 2) described in the SURF descriptor unique point F of all training images of extraction, carry out as follows:

2a) j width of cloth training image is divided into 8 * 8 pixel grid, each grid obtains the SURF SpeedUp Robust Feature unique point F of i width of cloth training image respectively by the graphical rule sampling of 16 and 25 pixel sizes _j

2b) according to step 2a) extract the SURF descriptor unique point F of all training images, wherein, F={F ₁..., F _j..., F _N, j ∈ [1, M], M is a number of training.

3. method according to claim 1, the visual vocabulary of the whole training sample of the acquisition described in the step (3) 350 dimensions wherein, carry out as follows:

3a) to each width of cloth training sample image; On 8 * 8 image grid; According to 16; 25 pixel size yardsticks, 15 SURF unique points that obtained by step (2) of random sampling are designated as

i and represent i width of cloth training image respectively;

3b) repeating step 3a), extracts the SURF unique point of all training samples, be designated as F ', define 350 cluster centres, utilize the k-means clustering method that SURF unique point similar among the F ' is carried out cluster, obtain the visual vocabulary of whole training sample 350 dimensions.

4. according to the said method of claim 1, wherein the described input of step (8) image to be detected utilizes the sorter that has obtained to confirm final testing result, carries out as follows:

(8a) input image to be detected; With the size in the image to be detected upper left corner zone that is 128 * 64 pixels as first scanning window; Every to 8 pixels of right translation or downwards 16 pixels of translation as a new scanning window; Obtain one group of scanning window thus, input step (7) gained sorter obtains the sorter mark of each scanning window;

(8b) judge whether comprised human body in the altimetric image,, then from all scanning windows that contains human body, find out the highest scanning window of sorter mark as main window if the scanning window of sorter output contains human body according to the sorter mark of scanning window;

(8c) main window and other human body windows are made up judgement, when other human body windows are in around the main window and overlapping greater than 1/2 the time, with this window and main window combination, the human body window after obtaining to make up;

(8d) the human body window after the reservation combination, deletion main window and all are participated in the human body window of combination;

(8e) if also have remaining human body window, then find out human body window that wherein the sorter mark is the highest again as main window, and repeating step (8b)-(8d);

(8f) on tested person's volume image, mark all testing results,, adopt rectangle frame to represent testing result, the human body that is detected is in the rectangle frame as by the altimetric image final human detection result.

Images