CN103942577A - Identity identification method based on self-established sample library and composite characters in video monitoring - Google Patents

Abstract

The invention discloses an identity identification method based on self-established sample libraries and composite characters in video monitoring. The method comprises the steps that firstly, preprocessing is conducted on an acquired video, foreground detection is conducted, so that moving object information is obtained, then face detection is conducted on the basis of the moving object information, the detected face is identified, if the currently detected face can not be identified, a user is inquired so as to identify the detected face, and the identified face is added to the face sample library; if the face can not be detected in foreground information, pedestrian detection is conducted, a detected pedestrian is traced, gait period detection is conducted on a traced pedestrian image sequence, features of detected gait information of a period are extracted and identified, if identification fails, gaits are classified in the same mode of user identification and added to the gait sample library. The identity identification method provides a solution for identity identification on the condition of a lack of training sample diversity or small samples.

Description

Personal identification method based on certainly setting up Sample Storehouse and composite character in video monitoring

Technical field

The present invention relates to Digital Image Processing, pattern-recognition and machine learning techniques field, particularly a kind of personal identification method based on certainly setting up Sample Storehouse and composite character.

Background technology

Living things feature recognition refers to and utilizes the intrinsic physiological characteristic of people itself or behavioural characteristic, it processed to and then differentiated a kind of technology of personal identification by computing machine.Biological characteristic is divided into physiological characteristic and behavioural characteristic.Physiological characteristic mainly comprises people's face, fingerprint, hand-type, palmmprint, ear type, DNA, iris, retina, skeleton etc.Behavioural characteristic comprises the action of signature, the attitude of the rhythm of keystroke, one's voice in speech, walking etc.Wherein, the recognition of face of physiological characteristic and the Gait Recognition of behavioural characteristic become in video monitoring system the main method for identification because having non-infringement.

Recognition of face refers to utilize analyzes the computer technology that relatively people's face visual signature information is carried out identity discriminating, it can be defined as: input rest image or video in scene to be checked, use and utilize the recognizer that face database " study " obtains to identify or verify a people or a plurality of people in scene.The advantage of recognition of face is its naturality and the feature of not discovered by individuality to be identified.So-called naturality, refers to that the biological characteristic utilizing when this recognition method is carried out human body identification with the mankind is identical.Be difficult for discovering and refer to that individuality to be identified is not easy to cause its attention in the process of identification, this makes this recognition methods not offensive, and is not easy by impersonation owing to being difficult for being realized.Gait refers to mode when people walk, and when walking, the posture at double-legged exercise and other positions of health is unique to a people.And this posture has relative stability, in regular hour scope, be not easy to change with under similar walking environment, people can judge the identity of walker accordingly.Adopt gait feature carry out identification have advantages of unique: first, while utilizing gait to identify, gather the distant of the video camera of gait information and individuality to be identified, so the method do not have infringement, also do not need Body contact, gait has the perceptibility of being difficult for.Secondly, gait is difficult to hide, and is subject to the various factors such as age, skeletal structure due to gait feature, pretends other people gait to be easy to show the cloven hoof, and gait has the camouflage of being difficult for property.The 3rd, while utilizing video image to carry out Gait Recognition, lower to the resolution requirement of image, Gait Recognition is not high to hardware requirement.The 4th, Gait Recognition can be identified in far range, accomplishes to give warning in advance.Based on these advantages, more and more higher to the attention rate of Research on Gait Recognition.

In sum, people's face and gait advantage when being applied to identification is comparatively obvious.Table 1 application scenario, relative merits, decipherment distance, whether need measured coordinate, to picture quality require and discrimination aspect compare.As can be seen from Table 1, recognition of face has similar application scenario and applicable elements with Gait Recognition, and they do not need testee's cooperation, can in the situation that not discovered, target not carried out to identification.On decipherment distance, Gait Recognition belongs to remote identification; Recognition of face can be used for moderate distance.Therefore on application scenarios, people's face and gait have the possible of fusion.Recognition of face is subject to the impact of the factors such as illumination, expression, attitude, change of age, and Gait Recognition is subject to the impact of stimulus, physiological change, psychological factor, clothes illumination condition, shelter etc., the influence factor of these two kinds of biological characteristics is different, and when wherein a kind of biological characteristic factor that is affected is disturbed, another kind of biological characteristic still can play a role effectively.Therefore, gait feature and face characteristic combine and carry out the new research direction that identification has become current pedestrian's identification.

Table 1 Gait Recognition and recognition of face comparison sheet

No matter existing identity recognizing technology utilizes face characteristic or gait feature, for guaranteeing that the recognition accuracy different shooting angles that utilization is set up in advance is mostly, the Sample Storehouse of the composition of sample of different illumination conditions, by feature extraction, utilize machine learning techniques to identify again.

The research of the identification combining for face characteristic and gait feature, application number is 200910073004.1 patent of invention (front face human body automatic identity recognition method under long-distance video) use Adaboost method detection pedestrian, if detect, just automatically open people's face module and gait module, to gait and people's face, adopt core principle component analysis (KPCA) to carry out feature extraction respectively, finally adopt the auxiliary gait feature of face characteristic to identify in the fusion method of decision level.The advantage of the method is, even if gait training sample is single sample, and facial image is a plurality of, so just from another one angle, has expanded the number of training sample, contributes to identification.This patent of invention only has the situation of for sample size in Gait Recognition process, in conjunction with a plurality of people's face samples, improves the accuracy rate of identification.

The discrimination of existing identity recognizing technology is affected by three factors mainly: the Feature Correspondence Algorithm of describing algorithm and identifying for Sample Storehouse, the feature of training.The diversity that wherein for the Sample Storehouse of training, the impact of discrimination is mainly comprised to the sample that Sample Storehouse comprises whether, the difference of environmental factor (as angle and illumination etc.) between sample and the input image sequence that comprises target to be identified, As time goes on and variation occurring etc. and gait when people's looks and walking.In addition, the recognition accuracy that utilizes as previously mentioned gait feature to carry out identification is starkly lower than face characteristic identification, its reason is, Gait Recognition belongs to behavioral characteristics identification category, when feature is described, need at least one continuous gait cycle image sequence, in extracting gait profile process, need to make gait sample background and gait image sequence to be identified up in background difference Different Effects that gait feature is described.

As can be seen here, how improving the diversity of sample or the in the situation that of small sample, complete identification and eliminate the difference causing due to different factors between sample and input data to be identified is one of current identity recognizing technology problem to be solved.

Summary of the invention

For the existing problem of existing identity recognizing technology, the object of the present invention is to provide a kind of personal identification method based on certainly setting up Sample Storehouse and composite character, with this, effectively improve the accuracy rate of identification.

In order to achieve the above object, the present invention adopts following technical scheme:

Personal identification method based on certainly setting up Sample Storehouse and composite character in video monitoring, the method first detects the image sequence of input by the good people's face of training in advance and pedestrian's sorter, the mode of utilizing man-machine interaction to the front face sample detecting and side gait sample carry out class indication as cognitive phase the sample for " training ", recycle these sample extraction people's face or gait feature trainable recognizer, detected other sample is carried out to class indication, if had, fail people's face or the gait of class indication, then inquire that user identifies.

In preferred version, while carrying out Gait Recognition, for consecutive image sequence frame at least comprise the data of a gait cycle.

Further, in detecting identifying, adopting recognition of face is that main Gait Recognition is that auxiliary mode is carried out identification, complete front face image do not detected in current image frame, carries out pedestrian detection and tracking and extracts gait feature and carry out identification.

Further, described personal identification method specifically comprises the steps:

(1) image sequence collecting is carried out to pre-service, after pre-service, obtain moving region information;

(2) in the moving region getting (being foreground information), carry out the detection of people's face, if people's face information detected, proceed to step (3); If people's face information do not detected, proceed to step (4);

(3) if people's face information detected, after normalization, utilize LBP operator extraction feature, and existing people's face sample characteristics compares identification, and export recognition result; If None-identified current detection to people's face information (do not mate with existing people's face sample or people's face Sample Storehouse for empty), inquire that user identifies, and the people's face sample having identified added in Sample Storehouse;

(4), if people's face do not detected, in moving region (being foreground information), carry out pedestrian detection;

(5) follow the tracks of the pedestrian who detects, obtain the image sequence of its walking, and carry out gait cycle detection;

(6) to following the tracks of the gait sequence of the gait cycle obtaining, extract gait feature, and compare identification with existing gait sample characteristics, and export recognition result; If the current gait feature extracting of None-identified (do not mate with existing gait sample or gait Sample Storehouse for empty), inquires that user identifies, and the gait information having identified is added in Sample Storehouse.

Further, the pretreatment operation in described step (1) comprises: contrast strengthens, edge strengthens, gray scale stretches, moving target extracts.

Further, the background subtraction method based on ViBe in described step (1) is extracted the moving target information in current scene, and when obtaining moving target information, the foreground mask of twice acquisition is carried out to logical “and” operation; The minimum area-encasing rectangle of the moving target obtaining while simultaneously also utilizing grabcut technology to extract in conjunction with moving target obtains complete moving target information.

Further, while carrying out the detection of people's face in described step (2), the sorter that in the Sample Storehouse that utilization is extracted, the Haar-like feature of sample obtains by Adaboost Algorithm for Training detects, and training process is used " bootstrapping " method.

Further, while carrying out recognition of face in described step (3), utilize the LBP operator of improved border circular areas, the people's face detecting is divided into some, then generate the LBP histogram of each piece, finally the LBP histogram of all is coupled together as feature histogram, and be described as according to carrying out Classification and Identification with this.

Further, while carrying out pedestrian detection in described step (4), sample in Sample Storehouse (being gait image) is normalized to the picture of 64 * 128 sizes, extract thus and obtain histograms of oriented gradients corresponding to sample in Sample Storehouse, and using this sorter that utilizes Adaboost Algorithm for Training to obtain as detected characteristics and detect.

Further, in described step (5), utilize meanshift algorithm to follow the tracks of the pedestrian who detects, to following the tracks of the gait image sequence obtaining, utilize each width gait profile diagram to calculate its swinging distance, and utilize odd number maximum point to determine gait cycle.

Further, while carrying out gait feature identification in described step (6), utilize the profile feature modeling profile barycenter of gait, and by profile diagram be transformed into take barycenter as round dot, level to the right as real axis, to hold up be upwards the complex coordinates system of the imaginary axis, utilizes PSA technology to move to profile or structural statistical presentation is described its characteristic and obtained PMS and carries out identification after conversion.

The present invention who forms according to above technical scheme, its tool has the following advantages:

The Sample Storehouse needing while utilizing people's face of detecting in monitor video or gait to set up cognitive phase training, reduces the low possibility of discrimination causing due to the difference of environmental factor due between the sample gathering and monitor video.

Utilize interactive mode to identify Sample Storehouse, avoid wrong identification that the non-diversity due to sample causes or the problem of None-identified.

Utilize side gait feature and front face feature to carry out identification, the situation of None-identified while having avoided due to distance or without front face information, simultaneously because Gait Recognition accuracy rate is lower than front face recognition accuracy, therefore when having reliable face characteristic, do not consider gait information, guarantee certain recognition accuracy.

Accompanying drawing explanation

Below in conjunction with the drawings and specific embodiments, further illustrate the present invention.

Fig. 1 is the schematic flow sheet of the inventive method;

Fig. 2 is two prospect masks and the logical “and” result design sketch that the background extracting technology by ViBe obtains;

Fig. 3 is the minimum area-encasing rectangle that obtains of the result extracted in conjunction with moving target and utilizes this information to cut apart complete moving target schematic diagram in conjunction with Grabcut technology;

Fig. 4 (a) is Haar-like linear feature schematic diagram;

Fig. 4 (b) is Haar-like edge feature schematic diagram;

Fig. 4 (c) is Haar-like point feature schematic diagram;

Fig. 4 (d) is Haar-like diagonal line feature schematic diagram;

Fig. 5 is the schematic diagram that original LBP operator calculates;

Fig. 6 is the schematic diagram of gait profile conversion complex coordinates system;

Fig. 7 is that a gait cycle carries out the curve map that the swinging distance that obtains after mean filter is drawn;

Fig. 8 is PMS (Procrustes Mean Shape, average shape) the value schematic diagram that four sections of different gait sequences of two different targets obtain by calculating.

Embodiment

For technological means, creation characteristic that the present invention is realized, reach object and effect is easy to understand, below in conjunction with concrete diagram, further set forth the present invention.

For video monitoring, the present invention adopts the personal identification method based on certainly setting up Sample Storehouse and composite character, the method detects the image sequence of input by the good people's face of training in advance and pedestrian's sorter, the mode of utilizing man-machine interaction to the front face sample detecting and side gait sample carry out class indication as cognitive phase the sample for " training ", recycle these sample extraction people's face or gait feature trainable recognizer, detected other sample is carried out to class indication, if have and fail people's face or the gait of class indication, then inquire that user identifies, and the people's face sample having identified or gait sample are added in corresponding Sample Storehouse, be used for training corresponding recognition of face device or Gait Recognition device.

Because Gait Recognition belongs to Dynamic Recognition, need the data of at least one gait cycle in consecutive image sequence frame could extract complete reliable gait feature and complete identification,

Because the correct recognition rata of recognition of face also will be apparently higher than Gait Recognition, therefore in the present invention, adopting recognition of face is that main Gait Recognition is that auxiliary mode is carried out identification, if that is: complete front face image do not detected in current image frame, carry out pedestrian detection and tracking and extract gait feature and carry out identification.

Referring to Fig. 1, it is depicted as the particular flow sheet that carries out identification based on above-mentioned principle.As seen from the figure, whole identifying is divided into following six aspects:

1. moving target extracts

First the image sequence collecting is carried out to pre-service, after pre-service, obtain moving region information.Wherein, pre-service comprises: (1) contrast strengthens; (2) edge strengthens; (3) gray scale stretches; (4) moving target extracts.

When obtaining moving target, the present invention utilizes the background subtraction method based on ViBe to extract the moving target in current scene, because ViBe technology is based on random theory, that is to say that the background model result of at every turn obtaining is all slightly different, while therefore obtaining in the present invention moving target information, the foreground mask of twice acquisition is carried out to logical “and” operation, reduce issuable error detection in foreground detection process, referring to Fig. 2, two prospect masks that it is depicted as that background extracting technology by ViBe obtains and logical “and” result, as can be seen from the figure the result of logical “and” visually will obviously be better than preoperative two the prospect masks of logical “and” aspect false drop rate.

In order to improve the integrality of moving target, the minimum area-encasing rectangle that the result of also extracting in conjunction with moving target in the present invention obtains, utilizes Grabcut technology to cut apart complete moving target, and design sketch as shown in Figure 3.

2. people's face detects;

In the moving region getting (being foreground information), carry out the detection of people's face, if people's face information detected, proceed to step (3); If people's face information do not detected, proceed to step (4).

In the present invention, the detection of people's face is that the sorter that utilizes the Haar-like feature of sample in the Sample Storehouse extracting to obtain by Adaboost Algorithm for Training detects, and the Haar-like feature adopting is (Fig. 4 a to Fig. 4 b) as shown in Figure 4.People's face used herein detects with pedestrian detection sorter " non-face " sample in training process and has used " bootstrapping (bootstrap) " method: one of model is only used the preliminary classification device of " people's face " sample and a small amount of " non-face " sample training to detect one group of image, and the image that is people's face the wrong detection of all non-face quilts adds " non-face " Sample Storehouse; Then " the people's face " that use obtains and " non-face " sample training are constructed new sorter and are re-started detection.Above process constantly circulates, until collected enough " non-face " samples.

Adaboost sorter is to be formed by the cascade of multilayer Weak Classifier, and the correct result of classifying out from ground floor sorter triggers second layer sorter, and the correct result of classifying out from the second layer triggers three-layer classification device, by that analogy.Wherein the result being denied of any one node output all can cause the detection of this subwindow to stop immediately.By the threshold value of every layer is set, make most people's faces can by and non-face can not passing through, near the layer of cascade classifier rear end, refused most non-face like this.Adaboost algorithm is trained same basic classification device (Weak Classifier) for different training sets, and the sorter then these being obtained on different training sets gathers, and forms a stronger final sorter (strong classifier).As long as each Weak Classifier classification capacity is better than random conjecture, when number trends towards infinite number, the error rate of strong classifier will trend towards zero.

In Adaboost algorithm, different training sets is to realize by adjusting the weight that each sample is corresponding.When initial, the weight that each sample is corresponding is identical, trains a basic classification device h under this sample distribution ₁(x).For h ₁(x) sample of wrong minute, increases the weight of its corresponding sample; And for the sample of correct classification, reduce its weight.Like this can be so that the sample of wrong minute highlight, and obtain a new sample distribution.The situation of dividing according to mistake is given h simultaneously ₁(x) weight, represents that the significance level of this basic classification device, mistake are got fewer weight larger.Under new sample distribution, again basic classification device is trained, obtain basic classification device h ₂and weight (x).The like, through T such circulation, just obtained the weight of T basic classification device and their correspondences.Finally this T basic classification device added up by corresponding weight, just obtained final desired strong classifier, specifically describe as follows.

Suppose that X represents sample space, Y represents the set of sample class sign, supposes it is two-value classification problem, limits Y={-1 here ,+1}.Make S={ (x _i, y _i) | i=1,2, L, m} is sample training collection, wherein x _i∈ X, y _i∈ Y.The weights of an initialization m sample, suppose sample distribution D _tfor being uniformly distributed: D _t(i)=1/m, D _t(i) be illustrated in every wheel and in iteration, be assigned to sample (x _i, y _i) weights.According to sample distribution D _t, by training set S is sampled, produce training set S _t.At training set S _tupper training classifier h _t, use sorter h _tto all sample classifications in former training set S.Obtain the sorter h of epicycle iteration _t: X → Y, and have error make α _t=ln[(1-ε _t)/ε _t]/2, by following formula, upgrade the weights of each sample:

$D_{t+1}\left(i\right)=\frac{D_t\left(i\right)}{Z_t}\×\left\{\begin{array}{cc}{e}^{-{\mathrm{\α}}_i},& \mathrm{if}{h}_t\left(x_i\right)=y_i\\ e^{{\mathrm{\α}}_i},& \mathrm{if}{h}_t\left(x_i\right)\≠y_i\end{array}\right.$

Wherein, Z _tbe a normalization factor, be used for guaranteeing ∑ _id _t+1(i)=1.Complete after T iteration, final prediction is output as:

$H\left(x\right)=\mathrm{sign}\left(\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\α}}_t{h}_t\left(x\right)\right)$

3. recognition of face

After people's face information being detected, be normalized the rear LBP operator extraction feature of utilizing improved border circular areas, compare identification with existing people's face sample characteristics, and export recognition result, if and existing people's face sample does not mate or people's face Sample Storehouse is sky, inquire that user identifies the people's face detecting, and the people's face sample having identified is added in Sample Storehouse.

In the present invention, utilize LBP (Local Binary Pattern) technology to identify front face, this technology is the earliest for describing Local textural feature, as shown in Figure 5, original LBP operator definitions is in 3 * 3 window, take window center pixel as threshold value, the gray-scale value of 8 adjacent pixels is compared with it, if its pixel value is greater than center pixel value, the position of this pixel is marked as 1, otherwise is 0.Like this, 8 points in 3 * 3 neighborhoods can produce the unsigned number of a 8bit, obtain the LBP value of this window, and this value is used for reflecting the texture information in this region.

LBP operator LBP after improvement _p,Rbe defined as, the border circular areas that radius is R contains P sampled point, and its expression formula is:

${\mathrm{LBP}}_{P,R}=\underset{P=0}{\overset{P-1}{\mathrm{\&Sigma;}}}s(g_p-g_c)2^P,s(g_p-g_c)=\left\{\begin{array}{cc}1,& g_p\&GreaterEqual;g_c\\ 0,& g_p<g_c\end{array}\right.$

Wherein, g _crepresent center (x in region _c, y _c) gray-scale value located, g _prepresent that the center that is evenly distributed on is for (x _c, y _c) radius is the gray-scale value of on R circumference P point.Carrying out people's face while describing, first people's face is divided into some, then generate the LBP histogram of each piece, finally the LBP histogram of all is coupled together as feature histogram, and be described as according to carrying out Classification and Identification with this.

4. pedestrian detection

If people's face do not detected, in foreground information, carry out pedestrian detection.In the present invention, pedestrian detection adopts AdaBoost Algorithm for Training sorter equally, and detected characteristics adopts histograms of oriented gradients (Histograms of Oriented Gradients, HOG).HOG descriptor is that the picture that the gait image of input is normalized to 64 * 128 sizes is divided, and the pixel region of 4 * 4 sizes is divided into a unit, carries out gradient statistics in each junior unit.

Gradient direction, in [0 °, 360 °], is divided into 9 grades by histogram, calculates the gradient orientation histogram with gradient magnitude weighting that each unit is corresponding, is expressed as the proper vector of one 9 dimension.Mesh merging is above become to large region, and 2 * 2 unit form a new region, and each region is 8 * 8 pixels.In order to comprise as far as possible many information, while dividing in the territory of original image, two adjacent regions are overlapped.Lap size is 50% of former region, and the picture of 64 * 128 pixels can be divided the region of 7 * 15 8 * 8 pixels so.The proper vector of all unit is tied, i.e. composing images characteristic of correspondence vector.

Therefore, the overall vector of the picture of 64 * 128 becomes the vector of one group of 3780 dimension.By the histogram ability of whole image, proper vector is normalized, normalization operator is:

${V_i}^{*}=V_i/\sqrt{\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{V_i}^2+\mathrm{\&epsiv;}}$

5. pedestrian follows the tracks of and gait cycle detection

The pedestrian that tracking detects, to obtain the image sequence of its walking, and carries out gait cycle detection.

In the present invention, utilize meanshift algorithm to carry out pedestrian's tracking, meanshift is a kind of Density Estimator algorithm, it moves to each point at the Local modulus maxima place of density function, the point (mode point) that density gradient is 0, and multidimensional Density Estimator can be expressed as:

$f\left(x\right)=\frac{1}{{\mathrm{nh}}^d}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}K\left(\frac{x-x_i}{h}\right)$

Scaling function K (x) should meet:

$\left\{\begin{array}{c}\underset{{\mathrm{reR}}^i}{\sup }\left|K\left(y\right)\right|<\&infin;\\ {\&Integral;}_{R^i}\left|K\left(y\right)\right|\mathrm{dy}<\&infin;\\ {\&Integral;}_{R^i}K\left(x\right)\mathrm{dx}=1\\ \underset{\left|\right|y\left|\right|\&RightArrow;\&infin;}{\lim }{\left|\right|y\left|\right|}^{d}K\left(y\right)=0\end{array}\right.$

Wherein, || * || represent Euclidean distance, R ^drepresent d dimension space, parameter h is the function of sample size n, should meet thereby guarantee the progressive unbiasedness of estimating.In order to ensure all square consistance of estimating, h need to satisfy condition in order to ensure global coherency, h need to satisfy condition when kernel function adopts following Epanechiov core, integrated square error is minimum.

$K_E\left(x\right)=\left\{\begin{array}{cc}\frac{1}{2}{c}_d^{-1}(d+2)(1-{\left|\right|x\left|\right|}^2),& \mathrm{if}\left|\right|x\left|\right|<1\\ 0,& \mathrm{if}\left|\right|x\left|\right|>1\end{array}\right.$

Wherein, c _dfor d dimension unit spheroid volume.Another kernel function being often used to is gaussian kernel: definition kernel function profile function k meet k:[0, ∞) → R, K (x)=k (|| x|| ²), therefore, Density Estimator function can be rewritten as following form:

$f\left(x\right)=\frac{1}{{\mathrm{nh}}^d}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}k\left({\left|\right|\frac{x-x_i}{h}\left|\right|}^2\right)$

Definition g (x)=-k'(x), wherein, at interval x ∈, [0, ∞) the upper minority finite point place that removes all exists the first order derivative of suppose k (x).Definition kernel function G (x)=Cg (|| x|| ²), wherein C is normaliztion constant.Utilize the differentiability of kernel function, Density gradient estimation is constantly equal to the gradient of Density Estimator, by following formula:

${\&dtri;f}_K\left(x\right)\&equiv;\&dtri;f\left(x\right)=\frac{2}{{\mathrm{nh}}^{d+2}}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}(x-x_i)k^{\&prime;}\left({\left|\right|\frac{x-x_i}{h}\left|\right|}^2\right)$

Can obtain $\frac{2}{{\mathrm{nh}}^{d+2}}g\left({\left|\right|\frac{x-x_i}{h}\left|\right|}^2\right)[\frac{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_{i}g\left({\left|\right|\frac{x-x_i}{h}\left|\right|}^2\right)}{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}g\left({\left|\right|\frac{x-x_i}{h}\left|\right|}^2\right)}-x],$ Wherein be not 0, latter one is referred to as meanshift vector, that is:

$M_{k,G}\left(x\right)=\frac{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_{i}g\left({\left|\right|\frac{x-x_i}{h}\left|\right|}^2\right)}{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}g\left({\left|\right|\frac{x-x_i}{h}\left|\right|}^2\right)}-x$

From this formula, be not difficult to find out, the density Estimation of ordering at x is expressed as with kernel function G:

$f_G\left(x\right)=\frac{C}{{\mathrm{nh}}^d}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}g\left({\left|\right|\frac{x-x_i}{h}\left|\right|}^2\right)$

Density gradient estimation becomes:

${\&dtri;f}_K\left(x\right)={\&dtri;f}_G\left(x\right)=\frac{2/C}{{\mathrm{nh}}^{d+2}}{M}_{k,G}\left(x\right)$

So this formula shows, at an x place, the Meanshift vector based on kernel function G (x) and the Density gradient estimation based on kernel function K (x) only differ from the scale-up factor of a constant.And gradient refers to the direction of variable density maximum.So meanshift vector also always points to the maximum direction that density increases.

Because the needed time of position is farthest stepped out forward to next same leg in the position that a gait cycle is defined as stepping out forward farthest from one leg, so in the present invention, gait cycle detection is to extract gait cycle by calculating the swinging distance of gait sequence.Each width gait profile diagram in a gait sequence is calculated to its swinging distance:

$\mathrm{sw}={\mathrm{\&Sigma;}}_{y=y_b}^{y_t\frac{(y_t-y_c)}{2}}{\mathrm{\&Sigma;}}_{x=x_l}^{x_r}|(x-x_c)\&times;\frac{I(x,y)}{255}|$

(x wherein _c, y _c) be the barycenter of side profile, y _bthe ordinate of human body lowermost end, y _tthe ordinate of human body top, x _lthe horizontal ordinate of the human body leftmost side, x _rbe the horizontal ordinate of the human body rightmost side, I (x, y) is the brightness value of current point.Referring to Fig. 7, shown in it, be the swinging distance carrying out after mean filter, as can be seen from the figure swinging distance has obvious periodicity, and the moment that each maximum value occurs is that two legs separates the maximum moment, in the present invention, use maximum value computation period, the position that odd number maximum point occurs.

6. Gait Recognition

To following the tracks of the gait sequence of the gait cycle obtaining, extract gait feature, and compare identification with existing gait sample characteristics, and export recognition result, if and existing gait sample does not mate or gait Sample Storehouse is sky, inquire that user identifies the gait feature obtaining, and the gait information having identified is added in Sample Storehouse.Its concrete identifying is as follows:

For every frame gait image, once after pedestrian's motion silhouette is extracted, can obtain the profile border of human body silhouette by the edge following algorithm based on connective, establish (x _i, y _i) be any point on border, the barycenter (x of human body contour outline _c, y _c) can be tried to achieve by following formula:

$\left\{\begin{array}{c}{x}_c=\frac{1}{K}\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{x}_i,\\ \mathrm{yc}=\frac{1}{K}\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{y}_i\end{array}\right.$

Wherein, K represents borderline sampling number.K is larger, and profile expression is more accurate, and corresponding calculated amount also can increase.For the borderline sampled point of more succinct expression, in the present invention, rectangular coordinate system is mapped to complex coordinates system, with barycenter (x _c, y _c) as the initial point in two-dimensional shapes space, take level to the right as real axis, straight up as the imaginary axis is set up new complex coordinates system.Under this coordinate system, contour shape to be separated in the counterclockwise direction to coiling and launched to become a pixel point set, borderline point can represent to be shaped as " x _i+ j * y _i" form, whole border in two-dimentional complex number space, also can be expressed as one mark order complex vector located, that is:

u＝[u ₁?u ₂?L?z _i?L?u _k] ^T,zi＝(x _i-x _c)+j×(y _i-y _c)

Wherein, j is imaginary unit.Human body contour outline figure after converted coordinate system as shown in Figure 6.

A given gait sequence with m width image, can obtain m individual similarly complex vector located.In order to obtain the compact representation of gait sequence, the present invention adopts PSA (Procrustes Shape Analysis, Statistical Shape analysis) technology to describe its characteristic to continuous variation campaign or the structural statistical presentation of human body contour outline.PSA method is calculated the formal shape PMS of average (Procrustes Mean Shape, average shape) of this sequence by following method.First calculate S _umatrix:

$S_u=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left(u_i{u}_i^{*}\right)/\left(u_i^{*}{u}_i\right)$

Wherein, subscript " * " represents complex conjugate transposition.To S _umatrix carries out svd, and the corresponding proper vector of eigenvalue of maximum obtaining is PMS, and referring to Fig. 8, it is the PMS schematic diagram of four gait sequences of two Different Individual.After obtaining PMS, it is carried out to Fourier spectrum analysis, using the amplitude spectrum of PMS as final feature, and utilize Euclidean distance to characterize the similarity of two different characteristics, identify.

The present invention solves the identification in video monitoring by above six steps.The ViBe technology of utilization based on random theory obtains foreground information, and the foreground extraction result of twice is carried out to logical “and” operation, after the minimum area-encasing rectangle of acquisition motion unit, in conjunction with grabcut technology, obtains complete foreground information; Utilize the face classification device that Haar-like feature obtains in conjunction with Adaboost Algorithm for Training to carry out the detection of people's face to foreground information; Utilize improved LBP operator that the people's face detecting is described and is identified, if the Sample Storehouse for trainable recognizer is that sky or None-identified are worked as forefathers' face, adopt the mode of man-machine interaction to identify the people's face detecting, and the people's face information having identified is added in Sample Storehouse; If people's face do not detected in foreground information, the pedestrian's sorter that utilizes HOG descriptor and Adaboost Algorithm for Training to obtain carries out pedestrian detection; Utilize Meanshift technology to follow the tracks of the pedestrian who detects, and carry out gait cycle detection; Utilize PMS that the gait information detecting is described and is identified, if gait Sample Storehouse is empty or current gait information None-identified, adopt the mode of man-machine interaction to identify the gait detecting.

More than show and described ultimate principle of the present invention, principal character and advantage of the present invention.The technician of the industry should understand; the present invention is not restricted to the described embodiments; that in above-described embodiment and instructions, describes just illustrates principle of the present invention; without departing from the spirit and scope of the present invention; the present invention also has various changes and modifications, and these changes and improvements all fall in the claimed scope of the invention.The claimed scope of the present invention is defined by appending claims and equivalent thereof.

Claims (10)

1. the personal identification method based on certainly setting up Sample Storehouse and composite character in video monitoring, it is characterized in that, described method first detects the image sequence of input by the good people's face of training in advance and pedestrian's sorter, the mode of utilizing man-machine interaction to the front face sample detecting and side gait sample carry out class indication as cognitive phase the sample for " training ", recycle these sample extraction people's face or gait feature trainable recognizer, detected other sample is carried out to class indication, if have and fail people's face or the gait of class indication, then inquire that user identifies.

2. the personal identification method based on certainly setting up Sample Storehouse and composite character in video monitoring according to claim 1, it is characterized in that, in detecting identifying, adopting recognition of face is that main Gait Recognition is that auxiliary mode is carried out identification, complete front face image in current image frame, do not detected, carry out pedestrian detection and tracking and extract gait feature and carry out identification; While carrying out Gait Recognition, for consecutive image sequence frame at least comprise the data of a gait cycle.

3. the personal identification method based on certainly setting up Sample Storehouse and composite character in video monitoring according to claim 2, is characterized in that, specifically the comprising the steps: of described personal identification method

(1) image sequence collecting is carried out to pre-service, after pre-service, obtain moving region information;

(2) in the moving region getting (being foreground information), carry out the detection of people's face, if people's face information detected, proceed to step (3); If people's face information do not detected, proceed to step (4);

(3) if people's face information detected, after normalization, utilize LBP operator extraction feature, and existing people's face sample characteristics compares identification, and export recognition result; If None-identified current detection to people's face information (do not mate with existing people's face sample or people's face Sample Storehouse for empty), inquire that user identifies, and the people's face sample having identified added in Sample Storehouse;

(4), if people's face do not detected, in moving region (being foreground information), carry out pedestrian detection;

(5) follow the tracks of the pedestrian who detects, obtain the image sequence of its walking, and carry out gait cycle detection;

(6) to following the tracks of the gait sequence of the gait cycle obtaining, extract gait feature, and compare identification with existing gait sample characteristics, and export recognition result; If the current gait feature extracting of None-identified (do not mate with existing gait sample or gait Sample Storehouse for empty), inquires that user identifies, and the gait information having identified is added in Sample Storehouse.

4. the personal identification method based on certainly setting up Sample Storehouse and composite character in video monitoring according to claim 3, it is characterized in that, the pretreatment operation in described step (1) comprises: contrast strengthens, edge strengthens, gray scale stretches, moving target extracts.

5. according to the personal identification method based on certainly setting up Sample Storehouse and composite character in the video monitoring described in claim 3 or 4, background subtraction method based on ViBe in described step (1) is extracted the moving target information in current scene, and when obtaining moving target information, the foreground mask of twice acquisition is carried out to logical “and” operation; The minimum area-encasing rectangle of the moving target obtaining while simultaneously also utilizing grabcut technology to extract in conjunction with moving target obtains complete moving target information.

6. the personal identification method based on certainly setting up Sample Storehouse and composite character in video monitoring according to claim 3, while carrying out the detection of people's face in described step (2), the sorter that in the Sample Storehouse that utilization is extracted, the Haar-like feature of sample obtains by Adaboost Algorithm for Training detects, and training process is used " bootstrapping " method.

7. the personal identification method based on certainly setting up Sample Storehouse and composite character in video monitoring according to claim 3, while carrying out recognition of face in described step (3), utilize the LBP operator of improved border circular areas, the people's face detecting is divided into some, then generate the LBP histogram of each piece, finally the LBP histogram of all is coupled together as feature histogram, and be described as according to carrying out Classification and Identification with this.

8. the personal identification method based on certainly setting up Sample Storehouse and composite character in video monitoring according to claim 3, while carrying out pedestrian detection in described step (4), sample in Sample Storehouse is normalized to the picture of 64 * 128 sizes, extract thus and obtain histograms of oriented gradients corresponding to sample in Sample Storehouse, and using this sorter that utilizes Adaboost Algorithm for Training to obtain as detected characteristics and detect.

9. the personal identification method based on certainly setting up Sample Storehouse and composite character in video monitoring according to claim 3, in described step (5), utilize meanshift algorithm to follow the tracks of the pedestrian who detects, to following the tracks of the gait image sequence obtaining, utilize each width gait profile diagram to calculate its swinging distance, and utilize odd number maximum point to determine gait cycle.

10. the personal identification method based on certainly setting up Sample Storehouse and composite character in video monitoring according to claim 3, while carrying out gait feature identification in described step (6), utilize the profile feature modeling profile barycenter of gait, and by profile diagram be transformed into take barycenter as round dot, level to the right as real axis, to hold up be upwards the complex coordinates system of the imaginary axis, utilizes PSA technology to move to profile or structural statistical presentation is described its characteristic and obtained PMS and carries out identification after conversion.