CN105955473A - Computer-based static gesture image recognition interactive system - Google Patents

Abstract

The invention provides a computer-based static gesture image recognition interactive system and belongs to the field of man-machine interaction. The system comprises a gesture image acquisition module, an image preprocessing module, an image element acquisition module, a farthest point characteristic information acquisition module, an area characteristic information acquisition module and a static gesture recognition module, wherein the gesture image acquisition module processes an image captured by a camera to obtain a standard gesture image; the image preprocessing module carries out bounding volume processing on the obtained standard gesture image and projecting the obtained gesture image to a standard image; the image element acquisition module acquires a centroid, a farthest point and a main direction of a gesture region in the standard image; the farthest point characteristic information acquisition module acquires the characteristic information of the farthest point of the twelve regions in the standard image; the area characteristic information acquisition module acquires area characteristic information of the standard image; and the static gesture recognition module recognizes a static gesture by using a PCA (Principal Component Analysis) algorithm.

Description

A kind of computer based static gesture image recognition interactive system

Technical field

The present invention relates to a kind of gesture recognition system, belong to field of human-computer interaction, be specifically related to a kind of computer based Static gesture image recognition interactive system.

Background technology

At present, gesture identification is divided into two kinds of gesture identification modes based on wearable device and view-based access control model.Based on wearing Wear the gesture identification of equipment mainly based on data glove as the gesture identification of input equipment, and the gesture identification of view-based access control model Then have only to single or multiple photographic head as input equipment hands.The advantage of gesture identification based on wearable device is to obtain Gesture data precision high, its shortcoming is that equipment price is expensive, uses inconvenience.And the gesture recognition system of view-based access control model is because calculating Process is complicated, obtains image relatively big by illumination effect, thus causes discrimination and real-time the most poor；Its advantage is equipment price Relatively low and do not disturb user behavior, study uses simple and flexible, natural alternately.

The gesture identification of view-based access control model mainly has following several method.Gesture identification based on neutral net, is characterized in Self-organizing, self study, anti-interference, because of indifferent to time Series Processing, it is used for static gesture identification；Based on hidden Ma Erke The gesture identification of husband's model, this is a kind of Statistic analysis models, the change in time and space of description hand signal that can be the most careful, generally For dynamic hand gesture recognition；Gesture identification method based on geometric properties, mainly chooses the geometric moment feature of gesture, edge wheel Wide feature or gesture area feature carry out feature point extraction to image, then carry out template matching (Model according to various distances Matching), such as Hausdorff distance and Euclidean distance etc..

E.Stergiopoulou etc. propose a kind of based on gesture self-adjusting, self-organizing, adaptive neutral net (Neural Gas network), this neutral net, by the extraction of the feature of gesture and process, the most successfully identifies Gesture；Heung-ll Suk etc. proposes a kind of dynamic bayesian network being identified the gesture in video, the method Can to gesture in video with similar in the case of gesture successfully identify；Noriko Yoshiike etc. are by using Big neutral net is partitioned into images of gestures, thus carries out gesture identification；Li Shaozhi etc. propose automatic encoding and principal component analysis (refer to: Wang Song, Xia Shaowei. a kind of Robust Principal Component Analysis (PCA) algorithm [J]. the system engineering theory and practice, 1998,18 (1): 9-13.) method combined is for the identification of U.S.'s representation language (American sign language, ASL), makes successfully Rate is promoted to 99.05% by 75%.

Daehwan Kim etc. are for gesture identification identification after Hand Gesture Segmentation in solution video dynamic hand gesture recognition not in time Property, it is proposed that the accumulation HMM algorithm of a kind of forward marker so that the segmentation of dynamic gesture and identification in video Carrying out, accuracy of identification is up to 95.42% simultaneously；Tim Dittmar etc. proposes a kind of HMM that converts for touching Touch gesture identification, respond well；Old Feng Sheng et al. proposes the HMM of a kind of real-time Hand Gesture Segmentation, successfully knows Not 20 kinds of gestures and discrimination are up to 90%.

But, the subject matter being currently based on the existence of visual gesture recognizer is: obtain not of uniform size, the image of image The anglec of rotation different, the problem such as Image Reversal problem and gesture identification real-time.

Summary of the invention

It is an object of the invention to solve images of gestures present in above-mentioned prior art rotate, upset and not of uniform size Problem, it is provided that a kind of computer based static gesture image recognition interactive system.

The present invention is achieved by the following technical solutions:

A kind of computer based static gesture image recognition interactive system, it is characterised in that including:

Images of gestures acquisition module: the image for being captured by photographic head is processed, it is thus achieved that the gesture figure of standard Picture；

Image pre-processing module: the images of gestures of the standard of acquisition is carried out bounding box process, and is projected into standard drawing Picture；

Pictorial element acquisition module: obtain the gesture area centre of form, solstics and the principal direction in standard picture；

Obtain solstics characteristic information module: obtain solstics, standard picture 12 region characteristic information；

Obtain area features information module: obtain the area features information of standard picture；

Static gesture identification module: use PCA algorithm to carry out static gesture identification.

As the further restriction to the technical program, described images of gestures acquisition module is achieved in that

Image for being captured by photographic head is used based on RGB and YCbCr color space complexion model gesture district Territory carries out dividing processing, obtains images of gestures；

Described images of gestures is carried out following process:

Twice eight field denoisings, the noise in eliminating image；

Corrosion treatmentCorrosion Science, makes the connected domain being slightly connected in images of gestures be divided into two independent connected domains；

Connected area disposal$, and calculate the area of each connected domain, bigger non-gesture area is background process, Ji Jiangfei Gesture area is set to black, and gesture area is set to redness；

Image after processing carries out expansion process, and reduction is through the image of corrosion treatmentCorrosion Science.

As the further restriction to the technical program, described image pre-processing module is achieved in that

Internally scan from each limit of image successively, be the limit on this limit of bounding box when scan line encounters images of gestures Boundary, the rectangle that four borders surround is the bounding box of this images of gestures；

Region outside bounding box is invalid data region, and the region within bounding box is valid data region；

Bounding box area image is mapped on standard picture.

As the further restriction to the technical program, described being mapped on standard picture by bounding box area image is this Sample realizes:

Accepted standard image size is 100*100, specifically comprises the following steps that

Step1. size and standard image size according to described bounding box are according to formula (1.1) calculating scaling ratio:

$\begin{array}{c}zoomX=newWidth/width\\ zommY=newHeight/height\end{array}---\left(1.1\right)$

Wherein zoomX, zoomY are respectively wide and high scaling ratio, and newWidth, newHeight are respectively standardization figure The length of side of picture, width, height are width and the height of source images；

Step2. the images of gestures in bounding box is zoomed in standard picture according to formula (1.2):

$\left[\begin{array}{c}{x}^{\′}\\ y^{\′}\end{array}\right]=\left[\begin{array}{cc}zoomX& 0\\ 0& zoomY\end{array}\right]\left[\begin{array}{c}x\\ y\end{array}\right]---\left(1.2\right)$

The coordinate of the pixel in image after wherein (x ', y ') is standardization, (x y) is the coordinate of pixel in source images Value.

As the further restriction to the technical program, described pictorial element acquisition module is achieved in that

Calculate abscissa and the vertical coordinate sum of each pixel of gesture area in standard picture the most respectively, and calculate Gesture area pixel number, represents the area of gesture area with this；

Step2. ask for barycentric coodinates according to asking for center of gravity formula (1.3), be the coordinate of the gesture area centre of form:

$\left\{\begin{array}{c}\stackrel{\‾}{x}=\frac{\underset{A}{\∫}ydA}{A}\\ \stackrel{\‾}{y}=\frac{\underset{A}{\∫}xdA}{A}\end{array}\right.---\left(1.3\right)$

Wherein, A represents images of gestures area,Represent centre of form abscissa value,Representing centre of form ordinate value, dA is integration Element；

Solstics is that images of gestures area pixel point is to centre of form distance point furthest；

Principal direction is the line connecting centroid point to solstics, and direction is to point to solstics from centroid point.

As the further restriction to the technical program, described acquisition solstics characteristic information module is achieved in that

Centered by the gesture area centre of form, with 30 degree of angles for the anglec of rotation, gesture area is divided into 12 regions, Qi Zhong One region is the region of 15 degree of angle compositions about principal direction, from the beginning of first area, counter clockwise direction is followed successively by second area To territory, No.12 District, the characteristic information step obtaining 12 solstics, region is as follows:

Step1. it is calculated the solstics in each region, and calculates the solstics distance to the centre of form, compare and obtain wherein Ultimate range；

Step2. ultimate range is equally divided into 5 sections, then obtains 5 packets, the distance centre of form nearest for first group, distance The centre of form farthest for the 5th group；

Calculate 12 solstics the most respectively and fall the number of times of 5 packets；

Step4. by the number of times falling 5 regions obtained divided by 12 to uniform 12 solstics, region characteristics；? To 12 solstics, the region characteristic vector U={ μ of a length of 5₁,μ₂,μ₃,μ₄,μ₅}。

As the further restriction to the technical program, described acquisition area features information module is achieved in that

Step1. calculate the distance in gesture solstics and the centre of form, and this range averaging is divided into 12 sections, then obtain 12 points Group, distance the centre of form nearest for first group, distance the centre of form farthest for the 12nd group；

Step2. with 12 region criteria for classifying images, and the centre of form in each region is calculated；

Do the centre of form of gesture area the most successively by the centre of form in 12 regions, calculate solstics now and principal direction；

Step4. calculate the distance in the centre of form and solstics, and be averaged and be divided into 5 parts, obtain 5 packets, by whole circumference It is divided into 12 parts, i.e. from the beginning of principal direction, to be divided into a region for direction every 30 degree of angles counterclockwise, so obtains 12 districts Territory, 5 packets and 12 regions intersections then obtain 60 boxed area, calculate gesture area and fall in the pixel of each boxed area The number of point, the area of the most each boxed area；

Calculate the maximum area often organized the most respectively, the most often organize the maximum area all divided by this group, i.e. obtain homogeneous The area features value changed.

As the further restriction to the technical program, described static gesture identification module is achieved in that

Step1, each gesture contains solstics, standard picture 12 region characteristic information, is saved as template, if often The characteristic vector of individual template is U={ μ₁,μ₂,μ₃,μ₄,μ₅, the characteristic vector of gesture to be identified is N={ η₁,η₂,η₃,η₄,η₅, Use Euclidean distance as the discrimination of two characteristic vectors, calculate gesture to be identified and template gesture according to formula (4.1) Discrimination diff={ β₁,β₂,…β₉}:

$diff=\underset{i=1}{\overset{5}{\Σ}}\left|\right|{\mathrm{\μ}}_i-{\mathrm{\η}}_i|{|}^2---\left(4.1\right)$

Being calculated 5 minimum discriminations, gesture to be identified is the one in these 5 kinds of gestures；

Step2, PCA gesture identification:

Structural feature gesture space: for the images of gestures of a width M × N, one size of the composition that is connected with tail by its head is D The column vector of=M × N-dimensional, D is exactly the dimension of images of gestures, that is to say the dimension of image space, if n is the number of training sample Mesh, x_jRepresent the characteristic information column vector that jth width images of gestures is formed, the then covariance matrix of institute's sample, formula (3.1) obtain Go out:

$S_r=\underset{j=1}{\overset{n}{\Σ}}(x_i-u){(x_j-u)}^T---\left(3.1\right)$

Wherein u is the average image vector of training sample, formula (3.2) draw:

$u=\frac{1}{n}\underset{j=1}{\overset{n}{\Σ}}{x}_j---\left(3.2\right)$

Make A=[x₁-u,x₂-u,...,x_n-u], then there is S_r=AA^T, its dimension is D × D；

According to the 5 kinds of gestures determined, gesture to be identified among these 5 kinds of gestures, these 5 kinds of handss in extraction feature matrix A Gesture characteristic of correspondence vector, these characteristic vectors are by eigenmatrix A '={ X new for composition₁,X₂,…X₅₀}^T, wherein X_iIt it is the i-th width Training sample image seeks character pair vector, equally image to be identified is carried out feature extraction, obtains characteristic vector B={x₁, x₂,…x_i, wherein i is by through information computing being taken out front i character vector, so obtaining gathering A ' and B two Individual set, uses Euclidean distance formula (4.2) to judge set B and gathers a degree of characteristic vector in A ':

${\mathrm{\θ}}_j=\underset{i=1}{\overset{i}{\Σ}}\left|\right|{\mathrm{\μ}}_i-x_i|{|}^2---\left(4.2\right)$

And calculate matching degree set diff_final={ θ₁,θ₂,…θ₄₉, θ₅₀, then ask for the minima of θ, minimum The images of gestures of θ value correspondence is final recognition result.

Compared with prior art, the invention has the beneficial effects as follows: the feature extracting method based on area that the present invention uses, Different from feature based on contours extract, need to extract gesture internal area information based on area gesture feature, and finally use PCA (Principal Component Analysis, PCA) carries out gesture identification, and is successfully identified, Solve images of gestures to rotate, overturn and problem not of uniform size.

Accompanying drawing explanation

Fig. 1 is through processing the image of the gesture area obtained

Fig. 2 processes the bounding box images of gestures obtained through OBBs algorithm

Fig. 3 original image is mapped as standard picture

The centre of form of Fig. 4 gesture

9 kinds of gestures to be identified for Fig. 5

The step block diagram of Fig. 6 the inventive method

Detailed description of the invention

Below in conjunction with the accompanying drawings the present invention is described in further detail:

Multiformity, polysemy, complex deformation, spatio-temporal difference are gesture have the feature thats.The present invention uses lift-launch Photographic head in Android platform catches the image comprising gesture area, the image caught is carried out images of gestures segmentation with Obtain images of gestures information.Conventional Hand Gesture Segmentation method has skin color segmentation method, background subtraction, grey level histogram detection method. Using single dividing method cannot obtain good segmentation effect images of gestures, therefore to obtain good segmentation effect needs Image is split in conjunction with multiple method simultaneously.The effect of images of gestures segmentation directly influences the identification essence of gesture identification Degree.

The present invention as shown in Figure 6, including with lower module:

1, images of gestures acquisition module

Image for being caught by photographic head is used based on RGB and YCbCr color space complexion model gesture area Carry out dividing processing, obtain images of gestures, now split the images of gestures obtained and contain noise and bigger non-gesture area.

First images of gestures obtained in the previous step is carried out twice eight field denoisings, through this step, the noise in image To be eliminated；Secondly image obtained in the previous step is carried out corrosion treatmentCorrosion Science, make the connected domain being slightly connected in images of gestures divide It is two independent connected domains, then images of gestures is done Connected area disposal$, and calculate the area of each connected domain, by bigger Non-gesture area does background process, will be set to black by non-gesture area, and gesture area is set to redness；Finally to the figure after processing As carrying out expansion process with reduction through the image of corrosion treatmentCorrosion Science.As it is shown in figure 1, background is set to black, gesture area is set to red Color.

Image pre-processing module: the images of gestures of the standard of acquisition is carried out bounding box process, and is projected into standard drawing Picture, specific as follows:

OBBs algorithm is a kind of method solving discrete point set Optimal packet confining space.Basic thought be with volume slightly larger and The simple solid of characteristic (referred to as bounding box) replaces the geometric object of complexity approx.The present invention uses Rectangular Bounding Volume.

Internally scan from each limit of image successively, be bounding box when scan line encounters images of gestures on this limit Border, the rectangle that four borders surround is the bounding box of this images of gestures.As shown in Figure 2.Region outside bounding box is nothing Effect data area, the region within bounding box is just valid data region.

For solving the images of gestures size of the seizure impact on accuracy of identification during gesture identification, reduce picture number According to, improve gesture identification speed, bounding box area image is mapped on standard picture, accepted standard image size of the present invention For 100*100.Image standardization step is as follows:

Step1. size and standard image size according to bounding box obtained in the previous step calculate scaling according to formula (1.1) Ratio；

$\begin{array}{c}zoomX=newWidth/width\\ zommY=newHeight/height\end{array}---\left(1.1\right)$

Wherein zoomX, zoomY are respectively wide and high scaling ratio, and newWidth, newHeight are respectively standardization figure The length of side of picture, width, height are width and the height of source images.

Step2. the images of gestures in bounding box is zoomed in standard picture according to formula (1.2).Scaled results such as Fig. 3 Shown in, in Fig. 3, original image is mapped as standard picture, and the left side of figure is original image, and the right side of figure is the standard picture after mapping.

$\left[\begin{array}{c}{x}^{\′}\\ y^{\′}\end{array}\right]=\left[\begin{array}{cc}zoomX& 0\\ 0& zoomY\end{array}\right]\left[\begin{array}{c}x\\ y\end{array}\right]---\left(1.2\right)$

The coordinate of the pixel in image after wherein (x ', y ') is standardization, (x y) is the coordinate of pixel in source images Value, zoomX, the zoomY in formula can pass through formula (1.1) and obtain.It is not that each point of original image maps to standard Change on image, but be mapped to standard picture from original image selected part point, the most both can retain the spy of source images of gestures Levy and can greatly reduce the data volume of images of gestures.

Pictorial element acquisition module: obtain the gesture area centre of form, solstics and the principal direction in standard picture；

The centre of form of two dimensional image, can be tried to achieve by the center of gravity asking for two dimensional image, because image as unit point does not has weight, this Time images of gestures the centre of form with center of gravity overlap, the step asking for two-dimensional gesture image center of gravity is as follows:

Calculate the most respectively the horizontal stroke of gesture area pixel in standard picture, vertical coordinate and, and calculate gesture area picture Vegetarian refreshments number, represents the area of gesture area with this；

Step2. basis is asked for center of gravity formula (1.3) and is asked for barycentric coodinates, is gesture area centre of form coordinate.Such as Fig. 4 institute Show.

$\left\{\begin{array}{c}\stackrel{\‾}{x}=\frac{\underset{A}{\∫}ydA}{A}\\ \stackrel{\‾}{y}=\frac{\underset{A}{\∫}xdA}{A}\end{array}\right.---\left(1.3\right)$

Wherein, A represents images of gestures area,Represent centre of form abscissa value,Representing centre of form ordinate value, dA is integration Element.

Solstics is that images of gestures area pixel point is to centre of form distance point furthest.As shown in Figure 4, center picture in Fig. 4 The centre of form that Grey Point is gesture of point, the Grey Point in the upper right corner is the solstics of this image.

Principal direction is the line connecting centroid point to solstics, and direction is to point to solstics from centroid point.Ask for image Principal direction is the rotational invariance in order to solve image, it is ensured that the feature obtained after gesture rotation and the feature phase not rotating acquirement With.

2, obtain solstics characteristic information module: obtain solstics, standard images of gestures 12 region characteristic information.

Centered by the centre of form, for the anglec of rotation, gesture area being divided into 12 regions with 30 degree of angles, wherein first area is The region of 15 degree of angle compositions about principal direction.From the beginning of first area, counter clockwise direction is followed successively by second area to the 12nd Region.The characteristic information step then obtaining 12 solstics, region is as follows:

Step1. it is calculated the solstics in each region, and calculates the solstics distance to centroid point, compare and obtain it In ultimate range；

Step2. ultimate range is equally divided into 5 sections, then obtains 5 packets, the distance centre of form nearest for first group, distance The centre of form farthest for the 5th group；

Calculate 12 solstics the most respectively and fall the number of times in 5 regions；

Step4. by the number of times falling 5 regions obtained divided by 12 to uniform 12 solstics, region characteristics.

12 solstics, the region characteristic vector U={ μ of a length of 5 will be obtained through above-mentioned steps₁,μ₂,μ₃,μ₄,μ₅}。

Obtain area features information module: obtain standard images of gestures area features information.

Obtain the area features information characteristic information as PCA gesture identification of images of gestures.The basic ideas of the method It is: images of gestures is divided 60 districts of composition by the concentric circular with gesture area centroid point as the center of circle and the ray with the center of circle as starting point Territory, adds up the number area as this region that each region comprises gesture area pixel respectively, and calculates 60 regions Maximum area, processes the area features data homogenization obtained, thus obtains area features information.Obtain area features information Step is as follows:

Step1. calculate gesture solstics and centroid point distance, and this range averaging is divided into 12 sections, then obtain 12 points Group, distance the centre of form nearest for first group, distance the centre of form farthest for the 12nd group；

Step2. divide images of gestures with 12 regions, and calculate the centre of form in each region；

Do the centre of form of gesture area the most successively by the centre of form in 12 regions, calculate solstics now and principal direction；

Step4. calculate the distance of the centre of form and solstics, be averaged and be divided into 5 parts, obtain 5 packets, whole circumference is divided It is 12 parts, i.e. from the beginning of principal direction, to be divided into a region for direction every 30 degree of angles counterclockwise, so obtains 12 districts Territory, 5 packets and 12 regions intersections then obtain 60 boxed area, calculate gesture area and fall in the pixel of each boxed area The number of point, the area of the most each boxed area；

The most so obtaining 12 groups, often group comprises the area of 60 boxed area.Calculate the maximum area often organized respectively, The most often organize the maximum area all divided by this group, the most i.e. can obtain the area features value of homogenization.

3. static gesture identification module: use PCA algorithm to carry out static gesture identification.

The staff identification of Based PC A algorithm is typically through three phases: first stage utilizes training image data construct Feature hands space；Second stage is the training stage, is mainly projected on feature hands subspace by training image；Last is Cognitive phase, projects to equally by images of gestures to be identified on feature hands subspace, and compared with the training image after projection Relatively, recognition result is finally drawn.

Feature gesture spatial configuration:

For the images of gestures of a secondary M × N, it is D=M × N-dimensional by its head with one size of composition that is connected with tail end to end Column vector.D is exactly the dimension of images of gestures, that is to say the dimension of image space.If n is the number of training sample, x_jRepresent jth The characteristic information column vector that width images of gestures is formed, then the covariance matrix of institute's sample, formula (3.1) can be had to draw.

$S_r=\underset{j=1}{\overset{n}{\Σ}}(x_i-u){(x_j-u)}^T---\left(3.1\right)$

Wherein u is the average image vector of training sample, can be drawn by formula (3.2).

$u=\frac{1}{n}\underset{j=1}{\overset{n}{\Σ}}{x}_j---\left(3.2\right)$

Make A=[x₁-u,x₂-u,...,x_n-u], then there is S_r=AA^T, its dimension is D × D.

According to Karhunen-Loeve transformation principle, needing the eigenmatrix tried to achieve is by AA^TThe characteristic vector corresponding to nonzero eigenvalue Composition.In view of directly calculating, amount of calculation is bigger, so using singular value decomposition [5] (Singular Value Decomposition, SVD) theorem.By solving A^TEigenvalue and the characteristic vector of A obtain AA^TEigenvalue and feature to Amount.

According to SVD theorem, make λ_i(i=1,2 ..., r) it is matrix A^TR the nonzero eigenvalue of A, ν_iFor A^TA corresponds to λ_i Characteristic vector, then AA^TOrthogonal normalizing characteristic vector μ_iObtained by formula (3.3).

${\mathrm{\μ}}_i=\frac{1}{\sqrt{{\mathrm{\λ}}_i}}{A}_{v_i},(i=1,2,...,r)---\left(3.3\right)$

The dimension of the characteristic vector so obtained is higher, in order to reduce dimension, uses and calculates information computing formula (3.4) i characteristic vector before the method choice of dimension is determined.Because image corresponding to these characteristic vectors is like staff, so This image is referred to as " feature hands ".Having had such a deep red n-dimensional subspace n being made up of " feature hands ", any sub-picture all may be used To project and to obtain one group of vector to it, this group vector parameter indicates the similarity degree of this image and " feature hands ", thus this Group vector can be as the foundation identifying staff.

$\frac{({\mathrm{\&lambda;}}_1+{\mathrm{\&lambda;}}_2+...+{\mathrm{\&lambda;}}_i)}{({\mathrm{\&lambda;}}_1+{\mathrm{\&lambda;}}_2+...+{\mathrm{\&lambda;}}_n)}>=0.99,(i<n)---\left(3.4\right)$

The feature extraction of training sample:

Feature extraction is exactly from measurement space RⁿTo feature space R^mMapping.Here measurement space is exactly gesture space, Two principles are followed in mapping, and one is the classification information that feature space must retain in measurement space, and two is feature space Dimension should be well below the dimension of measurement space.

PCA algorithm utilizes Karhunen-Loeve transformation to carry out feature extraction, and this conversion is a kind of data pressure meeting mentioned above principle Compression method, the ultimate principle of its feature extraction is: find out one group of m orthogonal vector in measurement space, it is desirable to this organizes vector The variance representing data that energy is maximum；Then former images of gestures vector is projected to this group orthogonal vector institute structure from n-dimensional space The m n-dimensional subspace n become, then projection coefficient is exactly the characteristic vector constituting former images of gestures, completes compression (the m ＜ of dimension simultaneously ＜ n).

The image sources that the present invention chooses is in the image acquisition of Android platform photographic head.Containing 9 kinds of handss in image set Gesture, every kind of gesture has 10 width images, and these images have following feature, and image background color is single, are conducive to being partitioned into gesture；Hands Gesture detail section has change in various degree, as the most equal in the angle between finger.Using all of 90 width images as training Image, constitutes the training set of 90 width images, and test set is the images of gestures obtained in real time.The concrete steps of feature extraction can Feature Selection is carried out by following described step.

Present invention PCA based on area features gesture identification is specific as follows:

1 training sample is roughly selected and is taken

Needing the gesture identified is 9 kinds, as shown in Figure 5.Each gesture contains solstics, standard images of gestures 12 region Characteristic information, and it is saved as template.Gesture to be identified also contains this feature, can carry out template matching and identify result.

If the characteristic vector of each template is U={ μ₁,μ₂,μ₃,μ₄,μ₅, the characteristic vector of gesture to be identified is N={ η₁, η₂,η₃,η₄,η₅, use Euclidean distance as the discrimination of two characteristic vectors, calculate gesture to be identified according to formula (4.1) Discrimination diff={ β with template gesture₁,β₂,…β₉}。

$diff=\underset{i=1}{\overset{5}{\&Sigma;}}\left|\right|{\mathrm{\&mu;}}_i-{\mathrm{\&eta;}}_i|{|}^2---\left(4.1\right)$

It is calculated 5 minimum discriminations, it is judged that final gesture is the one in these 5 kinds of gestures, for next step PCA Gesture identification provides preliminary and judges.

2PCA gesture identification:

Through extracting the eigenmatrix A of training sample above, according to the 5 kinds of gestures determined, gesture to be identified this 5 Planting among gesture, these 5 kinds of gesture characteristics of correspondence vectors in extraction feature matrix A, these characteristic vectors are by feature new for composition Matrix A '={ X₁,X₂,…X₅₀}^TWherein X_iIt is that the i-th width training sample image seeks character pair vector, equally to image to be identified Carry out feature extraction, obtain characteristic vector B={x₁,x₂,…x_iWherein i by through information computing being taken out front i special Levy characteristic vector.

Thus obtain gathering A ' and B two set, use Euclidean distance formula (4.2) to judge in set B and set A special Levy a degree of vector.The matching degree set diff calculated_final={ θ₁,θ₂,…θ₄₉, θ₅₀Ask for the minima of θ, minimum Images of gestures corresponding to θ value be final recognition result.

${\mathrm{\&theta;}}_j=\underset{i=1}{\overset{i}{\&Sigma;}}\left|\right|{\mathrm{\&mu;}}_i-x_i|{|}^2---\left(4.2\right)$

4. wireless communication module: the miscellaneous equipment that the gesture information of static gesture identification module identification is sent in network Carrying out information sharing or information is mutual, wireless communication module uses communication module of the prior art, does not repeats them here.

Technique scheme is one embodiment of the present invention, for those skilled in the art, at this On the basis of disclosure of the invention application process and principle, it is easy to make various types of improvement or deformation, it is not limited solely to this Inventing the method described by above-mentioned detailed description of the invention, the most previously described mode the most preferably, and does not have restriction The meaning of property.

Claims (8)

1. a computer based static gesture image recognition interactive system, it is characterised in that including:

Images of gestures acquisition module: the image for being captured by photographic head is processed, it is thus achieved that the images of gestures of standard；

Image pre-processing module: the images of gestures of the standard of acquisition is carried out bounding box process, and is projected into standard picture；

Pictorial element acquisition module: obtain the gesture area centre of form, solstics and the principal direction in standard picture；

Obtain solstics characteristic information module: obtain solstics, standard picture 12 region characteristic information；

Obtain area features information module: obtain the area features information of standard picture；

Static gesture identification module: use PCA algorithm to carry out static gesture identification.

Computer based static gesture image recognition interactive system the most according to claim 1, it is characterised in that: described Images of gestures acquisition module is achieved in that

Image for being captured by photographic head uses to enter gesture area based on RGB and YCbCr color space complexion model Row dividing processing, obtains images of gestures；

Described images of gestures is carried out following process:

Twice eight field denoisings, the noise in eliminating image；

Corrosion treatmentCorrosion Science, makes the connected domain being slightly connected in images of gestures be divided into two independent connected domains；

Connected area disposal$, and calculate the area of each connected domain, bigger non-gesture area is done background process, will non-gesture Region is set to black, and gesture area is set to redness；

Image after processing carries out expansion process, and reduction is through the image of corrosion treatmentCorrosion Science.

Computer based static gesture image recognition interactive system the most according to claim 2, it is characterised in that: described Image pre-processing module is achieved in that

Internally scan from each limit of image successively, be the border on this limit of bounding box when scan line encounters images of gestures, The rectangle that four borders surround is the bounding box of this images of gestures；

Region outside bounding box is invalid data region, and the region within bounding box is valid data region；

Bounding box area image is mapped on standard picture.

Computer based static gesture image recognition interactive system the most according to claim 3, it is characterised in that: described Bounding box area image is mapped on standard picture and is achieved in that

Accepted standard image size is 100*100, specifically comprises the following steps that

Step1. size and standard image size according to described bounding box are according to formula (1.1) calculating scaling ratio:

ZoomX=newWidth/width

(1.1)

ZommY=newHeight/height

Wherein zoomX, zoomY are respectively wide and high scaling ratio, and newWidth, newHeight are respectively standardized images The length of side, width, height are width and the height of source images；

Step2. the images of gestures in bounding box is zoomed in standard picture according to formula (1.2):

$\left[\begin{array}{c}{x}^{\&prime;}\\ y^{\&prime;}\end{array}\right]=\left[\begin{array}{cc}zoomX& 0\\ 0& zoomY\end{array}\right]\left[\begin{array}{c}x\\ y\end{array}\right]---\left(1.2\right)$

The coordinate of the pixel in image after wherein (x ', y ') is standardization, (x y) is the coordinate figure of pixel in source images.

Computer based static gesture image recognition interactive system the most according to claim 4, it is characterised in that: described Pictorial element acquisition module is achieved in that

Calculate abscissa and the vertical coordinate sum of each pixel of gesture area in standard picture the most respectively, and calculate gesture Area pixel point number, represents the area of gesture area with this；

Step2. ask for barycentric coodinates according to asking for center of gravity formula (1.3), be the coordinate of the gesture area centre of form:

$\left\{\begin{array}{c}\stackrel{\&OverBar;}{x}=\frac{\underset{A}{\&Integral;}ydA}{A}\\ \stackrel{\&OverBar;}{y}=\frac{\underset{A}{\&Integral;}xdA}{A}\end{array}\right.---\left(1.3\right)$

Wherein, A represents images of gestures area,Represent centre of form abscissa value,Representing centre of form ordinate value, dA is integral element；

Solstics is that images of gestures area pixel point is to centre of form distance point furthest；

Principal direction is the line connecting centroid point to solstics, and direction is to point to solstics from centroid point.

Computer based static gesture image recognition interactive system the most according to claim 5, it is characterised in that: described Obtain solstics characteristic information module to be achieved in that

Centered by the gesture area centre of form, with 30 degree of angles for the anglec of rotation, gesture area is divided into 12 regions, wherein the firstth district Territory is the region of 15 degree of angles composition about principal direction, from the beginning of first area, counterclockwise on be followed successively by second area to the Territory, No.12 District, the characteristic information step obtaining 12 solstics, region is as follows:

Step1. be calculated the solstics in each region, and calculate the solstics distance to the centre of form, compare obtain therein Big distance；

Step2. ultimate range is equally divided into 5 sections, then obtains 5 packets, the distance centre of form nearest for first group, the distance centre of form Farthest for the 5th group；

Calculate 12 solstics the most respectively and fall the number of times of 5 packets；

Step4. by the number of times falling 5 regions obtained divided by 12 to uniform 12 solstics, region characteristics；Grown Degree is 12 solstics, the region characteristic vector U={ μ of 5₁,μ₂,μ₃,μ₄,μ₅}。

Computer based static gesture image recognition interactive system the most according to claim 6, it is characterised in that: described Obtain area features information module to be achieved in that

Step1. calculate the distance in gesture solstics and the centre of form, and this range averaging is divided into 12 sections, then obtain 12 packets, away from From the centre of form nearest for first group, distance the centre of form farthest for the 12nd group；

Step2. with 12 region criteria for classifying images, and the centre of form in each region is calculated；

Do the centre of form of gesture area the most successively by the centre of form in 12 regions, calculate solstics now and principal direction；

Step4. calculate the distance in the centre of form and solstics, and be averaged and be divided into 5 parts, obtain 5 packets, whole circumference is divided into 12 parts, i.e. from the beginning of principal direction, to be divided into a region for direction every 30 degree of angles counterclockwise, so obtain 12 regions, 5 Individual packet and 12 regions intersect and then obtain 60 boxed area, calculate gesture area and fall at the pixel of each boxed area Number, the area of the most each boxed area；

Calculate the maximum area often organized the most respectively, the most often organize the maximum area all divided by this group, i.e. obtain homogenization Area features value.

Computer based static gesture image recognition interactive system the most according to claim 7, it is characterised in that: described Static gesture identification module is achieved in that

Step1, each gesture contains solstics, standard picture 12 region characteristic information, is saved as template, if each mould The characteristic vector of plate is U={ μ₁,μ₂,μ₃,μ₄,μ₅, the characteristic vector of gesture to be identified is N={ η₁,η₂,η₃,η₄,η₅, use Euclidean distance, as the discrimination of two characteristic vectors, calculates the differentiation of gesture to be identified and template gesture according to formula (4.1) Degree diff={ β₁,β₂,…β₉}:

$diff=\underset{i=1}{\overset{5}{\&Sigma;}}\left|\right|{\mathrm{\&mu;}}_i-{\mathrm{\&eta;}}_i|{|}^2---\left(4.1\right)$

Being calculated 5 minimum discriminations, gesture to be identified is the one in these 5 kinds of gestures；

Step2, PCA gesture identification:

Structural feature gesture space: for the images of gestures of a width M × N, one size of the composition that is connected with tail by its head is D=M The column vector of × N-dimensional, D is exactly the dimension of images of gestures, that is to say the dimension of image space, if n is the number of training sample, x_j Represent the characteristic information column vector that jth width images of gestures is formed, the then covariance matrix of institute's sample, formula (3.1) draw:

$S_r=\underset{j=1}{\overset{n}{\&Sigma;}}(x_i-u){(x_j-u)}^T---\left(3.1\right)$

Wherein u is the average image vector of training sample, formula (3.2) draw:

$u=\frac{1}{n}\underset{j=1}{\overset{n}{\&Sigma;}}{x}_j---\left(3.2\right)$

Make A=[x₁-u,x₂-u,...,x_n-u], then there is S_r=AA^T, its dimension is D × D；

According to the 5 kinds of gestures determined, gesture to be identified among these 5 kinds of gestures, these 5 kinds of gestures pair in extraction feature matrix A The characteristic vector answered, these characteristic vectors are by eigenmatrix A '={ X new for composition₁,X₂,…X₅₀}^T, wherein X_iIt it is the i-th width training Sample image seeks character pair vector, equally image to be identified is carried out feature extraction, obtains characteristic vector B={x₁,x₂,… x_i, wherein i is by through information computing being taken out front i character vector, so obtaining gathering A ' and B two collection Close, a degree of characteristic vector in employing Euclidean distance formula (4.2) judgement set B and set A ':

${\mathrm{\&theta;}}_j=\underset{i=1}{\overset{i}{\&Sigma;}}\left|\right|{\mathrm{\&mu;}}_i-x_i|{|}^2---\left(4.2\right)$

And calculate matching degree set diff_final={ θ₁,θ₂,…θ₄₉, θ₅₀, then ask for the minima of θ, minimum θ value Corresponding images of gestures is final recognition result.