CN106203446A - Three dimensional object recognition positioning method for augmented reality auxiliary maintaining system - Google Patents

Abstract

The present invention relates to a kind of three dimensional object recognition positioning method for augmented reality auxiliary maintaining system, the method includes: set up segmentation and the extraction of feature, the identification of object type and the location of object of the information database of three dimensional object, three dimensional object.The present invention is the three dimensional object character description method that a kind of overall situation edge histogram is combined with the local edge histogram of angle point guiding, describe method by this and two dimensional image of each three dimensional object different visual angles is carried out feature extraction, and using Gentle AdaBoost multi classifier that the characteristic vector extracted is trained, verification and measurement ratio that can be higher realizes the classification identification to three dimensional object.Machine learning is also combined by the present invention with template matching method, carries out coupling location by searching for little scope local data bank, improves location efficiency.

Description

Three dimensional object recognition positioning method for augmented reality auxiliary maintaining system

Technical field

The present invention relates to a kind of machine vision technique, a kind of for the three of augmented reality auxiliary maintaining system Dimensional object recognition positioning method.

Background technology

Carry out under augmented reality auxiliary maintaining environment detection, repair, dismount, assemble, safeguard, the master of the maintenance such as maintenance Body is to operate with personnel, and the maintenance objects such as complete machine, parts, parts is then the object of maintenance.If will increase Strong reality auxiliary maintaining system is applied to actual maintenance, and first have to solution keeps in repair the identification of maintenance objects in scene exactly Orientation problem, it is video camera Attitude estimation, Tracing Registration and the premise of deficiency and excess combination and basis.Therefore tie up for augmented reality Repair environment to carry out three dimensional object identification Position Research and have great importance.

What physical feature recognition methods actually solved is the identification orientation problem of the lower maintenance objects of various visual angles.Utilize some spy Levy, outward appearance or contour feature and affine invariants are modal various visual angles identification character description methods.As by various visual angles 2D Image SIFT feature point is mapped to the surface of 3D model and carries out 3D Object identifying, utilizes unchanged view angle profile relation and appearance information It is trained by machine learning algorithm, obtains the identification model of 3D object, extract affine invariant subspace feature and realize affine Change the match cognization of hypograph target, and contour feature combines with affine invariants point and is identified location etc., but Affine invariants point often requires that visual angle change scope can not be the biggest.

Summary of the invention

It is an object of the invention to provide a kind of three dimensional object identification location side for augmented reality auxiliary maintaining system Method, the problem the highest to the verification and measurement ratio of the classification identification of three dimensional object to solve existing location mode.

The object of the present invention is achieved like this: a kind of three dimensional object identification for augmented reality auxiliary maintaining system is fixed Method for position, comprises the following steps:

A, set up the information database of three dimensional object:

A-1, using to have the equipment component of information association as three dimensional object and right with augmented reality auxiliary maintaining process O should be labeled as₁,O₂,…,O_k；Set up the w that two kinds of visual angle change density of each three dimensional object is different_t×h_tTwo dimensional image template, its Middle a kind of template being spaced apart 10 ° for visual angle change, is denoted as TA, the another kind of template being spaced apart 2.5 ° for visual angle change, is denoted as TB；All templates TA and template TB of each three dimensional object are associated with in the visual angle corresponding to its two dimensional image template, and by regarding Angle classification is saved in template database；

Edge image TEA, TEB of a-2, template TA extracting each three dimensional object respectively and template TB, extracts each three-dimensional right Template TA of elephant and histogrammic characteristic vector FA in compound edge of template TB, FB, be associated with each self-corresponding template TA and mould In plate TB, and it is saved in information database；

A-3, O will be labeled as₁—O_kThe each self-corresponding compound edge histogram feature vector FA of three dimensional object and right The class label answered is sent into grader and is trained, it is thus achieved that all classification identification model M of three dimensional object_a, and store；Point After not removing histogrammic characteristic vector FA in compound edge of each three dimensional object, it is fed again into grader and is trained, To the identification model M not including deleted three dimensional object_i≠1,M_i≠2,…,M_i≠k, and store；

A-4, the compound edge histogrammic characteristic vector FB category of all three dimensional objects is set up retrieval K-D tree KB₁, KB₂,…,KB_k。

Grader described in step a-3 is to use Gentle AdaBoost multi classifier based on decision-making stub, wherein The number maximum occurrences of Weak Classifier is 100.

B, the segmentation of three dimensional object and the extraction of feature:

Image, from a two field picture I of imageing sensor, is filtered and extracts its binary edge map by b-1, acquisition E；Calculate the integral image II about binary edge map E_E；Calculate the integration about binary edge map E and be combined edge histogram Figure IH_E；

B-2, establishment width are w, height is the sliding window W1 of h, create width be (w+2), width be the slip of (h+2) Window W2；Sliding window W1 is nested in sliding window W2 between two parties, and carries out binary edge map E with sliding window W2 simultaneously Scanning, utilizes integral image II_ECalculate the pixel sum of sliding window W1 and sliding window W2；When sliding window W1 and sliding window The pixel sum of mouthful W2 more than 0 and time the difference of pixel sum is 0, target area image E not being blocked_W1, and forward step to b-3；If the scanning result of entire image all can not meet above-mentioned condition, then w value adds 2, and h value adds 2, and repeats this step Suddenly；When the height value of the width value of w >=image I and h >=image I, forward step b-1 to；

B-3, utilize IH_EObtain the compound edge histogram H of region sliding window W1_W, then perform step c.

C, the identification of object type:

Read and identify model M_a, by compound edge histogram H_WSend into read as input vector and identify model M_a, known Other result is combined edge histogram H_WRepresentative class label n.

D, the location of object:

D-1, with compound edge histogram H_WWith retrieval K-D tree KB_nPasteur's distance D1 value of middle characteristic vector is minimum to be depended on According to, at retrieval K-D tree KB_nMiddle retrieval obtains and is combined edge histogram H_WImmediate characteristic vector position p；

D-2, when Pasteur's distance D1 >=threshold tau₁Time, directly proceed to step d-3, otherwise by the binary edge of sliding window W1 Image E_W1Being normalized to width is w_t, height be h_tYardstick, and take template image TB according to characteristic vector position k_n,p, will be sliding The binary edge map E of dynamic window W1_W1With template image TB_n,pDo Chamfer distance coupling, obtain matching result D2；When D2 < Threshold tau₂Time, output template image TB_n,pCorresponding visual angle and the binary edge map E of sliding window W1_W1Normalization coefficient ε, Complete whole position fixing process and exit, otherwise forwarding step d-3 to；

D-3, reading identify model M_i≠n, by H_WM is sent into as input vector_i≠n, it is identified the classification that result is new Label n, and take retrieval K-D tree KB according to class label n_n, go to step d-1；When identifying model M_i≠1,M_i≠2,…,M_i≠kQuilt time After going through once, declaration positions unsuccessfully, goes to step d-2.

The histogrammic extracting method in compound edge in the present invention is:

For any piece image I, its any pixel I (x, horizontal gradient and the computing formula of vertical gradient y) be:

$\left\{\begin{array}{c}{G}_x(x,y)={\mathrm{mask}}_x\&CircleTimes;I(x,y)\\ G_y(x,y)={\mathrm{mask}}_y\&CircleTimes;I(x,y)\end{array}\right.---\left(1\right)$

Wherein, mask_xFor horizontal filter template, mask_yFor vertical filter template,For convolution operation；

(x, y) edge direction at place is defined as:

$\mathrm{\&theta;}(x,y)=\frac{180}{\mathrm{\&pi;}}\&CenterDot;arctan\frac{G_y(x,y)}{G_x(x,y)}---\left(2\right)$

It is interval interior that the edge direction obtained is quantized to K in the range of-180 °～180 °；Order:

b_k=[-180+ (k-1) 360/K ,-180+k 360/K] (3)

Wherein, k=1,2 ..., K, then edge orientation histogram is defined as formula (4) and formula (5)；

${\mathrm{\&psi;}}_k(x,y)=\left\{\begin{array}{cc}1,& \begin{array}{cc}if& \mathrm{\&theta;}(x,y)\&Element;b_k\end{array}\\ 0,& otherwise\end{array}\right.---\left(4\right)$

${\mathrm{GH}}_k=\underset{(x,y)\&Element;E}{\&Sigma;}{\mathrm{\&psi;}}_k(x,y)---\left(5\right)$

Wherein, binary edge map E is the edge image of image I；

With k as vertical coordinate, GH_kEdge histogram is set up for abscissa；

Extracting the angle point of image, centered by angle point, the edge histogram of statistics angle point neighborhood is as local edge Nogata Scheming, it is defined as:

${\mathrm{LH}}_k=\underset{(x,y)\&Element;E\&cap;C}{\&Sigma;}{\mathrm{\&psi;}}_k(x,y)---\left(6\right)$

Wherein, the region that C is formed by angle point and eight neighborhoods thereof；

Overall situation edge histogram feature is combined as the following formula with the local edge histogram feature of angle point guiding:

H_k=GH_k∪LH_k (7)

I.e. available as compound edge histogram feature.

The present invention is the three dimensional object that a kind of overall situation edge histogram is combined with the local edge histogram of angle point guiding Character description method, describes method by this and two dimensional image of each three dimensional object different visual angles is carried out feature extraction, and adopt Being trained, with Gentle AdaBoost multi classifier, the characteristic vector extracted, verification and measurement ratio that can be higher realizes three The classification identification of dimensional object.Machine learning is also combined by the present invention with template matching method, by searching for little scope local number Carry out coupling location according to storehouse, improve location efficiency.

Accompanying drawing explanation

Fig. 1 is the FB(flow block) of recognition positioning method of the present invention.

Fig. 2 is AdaBoost algorithm structure figure.

Fig. 3 is the 1st 6 class identification location object that experiment is targeted.

Fig. 4 is detection the 2nd class and the positioning result of the 5th class destination object.

Fig. 5 is detection the 5th class and the positioning result of the 6th class destination object.

Detailed description of the invention

Choose six classes image of component to be identified as identify object, the specific implementation process of the present invention is described.Such as Fig. 1 institute Showing, the present invention comprises the steps: towards the three dimensional object recognition positioning method strengthening system supplymentary maintenance system

1, the template database of object to be identified is built.

1-1, set up the template database of three dimensional object.

First, being fixed on by video camera on the position becoming 45 ° of pitching visual angle with maintenance objects, maintenance objects is placed on high-precision On degree turntable, driving turntable with 10 ° for interval stepping, the formation different viewpoint of video camera also triggers video camera and shoots.Right The template image recycling image processing software that shooting obtains carries out rotating in the face at interval with 10 °, template image and process After image collectively as the rough segmentation visual angle template of three dimensional object, be denoted as TA.

Secondly, driving turntable with 2.5 ° for interval stepping, viewpoint that formation video camera is different also triggers video camera and claps Take the photograph.The template image recycling image processing software obtaining shooting carries out rotating in the face at interval with 2.5 °, template image With the image after process collectively as the segmentation visual angle template of three dimensional object, it is denoted as TB.By template TA of all three dimensional objects and The corresponding visual angle of template TB is associated, and category is saved in data base.Thus obtain this maintenance objects multiple Template image under viewpoint.

1-2, set up the information database of three dimensional object.

Extract template TA and edge image TEA, TEB of template TB and compound edge histogram feature vector FA, FB respectively, And be associated with template TA and template TB, then preserve to information database.

The histogrammic extracting method in compound edge is:

For any piece image I, its any pixel I (x, horizontal and vertical gradient calculation formula y) is:

$\left\{\begin{array}{c}{G}_x(x,y)={\mathrm{mask}}_x\&CircleTimes;I(x,y)\\ G_y(x,y)={\mathrm{mask}}_y\&CircleTimes;I(x,y)\end{array}\right.---\left(1\right)$

Wherein, mask_xFor horizontal filter template, mask_yFor vertical filter template,For convolution operation.

(x, y) edge direction at place is defined as:

$\mathrm{\&theta;}(x,y)=\frac{180}{\mathrm{\&pi;}}\&CenterDot;arctan\frac{G_y(x,y)}{G_x(x,y)}---\left(2\right)$

It is interval interior that the edge direction obtained is quantized to K in the range of-180 °～180 °.Order:

b_k=[-180+ (k-1) 360/K ,-180+k 360/K] (3)

Wherein, k=1,2 ..., K, then edge orientation histogram is defined as formula (4) and (5).

${\mathrm{\&psi;}}_k(x,y)=\left\{\begin{array}{cc}1,& \begin{array}{cc}if& \mathrm{\&theta;}(x,y)\&Element;b_k\end{array}\\ 0,& otherwise\end{array}\right.---\left(4\right)$

${\mathrm{GH}}_k=\underset{(x,y)\&Element;E}{\&Sigma;}{\mathrm{\&psi;}}_k(x,y)---\left(5\right)$

Wherein: binary edge map E is the edge image of image I.

With k as vertical coordinate, GH_kEdge histogram is set up for abscissa.

Extracting the angle point of image, centered by angle point, the edge histogram of statistics angle point neighborhood is as local edge Nogata Scheming, it is defined as:

${\mathrm{LH}}_k=\underset{(x,y)\&Element;E\&cap;C}{\&Sigma;}{\mathrm{\&psi;}}_k(x,y)---\left(6\right)$

Wherein C is angle point and the region of eight neighborhood formation thereof.

Overall situation edge histogram feature is combined as compound edge with the local edge histogram feature of angle point guiding Histogram feature, such as formula (7):

H_k=GH_k∪LH_k (7)

1-3, set up thick visual angle divide three dimensional object identification model.

By three dimensional object O₁,O₂,…,O₆Compound edge histogram feature vector FA and correspondence class label send into Grader is trained, it is thus achieved that about the classification identification model M of whole three dimensional objects_a.Remove each three dimensional object respectively After histogrammic characteristic vector FA in compound edge, it is fed again into grader and is trained, do not included the most a certain classification Identify model M_i≠1,M_i≠2,…,M_i≠6。

Wherein, grader uses Gentle AdaBoost multi classifier based on decision-making stub, Gentle AdaBoost multi classifier is the improvement of AdaBoost grader, and AdaBoost grader obtains multiple weak typing by iteration Device is also combined into strong classifier, and its algorithm structure is as shown in Figure 2.

Strong classifier Y_MFormed by 100 Weak Classifier weighted arrays, update its weights omega by iteration_nMake weak typing Device error in classification is minimum.The weighter factor of each Weak Classifier represents the significance level of Weak Classifier:

${\mathrm{\&alpha;}}_m=\frac{1}{2}\ln \frac{1-e_m}{e_m}---\left(8\right)$

Wherein, e_mFor error in classification rate.The output codomain of Weak Classifier is set by Gentle AdaBoost multi classifier For [-1,1], in the way of one-to-many, carry out multicategory classification training.If the various visual angles of maintenance objects O are combined edge histogram feature For H_k(O), the jth Weak Classifier of its correspondence is h_j(O).Make y_iFor the label of i-th training sample, 1 represents positive sample ,-1 Represent negative sample, ω_iFor i-th training sample weight.Use decision-making stub structure Weak Classifier h_j(O) it is:

$h_j\left(O\right)=\left\{\begin{array}{cc}{a}_j,& H_k\left(O\right)\&GreaterEqual;{\mathrm{\&theta;}}_j\\ b_j,& otherwise\end{array}\right.---\left(9\right)$

Wherein, affine parameter:Offset parameter: And threshold θ_jNeed to make criterion function:

$J\left(h_j\right(O\left)\right)=\underset{i}{\&Sigma;}{(y_i-h_j(O\left)\right)}^2---\left(10\right)$

Value is minimum.The strong classifier that 100 Weak Classifiers are combined into is:

$H=\underset{j=1}{\overset{100}{\&Sigma;}}{h}_j\left(O\right)---\left(11\right)$

1-4, set up thin visual angle divide retrieving three-dimensional objects K-D tree.

The compound edge histogram feature vector FB category of all three dimensional objects is set up retrieval K-D tree KB₁,KB₂,…, KB₆。

2, the segmentation of three dimensional object and the rapid extraction of compound edge histogram feature.

2-1, acquisition image I as shown in Fig. 4 or Fig. 5, be filtered it and extract its binary edge map E.Calculate Integral image II about binary edge map E_E；Calculate the integration about compound edge histogram feature vector E and be combined edge Rectangular histogram IH_E.Wherein, to be combined the histogrammic computational methods in edge as follows for integral image and integration:

For a given two field picture, by initial point and coordinate (x, y) integration histogram in determined region is defined as:

$H(x,y,u)=\underset{x^{\&prime;}=0}{\overset{x}{\&Sigma;}}\underset{y^{\&prime;}=0}{\overset{y}{\&Sigma;}}Q(x^{\&prime;},y^{\&prime;},u)---\left(12\right)$

In formula, u is the index of rectangular histogram bin image, and d is the number of post image:

$Q(x^{\&prime;},y^{\&prime;},u)=\left\{\begin{array}{cc}1,& \begin{array}{cc}if& I(x^{\&prime;},y^{\&prime;})\&Element;u-thbin\end{array}\\ 0,& otherwise\end{array}\right.---\left(13\right)$

Further, the value of the integration histogram of each pixel can obtain according to following formula recursive calculation:

H (x, y, u)=H (x-1, y, u)+H (x, y-1, u)-H (x-1, y-1, u)+Q (x, y, u)., u=1 ..., d (14)

After integration histogram is set up, for any rectangular area R (x^-,y^-,x⁺,y⁺) rectangular histogram can be by:

H (R, u)=H (x⁺,y⁺,u)-H(x^-,y⁺,u)-H(x⁺,y^-,u)+H(x^-,y^-, u) u=1 ..., d (15)

Rapid extraction obtains.B is the number of post image, w and h is respectively width and the height of image, thus constitutes a w The cube of × h × b, for the histogrammic extraction of any subregion in w × h, only need to be by each post corresponding for subregion Image is according to simply adding reducing, and this accumulative integration histogram also can quickly obtain subregion and be combined edge Rectangular histogram.

2-2, create width and take advantage of the sliding window W1 of a height of w × h and width to take advantage of the sliding window W2 of a height of (w+2) × (h+2)；Sliding Dynamic window W1 is nested in W2 between two parties, and is scanned image with sliding window W2 simultaneously, calculates inside and outside window respectively and surrounds Pixel value in region and with difference.When the pixel sum both greater than 0 of two windows and the difference of pixel sum are 0, obtain not hidden Target area image E of gear_W1, go to step 2-3；If the scanning result of entire image all can not meet above-mentioned condition, the widest Degree w adds 2, and highly h adds 2, repeats this step operation；When the width of w >=image I, during h >=image I high, forward step 2-1 to.

2-3, utilize IH_EObtain the compound edge histogram H of region W1_W, go to step 3.

3, the identification of object type.

Read the identification model M about all 6 class three dimensional objects_a, by compound edge histogram H_WSend into as input vector Read and identify model M_a, it is identified result and is combined edge histogram H_WRepresentative class label n.

4, the location of object.

4-1, with compound edge histogram H_WWith retrieval K-D tree KB_nPasteur's distance D1 value of middle characteristic vector is minimum to be depended on According to, at retrieval K-D tree KB_nMiddle retrieval obtains and is combined edge histogram H_WImmediate characteristic vector position p.

4-2, when Pasteur's distance D1 >=threshold tau₁Time directly proceed to step 4-3, otherwise, by target area image E_W1Normalization To w_t×h_tYardstick, and take template image TB according to characteristic vector position k_n,p, by target area image E_W1With template image TB_n,p Do Chamfer distance coupling, obtain matching result D2；When D2 < threshold tau₂Time, output template image TB_n,pCorresponding visual angle and mesh Mark area image E_W1Normalization coefficient ε, complete position fixing process and exit operation, otherwise forwarding step 4-3 to.

4-3, reading identify model M i ≠ n, by compound edge histogram H_WAs input vector, send into Mi ≠ n, known The class label that other result is new is n；Go to step 4-1.When Mi ≠ 1, Mi ≠ 2 ..., after Mi ≠ 6 are traversed once, declaration Position unsuccessfully, go to step 2-2.

According to the actually detected testing result obtained shown in Fig. 4 and Fig. 5 of aforesaid way.

Claims (3)

1., for a three dimensional object recognition positioning method for augmented reality auxiliary maintaining system, it is characterized in that, including following step Rapid:

A, set up the information database of three dimensional object:

A-1, will there is the equipment component of information association as three dimensional object with augmented reality auxiliary maintaining process, and corresponding mark It is designated as O₁,O₂,…,O_k；Set up the w that two kinds of visual angle change density of each three dimensional object is different_t×h_tTwo dimensional image template, Qi Zhongyi Plant the template being spaced apart 10 ° for visual angle change, be denoted as TA, the another kind of template being spaced apart 2.5 ° for visual angle change, be denoted as TB；Will All templates TA of each three dimensional object and template TB are associated with in its visual angle corresponding to two dimensional image template, and by visual angle classification It is saved in template database；

Edge image TEA, TEB of a-2, template TA extracting each three dimensional object respectively and template TB, extracts each three dimensional object Histogrammic characteristic vector FA in compound edge of template TA and template TB, FB, be associated with each self-corresponding template TA and template TB In, and be saved in information database；

A-3, O will be labeled as₁—O_kThe each self-corresponding compound edge histogram feature vector FA of three dimensional object and correspondence Class label is sent into grader and is trained, it is thus achieved that all classification identification model M of three dimensional object_a, and store；Go respectively After histogrammic characteristic vector FA in compound edge of each three dimensional object, it is fed again into grader and is trained, obtain not Identification model M including deleted three dimensional object_i≠1,M_i≠2,…,M_i≠k, and store；

A-4, the compound edge histogrammic characteristic vector FB category of all three dimensional objects is set up retrieval K-D tree KB₁, KB₂,…,KB_k；

B, the segmentation of three dimensional object and the extraction of feature:

Image, from a two field picture I of imageing sensor, is filtered and extracts its binary edge map E by b-1, acquisition；Meter Calculate the integral image II about binary edge map E_E；Calculate the integration about binary edge map E and be combined edge histogram IH_E；

B-2, establishment width are w, height is the sliding window W1 of h, create width be (w+2), width be the sliding window of (h+2) W2；Sliding window W1 is nested in sliding window W2 between two parties, and sweeps binary edge map E with sliding window W2 simultaneously Retouch, utilize integral image II_ECalculate the pixel sum of sliding window W1 and sliding window W2；When sliding window W1 and sliding window The pixel sum of W2 more than 0 and time the difference of pixel sum is 0, target area image E not being blocked_W1, and forward step b-to 3；If the scanning result of entire image all can not meet above-mentioned condition, then w value adds 2, and h value adds 2, and repeats this step； When the height value of the width value of w >=image I and h >=image I, forward step b-1 to；

B-3, utilize IH_EObtain the compound edge histogram H of region sliding window W1_W, then perform step c；

C, the identification of object type:

Read and identify model M_a, by compound edge histogram H_WSend into read as input vector and identify model M_a, it is identified knot The compound edge histogram H of fruit_WRepresentative class label n；

D, the location of object:

D-1, with compound edge histogram H_WWith retrieval K-D tree KB_nPasteur's minimum foundation of distance D1 value of middle characteristic vector, At retrieval K-D tree KB_nMiddle retrieval obtains and is combined edge histogram H_WImmediate characteristic vector position p；

D-2, when Pasteur's distance D1 >=threshold tau₁Time, directly proceed to step d-3, otherwise by the binary edge map of sliding window W1 E_W1Being normalized to width is w_t, height be h_tYardstick, and take template image TB according to characteristic vector position k_n,p, by sliding window The binary edge map E of mouth W1_W1With template image TB_n,pDo Chamfer distance coupling, obtain matching result D2；When D2 < threshold value τ₂Time, output template image TB_n,pCorresponding visual angle and the binary edge map E of sliding window W1_W1Normalization coefficient ε, complete Whole position fixing processs also exit, and otherwise forward step d-3 to；

D-3, reading identify model M_i≠n, by H_WM is sent into as input vector_i≠n, it is identified the class label that result is new N, and take retrieval K-D tree KB according to class label n_n, go to step d-1；When identifying model M_i≠1,M_i≠2,…,M_i≠kIt is traversed one After secondary, declaration positions unsuccessfully, goes to step d-2.

Three dimensional object recognition positioning method for augmented reality auxiliary maintaining system the most according to claim 1, it is special Levying and be, the described histogrammic extracting method in compound edge is:

For any piece image I, its any pixel I (x, horizontal gradient and the computing formula of vertical gradient y) be:

$\left\{\begin{array}{c}{G}_x(x,y)={\mathrm{mask}}_x\&CircleTimes;I(x,y)\\ G_y(x,y)={\mathrm{mask}}_y\&CircleTimes;I(x,y)\end{array}\right.---\left(1\right)$

Wherein, mask_xFor horizontal filter template, mask_yFor vertical filter template,For convolution operation；

(x, y) edge direction at place is defined as:

$\mathrm{\&theta;}(x,y)=\frac{180}{\mathrm{\&pi;}}\&CenterDot;\arctan \frac{G_y(x,y)}{G_x(x,y)}---\left(2\right)$

It is interval interior that the edge direction obtained is quantized to K in the range of-180 °～180 °；Order:

b_k=[-180+ (k-1) 360/K ,-180+k 360/K] (3)

Wherein, k=1,2 ..., K, then edge orientation histogram is defined as formula (4) and formula (5)；

${\mathrm{\&psi;}}_k(x,y)=\left\{\begin{array}{cc}1,& \begin{array}{cc}if& \mathrm{\&theta;}(x,y)\&Element;b_k\end{array}\\ 0,& otherwise\end{array}\right.---\left(4\right)$

${\mathrm{GH}}_k=\underset{(x,y)\&Element;E}{\&Sigma;}{\mathrm{\&psi;}}_k(x,y)---\left(5\right)$

Wherein, binary edge map E is the edge image of image I；

With k as vertical coordinate, GH_kEdge histogram is set up for abscissa；

Extract image angle point, centered by angle point, statistics angle point neighborhood edge histogram as local edge histogram, its It is defined as:

${\mathrm{LH}}_k=\underset{(x,y)\&Element;E\&cap;C}{\&Sigma;}{\mathrm{\&psi;}}_k(x,y)---\left(6\right)$

Wherein, the region that C is formed by angle point and eight neighborhoods thereof；

Overall situation edge histogram feature is combined as the following formula with the local edge histogram feature of angle point guiding:

H_k=GH_k∪LH_k (7)

I.e. available as compound edge histogram feature.

Three dimensional object recognition positioning method for augmented reality auxiliary maintaining system the most according to claim 1, it is special Levying and be, grader described in step a-3 is to use Gentle AdaBoost multi classifier based on decision-making stub, the most weak The number maximum occurrences of grader is 100.