CN106886754A - Object identification method and system under a kind of three-dimensional scenic based on tri patch - Google Patents

Abstract

The invention discloses the object identification method under a kind of three-dimensional scenic based on tri patch.The method is on the basis of traditional two dimensional image recognition methods, in view of the triangle gridding storage mode of 3 D scene rebuilding information, 3-D view is carried out into tri patch division, counted by color neighborhood histogram feature, spatial pyramid multi-scale information and neighborhood space information are merged, it is trained with SVM, obtains disaggregated model and the meadow dough sheet in three-dimensional scenic is identified.Present invention also offers the object identification system under a kind of three-dimensional scenic based on tri patch, technical solution of the present invention solves the problems, such as in three-dimensional reconstruction scene that local message can not completely express the problem of characteristics of image in the doubt problem of identification region and image in identification, understands etc. significant for resurfacing and scene.

Description

Object identification method and system under a kind of three-dimensional scenic based on tri patch

Technical field

The invention belongs to technical field of computer vision, more particularly, to a kind of three-dimensional scenic based on tri patch Under object identification method and system.

Background technology

With the development of human society, image recognition has been developed rapidly as a particularly important technological meanses, its Goal in research is to confer to the visual capacity of the computer mankind, is passed to image cognition surrounding enviroment information, including environment-identification The geometry of middle objective, position and attitude etc..Image recognition needs integrated use computer science, pattern-recognition, machine The subject such as device vision and image understanding, and advance with the development of these subjects.

In general, image identification system is made up of following components.Part I is the acquisition of image information, to figure As being exactly that picture, egative film and character and graphic etc. are converted to electric signal in case subsequent treatment with photoelectric scanning device for identification. Part II is the pretreatment of image, and main purpose is to eliminate extraneous features and strengthen feature of interest.3rd step is that feature is carried Take, conventional feature includes color, texture, Scale invariant features transform (Scale-invariant feature Transform, SIFT), histogram of gradients feature (Histogram of Oriented Gradient, HOG) etc..Part IV It is to adjudicate or classify.Wherein training method can using support vector machine method (Support Vector Machine, SVM), Improved repetitive exercise method (Adaptive Boostint, AdaBoost) etc..

At present, three kinds of Three-dimension object recognition Research Thinkings of most main flow are：1) method based on model or geometry；2) it is based on The method of outward appearance and view；3) method based on local feature matching.

The principle of the method based on model or geometry is exactly to obtain the three-dimensional information of real goal and to letter using sensor Breath is analyzed treatment, obtains the description of a kind of surface, border and annexation.This method advantage is comparing intuitively and is easy to Understand, but the operand of the algorithm for generally using is all larger, and need artificial by CAD (Computer Aided Design, CAD) etc. software produce model.

Method based on view recognizes object by visual similarity, and identifying system design is relatively easy, without explicit Ground calculates object dimensional model.The method is generally divided into two steps：First, by different illumination conditions and three-dimensional body two The different attitudes showed in dimension image, automatically to learn expression or the training system of object；Then, at one unknown two Object is judged whether in dimension image.The restrictive condition of the method is that object interested must can be preferable with background Make a distinction, thus it is more sensitive the overlap between object, and need preferable image segmentation.But when object Geometric Modeling very It is difficult or when not being likely to be obtained, then can recognize three-dimensional body using the method based on view.

Method based on local feature matching mainly includes the local feature in point, line and face, and its advantage is, because visual angle The deformation of the object outward appearance that change causes, although global apparently extremely complex, can be by simple change on local yardstick Change to estimate；Simultaneously as being matched without all of local feature, this method has overlap and complex background feelings in object There is preferable robustness under condition.Because establishing interregional similitude, localization has also been accomplished in object identification；Two dimensional image In identification image pretreatment stage often through super-pixel segmentation into different regions, feature extraction then is carried out to region And classified, and three-dimensional scenic is due to its structural complexity, it is impossible to it is identified by dividing super-pixel region.

The content of the invention

For the disadvantages described above or Improvement requirement of prior art, the invention provides a kind of three dimensional field based on tri patch Object identification method and system under scape, its object is to three-dimensional scenic to be identified is divided into different size of tri patch, It is identified using the clarification of objective in tri patch, is thus extended two dimensional image recognition methods, proposes a kind of base In the meadow recognition methods of tri patch.

To achieve the above object, according to one aspect of the present invention, there is provided a kind of three-dimensional scenic based on tri patch Under object identification method, the method comprises the following steps：

(1) three-dimensional scenic to be identified is carried out into three-dimensional reconstruction, by the triangulation tri patch on three-dimensional scenic surface；

(2) the HSI features of all tri patch units are extracted；

(3) the H characteristic values and S characteristic values according to selected tri patch and the adjacent triangular faces on its shared side carry out Nogata Figure statistics, statistics is put into characteristic vector, by the feature of the characteristic vector of tri patch and its adjacent triangular faces to Amount is overlapped (in specification describe in detail), stack result as selected tri patch characteristic vector；

(4) characteristic vector of selected tri patch is input in SVM identification models, obtains recognition result.

Further, enter column hisgram in the step (3) to be specifically configured to：

It is 8 to set histogrammic group of number of H features value histogram and S characteristic values, and group is away from respectively H characteristic values span The 1/8 of 1/8 and S characteristic value spans.

Further, statistics is put into characteristic vector in the step (3) and is specially：

The H characteristic values of tri patch are carried out into statistics with histogram and obtains one group of 8 dimensional vector, by the S characteristic values of tri patch Carry out statistics with histogram and also obtain one group of 8 dimensional vector, 8 dimensional vectors of 8 dimensional vectors of H characteristic values and S characteristic values are joined end to end The characteristic vector of the dimension of composition 16.

Further, characteristic vector is overlapped specially in the step (3)：

Wherein, { H₁₁,H₁₂,H₁₃,...,H₁₈,S₁₁,S₁₂,...,S₁₈And { H₂₁,H₂₂,H₂₃,...,H₂₈,S₂₁,S₂₂,..., S₂₈Represent 16 characteristic vectors tieed up.

Further, the training process of SVM identification models is specially in the step (4)：

(41) setting positive sample and setting negative sample are collected as sample set, positive sample is chosen and is all object to be identified Two-dimension picture, negative sample choose without object to be identified two-dimension picture；

(42) it is consistent by the size adjusting of all samples, selects multiple default yardsticks to carry out tri patch to sample Divide, extract the HSI features of all tri patch；

(43) the H characteristic values and S characteristic values according to selected tri patch and the adjacent triangular faces on its shared side are carried out directly Side's figure statistics, statistics is put into characteristic vector, by the characteristic vector of tri patch and the feature of its adjacent triangular faces Vector is overlapped, stack result as selected tri patch characteristic vector；

(44) characteristic vector of all samples is input into SVM to be trained；The decision function of SVM is：

Wherein, L presentation classes sum, value 2；c_iRepresentative function coefficient, c_i∈R；γ represents the parameter of cross check, γ ∈R₊；Expression is chosen for the support vector of training；h⁰Represent the non-supported vector of training；Using histogrammic card side's distanceAs kernel function, it is defined as：

Wherein, N represents characteristic vector total quantity in sample set；Represent k-th characteristic vector it is corresponding support to Amount；h⁰K () represents k-th characteristic vector；

(45) SVM identification model parameters are obtained after training, SVM identification models are built.

It is another aspect of this invention to provide that there is provided the object identification system under a kind of three-dimensional scenic based on tri patch System, changing system is included with lower module：

Tri patch division module, for three-dimensional scenic to be identified to be carried out into three-dimensional reconstruction, by the three of three-dimensional scenic surface Angle mesh generation tri patch；

HSI characteristic extracting modules, the HSI features for extracting all tri patch units；

Characteristic vector computing module, for the H features according to selected tri patch and the adjacent triangular faces on its shared side Value and S characteristic values carry out statistics with histogram, and statistics is put into characteristic vector, by the characteristic vector of tri patch and its phase The characteristic vector of adjacent tri patch is overlapped, stack result as selected tri patch characteristic vector；

Identification module, for the characteristic vector of selected tri patch to be input in SVM identification models, obtains identification knot Really.

Further, enter column hisgram in the characteristic vector computing module to be specifically configured to：

It is 8 to set histogrammic group of number of H features value histogram and S characteristic values, and group is away from respectively H characteristic values span The 1/8 of 1/8 and S characteristic value spans.

Further, statistics is put into characteristic vector in the characteristic vector computing module and is specially：

The H characteristic values of tri patch are carried out into statistics with histogram and obtains one group of 8 dimensional vector, by the S characteristic values of tri patch Carry out statistics with histogram and also obtain one group of 8 dimensional vector, 8 dimensional vectors of 8 dimensional vectors of H characteristic values and S characteristic values are joined end to end The characteristic vector of the dimension of composition 16.

Further, characteristic vector is overlapped specially in the step (3)：

Wherein, { H₁₁,H₁₂,H₁₃,...,H₁₈,S₁₁,S₁₂,...,S₁₈And { H₂₁,H₂₂,H₂₃,...,H₂₈,S₂₁,S₂₂,..., S₂₈Represent 16 characteristic vectors tieed up.

Further, the training system of SVM identification models is included with lower unit in the identification module：

Sample collection unit, for collecting multiple positive samples and multiple negative samples as sample set, positive sample is chosen and is all The two-dimension picture of object to be identified, negative sample chooses the two-dimension picture without object to be identified；

Tri patch division unit, for being consistent by the size adjusting of all samples, selects multiple default yardsticks pair Sample carries out the division of tri patch, extracts the HSI features of all tri patch；

Characteristic vector computing unit, for the H features according to selected tri patch and the adjacent triangular faces on its shared side Value and S characteristic values carry out statistics with histogram, and statistics is put into characteristic vector, by the characteristic vector of tri patch and its phase The characteristic vector of adjacent tri patch is overlapped, stack result as selected tri patch characteristic vector；

Training unit, is trained for the characteristic vector of all samples to be input into SVM；The decision function of SVM is：

Wherein, L presentation classes sum, value 2；c_iRepresentative function coefficient, c_i∈R；γ represents the parameter of cross check, γ ∈R₊；Expression is chosen for the support vector of training；h⁰Represent the non-supported vector of training；Using histogrammic card side's distanceAs kernel function, it is defined as：

Wherein, N represents characteristic vector total quantity in sample set；Represent k-th characteristic vector it is corresponding support to Amount；h⁰K () represents k-th characteristic vector；

Model construction unit, for obtaining SVM identification model parameters after training, builds SVM identification models.

In general, by the contemplated above technical scheme of the present invention compared with prior art, it is special with following technology Levy and beneficial effect：

(1) technical solution of the present invention make use of the triangle gridding in the three-dimensional scenic for reconstructing, and three-dimensional scenic is divided into Multiple two dimension tri patch, it is to avoid super-pixel is carried out in three-dimensional scenic and extracts excessively complicated problem；

(2) technical solution of the present invention uses Analysis On Multi-scale Features information and neighboring information fusion, and easy and effective, discrimination is high, Three-dimensional scenic resurfacing can also be further used for and scene understands the problems such as analyzing.

Brief description of the drawings

Fig. 1 is embodiment of the present invention meadow recognition methods flow；

Fig. 2 is the flow chart of the SVM training of embodiment of the present invention meadow identification；

Fig. 3 a are embodiment of the present invention meadow sample H data statistic histograms；

Fig. 3 b are embodiment of the present invention sky sample H data statistic histograms.

Specific embodiment

In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and It is not used in the restriction present invention.As long as additionally, technical characteristic involved in invention described below each implementation method Not constituting conflict each other can just be mutually combined.

It is as shown in Figure 1 including the following step using the meadow recognition methods flow of the three-dimensional scenic of technical solution of the present invention Suddenly：

(1) three-dimensional scenic to be identified is carried out into three-dimensional reconstruction, by the triangulation tri patch on three-dimensional scenic surface.

(2) the HSI features of all tri patch units are extracted.

(3) the H characteristic values and S characteristic values according to selected tri patch and the adjacent triangular faces on its shared side carry out Nogata Figure statistics, statistics is put into characteristic vector, by the feature of the characteristic vector of tri patch and its adjacent triangular faces to Amount is overlapped, stack result as selected tri patch characteristic vector.

(4) characteristic vector of selected tri patch is input in SVM identification models, obtains recognition result.

It is illustrated in figure 2 the flow of the meadow SVM identification models using the structure three-dimensional scenic of technical solution of the present invention：

(1) 50 positive samples and 200 negative samples are collected, selects to be all the image on meadow as positive sample in order to carry out Training set is sampled and feature extraction, because the negative sample species to be covered is complicated, Pascal's data set (PASCAL is selected here VOC all kinds of non-meadow target in) is used as negative sample.

(2) it is 162*162 by the size adjusting of positive Negative training sample, training sample is made into regular right angle trigonometry dough sheet divides Treatment, the size of tri patch is 3 kinds, is respectively 3 pixels of the right angle length of side, 6 pixels and 9 pixels；Due to sample This image is RGB patterns, but due to RGB patterns at different conditions, the color Distribution value dispersion of RGB is measured, in rgb space two The non-linear relation of Euclidean distance and color distance of point, is unfavorable for carrying out the classification of color characteristic.To improve recognition effect, will Color represents the HSI color spaces that closer eye-observation mode is converted into from RGB color, so as to improve under different illumination The robustness of identifying system under environment.In HSI color spaces：H represents colourity；S represents saturation degree；I represents brightness.3 points It is I component by illumination effect maximum in amount, and H and S are typically the light absorption and reflectance signature of the raw material for constituting object Determine.In order to reduce influence of the illumination to accuracy of identification, when target is recognized, using H and S as the feature of extraction.From RGB to The conversion formula of HSI is as follows：

I=(R+G+B)/3

S=1-3min (R, G, B)/I

If B ＞ G, H=360-H

(3) the H characteristic values and S characteristic values according to selected tri patch and the adjacent triangular faces on its shared side carry out Nogata Figure statistics, statistics is put into characteristic vector, by the feature of the characteristic vector of tri patch and its adjacent triangular faces to Amount is overlapped, stack result as selected tri patch characteristic vector.H features value histogram and S characteristic value Nogatas are set The group number of figure be 8, group away from the 1/8 and S characteristic value spans of respectively H characteristic values span 1/8.By tri patch H characteristic values carry out statistics with histogram and obtain one group of 8 dimensional vector, and the S characteristic values of tri patch are carried out into statistics with histogram also obtains One group of 8 dimensional vector, the characteristic vector of the dimension of composition 16 that 8 dimensional vectors of 8 dimensional vectors of H characteristic values and S characteristic values are joined end to end.Will The characteristic vector of tri patch preserves into mat files and is used for SVM.It is as shown in Figure 3 a certain tri patch in sky sample H statistics with histogram results, are as shown in Figure 3 b the H statistics with histogram results of certain tri patch in the sample of meadow.

(4) characteristic vector of all samples is input into SVM to be trained；The decision function of SVM is：

Wherein, L presentation classes sum, value 2；c_iRepresentative function coefficient, c_i∈R；γ represents the parameter of cross check, γ ∈R₊；Expression is chosen for the support vector of training；h⁰Represent the non-supported vector of training；Using histogrammic card side's distanceAs kernel function, it is defined as：

Wherein, N represents characteristic vector total quantity in sample set；Represent k-th characteristic vector it is corresponding support to Amount；h⁰K () represents k-th characteristic vector.

Covariance matrix space is calculated according to histogrammic card side distance.Mirror image change is carried out to this covariance matrix space Get corresponding hyperplane matrix in return, that is to say, that it, by a mirror image for hyperplane reflection, is a kind of to be by a vector transformation Linear transformation, SVM is exactly that optimal hyperlane is constructed in feature space.Obtain covariance matrix and its corresponding hyperplane square Battle array, the characteristic coefficient of each feature is calculated according to the two matrixes, and characteristic coefficient is zoomed in and out to covariance matrix respectively. After being inverted to the covariance matrix after scaling, computation model parameter.The parameter of the SVM models that will be trained is stored in matrix file For the identifying call of follow-up three-dimensional scenic meadow.

Presently preferred embodiments of the present invention is the foregoing is only, is not intended to limit the invention, it is all in essence of the invention Any modification, equivalent and improvement made within god and principle etc., should be included within the scope of the present invention.

Claims (8)

1. a kind of meadow recognition methods of the three-dimensional scenic based on tri patch, it is characterised in that the method is comprised the following steps：

(1) three-dimensional scenic to be identified is carried out into three-dimensional reconstruction, by the triangulation tri patch on three-dimensional scenic surface；

(2) the HSI features of all tri patch units are extracted；

(3) the H characteristic values and S characteristic values according to selected tri patch and the adjacent triangular faces on its shared side enter column hisgram system Meter, statistics is put into characteristic vector, and the characteristic vector of the characteristic vector of tri patch and its adjacent triangular faces is entered Row superposition, stack result as selected tri patch characteristic vector；

(4) characteristic vector of selected tri patch is input in SVM identification models, obtains recognition result.

2. the meadow recognition methods of a kind of three-dimensional scenic based on tri patch according to claim 1, it is characterised in that Enter column hisgram in the step (3) to be specifically configured to：

It is 8 to set H features value histogram and S characteristic values histogrammic group of number, group away from respectively H characteristic values span 1/8 With the 1/8 of S characteristic value spans.

3. the meadow recognition methods of a kind of three-dimensional scenic based on tri patch according to claim 1, it is characterised in that Statistics is put into characteristic vector in the step (3) is specially：

The H characteristic values of tri patch are carried out into statistics with histogram and obtains one group of 8 dimensional vector, the S characteristic values of tri patch are carried out Statistics with histogram also obtains one group of 8 dimensional vector, and 8 dimensional vectors of 8 dimensional vectors of H characteristic values and S characteristic values are joined end to end composition The characteristic vector of 16 dimensions.

4. a kind of meadow recognition methods of the three-dimensional scenic based on tri patch according to claim 1,2 or 3, its feature It is that the training process of SVM identification models is specially in the step (4)：

(41) multiple positive samples and multiple negative samples are collected as sample set, positive sample chooses the X-Y scheme for being all object to be identified Piece, negative sample chooses the two-dimension picture without object to be identified；

(42) it is consistent by the size adjusting of all samples, selects multiple default yardsticks to carry out tri patch to sample and draw Point, extract the HSI features of all tri patch；

(43) the H characteristic values and S characteristic values according to selected tri patch and the adjacent triangular faces on its shared side enter column hisgram Statistics, statistics is put into characteristic vector, by the characteristic vector of tri patch and the characteristic vector of its adjacent triangular faces Be overlapped, stack result as selected tri patch characteristic vector；

(44) characteristic vector of all samples is input into SVM to be trained；The decision function of SVM is：

$C\left(h^0\right)=\stackrel{L}{\Σ}{c}_i\exp (-{\mathrm{\γd}}_{{\mathrm{\χ}}^2}^2(h^0,h_i^0\left)\right),$

Wherein, L presentation classes sum, value 2；c_iRepresentative function coefficient, c_i∈R；γ represents the parameter of cross check, γ ∈ R₊；Expression is chosen for the support vector of training；h⁰Represent the non-supported vector of training；Using histogrammic card side's distanceAs kernel function, it is defined as：

$d_{{\mathrm{\χ}}^2}^2(h^0,h_i^0)=\underset{k=1}{\overset{N}{\Σ}}\frac{{\left(h^0\right(k)-h_i^0(k\left)\right)}^2}{h^0\left(k\right)+h_i^0\left(k\right)}$

Wherein, N represents characteristic vector total quantity in sample set；Represent the corresponding support vector of k-th characteristic vector；h⁰ K () represents k-th characteristic vector；

(45) SVM identification model parameters are obtained after training, SVM identification models are built.

5. a kind of meadow identifying system of the three-dimensional scenic based on tri patch, it is characterised in that the system is included with lower module：

Tri patch division module, for three-dimensional scenic to be identified to be carried out into three-dimensional reconstruction, by the triangulation network on three-dimensional scenic surface Lattice divide tri patch；

HSI characteristic extracting modules, the HSI features for extracting all tri patch units；

Characteristic vector computing module, for H characteristic values and S according to selected tri patch and the adjacent triangular faces on its shared side Characteristic value carries out statistics with histogram, and statistics is put into characteristic vector, by the characteristic vector of tri patch adjacent with its three The characteristic vector of edged surface piece is overlapped, stack result as selected tri patch characteristic vector；

Identification module, for the characteristic vector of selected tri patch to be input in SVM identification models, obtains recognition result.

6. the meadow identifying system of a kind of three-dimensional scenic based on tri patch according to claim 5, it is characterised in that Enter column hisgram in the characteristic vector computing module to be specifically configured to：

It is 8 to set H features value histogram and S characteristic values histogrammic group of number, group away from respectively H characteristic values span 1/8 With the 1/8 of S characteristic value spans.

7. the meadow recognition methods of a kind of three-dimensional scenic based on tri patch according to claim 5, it is characterised in that Statistics is put into characteristic vector in the characteristic vector computing module is specially：

The H characteristic values of tri patch are carried out into statistics with histogram and obtains one group of 8 dimensional vector, the S characteristic values of tri patch are carried out Statistics with histogram also obtains one group of 8 dimensional vector, and 8 dimensional vectors of 8 dimensional vectors of H characteristic values and S characteristic values are joined end to end composition The characteristic vector of 16 dimensions.

8. a kind of meadow recognition methods of the three-dimensional scenic based on tri patch according to claim 5,6 or 7, its feature It is that the training system of SVM identification models is included with lower unit in the identification module：

Sample collection unit, for collecting multiple positive samples and multiple negative samples as sample set, positive sample is chosen to be all and waits to know The two-dimension picture of other object, negative sample chooses the two-dimension picture without object to be identified；

Tri patch division unit, for being consistent by the size adjusting of all samples, selects multiple default yardsticks to sample The division of tri patch is carried out, the HSI features of all tri patch are extracted；

Characteristic vector computing unit, for H characteristic values and S according to selected tri patch and the adjacent triangular faces on its shared side Characteristic value carries out statistics with histogram, and statistics is put into characteristic vector, by the characteristic vector of tri patch adjacent with its three The characteristic vector of edged surface piece is overlapped, stack result as selected tri patch characteristic vector；

Training unit, is trained for the characteristic vector of all samples to be input into SVM；The decision function of SVM is：

$C\left(h^0\right)=\stackrel{L}{\Σ}{c}_i\exp (-{\mathrm{\γd}}_{{\mathrm{\χ}}^2}^2(h^0,h_i^0\left)\right),$

Wherein, L presentation classes sum, value 2；c_iRepresentative function coefficient, c_i∈R；γ represents the parameter of cross check, γ ∈ R₊；Expression is chosen for the support vector of training；h⁰Represent the non-supported vector of training；Using histogrammic card side's distanceAs kernel function, it is defined as：

$d_{{\mathrm{\χ}}^2}^2(h^0,h_i^0)=\underset{k=1}{\overset{N}{\Σ}}\frac{{\left(h^0\right(k)-h_i^0(k\left)\right)}^2}{h^0\left(k\right)+h_i^0\left(k\right)}$

Wherein, N represents characteristic vector total quantity in sample set；Represent the corresponding support vector of k-th characteristic vector；h⁰ K () represents k-th characteristic vector；

Model construction unit, for obtaining SVM identification model parameters after training, builds SVM identification models.