CN103971377A - Building extraction method based on prior shape level set segmentation - Google Patents

Abstract

The invention provides a building extraction method based on prior shape level set segmentation in order to overcome the shortcomings that the conventional building extraction technology is sensitive to noise and low in extraction precision, and depends on a segmentation result. According to the basic idea of the building extraction method based on the prior shape level set segmentation, urban building extraction is carried out in two steps, wherein in the first step, a rough outline of a building in altitude data is extracted, and morphological filtering and small-region removal are performed on the altitude data to obtain a final building distribution binary graph, and in the second step, outline information of the building is extracted in an optical image of an unmanned aerial vehicle, the building distribution binary graph serving as prior information is introduced into a level set segmentation model, and an evolution initial position and a prior outline are supplied to the model, so that the segmentation result is a building outline information extraction result. The building extraction method has the advantages of high pertinence and high precision.

Description

The buildings extracting method of cutting apart based on prior shape level set

Technical field

The invention belongs to computer digital image and process and mode identification technology, relate generally to City Building recognition methods, be specifically related to a kind of buildings extracting method based on unmanned aerial vehicle remote sensing images.

Background technology

The extraction of City Building profile information is also not good at present, the algorithm that the scope of application is very wide.Traditional engineering is all to adopt artificial interpretation, manual mode of sketching, and operating efficiency is low, and labour intensity is large.Therefore the automatic extractive technique of contour of building information becomes the important subject of field of target recognition, and the main method of contour of building information automation extraction has at present:

1, the buildings extractive technique based on geometric properties

Common ground based on geometric properties is to utilize lower-level vision technology to extract the geometric elements such as straight line in image, circular arc, polygon, then with middle-and-high-ranking vision technique, target (Meng Yu, 2009) is expressed, is described as to these geometric elements.G.Roth and M.Levine have introduced the extracting method (Roth.G based on circular shape target, 1994), Lin, A.Huertas and R.Nevatia adopt the concept of consciousness grouping by line grouping and then extraction buildings (LinC, 1994), T.Kim and J.P.Muller, by setting up linear relation figure, utilize cost function minimum criteria to extract rectangle (Kim T, 1999); Zhang Yu etc. have proposed a kind of geometrical constraint and have cut apart with image the semiautomatic building extraction method (Zhang Yu, 2000) combining.

These class methods obtain good effect at present in the time detecting regular house and housing density rural image little and simple in structure, but for general pattern, also there are many problems: (1) is due to noise, shade and the impact of factor such as block, and often cannot obtain continuous edge or line segment; (2) a large amount of noise edge sharply increases the calculated amount of target detection, the testing result even leading to errors.

List of references:

Roth.G，Levine M.Geometric.primitive extraction using a geneticalgorithm[J].IEEE Trans.PAMI.1994，16(9)：906-910

Lin C，Huertas A，Necatia R.Detection of buildings using perceptualgrouping and shadows[C].Proc.Of IEEE computer science conference onComputer Vision and PatternRecognition，Seattle，Washington，USA，1 994(1)：62-69

Kim T，Muller J P.Development of a graph-based approach forbuilding detection[J].Image and Vision Computing，1999，17(1)：3-14.

Zhang Yu, Zhang Zuxun, Zhang Jianqing. the quick House Semi-automatic that geometrical constraint combines with Image Segmentation extracts [J]. Surveying & Cartography Scientific & Technological Univ., Wuhan's journal, 2000,25 (3): 238-242.

2, the buildings extractive technique based on stereo-picture

Traditional image capture method often uses two cameras to take same object, to obtain stereo-picture.Therefore, the automatic buildings in stereo-picture is identified also by the problem of extensive concern.

Fua and Hanson reach the object of checking selectively with stereo-picture.First produce the hypothesis of some buildingss according to piece image, then project in another piece image contrary this width image, verify the correctness (P.Fua, 1986.) of buildings hypothesis according to the relation between the new projected image producing and other piece image.Horaud and Skordas be detection of straight lines in image first, and then between two width images, finds straight line and combine and determine buildings (RaduHoraud, 1989).Wang and Schenk have also proposed a similarly algorithm based on feature.They utilize orthogonal this priori of building walls to improve the correctness (Zheng Wang, 1992) of testing result.

But the image of a scene is more, the information that it provides is also more, but also makes the cost of taking, take care of and safeguarding improve simultaneously.In addition, also there is the problem of how to synthesize the information comprising in multiple image.

List of references:

P.Fua and A.J.Hanson.Resegmentation. Using GenericShape：Locating General Cultural Objects，Technical Report[C].ArtificialIntelligence Center，SRI International，1 986.

Radu Horaud，Thomas Skordas.Stereo Correspondence ThroughFeature Grouping and Maximal Cliques[J].IEEE Trans.PAMI，1 989，11(11)：1168-1180

Zheng Wang and Toni Schenk.Urban Area SurfaceAanalysis[C].ISPRS XVII，Commission III，International Archives ofPhotogrammetry and Remote Sensing.1992，２9(3)：720-726

3, based on primitive buildings extractive technique

Based on primitive the common ground of profile information extractive technique be first to Image Segmentation Using, then to cutting apart the object of rear generation, sort out according to the spectrum of its subject area, texture, several how feature, to obtain building target.The use division merging methods such as Yanfeng Wei are cut apart city remote sensing image, find out the image category at buildings place, then depict the appearance profile (YanfengWei, 2004) of buildings by the method for finding construction zone main section.Jie Tian etc. has proposed a kind of urban architecture thing boundary extraction method based on image region segmentation, and first the method utilizes OO sorting technique to Remote Sensing Image Segmentation, thereby obtains buildings corresponding region in image.Then extract the edge of construction zone, on these bases, edge, find the principal direction of buildings, finally depict its regular profile (Jie Tian, 2004) according to the principal direction of buildings.Song et al. has proposed the buildings extraction algorithm based on region.First, the descriptive model (being mainly to utilize texture and shape) about building target by sample acquisition; By image division method, obtain over-segmentation image; Identify the cutting unit with the building model of previous definition with model identical; Merge these cutting units, and extract the straight-line segment relevant to these unit, utilize the profile of straight-line segment and merged unit to build the profile of buildings, and carried out verifying (Song Z., 2006) by shade and geometrical rule.

The advantage of this class methods maximum is to make full use of the information such as spectrum, texture, space and the structure of remote sensing image, to describe more accurately building target; Noise in image and too much details are caused extracting in result simultaneously and occur that " spiced salt " phenomenon has certain inhibition.Its shortcoming is that buildings extraction accuracy extremely depends on image segmentation precision.

List of references:

Yanfeng Wei，Zhongming Zhao and Jianghong Song.Urban buildingextraction from high-resolution satellite panchromatic image usingclustering and edge detection[C].IGARSS'04.2004(3)：2008-201

Jie Tian，Jinfei Wang，Peijun Shi.Urban Building BoundaryExtraction From IKONOS Imagery[C].The Canadian Remote SensingSociety'04 Oct.2004，poster session.

Song Z.，Pan C.，and Yang Q.A.Region-Based Approach toBuilding Detection in Densely Build-Up High Resolution SatelliteImage[C].ICIP2006：3225-3228

Summary of the invention

(1) goal of the invention

The object of the invention is: for the extraction of the three-dimensional information of the building demand based on remote sensing images, provide a kind of buildings extracting method of cutting apart based on prior shape level set, based on unmanned plane aeronautical data, can accurately extract the three-dimensional information of buildings.

(2) technical solution

Principle of the present invention is: utilize in altitude figures building feature obviously and unmanned plane optical image contour of building feature clearly, altitude figures is combined with and carries out buildings extraction with unmanned plane optical image.

Basic ideas of the present invention are: City Building is extracted, be divided into two parts, Part I carries out the extraction of the rough profile of buildings in altitude figures, and altitude figures is carried out to morphologic filtering, removing of small regions, obtains final buildings distribution binary map; Part II carries out the extraction of contour of building information in unmanned plane optical image, the buildings binary map that distributes is incorporated in level set parted pattern as prior imformation, for this model provides evolution initial position and priori profile, segmentation result is contour of building information extraction result.

Technical scheme of the present invention is:

Step S1 extracts buildings in elevation image;

The described buildings that extracts in elevation image refers to the thought according to regularization altitude figures, adopt multi-scale morphology filtering method that topographic entity and non-topographic entity are distinguished, and then by removing of small regions method, non-buildings key element is eliminated from non-topographic entity, finally obtain buildings and extract result.Detailed process is as follows:

(S11) elevation image is carried out to the multi-scale morphology filtering computing based on the gradient.First altitude figures is carried out to morphology opening operation, structural elements initial window size is set as MxM; Filtering for the first time weeds out the tiny atural object such as trees, electric pole.Altitude figures after treatment is carried out to gradient calculating, and the inspection gradient, whether lower than a certain threshold value, if not, increases structural elements window, again carries out filtering until its gradient lower than a certain threshold value, generates final topographic entity image.

(S12) altitude figures is carried out to regularization and binaryzation.Elevation image is deducted to topographic entity image and obtain non-topographic entity image, non-topographic entity now comprises trees, buildings etc.Then according to actual conditions, will think to build object point higher than the pixel of a certain threshold value, thereby complete the binaryzation process between buildings and non-buildings.

(S13) binaryzation image is carried out based on morphologic removing of small regions.First carry out the erosion operation that window is NxN, carry out zonule area statistics to processing rear image, reject the bin of area lower than T, finally result image is carried out the dilation operation of NxN, the buildings completing in elevation image extracts.

Step S2 carries out pre-service to Unmanned Aerial Vehicle Data

Unmanned plane image resolution is high, texture complexity, and details noise is more.And level set based on prior shape to cut apart be that edge based on image is cut apart, easily affected by noise, therefore need Unmanned Aerial Vehicle Data to carry out pre-service.Preprocessing process mainly comprises that noise filtering and edge strengthen two parts.Detailed process is as follows:

(S21) Unmanned Aerial Vehicle Data is carried out to the noise filtering based on anisotropy diffusion principle.The method can, in keeping atural object edge, be carried out noise filtering.

(S22) image after noise filtering is carried out to edge enhancing, the present invention selects the first order derivative of Gaussian function as the function of computed image gradient amplitude, then utilizes bounded inverse to carry out edge enhancing.

Step S3 extracts result based on the buildings in S1 and carries out obtaining of buildings prior shape.

Level set based on prior shape is cut apart, carried out the extraction of contour of building information in Unmanned Aerial Vehicle Data, but because the shape of buildings varies, the sample set of buildings prior shape is difficult to unified.But by manually delineating, be again bothersome, a to require great effort thing.Therefore the present invention is based on the sample set of the buildings result structure buildings extracting in S1.The prerequisite of carrying out this step is that the buildings in unmanned plane image is taking rectangle as main.Key step is as follows:

(S31) extraction of the buildings based in S1 result is set up the convex closure of each buildings, and convex closure construction method adopts better simply Graham method.Then set up its minimum boundary rectangle based on convex closure, its method adopting is rotary process.

(S32) taking minimum boundary rectangle as foundation, the in the situation that, center similar in rectangular aspect ratio being constant, this rectangle is carried out to convergent-divergent in proportion, Zoom method adopts the method for affined transformation to realize.The shape obtaining after convergent-divergent forms the prior shape sample set of this buildings, and this sample set can reflect average shape and the probability distribution of buildings preferably.

Step S4 adopts the level set based on prior shape to cut apart to carry out buildings extraction

In Unmanned Aerial Vehicle Data, carry out the extraction of contour of building information, the present invention adopts the level set dividing method based on prior shape, the method can be avoided near the similar noise effect of gray scale of buildings, can ignore again inner pseudo-object edge, thereby acquisition is similar with prior shape, and keep the buildings of real profile details to extract result.Key step is as follows:

(S41) the buildings prior shape sample set of setting up in S3 is carried out to dimension-reduction treatment, to eliminate the bulk redundancy information existing between sample, reduce the complicacy of calculating.The dimension reduction method of taking in the present invention is principal component analytical method (PCA), and the weighting of the use average shape that in sample set, objective contour can be similar to and a front m changing pattern carrys out approximate representation:

$\tilde{X}=\stackrel{\‾}{X}+\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_i{Y}_i---\left(1\right)$

Wherein for the average of sample set, Y _ibe i major component, α _ibe weight parameter, by adjusting weight parameter, can produce new Statistical Shape example, but that weight parameter can not change be too large, otherwise can produce relatively large deviation with the sample in sample set.Conventionally there is following restriction:

$-3\sqrt{{\mathrm{\λ}}_i}\≤{\mathrm{\α}}_i\≤-3\sqrt{{\mathrm{\λ}}_i},(i=\mathrm{1,2},...,m)---\left(2\right)$

(S42) carrying out level set based on the prior shape of setting up in S41 cuts apart.The basic thought that level set is cut apart is based on curve evolvement theory, and the method is utilized the Deformation Law of closed curve, and the energy function of definition tolerance closed curve deformation, makes closed curve approach gradually the edge of objective area in image by minimization of energy function.Profile information for buildings detects, and this process can be considered as to the evolutionary process of curve, and last curve lock fixes on buildings edge, and the profile information that completes buildings extracts.The energy function of its closed curve is expressed as follows:

$\mathrm{\μL}\left(\mathrm{\τ}\right)+\mathrm{vS}\left(\mathrm{\τ}\right)+{\mathrm{\λ}}_a\underset{\mathrm{inside}\left(\mathrm{\τ}\right)}{\∫}|I-C_a|\mathrm{dxdy}+{\mathrm{\λ}}_b\underset{\mathrm{outside}\left(\mathrm{\τ}\right)}{\∫}|I-C_b|\mathrm{dxdy}---\left(3\right)$

Wherein, E is the energy of closed curve, μ>=0, v>=0, λ _a>=0, λ _bthe>=0th, weighting function, τ refers to closed curve, and L (τ) is the length of τ, and S (τ) is the inner area of τ, and I is image pixel value, C _agray average in region, C _bbe region exterior gray average, inside (τ) is the picture point set of curve inside, and outside (τ) is the picture point set of curve outside.From above formula, in the time that τ develops to object boundary, energy E minimum.

But present stage, level set was cut apart two larger limitations of existence, and one is that level set parted pattern is only the setting that realizes parted pattern by adjusting parameter, and the priori of combining target profile, does not therefore have selective recognition to the target being partitioned into.Another limitation is that level set parted pattern is that edge based on image is cut apart; the image border that is about to converge to is as the geometric profile of target; but under actual conditions; the geometric profile of target might not be consistent with the edge of image, often there will be the edge extracting of noise effect image, and then affect the profile information extraction of target; simultaneously; because the shape constraining of parted pattern is few, often can be attracted by pseudo-edge or non-target, cause segmentation result poor.Therefore the present invention is incorporated into level set by the priori Statistical Shape of target and cuts apart in energy function, obtains new energy function, as follows:

E _all=E _f+βE _shape(β＞0) (4)

Wherein, E _allfor total energy function, E _ffor Fast Marching parted pattern energy function, E _shapefor the energy function of target prior shape, β is for adjusting parameter.

Known by analyzing, the target shape in sample set is always near an average shape, and we can think that target shape distribution meets Gaussian distribution approx, and hypothetical target is shaped as z, and the statistical distribution of target shape is:

$P\left(z\right)=\frac{1}{\sqrt{{\left(2\mathrm{\π}\right)}^m\left|S_m\right|}}\exp (-\frac{1}{2}{(z-z_0)}^T{S}_m^{-1}(z-z_0))---\left(5\right)$

Wherein, m is the number of the principal component chosen, S _mthe capable and front m row of the front m of covariance matrix, z ₀it is the average shape of target shape.Our prior imformation using this probability as target shape, joins in Fast Marching parted pattern, and guiding is cut apart to more rational Direction distortion.This is distributed on whole space is continuous, and it provides a non-zero statistical value for each shape z, can be obtained by above-mentioned distribution, and the energy function of target shape is:

$E_{\mathrm{shape}}=-\log \left(P\left(z\right)\right)=\frac{1}{2}{(z-z_0)}^T{S}_m^{-1}(z-z_0)---\left(6\right)$

From the above mentioned, the level set parted pattern total energy function based on prior shape is as follows:

$\begin{array}{c}{E}_{\mathrm{all}}=E_f+{\mathrm{\βE}}_{\mathrm{shape}}\\ =\mathrm{\μL}\left(\mathrm{\τ}\right)+\mathrm{vS}\left(\mathrm{\τ}\right)+{\mathrm{\λ}}_a\underset{\mathrm{inside}\left(\mathrm{\τ}\right)}{\∫}|I-C_a|\mathrm{dxdy}+{\mathrm{\λ}}_b\underset{\mathrm{outside}\left(\mathrm{\τ}\right)}{\∫}|I-C_b|\mathrm{dxdy}+\mathrm{\β}\frac{1}{2}{(z-z_0)}^T{S}_m^{-1}(z-z_0)\end{array}---\left(7\right)$

Wherein, μ>=0, v>=0, λ _a>=0, λ _b>=0, β>=0th, weighting function, τ refers to closed curve, and L (τ) is the length of τ, and S (τ) is the inner area of τ, and I is image pixel value, C _agray average in region, C _bbe region exterior gray average, inside (τ) is the picture point set of curve inside, and outside (τ) is the picture point set of curve outside.

As total energy function E _allreach hour, curve evolvement, to contour of building edge, completes contour of building information extraction.

(3) technique effect

The present invention compares and has following advantage and beneficial effect from existing technical scheme: City Building 3-D information fetching method of the present invention is different with other extracting method, the supplementary of the present invention using the elevation information based on unmanned plane image capturing as buildings position, reduce in a large number the false detection rate of City Building, and the elevation information extraction accuracy of its buildings is higher, can build for digital city, Disaster Assessment etc. provides effective reference data.

Brief description of the drawings

Fig. 1 is the buildings extracting method schematic diagram of cutting apart based on prior shape level set of the present invention

Fig. 2 is original instance data of the present invention

Fig. 3 is process and the result figure that carries out multi-scale morphology filtering based on Fig. 2 (b)

Fig. 4 is the buildings two-value distribution plan (white portion of image is as buildings) obtaining as threshold value taking 3.5 meters

Fig. 5 carries out removing of small regions buildings afterwards based on Fig. 4 to extract result

Fig. 6 carries out the result after the noise filtering based on anisotropy diffusion principle to unmanned plane RGB image

Fig. 7 carries out the result figure after edge enhancing based on Fig. 6

Fig. 8 is the result figure that asks for its minimum boundary rectangle based on a buildings in buildings distribution plan.(yellow square frame is minimum boundary rectangle)

Fig. 9 carries out based on the minimum boundary rectangle of buildings the buildings prior shape sample set that different proportion convergent-divergent obtains

Figure 10 is the probability distribution graph of the buildings prior shape obtained by principal component analysis (PCA)

Figure 11 is that the buildings of cutting apart based on prior shape level set extracts result figure

Embodiment

In conjunction with Fig. 1-Figure 11, method of the present invention is done to further detailed elaboration below:

Instance data is chosen at takes Jilin Province's Hunchun City aviation RGB chromatic image and corresponding elevation image by unmanned plane on August 10th, 2008, as shown in Figure 2, takes camera and selects slr camera Cannon5D, and the specific embodiment of the present invention is described.

Step S1 extracts buildings in elevation image;

This step is mainly the thought according to regularization DSM, adopt multi-scale morphology filtering method that topographic entity and non-topographic entity are distinguished, and then by removing of small regions method, non-buildings key element is eliminated from non-topographic entity, finally obtain buildings and extract result.Detailed process is as follows:

(S11) elevation image is carried out to the multi-scale morphology filtering computing based on the gradient.First altitude figures is carried out to morphology opening operation, structural elements initial window size is set as 3x3; Filtering for the first time weeds out the tiny atural object such as trees, electric pole.Altitude figures after treatment is carried out to gradient histogram calculation, whether histogrammic 95% pixel value of the inspection gradient is lower than 35 degree, if not, increase structural elements window 5x5, again carry out filtering, until histogrammic 95% pixel value of its gradient is lower than 35 degree, generate final topographic entity image, this example is in the time that structural elements window is 11x11, and histogrammic 95% pixel value of its gradient, lower than 35 degree, generates final topographic entity image.As shown in Figure 3

(S12) altitude figures is carried out to regularization and binaryzation.Elevation image is deducted to topographic entity image and obtain non-topographic entity image, non-topographic entity now comprises trees, buildings etc.Then according to actual conditions, the relative height of buildings all exceedes 3.5m, and this is highly the height of the ground floor of buildings in general.Therefore the present invention selects the decision threshold of 3.5m as construction zone, thereby completes the binaryzation process between buildings and non-buildings.As shown in Figure 4.

(S13) binaryzation image is carried out based on morphologic removing of small regions.First carry out the erosion operation that window is 3x3, carry out zonule pixel count statistics to processing rear image, reject pixel count lower than 10 bin, finally result image is carried out the dilation operation of 3x3, the buildings completing in elevation image extracts.As shown in Figure 5.

Step S2 carries out pre-service to Unmanned Aerial Vehicle Data

This step mainly comprises that noise filtering and edge strengthen two parts.Detailed process is as follows:

(S21) Unmanned Aerial Vehicle Data is carried out to the noise filtering based on anisotropy diffusion principle.The method can, in keeping atural object edge, be carried out noise filtering.As shown in Figure 6.

(S22) image after noise filtering is carried out to edge enhancing, the present invention selects the first order derivative of Gaussian function as the function of computed image gradient amplitude, then utilizes bounded inverse to carry out edge enhancing.As shown in Figure 7.

Step S3 extracts result based on the buildings in S1 and carries out obtaining of buildings prior shape.

The present invention is based on the sample set of the buildings result structure buildings extracting in S1.Key step is as follows:

(S31) extraction of the buildings based in S1 result is set up the convex closure of each buildings, and convex closure construction method adopts better simply Graham method.Then set up its minimum boundary rectangle based on convex closure, in its rotary process, each anglec of rotation is set as to 1 degree.As shown in Figure 8.

(S32) taking minimum boundary rectangle as foundation, the in the situation that, center similar in rectangular aspect ratio being constant, this rectangle is carried out to convergent-divergent in proportion, Zoom method adopts the method for affined transformation to realize.The shape obtaining after convergent-divergent forms the prior shape sample set of this buildings, and in this example, length and width scaling adopts 1: 1,0.9: 0.9,09: 0.7,0.9: 0.5,0.7: 0.9,07: 0.7 totally 6 scalings.As shown in Figure 9.

Step S4 adopts the level set based on prior shape to cut apart to carry out buildings extraction

In Unmanned Aerial Vehicle Data, carry out the extraction of contour of building information, the present invention adopts the level set dividing method based on prior shape, the method can be avoided near the similar noise effect of gray scale of buildings, can ignore again inner pseudo-object edge, thereby acquisition is similar with prior shape, and keep the buildings of real profile details to extract result.Key step is as follows:

(S41) the buildings prior shape sample set of setting up in S3 is carried out to principal component analysis (PCA), reduce the complicacy of calculating.The weighting of the use average shape that in sample set, objective contour can be similar to and a front m changing pattern carrys out approximate representation:

$\tilde{X}=\stackrel{\‾}{X}+\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_i{Y}_i---\left(8\right)$

Wherein for the average of sample set, Y _ibe i major component, α _ibe weight parameter, by adjusting weight parameter, can produce new Statistical Shape example.

In this example, the present invention has retained first three major component and the average shape of prior shape sample set, as shown in figure 10.

(S42) carrying out level set based on the prior shape of setting up in S41 cuts apart.Level set parted pattern total energy function based on prior shape is as follows:

$\begin{array}{c}{E}_{\mathrm{all}}=E_f+{\mathrm{\βE}}_{\mathrm{shape}}\\ =\mathrm{\μL}\left(\mathrm{\τ}\right)+\mathrm{vS}\left(\mathrm{\τ}\right)+{\mathrm{\λ}}_a\underset{\mathrm{inside}\left(\mathrm{\τ}\right)}{\∫}|I-C_a|\mathrm{dxdy}+{\mathrm{\λ}}_b\underset{\mathrm{outside}\left(\mathrm{\τ}\right)}{\∫}|I-C_b|\mathrm{dxdy}+\mathrm{\β}\frac{1}{2}{(z-z_0)}^T{S}_m^{-1}(z-z_0)\end{array}---\left(9\right)$

Wherein, μ>=0, v>=0, λ _a>=0, λ _b>=0, β>=0th, weighting function, τ refers to closed curve, and L (τ) is the length of τ, and S (τ) is the inner area of τ, and I is image pixel value, C _agray average in region, C _bbe region exterior gray average, inside (τ) is the picture point set of curve inside, and outside (τ) is the picture point set of curve outside.As total energy function E _allreach hour, curve evolvement, to contour of building edge, completes contour of building information extraction.Its buildings carries out after level set cuts apart, and acquired results as shown in figure 11.

The above is only the preferred embodiment of the present invention; it should be pointed out that for those skilled in the art, do not departing under the prerequisite of the technology of the present invention principle; can also make some improvement and modification, these improve and modification also should be considered as protection scope of the present invention.

Claims (5)

1. a buildings extracting method, is characterized in that, the step comprising has:

Step S1, in elevation image, extract buildings distribution plan;

Step S2, unmanned plane RGB color data is carried out to pre-service;

Step S3, carry out the extraction of buildings prior shape sample set based on buildings distribution plan;

Step S4, adopt level set based on prior shape to cut apart to carry out buildings extraction.

2. according to a kind of buildings extracting method described in right 1, it is characterized in that, the described concrete steps of extracting buildings distribution plan in elevation image are:

Elevation image is carried out to the multi-scale morphology filtering computing based on the gradient.Generate the final topographic entity image of its overall gradient lower than a certain threshold value.

Altitude figures is carried out to regularization and binaryzation, and its binary-state threshold is set as actual one deck depth of building, thereby completes the binaryzation process between buildings and non-buildings.

The removing of small regions that binaryzation image is carried out to combining form and area statistics, corroding between expansion, carries out area statistics, and object less area is removed, and completes the buildings extraction in elevation image.

3. according to a kind of buildings extracting method described in right 1, it is characterized in that, describedly unmanned plane RGB color data is carried out to pretreated concrete steps be:

Unmanned Aerial Vehicle Data is carried out to the noise filtering based on anisotropy diffusion principle.

Image after noise filtering is carried out to edge enhancing, and the function using the first order derivative of Gaussian function as computed image gradient amplitude, then utilizes bounded inverse to carry out edge enhancing.

4. according to a kind of buildings extracting method described in right 1, it is characterized in that, the concrete steps of carrying out the extraction of buildings prior shape sample set based on buildings distribution plan are:

Adopt Graham method to set up the convex closure of each buildings, then adopt rotary process to set up its minimum boundary rectangle.

Taking minimum boundary rectangle as foundation, the in the situation that, center similar in rectangular aspect ratio being constant, set up the buildings prior shape sample set that a series of pantograph ratios are different.This sample set can reflect average shape and the probability distribution of buildings preferably.

5. according to a kind of buildings extracting method described in right 1, it is characterized in that, the level set based on prior shape is cut apart concrete steps and is:

Buildings prior shape sample set is carried out to principal component analysis (PCA), and first three major component by sample set and average shape reflect the probability distribution of its contour of building.

The probability distribution of contour of building is incorporated into level set parted pattern mutually as energy, and on unmanned plane RGB, colored impact is carried out cutting apart based on the level set of prior shape, obtains final buildings and extracts result.