CN105069843A - Rapid extraction method for dense point cloud oriented toward city three-dimensional modeling - Google Patents

Abstract

The invention provides a rapid extraction method for dense point cloud oriented toward city three-dimensional modeling. The rapid extraction method comprises the steps of: 1) preparing data; 2) establishing a coordinate grid; 3) converting an image into a gray scale image; 4) judging whether a difference value between a gray value of each target pixel point and a gray value of a neighborhood pixel point is greater than a threshold value A, if so, regarding the target pixel point as a candidate angular point, and if not then rounding; 5) calculating a corner response function (CRF) value of the candidate angular point, and judging whether the CRF value of the candidate angular point is greater than a threshold B, is so regarding the candidate angular point as a feature point, and if not then rounding; 5) matching homonymy points; 7) generating seed points; 8) conducting regional diffusion; 9) conducting gross error elimination; 10) and generating dense point cloud to form a tin for establishing a three-dimensional model. Compared with the prior art, the rapid extraction method provided by the invention has the advantages of small error, high precision, high operation efficiency, wide application extension and the like.

Description

A kind of rapid extracting method of the point of density cloud towards cybercity construction

Technical field

The present invention relates to a kind of extracting method of point of density cloud, specifically a kind of rapid extracting method of the point of density cloud towards cybercity construction.

Background technology

Oblique photograph technology is the new and high technology that international Mapping remote sensing technology field developed recently gets up, by carrying multiple stage sensor (conventional is five lens cameras at present) on same flying platform, simultaneously gather image from different angles such as vertical, inclinations, obtain the information of ground object more complete and accurate.The image that vertical ground angle shot obtains is called positive (group image), and camera lens is taken the image obtained be called inclined tab (four group images) towards being formed an angle with ground, is inclination image; Wherein, inclination image has following feature: (1) can obtain the image at multiple viewpoint and visual angle, thus obtains more detailed side-information; (2) there is higher resolution and larger field angle; (3) same atural object has the image of multiple resolution; (4) inclination image atural object eclipse phenomena is more outstanding; For these features, oblique photograph measuring technique can be used for quick three-dimensional modeling.

The image corrected is utilized to complete the three-dimensional modeling of atural object, obtain positional information intuitively, need to carry out intensive data reduction to image, just can structure tin after point of density cloud extracts, generating three-dimensional models, and the acquisition of point of density cloud has been come by Image Matching.

In the leaching process of point of density cloud, general flow process obtains the unique point on image, coupling same place, determine object space Seed Points, form sparse some cloud, carry out elimination of rough difference after diffusion obtains comparatively intensive some cloud thus obtain the higher point of density cloud of precision, for three-dimensional modeling lays reliable basis.The leaching process of point of density cloud can be described as the geometry of a series of constraint condition, in this process may there is many errors, add up and can have an impact to the precision of three-dimensional modeling, in the flow process of data reduction, still have and need perfect place, and Seed Points regional diffusion is the severely afflicated area that error produces, needs the some cloud elimination of rough difference repeatedly after to diffusion, improve precision.

In addition, nowadays the research of intensive data reduction is got more and more, the method that data reduction relates to also constantly is being improved, as in feature point extraction, initial many employing Morevec algorithms, consideration pixel moves the change that shade of gray produces to specific several direction, this method can not consider the situation of all shade of gray changes around pixel, therefore easy missing feature point, has certain harmful effect to follow-up data reduction; Many employing Harris algorithm extract minutiaes now, the improvement on Morevec basis, can, to the change of all angle detecting shade of gray around pixel, make the extraction of unique point more reasonable.But Harris algorithm also also exists limitation, during to feature point extraction, comparatively slow on the time of detection, operational efficiency is low.

Harris algorithm is a kind of Corner Detection Algorithm based on image greyscale, be applicable to the feature point extraction of inclination image, angle is preferably within 45 degree, more responsive for dimensional variation, it is a kind of effective interest point detect algorithm, the thought of Harris algorithm is: on image, arrange a window centered by detected pixel, along all directions minute movement window, investigate the mean change situation of gray-scale intensity in window, when this changing value exceedes the threshold value of setting, then tested measuring point is extracted as unique point, namely thinks angle point.

Its summary of benefits gets up to have: 1. calculate simple: first order difference and the filtering of only using gray scale in Harris algorithm, simple to operate; 2. the point patterns extracted is even and reasonable: Harris algorithm calculates its interest value to each point in image, then in neighborhood, selects optimum point.Experiment shows, in the region that texture information is abundant, Harris algorithm can extract unique point useful in a large number, and in the few region of texture information, the unique point of extraction is then less; 3. stablize: in the computing formula of Harris algorithm, only relate to first order derivative, therefore insensitive to the rotation of image, grey scale change, noise effect and viewpoint change.

During embody rule:

1. neighborhood is smooth, then to move change with window very little for gray-scale intensity; (R gets the positive number of fractional value, can not be angle point)

If 2. central point neighborhood comprises edge, then time along horizontal edge direction moving window, gray-scale intensity change is very little, and during along vertical edge direction moving window, gray-scale intensity alters a great deal; (R gets the negative of large numerical value)

If 3. comprise angle point in central point neighborhood, then, time along any direction moving window, gray-scale intensity all has significant change; (R gets the positive number of large numerical value)

Gray-scale intensity change function:

$E\left(u,v\right)=\underset{x,y}{\Σ}w\left(x,y\right){\[I\left(x+u,y+v\right)-I\left(x,y\right)\]}^2$

In order to the window of searching belt angle point, the window that search pixel grey scale change is larger, so expectation maximization E (u, v).

Can cover the directive detection of institute by differentiating and carry out Taylor expansion, its gray-scale intensity variable quantity may be defined as:

$E\left(u,v\right)\≈\left[\begin{array}{cc}u& v\end{array}\right]\left(\underset{x,y}{\Σ}w(x,y)\left[\begin{array}{cc}{I}_x^2& I_x{I}_y\\ I_x{I}_y& I_y^2\end{array}\right]\right)\left[\begin{array}{c}u\\ v\end{array}\right]$

Ix represents the difference in x direction, and Iy represents the difference in y direction, w (x, y) Gaussian function

$M=\underset{x,y}{\Σ}w(x,y)\left[\begin{array}{cc}{I}_x^2& I_x{I}_y\\ I_x{I}_y& I_y^2\end{array}\right]$

$E(u,v)\≈\left[\begin{array}{cc}u& v\end{array}\right]M\left[\begin{array}{c}u\\ v\end{array}\right]$

R＝det(M)-k(trace(M)) ²

(k is empirical value, often gets 0.04 ~ 0.06)

det(M)＝λ ₁λ ₂

trace(M)＝λ ₁+λ ₂

λ 1, λ 2 are two eigenwerts of second-order matrix M.R is angle point response function, when the R value of target pixel points is greater than the threshold value (because the classification of R value is obvious, when easily determining threshold values) of setting, can choose as angle point.

Concrete steps:

1. calculate Grad Ix and Iy on image on each pixel horizontal and vertical direction, try to achieve autocorrelation matrix M;

2. gaussian filtering is carried out to the image after calculating, obtain filtered M;

3. the value of the angle point response function of each pixel is calculated;

4. the Local modulus maxima alternatively angle point of R is chosen;

5. set corresponding threshold value and extract angle point.

In summary, the shortcoming of prior art mainly concentrate on following some:

A) towards in the application of city rapid modeling, have impact on the precision of three-dimensional modeling at present based on the point of density cloud extracting method of inclination image, its extracting method flow process still To be improved;

B) small and medium-sized cities modeling work adopts inclination image modeling pattern probably to need the time of three to five months, the Harris algorithm adopted in multi-view images point of density cloud leaching process needs to carry out angle point response to each pixel and detects in the process detecting angle point, wherein can relate to a large amount of multiplyings, cause computing velocity slow, affect the efficiency of work.

C) have a lot of error points in the some cloud obtained after the diffusion of multi-view images intensive data reduction Seed Points, the existence of error point can make the geometric configuration of atural object change, and has larger impact to follow-up three-dimensional modeling.

In order to solve the problems of the technologies described above, the present invention proposes a kind of high precision, high-level efficiency and applied range

Towards the rapid extracting method of the point of density cloud of cybercity construction.

Summary of the invention

The object of the invention is the deficiency existed to overcome above-mentioned existing point of density cloud extracting method, proposing a kind ofly to optimize intensive data reduction flow process, improve unique point acquisition algorithm to improve operational efficiency and the high precision of elimination of rough difference, high-level efficiency and the rapid extracting method of the point of density cloud towards cybercity construction of applied range can be carried out repeatedly the some cloud after spreading.

To achieve these goals, present invention employs following technical scheme:

Towards a rapid extracting method for the point of density cloud of cybercity construction, specifically comprise the following steps:

1) data encasement, data specifically comprise image elements of exterior orientation, inclination aerial surveying camera camera parameter information after the multi-view images data of oblique aerial photography acquisition, empty three encryptions, described multi-view images data include metadata, include image resolution, image projecting coordinate;

2) on the vertical image in multi-view images data in step 1, coordinate grid is set up in the actual demand become more meticulous according to cybercity construction, in coordinate grid, the size of each grid unit is not less than image resolution, and described image resolution is identical with the image resolution of step 1;

3) multi-view images is converted to gray level image, and by gaussian filtering to the smoothing process of gray level image;

4) in gray level image, calculate the gray-scale value of each target pixel points and the gray-scale value of its neighborhood territory pixel point, the gray-scale value of the gray-scale value of each target pixel points and its neighborhood territory pixel point is compared formation difference, multiple difference forms a numerical range, the size of threshold values A is determined in this numerical range, then judge whether the difference of the gray-scale value of each target pixel points and the gray-scale value of its neighborhood territory pixel point is greater than threshold values A, if, then this target pixel points is extracted alternatively angle point, otherwise, directly cast out;

5) angle point response function (CRF) value of candidate angular in calculation procedure 4, judge whether angle point response function (CRF) value of this candidate angular is greater than the threshold values B preset, if so, then this candidate angular is extracted as unique point, otherwise, directly cast out;

6) mate same place, select, with reference to image and search image, on search image, to find same place by core line geometry constraint condition with reference to the unique point on image;

7) generate Seed Points, utilize the data in step 1 to carry out initial matching, obtain Seed Points;

8) regional diffusion, i.e. Seed Points diffusion, interpolation generates the volume coordinate of point to be located and normal vector, obtains relatively intensive some cloud, makes it evenly intensively be distributed on regular grid;

9) elimination of rough difference, tightly separates mode and the least square thought by multi-disc forward intersection, rejects some cloud rough error point; And it is complete to judge whether grid unit calculates, and if so, then enters step 10, otherwise, get back to step 8 and proceed regional diffusion;

10) generate point of density cloud, structure tin sets up three-dimensional model.

The present invention is first on the basis of available data, set up spatial grid, Seed Points is spread and is easy to control, then Harris innovatory algorithm is utilized to extract as unique point, to Feature Points Matching same place, after obtaining object space Seed Points by every a pair same place, carry out regional diffusion, the thought of least square is adopted to carry out elimination of rough difference to the error point after diffusion, after obtaining the higher Seed Points of precision, repeat region diffusion and elimination of rough difference, until some cloud precision is met, be comparatively evenly distributed in spatial grid; The wherein extraction algorithm of unique point, improves the efficiency of feature point extraction in oblique aerial point of density cloud leaching process; After Seed Points diffusion, adopt the thought of least square to carry out iterative elimination of rough difference to the point of density cloud obtained, substantially increase precision and the density of data reduction.The cloud data extracted by flow process of the present invention may be used for structure tin, entered a series of data processing, model building method carries out cybercity construction, the development of cybercity construction technology has adapted to urban development demand, the structure of three-dimensional model makes the actual conditions of atural object obtain reflecting more really, the foundation of urban three dimensional landscape, by expressing in brand-new mode and processing geospatial information, plays an important role in fields such as city planning, development of real estate, traffic administration, tourisms.

Preferably, coordinate grid is set up about in step 2, the scope of graticule mesh comprises all coverage of image at object space participating in coupling, the size of grid unit represents the least unit of achievement point of density cloud, utilize regular grid that plane is divided into little bin, make it in the plane along the X, the Y direction regular distribution that are parallel to object space; Wherein, the size of bin is according to mating the precision that will reach or determining according to the resolution generating DSM, and it is not less than the resolution of corresponding image; Meanwhile, the planimetric coordinates at each bin center is fixing, and initialization flat element does not possess height value attribute.

Preferably, candidate angular is chosen about step 4, setting threshold values A is as a S (i, j) parameter of target pixel points and neighborhood territory pixel point difference, by scanning gray level image, if by the difference of the target pixel points that scans and its neighborhood territory pixel point at (-s, s) between, then this target pixel points is gray scale similitude, namely refer to that the difference of the gray-scale value of target pixel points and neighborhood territory pixel point is less than the target pixel points of threshold values A, judge whether this point is candidate angular according to the number of gray scale similitude in eight neighborhood, and record all candidate angular S (i, j).

Preferably, about the determination of threshold values B in step 5, Harris angle point response function (CRF) is utilized to determine threshold value B in gray level image, specifically, target pixel points and its neighborhood territory pixel point grey scale change little, Harris angle point response function (CRF) value is the positive number of fractional value, is not angle point; When target pixel points comprises edge, Harris angle point response function (CRF) value is the negative of large numerical value, is not angle point; Target pixel points is large along grey scale change during any direction moving window, Harris angle point response function (CRF) value is the positive number of large numerical value, clearly, wherein threshold value B is greater than the positive number of fractional value to the size discrimination of above three kinds of situation Harris angle point response function (CRF) values.

Preferably, in step 6, adopt vertical image with reference to image, search image adopts the image on same course line, camera of the same name, if cannot obtain the image on same course line, camera of the same name, searches for image and adopts with attitude image.

Preferably, about step 7, after search image obtains the same place with reference to image, utilize the results of empty three encryptions, determine the inside and outside azimuth information of same place, being crossed according to geometry by two same places obtains object space point, then each object space point is mapped in corresponding grid according to its volume coordinate, represent different elevation informations by different colors, be mapped to the initial seed point of the object space point in grid as dense Stereo Matching, and each graticule mesh is just containing a Seed Points.

Preferably, about the regional diffusion of step 8, the diffusion of Seed Points is realized by spatial interpolation, utilize regretional analysis, least square method carries out matching to Seed Points, point to be located in window is refined, by the some constructed fuction relational expression of known space of points information, interpolation generates volume coordinate and the normal vector of point to be located.

Seed Points has had coordinate and elevation, is point to be located to the point of spatial information that do not have around its graticule mesh.

Preferably, about step 9 elimination of rough difference, the mode of tightly being separated by multi-disc forward intersection, the least square thought are to reject some cloud rough error points;

Concrete mathematical model:

$\left\{\begin{array}{c}(x-x_0)\[a_3(X-X_s)+b_3(Y-Y_s)+c_3(Z-Z_s)\]=\\ -f\[a_1(X-X_s)+b_1(Y-Y_s)+c_1(Z-Z_s)\]\\ (y-y_0)\[a_3(X-X_s)+b_3(Y-Y_s)+c_3(Z-Z_s)\]=\\ -f\[a_2(X-X_s)+b_2(Y-Y_s)+c_2(Z-Z_s)\]\end{array}\right.$

Arrangement can obtain:

$\left\{\begin{array}{c}{l}_{1}X+l_{2}Y+l_{3}Z-l_x=0\\ l_{4}X+l_{5}Y+l_{6}Z-l_y=0\end{array}\right.$

Wherein:

l ₁＝fa ₁+(x-x ₀)a ₃

l ₂＝fb ₁+(x-x ₀)b ₃

l ₁＝fc ₁+(x-x ₀)c ₃

l _x＝fa ₁X _s+fb ₁Y _s+fc ₁Z _s+(x-x ₀)a ₃X _s+(x-x ₀)b ₃Y _s+(x-x ₀)c ₃Z _s

l ₄＝fa ₂+(y-y ₀)a ₃

l ₅＝fb ₃+(y-y ₀)b ₃

l ₆＝fc ₂+(y-y ₀)c ₃

l _y＝fa ₂X _s+fb ₂Y _s+fc ₂Z _s+(y-y ₀)a ₃X _s+(y-y ₀)b ₃Y _s+(y-y ₀)c ₃Z _s

To with reference to a pair same place on video and search image, 4 above-mentioned linear equations can be listed, and unknown number number is 3, therefore can solve by least square method, if containing same object space point in n width image, then can ask X by 2n linear equation solution altogether, Y, Z3 unknown number.

This is a kind of space intersection method strictly, retrained by image number, owing to being solve linear equations, therefore does not also need the initial value of volume coordinate; In order to improve the precision of a cloud further, the constraint of the geometric relationships such as the atural object such as buildings, zebra stripes straight line, right angle can be utilized, further excluding gross error point.

The technique effect that the present invention adopts technique scheme to obtain is:

The present invention propose a kind of can to diffusion after some cloud carry out elimination of rough difference repeatedly, and improve unique point and obtain flow process to improve the high precision of operational efficiency, high-level efficiency and the rapid extracting method of the point of density cloud towards cybercity construction of applied range, concrete advantage is as follows:

1) error reduces: the present invention out adopts the thought of least square afterwards to the elimination of rough difference of a cloud at data reduction, the mode of the tight solution that crossed by front, and retrain in conjunction with the geometric relationship of buildings etc. the dynamics that improve rough error point and reject, and to the Seed Points repeat region diffusion extracted and elimination of rough difference, until some cloud precision is met, comparatively be evenly distributed in spatial grid, error is less.

2) efficiency improves: on the basis of rapid extraction flow process proposing a kind of point of density cloud towards cybercity construction, the Harris algorithm of feature extraction makes improvements, the multiplication that original algorithm relates to is many, the computing time taken is many, and the method for improvement extracts a part of candidate angular, adds additive operation, but angle point response function need not be calculated to each pixel, computing velocity is very fast, in the former methodical advantage of guarantee, improves the efficiency of feature point extraction.

3) application is expanded: the application makes moderate progress on the extraction rate and precision of point of density cloud, can be used for rapidly constructing three-dimensional model, the degree of becoming more meticulous of model is improved, and then promote the visualization of spatial information of city three-dimensional model, the directiveness work of the industries such as urban construction, planning and design, urban traffic, emergency command is strengthened.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the present embodiment point of density cloud extracting method;

Fig. 2 is the present embodiment coupling grid partition schematic diagram;

Fig. 3 is the present embodiment kernel pixels point and neighborhood territory pixel point intensity contrast schematic diagram;

Fig. 4 is the present embodiment core line geometry constraint schematic diagram;

Fig. 5 is that the present embodiment Seed Points obtains schematic diagram;

Fig. 6 is the schematic diagram of the present embodiment unit graticule mesh many seeds point;

Fig. 7 is the spatial information schematic diagram that the present embodiment regional diffusion generates point to be located;

Fig. 8 is the present embodiment buildings geometric relationship schematic diagram;

Fig. 9 is the present embodiment three-dimensional model schematic diagram.

Embodiment

Below in conjunction with accompanying drawing and embodiment, the present invention is described in further detail:

Embodiment: a kind of rapid extracting method of the point of density cloud towards cybercity construction, as shown in Figure 1, specifically comprises the following steps:

1) data encasement, data specifically comprise image elements of exterior orientation, inclination aerial surveying camera camera parameter information after the multi-view images data of oblique aerial photography acquisition, empty three encryptions, described multi-view images data include metadata, include image resolution, image projecting coordinate;

2) on the vertical image in multi-view images data in step 1, coordinate grid is set up in the actual demand become more meticulous according to cybercity construction, in coordinate grid, the size of each grid unit is not less than image resolution, and described image resolution is identical with the image resolution of step 1;

3) multi-view images is converted to gray level image, and by gaussian filtering to the smoothing process of gray level image;

4) in gray level image, calculate the gray-scale value of each target pixel points and the gray-scale value of its neighborhood territory pixel point, the gray-scale value of the gray-scale value of each target pixel points and its neighborhood territory pixel point is compared formation difference, multiple difference forms a numerical range, the size of threshold values A is determined in this numerical range, then judge whether the difference of the gray-scale value of each target pixel points and the gray-scale value of its neighborhood territory pixel point is greater than threshold values A, if, then this target pixel points is extracted alternatively angle point, otherwise, directly cast out;

5) angle point response function (CRF) value of candidate angular in calculation procedure 4, judge whether angle point response function (CRF) value of this candidate angular is greater than the threshold values B preset, if so, then this candidate angular is extracted as unique point, otherwise, directly cast out;

About the determination of threshold values B, utilize Harris angle point response function (CRF) to determine threshold value B in gray level image, specifically, target pixel points and its neighborhood territory pixel point grey scale change little, Harris angle point response function (CRF) value is the positive number of fractional value, is not angle point; When target pixel points comprises edge, Harris angle point response function (CRF) value is the negative of large numerical value, is not angle point; Target pixel points is large along grey scale change during any direction moving window, Harris angle point response function (CRF) value is the positive number of large numerical value, clearly, wherein threshold value B is greater than the positive number of fractional value to the size discrimination of above three kinds of situation Harris angle point response function (CRF) values;

6) mate same place, select, with reference to image and search image, on search image, to find same place by core line geometry constraint condition with reference to the unique point on image;

7) generate Seed Points, utilize the data in step 1 to carry out initial matching, obtain Seed Points;

8) regional diffusion, i.e. Seed Points diffusion, interpolation generates the volume coordinate of point to be located and normal vector, obtains relatively intensive some cloud, makes it evenly intensively be distributed on regular grid;

9) elimination of rough difference, tightly separates mode and the least square thought by multi-disc forward intersection, rejects some cloud rough error point; And it is complete to judge whether grid unit calculates, and if so, then enters step 10, otherwise, get back to step 8 and proceed regional diffusion;

10) generate point of density cloud, structure tin sets up three-dimensional model.(as shown in Figure 9)

First on the basis of available data, set up spatial grid, Seed Points is spread and is easy to control, then utilize Harris innovatory algorithm to extract as unique point, to Feature Points Matching same place, after obtaining object space Seed Points by every a pair same place, carry out regional diffusion, adopt the thought of least square to carry out elimination of rough difference to the error point after diffusion, after obtaining the higher Seed Points of precision, repeat region diffusion and elimination of rough difference, until some cloud precision is met, be comparatively evenly distributed in spatial grid; The wherein extraction algorithm of unique point, improves the efficiency of feature point extraction in oblique aerial point of density cloud leaching process; After Seed Points diffusion, adopt the thought of least square to carry out iterative elimination of rough difference to the point of density cloud obtained, substantially increase precision and the density of data reduction.

Furthermore, coordinate grid is set up about in step 2, the scope of graticule mesh comprises all coverage of image at object space participating in coupling, the size of grid unit represents the least unit of achievement point of density cloud, utilize regular grid that plane is divided into little bin, make it in the plane along the X, the Y direction regular distribution that are parallel to object space; Wherein, the size of bin is according to mating the precision that will reach or determining according to the resolution generating DSM, and it is not less than the resolution of corresponding image; Meanwhile, the planimetric coordinates at each bin center is fixing, and initialization flat element does not possess height value attribute.The graticule mesh set up will cover whole image, and image divides pixel cell with resolution, and image and graticule mesh set up corresponding relation, and thinner says, basic with corresponding to of grid unit and pixel; In the process of coupling, as shown in Figure 2, abstractively the plane after regular grid dividing elements can be regarded as the DSM of rasterizing, in plane, each flat element regards a picture dot of DSM as.

About choosing candidate angular and unique point, as shown in Figure 3, setting threshold values A is as a S (i, j) parameter of target pixel points and neighborhood territory pixel point difference, by scanning gray level image, if by the difference of the target pixel points that scans and its neighborhood territory pixel point at (-s, s) between, then this target pixel points is gray scale similitude, namely refer to that the difference of the gray-scale value of target pixel points and neighborhood territory pixel point is less than the target pixel points of threshold values A, judge whether this point is candidate angular according to the number of gray scale similitude in eight neighborhood, and record all candidate angular S (i, j), Harris angle point respective function (CRF) is utilized to determine threshold values B suitable in image, by to candidate angular S (i, j) scan, if angle point respective function (CRF) value is higher than threshold values B, then being defined as this candidate angular is unique point, preserve S (i, j).

The larger angle point of general Gauss's window is fewer, in order to understand better, for 9*9 rectangular window, harris algorithm and the algorithm after improving thereof is carried out to the analysis of time complexity.

Former Harris algorithm calculates the angle point response function of flash trimming each pixel out-of-bounds, Harris algorithm adopts the rectangle Gauss window calculated amount of 9*9 as follows: multiplication number of times in autocorrelation matrix M process: (9*9+1) * 3=246 time, addition number of times is: (9*9-1) * 3=240 time, calculation level (i, j) angle point response function multiplication 2 times, addition successively, multiplication 248 times, addition 241 times, autocorrelation matrix determinant det (t) 1 time altogether, diagonal sum trace (M) once.

Be the image of Height*Width for resolution, do not consider the impact on border, the time complexity of algorithm is that once+trace (M) is once for 248* (Height-boundary) * (Width-boundary) secondary multiplication+241* (Height-boundary) * (Width-boundary) sub-addition+1* (Height-boundary) * (Width-boundary)+det (M), because the time of a multiplication will far more than a sub-addition, the multiplying amount that Harris algorithm relates to is beaten, and it is slow that speed just shows.

The optimized algorithm that the application proposes mainly had done a candidate angular and has selected before calculating angle point response function, and what mainly do is additive operation, and operand is that 8 × (Height-boundary) * (Width-boundary) are secondary.Here due to increase is not multiplying, although add operand, but for follow-up Harris algorithm eliminates the pixel of most of non-angle point, generally can get rid of pixel ratio and account for more than 50%, make the CRF value not needing to calculate each pixel in subsequent algorithm, in the advantage ensureing original method, improve the efficiency of Corner Detection, save the time.

In step 6, adopt vertical image with reference to image, search image adopts the image on same course line, camera of the same name, if cannot obtain the image on same course line, camera of the same name, searches for image and adopts with attitude image.

To any one unique point on an image, its possible same place of acquisition on another image (refers to the point of position imaging on another image that actual geographic coordinate is identical, below be referred to as to search for image and candidate point) process, this process is exactly homotopy mapping process, usually some known conditions are utilized to reduce matching range a little, as the constraint of core line geometry.As shown in Figure 4, according to corresponding image rays to intersect principle, suppose that S1 and S2 is respectively the camera site of image I1 and I2, p1 is a certain unique point on image I1, its same place on I2 is p2, then corresponding image rays S1p1, S2p2 intersects at spatial point P, obvious S1p1, S2p2 and S1S2 tri-light are coplanar, and the same place p2 of p1 must drop on the intersection of this plane and image I2, and this intersection is called the core line of a p1 on image I2, utilize this core line can retrain the hunting zone of candidate point, core line geometry constraint that Here it is.Geometric relationship due to core line determines same place must be positioned at characteristic on corresponding epipolar line, so just can become the one-dimensional correlation problem along corresponding epipolar line search same place by robotization along the two-dimensional correlation problem of x, y direction search same place, thus decrease workload.

About step 7, as shown in Figure 5,6, after search image obtains the same place with reference to image, utilize the result of empty three encryptions, determine the inside and outside azimuth information of same place, to be crossed according to geometry by two same places and obtain object space point, then each object space point is mapped in corresponding grid according to its volume coordinate, different elevation informations is represented by different colors, be mapped to the initial seed point of the object space point in grid as dense Stereo Matching, and each graticule mesh is just containing a Seed Points.

About the regional diffusion of step 8, as shown in Figure 7, the diffusion of Seed Points is realized by spatial interpolation, utilize regretional analysis, least square method carries out matching to Seed Points, point to be located in window is refined, by the some constructed fuction relational expression of known space of points information, interpolation generates volume coordinate and the normal vector of point to be located.

About step 9 elimination of rough difference, the mode of tightly being separated by multi-disc forward intersection, the least square thought are to reject some cloud rough error points;

Concrete mathematical model:

$\left\{\begin{array}{c}(x-x_0)\[a_3(X-X_s)+b_3(Y-Y_s)+c_3(Z-Z_s)\]=\\ -f\[a_1(X-X_s)+b_1(Y-Y_s)+c_1(Z-Z_s)\]\\ (y-y_0)\[a_3(X-X_s)+b_3(Y-Y_s)+c_3(Z-Z_s)\]=\\ -f\[a_2(X-X_s)+b_2(Y-Y_s)+c_2(Z-Z_s)\]\end{array}\right.$

Arrangement can obtain:

$\left\{\begin{array}{c}{l}_{1}X+l_{2}Y+l_{3}Z-l_x=0\\ l_{4}X+l_{5}Y+l_{6}Z-l_y=0\end{array}\right.$

Wherein:

l ₁＝fa ₁+(x-x ₀)a ₃

l ₂＝fb ₁+(x-x ₀)b ₃

l ₁＝fc ₁+(x-x ₀) _c3

l _x＝fa ₁X _s+fb ₁Y _s+fc ₁Z _s+(x-x ₀)a ₃X _s+(x-x ₀)b ₃Y _s+(x-x ₀)c ₃Z _s

l ₄＝fa ₂+(y-y ₀)a ₃

l ₅＝fb ₃+(y-y ₀)b ₃

l ₆＝fc ₂+(y-y ₀)c ₃

l _y＝fa ₂X _s+fb ₂Y _s+fc ₂Z _s+(y-y ₀)a ₃X _s+(y-y ₀)b ₃Y _s+(y-y ₀)c ₃Z _s

To with reference to a pair same place on video and search image, 4 above-mentioned linear equations can be listed, and unknown number number is 3, therefore can solve by least square method, if containing same object space point in n width image, then can ask X by 2n linear equation solution altogether, Y, Z3 unknown number.

This is a kind of space intersection method strictly, retrained by image number, owing to being solve linear equations, therefore does not also need the initial value of volume coordinate; In order to improve the precision of a cloud further, the constraint of the geometric relationships such as the atural object such as buildings, zebra stripes straight line, right angle can be utilized, as shown in Figure 8, further excluding gross error point.

Also it should be noted that, all within spirit of the present invention and principle, any amendment made, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (8)

1., towards a rapid extracting method for the point of density cloud of cybercity construction, it is characterized in that specifically comprising the following steps:

1) data encasement, data specifically comprise image elements of exterior orientation, inclination aerial surveying camera camera parameter information after the multi-view images data of oblique aerial photography acquisition, empty three encryptions, described multi-view images data include metadata, include image resolution, image projecting coordinate;

2) on the vertical image in multi-view images data in step 1, coordinate grid is set up in the actual demand become more meticulous according to cybercity construction, in coordinate grid, the size of each grid unit is not less than image resolution, and described image resolution is identical with the image resolution of step 1;

3) multi-view images is converted to gray level image, and by gaussian filtering to the smoothing process of gray level image;

4) in gray level image, calculate the gray-scale value of each target pixel points and the gray-scale value of its neighborhood territory pixel point, the gray-scale value of the gray-scale value of each target pixel points and its neighborhood territory pixel point is compared formation difference, multiple difference forms a numerical range, the size of threshold values A is determined in this numerical range, then judge whether the difference of the gray-scale value of each target pixel points and the gray-scale value of its neighborhood territory pixel point is greater than threshold values A, if, then this target pixel points is extracted alternatively angle point, otherwise, directly cast out;

5) angle point response function (CRF) value of candidate angular in calculation procedure 4, judge whether angle point response function (CRF) value of this candidate angular is greater than the threshold values B preset, if so, then this candidate angular is extracted as unique point, otherwise, directly cast out;

6) mate same place, select, with reference to image and search image, on search image, to find same place by core line geometry constraint condition with reference to the unique point on image;

7) generate Seed Points, utilize the data in step 1 to carry out initial matching, obtain Seed Points;

8) regional diffusion, i.e. Seed Points diffusion, interpolation generates the volume coordinate of point to be located and normal vector, obtains relatively intensive some cloud, makes it evenly intensively be distributed on regular grid;

9) elimination of rough difference, tightly separates mode and the least square thought by multi-disc forward intersection, rejects some cloud rough error point; And it is complete to judge whether grid unit calculates, and if so, then enters step 10, otherwise, get back to step 8 and proceed regional diffusion;

10) generate point of density cloud, structure tin sets up three-dimensional model.

2. the rapid extracting method of a kind of point of density cloud towards cybercity construction according to claim 1, it is characterized in that: set up coordinate grid about in step 2, the scope of graticule mesh comprises all coverage of image at object space participating in coupling, the size of grid unit represents the least unit of achievement point of density cloud, utilize regular grid that plane is divided into little bin, make it in the plane along the X, the Y direction regular distribution that are parallel to object space; Wherein, the size of bin is according to mating the precision that will reach or determining according to the resolution generating DSM, and it is not less than the resolution of corresponding image; Meanwhile, the planimetric coordinates at each bin center is fixing, and initialization flat element does not possess height value attribute.

3. the rapid extracting method of a kind of point of density cloud towards cybercity construction according to claim 1, it is characterized in that: choose candidate angular about step 4, setting threshold values A is as a S (i, j) parameter of target pixel points and neighborhood territory pixel point difference, by scanning gray level image, if by the difference of the target pixel points that scans and its neighborhood territory pixel point at (-s, s) between, then this target pixel points is gray scale similitude, namely refer to that the difference of the gray-scale value of target pixel points and neighborhood territory pixel point is less than the target pixel points of threshold values A, judge whether this point is candidate angular according to the number of gray scale similitude in eight neighborhood, and record all candidate angular S (i, j).

4. the rapid extracting method of a kind of point of density cloud towards cybercity construction according to claim 1, it is characterized in that: about the determination of threshold values B in step 5, Harris angle point response function (CRF) is utilized to determine threshold value B in gray level image, specifically, target pixel points and its neighborhood territory pixel point grey scale change little, Harris angle point response function (CRF) value is the positive number of fractional value, is not angle point; When target pixel points comprises edge, Harris angle point response function (CRF) value is the negative of large numerical value, is not angle point; Target pixel points is large along grey scale change during any direction moving window, Harris angle point response function (CRF) value is the positive number of large numerical value, clearly, wherein threshold value B is greater than the positive number of fractional value to the size discrimination of above three kinds of situation Harris angle point response function (CRF) values.

5. the rapid extracting method of a kind of point of density cloud towards cybercity construction according to claim 1, it is characterized in that: in step 6, vertical image is adopted with reference to image, search image adopts the image on same course line, camera of the same name, if cannot obtain the image on same course line, camera of the same name, searches for image and adopts with attitude image.

6. the rapid extracting method of a kind of point of density cloud towards cybercity construction according to claim 1, it is characterized in that: about step 7, after search image obtains the same place with reference to image, utilize the result of empty three encryptions, determine the inside and outside azimuth information of same place, to be crossed according to geometry by two same places and obtain object space point, then each object space point is mapped in corresponding grid according to its volume coordinate, different elevation informations is represented by different colors, be mapped to the initial seed point of the object space point in grid as dense Stereo Matching, and each graticule mesh is not just containing a Seed Points.

7. the rapid extracting method of a kind of point of density cloud towards cybercity construction according to claim 1, it is characterized in that: about the regional diffusion of step 8, the diffusion of Seed Points is realized by spatial interpolation, utilize regretional analysis, least square method carries out matching to Seed Points, point to be located in window is refined, by the some constructed fuction relational expression of known space of points information, interpolation generates volume coordinate and the normal vector of point to be located.

8. the rapid extracting method of a kind of point of density cloud towards cybercity construction according to claim 1, is characterized in that: about step 9 elimination of rough difference, and the mode of tightly being separated by multi-disc forward intersection, the least square thought are to reject some cloud rough error points;

Concrete mathematical model:

$\left\{\begin{array}{c}(x-x_0)\[a_3(X-X_s)+b_3(Y-Y_s)+c_3(Z-Z_s)\]=\\ -f\[a_1(X-X_s)+b_1(Y-Y_s)+c_1(Z-Z_s)\]\\ (y-y_0)\[a_3(X-X_s)+b_3(Y-Y_s)+c_3(Z-Z_s)\]=\\ -f\[a_2(X-X_s)+b_2(Y-Y_s)+c_2(Z-Z_s)\]\end{array}\right.$

Arrangement can obtain:

$\left\{\begin{array}{c}{l}_{1}X+l_{2}Y+l_{3}Z-l_x=0\\ l_{4}X+l_{5}Y+l_{6}Z-l_y=0\end{array}\right.$

Wherein:

l ₁＝fa ₁+(x-x ₀)a ₃

l ₂＝fb ₁+(x-x ₀)b ₃

l ₁＝fc ₁+(x-x ₀)c ₃

l _x＝fa ₁X _s+fb ₁Y _s+fc ₁Z _s+(x-x ₀)a ₃X _s+(x-x ₀)b ₃Y _s+(x-x ₀)c ₃Z _s

l ₄＝fa ₂+(y-y ₀)a ₃

l ₅＝fb ₃+(y-y ₀)b ₃

l ₆＝fc ₂+(y-y ₀)c ₃

l _y＝fa ₂X _s+fb ₂Y _s+fc ₂Z _s+(y-y ₀)a ₃X _s+(y-y ₀)b ₃Y _s+(y-y ₀)c ₃Z _s

To with reference to a pair same place on video and search image, 4 above-mentioned linear equations can be listed, and unknown number number is 3, therefore can solve by least square method, if containing same object space point in n width image, then can ask X by 2n linear equation solution altogether, Y, Z3 unknown number.