CN106846290B - Stereoscopic parallax optimization method based on anti-texture cross and weight cross - Google Patents
Stereoscopic parallax optimization method based on anti-texture cross and weight cross Download PDFInfo
- Publication number
- CN106846290B CN106846290B CN201710038657.0A CN201710038657A CN106846290B CN 106846290 B CN106846290 B CN 106846290B CN 201710038657 A CN201710038657 A CN 201710038657A CN 106846290 B CN106846290 B CN 106846290B
- Authority
- CN
- China
- Prior art keywords
- parallax
- point
- cross
- information
- disparity map
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 57
- 238000005457 optimization Methods 0.000 title claims abstract description 20
- 238000001514 detection method Methods 0.000 claims abstract description 5
- 239000000284 extract Substances 0.000 claims abstract description 5
- 238000004422 calculation algorithm Methods 0.000 claims description 18
- 238000006116 polymerization reaction Methods 0.000 claims description 7
- 238000005314 correlation function Methods 0.000 claims description 2
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims 1
- 239000007787 solid Substances 0.000 abstract description 2
- 238000002474 experimental method Methods 0.000 description 7
- 230000008569 process Effects 0.000 description 6
- 230000011218 segmentation Effects 0.000 description 6
- 238000010586 diagram Methods 0.000 description 5
- 230000008859 change Effects 0.000 description 3
- 230000002776 aggregation Effects 0.000 description 2
- 238000004220 aggregation Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 238000012804 iterative process Methods 0.000 description 2
- 238000003909 pattern recognition Methods 0.000 description 2
- YMHOBZXQZVXHBM-UHFFFAOYSA-N 2,5-dimethoxy-4-bromophenethylamine Chemical compound COC1=CC(CCN)=C(OC)C=C1Br YMHOBZXQZVXHBM-UHFFFAOYSA-N 0.000 description 1
- 241001269238 Data Species 0.000 description 1
- 241000545067 Venus Species 0.000 description 1
- 230000004888 barrier function Effects 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/50—Image enhancement or restoration using two or more images, e.g. averaging or subtraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
- G06T2207/10012—Stereo images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The present invention discloses a kind of stereoscopic parallax optimization method based on anti-texture cross and weight cross.Mainly solve the problems, such as the inaccuracy of disparity map obtained by prior art parallax optimization method.Implementation step are as follows: 1) erroneous point detection is carried out to former horizontal parallax figure, obtain wrong point set;2) original color image is pre-processed, extracts its structural information;3) to each erroneous point, its support area is constructed using structural information;4) to the point of each support area, its parallax cost is calculated;5) support area is utilized, by calculating consecutive points weight, constructs weight cross;6) to each layer of optional parallax, it is polymerize by the parallax cost that weight cross carries out two step information transmitting up and down;7) it polymerize gross information content using parallax, selects the best parallax of erroneous point;8) (3)-(7) are repeated, until all erroneous points are updated.We are bright rapidly and accurately to carry out parallax optimization to original disparity map, can be used for binocular solid matching.
Description
Technical field
The invention belongs to computer vision fields, and in particular to a kind of stereoscopic parallax optimization method, can be used for it is unmanned,
3D tracking and intelligent robot system.
Background technique
Stereo matching is always one important research topic of computer vision field.It can matched color image from two width
Middle generation disparity map is widely applied very much in many fields, and such as unmanned, 3D is rebuild, 3D is tracked.
Existing Stereo Matching Algorithm is divided into two classes: Global Algorithm and local algorithm by Scharstein et al..The overall situation is calculated
Although method accuracy rate is good, computational complexity is very high.Local algorithm is high-efficient, but since real-time demand is extensive, becomes current
Main study subject.Local algorithm is generally divided into four steps: matching cost calculating, cost polymerization, disparity computation and parallax
Optimization.
In recent years, many local algorithms are suggested, but all local algorithms require by building support area come
Find the similitude of pixel.Document (K.Zhang et al., " Cross-based local stereo matching
using orthogonal integral images.”IEEE Trans.on Circuits and Systems for
Video Technology, 2009,19 (7): 1073-1079) propose a kind of sectional perspective matching algorithm based on cross.The calculation
Method can construct an accurate support area.Plum et al. is (referring to " On building an accurate stereo
matching system on graphics hardware.”IEEE Int’l Conf.on Computer Vision
Workshops.pp.467-474,2011) by the termination condition of change cross, improve the algorithm.But both algorithms are equal
It is the interference by texture information in color image based on color similarity and distance.Though parallax value is in image texture region nothing
Variation, but the depth that these regions would generally change, influence the extension of cross.And correct pixel in support area
Number is also fewer.More importantly it is that both methods can not identify object boundary, leads to obscurity boundary in final parallax.
Recently, document (Q.Yang, " A non-local cost aggregation method for stereo
matching,”Int’l.Conf.Computer Vision and Pattern Recognition,pp.1402-1409,
2012) a kind of polymerization based on tree construction is proposed.In the method, piece image is counted as the undirected of one four connection
Figure.A node in each point corresponding diagram in image.The similar adjacent node of every two is connected by a side.While being sorted simultaneously
And it is used to select the weight based in minimum spanning tree.In cost polymerization stage, information passes to adjacent node from a node.
After this, document X.Mei et al., " Segment-tree based cost aggregation for stereo
matching,”IEEE Int’l.Conf.Computer Vision and Pattern Recognition,pp.313-320,
2013 have also been proposed a kind of segmentation tree method, for improving the structure of tree.If this method is divided into a minimum spanning tree
Dry tree is as support area.But the also interference by texture region in image of both algorithms based on tree construction, figure
Although texture region as in will not influence parallax, but in building minimum spanning tree, and it is very big to will lead to consecutive points weight, most
Therefore the structure of small spanning tree changes, be blocked so as to cause information in texture region transmitting, cause result inaccurate.
Parallax optimization, is a step important in stereo algorithm, is more and more paid attention to.First three step of Stereo matching can
To obtain former disparity map.But due to there are occlusion area, texture region and highlight area, being deposited in former disparity map in actual measurement scene
In many erroneous points.Parallax optimization exactly detects and updates the process of these erroneous points.Season et al. one is proposed effectively to be based on
Stablize the parallax optimization algorithm of tree (referring to Y.Ji et al., " Disparity Refinement with Stability-
based Tree for Stereo Matching,”IEEE Intelligent Vehicles Symposium,pp.469-
474,2015), but the algorithm is based on minimum spanning tree, and to will lead to consecutive points weight very big for the interference of texture information, therefore meeting
Change the structure of minimum spanning tree, is blocked so as to cause information in texture region transmitting, causes result inaccurate.
Summary of the invention
It is an object of the invention to propose a kind of stereoscopic parallax optimization method based on anti-texture cross and weight cross,
To reduce influence of the interference of texture to support area and conventional tree structure, the accuracy rate of disparity map is improved.
To achieve the goals above, technical solution of the present invention includes the following steps:
(1) to former left disparity map dLWith former right disparity map dRErroneous point detection is carried out, former left disparity map d is obtainedLMiddle erroneous point
Set E;
(2) original color image I is pre-processed, that is, filters out the texture information in original color image I, retain structure
Information S;
(3) support area R is constructed:
The wrong point set E that (3a) utilizes step (1) to obtain, chooses an erroneous point p ∈ E, in structural information S, into
Row to left and right, above and below four direction anti-texture cross extend, that is, determine next point whether meet color similarity and away from
From similitude, if satisfied, then continuing to extend, otherwise, with vector, i.e. L continuity point behind the position is explored, point place is explored
Region, if the point stops extending in structure boundary, if continuing to extend on texture region;
(3b) obtains the orthogonal cross left arm L of erroneous point p by the way that the cross of four direction extends to left and right, above and below
(p), right arm R (p), upper arm U (p) and lower arm D (p);
(3c) is extended through both direction to the left and right to each point q belonged on upper arm U (p) and lower arm D (p), finds it
Horizontal arm H (q), and then obtain entire support area R;
(4) to the every bit t in the R of support area, in each layer of optional parallax d, its parallax cost C is calculatedd(t):
Wherein, dLIt (t) is parallax value of the point t in former left disparity map;D is optional parallax, is taken 0 between max=64
Integer, then every bit t can obtain max+1 layers of parallax cost Cd(t);
(5) weight cross T is constructed:
(5a) for support area R, according to elder generation left arm L (p), right arm R (p), upper arm U (p), lower arm D (p) in step (3),
The sequence of horizontal arm H (q) again calculates weight W between the consecutive points m and n of every armmn:
Wherein, Δ Ic=| Ic(m)-Ic(n) |, be structural information S in adjacent two o'clock m and n depth interval, Δ d=| dL
(m)-dL(n) |, for the parallax interval of adjacent two o'clock m and n in former left disparity map, c indicates R, and one in tri- channels G, B is logical
Road, λ ∈ [0,1], for the weight parameter for adjusting depth information and parallax information, τ3It is the thresholding for controlling parallax interval;
(5b) is by calculating the weight W between consecutive points m and nmn, weight cross T, i.e. a weighted undirected graph are obtained, benefit
With weight Wmn, weight cross T is traversed by Freud's algorithm, obtains the distance D of shortest path between any two nodes x and y
(x,y);
(6) parallax cost polymerize:
(6a) calculates the upward information content of each node m using weight cross T for each layer of optional parallax dWith downward information content
Wherein, F (n) indicates that the father node of n, F (m) indicate the father node of m, Cd(m) the parallax cost of expression point m, S (m,
N) two o'clock m, the similitude between n are indicated;
The each layer of (6b) for optional parallax d, the parallax polymerization gross information content of each node mAre as follows:
(7) for erroneous point p, it polymerize gross information content using the parallax of its each optional parallax d of layerSelection is best
Parallax value doptimal(p):
The parallax value of erroneous point p is updated to doptimal(p);
(8) step (3)-(7) are repeated, until all erroneous point p ∈ E are updated.
The present invention compared with prior art, has the following advantages:
(1) present invention is due to filtering out texture information therein, then anti-using the structural information in first extraction color image
In the building process of texture cross, using vector is explored, judge that cross extends to is texture region or structure boundary, so that
There are good anti-texture features in the support area of generation.
(2) present invention optimizes disparity map using there is the update method of weight cross, compared to the side of original minimum spanning tree
Method, weight cross structure are stablized, and not by the interference of texture information, allow parallax information and depth information well in weight
It is transmitted in cross.
(3) present invention only need to consider closer point similar from erroneous point in erroneous point update, and from erroneous point not phase
As farther away point, be not required to consider, so that computation complexity greatly reduces, operation time is shorter.
Detailed description of the invention
Fig. 1 is implementation flow chart of the invention;
Fig. 2 is the wrong point diagram that former horizontal parallax figure is generated in the present invention;
Fig. 3 be the present invention in original color image pretreatment after with the comparison result figure before pretreatment;
Fig. 4 is that the anti-texture cross in the present invention supports administrative division map;
Fig. 5 is the weight cross figure in the present invention;
Fig. 6 is the support area figure comparison diagram generated with the present invention and existing method;
Fig. 7 is the parallax optimization comparison diagram generated with the present invention and existing method.
Specific embodiment
The invention will be further described with table with reference to the accompanying drawing:
Parallax optimization, is the left and right two images for obtaining two horizontal positioned cameras, by Stereo Matching Algorithm
In row filtering processing, the left disparity map of not optimized original and former right disparity map is calculated, then the original disparity map is carried out
Parallax optimization, updates its erroneous point, obtains more accurate disparity map.
Referring to Fig.1, steps are as follows for realization of the invention:
The detection of step 1 erroneous point.
(1.1) former left disparity map d is utilizedLWith former right disparity map dR, erroneous point inspection is carried out by left and right method for detecting continuity
It surveys, i.e., for former left disparity map dLIn point p, determine whether to meet following formula:
|dL(p)-dR(p-dL(p)) | >=1,
If it is satisfied, then p is erroneous point, conversely, p is correct point;
(1.2) to former left disparity map dLIn all the points repeat step (1.1) operation, generation error point set E, such as Fig. 2
Shown, wherein Fig. 2 (a) is former left disparity map, and Fig. 2 (b) is former right disparity map, and Fig. 2 (c) is wrong point diagram, black in Fig. 2 (c)
Color dot is erroneous point.
Step 2 filters out the texture information in original color image I.
(2.1) original color image I is filtered using opposite total deviation method, extracts its structural information and obtains the first knot
Structure layer S1;
(2.2) first structure layer S is removed from original color image I1, obtain the first levels of detail D1=I-S1, continue to use phase
To total deviation method to the first levels of detail D1It is filtered, extracts D1Structural information obtain the second structure sheaf S2;
(2.3) according to first structure layer S1With the second structure sheaf S2, calculate the structural information S of original color image:
S=λ1×S1+λ2×S2,
Wherein, λ1For the first layer coefficients, λ2For the second layer coefficients;
The result of this step is as shown in Figure 3, wherein Fig. 3 (a) is original color image, and Fig. 3 (b) is resulting structures information
S, Fig. 3 (c) are one section of original color image, and Fig. 3 (d) is one section of resulting structures information S.
From figure 3, it can be seen that being effectively maintained by image preprocessing structure boundary, such as 2, circle in Fig. 3 c)
Point, Fig. 3 d) in 4 part of circle, and grain details are filtered out, such as 1 part of circle in Fig. 3 c), Fig. 3 d) in 3 part of circle.
Step 3, building support area R.
(3.1) color similarity is defined:
If two pixel m and n meet in structural information S: max (| Ic(m)-Ic(n)|C=R, G, B)<τ1, then the two
Pixel meets color similarity, otherwise, is unsatisfactory for color similarity, wherein Ic() is depth information in structural information S, c
For R, a channel in tri- channels G, B, τ1For color thresholding, τ1=20;
(3.2) distance similarity is defined:
If two pixel m and n meet in structural information S: Ω (m, n) < τ2, then the two pixels meet apart from phase
Like property, otherwise, it is unsatisfactory for distance similarity, wherein Euclidean distance of the Ω (m, n) between m and n, τ2For distance threshold,
τ2=25;
(3.3) the wrong point set E obtained using step (1), chooses an erroneous point p ∈ E, in structural information S, into
The anti-texture cross of four direction extends row to the left, to the right, upwards, downwards, that is, it is similar to determine whether next point meets color
Property and distance similarity, if satisfied, then continue to extend, otherwise, with exploring vector, i.e. L=5 continuity point behind the position, spy
Region where L continuity point of rope, if L continuity point in structure boundary, stops extending, if its on texture region,
Continue to extend;
(3.4) by the way that the cross of four direction extends to the left, to the right, upwards, downwards, the orthogonal cross of erroneous point p is obtained
Left arm L (p), right arm R (p), upper arm U (p) and lower arm D (p);
(3.5) to each point q belonged on upper arm U (p) and lower arm D (p), by the way that both direction extends to the left and right respectively,
Its horizontal arm H (q) is found, and then obtains entire support area R, as shown in Figure 4.
In Fig. 4, the point of grid filling is left arm, the right arm, upper and lower arms of p, and shadow spots are the horizontal arm of q, black color dots
For the point with p dissmilarity, there is the continuity point of white frame to explore vector.
Step 4 is to the every bit t in the R of support area, in each layer of optional parallax d, calculates its parallax cost Cd(t):
Wherein, dLIt (t) is parallax value of the point t in former left disparity map;D is optional parallax, is taken 0 between max=64
Integer, then every bit t can obtain max+1 layers of parallax cost Cd(t)。
Step 5 constructs weight cross T.
(5.1) for support area R, according to elder generation left arm L (p), right arm R (p), upper arm U (p), lower arm D in step (3)
(p), then the sequence of horizontal arm H (q) calculates weight W between the consecutive points m and n of every armmn:
Wherein, Δ Ic=| Ic(m)-Ic(n) | for the depth interval of adjacent two o'clock m and n in structural information S, Δ d=| dL
(m)-dL(n) | for the parallax interval of adjacent two o'clock m and n in former left disparity map, c indicates R, and one in tri- channels G, B is logical
Road, λ=0.8, for the weight parameter for adjusting depth information and parallax information, τ3It is the thresholding for controlling parallax interval, value τ3=
10;
(5.2) pixel of support area R is created as the point set V of figure, by the consecutive points m of arm every in the R of support area
It is attached between n, establishes the side collection Z of figure, and be assigned to weight W on side to itmn, correlation function K is established, it is undirected to generate weighting
Scheme (V, Z, K), such as Fig. 5, as weight cross T;
In weight cross T, central point is erroneous point p, child node, the i.e. node adjacent with the node and is had from central point
Biggish distance, father node, the i.e. node adjacent with the node and has with a distance from lesser from central point;Ten word terminals, that is, do not have
All child nodes of the node of child node, sub- cross, the i.e. node and the node until ten word terminals;
(5.3) weight W is utilizedmn, weight cross T is traversed by Freud's algorithm, is searched between any two nodes x and y
The distance D (x, y) of shortest path.
Step 6, polymerization parallax cost.
(6.1) the similitude S (m, n) in weight cross T between any two points m, n is calculated:
Wherein, σ=0.1, for the parameter for adjustable range, D (m, n) indicates two o'clock m, the distance of shortest path between n;
(6.2) the upward information transmitting of weight cross T:
Upward information transmitting is the process that information is transmitted to central point from ten word terminals, for each of optional parallax d
Layer, using weight cross T, calculates the upward information content of each node m
Wherein, F (n) represents the father node of n, Cd(m) the parallax cost of point m is indicated, between S (m, n) expression two o'clock m, n
Similitude.For the point on ten word terminals,The process is an iterative process, whole from first access cross
End starts iteration, until central node is accessed;
(6.3) the downward information transmitting of weight cross T:
Downward information transmitting is the process that information is transmitted to ten word terminals from central point, for each of optional parallax d
Layer, utilizes upward information contentCalculate the downward information content of each node m
For root nodeThe process is an iterative process, is changed since first accessing central point
Generation, until accessing ten word terminals.
For each layer of optional parallax d, according to downward information contentThe parallax polymerization for obtaining each node m is total
Information contentAre as follows:
Step 7, for erroneous point p, gross information content is polymerize according to the parallax of its each optional parallax d of layerSelection
Best parallax value:The parallax value of erroneous point p is updated to dopt(p)。
Step 8, step 3-7 is repeated, until all erroneous point p ∈ E are updated.
Effect of the invention is described further below with reference to experiment.
1. simulation parameter
2005 data sets are concentrated using middleburry normal data in this experiment, are in binocular solid matching
One of most common data set can be measured in disparity map, the error rate of whole pixels and unshielding part, including cones
Four groups of (375*450), teddy (375*450), venus (383*434), tuskuba (288*384) image datas.
2. emulation experiment content
The present invention and the method based on cross are respectively adopted in this experiment, segmentation tree method carries out parallax with tree method is stablized
The effect for comparing these four methods is judged in optimization by the vision of error rate and disparity map in disparity map.
Emulation experiment 1
Take first layer coefficient lambda1=0.8, second layer coefficient lambda2=0.1, the present invention is respectively adopted and is generated based on cross method
Support area, as a result such as Fig. 6, wherein Fig. 6 (a) is using the support generated based on cross method to the data table angle part cones
Administrative division map, Fig. 6 (b) are the support area figure generated using the present invention to the data table angle part cones, and Fig. 6 (c) is that use is based on
The support area figure that cross method generates cones data barrier portion, Fig. 6 (d) are using the present invention to cones data fence
The support area figure that part generates.
From fig. 6, it can be seen that based on the method for cross due to the interference by texture region in image, the support area of generation
Domain is often smaller, and in irregular shape, and the present invention has good anti-texture features, can preferably find the side of object
The support area on boundary, generation is more satisfactory.
Emulation experiment 2
The present invention and existing segmentation tree method is respectively adopted and stablizes tree method, parallax optimization is carried out to former disparity map, leads to
Cross the result for comparing these three methods using the error rate of disparity map.Wherein, the error rate calculation formula of disparity map are as follows:In formula, N is pixel sum, dcIt is the parallax that parallax optimization method generates disparity map
Value, dTIt is the parallax value of standard disparity map, δdIt is fault-tolerant thresholding.
The following Tables 1 and 2 of experimental result:
1 present invention of table and existing method are in fault-tolerant thresholding δdError rate compares when=1
2 present invention of table and existing method are in fault-tolerant thresholding δdError rate compares when=2
In Tables 1 and 2, " unshielding " is the error rate of de-occlusion region in disparity map, and " all " is complete in disparity map
The error rate in portion region, in fault-tolerant thresholding δdWhen=1, the present invention has 4 groups of results best in 8 groups of experimental results, in fault-tolerant thresholding δd
When=2, the present invention has 5 groups of results best in 8 groups of experimental results, it is seen that the present invention has preferable accuracy rate.
Emulation experiment 3
The present invention and existing segmentation tree method is respectively adopted and stablizes tree method, parallax optimization, knot are carried out to former disparity map
Fruit such as Fig. 7, wherein Fig. 7 (a) is the standard disparity map of cones data, and Fig. 7 (b) is to be generated with the present invention to cones data
Disparity map, Fig. 7 (c) are with the disparity map that generates to cones data of segmentation tree method, and Fig. 7 (d) is with stablizing tree method pair
The disparity map that cones data generate, Fig. 7 (e) are the standard disparity map of teddy data, and Fig. 7 (f) is with the present invention to teddy number
According to the disparity map of generation, Fig. 7 (g) is the disparity map generated with segmentation tree method to teddy data, and Fig. 7 (h) is with stable tree side
The disparity map that method generates teddy data.
As seen from Figure 7, the disparity map that the present invention generates has preferably accuracy rate, and it is more accurate that erroneous point updates, object
Sharpness of border.
It, can be in conclusion the stereoscopic parallax optimization method proposed by the present invention based on anti-texture cross and weight cross
Reduce influence of the interference of texture to support area and conventional tree structure, preferably retains the boundary of object, improve the standard of disparity map
True rate can rapidly and accurately optimize original disparity map parallax.
Claims (7)
1. a kind of stereoscopic parallax optimization method based on anti-texture cross and weight cross, comprising:
(1) to former left disparity map dLWith former right disparity map dRErroneous point detection is carried out, former left disparity map d is obtainedLMiddle mistake point set
E;
(2) original color image I is pre-processed, that is, filters out the texture information in original color image I, retain structural information
S;
(3) support area R is constructed:
(3a) utilizes step (1) obtained wrong point set E, chooses an erroneous point p ∈ E, in structural information S, carry out to
The anti-texture cross of left and right, upper and lower four direction extends, that is, determines whether next point meets color similarity and apart from phase
Like property, if satisfied, then continuing to extend, otherwise, with vector, i.e. L continuity point behind the position is explored, the area where the point is explored
Domain, if the point stops extending in structure boundary, if continuing to extend on texture region;
(3b) obtains the orthogonal cross left arm L (p) of erroneous point p by the way that the cross of four direction extends to left and right, above and below, right
Arm R (p), upper arm U (p) and lower arm D (p);
(3c) extends each point q belonged on upper arm U (p) and lower arm D (p) through both direction to the left and right, finds its level
Arm H (q), and then obtain entire support area R;
(4) to the every bit t in the R of support area, in each layer of optional parallax d, its parallax cost C is calculatedd(t):
Wherein, dLIt (t) is parallax value of the point t in former left disparity map;D is optional parallax, is taken 0 to the integer between max=64,
Then every bit t can obtain max+1 layers of parallax cost Cd(t);
(5) weight cross T is constructed:
(5a) for support area R, according to elder generation left arm L (p), right arm R (p), upper arm U (p), lower arm D (p) in step (3), then water
The sequence of flat arm H (q) calculates weight W between the consecutive points m and n of every armmn:
Wherein, Δ Ic=| Ic(m)-Ic(n) |, be structural information S in adjacent two o'clock m and n depth interval, Δ d=| dL(m)-dL
(n) |, for the parallax interval of adjacent two o'clock m and n in former left disparity map, c indicates R, a channel in tri- channels G, B, λ ∈
[0,1], for the weight parameter for adjusting depth information and parallax information, τ3It is the thresholding for controlling parallax interval;
(5b) is by calculating the weight W between consecutive points m and nmn, obtain weight cross T, i.e. a weighted undirected graph, exploitation right
Value Wmn, by Freud's algorithm traverse weight cross T, obtain shortest path between any two nodes x and y distance D (x,
y);
(6) parallax cost polymerize:
(6a) calculates the upward information content of each node m using weight cross T for each layer of optional parallax d
With downward information content
Wherein, F (n) indicates that the father node of n, F (m) indicate the father node of m, Cd(m) the parallax cost of point m is indicated, S (m, n) is indicated
Similitude between two o'clock m, n;
The each layer of (6b) for optional parallax d, the parallax polymerization gross information content of each node mAre as follows:
(7) for erroneous point p, it polymerize gross information content using the parallax of its each optional parallax d of layerSelect best parallax
Value dopt(p):
The parallax value of erroneous point p is updated to dopt(p);
(8) step (3)-(7) are repeated, until all erroneous point p ∈ E are updated.
2. the method as described in claim 1, which is characterized in that utilize former left disparity map d in step (1)LWith former right disparity map
dR, erroneous point detection is carried out by left and right method for detecting continuity, i.e., first for former left disparity map dLIn point p, determine whether full
Foot | dL(p)-dR(p-dL(p)) | >=1, if it is satisfied, then p is erroneous point, conversely, p is correct point;Again to former left disparity map dL
In all the points repeat this operation, generation error point set E.
3. the method as described in claim 1, which is characterized in that filter out the texture information in original color image I in step (2)
D2, it carries out as follows:
(2a) is filtered original color image I using opposite total deviation method, extracts structural information and obtains first structure layer
S1;
(2b) removes first structure layer S from original color image I1, obtain the first levels of detail D1=I-S1, continue to use relatively total
Deviation method is to the first levels of detail D1It is filtered, extracts D1Structural information obtain the second structure sheaf S2;
(2c) is according to first structure layer S1With the second structure sheaf S2, calculate original color image structural information S:
S=λ1×S1+λ2×S2,
Wherein, λ1For the first layer coefficients, λ2For the second layer coefficients.
4. the method as described in claim 1, which is characterized in that the color similarity in step (3a), is defined as: if knot
Two pixel m and n meet in structure information S: max (| Ic(m)-Ic(n)|C=R, G, B)<τ1, then the two pixels meet color
Otherwise similitude is unsatisfactory for color similarity, wherein Ic() is depth information in structural information S, and c R, G, B tri- logical
A channel in road, τ1For color thresholding, τ1=20.
5. the method as described in claim 1, which is characterized in that the distance similarity in step (3a), is defined as: if knot
Two pixel m and n meet in structure information S: Ω (m, n) < τ2, then the two pixels meet distance similarity, otherwise, are discontented with
Sufficient distance similarity, wherein Euclidean distance of the Ω (m, n) between m and n, τ2For distance threshold, τ2=25.
6. the method as described in claim 1, which is characterized in that construct weight cross T in step (5b), be support area R
Pixel be created as the point set V of figure, will be attached between the consecutive points m and n of arm every in the R of support area, establish figure
Side collection Z, and weight W on side is assigned to itmn, correlation function K is established, is generated weighted undirected graph (V, Z, K), as weight cross T.
7. the method as described in claim 1, which is characterized in that the similitude S (m, n) in step (6a), is defined as:Wherein, σ=0.1, for the parameter for adjustable range, D (m, n) indicates two o'clock m, between n
The distance of shortest path.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710038657.0A CN106846290B (en) | 2017-01-19 | 2017-01-19 | Stereoscopic parallax optimization method based on anti-texture cross and weight cross |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710038657.0A CN106846290B (en) | 2017-01-19 | 2017-01-19 | Stereoscopic parallax optimization method based on anti-texture cross and weight cross |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106846290A CN106846290A (en) | 2017-06-13 |
CN106846290B true CN106846290B (en) | 2019-10-11 |
Family
ID=59124837
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710038657.0A Active CN106846290B (en) | 2017-01-19 | 2017-01-19 | Stereoscopic parallax optimization method based on anti-texture cross and weight cross |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106846290B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110176037A (en) * | 2019-05-31 | 2019-08-27 | 东北大学 | A kind of target range Method of fast estimating driven for outdoor road auxiliary |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102136136A (en) * | 2011-03-17 | 2011-07-27 | 南京航空航天大学 | Luminosity insensitivity stereo matching method based on self-adapting Census conversion |
CN102930530A (en) * | 2012-09-26 | 2013-02-13 | 苏州工业职业技术学院 | Stereo matching method of double-viewpoint image |
CN104200453A (en) * | 2014-09-15 | 2014-12-10 | 西安电子科技大学 | Parallax image correcting method based on image segmentation and credibility |
CN104867135A (en) * | 2015-05-04 | 2015-08-26 | 中国科学院上海微系统与信息技术研究所 | High-precision stereo matching method based on guiding image guidance |
CN106228605A (en) * | 2016-07-29 | 2016-12-14 | 东南大学 | A kind of Stereo matching three-dimensional rebuilding method based on dynamic programming |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9754377B2 (en) * | 2014-08-15 | 2017-09-05 | Illinois Institute Of Technology | Multi-resolution depth estimation using modified census transform for advanced driver assistance systems |
-
2017
- 2017-01-19 CN CN201710038657.0A patent/CN106846290B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102136136A (en) * | 2011-03-17 | 2011-07-27 | 南京航空航天大学 | Luminosity insensitivity stereo matching method based on self-adapting Census conversion |
CN102930530A (en) * | 2012-09-26 | 2013-02-13 | 苏州工业职业技术学院 | Stereo matching method of double-viewpoint image |
CN104200453A (en) * | 2014-09-15 | 2014-12-10 | 西安电子科技大学 | Parallax image correcting method based on image segmentation and credibility |
CN104867135A (en) * | 2015-05-04 | 2015-08-26 | 中国科学院上海微系统与信息技术研究所 | High-precision stereo matching method based on guiding image guidance |
CN106228605A (en) * | 2016-07-29 | 2016-12-14 | 东南大学 | A kind of Stereo matching three-dimensional rebuilding method based on dynamic programming |
Non-Patent Citations (2)
Title |
---|
Cross-Based Local Stereo Matching Using Orthogonal Integral Images;K. Zhang等;《IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY》;20090731;第1073-1079页 * |
基于改进Census变换和动态规划的立体匹配算法;祝世平 等;《光学学报》;20160430;第36卷(第4期);第1-9页 * |
Also Published As
Publication number | Publication date |
---|---|
CN106846290A (en) | 2017-06-13 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112435325B (en) | VI-SLAM and depth estimation network-based unmanned aerial vehicle scene density reconstruction method | |
CN103226821B (en) | Stereo matching method based on disparity map pixel classification correction optimization | |
CN104867135B (en) | A kind of High Precision Stereo matching process guided based on guide image | |
CN105654548B (en) | A kind of a lot of increment type three-dimensional rebuilding methods based on extensive unordered image | |
CN108596975B (en) | Stereo matching algorithm for weak texture region | |
CN101625768B (en) | Three-dimensional human face reconstruction method based on stereoscopic vision | |
CN109658445A (en) | Network training method, increment build drawing method, localization method, device and equipment | |
CN104820991B (en) | A kind of multiple soft-constraint solid matching method based on cost matrix | |
Deng et al. | Noisy depth maps fusion for multiview stereo via matrix completion | |
Chen et al. | Transforming a 3-d lidar point cloud into a 2-d dense depth map through a parameter self-adaptive framework | |
CN108921864A (en) | A kind of Light stripes center extraction method and device | |
Yang et al. | Non-parametric depth distribution modelling based depth inference for multi-view stereo | |
Hu et al. | Stereo matching using weighted dynamic programming on a single-direction four-connected tree | |
CN104966290B (en) | A kind of adaptive weighting solid matching method based on SIFT description | |
CN104331890B (en) | A kind of global disparity method of estimation and system | |
CN113989758A (en) | Anchor guide 3D target detection method and device for automatic driving | |
Kostková et al. | Stratified Dense Matching for Stereopsis in Complex Scenes. | |
CN115601406A (en) | Local stereo matching method based on fusion cost calculation and weighted guide filtering | |
CN109887008A (en) | Based on front and back to smooth and O (1) complexity parallax solid matching method, device and equipment | |
CN113780389B (en) | Deep learning semi-supervised dense matching method and system based on consistency constraint | |
Yin et al. | A new apple segmentation and recognition method based on modified fuzzy C-means and hough transform | |
CN106846290B (en) | Stereoscopic parallax optimization method based on anti-texture cross and weight cross | |
CN110942102B (en) | Probability relaxation epipolar matching method and system | |
CN104408710B (en) | Global parallax estimation method and system | |
CN107194334B (en) | Video satellite image dense Stereo Matching method and system based on optical flow estimation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |