CN107133977A - A kind of quick stereo matching process that model is produced based on probability - Google Patents
A kind of quick stereo matching process that model is produced based on probability Download PDFInfo
- Publication number
- CN107133977A CN107133977A CN201710351601.0A CN201710351601A CN107133977A CN 107133977 A CN107133977 A CN 107133977A CN 201710351601 A CN201710351601 A CN 201710351601A CN 107133977 A CN107133977 A CN 107133977A
- Authority
- CN
- China
- Prior art keywords
- parallax
- image
- point
- probability
- matching
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
- G06T7/33—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/13—Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/143—Segmentation; Edge detection involving probabilistic approaches, e.g. Markov random field [MRF] modelling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/50—Depth or shape recovery
- G06T7/529—Depth or shape recovery from texture
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Probability & Statistics with Applications (AREA)
- Software Systems (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of quick stereo matching process that model is produced based on probability, including, the feature of each pixel of one group of image captured by extraction binocular camera；Sampled according to fixed intervals and obtain the sampling characteristic point of an image, and match the characteristic point on another image, filter out stable match point；Trigonometric ratio segmentation is carried out to image, the corresponding triangle set of image is obtained；It is determined that the parameter of the parallax and the delta-shaped region disparity plane on each Atria summit；The prior probability of each parallax is obtained using plane parameter, the posterior probability of parallax is calculated with reference to Matching power flow, the initial parallax figure of image is further obtained；Optimize initial parallax figure.The invention also discloses the application of the above method.The method of technical solution of the present invention can obtain the stable matching point of more robust there is provided a kind of matching pixel confidence balancing method of more robust；Stereo matching can also be quickly carried out, improves matching effect.
Description
Technical field
The invention belongs to computer stereo vision field, and in particular to a kind of quick stereo that model is produced based on probability
Method of completing the square.
Background technology
Computer stereo vision is mainly studied how from multiple image, obtains distance (depth) letter of object in scene
Breath.Binocular stereo vision is one of most basic problem in computer stereo vision, for from by two cameras shoot it is same
In the two images of scene, the depth information of scene is recovered.The main flow of binocular stereo vision is：Obtain image, demarcation
Five parts such as video camera, correction chart picture, Stereo matching and three-dimensional reconstruction.Wherein Stereo matching is the core of whole algorithm
Point, the dense disparity map to generate, wherein parallax are the abscissa difference of same place in two figures of left and right.Stereo matching is whole
It is most difficult in individual algorithm also most to influence a part of three-dimensional reconstruction quality.
According to Binocular Vision Principle, if it is possible to two match points are determined in image coordinate system, and know that its is respective
Image coordinate, then be obtained with the depth information of correspondence spatial point.Therefore, the key for realizing Depth Information Acquistion is to obtain
A matching pair of the spatial point in left images plane, and stereo matching problem is then to realize the key that depth is obtained.In reality
In the operation of border, binocular solid matching implementation will consider factors, and with indicators of overall performances such as computation complexity and stability
To weigh the feasibility and validity of scheme implementation.
Current Stereo Matching Algorithm is broadly divided into four steps：Matching power flow calculating, cost polymerization, disparity computation, parallax
Optimization.It is that the optimization related to all point progress in figure is calculated because cost polymerize to disparity computation, therefore the time spends ratio
It is larger.The fast algorithm of current main flow mainly optimizes to the two step progressive, and is generally accelerated using GPU.Another
In a kind of Stereo matching thinking, first find out on a small quantity stable match point, author assumes stable matching point to not yet matching at present
Point has priori directive function, therefore produces model to solve the parallax that they are optimal by probability.This procedure avoids
Energy solution procedure in cost polymerization and disparity computation, therefore algorithm speed is quickly.But, in this algorithm, stable
Selection with point is most important to last result, and the selection mode of author does not ensure that the accuracy of these points.
The content of the invention
For the disadvantages described above or Improvement requirement of prior art, the invention provides a kind of pair that model is produced based on probability
Mesh image depth information Fast Stereo Matching Algorithm.For this problem of recovery depth information from the binocular image corrected
For, the method for technical solution of the present invention can be obtained more there is provided a kind of matching pixel confidence balancing method of more robust
Plus the stable matching point of robust；The method of technical solution of the present invention can also quickly carry out Stereo matching, improve matching effect.
To achieve the above object, according to one aspect of the present invention there is provided it is a kind of based on probability produce model it is quick
Solid matching method, it is characterised in that including,
A kind of quick stereo matching process that model is produced based on probability, it is characterised in that including,
S1 extracts the feature of each pixel of one group of image captured by binocular camera；
S2 obtains sampling characteristic point to an image in one group of image according to fixed intervals sampling, special according to the sampling
The characteristic point on another image in this group of image of Point matching is levied, stable matching point is obtained after screening；
S3 carries out trigonometric ratio segmentation to image respectively using the stable matching point, and the corresponding triangle collection of image is obtained respectively
Close；
S4 utilizes the parallax of stable matching point in described image, it is determined that the parallax on each Atria summit, enters one
Step ground, obtains the parameter for being fitted the delta-shaped region disparity plane；
S5 obtains the prior probability of each parallax using plane parameter, in conjunction with each parallax for the pixel
Matching power flow calculates likelihood probability, the comprehensive posterior probability for obtaining some parallax, further to obtain the initial of whole image
Disparity map；
S6 optimizes to obtain the disparity map of better quality to the initial parallax figure.
In technical solution of the present invention, it is preferred to use be that one group of image that binocular camera is shot carries out depth information meter
Calculate.Binocular camera is usually to be combined by two monocular cameras arranged in parallel, differential seat angle is caused by simulating human eye, with this
Come the effect for reaching three-dimensional imaging or detecting the depth of field.That is, two monocular cameras arranged in parallel can also realize it is double
Effect of mesh camera.The once shooting of binocular camera, two monocular cameras for actually constituting binocular camera are same to target
Shi Jinhang is shot, and one group of image can be once obtained for same target.Because binocular camera simulation is human eye differential seat angle, because
This this group of image has different visual angles.By being matched to one group of image with different visual angles, it can obtain wherein
Depth information.
In technical solution of the present invention, what is extracted in step S1 is that two monocular cameras for constituting binocular camera are shot respectively
The feature of each pixel in picture, by being matched to every group of image characteristic point captured by each camera, can be obtained
The range information of object, i.e. depth information in scene.When being matched, an image is chosen, accordingly, it is at least right
The image from another angle shot should be opened in the presence of one.After the two progress sampling Feature Points Matching, wherein matching error is rejected
Characteristic point, it is possible to obtain stable match point.Multiple match points whole image can be divided into multiple triangle sets into
Set, by calculating these vertex of a triangle parallaxes, the parameter of the delta-shaped region disparity plane can be obtained.
It is possible to further using the plane parameter and Matching power flow of triangle, obtain the prior probability of each parallax, likelihood probability and
Posterior probability, the final initial parallax figure for obtaining whole image.Initial parallax figure is after optimization, as target disparity map.
In technical solution of the present invention, key step has, and extracts image characteristic point, screening and obtains stable match point, to figure
As carrying out trigonometric ratio segmentation, digital simulation disparity plane parameter, model produced based on probability obtain the disparity map of whole image with
And disparity map optimization.
As the optimal technical scheme of the present invention, step S1 is specifically included,
S11 calculates the edge gradient of described image respectively using arithmetic operators；
If S12 chooses the gradient information done in each neighborhood of pixels, the characteristic vector of current point is combined into.
When feature extraction is specifically carried out, the edge gradient of image is calculated respectively using arithmetic operators, and
If choosing the gradient information done in adjacent domain for each pixel, the characteristic vector of current point is formed.That is, current
The characteristic vector of pixel, is the gradient information synthesis group in the corresponding edge gradient value of each pixel and adjacent area
Into.
As the optimal technical scheme of the present invention, step S2 is specifically included,
S21 chooses a sampling characteristic point, calculates its Matching power flow between the match point in certain disparity range；
S22 investigates Matching power flow of the sampling characteristic point under each parallax, finds out its minimum value, sub-minimum and second local
Minimum cost value, to calculate the stability of sampling characteristic point；Retain the sampling characteristic point for meeting stable condition, its parallax value is to take
Obtain the parallax value of minimum cost value；
S23 finds its corresponding match point according to the parallax of sampling characteristic point, investigates matching of the match point under each parallax
Cost, finds out its minimum value, sub-minimum and the second local minimum cost value, to calculate the stability of match point；Retain and meet
The match point of stable condition, its parallax value is the parallax value for obtaining minimum cost value；
If the parallax of the corresponding match point of S24 samplings characteristic point is consistent, it is determined that the sampling characteristic point and its correspondence
Match point be stable matching point.
During stable matching point is confirmed, most critical is exactly to find the minimum parallax value of Matching power flow.In whole
Pixel in, according to it is certain rule choose several sampled points, calculate sampled point Matching power flow.Matching generation is filtered out successively
Minimum value, sub-minimum and the second local minimum of valency, judge whether current sampling point meets corresponding steady using above-mentioned value
Fixed condition, if meeting, retains this sampled point, and confirms the parallax value that the parallax value for causing Matching power flow minimum is the point.
For the sampled point, its match point on another image is confirmed, and judge whether the match point meets stable condition, it is stable
Then retain the point.Calculating obtains sampled point with after the parallax value of corresponding match point, judging whether the two parallax is consistent, if one
Cause, then the two matching relationship of explanation is stable, and the corresponding match point of the sampled point is stable matching point.
As the optimal technical scheme of the present invention, step S5 is specifically included,
The parameter of S51 planes according to where any pixel point, calculating obtains the parallax value of point in the plane；
S52 sets up probability distribution centered on the parallax value, for determining any one parallax value in scope, according to institute
State the prior probability that probability distribution calculates the parallax；
S53 calculates the Matching power flow between the match point in pixel and certain disparity range, further, calculates it seemingly
Right probability；
S54, which is calculated, to be obtained determining the posterior probability in scope under each parallax, chooses the parallax of maximum probability as the point most
Whole parallax；
S55 obtains the parallax of whole pixels in described image respectively, further, and the parallax of described image is obtained respectively
Figure.
As the optimal technical scheme of the present invention, probability distribution is preferably superimposed with one uniformly by a Gaussian Profile
Distribution is formed.
Triangle obtained by splitting for trigonometric ratio, the calculating of its plane parameter needs to utilize Atria summit
Information, including coordinate and parallax, the disparity plane parameter of any Delta Region can be solved according to these information.It is flat according to parallax
Face parameter, can solve the parallax value of a certain pixel in the plane；Probability distribution can be set up centered on the parallax value.
In technical solution of the present invention, probability distribution, which is preferably superimposed with one by a Gaussian Profile and is uniformly distributed, to be formed.It is general using this
Rate is distributed, and for any one parallax value determined in scope, can calculate its probability, that is, for the parallax
Prior probability.Then, in the disparity range of a determination, the Matching power flow between pixel and match point is calculated, it is counted
Calculation method is consistent with the computational methods in above-mentioned steps S2.Calculating is obtained after the Matching power flow between pixel and match point, can
Further to calculate the likelihood probability for obtaining a certain parallax.One pixel can be by elder generation in the disparity range of a determination
Test probability and likelihood probability calculates the corresponding posterior probability of each parallax, choose the parallax of wherein maximum probability as the point most
Whole parallax.Piece image has multiple pixels, and all pixels point is chosen after suitable parallax, can obtained for a certain tool
One disparity map of body image, the disparity map is initial parallax figure.
As the optimal technical scheme of the present invention, step S6 is specifically included,
S61 carries out cross checking to initial parallax figure；
S62 chooses the current parallax of other parallax values replacement of its adjacent domain to the point do not verified by cross；
S63 is filtered to disparity map using wave filter, removes noise spot, obtains final disparity map.
Initial parallax figure needs that final disparity map could be obtained by a series of optimization process.Specifically, it is exactly
Need to verify initial parallax figure, i.e., whether each self-corresponding parallax size of disparity map terminal between determination different images
Meet the condition of cross validation.Meet intersection for by the match point of cross validation condition, not choosing in its adjacent domain and test
Card condition, and the minimum parallax of numerical value in the region, replace current parallax.
It is based on according to another aspect of the present invention there is provided one kind described in a kind of any one of application claim 1~6
The system that probability produces the quick stereo matching process of model, it is characterised in that including,
Image extraction unit, the feature of each pixel of image captured by binocular camera arranged in parallel for extracting；
Characteristic matching unit, for obtaining sampling characteristic point according to fixed intervals sampling to wherein one image, according to institute
The characteristic point on sampling another image of Feature Points Matching is stated, stable match point is obtained after screening；
Image segmentation unit, for carrying out trigonometric ratio segmentation to image respectively using the match point, obtains image respectively
Corresponding triangle set；
Parameter fitting unit, for the parallax using stable matching point in described image, it is determined that each Atria top
Parallax on point, further, obtains the parameter for being fitted the delta-shaped region disparity plane；
Parallax acquiring unit, for for the pixel, the prior probability of each parallax to be obtained using plane parameter, then
Likelihood probability, the comprehensive posterior probability for obtaining some parallax, further to obtain are calculated with reference to Matching power flow under each parallax
The initial parallax figure of whole image；
Parallax optimizes unit, for optimizing the initial parallax figure to obtain the disparity map of better quality.
According to another aspect of the present invention there is provided a kind of storage device, wherein a plurality of instruction that is stored with, the instruction
Suitable for being loaded by processor and being performed：
S1 extracts the feature of each pixel of image captured by binocular camera arranged in parallel；
S2 obtains sampling characteristic point to wherein one image according to fixed intervals sampling, according to the sampling Feature Points Matching
Characteristic point on another image, obtains stable match point after screening；
S3 carries out trigonometric ratio segmentation to image respectively using the match point, and the corresponding triangle set of image is obtained respectively；
S4 utilizes the parallax of stable matching point in described image, it is determined that the parallax on each Atria summit, enters one
Step ground, obtains the parameter for being fitted the delta-shaped region disparity plane；
S5 obtains the prior probability of each parallax using plane parameter, in conjunction with each parallax for the pixel
Matching power flow calculates likelihood probability, the comprehensive posterior probability for obtaining some parallax, further to obtain the initial of whole image
Disparity map；
S6 optimizes to obtain the disparity map of better quality to the initial parallax figure.
According to another aspect of the present invention, there is provided a kind of terminal, it is characterised in that including：
Processor, is adapted for carrying out each instruction；
Storage device, suitable for storing a plurality of instruction, wherein the instruction is applied to be loaded and performed by processor：
S1 extracts the feature of each pixel of image captured by binocular camera arranged in parallel；
S2 obtains sampling characteristic point to wherein one image according to fixed intervals sampling, according to the sampling Feature Points Matching
Characteristic point on another image, obtains stable match point after screening；
S3 carries out trigonometric ratio segmentation to image respectively using the match point, and the corresponding triangle set of image is obtained respectively；
S4 utilizes the parallax of stable matching point in described image, it is determined that the parallax on each Atria summit, enters one
Step ground, obtains the parameter for being fitted the delta-shaped region disparity plane；
S5 obtains the prior probability of each parallax using plane parameter, in conjunction with each parallax for the pixel
Matching power flow calculates likelihood probability, the comprehensive posterior probability for obtaining some parallax, further to obtain the initial of whole image
Disparity map；
S6 optimizes to obtain the disparity map of better quality to the initial parallax figure.
In general, by the contemplated above technical scheme of the present invention compared with prior art, with following beneficial effect
Really：
1) method of technical solution of the present invention, model is produced using probability, the probability of each parallax of pixel is calculated, by probability
Maximum parallax value is designated as the parallax of pixel, and drawing the method for disparity map of whole image, there is provided a kind of more robust
Pixel confidence balancing method is matched, the stable matching point of more robust can be obtained, the standard of binocular depth information calculating is improved
True property.
2) method of technical solution of the present invention, using the method for sampling matching, chooses several according to certain rule and adopts
Sampling point and its match point in another image, and wherein stable match point is further filtered out, for stable sampled point
Matching power flow is carried out with match point to calculate and parallax probability calculation, Stereo matching can be quickly carried out, and improves matching effect.
Brief description of the drawings
Fig. 1 is the schematic flow sheet schematic diagram of the Fast Stereo Matching Algorithm of technical solution of the present invention embodiment；
Fig. 2 is the characteristic vector make schematic diagram of technical solution of the present invention embodiment；
Fig. 3 is the Matching power flow schematic diagram of technical solution of the present invention embodiment；
Fig. 4 is the trigonometric ratio segmentation schematic diagram of technical solution of the present invention embodiment.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.As long as in addition, technical characteristic involved in each embodiment of invention described below
Not constituting conflict each other can just be mutually combined.With reference to embodiment, the present invention is described in more detail.
The Fast Stereo Matching Algorithm that model is produced based on probability that the present invention is provided, flow is as shown in figure 1, including feature
Extract, stable matching point is calculated, trigonometric ratio is split, planar delta is fitted, calculates dense parallax and parallax Optimization Steps；Tie below
Example is closed to be specifically described this Fast Stereo Matching Algorithm for producing model based on probability that the present invention is provided.
The Fast Stereo Matching Algorithm provided in an embodiment of the present invention that model is produced based on probability, is comprised the following steps that：
(1) feature extraction.Two images captured by selection binocular camera arranged in parallel, it is contemplated that binocular camera is in itself
Or it is made up of two monocular cameras, with reference to the characteristics of binocular image, is named as left image and right image.To left images
Feature extraction is carried out respectively, and it is comprised the following steps that：
(1.1) lower column processing is carried out respectively according to the different of input picture：If input picture is RGB color image,
It is all converted to single channel gray level image, if input picture is gray level image, the gray level image, note left and right is directly used
Figure is respectively I1,I2。
(1.2) convolution operation is carried out using Sobel operators respectively to left images, to carry out edge gradient extraction to it,
Obtain x, the gradient image in y directions, now the two of Sobel operators group weight matrix be respectively：
After convolution operation terminates, then by the range of all Grad Linear Mappings calculated to [0,255], specifically map
Method is that above-mentioned Grad is multiplied by into 1/4 adds 128；The Grad of two gradient images is designated as G respectivelyx, Gy。
(1.3) Grad calculated by previous step, is the characteristic vector that each pixel builds one 16 dimension, specifically
Construction method is：In each pixel correspondence GxAnd GyIn the 5*5 neighborhoods of position, choose certain point and be used as characteristic value, such as Fig. 2
It is shown, 12 points are chosen in x direction gradients, digitized representation point in y direction gradients choose 4 points, figure is in feature
Order in vector.If choosing the gradient information done in neighborhood for each pixel, current point k characteristic vector F is combined intok,
Remember all pixels point eigenvalue cluster into space be F.
(2) stable matching point is calculated.Left figure is sampled according to fixed intervals, sampled point is matched and screened, is obtained
Stable match point, its idiographic flow is as follows：
(2.1) 5 pixel samplings, i.e. selection coordinate are (0,0), (0,5), (0,10) ... at regular intervals in left figure
(15,5) ... (325,200) ... such point is used as investigation object.If setting current sampling point as S(l)(x, y), calculate its with
Parallax d ∈ [dmin,dmax] in the range of match point S(r)(x-d, y) between Matching power flow, computational methods are：Choose S(l)And S(r)The neighborhood point comparative feature of position four, and its characteristic difference is added up：
Wherein p and q are respectively S(l)And S(r)4 neighborhood points.
(2.2) S is investigated(l)Matching power flow of the sampled point under each parallax, finds out its minimum cost value c0, secondary small cost
Value c1, the second local minimum cost value c2, c0,c1,c2Implication refers to such as Fig. 3.Investigate whether it meets following condition：
If all meeting above-mentioned condition, S(l)Stable condition is met, retains the point, its parallax value is acquirement minimum cost
The parallax value of value：
(2.3) to the left figure sampled point of reservation, the match point of right figure is found according to its parallax, the match point of right figure is carried out
Investigation in step (2.2), if meeting stable condition, retains.
(2.4) if left figure sampled point is consistent with the parallax of its right figure match point, i.e.,：
|d(S(l))-d(S(r))|≤2
Then the two matching relationship of explanation is stable, and the sampled point and its match point are stable matching point.
(2.5) all stable matching points are recorded.
(3) trigonometric ratio is split.Using stable matching point in left figure, right figure, progress preferably is used to left images respectively
Delaunay trigonometric ratio partitioning algorithms, obtain triangle a set T, such as Fig. 4 respectively.
(4) planar delta is fitted.Using the parallax of stable matching point in left figure, right figure, it is determined that each Atria top
Parallax on point, so as to obtain the parameter for being fitted the delta-shaped region disparity plane, detailed process is as follows：
(4.1) i-th of Delta Region T is assumediVertex information is：A(x1,y1,d1),B(x2,y2,d2),C(x3,y3,d3),
Wherein x, y are respectively coordinate, and d is parallax, sets up equation group：
(4.2) solve equation group and obtain i-th of Delta Region disparity plane parameter θi={ ai,bi,ci}。
(4.3) above-mentioned solution is carried out for all triangles, its plane parameter is obtained respectively.
(5) parallax that model calculating is each put is produced based on probability.For each pixel, obtain every using plane parameter
The prior probability of individual parallax, likelihood probability, the comprehensive posteriority for obtaining some parallax are calculated in conjunction with Matching power flow under each parallax
Probability.Its process is as follows：
(5.1) for T in i-th of triangle in left figureiPixelCalculated using previous step
Plane parameter, can calculate and obtain the parallax value d of the point in the planek0=aixk+biyk+ci。
(5.2) with dk0Centered on set up a probability distribution, this probability distribution is superimposed with one by a Gaussian Profile
Even distribution, for [dmin,dmax] in the range of any one parallax value dkIts probability can be calculated according to this probability distribution, i.e.,
The prior probability of the parallax：
Wherein γ=5 and σ=1 are the parameter manually set.
(5.3) calculateWith parallax dk∈[dmin,dmax] in the range of match pointBetween matching generation
Valency, calculation is identical with step (2), so as to obtainThen parallax dkColor likelihood be：
Wherein parameter beta=1.
(5.4) posterior probability under each parallax is calculated：
In dk∈[dmin,dmax] in the range of choose maximum probability parallax be used as the final parallax of the point.Walked more than repeating
Suddenly, it is that all pixels point chooses suitable parallax, obtains the disparity map of left figure.Using right figure as object is investigated, (3)-(5) step is repeated
The rapid disparity map for obtaining right figure.
(6) parallax optimizes.Initial parallax figure is optimized to obtain the disparity map of better quality.Its process is as follows：
(6.1) cross is verified
Horizontal parallax figure is compared, it is assumed that the parallax of left disparity map terminal A (x, y) is d1, then in its right disparity map
Corresponding point is B (x-d1, y), investigate B points corresponding parallax d in right disparity map2If, | d1-d2|≤ε, (set ε=2) then
Think that A points are verified by cross, otherwise do not pass through.
(6.2) filling cavity
Cross validation condition is met in its adjacent domain for by the match point of cross validation condition, not choosing, and
The minimum parallax of numerical value in the region, replaces current parallax, to fill up the vacancy of the match point.Because being unsatisfactory for cross validation
Point is typically at occlusion area, and the true parallax of occlusion area is usually background parallax, and the minimum parallax of numerical value can be with here
It is used as background parallax.
(6.3) disparity map is filtered
Disparity map is filtered using median filter, noise spot is removed, final disparity map is obtained.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, it is not used to
The limitation present invention, any modifications, equivalent substitutions and improvements made within the spirit and principles of the invention etc., it all should include
Within protection scope of the present invention.
Claims (9)
1. a kind of quick stereo matching process that model is produced based on probability, it is characterised in that including,
S1 extracts the feature of each pixel of one group of image captured by binocular camera；
S2 obtains sampling characteristic point to an image in one group of image according to fixed intervals sampling, according to the sampling characteristic point
The characteristic point on another image in this group of image is matched, stable matching point is obtained after screening；
S3 carries out trigonometric ratio segmentation to image respectively using the stable matching point, and the corresponding triangle set of image is obtained respectively；
S4 utilizes the parallax of stable matching point in described image, it is determined that the parallax on each Atria summit, further,
Obtain the parameter for being fitted the delta-shaped region disparity plane；
S5 obtains the prior probability of each parallax using plane parameter for the pixel, is matched in conjunction with each parallax
Cost calculates likelihood probability, the comprehensive posterior probability for obtaining some parallax, further to obtain the initial parallax of whole image
Figure；
S6 optimizes to obtain the disparity map of better quality to the initial parallax figure.
2. a kind of quick stereo matching process that model is produced based on probability according to claim 1, wherein, the step
S1 is specifically included,
S11 calculates the edge gradient of described image respectively using arithmetic operators；
If S12 chooses the gradient information done in each neighborhood of pixels, the characteristic vector of current point is combined into.
3. a kind of quick stereo matching process that model is produced based on probability according to claim 1 or 2, wherein, it is described
Step S2 is specifically included,
S21 chooses a sampling characteristic point, calculates it with determining the Matching power flow between the match point in disparity range；
S22 investigates Matching power flow of the sampling characteristic point under each parallax, finds out its minimum value, sub-minimum and the second Local Minimum
Cost value, to calculate the stability of sampling characteristic point；Retain the sampling characteristic point for meeting stable condition, its parallax value is to obtain most
The parallax value of small cost value；
S23 finds its corresponding match point according to the parallax of sampling characteristic point, investigates Matching power flow of the match point under each parallax,
Its minimum value, sub-minimum and the second local minimum cost value are found out, to calculate the stability of match point；Reservation meets stablizing bar
The match point of part, its parallax value is the parallax value for obtaining minimum cost value；
If the parallax of the corresponding match point of S24 samplings characteristic point is consistent, it is determined that corresponding with its of the sampling characteristic point
It is stable matching point with point.
4. according to it is according to any one of claims 1 to 3 it is a kind of based on probability produce model quick stereo matching process, its
In, the step S5 is specifically included,
S51 is calculated according to the plane parameter of any pixel point and is obtained the parallax value of point in the plane；
S52 sets up probability distribution centered on the parallax value, for determining any one parallax value in scope, according to described general
Rate is distributed the prior probability for calculating the parallax；
S53 calculates the Matching power flow between the match point in pixel and certain disparity range, further, calculates its likelihood general
Rate；
S54 is calculated obtains determining the posterior probability in scope under each parallax, and the parallax for choosing maximum probability is final as the point
Parallax；
S55 obtains the parallax of whole pixels in described image respectively, further, and the disparity map of described image is obtained respectively.
5. a kind of quick stereo matching process that model is produced based on probability according to claim 4, wherein, it is described general
Rate distribution, which is preferably superimposed with one by a Gaussian Profile and is uniformly distributed, to be formed.
6. according to it is according to any one of claims 1 to 5 it is a kind of based on probability produce model quick stereo matching process, its
In, the step S6 is specifically included,
S61 carries out cross checking to initial parallax figure；
S62 chooses the less value of its close region parallax and replaces current parallax to the point do not verified by cross；
S63 is filtered to disparity map using wave filter, removes noise spot, obtains final disparity map.
7. what it is using a kind of quick stereo matching process that model is produced based on probability described in any one of claim 1~6 is
System, it is characterised in that including,
Image extraction unit, the feature of each pixel of image captured by binocular camera arranged in parallel for extracting；
Characteristic matching unit, for obtaining sampling characteristic point according to fixed intervals sampling to wherein one image, is adopted according to described
Characteristic point on another image of sample Feature Points Matching, obtains stable match point after screening；
Image segmentation unit, for carrying out trigonometric ratio segmentation to image respectively using the match point, obtains image correspondence respectively
Triangle set；
Parameter fitting unit, for the parallax using stable matching point in described image, it is determined that on each Atria summit
Parallax, further, obtain be fitted the delta-shaped region disparity plane parameter；
Parallax acquiring unit, for for the pixel, the prior probability of each parallax to be obtained using plane parameter, in conjunction with
Matching power flow calculates likelihood probability under each parallax, and the comprehensive posterior probability for obtaining some parallax is whole further to obtain
The initial parallax figure of image；
Parallax optimizes unit, for optimizing the initial parallax figure to obtain the disparity map of better quality.
8. a kind of storage device, wherein a plurality of instruction that is stored with, the instruction is applied to be loaded and performed by processor：
S1 extracts the feature of each pixel of image captured by binocular camera arranged in parallel；
S2 obtains sampling characteristic point to wherein one image according to fixed intervals sampling, another according to the sampling Feature Points Matching
The characteristic point on image is opened, stable match point is obtained after screening；
S3 carries out trigonometric ratio segmentation to image respectively using the match point, and the corresponding triangle set of image is obtained respectively；
S4 utilizes the parallax of stable matching point in described image, it is determined that the parallax on each Atria summit, further,
Obtain the parameter for being fitted the delta-shaped region disparity plane；
S5 obtains the prior probability of each parallax using plane parameter for the pixel, is matched in conjunction with each parallax
Cost calculates likelihood probability, the comprehensive posterior probability for obtaining some parallax, further to obtain the initial parallax of whole image
Figure；
S6 optimizes to obtain the disparity map of better quality to the initial parallax figure.
9. a kind of terminal, it is characterised in that including：
Processor, is adapted for carrying out each instruction；
Storage device, suitable for storing a plurality of instruction, wherein the instruction is applied to be loaded and performed by processor：
S1 extracts the feature of each pixel of image captured by binocular camera arranged in parallel；
S2 obtains sampling characteristic point to wherein one image according to fixed intervals sampling, another according to the sampling Feature Points Matching
The characteristic point on image is opened, stable match point is obtained after screening；
S3 carries out trigonometric ratio segmentation to image respectively using the match point, and the corresponding triangle set of image is obtained respectively；
S4 utilizes the parallax of stable matching point in described image, it is determined that the parallax on each Atria summit, further,
Obtain the parameter for being fitted the delta-shaped region disparity plane；
S5 obtains the prior probability of each parallax using plane parameter for the pixel, is matched in conjunction with each parallax
Cost calculates likelihood probability, the comprehensive posterior probability for obtaining some parallax, further to obtain the initial parallax of whole image
Figure；
S6 optimizes to obtain the disparity map of better quality to the initial parallax figure.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710351601.0A CN107133977A (en) | 2017-05-18 | 2017-05-18 | A kind of quick stereo matching process that model is produced based on probability |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710351601.0A CN107133977A (en) | 2017-05-18 | 2017-05-18 | A kind of quick stereo matching process that model is produced based on probability |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107133977A true CN107133977A (en) | 2017-09-05 |
Family
ID=59731808
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710351601.0A Pending CN107133977A (en) | 2017-05-18 | 2017-05-18 | A kind of quick stereo matching process that model is produced based on probability |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107133977A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108921020A (en) * | 2018-05-29 | 2018-11-30 | 北大青鸟环宇消防设备股份有限公司 | A kind of fire image-recognizing method and its device |
| CN109684932A (en) * | 2018-11-30 | 2019-04-26 | 华南农业大学 | A kind of pallet method for recognizing position and attitude based on binocular vision |
| CN110246169A (en) * | 2019-05-30 | 2019-09-17 | 华中科技大学 | A kind of window adaptive three-dimensional matching process and system based on gradient |
| CN110349198A (en) * | 2018-04-02 | 2019-10-18 | 联发科技股份有限公司 | Solid matching method and corresponding Stereo matching device |
| CN113141495A (en) * | 2020-01-16 | 2021-07-20 | 纳恩博（北京）科技有限公司 | Image processing method and device, storage medium and electronic device |
| CN114331919A (en) * | 2022-03-09 | 2022-04-12 | 合肥的卢深视科技有限公司 | Depth recovery method, electronic device, and storage medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102819843A (en) * | 2012-08-08 | 2012-12-12 | 天津大学 | Stereo image parallax estimation method based on boundary control belief propagation |
| CN103440653A (en) * | 2013-08-27 | 2013-12-11 | 北京航空航天大学 | Binocular vision stereo matching method |
| CN105513064A (en) * | 2015-12-03 | 2016-04-20 | 浙江万里学院 | Image segmentation and adaptive weighting-based stereo matching method |
-
2017
- 2017-05-18 CN CN201710351601.0A patent/CN107133977A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102819843A (en) * | 2012-08-08 | 2012-12-12 | 天津大学 | Stereo image parallax estimation method based on boundary control belief propagation |
| CN103440653A (en) * | 2013-08-27 | 2013-12-11 | 北京航空航天大学 | Binocular vision stereo matching method |
| CN105513064A (en) * | 2015-12-03 | 2016-04-20 | 浙江万里学院 | Image segmentation and adaptive weighting-based stereo matching method |
Non-Patent Citations (2)
| Title |
|---|
| 李娇 等: "一种基于贝叶斯理论的高效立体匹配方法", 《激光与光电子学进展》 * |
| 梁峰 等: "基于贝叶斯模型的快速立体匹配算法", 《计算机工程与设计》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110349198A (en) * | 2018-04-02 | 2019-10-18 | 联发科技股份有限公司 | Solid matching method and corresponding Stereo matching device |
| CN108921020A (en) * | 2018-05-29 | 2018-11-30 | 北大青鸟环宇消防设备股份有限公司 | A kind of fire image-recognizing method and its device |
| CN109684932A (en) * | 2018-11-30 | 2019-04-26 | 华南农业大学 | A kind of pallet method for recognizing position and attitude based on binocular vision |
| CN109684932B (en) * | 2018-11-30 | 2023-05-23 | 华南农业大学 | Binocular vision-based tray pose recognition method |
| CN110246169A (en) * | 2019-05-30 | 2019-09-17 | 华中科技大学 | A kind of window adaptive three-dimensional matching process and system based on gradient |
| CN113141495A (en) * | 2020-01-16 | 2021-07-20 | 纳恩博（北京）科技有限公司 | Image processing method and device, storage medium and electronic device |
| CN113141495B (en) * | 2020-01-16 | 2023-03-24 | 纳恩博（北京）科技有限公司 | Image processing method and device, storage medium and electronic device |
| CN114331919A (en) * | 2022-03-09 | 2022-04-12 | 合肥的卢深视科技有限公司 | Depth recovery method, electronic device, and storage medium |
| CN114331919B (en) * | 2022-03-09 | 2022-05-31 | 合肥的卢深视科技有限公司 | Depth recovery method, electronic device, and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107133977A (en) | A kind of quick stereo matching process that model is produced based on probability | |
| CN106780590B (en) | Method and system for acquiring depth map | |
| CN106651942B (en) | Three-dimensional rotating detection and rotary shaft localization method based on characteristic point | |
| CN110148181A (en) | A kind of general binocular solid matching process | |
| CN104346608B (en) | Sparse depth figure denseization method and apparatus | |
| EP3528209A1 (en) | Method and device for determining external parameter of stereoscopic camera | |
| CN106023230B (en) | A kind of dense matching method of suitable deformation pattern | |
| CN109690620A (en) | Threedimensional model generating means and threedimensional model generation method | |
| CN107798702B (en) | Real-time image superposition method and device for augmented reality | |
| CN106228528B (en) | A kind of multi-focus image fusing method based on decision diagram and rarefaction representation | |
| CN106296698B (en) | A kind of lightning 3-D positioning method based on stereoscopic vision | |
| WO2014073670A1 (en) | Image processing method and image processing device | |
| JP6020471B2 (en) | Image processing method, image processing apparatus, and image processing program | |
| CN111105452B (en) | Binocular vision-based high-low resolution fusion stereo matching method | |
| CN111062131A (en) | Power transmission line sag calculation method and related device | |
| US20130287291A1 (en) | Method of processing disparity space image | |
| CN109658497B (en) | Three-dimensional model reconstruction method and device | |
| CN107341804B (en) | Method and device for determining plane in point cloud data, and method and equipment for image superposition | |
| CN107560592A (en) | A kind of precision ranging method for optronic tracker linkage target | |
| US10582179B2 (en) | Method and apparatus for processing binocular disparity image | |
| CN113810611B (en) | Data simulation method and device for event camera | |
| CN106575447A (en) | Constructing a 3D structure | |
| JP2012194751A (en) | Image processing method, image processing system and computer program | |
| CN108109148A (en) | Image solid distribution method, mobile terminal | |
| CN112580434B (en) | Face false detection optimization method and system based on depth camera and face detection equipment |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170905 |
|
| WD01 | Invention patent application deemed withdrawn after publication |