CN107316324A - Method based on the CUDA real-time volume matchings realized and optimization - Google Patents
Method based on the CUDA real-time volume matchings realized and optimization Download PDFInfo
- Publication number
- CN107316324A CN107316324A CN201710352967.XA CN201710352967A CN107316324A CN 107316324 A CN107316324 A CN 107316324A CN 201710352967 A CN201710352967 A CN 201710352967A CN 107316324 A CN107316324 A CN 107316324A
- Authority
- CN
- China
- Prior art keywords
- pixel
- cuda
- super
- real
- optimization
- 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.)
- Granted
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
- G06T1/00—General purpose image data processing
- G06T1/20—Processor architectures; Processor configuration, e.g. pipelining
-
- 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
- G06T7/55—Depth or shape recovery from multiple 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
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20228—Disparity calculation for image-based rendering
Abstract
The present invention relates to the technical field of computer vision, the method for matching and optimizing more particularly, to the real-time volume realized based on CUDA.The present invention is a kind of to handle to carry out left and right input picture intensive Stereo matching, and the method for generating real-time disparity map using CUDA parallelizations.Including：Do census conversions to the figure of left and right two, generate a character string, with Hamming code away from obtaining an initial cost, each paths of pixel 8 of Dynamic Programming take a most short path and to obtain final cost, obtain an initial intensive disparity map；Super-pixel segmentation is carried out using k means algorithms to left figure, super-pixel flat blocks one by one is obtained, is optimized using super-pixel plane fitting to optimize initial parallax.Accelerate parallel the invention further relates to multitask GPU, and in particular in NVIDA CUDA framework Parallel Implementation multiple tasks, belong to GPGPU calculating fields.The calculating time is greatly reduced by the optimization of GPU multiple threads, real-time disparity map is obtained.
Description
Technical field
The present invention relates to the technical field of computer vision, more particularly, to the real-time volume realized based on CUDA
The method matched somebody with somebody and optimized.
Background technology
Stereoscopic vision is an important topic in computer vision, and it can cause people to get in object and scene
Depth information, be later stage 3D rebuild and content analysis basis.Stereo matching is one of key technology of stereoscopic vision, is passed through
The left and right picture that the binocular camera demarcated is shot, after being alignd through overcorrection, Stereo matching is direct acquisition parallax information
Final steps.
The direct purpose of Stereo matching is to obtain accurate parallax information, and the accuracy that matches and it is ageing be to weigh
The standard of method of completing the square, how to balance precision and it is ageing be an important problem.The solid matching method point of current main flow
For local matching, three kinds of half global registration and global registration, precision increases successively, while the time of consumption also increases successively.
Half global registration is preferable in precision and ageing balance, therefore we use half global registration, and puts forth effort to carry
High accurancy and precision, reduces run time.
In prospect and the marginal portion of background, precision is difficult to hold, and how to accept or reject and obtains in prospect expansion and noise
Good effect is to improve the key factor of accuracy.
It is the important means of front and rear scape edge segmentation by carrying out super-pixel plane segmentation to image, and dividing method
It is one of study hotspot of computer vision field, also there is extensive application in image procossing, Object identifying.It is accurate by quick
True super-pixel flat blocks segmentation, fitting significantly improves the accuracy of parallax.
GPGPU is the technology that GPU carries out large-scale calculations, and CUDA is the GPGPU architecture that NVIDA companies provide, GPU tools
There are the floating-point operation ability and memory bandwidth far above CPU, there is the concurrency of height, can only be one by one different from CPU programs
Order perform, GPU hardware design supports multitask to share, and program can simultaneously be run by multiple threads.
Therefore program is operated on GPU by CUDA, can greatly shortens run time, improved ageing.
Not yet find possessing pinpoint accuracy and the patent or document of real-time disparity map can be exported at present.
The content of the invention
There is provided the real-time volume realized based on CUDA at least one defect for overcoming described in above-mentioned prior art by the present invention
Matching and the method for optimization, are to take into account accurate and real-time solid matching method, add super-pixel segmentation to be fitted parallax
Plane improves accuracy, reduces error hiding probability, make use of CUDA multi-threaded parallelizations to design, and the algorithm routine of script is calculated
Time is reduced to a millimeter rank by second rank, can reach real-time dense Stereo Matching.
Technical scheme：Method based on the CUDA real-time volume matchings realized and optimization, wherein, including it is following
Step：
S1. a frame size is selected, being census except each pixel of marginal portion to left images is converted to character
Sequence；
S2. in given disparity range, each character string in the range of left figure pixel, traversal right figure is selected, is calculated
Hamming code is away from obtaining initial cost；
S3. path polymerize, and finds each path polymerizing value minimum put by Dynamic Programming, obtains energy function；
S4. the parallax value corresponding to energy function minimum is chosen with WTA；
S5. left and right consistency check carries out later stage verification, obtains initial dense disparity map；
S6. each pixel of left figure is converted into CIElab color spaces by RGB color；
S7. super-pixel flat blocks are marked off by k-means aggregating algorithms, iteration merges super picture several times up to convergence
Plain flat blocks；
S8. to each super-pixel plane multiple repairing weld, Calculation Plane parameter；
S9. calculated with obtained plane parameter and obtain new parallax value；
S10. interpolation arithmetic is done to new parallax, reduces black block, smooth disparity figure.
Programmed algorithm based on CUDA is all applied in above step.
Compared with prior art, beneficial effect is：In the present invention, CUDA streams are used for multiple times, shared drive block, thread is same
Step, multithreading shares paralell design, and optimizes data structure, all greatly improves operation efficiency.To left images
Census conversions are carried out, Hamming code are calculated away from cost is drawn, the energy value being polymerize with dynamic programming path optimizing obtains initial
Dense disparity map, based on the understanding to super-pixel plane, obtain each super-pixel block with k-means clustering convergences, will be each
Super-pixel block regards a plane as, and we, which calculate plane parameter, is used to be fitted new parallax value.Whole a set of flow because
There is the addition of super-pixel segmentation and plane fitting so that disparity map result accuracy is improved, and select parallel using CUDA
Changing programming also causes real-time to be guaranteed, therefore the present invention has accomplished good balance in accuracy and ageing aspect,
This does not have other document from present's view and patent was accomplished.
Brief description of the drawings
Fig. 1 represents the flow chart of a whole set of algorithm.
Fig. 2 represents the distribution of thread lattice and calculation when path polymerization flows parallel using 8 CUDA.
Fig. 3 represents the segmentation figure that k-means polymerizations are obtained.
Fig. 4 represents that the disparity map ultimately generated renders figure.
Embodiment
Accompanying drawing being given for example only property explanation, it is impossible to be interpreted as the limitation to this patent；It is attached in order to more preferably illustrate the present embodiment
Scheme some parts to have omission, zoom in or out, do not represent the size of actual product；To those skilled in the art,
Some known features and its explanation may be omitted and will be understood by accompanying drawing.Being given for example only property of position relationship described in accompanying drawing
Explanation, it is impossible to be interpreted as the limitation to this patent.
CUDA manages thread using thread lattice, and each copy of kernel function can be by built-in variable blockIdx come really
Determine the index of thread block, gridDim obtains the quantity of thread block, and blockDim obtains the quantity of thread, using different
GridDim and blockDim are the keys of the different parallelization degree of acquisition.When needing multiple thread blocks and each thread block again
During comprising multiple threads, index computational methods are：Int tid=threadIdx.x+blockIdx.x*blockDim.x；Obtain
Correct index tid, which can just navigate to us, needs thread to be processed, it is established that mapping relations, is specifically shown in Fig. 2
It is read-write in GPU internal storage structure to have：Register, shared drive and global memory.CUDA context switching machines
System is very efficient, almost zero-overhead, and the speed for accessing register is very fast, and register is preferentially used so should try one's best.
CUDA shared drive buffering areas are resided on physics GPU, so access speed is also quickly, the presence of shared drive can make operation
The intercommunication of multiple threads in thread block.Global memory is because GPU and CPU can carry out write operation, speed to it
It is slower.Therefore in CUDA programmings, the various internal memory modes of reasonable employment can effectively improve efficiency and the speed of service.
CUDA can use thread synchronization mechanism _ _ syncthreads ()；Thread in thread block is synchronized.The letter
Number may insure that all threads in same thread block can be just performed under it after all sentences before having performed the sentence
One sentence, can typically arrange in pairs or groups with shared drive and occur.
We also use the mode of CUDA streams, and CUDA flow tables show a GPU operation queue, and the operation in the queue
Performed with the sequencing for being added to queue.The parallel of task level can be realized using CUDA streams, can be simultaneously between kernel function
Row processing, can also be in main frame and exchanged between equipment data while GPU is performing kernel function.
In step one, what we chose is n*m block size progress census conversions.In CUDA programs, 16* have chosen
16 thread block, to improve efficiency with shared drive behind facilitating, the width of picture is width, and height is height, selection
Kernel sizes are ((width+15)/16, (height+15)/16) respectively.The size of shared drive is (n+16) * (m+16),
Data are first copied into shared drive from global memory, census conversions, synchronizing thread after conversion, by character string knot is then carried out
Fruit is stored in gpu global memories.The thought of Census conversions is the pixel that will be chosen, it is assumed that the gray value put for (x, y)
It is compared with the gray value of each pixel of its neighborhood block size, it is bigger than (x, y) to be set to 0, it is small for 1.
In step 2, the disparity range that we choose is d, and the thread block of 32*8 sizes is chosen in cuda, uses (32+
D) shared drive of * 8 sizes, first copies to the character string that each pixel is obtained in step one altogether from gpu global memories
Enjoy in internal memory, corresponding character string in d disparity range is then looked for from right figure, two pairs of character sequences are calculated by xor operation
The Hamming code of row away from.The Hamming code quantity different away from two equal length word correspondence positions are represented, we represent two with d (x, y)
Hamming distance between word x, y.Two character strings are carried out with XOR, and statistical result is 1 number, then this number is just
Hamming code away from.D parallax is traveled through, selection Hamming code is stored into global memory away from minimum parallax, obtain what we needed
Initial cost values.
In step 3, path polymerization is so as to finding the energy function of minimum to determine to regard in order to optimize energy function
Difference.
Lr (p, d)=C (p, d)+min (Lr (p-r, d),
Lr(p-r,d-1)+P1,Lr(p-r,d+1)+P1,
In above-mentioned formula, C (p, d) be exactly last step we obtain p points parallax be d when to cost values.
And Lr (p, d) then represents that in the parallax on r paths of making an inventory be d path cost and value, the cost polymerization of each point is expressed as
Value is that " (the current parallax cost polymerizing value of path consecutive points, the disparity difference of path consecutive points is 1 generation to current cost+min
Valency polymerizing value+P1, the parallax interpolation of path consecutive points is more than 1 minimum cost polymerizing value+P2) parallaxes of-path consecutive points inserts
Minimum cost polymerizing value of the value more than 1 ", last that is set to prevent numeral excessive.Wherein P1, P2 are phase respectively
Adjacent pixel parallax value difference is 1, and the penalty coefficient more than 1, and it is to ensure that disparity map is put down to add the two regularization terms
It is sliding to keep edge simultaneously.P1 is the adaptation to object out-of-flatness surface, and P2 is (is probably multiple objects) discontinuous to gray scale
Adapt to.
The image want in 2D is found so that the problem of minimum disparity map of cost polymerizing value is individual NP-hard, then individually
The path of the selection such as (such as from left to right, from right to left) match point causes cost polymerizing value minimum (i.e. finally on the r of certain direction
L is minimum), the problem reforms into soluble in polynomial time.We are using the dynamic programming algorithm with global nature
Minimum value is calculated to optimize.To make p cost want minimum, then premise must be the Least-cost of the point q in neighborhood,
Q wants Least-cost, then must assure that q neighborhood point m Least-cost, so hands on to divide and rule and draws minimum
Value.
If we solve only along every a line, then constraint in the ranks consider completely less than, q be p neighborhood point its
This real when is weakened to have arrived the left-hand point or right-hand point that q is p, such to seek excellent effect difference and occur that long-tail is imitated
Should.Then, we select to optimize using 8 directions path.Namely r=8, end product is the polymerizing value of 8 paths,
The energy function that namely we require.
In the program that we realize, 8 CUDA streams are used, while kernel functions are performed, can also have been transmitted
Data, and each path can carry out computing simultaneously, and to improve efficiency, the block of 4 32*8 sizes has been used respectively,
The grid of (height/8,1) size comes calculated level path and vertical-path, has used 4 32*8 size block,
The grid of ((height+width)/8,1) size calculates oblique path, has used the shared drive of 32*8 sizes to store picture
The path polymerization cost values of vegetarian refreshments, when asking for minimum value using parallel stipulations and the method for syncthreads thread synchronizations, pole
The big differentiation for reducing thread beam, and repeatedly instructed in the circulating cycle with pragma unroll so that circulation is sufficiently spread out parallel
Operation, improves compiler efficiency, and Fig. 2 displayings are flowed using 8 CUDA, the pass of parallel operation and memory headroom distribution between kernel
System's figure.
In step 4, WTA (Winner-take-all) is a kind of competition learning rule, for learning without tutor, used in this
In refer to that only the smallest of energy value can be activated and use, that is, choose the minimum parallax value of energy value.Still selected in this program
The thread block of 32*8 sizes, the shared drive block of 128*8 sizes, first by the energy function obtained in previous step from global memory
Copy in shared drive, then compare successively, minimum result is carried out into parallel stipulations to reduce branch, then by parallax value d
It is stored in new global memory.
In step 5, follow-up treatment measures are often used in disparity map：Left and right consistency detection.Block is only
Appear in piece image, and those points that can't see in another width figure.Block is typically a little one piece of continuous region, LRC
Check effect is to realize that occlusion detection obtains the corresponding shielded image of left figure.For a point p in left figure, that tries to achieve regards
Difference is d1, then corresponding points of the p in right figure should be (p-d1), and the parallax value of (p-d1) is denoted as d2.If | d1-d2 |>
Threshold, p are labeled as blocking a little, using 16*16 block and (width/16, height/16) size in this program
Grid, left images are traveled through, each thread one pixel of correspondence, if | d1-d2 |>1, then parallax is set to 0 guarantor
There is global memory
In step 6, the stage of super-pixel segmentation plane has been come into.Use SLIC methods in the present invention, i.e., it is simple
Linear iteraction cluster.The foundation of sub-clustering is the color similarity and propinquity between pixel, and coloured image is converted into
Then 5 dimensional feature vectors are constructed distance metric by 5 dimensional feature vectors under CIELAB color spaces and XY coordinates, to figure
As pixel carries out the process of Local Clustering.Therefore first have to image being transformed into CIElab color spaces from rgb color space.It is first
First RGB to XYZ conversion：
[X, Y, Z]=[M] * [R, G, B], wherein M are a 3x3 matrixes：
XYZ to Lab conversion：
L=116*f (Y1) -16
A=500* (f (X1)-f (Y1))
B=200* (f (Y1)-f (Z1))
Wherein f is the correction function of a similar Gamma function：
Work as x>When 0.008856, f (x)=x^ (1/3)；
Work as x<When=0.008856, f (x)=(7.787*x)+(16/116)
X1, Y1, Z1 are the XYZ values after linear normalization respectively.
In CUDA programs, the thread block of 16*16 sizes is used so that each pixel corresponds to a thread line
Journey, each thread carries out above-mentioned function change, and computational efficiency substantially achieves theoretical peak, afterwards again protects obtained value
In the lab global variables that there is gpu.
In step 7, k-means algorithms be a kind of simple efficient and widely used cluster clustering method its be mainly
The algorithm of data aggregation is calculated, mainly by constantly taking the algorithm from the nearest average of seed point.By picture initialization segmentation n blocks
Super-pixel, and n cluster heart point lab information and xy coordinates are stored in gpu global memories, using the block of w*h sizes, often
One pixel of individual thread correspondences, all searches the cluster heart point information of 2w*2h regions class, calculates lab color informations and two points
Euclidean distance between position, each thread, which updates the cluster heart point of minimum Euclidean distance, copies to global memory, iteration
Repeatedly until convergence；
Dlab=(Plab.x-Clab.x)2+(Plab.y-Clab.y)2+(Plab.z-Clab.z)2
Dxy=(Pxy.x-Cxy.x)2+(Pxy.y-Cxy.y)2
D=Dlab+w*Dxy
Above formula DlabRepresent that P points are put down with the Euclidean distance that C points in 2w*2h regions are cluster heart point in lab color spaces
Fang He, DxyThe Euclidean distance quadratic sum of P points and cluster heart point C coordinate in picture is represented, w is a weight parameter, each picture
The information of cluster heart point after element renewal is planar tags.One threshold value is set, and the super-pixel plane less than this threshold value can be with
The average Lab color distances of the plane computations of surrounding, the minimum carry out plane fusion of distance.Fig. 3 is to have carried out k-means and melted
Super-pixel segmentation figure after conjunction.
In step 8, label has been beaten to each pixel in image in step 7, has shown that each pixel has pair
The super-pixel block answered.Known three points can represent a plane, so in each 3 parallax values of super-pixel block stochastical sampling not
For 0 and different point plane equation is determined by three points sampled：
A=(x1 × z2-x2 × z1) × (y2 × z3-y3 × z2)-(x2 × z3-x3 × z2) × (y1 × z2-y2 × z1)
B=y1 × z2-y2 × z1；
P0=((z2 × d1-z1 × d2) × (y2 × z3-y3 × z2)-(z3 × d2-z2 × d3) × (y1 × z2-y2 ×
z1))/A；
P1=(z3 × d2-z2 × d3-P0 × (x2 × z3-x3 × z2))/B；
P2=(d1-P0 × xi-P1×yi)/zi；
What above formula was represented is the equation of Calculation Plane parameter, if A<0, then plane parameter P0=0, P1=0, P2=-
If 1. B<0, then
P1=(z3 × d2-z2 × d3-P0 × (x2 × z3-x3 × z2))/B；
P2 parameters depend on ZiSize whether be less than 0, initialization for the first time seek plane parameter when, Z is 1, (x, y,
D) X-coordinate of pixel, Y-coordinate and parallax value are represented respectively.
After plane parameter is obtained, you can try to achieve plane equation：S=P0 × x+P1 × y+P2；Remaining planar point is used
Equation is fitted, and finds out the number of match point in a plane, and circulation is multiple, is chosen the most plane parameter of match point, is protected
It is stored in result.All match points are added up, new parameter is drawn, the plane parameter for carrying out a new round is calculated, and will
Final result is stored in data structure.
In step 9, we have been obtained for the label and plane parameter of each pixel in image in step 8, because
One flat blocks can be thought that parallax value is varied less for super-pixel flat blocks segmentation let us, therefore we use correspondence
Plane parameter equation replace initial parallax value, so as to reach the effect of optimization.New parallactic equation is exactly：D=P0 × x
+P1×y+P2；If d is in [0,255] is interval, just original initial parallax value is replaced with this new d value.Note
On the edge line of super-pixel flat blocks, it is not replaced, still using initial parallax value.
In step 10, row interpolation is entered to disparity map.In order that disparity map is more smooth, we are to pixel that parallax value is 0
Carry out interpolation processing.The interval that parallax value is 0 is found first, it is assumed that is [start, end], is then compared start-1 and end+1
The parallax value of point, is inserted into [start, end] interval by that less parallax value.By interpolation arithmetic, on the ground that parallax is 0
Square transition is more smooth, and what the black surround that the edge split in front and rear scape is caused because of prospect expansion also can be relatively good is smooth, carries
The high accuracy for regarding illustration.
In the present invention, CUDA streams are used for multiple times in we, and shared drive block, thread synchronization, multithreading is shared parallelization and set
Meter, and data structure is optimized, all greatly improve operation efficiency.Census conversions are being carried out to left images, the Chinese is calculated
Plain code is away from cost is drawn, the energy value being polymerize with dynamic programming path optimizing obtains initial dense disparity map, based on to super
The understanding of pixel planes, each super-pixel block is obtained with k-means clustering convergences, is regarded each super-pixel block as one and is put down
Face, we, which calculate plane parameter, is used to be fitted new parallax value.Whole a set of flow is because have super-pixel segmentation and plane to intend
The addition of conjunction so that disparity map result accuracy is improved, and select to use CUDA parallel programmings also so that real-time is obtained
Guarantee has been arrived, therefore the present invention has accomplished good balance in accuracy and ageing aspect, this is not different from present's view
Document and patent accomplished.
Obviously, the above embodiment of the present invention is only intended to clearly illustrate example of the present invention, and is not pair
The restriction of embodiments of the present invention.For those of ordinary skill in the field, may be used also on the basis of the above description
To make other changes in different forms.There is no necessity and possibility to exhaust all the enbodiments.It is all this
Any modifications, equivalent substitutions and improvements made within the spirit and principle of invention etc., should be included in the claims in the present invention
Protection domain within.
Claims (2)
1. the method based on the CUDA real-time volume matchings realized and optimization, it is characterised in that comprise the following steps：
S1. a frame size is selected, being census except each pixel of marginal portion to left images is converted to character sequence
Row；
S2. in given disparity range, each character string in the range of left figure pixel, traversal right figure is selected, Hamming is calculated
Code distance, obtains initial cost；
S3. path polymerize, and finds each path polymerizing value minimum put by Dynamic Programming, obtains energy function；
S4. the parallax value corresponding to energy function minimum is chosen with WTA；
S5. left and right consistency check carries out later stage verification, obtains initial dense disparity map；
S6. each pixel of left figure is converted into CIElab color spaces by RGB color；
S7. super-pixel flat blocks are marked off by k-means aggregating algorithms, iteration is several times up to convergence, and it is flat to merge super-pixel
Face block；
S8. to each super-pixel plane multiple repairing weld, Calculation Plane parameter；
S9. calculated with obtained plane parameter and obtain new parallax value；
S10. interpolation arithmetic is done to new parallax, reduces black block, smooth disparity figure.
2. the method according to claim 1 based on the CUDA real-time volume matchings realized and optimization, it is characterised in that：Institute
In the step of stating, the programmed algorithm based on CUDA is all applied in above step.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710352967.XA CN107316324B (en) | 2017-05-18 | 2017-05-18 | Method for realizing real-time stereo matching and optimization based on CUDA |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710352967.XA CN107316324B (en) | 2017-05-18 | 2017-05-18 | Method for realizing real-time stereo matching and optimization based on CUDA |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107316324A true CN107316324A (en) | 2017-11-03 |
| CN107316324B CN107316324B (en) | 2020-08-04 |
Family
ID=60183844
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710352967.XA Active CN107316324B (en) | 2017-05-18 | 2017-05-18 | Method for realizing real-time stereo matching and optimization based on CUDA |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107316324B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110310220A (en) * | 2018-03-27 | 2019-10-08 | 广东瑞图万方科技股份有限公司 | A kind of half global real-time volume matching process |
| CN111508068A (en) * | 2020-04-20 | 2020-08-07 | 华中科技大学 | Three-dimensional reconstruction method and system applied to binocular endoscope image |
| CN112233136A (en) * | 2020-11-03 | 2021-01-15 | 上海西井信息科技有限公司 | Method, system, equipment and storage medium for alignment of container trucks based on binocular recognition |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102572485A (en) * | 2012-02-02 | 2012-07-11 | 北京大学 | Self-adaptive weighted stereo matching algorithm, stereo display and collecting device and system |
| US20130070060A1 (en) * | 2011-09-19 | 2013-03-21 | Pelican Imaging Corporation | Systems and methods for determining depth from multiple views of a scene that include aliasing using hypothesized fusion |
| CN104732505A (en) * | 2015-03-18 | 2015-06-24 | 浙江大学 | Method for generating hiding images capable of recommending hiding positions |
| CN105847783A (en) * | 2016-05-17 | 2016-08-10 | 武汉鸿瑞达信息技术有限公司 | 3D video display and interaction method based on stream media and device |
| CN106570901A (en) * | 2016-11-11 | 2017-04-19 | 浙江大学 | CUDA-based binocular depth information recovery acceleration method |
-
2017
- 2017-05-18 CN CN201710352967.XA patent/CN107316324B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130070060A1 (en) * | 2011-09-19 | 2013-03-21 | Pelican Imaging Corporation | Systems and methods for determining depth from multiple views of a scene that include aliasing using hypothesized fusion |
| CN102572485A (en) * | 2012-02-02 | 2012-07-11 | 北京大学 | Self-adaptive weighted stereo matching algorithm, stereo display and collecting device and system |
| CN104732505A (en) * | 2015-03-18 | 2015-06-24 | 浙江大学 | Method for generating hiding images capable of recommending hiding positions |
| CN105847783A (en) * | 2016-05-17 | 2016-08-10 | 武汉鸿瑞达信息技术有限公司 | 3D video display and interaction method based on stream media and device |
| CN106570901A (en) * | 2016-11-11 | 2017-04-19 | 浙江大学 | CUDA-based binocular depth information recovery acceleration method |
Non-Patent Citations (1)
| Title |
|---|
| HYEONG-JUN CHO ET AL: "On high-computational disparity acquisition of HD image using CUDA implementation for stereo matching", 《2015 INTERNATIONAL SOC DESIGN CONFERENCE》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110310220A (en) * | 2018-03-27 | 2019-10-08 | 广东瑞图万方科技股份有限公司 | A kind of half global real-time volume matching process |
| CN111508068A (en) * | 2020-04-20 | 2020-08-07 | 华中科技大学 | Three-dimensional reconstruction method and system applied to binocular endoscope image |
| CN112233136A (en) * | 2020-11-03 | 2021-01-15 | 上海西井信息科技有限公司 | Method, system, equipment and storage medium for alignment of container trucks based on binocular recognition |
| CN112233136B (en) * | 2020-11-03 | 2021-10-22 | 上海西井信息科技有限公司 | Method, system, equipment and storage medium for alignment of container trucks based on binocular recognition |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107316324B (en) | 2020-08-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zou et al. | Df-net: Unsupervised joint learning of depth and flow using cross-task consistency | |
| CN107481279B (en) | Monocular video depth map calculation method | |
| WO2015139574A1 (en) | Static object reconstruction method and system | |
| CN104616286B (en) | Quick semi-automatic multi views depth restorative procedure | |
| Chandraker et al. | Globally optimal algorithms for stratified autocalibration | |
| Larsson et al. | Rank minimization with structured data patterns | |
| CN107316324A (en) | Method based on the CUDA real-time volume matchings realized and optimization | |
| CN110310220A (en) | A kind of half global real-time volume matching process | |
| WO2018214086A1 (en) | Method and apparatus for three-dimensional reconstruction of scene, and terminal device | |
| CN106952304A (en) | A kind of depth image computational methods of utilization video sequence interframe correlation | |
| CN108961385A (en) | A kind of SLAM patterning process and device | |
| Chang et al. | Efficient stereo matching on embedded GPUs with zero-means cross correlation | |
| CN103295223B (en) | For the apparatus and method using observability energy model estimating disparity | |
| CN117216591A (en) | Training method and device for three-dimensional model matching and multi-modal feature mapping model | |
| Yao et al. | AGO: accelerating global optimization for accurate stereo matching | |
| Madhusudana et al. | Revisiting dead leaves model: Training with synthetic data | |
| CN112116640B (en) | Binocular stereo matching method based on OpenCL | |
| CN110059651B (en) | Real-time tracking and registering method for camera | |
| CN109509218B (en) | The method, apparatus of disparity map is obtained based on FPGA | |
| Garcia et al. | 3D reconstruction with projective octrees and epipolar geometry | |
| Walia et al. | A computationally efficient framework for 3D warping technique | |
| Mezenceva et al. | The Study of the Semi-Global Block Matching Algorithm Implementing Parallel Calculation with GPU | |
| US20220165029A1 (en) | Computer Vision Systems and Methods for High-Fidelity Representation of Complex 3D Surfaces Using Deep Unsigned Distance Embeddings | |
| Muresan et al. | Patch warping and local constraints for improved block matching stereo correspondence | |
| Zhang et al. | An adaptive pose tracking method based on sparse point clouds matching |
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 |