CN110533710A - A kind of method and processing unit of the binocular ranging algorithm based on GPU - Google Patents
A kind of method and processing unit of the binocular ranging algorithm based on GPU Download PDFInfo
- Publication number
- CN110533710A CN110533710A CN201910779546.4A CN201910779546A CN110533710A CN 110533710 A CN110533710 A CN 110533710A CN 201910779546 A CN201910779546 A CN 201910779546A CN 110533710 A CN110533710 A CN 110533710A
- Authority
- CN
- China
- Prior art keywords
- cost
- value
- polymerizing
- polymerizing value
- polymerize
- 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
- 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
- G06T2207/10028—Range image; Depth image; 3D point clouds
-
- 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 method and processing unit of the embodiment of the invention discloses a kind of binocular ranging algorithm based on GPU improve the real-time of binocular depth cognition technology for promoting the operation efficiency of image matching algorithm in binocular vision.The method comprise the steps that obtaining the first image data and second picture data, the first image data and second picture data are respectively by different camera acquisitions；According to the first image data and second picture data, cost calculating is carried out, cost value is obtained；It is polymerize according to the synchronous cost for carrying out first direction, second direction and third direction of cost value and is calculated, obtains the cost polymerizing value of the cost polymerizing value of first direction, the cost polymerizing value of second direction and third direction；It is polymerize according to the cost that cost value carries out fourth direction and is calculated, obtains the cost polymerizing value of fourth direction；According to the cost polymerizing value of first direction, the cost polymerizing value of second direction, the cost polymerizing value of the cost polymerizing value of third direction and fourth direction, parallax value is determined.
Description
Technical field
The present invention relates to field of image processing more particularly to a kind of methods and processing of the binocular ranging algorithm based on GPU
Device.
Background technique
With the development of science and technology, unmanned robot is being widely used in all fields, and unmanned robot has
One common demand, needing to adjust the distance is perceived.Common distance measuring method mainly has two major classes at present: initiative range measurement and super
Sound ranging, infrared distance measurement, binocular distance measurement etc..Although initiative range measurement mode principle is relatively simple, real-time is higher,
Vulnerable to the influence of object reflecting surface, ambient environment etc., therefore initiative range measurement mode is not made in unmanned machine domain variability
For main distance measuring method.
Binocular distance measurement obtains scene images by two cameras, using different scenery two cameras it
Between imaging position it is different, try to calculate parallax, then calculate final distance according to the parallaxometer estimated.It is existing double
Visually feel location algorithm, due to having biggish calculation amount in the images match stage, real-time is relatively difficult to ensure, can not will be double
Mesh vision technique is applied well in unmanned robot.
Summary of the invention
The method and processing unit of the embodiment of the invention provides a kind of binocular ranging algorithm based on GPU, for being promoted
The operation efficiency of image matching algorithm in binocular vision improves the real-time of binocular depth cognition technology.
In view of this, first aspect present invention provides a kind of method of binocular ranging algorithm based on GPU, may include:
Obtain the first image data and second picture data, first image data and second picture data difference
It is obtained by different cameras；
According to first image data and the second picture data, cost calculating is carried out, cost value is obtained；
It is polymerize according to the synchronous cost for carrying out first direction, second direction and third direction of the cost value and is calculated, is obtained
The cost polymerizing value of the cost polymerizing value of first direction, the cost polymerizing value of second direction and third direction；
It is polymerize according to the cost that the cost value carries out fourth direction and is calculated, obtains the cost polymerizing value of fourth direction；
According to the cost polymerizing value of the first direction, the cost polymerizing value of the second direction, the third direction
The cost polymerizing value of cost polymerizing value and the fourth direction, determines parallax value.
Optionally, in some embodiments of the invention, described according to first image data and the second picture
Data carry out cost calculating, obtain cost value, comprising:
According to first image data and the second picture data, cost meter is carried out by the block of different arrangements
It calculates, obtains cost value, wherein one pixel of per thread alignment processing in each block.
Optionally, in some embodiments of the invention, described to carry out first direction, second according to the cost value is synchronous
The cost of direction and third direction polymerization calculate, obtain the cost polymerizing value of first direction, the cost polymerizing value of second direction and
The cost polymerizing value of third direction, comprising:
It is synchronous to carry out first direction, second direction and the by SGM binocular image matching algorithm according to the cost value
The cost in three directions, which polymerize, to be calculated, and the cost polymerizing value of first direction, the cost polymerizing value of second direction and third direction are obtained
Cost polymerizing value；
The cost for carrying out fourth direction according to the cost value, which polymerize, to be calculated, and the cost polymerization of fourth direction is obtained
Value, comprising:
According to the cost value, by SGM binocular image matching algorithm, the cost polymerization for carrying out fourth direction is calculated, is obtained
To the cost polymerizing value of fourth direction.
Optionally, in some embodiments of the invention, described to carry out first direction, second according to the cost value is synchronous
The cost of direction and third direction polymerization calculate, obtain the cost polymerizing value of first direction, the cost polymerizing value of second direction and
The cost polymerizing value of third direction, the cost for carrying out fourth direction according to the cost value, which polymerize, to be calculated, and obtains four directions
To cost polymerizing value, comprising:
According to the cost value, the cost value of first direction, the cost value of second direction, the cost value of third direction are determined
With the cost value of fourth direction；
According to the cost value of the first direction, the cost value of the cost value of the second direction and the third direction,
By butterfly sort algorithm, the synchronous cost for carrying out first direction, second direction and third direction, which polymerize, to be calculated, and obtains first party
To cost polymerizing value, the cost polymerizing value of the cost polymerizing value of second direction and third direction；
According to the cost value of the fourth direction, by butterfly sort algorithm, the cost polymerization for carrying out fourth direction is calculated,
Obtain the cost polymerizing value of fourth direction.
Optionally, in some embodiments of the invention, the cost polymerizing value according to the first direction, described
The cost polymerizing value of the cost polymerizing value in two directions, the cost polymerizing value of the third direction and the fourth direction determines view
Difference, comprising:
By the cost polymerizing value of the first direction, the cost polymerizing value of the second direction, the third direction generation
The cost polymerizing value of valence polymerizing value and the fourth direction, adds up in the case where different parallax values, obtain with it is described not
The corresponding cumulative polymerizing value with parallax value；
By butterfly sort algorithm, the minimum value in determining cumulative polymerizing value corresponding with the different parallax values is parallax
Value.
Second aspect of the present invention provides a kind of processing unit, may include:
Module is obtained, for obtaining the first image data and second picture data, first image data and described the
What two image datas were obtained by different cameras respectively；
Processing module, for carrying out cost calculating, obtaining according to first image data and the second picture data
Cost value；It is polymerize according to the synchronous cost for carrying out first direction, second direction and third direction of the cost value and is calculated, obtains the
The cost polymerizing value of the cost polymerizing value in one direction, the cost polymerizing value of second direction and third direction；According to the cost value
The cost for carrying out fourth direction, which polymerize, to be calculated, and the cost polymerizing value of fourth direction is obtained；It is poly- according to the cost of the first direction
Conjunction value, the cost polymerizing value of the second direction, the cost of the cost polymerizing value of the third direction and the fourth direction are poly-
Conjunction value, determines parallax value.
Optionally, in some embodiments of the invention,
The processing module is specifically used for passing through difference according to first image data and the second picture data
The block of arrangement carries out cost calculating, obtains cost value, wherein one picture of per thread alignment processing in each block
Element.
Optionally, in some embodiments of the invention,
The processing module is specifically used for according to the cost value, synchronous to carry out by SGM binocular image matching algorithm
The cost of first direction, second direction and third direction, which polymerize, to be calculated, and the cost polymerizing value of first direction, second direction are obtained
The cost polymerizing value of cost polymerizing value and third direction；It is carried out according to the cost value by SGM binocular image matching algorithm
The cost of fourth direction, which polymerize, to be calculated, and the cost polymerizing value of fourth direction is obtained.
Optionally, in some embodiments of the invention,
The processing module is specifically used for determining the cost value of first direction, the generation of second direction according to the cost value
Value, the cost value of the cost value of third direction and fourth direction；According to the cost value of the first direction, the second direction
Cost value and the third direction cost value, it is synchronous to carry out first direction, second direction and the by butterfly sort algorithm
The cost in three directions, which polymerize, to be calculated, and the cost polymerizing value of first direction, the cost polymerizing value of second direction and third direction are obtained
Cost polymerizing value；The cost polymerization of fourth direction is carried out by butterfly sort algorithm according to the cost value of the fourth direction
It calculates, obtains the cost polymerizing value of fourth direction.
Optionally, in some embodiments of the invention,
The processing module, specifically for gathering the cost of the cost polymerizing value of the first direction, the second direction
The cost polymerizing value of conjunction value, the cost polymerizing value of the third direction and the fourth direction, in the case where different parallax values
It adds up, obtains cumulative polymerizing value corresponding from the different parallax values；Pass through butterfly sort algorithm, the determining and difference
Minimum value in the corresponding cumulative polymerizing value of parallax value is parallax value.
Third aspect present invention provides a kind of processing unit, may include:
Transceiver, processor, memory, wherein the transceiver, the processor and the memory are connected by bus
It connects；
The memory, for storing operational order；
The transceiver, for obtaining the first image data and second picture data, first image data and described
What second picture data were obtained by different cameras respectively；
The processor, for calling the operational order, execute such as first aspect present invention and first aspect is any can
The step of selecting the method for the binocular ranging algorithm described in implementation based on GPU.
Fourth aspect present invention provides a kind of readable storage medium storing program for executing, is stored thereon with computer program, which is characterized in that institute
It states and realizes when computer program is executed by processor such as institute in first aspect present invention and any optional implementation of first aspect
The step of method for the binocular ranging algorithm based on GPU stated.
As can be seen from the above technical solutions, the embodiment of the present invention has the advantage that
In embodiments of the present invention, the first image data and second picture data, first image data and institute are obtained
State what second picture data were obtained by different cameras respectively；According to first image data and the second picture number
According to progress cost calculating obtains cost value；First direction, second direction and third direction are carried out according to the cost value is synchronous
Cost polymerize calculate, obtain the cost polymerizing value, the cost polymerizing value of second direction and the cost of third direction of first direction
Polymerizing value；It is polymerize according to the cost that the cost value carries out fourth direction and is calculated, obtains the cost polymerizing value of fourth direction；According to
The cost polymerizing value of the first direction, the cost polymerizing value of the second direction, the cost polymerizing value of the third direction and
The cost polymerizing value of the fourth direction, determines parallax value.The present invention utilizes GPU concurrent operation, is suitable for the spy of large-scale calculations
GPU is introduced into binocular ranging algorithm by point, promotes the operation efficiency of image matching algorithm in binocular vision, and is promoted double
The real-time of mesh matching algorithm.
Detailed description of the invention
Technical solution in order to illustrate the embodiments of the present invention more clearly, below will be to institute in embodiment and description of the prior art
Attached drawing to be used is needed to be briefly described, it should be apparent that, the accompanying drawings in the following description is only some implementations of the invention
Example, can also be obtained according to these attached drawings other attached drawings.
Fig. 1 is that cost polymerize the schematic diagram merged with disparity computation in the embodiment of the present invention；
Fig. 2 is one embodiment schematic diagram of the method for the binocular ranging algorithm based on GPU in the embodiment of the present invention；
Fig. 3 A is a time diagram of the stream of GPU in the embodiment of the present invention；
Fig. 3 B is the design diagram that cost calculates in the embodiment of the present invention；
Fig. 3 C is a design diagram of cost polymerization in the embodiment of the present invention；
Fig. 3 D is a schematic diagram of cost calculation optimization in the embodiment of the present invention；
Fig. 3 E is the schematic diagram that the array Shared_base length of benchmark pixel is stored in the embodiment of the present invention；
Fig. 3 F is a schematic diagram of butterfly sort algorithm in the embodiment of the present invention；
Fig. 4 is one embodiment schematic diagram of processing unit in the embodiment of the present invention；
Fig. 5 is one embodiment schematic diagram of processing unit in the embodiment of the present invention.
Specific embodiment
The method and processing unit of the embodiment of the invention provides a kind of binocular ranging algorithm based on GPU, for being promoted
The operation efficiency of image matching algorithm in binocular vision improves the real-time of binocular depth cognition technology.
In order to enable those skilled in the art to better understand the solution of the present invention, below in conjunction in the embodiment of the present invention
Attached drawing, technical solution in the embodiment of the present invention are described, it is clear that described embodiment is only present invention a part
Embodiment, instead of all the embodiments.Based on the embodiments of the present invention, it should fall within the scope of the present invention.
The present invention utilizes graphics processor (Graphics Processing Unit, GPU) concurrent operation, is suitable for extensive
GPU is introduced into binocular ranging algorithm by the characteristics of calculating, to promote the real-time of binocular ranging algorithm.But it is double in order to allow
Mesh matching algorithm plays better effect, is redesigned to the realization on GPU.It is directed to the fortune of binocular ranging algorithm
Calculation process devises a series of GPU prioritization scheme.First on total algorithm framework, in order to promote Data duplication utilization rate,
Cost polymerize by the present invention to be merged together with disparity computation, as shown in Figure 1, for cost polymerization and view in the embodiment of the present invention
Difference calculates a schematic diagram of fusion.
Below by way of examples, technical solution of the present invention is described further, as shown in Fig. 2, for the present invention
One embodiment schematic diagram of the method for binocular ranging algorithm in embodiment based on GPU may include:
201, the first image data and second picture data, first image data and the second picture data are obtained
It is obtained respectively by different cameras.
It is understood that first image data and the second picture data are obtained by different cameras respectively
, then the first image data can be and be obtained by left camera, second picture data, which can be, to be obtained by right camera.
It can be directed to the different operation process of different binocular ranging algorithms, design different GPU resource allocation plans.Such as Fig. 3 A institute
Show, is a time diagram of the stream (stream) of GPU in the embodiment of the present invention.It is parallel that the present invention devises 3 streams
Thinking, promote the operational efficiency of matching algorithm.As shown in Figure 3A, in cost polymerization stage, allow GPU while calculating 3 directions
Cost.Since the present invention is merged cost polymerization and disparity computation, it is therefore necessary to ensure other in cost polymerization
After the operation in direction is all finished, the operation of disparity computation could be executed, and the task in the same stream follow by
The principle that sequence executes, Executing Cost polymerization ' ↑ ' Shi Yiding ensure that the cost polymerization in first three direction calculates and completed.
It is easy to carry out concurrent operation due to the single calculating step very simple of calculating cost in cost calculation stages, because
This fairly simple practical and efficient mentality of designing is that each thread is allowed to be responsible for a pixel.In this way, an instruction cycle can
Directly to have handled a width picture.Therefore, the direction x for designing two dimension a block, each block first includes 32 threads,
The direction y includes 32 threads, wherein it is understood that the Thread Count of a 32 exactly thread beams (warp).Then, if
A two-dimensional thread lattice (grid) is set, the direction x of grid will include cols/blockDim.x blocks, the side y of grid
To will include cols/blockDim.y blocks.As shown in Figure 3B, the design calculated for cost in the embodiment of the present invention
Schematic diagram.
In cost polymerization stage, the optimal parallax of two pixels of each Blocks design treatment, and handed over to promote data
Mutual efficiency, using the thread (thread) in the same warp can directly mutual shared data principle, by a picture
The calculating of all parallax sizes is completed with a warp on certain plain direction.One warp maximum includes 32 thread, this hair
The calculating of a pixel whole parallax cost is completed in bright design with a warp, therefore is handled two pixels and needed two in total
Warp, therefore include 64 threads in a blocks.Whole parallax costs is calculated by 32 threads in order to meet, here
It designs per thread and handles MAX_DISPARITY/32 parallax.Whole thinking is calculated similar to cost, each block processing
Two row pixels gradually handle other pixels in a line using for circulation in block, and resource results distribution is as shown in Figure 3 C,
Fig. 3 C is a design diagram of cost polymerization in the embodiment of the present invention.
It should be noted that GPU architecture is not also identical for different directions, but general thought is all each grid
It handles a line or a column, each blocks handles two pixels.I.e. by the binocular vision matching optimization algorithm based on GPU,
The operation efficiency of binocular vision matching algorithm is obviously improved.Illustratively, tall and handsome up to picture processing speed on TX2 processor
Degree reaches 42FPS, can be applied in unmanned plane obstacle avoidance system.
202, according to first image data and the second picture data, cost calculating is carried out, cost value is obtained.
It is described to carry out cost calculating according to first image data and the second picture data, obtain cost value, it can
To include: to carry out cost meter by the block of different arrangements according to first image data and the second picture data
It calculates, obtains cost value, wherein one pixel of per thread alignment processing in each block.
Cost is calculated below and carries out a brief description, can be used in the embodiment of the present invention centrosymmetric
Census transformation carries out cost calculating, and this method can subtract in the case where the influence of anti-light line can be good as traditional census
Few a certain amount of stored memory.The present invention utilizes the concurrency of GPU, realizes optimization to centrosymmetric census transformation, has
Body optimum ideals are as follows:
A two-dimentional thread block (block) is designed first, and the direction x of each block includes 32 threads, and the direction y includes
32 threads.It is because this is exactly the Thread Count of a warp it is understood why selecting 32.Then, it is provided with
The direction x of one two-dimensional grid, grid will include cols/blockDim.x blocks, and the direction y of grid will include
Cols/blockDim.y blocks.
Centrosymmetric census is converted, due to individually calculating step very simple, is easy to carry out concurrent operation, because
This, fairly simple practical and efficient mentality of designing is that each thread is allowed to be responsible for a pixel.In this way, an instruction cycle can
Directly to have handled a width picture.The resource structures of GPU are as shown in Figure 3B, are 640* with the photo resolution that camera acquires
It is illustrated for 480.Here the direction x for designing two dimension a block, each block first includes 32 threads, the direction y
Comprising 32 threads, why 32 are selected, is because this is exactly the Thread Count of a warp.Then, setting one is two-dimensional
The direction x of grid, grid will include cols/blockDim.x blocks, and the direction y of grid will include cols/
BlockDim.y blocks, by taking 640*480 as an example, the direction y of cols/32=20, the direction the x blocks, grid of grid will
Include row/32=15 blocks.It handles in this way, each thread can just be allowed to handle a pixel.
Followed by the calculating of cost, the entire matching cost space known to the theory of binocular vision is W × H × D, therefore this
Invention is also designed according to this thinking.It as shown in Figure 3D, is a signal of cost calculation optimization in the embodiment of the present invention
Figure.
Since the space that traditional cost calculates is W × H × D, and parallax cost of some pixel under some direction its
Real is also a pixel cost, therefore traditional cost calculates and realizes will there is a large amount of memory redundancy.In order to be promoted to greatest extent
Data duplication utilization rate reduces data storage capacity, and there is no allow per thread all to handle one as census transformation by the present invention
A pixel, but allow each blocks processing one-row pixels utilizes for circulation in block gradually to handle other in a line
Whole parallaxes of pixel, each pixel are disposably obtained by threads.I.e. each grid will distribute H blocks, each
Blocks distributes D threads, and for circulation in block will repeat W times.
Illustratively, 480 block can be designed, design 128 threads in each block.In this way each
Blocks is responsible for the pixel that a line handles a line.In this case, the corresponding variable data structure of the present invention is also different from tradition side
Case, the array Shared_base length for storing benchmark pixel is D, and stores the array Shared_match long of pixel to be compared
Degree is 2D, and shown in the following Fig. 3 E of data structure, Fig. 3 E is the array Shared_base that benchmark pixel is stored in the embodiment of the present invention
The schematic diagram of length.And then the cost of 128 pixels of the last period of the previous D storage of Shared_match, the latter D are stored
Next section of 128 pixels cost.
203, it is polymerize according to the synchronous cost for carrying out first direction, second direction and third direction of the cost value and is calculated,
Obtain the cost polymerizing value of the cost polymerizing value of first direction, the cost polymerizing value of second direction and third direction.
It is described that calculating is polymerize according to the synchronous cost for carrying out first direction, second direction and third direction of the cost value,
The cost polymerizing value of the cost polymerizing value of first direction, the cost polymerizing value of second direction and third direction is obtained, may include:
It synchronizes by SGM binocular image matching algorithm according to the cost value and carries out first direction, second direction and third direction
Cost polymerization calculates, and the cost for obtaining the cost polymerizing value of first direction, the cost polymerizing value of second direction and third direction is poly-
Conjunction value.
It should be noted that polymerizeing for cost, polymerization theory of the invention uses SGM (semi-global
Maching) binocular image matching algorithm needs point multiple directions to carry out cost polymerization in this process.Therefore, in order into
One step promotes operation efficiency, and a plurality of stream can will be used to participate in cost and polymerize.Illustratively, in this example, only gather
The cost of four direction is closed, so the cost polymerizing value of ' → ' is handled by stream0, the cost polymerizing value of ' ↓ ' is by stream1
Reason, the cost polymerizing value of ' → ' are handled by stream2.But consider for data user rate is increased, the present invention is not by ' ↑ '
Cost polymerizing value be divided into stream3, but stream1 is used, because in order to improve memory usage, in progress ' ↑ ' number side
To while calculating, the processor active task of the optimal parallax of calculating will be completed, it is therefore necessary to ensure that the operation in other directions all carries out
Step operation could be executed by finishing, and the task in the same stream follows the principle executed in order, therefore is formed 3
A direction parallel computation, the cost of four direction such as calculate at the mentality of designing for carrying out parallax optimization after they have been calculated again.
It is calculated since SGM algorithm needs to complete following formula:
Lr(p, d)=C (p, d)+min (Lr(p-r, d),
Lr(p-r, d-1)+P1
Lr(p-r, d+1)+P1,
Wherein, in above-mentioned formula, meaning indicated by parameters is as follows:
Lr (p, d): the cost polymerizing value in the direction r of some match point P；
C (p, d): the matching cost value of some match point P；
Lr (p-r, d): the matching cost polymerizing value on some match point under the same disparity of a match point；
Lr (p-r, d-1): the parallax of a match point subtracts matching cost polymerizing value on some match point；
Lr (p-r, d+1): the parallax of a match point adds matching cost polymerizing value on some match point；
MinLr (p-r, i): cost polymerization on some match point under all parallaxes of a match point minimum value
(formula of back k is same)；
P1, P2: adjustable parameters compensate parallax, for being finely adjusted to algorithm.
The cost of one pixel needs to be related to neighbor pixel, needs to compare the generation of the adjacent parallax of neighbor pixel
Valence, the minimum value of the cost of whole parallaxes of the current parallax cost and neighbor pixel of neighbor pixel, therefore deposit here
In more data reusings and data interaction, and traditional scheme equally allows per thread to handle a parallax cost, can because
Data communication between different threads and reduce parallel operational efficiency, the invention proposes a kind of more efficient processing scheme, needles
To different directions, there is different architecture design strategies, by taking ' → ' direction as an example, layout strategy can be with reference to shown in Fig. 3 C.
204, it is polymerize according to the cost that the cost value carries out fourth direction and is calculated, obtains the cost polymerization of fourth direction
Value.
The cost for carrying out fourth direction according to the cost value, which polymerize, to be calculated, and the cost polymerization of fourth direction is obtained
Value may include:, by SGM binocular image matching algorithm, to carry out the cost polymerization meter of fourth direction according to the cost value
It calculates, obtains the cost polymerizing value of fourth direction.
205, according to the cost polymerizing value of the first direction, the cost polymerizing value of the second direction, the third party
To cost polymerizing value and the fourth direction cost polymerizing value, determine parallax value.
The cost polymerizing value according to the first direction, the cost polymerizing value of the second direction, the third party
To cost polymerizing value and the fourth direction cost polymerizing value, determine parallax value, may include: by the first direction
Cost polymerizing value, the cost polymerizing value of the second direction, the cost polymerizing value of the third direction and the fourth direction
Cost polymerizing value adds up in the case where different parallax values, obtains cumulative polymerizing value corresponding from the different parallax values；
By butterfly sort algorithm, the minimum value in determining cumulative polymerizing value corresponding with the different parallax values is parallax value.
It is understood that the optimal parallax of two pixels of each Blocks design treatment, and in order to promote data interaction
Efficiency, using the thread in the same warp can directly mutual shared data principle, by a pixel direction
The calculating of upper all parallax sizes is completed with a warp.And TX2 platform Pascal (pascall) framework is next at present
Warp maximum includes 32 thread, and the calculating of a pixel whole parallax cost is completed in present invention design with a warp, therefore
It handles two pixels and needs two warp in total, therefore include 64 threads in a blocks.In order to meet through 32 lines
Journey calculates whole parallax costs, designs per thread here and handles MAX_DISPARITY/32 parallax.Whole thinking will
It is calculated similar to cost, each block handles two row pixels, and other in a line are gradually handled using for circulation in block
Pixel.Therefore an one-dimensional grid will be designed, each grid includes the blocks of rows/2.With the image procossing of 640*480
For, this grid will include 480/2=239 blocks, and each blocks will include 64 threads, per thread processing
The cost of 128/32=4 parallax.
According to SGM algorithm, need to calculateAnd d*=mindThe value of S (p, d).Conventional method is to pass through
Common sort method, to obtain minimum value, such as bubbling method, at least needs to arrange n × (n-1)/2 in several inter-group orderings.
And for GPU, it can achieve higher efficiency, the present invention devises a kind of butterfly sort algorithm, using same
Thread in warp can directly mutual shared data principle, pass through _ shuf_xor_sync instruction, significant less sequence
Number.That value carries out data exchange between adjacent thread first, by the legacy data in script thread with new data progress size compared with,
Realization finds out data maximums between finding out adjacent thread, and such data volume will reduce half；Then, be interval 1 cross-thread into
Row data exchange compares by this wheel, realizes that the minimum value for finding out continuous 4 cross-threads, such data volume will reduce half again；
Continue to compare, to the last until the data reduction of 32 threads to 1 data.This method can reduce number of comparisons,
Computational complexity is reduced, by this butterfly sort operation, is shifted every time, the negligible amounts half for needing to compare, therefore
It only needs to compareIt is secondary, significantly less number of comparisons.Fig. 3 F is one of butterfly sort algorithm in the embodiment of the present invention
A schematic diagram.
Illustratively, judge the minimum disparity correspondence cost in 128 parallax points.Due to thread process 4 views
The matching cost of difference, therefore first determine whether the minimum value for comparing 4 parallaxes, need to carry out 3 comparisons in total.After this step,
128 values will narrow down to 32 values and need to compare size, this 32 values are respectively in 32 different threads.Then, to line
Journey uses shfl_xor_sync (val, 2), exactly swaps across 1 grid, i.e. the value of thread 2 has been placed to thread 0
Position similarly allows them to carry out size comparison, obtains the minimum value of thread 0,1,2,3, similarly other threads, needs in this way
The val value compared is reduced half again, is left 8 val values.Similarly, it is shifted 8 times in displacement 4 times later, displacement 16 times, gradually
The size of completeer 32 threads.
In embodiments of the present invention, depth optimization has been carried out for binocular vision matching algorithm, has proposed a series of be based on
The optimizing structure design scheme of GPU.It proposes aiming at the problem that SGM algorithm finds smallest match cost, proposes a kind of butterfly
Sort algorithm only needs operationIt is secondary to find out optimal parallax.Increase the utilization rate that cost calculates, the present invention redesigns
The data structure of cost, can protect guarantee data reusing to the greatest extent.
The embodiment of the invention provides it is a kind of new based on graphics processor (Graphics Processing Unit,
GPU binocular vision image matching method), can greatly promote operation efficiency, under the premise of matching precision is constant, reduce
Handle the time.The operation efficiency of image matching algorithm in binocular vision can be promoted, the reality of binocular depth cognition technology is improved
Shi Xing.
As shown in figure 4, may include: for one embodiment schematic diagram of processing unit in the embodiment of the present invention
Module 401 is obtained, for obtaining the first image data and second picture data, first image data and described
What second picture data were obtained by different cameras respectively；
Processing module 402, for carrying out cost calculating according to first image data and the second picture data,
Obtain cost value；It is polymerize according to the synchronous cost for carrying out first direction, second direction and third direction of the cost value and is calculated, is obtained
Cost polymerizing value, the cost polymerizing value of the cost polymerizing value of second direction and third direction to first direction；According to the generation
The cost that value carries out fourth direction, which polymerize, to be calculated, and the cost polymerizing value of fourth direction is obtained；According to the generation of the first direction
The generation of valence polymerizing value, the cost polymerizing value of the second direction, the cost polymerizing value of the third direction and the fourth direction
Valence polymerizing value, determines parallax value.
Optionally, in some embodiments of the invention,
Processing module 402 is specifically used for passing through different rows according to first image data and the second picture data
The block of cloth carries out cost calculating, obtains cost value, wherein one pixel of per thread alignment processing in each block.
Optionally, in some embodiments of the invention,
Processing module 402 is specifically used for according to the cost value, synchronous to carry out the by SGM binocular image matching algorithm
The cost in one direction, second direction and third direction, which polymerize, to be calculated, and the cost polymerizing value of first direction, the generation of second direction are obtained
The cost polymerizing value of valence polymerizing value and third direction；According to the cost value, by SGM binocular image matching algorithm, the is carried out
The cost in four directions, which polymerize, to be calculated, and the cost polymerizing value of fourth direction is obtained.
Optionally, in some embodiments of the invention,
Processing module 402 is specifically used for determining the cost value of first direction, the generation of second direction according to the cost value
Value, the cost value of the cost value of third direction and fourth direction；According to the cost value of the first direction, the second direction
Cost value and the third direction cost value, it is synchronous to carry out first direction, second direction and the by butterfly sort algorithm
The cost in three directions, which polymerize, to be calculated, and the cost polymerizing value of first direction, the cost polymerizing value of second direction and third direction are obtained
Cost polymerizing value；The cost polymerization of fourth direction is carried out by butterfly sort algorithm according to the cost value of the fourth direction
It calculates, obtains the cost polymerizing value of fourth direction.
Optionally, in some embodiments of the invention,
Processing module 402, specifically for the cost of the cost polymerizing value of the first direction, the second direction to polymerize
Value, the cost polymerizing value of the third direction and the fourth direction cost polymerizing value, in the case where different parallax values into
Row is cumulative, obtains cumulative polymerizing value corresponding from the different parallax values；By butterfly sort algorithm, determine and the different views
Minimum value in the corresponding cumulative polymerizing value of difference is parallax value.
As shown in figure 5, may include: for one embodiment schematic diagram of processing unit in the embodiment of the present invention
Transceiver 501, processor 502, memory 503, wherein transceiver 501, processor 502 and memory 503 pass through
Bus connection；It is understood that transceiver 501 can be image capture device.
Memory 503, for storing operational order；
Transceiver 501, for obtaining the first image data and second picture data, first image data and described the
What two image datas were obtained by different cameras respectively；
Processor 502, for calling the operational order, execution following steps:
Obtain the first image data and second picture data, first image data and second picture data difference
It is obtained by different cameras；
According to first image data and the second picture data, cost calculating is carried out, cost value is obtained；
It is polymerize according to the synchronous cost for carrying out first direction, second direction and third direction of the cost value and is calculated, is obtained
The cost polymerizing value of the cost polymerizing value of first direction, the cost polymerizing value of second direction and third direction；
It is polymerize according to the cost that the cost value carries out fourth direction and is calculated, obtains the cost polymerizing value of fourth direction；
According to the cost polymerizing value of the first direction, the cost polymerizing value of the second direction, the third direction
The cost polymerizing value of cost polymerizing value and the fourth direction, determines parallax value.
Optionally, in some embodiments of the invention, processor 502 execute as follows for calling the operational order
Step:
According to first image data and the second picture data, cost meter is carried out by the block of different arrangements
It calculates, obtains cost value, wherein one pixel of per thread alignment processing in each block.
Optionally, in some embodiments of the invention, processor 502 execute as follows for calling the operational order
Step:
It is synchronous to carry out first direction, second direction and the by SGM binocular image matching algorithm according to the cost value
The cost in three directions, which polymerize, to be calculated, and the cost polymerizing value of first direction, the cost polymerizing value of second direction and third direction are obtained
Cost polymerizing value；
According to the cost value, by SGM binocular image matching algorithm, the cost polymerization for carrying out fourth direction is calculated, is obtained
To the cost polymerizing value of fourth direction.
Optionally, in some embodiments of the invention, processor 502 execute as follows for calling the operational order
Step:
According to the cost value, the cost value of first direction, the cost value of second direction, the cost value of third direction are determined
With the cost value of fourth direction；
According to the cost value of the first direction, the cost value of the cost value of the second direction and the third direction,
By butterfly sort algorithm, the synchronous cost for carrying out first direction, second direction and third direction, which polymerize, to be calculated, and obtains first party
To cost polymerizing value, the cost polymerizing value of the cost polymerizing value of second direction and third direction；
According to the cost value of the fourth direction, by butterfly sort algorithm, the cost polymerization for carrying out fourth direction is calculated,
Obtain the cost polymerizing value of fourth direction.
Optionally, in some embodiments of the invention, processor 502 execute as follows for calling the operational order
Step:
By the cost polymerizing value of the first direction, the cost polymerizing value of the second direction, the third direction generation
The cost polymerizing value of valence polymerizing value and the fourth direction, adds up in the case where different parallax values, obtain with it is described not
The corresponding cumulative polymerizing value with parallax value；By butterfly sort algorithm, determining cumulative polymerization corresponding from the different parallax values
Minimum value in value is parallax value.
In the above-described embodiments, can come wholly or partly by software, hardware, firmware or any combination thereof real
It is existing.When implemented in software, it can entirely or partly realize in the form of a computer program product.
The computer program product includes one or more computer instructions.Load and execute on computers the meter
When calculation machine program instruction, entirely or partly generate according to process or function described in the embodiment of the present invention.The computer can
To be general purpose computer, special purpose computer, computer network or other programmable devices.The computer instruction can be deposited
Storage in a computer-readable storage medium, or from a computer readable storage medium to another computer readable storage medium
Transmission, for example, the computer instruction can pass through wired (example from a web-site, computer, server or data center
Such as coaxial cable, optical fiber, Digital Subscriber Line (DSL)) or wireless (such as infrared, wireless, microwave) mode to another website
Website, computer, server or data center are transmitted.The computer readable storage medium can be computer and can deposit
Any usable medium of storage either includes that the data storages such as one or more usable mediums integrated server, data center are set
It is standby.The usable medium can be magnetic medium, (for example, floppy disk, hard disk, tape), optical medium (for example, DVD) or partly lead
Body medium (such as solid state hard disk Solid State Disk (SSD)) etc..
It is apparent to those skilled in the art that for convenience and simplicity of description, the system of foregoing description,
The specific work process of device and unit, can refer to corresponding processes in the foregoing method embodiment, and details are not described herein.
In several embodiments provided by the present invention, it should be understood that disclosed system, device and method can be with
It realizes by another way.For example, the apparatus embodiments described above are merely exemplary, for example, the unit
It divides, only a kind of logical function partition, there may be another division manner in actual implementation, such as multiple units or components
It can be combined or can be integrated into another system, or some features can be ignored or not executed.Another point, it is shown or
The mutual coupling, direct-coupling or communication connection discussed can be through some interfaces, the indirect coupling of device or unit
It closes or communicates to connect, can be electrical property, mechanical or other forms.
The unit as illustrated by the separation member may or may not be physically separated, aobvious as unit
The component shown may or may not be physical unit, it can and it is in one place, or may be distributed over multiple
In network unit.It can select some or all of unit therein according to the actual needs to realize the mesh of this embodiment scheme
's.
It, can also be in addition, the functional units in various embodiments of the present invention may be integrated into one processing unit
It is that each unit physically exists alone, can also be integrated in one unit with two or more units.Above-mentioned integrated list
Member both can take the form of hardware realization, can also realize in the form of software functional units.
If the integrated unit is realized in the form of SFU software functional unit and sells or use as independent product
When, it can store in a computer readable storage medium.Based on this understanding, technical solution of the present invention is substantially
The all or part of the part that contributes to existing technology or the technical solution can be in the form of software products in other words
It embodies, which is stored in a storage medium, including some instructions are used so that a computer
Equipment (can be personal computer, server or the network equipment etc.) executes the complete of each embodiment the method for the present invention
Portion or part steps.And storage medium above-mentioned includes: USB flash disk, mobile hard disk, read-only memory (ROM, Read-Only
Memory), random access memory (RAM, Random Access Memory), magnetic or disk etc. are various can store journey
The medium of sequence code.
The above, the above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations；Although referring to before
Stating embodiment, invention is explained in detail, those skilled in the art should understand that: it still can be to preceding
Technical solution documented by each embodiment is stated to modify or equivalent replacement of some of the technical features；And these
It modifies or replaces, the spirit and scope for technical solution of various embodiments of the present invention that it does not separate the essence of the corresponding technical solution.
Claims (10)
1. a kind of method of the binocular ranging algorithm based on GPU characterized by comprising
The first image data and second picture data are obtained, first image data and the second picture data are not respectively by
What same camera obtained；
According to first image data and the second picture data, cost calculating is carried out, cost value is obtained；
It is polymerize according to the synchronous cost for carrying out first direction, second direction and third direction of the cost value and is calculated, obtains first
The cost polymerizing value of the cost polymerizing value in direction, the cost polymerizing value of second direction and third direction；
It is polymerize according to the cost that the cost value carries out fourth direction and is calculated, obtains the cost polymerizing value of fourth direction；
According to the cost polymerizing value of the first direction, the cost polymerizing value of the second direction, the cost of the third direction
The cost polymerizing value of polymerizing value and the fourth direction, determines parallax value.
2. the method according to claim 1, wherein described according to first image data and second figure
Sheet data carries out cost calculating, obtains cost value, comprising:
According to first image data and the second picture data, cost calculating is carried out by the block of different arrangements, is obtained
To cost value, wherein one pixel of per thread alignment processing in each block.
3. the method according to claim 1, wherein it is described according to the cost value it is synchronous carry out first direction,
Second direction and the polymerization of the cost of third direction calculate, and obtain cost polymerizing value, the polymerization of the cost of second direction of first direction
The cost polymerizing value of value and third direction, comprising:
It is synchronous to carry out first direction, second direction and third party by SGM binocular image matching algorithm according to the cost value
To cost polymerize calculate, obtain the cost polymerizing value, the cost polymerizing value of second direction and the generation of third direction of first direction
Valence polymerizing value；
The cost for carrying out fourth direction according to the cost value, which polymerize, to be calculated, and is obtained the cost polymerizing value of fourth direction, is wrapped
It includes:
According to the cost value, by SGM binocular image matching algorithm, the cost polymerization for carrying out fourth direction is calculated, and obtains the
The cost polymerizing value in four directions.
4. method according to any one of claim 1-3, which is characterized in that described to be carried out according to the cost value is synchronous
The cost of first direction, second direction and third direction, which polymerize, to be calculated, and the cost polymerizing value of first direction, second direction are obtained
The cost polymerizing value of cost polymerizing value and third direction, the cost for carrying out fourth direction according to the cost value polymerize meter
It calculates, obtains the cost polymerizing value of fourth direction, comprising:
According to the cost value, the cost value of first direction, the cost value of second direction, the cost value of third direction and are determined
The cost value in four directions；
According to the cost value of the first direction, the cost value of the cost value of the second direction and the third direction, pass through
Butterfly sort algorithm, the synchronous cost for carrying out first direction, second direction and third direction, which polymerize, to be calculated, and obtains first direction
The cost polymerizing value of cost polymerizing value, the cost polymerizing value of second direction and third direction；
According to the cost value of the fourth direction, by butterfly sort algorithm, the cost polymerization for carrying out fourth direction is calculated, is obtained
The cost polymerizing value of fourth direction.
5. method according to any one of claim 1-3, which is characterized in that the cost according to the first direction
Polymerizing value, the cost polymerizing value of the second direction, the cost of the cost polymerizing value of the third direction and the fourth direction
Polymerizing value determines parallax value, comprising:
The cost polymerizing value of the first direction, the cost polymerizing value of the second direction, the cost of the third direction are gathered
The cost polymerizing value of conjunction value and the fourth direction, adds up in the case where different parallax values, obtains and the different views
The corresponding cumulative polymerizing value of difference；
By butterfly sort algorithm, the minimum value in determining cumulative polymerizing value corresponding with the different parallax values is parallax value.
6. a kind of processing unit characterized by comprising
Module is obtained, for obtaining the first image data and second picture data, first image data and second figure
What sheet data was obtained by different cameras respectively；
Processing module, for carrying out cost calculating, obtaining cost according to first image data and the second picture data
Value；It is polymerize according to the synchronous cost for carrying out first direction, second direction and third direction of the cost value and is calculated, obtains first party
To cost polymerizing value, the cost polymerizing value of the cost polymerizing value of second direction and third direction；It is carried out according to the cost value
The cost of fourth direction, which polymerize, to be calculated, and the cost polymerizing value of fourth direction is obtained；According to the cost polymerizing value of the first direction,
The cost polymerizing value of the cost polymerizing value of the second direction, the cost polymerizing value of the third direction and the fourth direction,
Determine parallax value.
7. processing unit according to claim 6, which is characterized in that
The processing module is specifically used for passing through different arrangements according to first image data and the second picture data
Block carry out cost calculating, obtain cost value, wherein one pixel of per thread alignment processing in each block；
The processing module is specifically used for according to the cost value, synchronous to carry out first by SGM binocular image matching algorithm
The cost in direction, second direction and third direction, which polymerize, to be calculated, and cost polymerizing value, the cost of second direction of first direction are obtained
The cost polymerizing value of polymerizing value and third direction；The 4th is carried out by SGM binocular image matching algorithm according to the cost value
The cost in direction, which polymerize, to be calculated, and the cost polymerizing value of fourth direction is obtained.
8. processing unit according to claim 6 or 7, which is characterized in that
The processing module is specifically used for determining the cost value of first direction, the cost of second direction according to the cost value
Value, the cost value of the cost value of third direction and fourth direction；According to the cost value of the first direction, the second direction
The cost value of cost value and the third direction, it is synchronous to carry out first direction, second direction and third by butterfly sort algorithm
The cost in direction, which polymerize, to be calculated, and the cost polymerizing value of first direction, the cost polymerizing value of second direction and third direction are obtained
Cost polymerizing value；The cost polymerization meter of fourth direction is carried out by butterfly sort algorithm according to the cost value of the fourth direction
It calculates, obtains the cost polymerizing value of fourth direction；
The processing module, specifically for by the cost polymerizing value of the cost polymerizing value of the first direction, the second direction,
The cost polymerizing value of the third direction and the cost polymerizing value of the fourth direction, carry out tired in the case where different parallax values
Add, obtains cumulative polymerizing value corresponding from the different parallax values；By butterfly sort algorithm, determine and the different parallax values
Minimum value in corresponding cumulative polymerizing value is parallax value.
9. a kind of processing unit characterized by comprising
Transceiver, processor, memory, wherein the transceiver, the processor and the memory are connected by bus；
The memory, for storing operational order；
The transceiver, for obtaining the first image data and second picture data, first image data and described second
What image data was obtained by different cameras respectively；
The processor executes according to any one of claims 1 to 5 based on GPU's for calling the operational order
The step of method of binocular ranging algorithm.
10. a kind of readable storage medium storing program for executing, is stored thereon with computer program, which is characterized in that the computer program is processed
The step of device realizes the method for the binocular ranging algorithm according to any one of claims 1 to 5 based on GPU when executing.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910779546.4A CN110533710B (en) | 2019-08-22 | 2019-08-22 | Method and processing device for binocular matching algorithm based on GPU |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910779546.4A CN110533710B (en) | 2019-08-22 | 2019-08-22 | Method and processing device for binocular matching algorithm based on GPU |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110533710A true CN110533710A (en) | 2019-12-03 |
| CN110533710B CN110533710B (en) | 2023-07-14 |
Family
ID=68664126
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910779546.4A Active CN110533710B (en) | 2019-08-22 | 2019-08-22 | Method and processing device for binocular matching algorithm based on GPU |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110533710B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112256431A (en) * | 2020-10-23 | 2021-01-22 | 展讯通信(天津)有限公司 | Cost aggregation method and device, storage medium and terminal |
| CN112823378A (en) * | 2020-07-03 | 2021-05-18 | 深圳市大疆创新科技有限公司 | Image depth information determination method, device, equipment and storage medium |
| CN112889030A (en) * | 2020-07-03 | 2021-06-01 | 深圳市大疆创新科技有限公司 | Image processing method, integrated circuit, device and equipment |
| CN115063619A (en) * | 2022-08-18 | 2022-09-16 | 北京中科慧眼科技有限公司 | Cost aggregation method and system based on binocular stereo matching algorithm |
Citations (6)
| 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 |
| EP2860695A1 (en) * | 2013-10-14 | 2015-04-15 | Ricoh Company, Ltd. | Method and apparatus for identifying noise in disparity depth image |
| CN107481271A (en) * | 2017-07-25 | 2017-12-15 | 成都通甲优博科技有限责任公司 | A kind of solid matching method, system and mobile terminal |
| WO2018086348A1 (en) * | 2016-11-09 | 2018-05-17 | 人加智能机器人技术（北京）有限公司 | Binocular stereo vision system and depth measurement method |
| CN108174176A (en) * | 2017-12-22 | 2018-06-15 | 洛阳中科众创空间科技有限公司 | A kind of high-precision disparity computation accelerated method based on GPU |
| CN109978934A (en) * | 2019-03-04 | 2019-07-05 | 北京大学深圳研究生院 | A kind of binocular vision solid matching method and system based on matching cost weighting |
-
2019
- 2019-08-22 CN CN201910779546.4A patent/CN110533710B/en active Active
Patent Citations (6)
| 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 |
| EP2860695A1 (en) * | 2013-10-14 | 2015-04-15 | Ricoh Company, Ltd. | Method and apparatus for identifying noise in disparity depth image |
| WO2018086348A1 (en) * | 2016-11-09 | 2018-05-17 | 人加智能机器人技术（北京）有限公司 | Binocular stereo vision system and depth measurement method |
| CN107481271A (en) * | 2017-07-25 | 2017-12-15 | 成都通甲优博科技有限责任公司 | A kind of solid matching method, system and mobile terminal |
| CN108174176A (en) * | 2017-12-22 | 2018-06-15 | 洛阳中科众创空间科技有限公司 | A kind of high-precision disparity computation accelerated method based on GPU |
| CN109978934A (en) * | 2019-03-04 | 2019-07-05 | 北京大学深圳研究生院 | A kind of binocular vision solid matching method and system based on matching cost weighting |
Non-Patent Citations (3)
| Title |
|---|
| D. HERNANDEZ-JUAREZ: "Embedded real-time stereo estimation via Semi-Global Matching on the GPU", 《INTERNATIONAL CONFERENCE ON COMPUTATIONAL SCIENCE 2016, ICCS 2016》 * |
| D. HERNANDEZ-JUAREZ: "Embedded real-time stereo estimation via Semi-Global Matching on the GPU", 《INTERNATIONAL CONFERENCE ON COMPUTATIONAL SCIENCE 2016, ICCS 2016》, 1 June 2016 (2016-06-01), pages 143 - 153, XP029565684, DOI: 10.1016/j.procs.2016.05.305 * |
| 符立梅;彭国华;: "基于自适应阻尼因子渗透滤波器的匹配代价聚合算法", 计算机辅助设计与图形学学报, no. 05, pages 167 - 174 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112823378A (en) * | 2020-07-03 | 2021-05-18 | 深圳市大疆创新科技有限公司 | Image depth information determination method, device, equipment and storage medium |
| CN112889030A (en) * | 2020-07-03 | 2021-06-01 | 深圳市大疆创新科技有限公司 | Image processing method, integrated circuit, device and equipment |
| WO2022000456A1 (en) * | 2020-07-03 | 2022-01-06 | 深圳市大疆创新科技有限公司 | Image processing method and apparatus, integrated circuit, and device |
| WO2022000458A1 (en) * | 2020-07-03 | 2022-01-06 | 深圳市大疆创新科技有限公司 | Image depth information determination method and apparatus, device and storage medium |
| CN112256431A (en) * | 2020-10-23 | 2021-01-22 | 展讯通信(天津)有限公司 | Cost aggregation method and device, storage medium and terminal |
| CN115063619A (en) * | 2022-08-18 | 2022-09-16 | 北京中科慧眼科技有限公司 | Cost aggregation method and system based on binocular stereo matching algorithm |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110533710B (en) | 2023-07-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110533710A (en) | A kind of method and processing unit of the binocular ranging algorithm based on GPU | |
| Huang et al. | M3VSNet: Unsupervised multi-metric multi-view stereo network | |
| US8330763B2 (en) | Apparatus and method for volume rendering on multiple graphics processing units (GPUs) | |
| Zlateski et al. | ZNN--A Fast and Scalable Algorithm for Training 3D Convolutional Networks on Multi-core and Many-Core Shared Memory Machines | |
| AU2017338783A1 (en) | Efficient data layouts for convolutional neural networks | |
| CN110197270A (en) | Integrated circuit chip device and Related product | |
| CN103414853A (en) | Device and method for stabilizing video image sequence capable of doing multi-degree of freedom movement in real time | |
| CN110458957A (en) | A kind of three-dimensional image model construction method neural network based and device | |
| Ummenhofer et al. | Point-based 3D reconstruction of thin objects | |
| CN110335344A (en) | Three-dimensional rebuilding method based on 2D-3D attention mechanism neural network model | |
| CN104240229A (en) | Self-adaptation polarline correcting method based on infrared binocular camera | |
| CN110310220A (en) | A kind of half global real-time volume matching process | |
| CN109801325A (en) | A kind of Binocular Stereo Vision System obtains the method and device of disparity map | |
| Chang et al. | Efficient stereo matching on embedded GPUs with zero-means cross correlation | |
| KR102586173B1 (en) | Processor and control methods thererof | |
| CN106484532B (en) | GPGPU parallel calculating method towards SPH fluid simulation | |
| Liu et al. | When epipolar constraint meets non-local operators in multi-view stereo | |
| Happ et al. | A parallel image segmentation algorithm on GPUs | |
| Goulermas et al. | A collective-based adaptive symbiotic model for surface reconstruction in area-based stereo | |
| CN107316324A (en) | Method based on the CUDA real-time volume matchings realized and optimization | |
| CN115221103A (en) | Computing device, data processing method and related product | |
| TWI787430B (en) | Integrated circuit chip apparatus, chip, electronic device, and computing method of neural network | |
| CN103400390B (en) | The hardware acceleration structure of variable supporting zone Stereo Matching Algorithm | |
| Chang et al. | A GPU accelerator for domain transformation-based stereo matching | |
| Zhang et al. | A near real-time color stereo matching method for GPU |
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 |