CN106408609B - A kind of parallel institution end movement position and posture detection method based on binocular vision - Google Patents
A kind of parallel institution end movement position and posture detection method based on binocular vision Download PDFInfo
- Publication number
- CN106408609B CN106408609B CN201610819289.9A CN201610819289A CN106408609B CN 106408609 B CN106408609 B CN 106408609B CN 201610819289 A CN201610819289 A CN 201610819289A CN 106408609 B CN106408609 B CN 106408609B
- Authority
- CN
- China
- Prior art keywords
- parallel institution
- point
- matching
- image
- algorithm
- 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.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
-
- 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
- G06T2207/20021—Dividing image into blocks, subimages or windows
-
- 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
- G06T2207/20172—Image enhancement details
- G06T2207/20182—Noise reduction or smoothing in the temporal domain; Spatio-temporal filtering
Abstract
The invention discloses a kind of parallel institution end movement position and posture detection method based on binocular vision, firstly, collected parallel institution image is carried out the pretreatment of the image denoising based on wavelet transformation；Then, characteristic matching is carried out to parallel institution image using Harris-SIFT algorithm.The matching algorithm passes through Harris operator extraction image characteristic point first, and the characteristic point for recycling SIFT feature description to go out Harris operator extraction matches；Then, propose that a kind of new purification algorithm purifies matching result, which is taken by piecemeal a little and in advance takes a method for checking computations temporary pattern；Finally, the parallel institution End features point after matching purification is found out mechanism end three-dimensional pose by coordinate transform to binocular vision model is brought into.Image processing time can be greatly lowered in the present invention, and the another new purification algorithm by being proposed purifies matching result, further increase matching accuracy, so that the real-time and precision of the detection of parallel institution end pose are all improved.
Description
Technical field
The present invention relates to the pose detection systems based on machine vision, more particularly to are based on binocular vision, are directed to parallel machine
The detection method of structure end movement pose.
Background technique
In parallel institution control, end movement pose is the important parameter for reflecting mechanism kinematic state, accurately measures machine
The end movement pose of structure, which can effectively avoid, resolves brought error by kinematics model, this will be conducive to realize parallel institution
High performance control.Relative to other detection means, machine vision has many advantages, such as non-contact, strong applicability, high performance-price ratio, especially
It is suitable for having, movement multiple degrees of freedom, motion profile is complicated, is difficult to the parallel institution end movement pose directly detected detection.
Vision pose detection system is commonly divided into monocular, binocular and more visual inspection examining systems.Due to the robustness of binocular vision algorithm
It is better than monocular algorithm with accuracy, and the real-time of its algorithm has great advantages relative to multi-vision visual, therefore, binocular
Vision has obtained relatively broad application.For the end movement pose test problems of parallel institution, using based on binocular vision
How that the collected parallel institution image of vision system is quick and accurate after image procossing the difficult point of detection method is
Ground obtains the end posture information of the mechanism.In binocular vision pose detection process, Stereo matching process is mostly important, matching
Speed and precision directly determine the speed and precision of entire pose detection system, but since binocular vision algorithm is complicated, illumination
The extraneous factors such as variation, the reflective, noise jamming of background can cause the images match of binocular vision system to be difficult to accurately realize, cause
The speed and precision of vision detection system is affected.
Document " a kind of novel parallel robot pose stereo vision detection system " (Wu Difei, computer such as fourth immortality
Engineering and application, 2007,43 (33): 190-192) establish a kind of binocular vision for being based on Scale invariant features transform (SIFT)
Detection system carries out pose detection to parallel robot, and with the robustness of the simulation results show detection algorithm and feasible
Property, but the SIFT algorithm as used by the detection system matching process can generate the biggish feature vector of a large amount of dimensions, increase
The complexity of algorithm, the real-time for causing pose to detect are affected；Document " Harris-SIFT algorithm and its vertical in binocular
Application in body vision " (the Journal of UEST of China such as Zhao Qinjun, Zhao Dongbiao, 2010,39 (4): 546-550) combination
Harris conspicuousness and SIFT description, propose that a kind of Harris-SIFT algorithm applies in binocular vision matching, and test
The validity of method is demonstrate,proved, this method effectively increases the real-time of matching algorithm, but the angle point institute detected by Harris
Characteristic point information contained is detected less than SIFT containing information, Harris-SIFT algorithm has erroneous matching and error matching.
For error hiding problem existing for Harris-SIFT algorithm, it is contemplated that carried out using purification algorithm to matching result
Purification.Conventional purification algorithm calculates the object module of matching double points, then benefit generally according to one group of data set comprising abnormal point
With left point come testing model, to obtain valid data sample, therefore, the key for purifying algorithm is to be quickly found out correctly
Object module.However, the acquisition of object module usually has some problems: when such as selection random sample, there are two candidates
It puts hypotelorism and is considered as a point, so as to cause the problem for the object module inaccuracy sought；And select every time with
Machine sample set will find its supported collection for corresponding to candidate family, this, which will lead to the searching of overspending time, exists compared with multiple error
It observes on the corresponding support point set of data set.
Summary of the invention
In view of above-mentioned binocular vision detection there are the problem of, the present invention propose it is a kind of based on binocular vision parallel institution end
End motion position and posture detection method, this method propose a kind of new matching algorithm, algorithm is real by quick Harris-SIFT algorithm
Existing Image Feature Point Matching, and propose that a kind of new purification algorithm purifies matching result, to solve traditional binocular view
Feel that algorithm is complicated, spatial match is difficult to accurately realize, the problem that caused detection real-time is poor, detection accuracy is not high.
The technical solution adopted by the present invention is that using following steps:
A kind of parallel institution end movement position and posture detection method based on binocular vision, includes the following steps:
Step 1) acquires the original image of parallel institution using binocular camera, and is based on to collected mechanism image
The image denoising of wavelet transformation pre-processes；
Step 2) extracts mechanism characteristics of image using the corner extraction based on Harris operator；
Step 3) carries out characteristic matching to parallel institution image using Harris-SIFT algorithm；The matching algorithm is logical first
Harris operator extraction image characteristic point is crossed, SIFT feature description is recycled to carry out the characteristic point that Harris operator extraction goes out
Matching；
There is erroneous matching and error matching Harris-SIFT algorithm in step 4), calculated using new purification
Method does further purification processes to matching result: it is taken by piecemeal and a checking computations temporary pattern a little and is in advance taken to improve purification algorithm,
Wherein, piecemeal takes a little: taking the mode for taking a match point in image block and each piece at random；The interim mould of checking computations is taken in advance
Type: randomly selecting 9 matchings pair, and with 8 matchings to finding out temporary pattern parameter, and the 9th matching is to being used to verify interim mould
The correctness of type quickly determines iconic model with this；
Step 5) extracts End features point by image procossing and calculates the End features point according to Binocular Vision Principle
Three-dimensional coordinate, and calculate parallel institution terminal angle angle, finally obtain the parallel institution end posture information.
Further, in the step 2), the corner extraction of Harris operator can call directly the realization of OpenCV library function,
In the detection of Harris characteristic point, the second moment of pixel be may be expressed as:
I in formulax、IyIndicate gradient of the gray scale in the direction x and y of pixel；IxyFor second order mixing local derviation；Pixel
Harris receptance function are as follows:
R=det M-k (traceM)2
Wherein det M is the determinant of matrix M, and traceM is the authentic work of matrix M；K is empirical, usual value range
0.04~0.06；Given threshold t calculates response R according to above-mentioned Harris operator, if R < t, which is characterized a little.
Further, in the step 4), the purification algorithm calculates matching by one group of data set comprising abnormal point
The object module of point pair recycles left point to carry out testing model, to obtain valid data sample；Specific step is as follows:
4.1) picture is divided by b × b block according to the coordinate system of match point in the first width figure, wherein b > 3, at random
It chooses in 9 mutually different piece and each piece and randomly selects a match point, constitute the random sample collection S of 9 matching double points；
4.2) 8 matching double points are randomly selected by random sample collection S and finds out interim candidate family F；
4.3) detect the 9th point to whether the supported collection for being interim candidate family F: if it is, this interim candidate family F
For candidate family D；Otherwise, reselect 9 pairs of match points, repeat 4.1), 4.2) process；
4.4) all matching double points are detected by candidate family D and threshold value k, obtains the supported collection logarithm m of candidate family D；
4.5) if m >=threshold value k, obtains object module G；Otherwise, reselect 9 pairs of match points, repeat 4.1), 4.2),
4.3), 4.4) process；
4.6) object module G is optimized, it is final to determine object module H；
4.7) selection random sample collection, counter all add 1 every time, if repeating not find model parameter threshold value k times, eventually
Only program.
Further, in the step 5), according to projection theorem find out parallel institution End features point three-dimensional coordinate (X, Y,
Z), then by camera rotation matrix R, parallel institution terminal angle angle is found out；Wherein parallel institution terminal position information is as follows:
Pw=(ATA)-1ATB
Wherein,(u1,
v1)、(u2,v2) End features point is respectively indicated in the image coordinate of left and right camera subpoint；Indicate the i-th row of left and right projection matrix, jth column；
Meanwhile parallel institution terminal angle angle are as follows:
Wherein, Rij(i=1,2,3；J=1,2,3) indicate that the i-th row of spin matrix R, jth column, α are to bow around what X-axis rotated
The elevation angle, unit are as follows: rad, β are the roll angle rotated around Y-axis, unit are as follows: rad, γ are the course angle rotated about the z axis, unit are as follows:
rad。
Further, in the step 5), further include, when three-dimensional reconstruction, it is necessary first to obtain built binocular vision model
Camera parameters, the camera parameters can be obtained by camera calibration；Using Zhang Zhengyou plane reference method to binocular vision system
System carries out camera calibration.
Further, the Zhang Zhengyou plane reference method process are as follows: then fixed camera first takes pictures to scaling board,
Scaling board image is inputted into computer again and obtains the pixel coordinate of each lattice point on scaling board, then known to scaling board three
It ties up coordinate and is loaded into computer, by the way that the pixel coordinate of scaling board and three-dimensional coordinate are substituted into peg model, solve video camera
Inside and outside parameter can obtain the video camera of binocular vision system further according to following formula by the calibration respectively of left and right cameras
Parameter:
In formula, Rl、TlAnd Rr、TrRespectively indicate the spin matrix and translation matrix of left and right cameras.
The present invention proposes a kind of parallel institution end movement position and posture detection method based on binocular vision, by using above-mentioned
After technical solution, have the advantages that
1, the present invention realizes Stereo matching using Harris-SIFT algorithm in the Stereo matching stage, simple by algorithm
Harris operator extraction image characteristic point recycles SIFT feature description to match image, so that matching result has both
Real-time and stability；
2, the present invention proposes that a kind of new purification is calculated for the matching of error existing for matching algorithm and erroneous matching problem
Method takes the more accurate iconic model parameter of acquisition by piecemeal and takes a checking computations temporary pattern with quick determining figure in advance
Purification algorithm is improved as the mode of model, to purify Harris-SIFT matching result, Harris-SIFT algorithm is rejected and deposits
Mismatching point improve matched accuracy while guaranteeing matching algorithm real-time, thus improve mechanism end three-dimensional
The precision of pose detection.
Detailed description of the invention
Below in conjunction with the drawings and specific embodiments, invention is further described in detail.
Fig. 1 is a kind of parallel institution end pose detection system flow chart based on binocular vision proposed by the invention.
Fig. 2 is using the parallel institution image denoising effect figure based on small wave converting method.Wherein: a is noisy figure
Picture, b are the image after denoising.
Fig. 3 is Harris feature point extraction figure.
Fig. 4 is the characteristic point Stereo matching effect picture based on Harris-SIFT algorithm.
Fig. 5 is to the effect picture after matching result purification.
Fig. 6 is the binocular vision imaging schematic diagram of the method for the present invention.
Fig. 7 is built upon position fixing system and the moving coordinate system of mechanism platform.
Fig. 8 is the method for the present invention experimental result and laser range finder and the mechanism end that electronic compass measurement result compares
Hold the tracking error figure of pose parameter.
Specific embodiment
The present invention provides a kind of parallel institution end position and posture detection method based on binocular vision, to solve binocular vision
The problem of matching process real-time difference, the present invention realize Stereo matching using Harris-SIFT algorithm；And it is directed to Harris-
There is erroneous matching and error matching in SIFT algorithm, the present invention proposes a kind of purification algorithm, to reject Harris-SIFT
Mismatching point existing for algorithm improves matched accuracy while guaranteeing matching algorithm real-time, so that pose is examined
The real-time and precision of survey are all improved.Specific step is as follows:
1, referring to Fig. 1, parallel institution original image is acquired by binocular vision or so camera.It is obtained in real time by binocular camera
Take parallel institution image information, camera should adjust suitable position to guarantee in parallel institution motion process, parallel institution
End is always within the scope of camera fields of view.
2, the collected original image of left and right cameras referring to Fig. 1 and Fig. 2, is subjected to the image denoising based on wavelet transformation
Pretreatment operation.Detection system Interface Development is carried out by using Visual Studio Integrated Development Tool, and uses C/
C Plus Plus programming, in combination with OpenCV (open-source cross-platform computer vision library), realizes the programming of each module correlation function.
Denoising process pre-processes original mechanism image using the method for image wavelet threshold denoising (WaveDec ()), by setting
It sets threshold value (g_HighPassFilterValue), then threshold process is carried out to high-frequency wavelet coefficient, small echo finally is carried out to image
It reconstructs (WaveRec ()), realizes image denoising.
3, referring to Fig. 1 and Fig. 3, mechanism picture characteristics is extracted using the corner extraction based on Harris operator.
In order to guarantee the rapidity and stability of feature extraction, it is contemplated that parallel institution image is rigid image and there are multiple angle points,
And Harris Corner Detection not only calculate it is simple, stablize, but also be not easy to be illuminated by the light, therefore the influence of rigid body geometric deformation uses
Harris Corner Detection Algorithm carries out feature point extraction to mechanism image.Harris feature point extraction can call directly the library OpenCV
((conerHarris ()) realizes that in the detection of Harris characteristic point, the second moment of pixel be may be expressed as: to function
I in formulax、IyIndicate gradient of the gray scale in the direction x and y of pixel；IxyFor second order mixing local derviation.Therefore, pixel
The Harris receptance function of point are as follows:
R=det M-k (traceM)2 (2)
Wherein det M is the determinant of matrix M, and traceM is the authentic work of matrix M；K is empirical, usual value range
0.04~0.06.Given threshold t calculates response R according to above-mentioned Harris operator, if R < t, which is characterized a little.
4, referring to Fig. 1 and Fig. 4, characteristic point Stereo matching is realized using Harris-SIFT algorithm.By using Harris spy
Sign point replaces the extreme point in SIFT, and defines principal direction for each characteristic point, generates feature vector description according to characteristic point,
Feature Points Matching is carried out further according to the Euclidean distance of SIFT matching algorithm, by Harris Corner Detection Algorithm and SIFT Scale invariant
Eigentransformation organically combines, and the reality of SIFT operator not only can be improved there are two types of the common trait of algorithm in the characteristic point tool made
Shi Xing, and the stability of characteristic point can be improved.Characteristic matching based on Harris-SIFT algorithm, which needs to program, to be realized, program
Following formula is used when writing:
Wherein: formula (3) is characterized Taylor expansion of the DoG operator in scale space of point x, D (x) representation space scale
Function, Δ x are characterized the offset of point x；Formula (4) is characterized the offset of an exact position；Formula (5) indicates non-edge point
Rule of judgment, r be setting threshold value, H be 2 × 2 Hessian matrix, Tr (H) be H-matrix authentic work, Det (H) be H square
The determinant of battle array；The Hessian matrix that formula (6) is 2 × 2, DxxIndicate the second order local derviation of the DOG operator of characteristic point in the x-direction
Number, DyyIndicate the second-order partial differential coefficient of the DOG operator of characteristic point in the y-direction, DxyIndicate the DOG operator of characteristic point along the direction xy
Second-order partial differential coefficient；Formula (7) indicates that the modulus value m (x, y) and direction θ (x, y) of Gauss gradient at characteristic point, L (x, y) indicate feature
Scale space value where point.
5, referring to Fig. 1 and Fig. 5, matching result is purified using new purification algorithm, rejects error hiding pair.For
There is erroneous matching and error matching in Harris-SIFT algorithm, the present invention is using new purification algorithm to matching result
Further purification processes are done, to reject Mismatching point existing for Harris-SIFT algorithm, are guaranteeing the same of matching algorithm real-time
When, matched accuracy is improved, to improve the precision of mechanism end three-dimensional pose detection.When carrying out purification processes, due to mentioning
The key of pure algorithm is to be quickly found out correct object module, and object module is by the candidate family most comprising interior points
Optimize and obtain, in order to quickly find out candidate family parameter, the present invention randomly selects 9 matchings pair, wherein 8 matchings are to being used to
Temporary pattern parameter is found out, and the 9th matches to the correctness for being used to verify temporary pattern, to quickly determine object module；
Meanwhile the present invention take take the mode of a match point to avoid in image block and each piece at random there are two match point distances
Cross close and be considered as a point, caused by fundamental matrix inaccuracy problem.Obtaining object module, specific step is as follows:
[1] picture is divided by b × b (b > 3) block according to the coordinate system of match point in the first width figure, randomly selects 9
A match point is randomly selected in a mutually different piece and each piece, constitutes the random sample collection S of 9 matching double points.
[2] 8 matching double points are randomly selected by random sample collection S and finds out interim candidate family F.
[3] detect the 9th point whether the supported collection for being interim candidate family F: be that then this interim candidate family F is candidate
Model D；Otherwise, 9 pairs of match points are reselected, [1], [2] process are repeated.
[4] all matching double points are detected by candidate family D and threshold value k, obtains the supported collection logarithm m of candidate family D.
[5] if m >=threshold value k, obtains object module G；Otherwise, reselect 9 pairs of match points, repeat [1], [2], [3],
[4] process.
[6] object module G is optimized, it is final to determine object module H.
[7] selection random sample collection, counter all add 1 every time, if repeating not find model parameter threshold value k times, eventually
Only program.
6, referring to Fig. 1 and Fig. 6, three-dimensional reconstruction is realized according to Binocular Vision Principle, by the end that early period, image procossing extracted
End matching double points bring binocular vision model into, final to obtain parallel institution end posture information.When three-dimensional reconstruction, it is necessary first to
To the camera parameters of built binocular vision model, which can be obtained by camera calibration, and the present invention, which uses, to open
Positive friend's standardization carries out camera calibration to binocular vision system, according to Zhang Zhengyou plane reference process: fixed camera first, so
It takes pictures afterwards to scaling board, then scaling board image is inputted into computer and obtains the pixel coordinate of each lattice point on scaling board,
Then the known three-dimensional coordinate of scaling board is loaded into computer, by the way that the pixel coordinate of scaling board and three-dimensional coordinate are substituted into calibration
Model solves the inside and outside parameter of video camera, by the calibration respectively of left and right cameras, can obtain binocular further according to formula (8)
The camera parameters of vision system:
In formula, Rl、TlAnd Rr、TrRespectively indicate the spin matrix and translation matrix of left and right cameras.
Binocular vision projection matrix can be obtained after the completion of camera calibration.It is assumed that parallel institution End features point P is on a left side
The picture point of right video camera is respectively Pl and Pr, and calibration is completed in left and right cameras and their projection matrix is respectively Ml, Mr, then
Had according to projection theorem:
Wherein sl, srIt indicates scale factor, above formula expansion is then had:
In formula, (u1,v1, 1), (u2,v2, 1) and respectively indicate subpoint PlAnd PrHomogeneous coordinates；(X, Y, Z, 1) is end
Homogeneous coordinates of the characteristic point under world coordinate system；Indicate the i-th of left and right projection matrix
Row, jth column.
Following equation can be obtained after formula (10) and formula (11) simultaneous:
It enables:Then may be used
Parallel institution terminal position three-dimensional coordinate P is found out using least square methodW, it is expressed as follows:
Pw=(ATA)-1ATB (13)
Spin matrix is expressed as shown in formula (14) again:
Then parallel institution terminal angle angle is shown in formula (15):
In formula, α is the pitch angle (unit are as follows: rad) rotated around X-axis, β be rotated around Y-axis roll angle (unit are as follows:
Rad), γ is the course angle (unit are as follows: rad) rotated about the z axis.
So far, the end pose parameter solution of parallel institution is completed.
Embodiment
The present invention focuses on to propose a kind of new parallel institution end movement position and posture detection method based on binocular vision, with
Solving traditional binocular vision algorithm, real-time is poor, detection accuracy is not high asks present in the detection of parallel institution end pose
Topic.Below using a kind of novel 3-DOF parallel institution end as test object, the specific embodiment of the detection method is as follows:
1, parallel institution original image is acquired.Novel 3-DOF parallel institution image is acquired by dimension view binocular camera,
In, camera model MV-1300FM, camera lens model AFT-0814MP, camera install the front of parallel institution, and height off the ground
It can flexible modulation with angle.In addition, this detection system uses computer operating system for Windows7, processor model is
Intel (R) Core (TM) 2 Duo, dominant frequency 2.66GHz inside saves as 2GB.
2, the denoising based on wavelet transformation is carried out to original image.In order to reduce noise to parallel institution end pose
Adverse effect brought by detecting, pre-processes original image using the denoising method based on wavelet transformation, passes through denoising
Processing effectively inhibits picture noise, and so as to next step feature extraction operation, it is as shown in Figure 2 to denoise effect.
3, mechanism picture characteristics is extracted using the corner extraction based on Harris operator.In view of novel 3-
DOF parallel institution has apparent rigid body feature, and the detection of Harris characteristic point is not only in rigid body geometry deformation and brightness change
Aspect has high reinspection rate, and algorithm is simple, stablizes, therefore chooses Harris Corner Detection Algorithm and carry out spy to mechanism image
Sign point extracts.The Harris receptance function of pixel are as follows:
R=det M-k (traceM)2 (16)
Wherein det M is the determinant of matrix M, and traceM is the authentic work of matrix M；K is empirical, usual value range
0.04~0.06.Given threshold t calculates response R according to above-mentioned Harris operator, if R < t, which is characterized a little, this reality
It applies example and takes empirical k=0.04, and directly ((conerHarris ()) realizes novel 3- by calling OpenCV library function
DOF parallel institution feature point extraction, novel 3-DOF parallel institution feature point extraction figure are as shown in Figure 3.
4, characteristic point Stereo matching is realized using Harris-SIFT algorithm.The present embodiment using Harris-SIFT algorithm into
The novel 3-DOF parallel institution Image Feature Point Matching of row.The matching algorithm replaces the extreme value in SIFT using Harris characteristic point
Point realizes the accurate positioning to characteristic point using the method for being fitted three-dimensional quadratic function, and rejects the low comparison to noise-sensitive
It spends point and is difficult to the point in image edge positioned, to filter out SIFT feature of the stable point as image.It is based on
The Stereo matching of Harris-SIFT algorithm is divided into following steps:
(1) Taylor expansion is carried out to the DoG operator of characteristic point x:
Wherein, Δ x is characterized the offset of point x, and when x is the extreme point of DoG operator, the single order local derviation of D (x) is 0, then
The offset Δ x of characteristic point exact position can be acquired:
Exact position and the scale that measured point is obtained after successive ignition, the point can be acquired by being substituted into formula (16)
DoG value simultaneously takes its absolute value, sets contrast threshold Tc, to reject the low contrast that contrast absolute value is less than threshold value Tc
Point.
And the rejecting for edge point, it can be realized by the principal curvatures of DoG extreme value of a function point with threshold value comparison, in general,
The DoG extreme value of a function point at edge is bigger than the principal curvatures of non-edge point, therefore, can pick the edge point that principal curvatures ratio is greater than threshold value
It removes.Non- edge point meets shown in relationship such as formula (19):
Wherein, r is the threshold value of setting, and the present embodiment takes r=10；The Hessian matrix that H is 2 × 2, H-matrix are indicated such as formula
(20) shown in:
(2) characteristic point direction is determined.The gradient-norm of L (x, y) of setting up an office is m (x, y), and direction is θ (x, y), the then Gauss
Gradient-norm and direction are indicated as shown in formula (21):
The gradient distribution of key point neighborhood territory pixel is counted, and chooses direction of the main peak value of histogram as key point.Extremely
This, each characteristic point has 3 important information: position, scale and direction.
(3) Feature Descriptor is generated.It generates Feature Descriptor and needs to rotate to be reference axis characteristic point direction, and uniformly draw
The subregion that 16 × 16 pixel regions for dividing this feature point adjacent are 84 × 4, then calculates 8 on each 4 × 4 subregion
The histogram of gradients in a direction, characteristic point each so just can produce the SIFT feature vector that 128 numbers form 128 dimensions accordingly,
The feature vector of generation is subjected to length normalization method, the influence of illumination variation can be further reduced.
(4) after Feature Descriptor generates, Feature Points Matching can be carried out according to the Euclidean distance of SIFT matching algorithm.Specifically
It is accomplished by firstly, all characteristic points of a certain characteristic point p and right image to be matched in left image are carried out Euclidean respectively
Distance calculates；Then, take out right image in p point Euclidean distance recently and time closely two point p1 and p2；Finally, calculating nearest
Euclidean distance is with time ratio of nearly Euclidean distance and compared with a threshold value, if the ratio is less than threshold value, successful match, this
When, (p1, p2) is a matching double points of image sequence, conversely, then it fails to match.The present embodiment given threshold is 0.5, is based on
The novel 3-DOF parallel institution matching effect of Harris-SIFT algorithm is as shown in Figure 4.
5, Mismatching point present in matching result is rejected using purification algorithm proposed by the present invention.Carry out purification processes
When, it can include the data set of abnormal point according to one group, calculate the object module of matching double points, left point is recycled to examine mould
Type, to obtain valid data sample.It, will according to the coordinate system of match point in left figure in order to quickly find out candidate family parameter
Picture is divided into 16 pieces, randomly selects and randomly selects a match point in 9 mutually different piece and each piece, from 9
8 matchings are randomly choosed to being used to find out temporary pattern parameter with centering, and the 9th matching is to being used to verifying temporary pattern just
True property, to quickly determine object module.Purify matching result effect as shown in figure 5, object module H selection and Sampson
Distance is expressed as follows:
The selection of object module H between (1) two image: 8 matching double points are randomly choosed in such a way that the above piecemeal takes a little
Interim candidate family F is calculated, and the Sampson distance d, given threshold k that find out matching double points corresponding to model F will meet d < k
The point of condition advanced optimizes to obtain object module H as interior point, and by the most candidate family F of interior points, finally by
Object module H reevaluates match point, rejects error hiding pair.Projective transformation model is indicated as shown in formula (22):
In formula, (x, y, 1) and (x ', y ', 1) respectively indicates the homogeneous coordinates of matching double points m (x, y) and m (x ', y ').
(2) expression of Sampson distance d: for any matching double points m (x, y) and m (x ', y '), in candidate mould
Sampson distance d in type F may be expressed as:
6, novel 3-DOF parallel institution end movement pose is solved.According to novel 3-DOF parallel institution movement characteristic, build
The position fixing system of vertical mechanism platform and moving coordinate system are as shown in Figure 7, wherein and { B }={ O-XYZ }, { T }={ O'-X'Y'Z'} divides
Not Biao Shi position fixing system and moving coordinate system, then the variation of parallel institution end pose is represented by moving platform central point O' point
Pose variation.The present embodiment realizes the camera calibration of binocular vision system using Zhang Zhengyou calibration method, and passes through image early period
It is as follows that processing and later period coordinate transform find out novel 3-DOF parallel institution terminal position information:
Pw=(ATA)-1ATB (24)
Wherein,
(u1,v1)、(u2,v2) respectively indicate End features point in the image coordinate of left and right camera subpoint,Indicate the i-th row of left and right projection matrix, jth column.
Meanwhile novel 3-DOF parallel institution terminal angle angle is shown in formula (25):
Since this parallel institution does three-degree-of-freedom motion, it can be found out in Y-axis, the translation displacements of Z axis, and around Y
The roll angle β of axis, then mechanism pose parameter is (Y, Z, β).
The present embodiment timing acquiring 12 opens image, then finds out each moment mechanism three-dimensional pose parameter, and by result with
The mechanism end actual bit that laser range finder INR-Laser Scanner H and electronic compass Honeywell-HMR3100 are measured
Appearance parameter compares, and tracking error is as shown in Figure 8.Analysis chart 8 as it can be seen that the tracking error of measurement point P1 and P7 in the Y direction compared with
Greatly, and P4 and P10 is larger in the tracking error of Z-direction, this is related with the symmetry of mechanism and running track, since mechanism is upper
Measurement point direction is stated to change so that velocity and acceleration changes greatly, the influence of additional mechanism inertia, cause error compared with
Greatly, this shows that the parallel institution end pose motion state that experiment measurement obtains is consistent with actual motion state；Meanwhile according to
The maximum deviation that tracking error can calculate each measurement point is respectively Y-direction: 0.622mm, Z-direction: 0.782mm, β rotation
Angle: 0.677 °, illustrate that detection has degree of precision.The experimental results showed that the pose proposed by the invention based on binocular vision
Detection method can preferably realize the real-time detection of parallel institution end movement pose.
To sum up, a kind of parallel institution end movement position and posture detection method based on binocular vision of the invention, firstly, will adopt
The parallel institution image collected carries out the pretreatment of the image denoising based on wavelet transformation, is mentioned with eliminating picture noise to subsequent feature
Influence caused by taking；Then, characteristic matching is carried out to parallel institution image using Harris-SIFT algorithm.Different from general
Matching algorithm, the matching algorithm pass through Harris operator extraction image characteristic point first, recycle SIFT feature description right
The characteristic point that Harris operator extraction goes out is matched, so that matching result has both real-time and stability；It then, is further
Matching accuracy is improved, proposes that a kind of new purification algorithm purifies matching result, which is taken a little by piecemeal
The method for taking checking computations temporary pattern in advance solves object module in purification algorithm and obtains that time-consuming, model solution is inaccurate
The problem of；Finally, the parallel institution End features point after matching purification is asked by coordinate transform binocular vision model is brought into
Mechanism end three-dimensional pose out.Parallel institution end position and posture detection method based on binocular vision proposed by the invention, due to
Pass through Harris operator extraction image characteristic point first in matching stage, SIFT feature description is recycled to mention Harris operator
The characteristic point of taking-up is matched, and image processing time can be greatly lowered, it is another by the new purification algorithm that is proposed to
Purified with result, can further improve matching accuracy so that parallel institution end pose detection real-time and
Precision is all improved.
In the description of this specification, reference term " one embodiment ", " some embodiments ", " illustrative examples ",
The description of " example ", " specific example " or " some examples " etc. means specific features described in conjunction with this embodiment or example, knot
Structure, material or feature are included at least one embodiment or example of the invention.In the present specification, to above-mentioned term
Schematic representation may not refer to the same embodiment or example.Moreover, specific features, structure, material or the spy of description
Point can be combined in any suitable manner in any one or more of the embodiments or examples.
Although an embodiment of the present invention has been shown and described, it will be understood by those skilled in the art that: not
A variety of change, modification, replacement and modification can be carried out to these embodiments in the case where being detached from the principle of the present invention and objective, this
The range of invention is defined by the claims and their equivalents.
Claims (6)
1. a kind of parallel institution end movement position and posture detection method based on binocular vision, characterized in that include the following steps:
Step 1) acquires the original image of parallel institution using binocular camera, and carries out collected mechanism image based on small echo
The image denoising of transformation pre-processes；
Step 2) extracts mechanism characteristics of image using the corner extraction based on Harris operator；
Step 3) carries out characteristic matching to parallel institution image using Harris-SIFT algorithm；Matching algorithm passes through first
Harris operator extraction image characteristic point, the characteristic point progress for recycling SIFT feature description to go out Harris operator extraction
Match；
There is erroneous matching and error matching Harris-SIFT algorithm in step 4), using new purification algorithm pair
Matching result does further purification processes: being taken by piecemeal and a checking computations temporary pattern a little and is in advance taken to improve purification algorithm, wherein
Piecemeal takes a little: taking the mode for taking a match point in image block and each piece at random；A checking computations temporary pattern is taken in advance: with
Machine chooses 9 matchings pair, and with 8 matchings to finding out temporary pattern parameter, and the 9th matching is to being used to verifying temporary pattern just
True property quickly determines iconic model with this；
Step 5) extracts three that End features point calculates the End features point according to Binocular Vision Principle, by image procossing
Coordinate is tieed up, and calculates parallel institution terminal angle angle, finally obtains the parallel institution end posture information.
2. a kind of parallel institution end movement position and posture detection method based on binocular vision according to claim 1, special
Sign is: in the step 2), the corner extraction of Harris operator can call directly the realization of OpenCV library function, in Harris spy
In sign point detection, the second moment of pixel be may be expressed as:
I in formulax、IyIndicate gradient of the gray scale in the direction x and y of pixel；The Harris receptance function of pixel are as follows:
R=detM-k (traceM)2
Wherein detM is the determinant of matrix M, and traceM is the authentic work of matrix M；K is empirical, value range 0.04~
0.06；Given threshold t calculates response R according to above-mentioned Harris operator, if R < t, which is characterized a little.
3. a kind of parallel institution end movement position and posture detection method based on binocular vision according to claim 1, special
Sign is: in the step 4), the purification algorithm calculates the target of matching double points by one group of data set comprising abnormal point
Model recycles left point to carry out testing model, to obtain valid data sample；Specific step is as follows:
4.1) picture is divided by b × b block according to the coordinate system of match point in the first width figure, wherein b > 3 randomly select 9
A match point is randomly selected in a mutually different piece and each piece, constitutes the random sample collection S of 9 matching double points；
4.2) 8 matching double points are randomly selected by random sample collection S and finds out interim candidate family F；
4.3) detect the 9th point to whether the supported collection for being interim candidate family F: if it is, this interim candidate family F be wait
Modeling type D；Otherwise, reselect 9 pairs of match points, repeat 4.1), 4.2) process；
4.4) all matching double points are detected by candidate family D and threshold value t, obtains the supported collection logarithm m of candidate family D；
4.5) if m >=threshold value t, obtains object module G；Otherwise, reselect 9 pairs of match points, repeat 4.1), 4.2), 4.3),
4.4) process；
4.6) object module G is optimized, it is final to determine object module H；
4.7) selection random sample collection, counter all add 1 every time, if repeating not find model parameter, terminator k times.
4. a kind of parallel institution end movement position and posture detection method based on binocular vision according to claim 1, special
Sign is: in the step 5), finding out parallel institution End features point three-dimensional coordinate (X, Y, Z) according to projection theorem, then by taking the photograph
Camera spin matrix R finds out parallel institution terminal angle angle；Wherein parallel institution terminal position information is as follows:
Pw=(ATA)-1ATB
Wherein,(u1,v1)、
(u2,v2) End features point is respectively indicated in the image coordinate of left and right camera subpoint；Table
Show the i-th row of left and right projection matrix, jth column；
Meanwhile parallel institution terminal angle angle are as follows:
Wherein, Rij(i=1,2,3；J=1,2,3) the i-th row of spin matrix R is indicated, jth column, α is the pitching rotated around X-axis
Angle, unit are as follows: rad, β are the roll angle rotated around Y-axis, unit are as follows: rad, γ are the course angle rotated about the z axis, unit are as follows:
rad。
5. a kind of parallel institution end movement position and posture detection method based on binocular vision according to claim 1, special
Sign is: in the step 5), further including, when three-dimensional reconstruction, it is necessary first to the camera parameters of built binocular vision model are obtained,
The camera parameters can be obtained by camera calibration；Video camera is carried out to binocular vision system using Zhang Zhengyou plane reference method
Calibration.
6. a kind of parallel institution end movement position and posture detection method based on binocular vision according to claim 5, special
Sign is: the Zhang Zhengyou plane reference method process are as follows: then fixed camera first takes pictures to scaling board, then by scaling board
Image input computer and the pixel coordinate for obtaining each lattice point on scaling board, are then loaded into the known three-dimensional coordinate of scaling board
Computer solves the inside and outside parameter of video camera, passes through by the way that the pixel coordinate of scaling board and three-dimensional coordinate are substituted into peg model
The calibration respectively for crossing left and right cameras can obtain the camera parameters spin matrix of binocular vision system further according to following formula
R and translation matrix T:
In formula, Rl、TlAnd Rr、TrRespectively indicate the spin matrix and translation matrix of left and right cameras.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610819289.9A CN106408609B (en) | 2016-09-13 | 2016-09-13 | A kind of parallel institution end movement position and posture detection method based on binocular vision |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610819289.9A CN106408609B (en) | 2016-09-13 | 2016-09-13 | A kind of parallel institution end movement position and posture detection method based on binocular vision |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106408609A CN106408609A (en) | 2017-02-15 |
| CN106408609B true CN106408609B (en) | 2019-05-31 |
Family
ID=57999227
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610819289.9A Expired - Fee Related CN106408609B (en) | 2016-09-13 | 2016-09-13 | A kind of parallel institution end movement position and posture detection method based on binocular vision |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106408609B (en) |
Families Citing this family (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106875440A (en) * | 2017-02-22 | 2017-06-20 | 神华集团有限责任公司 | The computational methods and device of coal pile volume |
| CN107133983B (en) * | 2017-05-09 | 2019-07-02 | 河北科技大学 | Bundled round steel end face binocular vision system and space orientation and method of counting |
| CN107186548B (en) * | 2017-06-08 | 2018-12-18 | 大连理工大学 | A kind of five-axle number control machine tool rotating shaft geometric error detection method |
| CN107443369A (en) * | 2017-06-25 | 2017-12-08 | 重庆市计量质量检测研究院 | A kind of robotic arm of the inverse identification of view-based access control model measurement model is without demarcation method of servo-controlling |
| CN107680060A (en) * | 2017-09-30 | 2018-02-09 | 努比亚技术有限公司 | A kind of image distortion correction method, terminal and computer-readable recording medium |
| CN108416791B (en) * | 2018-03-01 | 2021-07-23 | 燕山大学 | Binocular vision-based parallel mechanism moving platform pose monitoring and tracking method |
| CN108460707B (en) * | 2018-03-12 | 2022-04-15 | 林为庆 | Intelligent supervision method and system for homework of students |
| CN108668637A (en) * | 2018-04-25 | 2018-10-19 | 江苏大学 | A kind of machine vision places grape cluster crawl independent positioning method naturally |
| CN109261528A (en) * | 2018-09-03 | 2019-01-25 | 广州铁路职业技术学院（广州铁路机械学校） | Express delivery sorting method and device based on binocular vision |
| CN110320100B (en) * | 2019-07-03 | 2022-04-01 | 中北大学 | Double-camera Brinell hardness measuring device and measuring method based on machine vision |
| CN111383272B (en) * | 2020-02-24 | 2023-07-21 | 江苏大学 | Binocular vision fruit sorting parallel robot vision blind area tail end pose detection method |
| CN111915680A (en) * | 2020-09-01 | 2020-11-10 | 山东新一代信息产业技术研究院有限公司 | Robot positioning method, system, electronic device and computer readable medium |
| CN112967319A (en) * | 2021-03-31 | 2021-06-15 | 交通运输部天津水运工程科学研究所 | Block motion real-time detection method based on feature point identification |
| CN113160332A (en) * | 2021-04-28 | 2021-07-23 | 哈尔滨理工大学 | Multi-target identification and positioning method based on binocular vision |
| CN113500584B (en) * | 2021-07-15 | 2022-06-28 | 西北工业大学 | Tail end error correction system and method of three-degree-of-freedom parallel robot |
| CN113689555B (en) * | 2021-09-09 | 2023-08-22 | 武汉惟景三维科技有限公司 | Binocular image feature matching method and system |
| CN114523471B (en) * | 2022-01-07 | 2023-04-25 | 中国人民解放军海军军医大学第一附属医院 | Error detection method based on association identification and robot system |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103955931A (en) * | 2014-04-29 | 2014-07-30 | 江苏物联网研究发展中心 | Image matching method and device |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130287256A1 (en) * | 2012-04-30 | 2013-10-31 | Telibrahma Convergent Communications Private Limited | Method and system for real time image recognition on a mobile device |
-
2016
- 2016-09-13 CN CN201610819289.9A patent/CN106408609B/en not_active Expired - Fee Related
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103955931A (en) * | 2014-04-29 | 2014-07-30 | 江苏物联网研究发展中心 | Image matching method and device |
Non-Patent Citations (2)
| Title |
|---|
| Harris-SIFT算法及其在双目立体视觉中的应用;赵钦君 等;《电子科技大学学报》;20100731;第39卷(第4期);第546-550页 |
| Simplifying the Kinematic Calibration of Parallel Mechanisms Using Vision-Based Metrology;P. Renaud et al;《IEEE Transactions on Robotics》;20061231;第22卷(第1期);第12-22页 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106408609A (en) | 2017-02-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106408609B (en) | A kind of parallel institution end movement position and posture detection method based on binocular vision | |
| CN106651942B (en) | Three-dimensional rotating detection and rotary shaft localization method based on characteristic point | |
| CN105069743B (en) | Detector splices the method for real time image registration | |
| CN107392964B (en) | The indoor SLAM method combined based on indoor characteristic point and structure lines | |
| CN104981105B (en) | A kind of quickly accurate detection and method for correcting error for obtaining element central and deflection angle | |
| KR101532864B1 (en) | Planar mapping and tracking for mobile devices | |
| CN106204574B (en) | Camera pose self-calibrating method based on objective plane motion feature | |
| CN110533722A (en) | A kind of the robot fast relocation method and system of view-based access control model dictionary | |
| CN105701827A (en) | Method and device for jointly calibrating parameters of visible light camera and infrared camera | |
| CN101383899A (en) | Video image stabilizing method for space based platform hovering | |
| CN107657644B (en) | Sparse scene flows detection method and device under a kind of mobile environment | |
| CN108898634A (en) | Pinpoint method is carried out to embroidery machine target pinprick based on binocular camera parallax | |
| CN110223355B (en) | Feature mark point matching method based on dual epipolar constraint | |
| CN104036542A (en) | Spatial light clustering-based image surface feature point matching method | |
| CN109341668A (en) | Polyphaser measurement method based on refraction projection model and beam ray tracing method | |
| CN101493936B (en) | Multi- resolution non-rigid head medicine image registration method based on image edge | |
| CN106056121A (en) | Satellite assembly workpiece fast-identification method based on SIFT image feature matching | |
| CN109584157A (en) | Object plane degree measurement method and device, storage medium, electronic metering equipment | |
| Cvišić et al. | Enhanced calibration of camera setups for high-performance visual odometry | |
| CN110363706A (en) | A kind of large area bridge floor image split-joint method | |
| JP4597509B2 (en) | Pattern inspection apparatus and pattern inspection method | |
| CN116935013A (en) | Circuit board point cloud large-scale splicing method and system based on three-dimensional reconstruction | |
| Geng et al. | Machine vision detection method for surface defects of automobile stamping parts | |
| CN109373901A (en) | Method for calculating center position of hole on plane | |
| Ma et al. | WeldNet: A deep learning based method for weld seam type identification and initial point guidance |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190531 Termination date: 20190913 |