WO2013086678A1 - 平面模型下点匹配与位姿同步确定方法及计算机程序产品 - Google Patents

平面模型下点匹配与位姿同步确定方法及计算机程序产品 Download PDF

Info

Publication number
WO2013086678A1
WO2013086678A1 PCT/CN2011/083850 CN2011083850W WO2013086678A1 WO 2013086678 A1 WO2013086678 A1 WO 2013086678A1 CN 2011083850 W CN2011083850 W CN 2011083850W WO 2013086678 A1 WO2013086678 A1 WO 2013086678A1
Authority
WO
WIPO (PCT)
Prior art keywords
pose
image
solution
iterative
rough
Prior art date
Application number
PCT/CN2011/083850
Other languages
English (en)
French (fr)
Inventor
张广军
魏振忠
王巍
Original Assignee
北京航空航天大学
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 北京航空航天大学 filed Critical 北京航空航天大学
Priority to US14/360,388 priority Critical patent/US9524555B2/en
Priority to PCT/CN2011/083850 priority patent/WO2013086678A1/zh
Publication of WO2013086678A1 publication Critical patent/WO2013086678A1/zh

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/70Determining position or orientation of objects or cameras
    • G06T7/73Determining position or orientation of objects or cameras using feature-based methods
    • G06T7/75Determining position or orientation of objects or cameras using feature-based methods involving models
    • G06T3/147
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10004Still image; Photographic image
    • G06T2207/10012Stereo images
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N13/00Stereoscopic video systems; Multi-view video systems; Details thereof
    • H04N13/20Image signal generators
    • H04N13/204Image signal generators using stereoscopic image cameras
    • H04N13/239Image signal generators using stereoscopic image cameras using two 2D image sensors having a relative position equal to or related to the interocular distance

Definitions

  • Model-based monocular visual pose estimation involves the problem of camera positioning, mainly under the condition that the internal parameters of the camera are known, given a set of features (points, lines, quadratic curves, etc.) in the world coordinate system. Describe the projection on the image and obtain the rigid body transformation relationship between the camera coordinate system and the object coordinate system.
  • Model-based monocular visual pose estimation is an important research topic in computer vision, and it has important applications in visual navigation, pattern recognition, camera calibration and many other aspects.
  • the commonly used pose calculation methods are mainly divided into two categories: analytical methods and numerical methods.
  • analytical methods are the algorithm based on plane point feature proposed by Gerald Schweighofer and Axel Pinz.
  • the numerical methods include: O.Oberkampf, D.DeMenthon, LSDavis, iterative solution algorithm based on plane point feature, Tsai Planar target calibration algorithm, Zhang Zhengyou's planar target calibration algorithm, etc., but this method must be performed under the condition that the model feature and its image feature correspondence are known. For example, the correspondence of point features or the correspondence of line features is known.
  • a type of pose solving algorithm can simultaneously perform feature matching and pose synchronization, and a representative algorithm such as SoftPosit algorithm.
  • the pose is solved under the condition that the correspondence between the object point and the image point is unknown, and the existing method is only for For 3D objects, for example: SoftPoist algorithm in numerical algorithm is based on 3D model for point matching and pose synchronization. However, when this method is used for point matching and pose synchronization in plane model, it is solved.
  • the problem of "position redundancy" is encountered in the process, namely: The estimated pose solution increases exponentially with the iterative solution, making the solution impossible.
  • the present invention provides a method for determining point matching and pose synchronization in a planar model, the method comprising:
  • a rough sketch of two poses is obtained by rough pose estimation
  • the final pose accurate solution is selected from the pose to be selected.
  • the iterative search by the extended TsPose algorithm to obtain the pose to be selected further includes:
  • the pose estimation solution includes the determined rotation matrix, translation vector, and/or matching matrix; and the pose estimation solution obtained when the iterative convergence condition is satisfied is the pose candidate to be selected.
  • the above method further includes:
  • the iterative search process After the iterative search process is finished, it is judged whether the initial iteration values of the two groups have been iterated. If the initial iteration value is not iterated, the initial iteration value without the iteration is the initial iteration value, and the TsPose algorithm is extended. The iterative search results in a pose to be selected.
  • the iterative convergence condition is: the uncertainty is greater than the ending uncertainty; wherein, the uncertainty is a product of the initial uncertainty A and the updated frequency A of the uncertainty; and/or, the iterative convergence The condition is: when the objective function value is less than a preset value indicating that the iteration cannot be found, the iterative search process ends when the iteration convergence condition is satisfied.
  • the accurate solution based on the pose and the rough solution, and selecting the final output from the pose to be selected further includes:
  • the objective function value obtained as the initial iteration value is roughly obtained according to each set of the poses, and the pose position corresponding to the smaller objective function value is to be The selection is selected as the exact interpretation of the pose of the final output.
  • a computer program product for implementing point matching and pose synchronization determination in a planar model
  • the product comprising a computer usable medium having a computer readable program code
  • the computer readable program Encoding is used to perform a method for determining a pose under a planar model; the method comprises: obtaining two sets of pose rough solutions by rough pose estimation; using the rough solution of each set as the initial iteration value, and iterating by extending the TsPose algorithm The search obtains the pose candidate to be selected; based on the pose rough solution, the final pose accurate solution is selected from the pose candidate to be selected.
  • the iterative search by the extended TsPose algorithm to obtain the pose to be selected further includes:
  • the pose estimation solution includes the determined rotation matrix, translation vector, and/or matching matrix; and the pose estimation solution obtained when the iterative convergence condition is satisfied is the pose candidate to be selected.
  • the method further includes: After the iterative search process is finished, it is judged whether the initial iteration values of the two groups have been iterated. If the initial iteration value is not iterated, the initial iteration value without the iteration is the initial iteration value, and the TsPose algorithm is extended. The iterative search results in a pose to be selected.
  • the iterative convergence condition is: the uncertainty is greater than the ending uncertainty; wherein the uncertainty is the product of the initial uncertainty A and the update frequency A of the uncertainty; and/or, the iteration converges
  • the condition is: when the objective function value is less than a preset value indicating that the iteration cannot be found, the iterative search process ends when the iteration convergence condition is satisfied.
  • the accurate solution based on the pose is roughly solved, and the final output is selected from the pose to be selected, further comprising:
  • the objective function value obtained as the initial iteration value is roughly obtained according to the set pose, and the pose corresponding to the smaller objective function value is selected and selected.
  • the exact pose of the pose as the final output.
  • the present invention has the following advantages:
  • the invention provides a method for determining point matching and pose synchronization in a planar model, the method comprising: using a clustering method to obtain a rough solution of the pose, defining a pose search space, and solving the iterative search.
  • the "position redundancy" problem ; the second step, under the two-dimensional projective transformation, using the extended TsPose algorithm, through the iterative search to obtain more accurate point matching and pose results.
  • the TsPose algorithm is an algorithm for accurately determining a pose, that is, a facet sliding algorithm
  • the extended TsPose algorithm is a combination of an algorithm for point matching such as the Softassign algorithm and for accurate pose determination. Algorithms such as the TsPose algorithm.
  • FIG. 1 is a schematic diagram of a coordinate system and an attitude angle
  • Figure 2 shows the size of the tapped corners in the "seed pair" from small to large
  • Figure 3 is a graph of the objective function
  • Figure 4a shows a schematic diagram of matching of object clusters and pixel clusters
  • FIG. 4b is a schematic diagram showing the matching of the transformed object point cluster and the image point cluster
  • Figure 5a is a schematic diagram of an embodiment of a method for rough estimation of the present invention.
  • Figure 5b is a schematic view of an embodiment of a method for accurately locating a pose of the present invention
  • FIG. 6 is a schematic illustration of an iterative search process of the present invention.
  • DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The basic idea of the present invention is: a two-step method for solving point matching and pose synchronization using a planar model: In the first step, a rough pose is obtained to define a pose search space, and to solve the "position" of the iterative search. Redundant "problem; second step, combining point matching algorithm with precise pose acquisition algorithm, using the pose rough solution as the initial solution of the iterative search and the benchmark for selecting the pose estimation solution in the iterative search process, and then obtaining the iterative search More precise point matching and pose exact solutions.
  • the pose rough solution includes a coarse attitude angle and a coarse translation vector Z coordinate value; the pose accurate solution includes an attitude angle and a translation vector.
  • rough poses of two sets of poses are obtained by rough pose estimation; respectively, the rough solution of each pose is used as the initial iteration value, and the two sets of poses are selected by the extended TsPose algorithm. Based on the objective function value, a set of pose-to-select solutions is selected as the final pose-accurate solution.
  • One step - a method of rough pose estimation, the method comprising the following steps:
  • the number of neighbor points to be selected is set to k; the Euclidean distance between the object point and all other object points is calculated, and the calculated Euclidean distances are arranged from small to large, and the first k nearest ones are selected.
  • the neighborhood point and the object point form a cluster of object points; wherein, there are a total of k+1 points in the cluster of object points.
  • the image point There are a total of k+1 points in the cluster.
  • the number of object points is m
  • the number of image points is n
  • the object points have m object point clusters
  • the image points have n image point clusters
  • the object cluster data is subtracted from the respective centroids, and the whitening transformation is performed as shown in the following formula (1-3); accordingly, the pixel cluster data is respectively subtracted from the respective centroids, and the following formula (1) 3) The whitening transformation shown.
  • the polar coordinates of the transformed object point cluster are calculated, and the order of the points in the cluster of the object points is arranged according to the size of the polar angle; similarly, the polar coordinates of the transformed image point cluster are calculated, and Arrange the points in the cluster of point clusters from small to large according to the size of the polar angle;
  • the total k+1 circular arrangement of each pixel cluster is calculated and the subspace distance of the object clusters sorted in step S103, and the minimum distance of the subspace is recorded, and the corresponding object cluster is a dot cluster, and a cluster of pixel points corresponding to a value at which the object point is paired with the smallest distance of the subspace, repeating the above steps for all the object clusters, and obtaining a subspace distance matrix composed of subspace distances of mxn dimensions, where, The rows of the spatial distance matrix represent clusters of object points, and the columns represent clusters of image points;
  • S105 Set a threshold d of the subspace distance, and select a point cluster and a cluster of pixel points corresponding to the threshold d in the subspace distance matrix;
  • the possible matching relationship can be reduced, and the transformed data points can be used to obtain the correct matching relationship, thereby reducing the amount of calculation.
  • the data in both the object cluster and the image cluster is six, as shown in the dotted circle in Figure 4a.
  • the dots in the left circle represent the points in the cluster, and in the right circle.
  • the points represent points within the point cluster.
  • the transformation relationship between the correctly matched points in the two point clusters is an affine transformation.
  • the row calculation selects a matching pair that meets a certain threshold size according to a predetermined threshold value of the determination criterion, and each possible matching relationship is calculated according to the determination criterion, so the calculation amount is quite large.
  • the affine transformation is calculated by using the unconverted data of the selected object point cluster and the image point cluster pair, and two sets of rotation matrices are calculated according to the formula (2-10), and the third column of the rotation matrix is obtained.
  • the k-means clustering method is used to cluster the attitude angle and the translation vector z-coordinate, and the rough values of the two sets of coarse attitude angles and translation vector z-coordinates are obtained.
  • the pose rough solution includes rough values of two sets of coarse posture angles and translation vector Z coordinates.
  • the coarse attitude angle includes a coarse yaw angle and a coarse pitch angle.
  • each object point cluster and Each pixel cluster data is transformed.
  • the transformation step is divided into three steps:
  • the object cluster data in the seed pair is subtracted from its centroid, and the image cluster data in the seed pair is subtracted from its centroid.
  • the seed pair refers to the selected object point cluster and the image point cluster pair.
  • A is a 2x2 matrix and t is a 2x1 vector.
  • the centroid of the object point is m p
  • the centroid of the image point is m q
  • step 1) there is only one linear transformation relationship between the object point cluster and the image point cluster.
  • the object clusters are arranged in accordance with the polar angles from small to large.
  • the dimension space formed by the cluster of structure points is ⁇ 2 ] , ⁇ ⁇ y wl , y wk f ;
  • Step 2 - A method of accurate pose estimation, the method comprising the following steps:
  • the homogeneous coordinate of a given image point is divided by its modulus, that is, normalized, and a set of vectors with a modulus of 1 is obtained, which is simply referred to as a "line of sight bundle".
  • the setting of the iterative convergence condition includes: determining an initial uncertainty degree, ending an uncertainty degree, an update frequency A of the uncertainty, and setting an initial value of the matching matrix M that represents the matching relationship.
  • the theoretical value of the objective function (2-7) is 0. Therefore, a constant ⁇ with a small value can be set in advance.
  • the value of the objective function (2-7) is smaller than the constant indicating that the matching point cannot be found.
  • f the iterative search process is exited.
  • an accurate pose acquisition algorithm such as the extended TsPose algorithm is used, and the rough estimate of the two sets of poses obtained by the coarse pose estimation method, that is, the rough pose angle and the translation vector z coordinate value are respectively taken as the roughest pose acquisition algorithm. Extend the initial iteration value of the TsPose algorithm.
  • this embodiment does not limit the manner in which the initial iteration value is obtained.
  • the embodiment may adopt multiple acquisition methods, for example: input an initial iteration value, that is, a set of poses and rough solutions for each iteration. Or, the two sets of poses are roughly input at the same time, and are called separately at each iteration; or, after the rough pose estimation is completed, the poses obtained by the rough pose estimation are roughly solved. It is supplied to the device that performs the accurate pose estimation, or the device that performs the accurate pose estimation actively picks up the pose and the like from the device that performs the accurate pose estimation.
  • a set of roughly estimated rough attitude angles is used to calculate the rotation matrix using the Rodrigues formula, and the calculated rotation matrix is used as the initial estimation of the rotation matrix in the iterative search of the extended TsPose algorithm.
  • the distributed random number produces the X coordinate and y of the translation vector Coordinate values, Z coordinate values are given by the coarse pose estimation algorithm.
  • the coarse attitude angle includes a coarse pitch angle and a coarse yaw angle.
  • steps S203-S204 a rotation matrix R and a translation vector T are obtained, namely:
  • the superscript 0 indicates the 0th iteration, which means the iteration initial.
  • the superscript indicates the number of iterations.
  • the position of the tangent plane 1 3 ⁇ 4 and the "line of sight" obtained from the data step of normalizing the given image point) are obtained by the formula (2-2), (2-3) k
  • S206 Acquire a matching matrix by using a “image slice” image point, and update the obtained matching matrix: determine an objective function by using the updated matching matrix;
  • the superscript indicates the initial value before the Sinkhorn algorithm.
  • the matching matrix is updated using the Sinkhorn algorithm. Let the updated matching matrix element be, and then calculate the objective function (2-7), namely:
  • h is given by (2-3'), (2-4,), and in the subsequent iterations, hh 2 is determined by the formula (2-8), (2-9) and (2-11) are calculated.
  • the affine transformation relationship between the image point and the object point is updated by an accurate pose determination algorithm such as the extended TsPose algorithm, and two sets of rotation matrices are calculated by the formula (2-10), and the third column of the two sets of rotation matrices They are respectively used as two normal "image planes" tangent tangent to the unit sphere, and the two sets of estimated second attitude angles are calculated by the normal vector.
  • a set of attitude angles is calculated, set to ⁇ ) (representing the pitch angle, representing the yaw angle), and a set of attitude angles is calculated, which is set to (representing the pitch angle, ⁇ represents the yaw angle), and the step S203 is obtained.
  • the attitude angle in a set of poses is (, represents the pitch angle, which represents the yaw angle).
  • the pose solution includes: an attitude angle and a translation vector. It should be noted that the corresponding attitude angle can be calculated according to the rotation matrix. Therefore, the pose solution can also be considered to include: a rotation matrix and a translation vector. Furthermore, a significant advantage of an embodiment of the present invention is that point matching can be achieved while performing pose determination, so the final output results include a rotation matrix, a translation vector, and a matching matrix.
  • the pose rough estimation algorithm the yaw angle in the pose rough solution is obtained, and a set of "image planes, . . . is determined by ⁇ ' ⁇ and (in the case of the iteration) "image plane" image point q ' construct a set of points Set, recorded as:
  • Vi j (2-14) is related to the rigid body transformation of the camera coordinate system and the world coordinate system: (2-16)
  • step S207 R in the formula (2-16) has been obtained in step S207, and the object point coordinate ⁇ TM ' obtained in this step has been known.
  • the translation vector ⁇ can be solved by minimizing the sum of squared errors:
  • the intersection of the line of sight and the tangent plane of the two unit balls is obtained, and two sets of intersection points are obtained.
  • the first two dimensions of each group of intersections are the coordinates of the new "image-like" image points, and the third dimension and the "old" image points in step S206 form a three-dimensional data.
  • the following formula is used. (2-18) Calculate two sets of translation vectors. In this step, select a new image plane position and a new "image cut" image point.
  • the rough value of the rough pose angle and the translation vector z coordinate of each set of the poses is used as an initial iteration value, and two new "image planes" are determined, and according to the two new ones.
  • Image section obtains two sets of pose estimation solutions, namely pose estimation solution 1 and pose estimation solution 2. It should be noted that one of two sets of pose estimation solutions is selected based on the coarse yaw angle, the pose estimation solution including a rotation matrix and a translation vector.
  • step S209 determining whether the iterative convergence condition is satisfied, if the iterative convergence condition is satisfied, ending the iteration, performing step S210; if the iterative convergence condition is not satisfied, returning to step S206 according to the new "image plane" image point;
  • step S210 determining whether the iterative convergence condition is satisfied, that is, determining whether the value of the objective function (2-7) is smaller than a predetermined parameter.
  • step S203 the yaw angles in the two estimated second attitude angles are respectively subtracted from the coarse yaw angles in the coarse attitude angles input in step S203, and the corresponding difference absolute values 'j' are selected.
  • the tangent plane is used as the new image plane position and the corresponding image point as a new "image plane" image point, and the process returns to step S206 until the preset end condition is reached.
  • steps S206-S209 are re-executed, and so on, until the preset end condition is reached.
  • the preset end condition refers to an exit condition of the iterative search process, that is, when the uncertainty is >, the iterative search process ends, or when the value of the objective function (2-7) is less than a predetermined parameter. When f, exit the iterative search process.
  • the pose estimation solution obtained when the iterative convergence condition is satisfied can be called the pose candidate 1; similarly, the iterative search can be obtained by another set of pose rough solutions to obtain the pose candidate. 2.
  • the iterative search shown in FIG. 6 is completed by the two sets of pose-like rough solutions obtained in step S203, respectively.
  • two iterative searches are performed for the two sets of poses and the left search branch and the right search branch respectively.
  • the pose solution output of the left search branch is ⁇ 1 ⁇
  • the point set is matched.
  • the result is M i
  • the objective function value is the pose solution of the right search branch output is ⁇ R T2 ⁇
  • the matching result of the point set is M 2
  • the objective function value is £ 2 .
  • ⁇ ⁇ the last bit of the algorithm is output.
  • the solution is that the matching result of the point set is; otherwise, the pose solution output by the algorithm is ⁇ R 2 ' T 2 ) , the matching result of the point set is M 2 .
  • step S211 is performed;
  • step S203 If it is determined that the initial value of the iteration has not been calculated, the process returns to step S203, and another set of poses is used as the initial iteration value, and a new calculation for solving the pose exact solution is started.
  • the objective function value corresponding to the initial iteration value is obtained, and the pose-accurate solution of the final output is selected from the two sets of poses to be selected.
  • the rough solution of each set of poses is recorded as the objective function value corresponding to the initial iteration value, and the pose corresponding to the objective pose value is selected as the bit pose candidate solution.
  • the pose is precisely solved to the final output.
  • the objective function is a quadratic function that characterizes the relationship between the object point and the pixel matching relationship, the object point and the image point change, and is used to select the pose solution of the final output.
  • the output pose is selected as the final pose exact solution, otherwise the output pose is selected. 2 as the final pose accurate solution.
  • the optical center of the camera is O
  • the focal length is 1
  • the image plane is
  • the origin of the camera coordinate system is established in the optical center
  • the X-axis and _-axis of the camera coordinate system are parallel to the length and width of the image plane, respectively.
  • the z-axis of the camera coordinate system is along the direction of the optical axis and points to the front of the camera.
  • the standard is used to indicate.
  • the plane model is known
  • the z-axis of the world coordinate system is along the normal direction of the plane
  • the world coordinate system is known
  • the X-axis, the j-axis, and the z-axis satisfy the right-hand rule.
  • the unit vector oq oq /
  • the matching points of some feature points may be lost, that is, there are feature points that cannot be matched.
  • the problem in order to represent such a situation, add a row and a column to the matrix M, that is, the "2+1"hash" and the +1st row hash.
  • equation (2-6) in order to solve the matching coefficient matrix and the affine transformation parameters at the same time, and to match the matrix elements as weighting parameters, the objective function formula of equation (2-6) can be modified accordingly as follows:
  • the matching matrix M is updated by the sinkhorn algorithm.
  • the objective function ⁇ ⁇ is the quadratic function of the unknown vector and h 2 , and the objective function J ⁇ pair vector And the partial derivatives of 11 2 are to seek the objective function at each iteration.
  • the translation vector T can be solved by minimizing the sum of squared errors
  • the extended TsPose algorithm is a combination of an algorithm for point matching, such as the Softassign algorithm, and an algorithm for accurate pose determination, such as the TsPose algorithm, in which the matching relationship is characterized.
  • the matrix M is also a value that needs to be continuously estimated in the iterative algorithm.
  • the extended TsPose algorithm is the TsPose algorithm.
  • the digitally simulated camera has a focal length of 769 pixels, an image plane size of 512x512, an image center of (256, 256), and a pixel size of 0.0076 mm/pixel.
  • the plane object points and false points selected by the experiment are distributed in a square with a side length of 200 mm, subject to uniform distribution, the number of object points is 50, and the ratio of false points to object points is 1:2. The number is 75, and the image point adds Gaussian noise with a standard deviation of 1 pixel.
  • the pitch angle is set to 65 degrees
  • the azimuth is set to 60 degrees
  • the panning vector is set to [0.5; 1; 1600] (mm).
  • the iteration maximum of the extended TsPose algorithm is set to 50 times.
  • Step 1 Set the number of domain points to 7, and generate 50 object clusters for 50 object points and 70 image point clusters for 70 image points according to the Euclidean distance criterion.
  • Step 2 Perform coordinate transformation on 50 clusters of object points, ie: subtract their centroids respectively, then perform "whitening transformation", calculate the polar coordinates of the cluster of object points, and arrange the clusters of points from small to large according to the size of the polar angle. The order of the points, similarly, is also handled as such for the data of the point cluster.
  • Step 3 Calculate the subspace distance of the total of 8 ring arrangement combinations of each pixel cluster and the ordered object point cluster, and establish a subspace distance matrix.
  • Step 4 Set the threshold of the subspace distance to 2, and select the object point cluster and the image point cluster corresponding to the threshold 2 in the subspace distance matrix.
  • Step 5 According to the selected object point cluster and the image point cluster, calculate the affine transformation using the data before the transformation, and obtain the rough solution of the object point cluster and the image point cluster of each group, and use k-means clustering. Method, set the category to 2, cluster all the pose solutions, and get two sets of poses and rough solutions.
  • Step 6 The two sets of poses are used as the initial solution of the extended TsPose algorithm and the benchmark of the selected pose estimation solution in the iterative search process. Two sets of poses to be selected and two sets of matching matrices are selected by iterative search. solution.
  • Step 7 If the objective function value corresponding to the solution 1 to be selected is smaller than the target function value corresponding to the solution 2 to be selected, the output pose is selected as the final pose exact solution, and the matching matrix is selected. As the final matching matrix solution, otherwise the output pose is selected as the final pose exact solution, and the matching matrix candidate 2 is chosen as the final matching matrix solution.
  • an objective function graph of an example is given below. It can be seen from Fig. 3 that, at the end stage of the iterative search by the accurate pose estimation algorithm, the objective function value is basically no longer transformed, because the pose estimation solution is selected based on the pose-based rough solution in each iterative search. Therefore, the accurate pose estimation algorithm can be considered to be limited to the search for the rough solution of the pose, so when the pose rough solution is far from the true pose solution, the search in its field will not Obtain an accurate solution of the good pose, and at the same time, the corresponding objective function value is not less than the objective function value corresponding to the search for the pose in the rough pose similar to the real pose.
  • the pose solution of the output of the program is: the elevation angle is 66.81 degrees, the azimuth angle is 60 degrees, the translation vector is [-0.03; 1.26; 1611.40] (mm), the pitch angle error is 1.81 degrees, and the azimuth error is 0 degrees.
  • the relative error of the distance is 0.72%, and the matching accuracy is 100%.
  • the embodiment of the invention provides a method for point matching and pose synchronization calculation based on a plane model.
  • This embodiment can solve the problem of estimating the pose solution with the iterative algorithm when performing point matching and pose synchronization under two-dimensional projective transformation. Carry out an exponential growth problem.
  • the pose calculation method of the present embodiment is based on a set of plane points and is independent of the geometry of the planar object, the conditional limitation on the pose calculation of the spatial plane object can be weakened, and the method disclosed in the embodiment is occluded, In a messy environment, and in the presence of different levels of noise, it can also exhibit better robustness and has good application value.
  • the above steps of the present invention can be implemented by a general-purpose computing device, which can be concentrated on a single computing device or distributed over a network composed of multiple computing devices; Alternatively, they may be implemented by program code executable by the computing device, such that they may be stored in the storage device by the computing device, or they may be fabricated into individual integrated circuit modules, or multiple of them Modules or steps are made in a single integrated circuit module.
  • the invention is not limited to any particular combination of hardware and software.
  • the present invention also provides a computer program product for implementing point matching and pose synchronization determination in a planar model, the product comprising a computer usable medium having a computer readable program code, the computer readable program code Used to perform a planar model to determine pose Method; the method includes:
  • a rough sketch of two poses is obtained by rough pose estimation
  • the final pose accurate solution is selected from the pose to be selected.
  • the iterative search by the extended TsPose algorithm to obtain the pose to be selected further includes:
  • the pose estimation solution includes the determined rotation matrix, translation vector, and/or matching matrix; and the pose estimation solution obtained when the iterative convergence condition is satisfied is the pose candidate to be selected.
  • the method adopted in the above products further includes:
  • the iterative search process After the iterative search process is finished, it is determined whether the initial iteration values of the two groups have been iterated. If the initial iteration value is not iterated, the initial iteration value without the iteration is initial.
  • the initial iteration value is obtained through an iterative search by extending the TsPose algorithm to obtain a pose candidate.
  • the iterative convergence condition is: the uncertainty is greater than the ending uncertainty; wherein the uncertainty is the product of the initial uncertainty A and the update frequency A of the uncertainty; and/or, the iteration converges
  • the condition is: the target function value is less than a preset value indicating that the matching point parameter cannot be found;
  • the accurate solution based on the pose is roughly solved, and the final output is selected from the pose to be selected, further comprising:
  • the objective function value obtained as the initial iteration value is roughly obtained according to the set pose, and the pose corresponding to the smaller objective function value is selected and selected.
  • the exact pose of the pose as the final output.

Abstract

本发明提供一种平面模型下点匹配与位姿同步确定的方法及计算机程序产品,该方法包括:通过粗略位姿估计得到两组位姿粗略解;以每组所述位姿粗略解作为初始迭代值,通过扩展TsPose算法经迭代搜索得到两组位姿待选解;基于目标函数,选出最终输出的位姿精确解。因此,本发明解决了二维射影变换下进行点匹配与位姿同步解算时估计位姿解随迭代算法的进行呈指数增长的问题,位姿的计算是基于平面点集,且与平面物体的几何形状无关,减弱了对空间平面物体位姿计算的条件限制,该方法在闭塞、混乱的环境下,以及存在不同级别的噪声的情况下也能表现出较好的鲁棒性,具有很好的应用价值。

Description

平面模型下点匹配与位姿同步确定方法及计算机程序产品 技术领域 本发明涉及基于模型的单目视觉位姿估计, 尤其涉及一种平面模型下 点匹配与位姿同步确定的方法及计算机程序产品。 背景技术 基于模型的单目视觉位姿估计涉及摄像机定位的问题, 主要是在摄像 机的内参已知的条件下, 给定一组特征(点、 线、 二次曲线等)在世界坐 标系下的描述及其在图像上的投影, 求得摄像机坐标系与物体坐标系之间 的刚体变换关系。 基于模型的单目视觉位姿估计是计算机视觉中的一个重 要研究课题, 在视觉导航、 模式识别、 摄像机标定等诸多方面都有重要应 用。
目前, 常用的位姿计算方法主要分为两大类: 解析方法和数值方法。 其中, 比较有代表性的解析方法为 Gerald Schweighofer、 Axel Pinz提出的 基于平面点特征的算法; 数值方法包括: O.Oberkampf, D.DeMenthon、 L.S.Davis提出的基于平面点特征的迭代求解算法, Tsai的平面靶标标定算 法, 张正友的平面靶标标定算法等, 但是该类方法必须要在模型特征与其 像特征对应关系已知的条件下进行, 例如, 点特征的对应关系或者线特征 的对应关系已知。 在数值方法中, 有一类位姿求解算法可以同时完成特征 匹配与位姿的同步解算, 具有代表性的算法如 SoftPosit算法。
在物点与像点的对应关系未知的条件下求解位姿, 现有方法只有针对 三维物体的, 例如: 数值算法中的 SoftPoist算法是基于三维模型的点匹配 与位姿同步求解算法, 但是, 将这类方法用于平面模型下的点匹配与位姿 同步解算时, 在求解过程中会遇到 "位姿冗余" 问题, 即: 估计位姿解随 着迭代求解的进行呈指数增长, 使得求解无法进行下去。
由此可见, 现有的一些空间平面物体的位姿计算必须在一些特定的条 件下才能进行, 因而位姿计算适用的范围受到了限制。 发明内容 针对二维射影变换下进行点匹配与位姿同步解算时估计位姿解随迭代 算法的进行呈指数增长的问题, 本发明有必要提供一种平面模型下点匹配 与位姿同步确定的方法及计算机程序产品, 以解决上述的技术问题。
为解决上述问题, 本发明提供一种平面模型下点匹配与位姿同步的确 定方法, 该方法包括:
通过粗略位姿估计得到两组位姿粗略解;
以每组所述位姿粗略解作为初始迭代值, 通过扩展 TsPose算法经迭代 搜索得到位姿待选解;
基于所述位姿粗略解, 从所述位姿待选解中选出最终输出的位姿精确 解。
上述方法中, 所述通过扩展 TsPose算法经迭代搜索得到位姿待选解进 一步包括:
以每组所述位姿粗略解作为初始迭代值, 通过计算得到的初始 "像切 面" 像点获取匹配矩阵, 并对该匹配矩阵进行更新, 利用更新后的匹配矩 阵获取目标函数值;
更新仿射变换关系, 并根据更新后的仿射变换关系来得到旋转矩阵, 根据位姿粗略解中的粗略偏航角选择一个新的像切面和旋转矩阵, 并利用 所选择的新的像切面确定平移向量和新的 "像切面" 像点;
判断是否满足迭代收敛条件, 如果满足迭代收敛条件, 则结束迭代搜 索过程; 如果不满足迭代收敛条件, 则所确定的新的 "像切面" 像点进行 新一轮的迭代搜索, 直到满足迭代收敛条件;
所述位姿估计解包括所确定的旋转矩阵、 平移向量和 /或匹配矩阵; 满 足迭代收敛条件时得到的位姿估计解为所述位姿待选解。
进一步地, 上述方法还包括:
在结束迭代搜索过程之后, 判断两组所述初始迭代值是否均已进行过 迭代, 若还有初始迭代值未进行迭代, 则以未进行迭代的初始迭代值为初 始迭代值, 通过扩展 TsPose算法经迭代搜索得到位姿待选解。
上述方法中, 所述迭代收敛条件为: 不确定度 大于结束不确定度; 其 中, 不确定度 为初始不确定度 A与不确定度的更新频率 A之积; 和 /或, 所述迭代收敛条件为: 所述目标函数值小于预先设置的表示找不到匹 任一所述迭代收敛条件得到满足时, 迭代搜索过程结束。
上述方法中, 所述基于所述位姿粗略解, 从所述位姿待选解中选出最 终输出的位姿精确解进一步包括:
在两组初始迭代值均完成迭代搜索之后, 根据每组所述位姿粗略解作 为初始迭代值时得到的目标函数值, 将较小的目标函数值所对应的位姿待 选解选出, 作为最终输出的位姿精确解。
根据本发明的另一个方面, 还提供一种用于平面模型下实现点匹配与 位姿同步确定的计算机程序产品, 该产品内包括具有计算机可读程序编码 的计算机可用媒介, 该计算机可读程序编码用于执行平面模型下确定位姿 的方法; 该方法包括: 通过粗略位姿估计得到两组位姿粗略解; 以每组所 述位姿粗略解作为初始迭代值, 通过扩展 TsPose算法经迭代搜索得到位姿 待选解; 基于所述位姿粗略解, 从所述位姿待选解中选出最终输出的位姿 精确解。
上述产品中, 所述通过扩展 TsPose算法经迭代搜索得到位姿待选解进 一步包括:
以每组所述位姿粗略解作为初始迭代值, 通过计算得到的初始 "像切 面" 像点获取匹配矩阵, 并对该匹配矩阵进行更新, 利用更新后的匹配矩 阵获取目标函数值;
更新仿射变换关系, 并根据更新后的仿射变换关系来得到旋转矩阵, 根据位姿粗略解中的粗略偏航角选择一个新的像切面和旋转矩阵, 并利用 所选择的新的像切面确定平移向量和新的 "像切面" 像点;
判断是否满足迭代收敛条件, 如果满足迭代收敛条件, 则结束迭代搜 索过程; 如果不满足迭代收敛条件, 则所确定的新的 "像切面" 像点进行 新一轮的迭代搜索, 直到满足迭代收敛条件;
所述位姿估计解包括所确定的旋转矩阵、 平移向量和 /或匹配矩阵; 满 足迭代收敛条件时得到的位姿估计解为所述位姿待选解。
上述产品中, 所述方法还包括: 在结束迭代搜索过程之后, 判断两组所述初始迭代值是否均已进行过 迭代, 若还有初始迭代值未进行迭代, 则以未进行迭代的初始迭代值为初 始迭代值, 通过扩展 TsPose算法经迭代搜索得到位姿待选解。
上述产品中, 所述迭代收敛条件为: 不确定度 大于结束不确定度; 其 中, 不确定度 为初始不确定度 A与不确定度的更新频率 A之积; 和 /或, 所述迭代收敛条件为: 所述目标函数值小于预先设置的表示找不到匹 任一所述迭代收敛条件得到满足时, 迭代搜索过程结束。
上述产品中, 所述基于所述位姿粗略解, 从所述位姿待选解中选出最 终输出的位姿精确解进一步包括:
在两组初始迭代值均完成迭代搜索之后, 根据每组所述位姿粗略解作 为初始迭代值时得到的目标函数值, 将较小的目标函数值所对应的位姿待 选解选出, 作为最终输出的位姿精确解。
与现有技术相比, 本发明具有以下优势:
本发明提出了一种平面模型下点匹配与位姿同步确定方法, 该方法包 括: 第一步利用一种聚类方法获得位姿的粗略解, 限定了位姿搜索空间, 解决了迭代搜索的 "位姿冗余" 问题; 第二步, 在二维射影变换下, 利用 扩展 TsPose算法,通过迭代搜索获得更加精确的点匹配和位姿结果。这里, 所述 TsPose算法是一种用于精确确定位姿的算法, 即像切面滑动算法; 所 述扩展 TsPose算法是一种结合用于点匹配的算法如 Softassign算法与用于 精确位姿确定的算法如 TsPose算法。
进一步来讲, 本发明不仅能够解决二维射影变换下进行点匹配与位姿 同步解算时, 估计位姿解随迭代算法的进行呈指数增长的问题; 而且, 由 于本发明的位姿计算方法是基于平面点集, 且与平面物体的几何形状无关, 因此, 能够减弱对空间平面物体位姿计算的条件限制, 本发明的方法在遮 挡、 杂乱的环境下, 以及存在不同级别的噪声的情况下也能表现出较好的 鲁棒性, 具有很好的应用价值。 附图说明 图 1是坐标系及姿态角示意图;
图 2是 "种子对" 中的点按极角的大小由小到大排列;
图 3是目标函数曲线图;
图 4a示出了物点簇与像点簇的匹配示意图;
图 4b示出了变换后的物点簇与像点簇的匹配示意图;
图 5a是本发明粗略估计的方法实施例示意图;
图 5b是本发明精确位姿的方法实施例示意图;
图 6是本发明迭代搜索过程的示意图。 具体实施方式 本发明的基本思想在于: 采用平面模型下点匹配与位姿同步求解的两 步法: 第一步, 得到位姿粗略解, 以限定位姿搜索空间, 解决迭代搜索的 "位姿冗余" 问题; 第二步, 结合点匹配算法与精确位姿获取算法, 将位 姿粗略解作为迭代搜索的初始解和在迭代搜索过程中选择位姿估计解的基 准, 进而通过迭代搜索获得更加精确的点匹配和位姿精确解。 其中, 所述位姿粗略解包括粗略姿态角和粗略平移向量 Z 坐标值; 所 述位姿精确解包括姿态角和平移向量。
下面结合附图及实施例, 对本发明的平面模型下点匹配与位姿同步的 方法进行说明。
在本实施例中, 通过粗略位姿估计得到两组位姿粗略解; 分别以每组 所述位姿粗略解作为初始迭代值, 通过扩展 TsPose算法经迭代搜索得到两 组位姿待选解, 基于目标函数值, 选出一组位姿待选解作为最终输出的位 姿精确解。
笫一步——粗略位姿估计的方法, 该方法包括以下步驟:
5101 , 分别针对每个物点与像点, 选择物点簇和像点簇;
本步驟中,设定待选出的邻域点的个数为 k; 计算物点与其它所有物点 的欧氏距离, 将计算得到的欧式距离由小到大排列, 选取前 k个最近的邻 域点与该物点组成一个物点簇; 其中, 该物点簇中共有 k+1个点。
类似地, 计算像点与其它像点的欧氏距离, 将计算得到的欧式距离由 小到大排列, 选取前 k个最近的像点与该像点组成一个像点簇; 其中, 该 像点簇中共有 k+1个点。 本实施例中, 设物点的个数为 m个, 像点的个数 为 n个, 则相应地, 物点有 m个物点簇, 像点有 n个像点簇;
5102, 分别对每一个物点簇数据和每一个像点簇数据进行变换; 本步驟中, 所述分别对每一个物点簇和每一个像点簇数据进行的变换 主要是指: 使每个物点簇数据分别减去各自的质心, 并作如下述公式( 1-3 ) 所示的白化变换; 相应地, 使像点簇数据分别减去各自的质心, 并作如下 述公式(1-3 )所示的白化变换。 S103 , 对变换后的物点簇和像点簇内的点进行排序;
本步驟中, 计算变换后的物点簇的极坐标, 并按照极角的大小由小到 大排列物点簇内的点的顺序; 同理, 计算变换后的像点簇的极坐标, 并按 照极角的大小由小到大排列像点簇内的点的顺序;
S104, 计算子空间距离, 建立子空间距离矩阵;
本步驟中,计算每个像点簇的共 k+1个环形排列组合与在步驟 S103中 已排序的物点簇的子空间距离, 记录下子空间距离最小的值、 相应的物点 簇内的点排列、 以及与物点配对该子空间距离最小的值对应的像点簇, 对 所有的物点簇重复以上的步驟, 得到一个 mxn维的由子空间距离组成的子 空间距离矩阵, 其中, 子空间距离矩阵的行表示物点簇、 列表示像点簇;
S105 , 设置子空间距离的阈值 d, 选取子空间距离矩阵中小于该阈值 d 所对应的物点簇及像点簇对;
利用步驟 S102-S105对物点簇与像点簇内的数据进行变换,能够减少可 能的匹配关系, 利用变换后的数据点获取正确的匹配关系, 进而减少计算 量。
例如, 某物点簇和像点簇内的数据都是 6个, 如图 4a所示的虚线圆中 的点集, 图中左边圆中的点代表物点簇内的点, 右边圆中的点代表像点簇 内的点, 此时, 两个点集簇内正确匹配的点之间的变换关系为仿射变换。 物点簇内的点与像点簇内的点有 6! = 720个匹配可能性, 例如给定物点簇内 的顺序为 "a-b-c-d-e-Γ ,那么像点簇内的点与物点簇内的点匹配的可能顺序 为 "A-B-C-D-E-F"、 "A-C-B-D-E-F"、 "A-D-C-B-E-F" 等。 根据给出的一种 判定正确匹配关系的判定准则, 那么需要对 720个匹配对按照判定准则进 行计算, 根据预先设定好的判定准则阈值, 选择出符合一定阈值大小的匹 配对, 对每一个可能的匹配关系都要根据判定准则进行计算, 所以计算量 是相当大的。 经过 S102-S105变换后, 左右两个点集内 (即图中虚线圆内) 正确匹配的点集之间的变换关系为旋转变换关系, 如图 4b所示, 所以此时 可能的匹配关系就从 6! = 720次减少到了 6次, 由于可能的匹配关系大大减 少, 利用判定准则只需要对这 6次可能的匹配关系进行计算判定, 计算量 也相应大大减少了。
5106, 根据选取的物点簇及像点簇对, 利用其未变换前的数据计算仿 射变换, 获得初步的位姿粗略解;
在得到正确的匹配关系后, 还是需要利用未变换前的数据来计算仿射 变换。 本步驟中, 利用选取的物点簇及像点簇对的未变换前的数据计算仿 射变换, 并根据公式(2-10 )所示计算得到两组旋转矩阵, 由旋转矩阵的第 三列计算得到两组姿态角, 即俯仰角和偏航角, 同时由仿射变换关系, 由 公式(2-11 )计算得到平移向量 z坐标的粗略值;
5107, 利用 k均值聚类方法, 对初步得到的每一个位姿粗略解聚类, 得到位姿粗略解。
本步驟中, 利用 k均值聚类方法将姿态角、 平移向量 z坐标进行聚类, 得到两组粗略姿态角和平移向量 z 坐标的粗略值。 其中, 所述位姿粗略解 包括两组粗略姿态角和平移向量 Z 坐标的粗略值。 所述粗略姿态角包括粗 略偏航角和粗略俯仰角。
粗略位姿估计方法的实施例
本实施例中, 通过选取得到物点簇和像点簇之后, 对每一个物点簇和 每一个像点簇数据作变换, 该变换的步驟分为三步:
1)种子对中的物点簇数据减去其质心, 种子对中的像点簇数据减去其 质心。 这里, 所述种子对是指所选择的物点簇及像点簇对。 设物点为 Pi, 像点为 两者存在仿射变换关系 q; = A ^t (1-1)
其中, A是一个 2x2的矩阵, t是一个 2x1的向量。 设物点的质心为 mp, 像点的质心是 mq, 贝' J
¾ίΡί ( 1-2)
其中,
Figure imgf000012_0001
i=Pi-mp。 所以, 经过了步驟 1), 物点簇与像点簇之间 仅存在一个线性变换关系。
2)对物点 和像点 £分别作 "白化变换", 即: p =s p 2p, , =sq 2q' ( 1-3)
其中, 是像点 的方差 t
Figure imgf000012_0002
3)设变换后的物点 p;与像点 q;存在的线性变换为 , 显然 = S^ q As p , 计算物点 p;的极坐标 计算像点 q;的极坐标( , ;), 如图 2所 示,将物点 p;按照其极角 的大小,由小到大排列,将像点 q;按照其极角 0q; 的大小, 由小到大排列。 需要说明的是, 能够容易证明得出 X是正交矩阵。
经过上述步驟 1) -3) 的变换后, 物点与像点的变换关系为:
Figure imgf000012_0003
其中, q:. = ("i, 是像点坐标, P; =(Xot, yOT)T是物点坐标, R是一 个 2x2的旋转矩阵。 如图 4b所示, 其示出了经过步驟 1 ) -3 )变换后的物点 簇与像点簇的匹配示意图, 左边圆中的点代表物点簇内的点, 右边圆中的 点代表像点簇内的点, 此时, 两个点集簇内正确匹配的点之间的变换关系 为旋转变换。 其中, T表示向量或者矩阵的转置。
本实施例中, 如图 4b所示, 物点簇内有给定按照极角从小到大排列的
6 个点, 那么构造物点簇张成的 维子空间为
Figure imgf000013_0001
ν2] , νι
Figure imgf000013_0002
ywl, ywkf ; 同理, 像点 簇 的 一 个 子 空 间 相 应 的 维 子 空 间 为 S2=spim[w1 w2] , wx = (u u2, ···, uk) , w2 ={v v2/ vk) 。 那么, 子 间 巨离 的定义如下:
其中, 是子空间 Si的投影矩阵, P2是子空间82的投影矩阵, 矩阵范数为 2-范数, 即: 矩阵的最大奇异值。投影矩阵的求解可通过现有的计算方法实 现, 此处则不对其进行详细描述。
笫二步——精确位姿估计的方法, 该方法包括以下步驟:
5201, 归一化处理给定像点的数据;
本步驟中, 将给定像点的齐次坐标除以其模, 即作归一化处理, 得到 一组模为 1的向量, 简称为 "视线束"。
5202, 设置迭代收敛条件
其中, 所述迭代收敛条件的设置包括: 确定初始不确定度 °, 结束不 确定度^ , 不确定度的更新频率 A , 设置表征匹配关系的匹配矩阵 M的初 始值。 其中, 不确定度 +1= χΑ, 当 +1>Α时, 则确定迭代结束。 正确时, 目标函数 (2-7)的理论值为 0, 所以, 可以预先设置一个值很小的常 数 ε , 当目标函数(2-7 )的值小于所设置表示找不到匹配点的常数 f时, 则 退出迭代搜索过程。
需要说明的是, 同时采用上述两种迭代收敛条件, 无论哪个条件先达 到, 都可以结束本次迭代搜索。
5203 , 获取精确位姿获取算法的初始迭代值;
本步驟中, 采用精确位姿获取算法如扩展 TsPose算法, 分别将第一步 中通过粗略位姿估计方法得到的两组位姿粗略解即粗略姿态角和平移向量 z坐标值的粗略估计值作为扩展 TsPose算法的初始迭代值。
需要说明的是, 本实施例不对获取初始迭代值的方式进行限制, 换句 话说, 本实施例可以采用多种获取方式, 例如: 每次迭代时输入一个初始 迭代值即一组位姿粗略解; 或者, 一次性将两组位姿粗略解同时输入, 每 次迭代时分别进行调用; 或者, 在粗略位姿估计完成之后, 由执行粗略位 姿估计的装置将其所得到的位姿粗略解输送给执行精确位姿估计的装置 , 或者由执行精确位姿估计的装置向执行粗略位姿估计的装置主动调取位姿 粗略解等。
5204 , 计算旋转矩阵, 并将计算得到的旋转矩阵作为扩展 TsPose算法 的迭代搜索中旋转矩阵初始估计值;
本步驟中, 由一组粗略估计的粗略姿态角, 利用罗德里格斯公式计算 得到旋转矩阵, 并将计算得到的旋转矩阵作为扩展 TsPose算法的迭代搜索 中旋转矩阵初始估计值, 由 0-1 分布的随机数产生平移向量的 X坐标和 y 坐标值, Z坐标值由粗略位姿估计算法给出。 其中, 所述粗略姿态角包括粗 略俯仰角和粗略偏航角。
S205, 确定初始 "像切面" 像点;
在步驟 S203-S204中, 得到旋转矩阵 R和平移向量 T, 即:
Figure imgf000015_0001
T= f , ί,〃 t ( 2-2' )
由 (2-Γ)、 (2-2')得表征仿射变换关系的两个列向量:
Figure imgf000015_0002
其中, 上标 = 0表示第 0次迭代, 意思是迭代初始, 在精确位姿估计 方法中, 上标都表示迭代的次数。
物点在世界坐标系中的齐次坐标为^„. =( ^, ywi> 1)τ, 由(1)得迭 代初始的切平面为 l =(r13, r23, r33f , k = 0 , 因为切平面的位置 1¾和 "视 线束"(由归一化给定像点的数据步驟得到)均得到, 由公式(2-2)、 (2-3) k 卜 k
计算得到 "像切面" 像点 q!' = , y{ , k = 0
S206, 利用 "像切面" 像点获取匹配矩阵, 更新所获取的匹配矩阵: 利用更新后的匹配矩阵确定目标函数;
本步驟中, 设: + h *p -y.
Figure imgf000015_0003
( 2-5' ) 从而, 计算匹配矩阵:
0 [β^-α)
m-: = r 1 '
l] 1 ( 2-6' ) 这里, 的上标表示进行 Sinkhorn算法前的初值。利用 Sinkhorn算法 更新匹配矩阵。 设更新后的匹配矩阵元素为 , 然后计算目标函数(2-7), 即:
=
Figure imgf000016_0001
其中, hx =λ(α , d\2 ' )
需要注意的是, 在第一次迭代时, h 由 (2-3')、 (2-4,) 式给出, 在以后的迭代中, h h2、 都是由公式(2-8)、 (2-9) 以及 (2-11 )计算 得到。
S207 , 更新仿射变换关系, 并根据更新后的仿射变换关系确定旋转矩 阵, 根据获取的粗略偏航角判据选择一个新的像切面和旋转矩阵;
本步驟中, 通过精确位姿确定算法如扩展 TsPose算法, 更新像点与物 点的仿射变换关系, 并由公式(2-10)计算得到两组旋转矩阵, 两组旋转矩 阵的第三列分别作为与单位球相切的两个新的 "像切面" 的法向量, 并由 法向量计算得到两组估计的第二姿态角。 利用 sinkhorn算法不断更新物点 与像点的匹配矩阵。 例如: 根据公式 ( 2-8 )、 ( 2-9 ) , 设 = (/iu, h12, h
ΙΤ-2 - (^^21, , ) , 己
Figure imgf000017_0001
, 则
'11 =h ' n 11, a 1? h * 11_2?., a ?1 h ' 2.Ί, * 2..
由公式(2-11 )得 , 因
Figure imgf000017_0002
则可以得矩阵 A, 由公式(2-10)可 得两个旋转矩阵 ^和^, 设^的第三列为 R2的第三列为 12, ^和12分 别代表两个新的 "像切面" 的表示, 如图 2所示, 要从新的 "像切面" ^和 12选择一个继续迭代搜索, 选择的方式如下:
由 计算得到一组姿态角,设为 ^)( 代表俯仰角, 代表偏航角 ), 由 计算得到一组姿态角, 设为 ( 代表俯仰角, ^代表偏航角), 设步驟 S203获取的一组位姿粗略解中的姿态角为 、 ( 代表俯仰角, 代表偏航角), 如果 - <1^- , 则保留 ι1 即新的 "像切面" 的表示为 lk+1 = 1: , 同时也保留旋转矩阵 作为迭代第 fc+i步的旋转矩阵估计值, 即 Rfc+1=R1; 反之, 则保留 12, 即新的 "像切面" 的表示为 1+1=12, 保留旋 转矩阵 R 2作为迭代第 + 1步的旋转矩阵估计值, 即 Rfc+1 = R2
本实施例中, 所述位姿解包括: 姿态角和平移向量。 需要说明的是, 根据旋转矩阵可以计算得到相应的姿态角。 因此, 也可以认为位姿解包括: 旋转矩阵和平移向量。 此外, 作为本发明实施例的一个突出的优势在于, 在进行位姿求解的同时能够实现点匹配, 所以, 最后的输出结果包括旋转 矩阵、 平移向量以及匹配矩阵。
S208, 利用新的像切面, 确定平移向量和新的 "像切面" 像点; 这里, 将 lfc+1代入公式(2-2)、 (2-3), 即: r+i. ^+iopj = i
Figure imgf000018_0001
根据位姿粗略估计算法得到位姿粗略解中的偏航角, 确定一组 "像切 面,,。 由 ν'· 和(第 步迭代时的) "像切面" 像点 q '构造一组点集, 记作:
、k -k fc+l、T
Qi中, Vi j (2-14) 由摄像机坐标系与世界坐标系的刚体变换关系得: (2-16)
公式( 2-16 )中的 、 R已在步驟 S207中得到 , 已在此步驟中得到 物点坐标 ρ™'已知。
平移向量 τ可以通过下面最小化误差平方和的方式求解得到:
Figure imgf000018_0002
由于 (2-17) 式是一个平移向量 T的二次函数, 当旋转矩阵 R给定后 平移向量 T可以给出解析解:
Figure imgf000018_0003
所以, 由点集 {^Q^^1''"'"}、 点集 {P 、 旋转矩阵 R 可解得:
Figure imgf000018_0004
(2-19) 可以看到, 迭代搜索过程中的旋转矩阵和平移向量的求解如图 6所示。 每次迭代都会有两组位姿估计解,以所确定的新的 "像切面 "像点继续进行迭 代搜索, 直到达到迭代收敛条件。
本步驟中, 求视线束与两个单位球的切平面的交点, 得到两组交点。 每组交点的前两维作为新的 "像切面"像点的坐标, 第三维与步驟 S206中的 "旧"像点组成一个三维数据, 利用该三维数据与物点坐标, 由下述的公式 ( 2-18 )计算得到两组平移向量。本步驟中,选择新的像平面位置及新的 "像 切面" 像点。
本实施例中, 以每组所述位姿粗略解即粗略姿态角和平移向量 z 坐标 的粗略值作为初始迭代值, 确定两个新的 "像切面", 并根据所述两个新的
"像切面"得到两组位姿估计解即位姿估计解 1和位姿估计解 2。 需要指出 的是, 根据粗略偏航角来选择两组位姿估计解中的一个, 所述位姿估计解 包括旋转矩阵和平移向量。
S209, 判断是否满足迭代收敛条件, 如果满足迭代收敛条件, 则结束 迭代, 执行步驟 S210; 如果不满足迭代收敛条件, 则根据新的 "像切面" 像点返回步驟 S206;
本步驟中, 可分别采用以下两种方式判断迭代收敛条件是否得到满足:
( 1 )判断迭代收敛条件是否得到满足, 即判断不确定度^ ^ = ^ X^» 是否大于结束不确定度^ ; 若 +1 >A , 则退出迭代搜索过程, 执行步驟 S210; 否则, 回到步驟 S206。 或者
( 2 )判断迭代收敛条件是否得到满足, 即判断目标函数( 2-7 )的值是 否小于预先给定的参数 当目标函数(2-7 )的值小于预先给定的参数 f时, 退出迭代搜索过程, 执行步驟 S210; 否则, 回到步驟 S206。
在本步驟中, 将两组估计的第二姿态角中的偏航角分别与步驟 S203所 输入的所述粗略姿态角中的粗略偏航角相减, 选择差绝对值 ' j、的所对应的 切平面作为新的像平面位置、所对应的像点作为新的"像切面"像点,返回并 执行步驟 S206, 直到达到预设的结束条件。
由公式(8 )得新的像切面像点为:
Figure imgf000020_0001
本步驟中, 以新的像平面位置、新的"像切面 "像点作为新的迭代值, 重 新执行步驟 S206-S209, 如此反复, 直到达到预设的结束条件。 所述预设的 结束条件是指迭代搜索过程的退出条件, 即: 当不确定度 > 时, 则本次 迭代搜索过程结束, 或者当目标函数(2-7 )的值小于预先给定的参数 f时, 退出迭代搜索过程。
那么, 经过 n次迭代, 满足迭代收敛条件时得到的位姿估计解, 可以 称之为位姿待选解 1 ; 同样, 采用另一组位姿粗略解进行迭代搜索可以得到 位姿待选解 2。
下面结合图 6, 对本实施例的迭代搜索过程作进一步说明:
分别通过步驟 S203中获取的两组位姿粗略解, 完成图 6所示的迭代搜 索。 这里, 设两组位姿粗略解分别进行的两个迭代搜索分别为左搜索分支 和右搜索分支, 如图 6所示, 设左搜索分支输出的位姿解为^ 1^ , 点集 的匹配结果为 Mi , 目标函数值为 右搜索分支输出的位姿解为 {R T2}, 点集的匹配结果为 M2 , 目标函数值为 £ 2 , 若^ ^ , 则该算法最后输出的 位姿解为 , 点集的匹配结果为 ; 反之, 则该算法输出的位姿解为 {R2 ' T 2 ) , 点集的匹配结果为 M2
5210, 记录下该组位姿粗略解作为初始迭代值对应的目标函数值, 并 判断两组初始迭代值是否均已进行过计算, 若还有初始值未计算过则返回 步驟 S203; 若均已计算完毕, 则执行步驟 S211 ;
若判断得知, 还有迭代初始值未计算过则返回步驟 S203 , 将另一组位 姿粗略解作为初始迭代值, 开启新一次求解位姿精确解的计算。
5211 , 根据每组位姿粗略解作为初始迭代值对应的目标函数值, 从迭 代得到两组位姿待选解中选出最终输出的位姿精确解。
本步驟中, 每次退出迭代搜索过程后, 记录下每组位姿粗略解作为初 始迭代值对应的目标函数值, 选择目标函数值小的位姿粗略解所对应的位 姿待选解作为位姿精确解, 以最终输出。 其中, 所述目标函数是表征物点 和像点匹配关系、 物点和像点变化关系的二次函数, 用于选出最终输出的 位姿解。
例如: 若位姿待选解 1对应的目标函数值小于位姿待选解 2对应的目 标函数值, 则输出位姿待选解 1 作为最终的位姿精确解, 否则输出位姿待 选解 2作为最终的位姿精确解。
精确位姿估计方法的实施例
本文的扩展 TsPose算法是在平面模型已知, 摄像机已标定的条件下得 出的, 除非特别说明, 算法所指的像点均是归一化像点, 即焦距 / = 1。
如图 1所示, 摄像机的光心为 O , 焦距为 1 , 像平面为 , 摄像机坐标 系的原点建立在光心, 摄像机坐标系的 X轴、 _ 轴分别与像平面的长与宽 平行,摄像机坐标系的 z轴沿着光轴的方向并指向摄像机的前方,摄像机坐 标系用 表示。 平面模型已知, 世界坐标系的 z轴沿着平面的法线方向,世界坐标系的
X轴、 j轴与 z轴满足右手法则, 世界坐标系用 owxw_ywzw表示, 由于物点是 共面点, 不失一般性, 欧氏坐标用 =( ;., ywi, ο)τ表示, 齐次坐标用
Vwi = (^i ywi if 表示, 物点在摄像机坐标系下的欧氏坐标用 p{=(xif yir f表示, 其中 = 1,2,..., w , "表示物点的个数, 像点的欧 氏坐标表示为 q,. =( ,., Vi)T , 像点的齐次坐标用 = vir if表示。
设。 q , vir 1)Γ, i = l,2,...,n, 单位矢量 oq =oq /||oq」| , 像平面 τ的单位法向量为 1 = , ly, lzf。 由于该文中像平面 τ是单位球 S 的切平面且又是成像面, 所以, 像平面 在本文中称为像切面, 像切面 可 以表示为:
X
1 , 1 , 1 y (2-1 )
z 单位矢量 oq 在的射线与像切面 r的交点用 V,=(x,, Vl, ζ,) 表示, 可由式 (2-2)、 (2-3)求得:
Figure imgf000022_0001
(2-3)
其中, i¾, = l,2,...,w代表放缩系数。 为了与给定的像点 q区分, 以 表示新的像点, 简称为 "像切面" 像 点, "像切面,, 像点的坐标是点 X的前两个分量, 即 =(^, yt , 相应 的齐次坐标为
Figure imgf000022_0002
支设像切面 τ经过第 步移动, 即算法的 第 次迭代搜索过程, 像切面 的单位法向量为 lk, 得到 "像切面"像点 Ϊ (上标 表示像切面 经过了第 次移动, 下标第 个 "像切面" 像点。), 由 "像切面" 像点 和物点^ ^, 建立近似仿射变换关系:
Figure imgf000023_0004
Figure imgf000023_0001
(2-4)可以简写为:
Figure imgf000023_0002
其中, 是一个比例常数, H , f,矩阵 A , 向量
^21 ^ 22 = (b b2f。 为了求解(2-5 ) 式中的仿射变换参数, 建立如下所示的目标函数式:
£fc = ∑| (hi,p- -Ji ) (2-6)
Figure imgf000023_0003
其中, 1^= (flu, a12, bt) , h2 = λ(α21, a22, b2 ) 。
当点的匹配关系未知时, 式( 2-6 )的目标函数 E应该相应地作些修改。 下面首先介绍一下表征匹配关系的矩阵匹配 M , 假设物点的个数是 个, 像点的个数是《2个, 由于点的对应关系不知道, 那么, 每一个像点 ( i = \,2,...,n2 )可以和任意一个物点 pwj ( 7 = 1,2,...,^ )进行匹配, 以 ( i = \,2,...,n2 , 二工 ,…, )表示匹 己关系, =1表示两点匹 己, my=0表 示不匹配, 矩阵 M是一个置换矩阵( permutation matrix ), 即: 矩阵 M元素 的值只能是 0或者 1, 且矩阵 M的每行和每列元素有且只有一个元素为 1, 这是一个对匹配矩阵进行 "双向约束" 的条件。 由于遮挡或者图像提取算 法等等的原因, 某些特征点的匹配点可能丟失, 即存在无法匹配的特征点 的问题, 为了表示这类情况,给矩阵 M增加一行和一列, 即第 "2+1 "散行" 和第 +1行散列。 这样, i i+1 =1表示第 个像点找不到匹配的物点, m¾+li. = 1表示第 j个物点找不到匹配的像点。
现在, 为了同时求解匹配系数矩阵以及仿射变换参数, 以匹配矩阵元 素作为加权参数, 式(2-6) 的目标函数式可以相应修改如下:
£fc =∑∑( _")
i=l j=l
, 〜 fc、2 /, 〜 、2 (2-7) i=l j=l V
其中, !^二 ^^, 12, bx) , h2 = λ α21, a22, b2) , "是一个很小的 常数。
利用 sinkhorn算法更新匹配矩阵 M , 具体的方法是: a) 初始化匹配矩阵 M ,令^ = 7X e^Xd" , l≤i≤m, 1< j≤n , w,,K+1,l≤i≤m + l分配一个很小的常数, <j< " + 1分配一个很小 的常数;
b) 对 矩 阵 M 的 每 一 行 的 元 素 进 行 归 一 化 ,
Figure imgf000024_0001
C 矩 阵 M 的 每 一 列 的 元 素 进 行 m Π .. = : ~~ m—-- ^- , 71 < z < m + 1, 71 < J j≤n
m ..
y
目标函数^ ^是未知向量 和 h2的二次函数, 设目标函数 J ^对向量 和112的偏导数分别为 在每次迭代时, 要寻求使目标函数
Figure imgf000025_0001
值最小化的向量 hl和 h2 , 这样的向量需要满足的条件是^" = 0且 dEk
= 0 , 当匹配系数 给定后, 和112的最优解可以以解析的形式表 达如下:
阵 A'可以计
Figure imgf000025_0002
算得到两个正交矩阵 R^PR2 , 以及比例常数 , 其中正交矩阵 R pR2左上 角的 2x2的矩阵是矩阵 A, 计算公式如下:
¾1
"21
a31
Figure imgf000025_0003
其中, S = -sign (
Figure imgf000025_0004
), sign表示取符号操作, 矩阵 的第 1行、 第 2行与第 3行满足右手定律, Sr =an 2 + an 2 +a2l 2 +a22 2 0 由于像平面 τ的单位法向量 1沿光轴的正向, 表示的是垂直于物平面 的矢量, 所以, 取大于 0的值, 即
(2-12)
Figure imgf000026_0001
设 和12分别为旋转矩阵!^和!^的第三列, 即代表两个新的 "像切面" 的表示, 由 计算得到一组姿态角, 设为( ' 、 ( 代表俯仰角, 代表偏 航角), 由 计算得到一组姿态角, 设为( ' ( 代表俯仰角, ^代表偏 航角), 设步驟 S203获 一组位姿粗略解中的姿态角为 、 ( 代表俯 仰角, 代表偏航角),
Figure imgf000026_0002
, 反之, lfc+1=l2
然后, 将得到的 lfc+1分别代入(2-2)、 (2-3) 式, 即:
lfc+1. +1op!) = l
Figure imgf000026_0003
由 和(第 步迭代时的) "像切面" 像点 q '构造一组点集, 记作:
k ^k+l
(2-14) 由摄像机坐标系与世界坐标系的刚体变换关系得:
Q卜 RP∞+T (2-16)
平移向量 T可以通过下面最小化误差平方和的方式求解得到
Figure imgf000026_0004
i=l (2-17) 由于 (2-17 ) 式是一个平移向量 τ的二次函数, 当旋转矩阵 R给定后, 平移向量 τ可以给出解析解:
T = -∑(^Q, -RpOT )
w '=i ( 2-18 ) 所以, 由点集 可解得:
Figure imgf000027_0001
需要说明的是, 本实施例中, 所述扩展 TsPose算法是一种结合用于点 匹配的算法如 Softassign算法与用于精确位姿确定的算法如 TsPose算法, 在该算法中, 表征匹配关系的矩阵 M也是迭代算法中需要不断估计的值, 当匹配矩阵 M始终为常矩阵一单位矩阵时, 扩展 TsPose算法即为 TsPose 算法。
数字仿真实验实施例
数字仿真的摄像机的焦距为 769个像素,像面大小为 512x512,像面中 心是(256, 256 ) , 像元尺寸为 0.0076毫米 /像素。 实验选取的平面物点和 虚假点均分布在边长为 200 毫米的正方形内, 服从均匀分布, 物点的个数 为 50个, 虚假点与物点的比为 1:2, 像点的个数为 75个, 像点添加了标准 差为 1个像素的高斯噪声。 俯仰角设定为 65度, 方位角设定为 60度, 平 移向量设定为 [0.5; 1;1600] (毫米)。 扩展 TsPose算法的迭代最大值设置为 50 次。
步驟 1 : 设定领域点的个数为 7, 根据欧式距离准则, 分别针对 50个 物点生成 50个物点簇, 针对 70个像点生成 70个像点簇。 步驟 2: 对 50个物点簇进行坐标变换, 即: 分别减去其质心, 然后进 行 "白化变换" , 计算物点簇的极坐标, 按照极角大小由小到大排列物点 簇内的点的顺序, 类似地, 对像点簇的数据也作如此的处理。
步驟 3:计算每个像点簇的共 8个环形排列组合与已排序的物点簇的子 空间距离, 建立子空间距离矩阵。
步驟 4: 设置子空间距离的阈值为 2, 选取子空间距离矩阵中小于阈值 2所对应的物点簇与像点簇。
步驟 5: 根据选取的物点簇与像点簇, 利用其未变换前的数据计算仿射 变换, 得到每一组匹配的物点簇与像点簇的位姿粗略解, 利用 k均值聚类 方法, 设定类别为 2, 对所有的位姿解进行聚类, 得到两组位姿粗略解。
步驟 6: 将两组位姿粗略解分别作为扩展 TsPose算法的迭代初始解和 迭代搜索过程中选择位姿估计解的基准, 通过迭代搜索得到两组位姿待选 解和两组匹配矩阵待选解。
步驟 7:若位姿待选解 1对应的目标函数值小于位姿待选解 2对应的目 标函数值, 则输出位姿待选解 1作为最终的位姿精确解, 匹配矩阵待选解 1 作为最终的匹配矩阵解, 否则输出位姿待选解 2作为最终的位姿精确解, 匹配矩阵待选解 2作为最终的匹配矩阵解。
如图 3所示, 下面给出一个实例的目标函数曲线图。 由图 3可以看到, 在精确位姿估计算法进行迭代搜索的结束阶段, 目标函数值基本不再发生 变换, 由于在每次迭代搜索时, 位姿估计解是基于位姿粗略解进行选择的, 所以精确位姿估计算法可以认为被限定在位姿粗略解的领域内进行搜索, 所以, 当位姿粗略解与真实位姿解相差很远时, 在它的领域进行搜索不会 得到好的位姿精确解, 同时, 它所对应的目标函数值也不会小于在与真实 位姿解相差不多的位姿粗略解领域搜索时所对应的目标函数值。 该程序的 输出的位姿解为: 俯仰角为 66.81 度, 方位角 60 度, 平移向量为 [-0.03;1.26;1611.40] (毫米), 俯仰角误差为 1.81度, 方位角误差为 0度, 距 离相对误差为 0.72%, 匹配正确率为 100%。
本发明实施例提供一种基于平面模型的点匹配与位姿同步解算的方 法, 本实施例能够解决二维射影变换下进行点匹配与位姿同步解算时估计 位姿解随迭代算法的进行呈指数增长的问题。 而且, 由于本实施例的位姿 计算方法是基于平面点集, 且与平面物体的几何形状无关, 所以能够减弱 对空间平面物体位姿计算的条件限制, 本实施例所公开的方法在遮挡、 杂 乱的环境下, 以及存在不同级别的噪声的情况下也能表现出较好的鲁棒性, 具有很好的应用价值。
显然, 本领域的技术人员应该明白, 上述的本发明的各步驟可以用通 用的计算装置来实现, 它们可以集中在单个的计算装置上, 或者分布在多 个计算装置所组成的网络上; 可选地, 它们可以用计算装置可执行的程序 代码来实现, 从而, 可以将它们存储在存储装置中由计算装置来执行, 或 者将它们分别制作成各个集成电路模块, 或者将它们中的多个模块或步驟 制作成单个集成电路模块来实现。 这样, 本发明不限制于任何特定的硬件 和软件结合。
为实现上述方法, 本发明还提供一种用于平面模型下实现点匹配与位 姿同步确定的计算机程序产品, 该产品内包括具有计算机可读程序编码的 计算机可用媒介, 该计算机可读程序编码用于执行平面模型下确定位姿的 方法; 该方法包括:
通过粗略位姿估计得到两组位姿粗略解;
以每组所述位姿粗略解作为初始迭代值, 通过扩展 TsPose算法经迭代 搜索得到位姿待选解;
基于所述位姿粗略解, 从所述位姿待选解中选出最终输出的位姿精确 解。
上述产品中, 所述通过扩展 TsPose算法经迭代搜索得到位姿待选解进 一步包括:
以每组所述位姿粗略解作为初始迭代值, 通过计算得到的初始 "像切 面" 像点获取匹配矩阵, 并对该匹配矩阵进行更新, 利用更新后的匹配矩 阵获取目标函数值;
更新仿射变换关系, 并根据更新后的仿射变换关系来得到旋转矩阵, 根据位姿粗略解中的粗略偏航角选择一个新的像切面和旋转矩阵, 并利用 所选择的新的像切面确定平移向量和新的 "像切面" 像点;
判断是否满足迭代收敛条件, 如果满足迭代收敛条件, 则结束迭代搜 索过程; 如果不满足迭代收敛条件, 则所确定的新的 "像切面" 像点进行 新一轮的迭代搜索, 直到满足迭代收敛条件;
所述位姿估计解包括所确定的旋转矩阵、 平移向量和 /或匹配矩阵; 满 足迭代收敛条件时得到的位姿估计解为所述位姿待选解。
上述产品中所采用的所述方法还包括:
在结束迭代搜索过程之后, 判断两组所述初始迭代值是否均已进行过 迭代, 若还有初始迭代值未进行迭代, 则以未进行迭代的初始迭代值为初 始迭代值, 通过扩展 TsPose算法经迭代搜索得到位姿待选解。 上述产品中, 所述迭代收敛条件为: 不确定度 大于结束不确定度; 其 中, 不确定度 为初始不确定度 A与不确定度的更新频率 A之积; 和 /或, 所述迭代收敛条件为: 所述目标函数值小于预先设置的表示找不到匹 配点参数 £ ;
任一所述迭代收敛条件得到满足时, 迭代搜索过程结束。
上述产品中, 所述基于所述位姿粗略解, 从所述位姿待选解中选出最 终输出的位姿精确解进一步包括:
在两组初始迭代值均完成迭代搜索之后, 根据每组所述位姿粗略解作 为初始迭代值时得到的目标函数值, 将较小的目标函数值所对应的位姿待 选解选出, 作为最终输出的位姿精确解。
以上所述, 仅为本发明的较佳实施例而已, 并非用于限定本发明的保 护范围。

Claims

权利要求书
1、 一种平面模型下点匹配与位姿同步的确定方法, 其特征在于, 该方 法包括:
通过粗略位姿估计得到两组位姿粗略解;
以每组所述位姿粗略解作为初始迭代值, 通过扩展 TsPose算法经迭代 搜索得到位姿待选解;
基于所述位姿粗略解, 从所述位姿待选解中选出最终输出的位姿精确 解。
2、 根据权利要求 1 所述的方法, 其特征在于, 所述通过扩展 TsPose 算法经迭代搜索得到位姿待选解进一步包括:
以每组所述位姿粗略解作为初始迭代值, 通过计算得到的初始 "像切 面" 像点获取匹配矩阵, 并对该匹配矩阵进行更新, 利用更新后的匹配矩 阵获取目标函数值;
更新仿射变换关系, 并根据更新后的仿射变换关系来得到旋转矩阵, 根据位姿粗略解中的粗略偏航角选择一个新的像切面和旋转矩阵, 并利用 所选择的新的像切面确定平移向量和新的 "像切面" 像点;
判断是否满足迭代收敛条件, 如果满足迭代收敛条件, 则结束迭代搜 索过程; 如果不满足迭代收敛条件, 则所确定的新的 "像切面" 像点进行 新一轮的迭代搜索, 直到满足迭代收敛条件;
所述位姿估计解包括所确定的旋转矩阵、 平移向量和 /或匹配矩阵; 满 足迭代收敛条件时得到的位姿估计解为所述位姿待选解。
3、 根据权利要求 2所述的方法, 其特征在于, 该方法还包括: 在结束迭代搜索过程之后, 判断两组所述初始迭代值是否均已进行过 迭代, 若还有初始迭代值未进行迭代, 则以未进行迭代的初始迭代值为初 始迭代值, 通过扩展 TsPose算法经迭代搜索得到位姿待选解。
4、 根据权利要求 2所述的方法, 其特征在于,
所述迭代收敛条件为: 不确定度 大于结束不确定度; 其中, 不确定度 为初始不确定度 A与不确定度的更新频率 之积; 和 /或,
所述迭代收敛条件为: 所述目标函数值小于预先设置的表示找不到匹 配点参数 £ ;
任一所述迭代收敛条件得到满足时, 迭代搜索过程结束。
5、 根据权利要求 2、 3或 4所述的方法, 其特征在于, 所述基于所述 位姿粗略解, 从所述位姿待选解中选出最终输出的位姿精确解进一步包括: 在两组初始迭代值均完成迭代搜索之后, 根据每组所述位姿粗略解作 为初始迭代值时得到的目标函数值, 将较小的目标函数值所对应的位姿待 选解选出, 作为最终输出的位姿精确解。
6、 一种用于平面模型下实现点匹配与位姿同步确定的计算机程序产 品, 其特征在于, 该产品内包括具有计算机可读程序编码的计算机可用媒 介, 该计算机可读程序编码用于执行平面模型下确定位姿的方法; 该方法 包括:
通过粗略位姿估计得到两组位姿粗略解;
以每组所述位姿粗略解作为初始迭代值, 通过扩展 TsPose算法经迭代 搜索得到位姿待选解;
基于所述位姿粗略解, 从所述位姿待选解中选出最终输出的位姿精确 解。
7、 根据权利要求 6 所述的产品, 其特征在于, 所述通过扩展 TsPose 算法经迭代搜索得到位姿待选解进一步包括:
以每组所述位姿粗略解作为初始迭代值, 通过计算得到的初始 "像切 面" 像点获取匹配矩阵, 并对该匹配矩阵进行更新, 利用更新后的匹配矩 阵获取目标函数值;
更新仿射变换关系, 并根据更新后的仿射变换关系来得到旋转矩阵, 根据位姿粗略解中的粗略偏航角选择一个新的像切面和旋转矩阵, 并利用 所选择的新的像切面确定平移向量和新的 "像切面" 像点;
判断是否满足迭代收敛条件, 如果满足迭代收敛条件, 则结束迭代搜 索过程; 如果不满足迭代收敛条件, 则所确定的新的 "像切面" 像点进行 新一轮的迭代搜索, 直到满足迭代收敛条件;
所述位姿估计解包括所确定的旋转矩阵、 平移向量和 /或匹配矩阵; 满 足迭代收敛条件时得到的位姿估计解为所述位姿待选解。
8、 根据权利要求 7所述的产品, 其特征在于, 所述方法还包括: 在结束迭代搜索过程之后, 判断两组所述初始迭代值是否均已进行过 迭代, 若还有初始迭代值未进行迭代, 则以未进行迭代的初始迭代值为初 始迭代值, 通过扩展 TsPose算法经迭代搜索得到位姿待选解。
9、 根据权利要求 7所述的产品, 其特征在于,
所述迭代收敛条件为: 不确定度 大于结束不确定度; 其中, 不确定度 为初始不确定度 A与不确定度的更新频率 之积; 和 /或,
所述迭代收敛条件为: 所述目标函数值小于预先设置的表示找不到匹 配点参数 £ ;
任一所述迭代收敛条件得到满足时, 迭代搜索过程结束。
10、 根据权利要求 7、 8或 9所述的产品, 其特征在于, 所述基于所述 位姿粗略解, 从所述位姿待选解中选出最终输出的位姿精确解进一步包括: 在两组初始迭代值均完成迭代搜索之后, 根据每组所述位姿粗略解作 为初始迭代值时得到的目标函数值, 将较小的目标函数值所对应的位姿待 选解选出, 作为最终输出的位姿精确解。
PCT/CN2011/083850 2011-12-12 2011-12-12 平面模型下点匹配与位姿同步确定方法及计算机程序产品 WO2013086678A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
US14/360,388 US9524555B2 (en) 2011-12-12 2011-12-12 Method and computer program product of the simultaneous pose and points-correspondences determination from a planar model
PCT/CN2011/083850 WO2013086678A1 (zh) 2011-12-12 2011-12-12 平面模型下点匹配与位姿同步确定方法及计算机程序产品

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2011/083850 WO2013086678A1 (zh) 2011-12-12 2011-12-12 平面模型下点匹配与位姿同步确定方法及计算机程序产品

Publications (1)

Publication Number Publication Date
WO2013086678A1 true WO2013086678A1 (zh) 2013-06-20

Family

ID=48611793

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2011/083850 WO2013086678A1 (zh) 2011-12-12 2011-12-12 平面模型下点匹配与位姿同步确定方法及计算机程序产品

Country Status (2)

Country Link
US (1) US9524555B2 (zh)
WO (1) WO2013086678A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110047106A (zh) * 2019-03-08 2019-07-23 深圳大学 用于增强现实的相机姿态估计方法、装置、设备及介质
CN110954083A (zh) * 2018-09-26 2020-04-03 苹果公司 移动设备的定位
CN111563882A (zh) * 2020-04-17 2020-08-21 广州番禺职业技术学院 一种药瓶倾斜校准与姿态调整机械结构装置及方法
CN111784746A (zh) * 2020-08-10 2020-10-16 上海高重信息科技有限公司 一种鱼眼镜头下行人多目标跟踪方法、装置及计算机系统

Families Citing this family (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104463953B (zh) * 2014-11-11 2017-06-16 西北工业大学 基于惯性测量单元与rgb‑d传感器的三维重建方法
CN108399377B (zh) * 2018-02-08 2022-04-08 北京理工大学 一种基于模式分类的光学定位方法
CN110544280B (zh) * 2018-05-22 2021-10-08 腾讯科技(深圳)有限公司 Ar系统及方法
US10817604B1 (en) 2018-06-19 2020-10-27 Architecture Technology Corporation Systems and methods for processing source codes to detect non-malicious faults
US10749890B1 (en) 2018-06-19 2020-08-18 Architecture Technology Corporation Systems and methods for improving the ranking and prioritization of attack-related events
US11295532B2 (en) 2018-11-15 2022-04-05 Samsung Electronics Co., Ltd. Method and apparatus for aligning 3D model
US11429713B1 (en) 2019-01-24 2022-08-30 Architecture Technology Corporation Artificial intelligence modeling for cyber-attack simulation protocols
US11128654B1 (en) 2019-02-04 2021-09-21 Architecture Technology Corporation Systems and methods for unified hierarchical cybersecurity
US10832444B2 (en) * 2019-02-18 2020-11-10 Nec Corporation Of America System and method for estimating device pose in a space
US11451581B2 (en) 2019-05-20 2022-09-20 Architecture Technology Corporation Systems and methods for malware detection and mitigation
US10930012B2 (en) * 2019-05-21 2021-02-23 International Business Machines Corporation Progressive 3D point cloud segmentation into object and background from tracking sessions
US11403405B1 (en) 2019-06-27 2022-08-02 Architecture Technology Corporation Portable vulnerability identification tool for embedded non-IP devices
US11444974B1 (en) 2019-10-23 2022-09-13 Architecture Technology Corporation Systems and methods for cyber-physical threat modeling
CN112580223B (zh) * 2019-12-30 2022-08-02 北京航空航天大学 一种基于选区应力判据的双金属界面自适应加载模拟方法
US11503075B1 (en) 2020-01-14 2022-11-15 Architecture Technology Corporation Systems and methods for continuous compliance of nodes
CN113744277A (zh) * 2020-05-29 2021-12-03 广州汽车集团股份有限公司 一种基于局部路径优化的视频去抖方法及系统

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20080298672A1 (en) * 2007-05-29 2008-12-04 Cognex Corporation System and method for locating a three-dimensional object using machine vision
CN101377812A (zh) * 2008-07-11 2009-03-04 北京航空航天大学 一种空间平面物体位姿识别方法

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
AU2001250802A1 (en) * 2000-03-07 2001-09-17 Sarnoff Corporation Camera pose estimation
US7706603B2 (en) * 2005-04-19 2010-04-27 Siemens Corporation Fast object detection for augmented reality systems
US7844106B2 (en) * 2007-04-23 2010-11-30 Mitsubishi Electric Research Laboratories, Inc Method and system for determining poses of objects from range images using adaptive sampling of pose spaces
US20090110267A1 (en) * 2007-09-21 2009-04-30 The Regents Of The University Of California Automated texture mapping system for 3D models
US8600192B2 (en) * 2010-12-08 2013-12-03 Cognex Corporation System and method for finding correspondence between cameras in a three-dimensional vision system
US10033979B2 (en) * 2012-03-23 2018-07-24 Avigilon Fortress Corporation Video surveillance systems, devices and methods with improved 3D human pose and shape modeling
US20150371440A1 (en) * 2014-06-19 2015-12-24 Qualcomm Incorporated Zero-baseline 3d map initialization

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20080298672A1 (en) * 2007-05-29 2008-12-04 Cognex Corporation System and method for locating a three-dimensional object using machine vision
CN101377812A (zh) * 2008-07-11 2009-03-04 北京航空航天大学 一种空间平面物体位姿识别方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
CHEN YIFENG: "PLANAR POSE ESTIMATION BAESED ON MONOCULAR VISION", SCIENCE & TECHNOLOGY INFORMATION, vol. 17, 20 June 2011 (2011-06-20), pages 33 - 34 *

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110954083A (zh) * 2018-09-26 2020-04-03 苹果公司 移动设备的定位
CN110954083B (zh) * 2018-09-26 2023-11-24 苹果公司 移动设备的定位
CN110047106A (zh) * 2019-03-08 2019-07-23 深圳大学 用于增强现实的相机姿态估计方法、装置、设备及介质
CN111563882A (zh) * 2020-04-17 2020-08-21 广州番禺职业技术学院 一种药瓶倾斜校准与姿态调整机械结构装置及方法
CN111563882B (zh) * 2020-04-17 2023-04-18 广州番禺职业技术学院 一种药瓶倾斜校准与姿态调整机械结构装置的控制方法
CN111784746A (zh) * 2020-08-10 2020-10-16 上海高重信息科技有限公司 一种鱼眼镜头下行人多目标跟踪方法、装置及计算机系统
CN111784746B (zh) * 2020-08-10 2024-05-03 青岛高重信息科技有限公司 一种鱼眼镜头下行人多目标跟踪方法、装置及计算机系统

Also Published As

Publication number Publication date
US20140321735A1 (en) 2014-10-30
US9524555B2 (en) 2016-12-20

Similar Documents

Publication Publication Date Title
WO2013086678A1 (zh) 平面模型下点匹配与位姿同步确定方法及计算机程序产品
CN106909877B (zh) 一种基于点线综合特征的视觉同时建图与定位方法
Sim et al. Recovering camera motion using l\infty minimization
Hartley et al. PowerFactorization: 3D reconstruction with missing or uncertain data
Gotardo et al. Computing smooth time trajectories for camera and deformable shape in structure from motion with occlusion
Hartley et al. Optimal algorithms in multiview geometry
Mirzaei et al. Globally optimal pose estimation from line correspondences
Bellekens et al. A benchmark survey of rigid 3D point cloud registration algorithms
Aftab et al. Convergence of iteratively re-weighted least squares to robust m-estimators
JP5012615B2 (ja) 情報処理装置、および画像処理方法、並びにコンピュータ・プログラム
Haner et al. Covariance propagation and next best view planning for 3d reconstruction
Zhao et al. Parallaxba: bundle adjustment using parallax angle feature parametrization
US20130080111A1 (en) Systems and methods for evaluating plane similarity
EP3086285A1 (en) Method of camera calibration for a multi-camera system and apparatus performing the same
CN112444246A (zh) 高精度的数字孪生场景中的激光融合定位方法
Sommer et al. Continuous-time estimation of attitude using b-splines on lie groups
Horn et al. Online extrinsic calibration based on per-sensor ego-motion using dual quaternions
Dubbelman et al. Efficient trajectory bending with applications to loop closure
Kukelova et al. Hand-eye calibration without hand orientation measurement using minimal solution
Aravkin et al. Student's t robust bundle adjustment algorithm
Comport et al. Kinematic sets for real-time robust articulated object tracking
Jia et al. Low-rank matrix fitting based on subspace perturbation analysis with applications to structure from motion
Luong et al. Consistent ICP for the registration of sparse and inhomogeneous point clouds
Guillemaut et al. Using points at infinity for parameter decoupling in camera calibration
Lee et al. Robust uncertainty-aware multiview triangulation

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 11877274

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 14360388

Country of ref document: US

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 11877274

Country of ref document: EP

Kind code of ref document: A1