CN113920180B - Point cloud registration optimization method based on normal distribution transformation hypothesis verification - Google Patents
Point cloud registration optimization method based on normal distribution transformation hypothesis verification Download PDFInfo
- Publication number
- CN113920180B CN113920180B CN202111487497.0A CN202111487497A CN113920180B CN 113920180 B CN113920180 B CN 113920180B CN 202111487497 A CN202111487497 A CN 202111487497A CN 113920180 B CN113920180 B CN 113920180B
- Authority
- CN
- China
- Prior art keywords
- point cloud
- grid
- delta
- function
- target
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
- G06T7/33—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods
- G06T7/344—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods involving models
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10028—Range image; Depth image; 3D point clouds
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
Abstract
The invention provides a point cloud registration optimization method based on NDT hypothesis verification, which comprises the following steps of: acquiring reference point clouds, performing three-dimensional grid division on the reference point clouds and calculating normal distribution parameters of the point clouds in each grid; constructing a local target function of the point cloud in the grid to calculate a check index delta of the NDT hypothesis; designing a conversion function to convert the delta into the weight of the three-dimensional grid in the global objective function, calculating the grid to which the target point belongs according to the three-dimensional coordinates of the target point cloud to obtain the corresponding weight, and substituting the weight into the global objective function; and finally, solving a global objective function with weight to obtain a point cloud registration result. By adopting the method, the global target function can be assisted to be converged to a more coincident state of the reference point cloud and the target point cloud through the weight of the target point cloud, and the stability and the accuracy of point cloud registration are improved.
Description
Technical Field
The invention belongs to the technical field of point cloud registration, and particularly relates to a point cloud registration optimization method based on normal distribution transformation hypothesis verification.
Background
The point cloud data is a set of vectors in a three-dimensional coordinate system, and the vectors may include information such as color, time, reflection intensity, and the like in addition to three-dimensional coordinates. The point cloud is generally obtained by scanning the surrounding environment by a laser radar, a depth camera and other equipment. In the field of mobile robots, a point cloud is often used to represent the surface geometry of a scene surrounding the robot.
Point cloud registration refers to: the target point cloud and the reference point cloud under different time or view angles are unified into a complete point cloud originally under the coordinate system of the reference point cloud through rotation and translation motion in a three-dimensional space. The point cloud registration is an important component of the mapping and positioning functions of the mobile robot, and determines the stability and accuracy of the robot during motion.
Martin Magnusson, university of Irela Blu, Sweden[1]Firstly, providing an original edition NDT algorithm applied to three-dimensional point cloud registration, assuming that the three-dimensional geometric shape of local point cloud conforms to multivariate normal distribution, calculating the mean value and covariance matrix of the local point cloud in a three-dimensional grid, and converting the point cloud in original discrete distribution into normal distribution expression with first-order and second-order derivatives; then, the method designs a global objective function for representing the registration state of the reference point cloud and the target point cloud based on normal distribution transformation, and solves the global objective function by utilizing a Levenberg-Marquardt algorithm to obtain a point cloud registration result. However, the preset using equipment of the original NDT point cloud registration algorithm is mainly an underground mining robot, and the geometric surface of an underground mine hole usually has more continuous line and surface characteristics and is uniformly distributed, so that point cloud data corresponding to the mine hole is more in line with the assumption of NDT midpoint cloud normal distribution. However, as the NDT point cloud registration algorithm is more and more applied to ground or aerial robots, the original local point cloud in the grid conforms to the assumption of normal distribution, and the three-dimensional surface shapes of various objects in the nature cannot be covered, so that the original NDT point cloud registration algorithm cannot be stably performed in some disordered scenes.
Ehsan Javanmandi of university of Tokyo, Japan[2]In original NDT point cloud registration, the size of a three-dimensional grid cannot be adjusted in a self-adaptive mode according to the change of the environment, when the size of the grid is large, NDT parameters of the grid can represent more point clouds, the calculation complexity is reduced, but detail information of the point clouds can be ignored, and the registration accuracy is reduced; when the grid size is small, the complexity of the map is increased, and the computational complexity is increased. Therefore, the method designs 10 evaluation factors for confirming the optimal grid resolution at each position off-line, reduces the complexity of the map and does not influence the point cloud registration precision. However, this method needs to recalculate the evaluation factor at each position of the map, and the steps are complicated and only suitable for point cloud registration of a specific small scene.
[1] Martin Magnusson. The Three-Dimensional Normal-Distributions Transform- an Efficient Representation for Registration,Surface Analysis, and Loop Detection[D]. Örebro: Örebro University, 2009。
[2] Javanmardi E , Javanmardi M , Y Gu, et al. Adaptive Resolution Refinement of NDT Map Based on Localization Error Modeled by Map Factors[C]. 2018 IEEE International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2018。
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a point cloud registration optimization method based on normal distribution transformation hypothesis verification. The technical conception of the invention is as follows:
in view of the "extremum registration" assumption present in the original NDT algorithm: under the premise of assuming that local point clouds in a three-dimensional grid of the reference point clouds accord with normal distribution, a global objective function is designed to represent the registration state between the target point clouds and the reference point clouds, the independent variables of the global objective function are three-dimensional rotation and translation motion between the target point clouds and the reference point clouds, and when the output of the global objective function is a global minimum value, the values corresponding to the independent variables are the optimal point cloud registration result.
Based on the assumption of the original NDT, the invention calculates the hypothesis verification index delta from two angles of 'extreme convergence point' and 'average probability score':
1) constructing a local target function of each three-dimensional grid in the reference point cloud, recording the corresponding rotational translation motion of the local target function when the local target function reaches a minimum convergence point as p1, and taking the modulus value of p1 as a check index delta1；
2) Calculating the minimum value of the local target function of each three-dimensional grid in the reference point cloud, and dividing the minimum value by the number of the point clouds to obtain a check index delta2。
Will delta1,δ2And obtaining a complete three-dimensional grid check index delta after combination. In order to enable the global target function to more accurately represent the registration state of the reference point cloud and the target point cloud, the delta needs to be converted into the weight of each three-dimensional grid in the global target function, and when the global target function is used, the global target function can more accurately represent the registration state of the reference point cloud and the target point cloudWhen the function is solved, the grid with high weight provides more constraint information, and the grid with low weight provides relatively less constraint information, so that the calculated point cloud result is more stable and reliable.
Therefore, the invention provides a point cloud registration optimization method based on normal distribution transformation hypothesis verification, which comprises the following steps:
(1) acquiring input reference point cloud, and performing three-dimensional grid division on the reference point cloud according to the resolution of a fixed size to obtain a set V, V = { V } of three-dimensional grids1,v2,...,vnN is the number of the stereoscopic grids;
(2) for each stereo grid in the set V, calculating V when the stereo grid conforms to normal distribution transformationμAnd vΣWherein v isμMean value of coordinates, v, representing the point cloud in a three-dimensional gridΣA covariance matrix representing coordinates of the point cloud within the grid;
(3) based on vμAnd vΣCalculating a check index delta for quantifying the coincidence of the three-dimensional grid with the NDT hypothesis degree, and adjusting the resolution of the three-dimensional grid according to the magnitude of a modulus of the delta;
(4) designing a conversion function, and converting the index delta into the weight of the grid in the point cloud registration target function through the conversion function;
(5) constructing a global target function of registration of the weighted target point cloud and the reference point cloud based on an NDT model of the reference point cloud three-dimensional grid;
(6) and substituting the target point cloud into a global target function for registering the target point cloud and the reference point cloud, and iteratively solving the limit of the function to obtain a point cloud registration result.
Further, in the above step (3), based on vμAnd vΣAnd constructing a local objective function of the stereoscopic grid. The arguments of the local objective function are the three-dimensional rotational and translational motions and the output is the registration probability score of the target point cloud at the stereo grid and the reference point cloud. If the stereo grid conforms to the NDT "extreme registration" assumption, then the reference point cloud is substituted into the local objective function, which should converge to the global extreme of the probability score when the rotation and translation motion is 0At small values. Conversely, when the stereo grid does not conform to the NDT assumption, the probability score of the local objective function cannot converge to a global minimum when the rotation and translation motion is 0. Therefore, the invention calculates the index delta from the angle of the 'extremum convergence point' of the local objective function1Calculating the index delta from the angle of' average probability score2Will delta1And delta2And obtaining the final check index delta after combination.
Further, in the step (3), the check index δ is divided into δ1And delta2Two parts.
δ can be calculated as follows1: first according to vμAnd vΣEstablishing an NDT model of point clouds in the three-dimensional grid, substituting all point cloud data in the grid into the NDT model to obtain a local objective function, solving the limit of the local objective function, recording the rotation and translation motions corresponding to the minimum value as p1, wherein the module value of p1 is the check index delta1。δ1The smaller the size, the more accurate the NDT model is to model the point cloud data, i.e., the more the stereo grid conforms to the NDT assumption.
δ can be calculated as follows2: substituting p1 into the local objective function of the three-dimensional grid, calculating the total probability score of the point clouds in the grid, and dividing the total score by the number of the point cloud data to obtain the verification index delta2. If the stereo grid conforms to the NDT "extreme registration" assumption, then δ2Is the global minimum, δ, of the local objective function2The smaller the grid, the more compliant the NDT assumption is.
Further, in step (3), the larger the resolution of the stereo grid is, the more the point cloud data inside the grid conforms to the normal distribution model in the NDT assumption, but the more the calculation amount of point cloud registration is. The resolution of the stereoscopic grid can be automatically adjusted according to the module value of the check index delta. Firstly, a threshold value theta is set according to the working environment of the robot, then the magnitude of delta and theta is compared through a module function Norm, and when Norm (delta) > = theta, the fact that the stereo grid does not meet the NDT extreme value registration assumption is shown, and the subsequent point cloud registration does not work. At this time, the original grid needs to be divided into a plurality of smaller grids to improve the resolution, so as to extract more detailed information of the point cloud, and to recalculate the δ of the divided grid.
Further, in step (4), the importance of each stereo grid in the reference point cloud in the point cloud registration global objective function is related to the extent to which the grid conforms to the NDT assumption. The higher the degree to which the stereo grid conforms to the NDT assumption, the more accurate the global objective function characterizes the registration state of the reference point cloud and the target point cloud. Since the modulus of the check index δ reflects the degree that the stereoscopic grid conforms to the NDT assumption, in order to enable the global objective function of point cloud registration to more accurately represent the registration state between the reference point cloud and the target point cloud, a conversion function needs to be designed to convert the modulus of δ into the weight of the stereoscopic grid in the global objective function.
Further, in step (4), the output of the transfer function is the weight w of the stereo grid, the input is the Norm (δ) of δ, and the transfer requirement of the sum w of the Norm (δ) is analyzed as follows: the smaller the Norm (δ), the more the grid conforms to the NDT assumption, the larger the w need, the negative correlation of Norm (δ) need and w; norm (δ) is minimum 0, w is the maximum value needed; in step (3), when Norm (δ) > = θ, the grid will be subdivided, and w needs to be the minimum value at this time.
Further, in step (4), the transfer function may be designed according to the following specific method: firstly, the value range of the weight w is limited between a minimum value 0 and a maximum value t, wherein t is larger than 0. When Norm (δ) =0, the degree to which the stereoscopic grid conforms to the NDT assumption is the highest, and w takes the maximum value t; when Norm (δ) > = θ, the stereo grid does not work for subsequent point cloud registration, w takes the minimum value of 0, i.e., the transfer function needs to pass through the keypoints (0, t) and (θ, 0). The addition of the coordinates of the key points (Norm (δ), w) satisfying the negative correlation condition can then be continued between 0 and θ, the smaller the Norm (δ), the larger w, such as: key point coordinates (0.1, 0.8), (0.3, 0.4), etc. Finally, in order to calculate the corresponding weight w for any input Norm (δ), curve fitting needs to be performed on the coordinates of the above key points to obtain the functional relationship between Norm (δ) and w. Commonly used fitting functions are: the invention selects a polynomial function which can obtain accurate fitting, and needs to be noted that when Norm (delta) > = theta, w is always 0.
Further, in step (5), the weighted global objective function is constructed according to the following method: taking the translation and rotation motion of the target point cloud as variables to be optimized of a global target function, expressing the variables with a symbol T, traversing each data point in the target point cloud, carrying out three-dimensional transformation on each data point according to T, calculating a three-dimensional grid to which each data point belongs according to the transformed three-dimensional coordinates, and calculating the v of the three-dimensional gridμAnd vΣAnd obtaining an NDT model of the grid, and obtaining the weight w of the grid according to a transfer function. And substituting each data point into the respective NDT model and multiplying by the weight w to obtain the probability score of each data point. And accumulating and summing the probability scores of all the data points in the target point cloud to obtain a global target function of point cloud registration.
Further, in step (6), the global objective function is solved iteratively according to the following method: substituting the target point cloud, the initial rotation and translation motion T of the target point cloud in the global target function and the weight of the target point cloud into the global target function, and then calculating the optimal motion variable quantity Δ T of the target point cloud and the reference point cloud by using an extremum solving method. When the mode value of the max T is larger than the threshold value of the extremum solving method, it is described that the global function is not converged, the value of the initial motion T needs to be updated to T +. T, the target point cloud is subjected to three-dimensional transformation according to T +. T, the weight of each data point in the target point cloud is recalculated and substituted into the global target function, and the optimal motion variable quantity Δ T is calculated by the extremum solving method again. When the mode value of the T is smaller than the threshold value, the global objective function is described to be converged, and the point cloud registration result at this time is T +. Δ T.
According to the scheme, the registration state between the reference point cloud and the target point cloud can be more accurately represented by the global target function through the weight of each three-dimensional grid in the global target function, when the global target function is subjected to iterative solution, the grid with high weight better conforms to the NDT assumption, more constraint information can be provided, and conversely, the grid with low weight can provide relatively less constraint information, so that the global target function can be converged to a more reasonable minimum value state, and the corresponding point cloud registration result is more stable and reliable.
Drawings
In order to more clearly illustrate the technical solution of the present invention, the drawings used in the description of the embodiments will be briefly described below.
FIG. 1 is a schematic flow chart of a point cloud registration optimization method based on a normal distribution transformation hypothesis verification according to an embodiment of the present invention;
FIG. 2 is a schematic diagram illustrating a reference point cloud three-dimensional grid division according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a point cloud distribution shape and corresponding normal distribution parameters in a three-dimensional grid according to an embodiment of the present invention;
FIG. 4 is a diagram illustrating a cloud hypothesis verification index δ for computing points in a three-dimensional grid according to an embodiment of the present invention1A schematic diagram of (a);
FIG. 5 is a diagram illustrating the conversion of δ into a weight w by various conversion functions according to an embodiment of the present invention;
FIG. 6 is an output trace effect diagram of an original NDT point cloud registration method in accordance with an embodiment of the present invention;
fig. 7 and 8 are output trajectory effect diagrams of the point cloud registration method after NDT hypothesis verification according to the embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
As shown in fig. 1, the point cloud registration optimization method based on the verification of the assumption of normal distribution transformation of the present embodiment includes the following steps:
s1, obtaining a reference point cloud input at a certain moment, and performing three-dimensional grid division on the reference point cloud according to the resolution of a fixed size to obtain a set V, V = { V } of three-dimensional grids1,v2,...,vnN is the number of the stereoscopic grids.
S2, for each stereoscopic grid in the set V, calculating V when the stereoscopic grid conforms to normal distribution transformationμAnd vΣWherein v isμMean value of coordinates, v, representing the point cloud in a three-dimensional gridΣA covariance matrix representing coordinates of the point cloud within the grid.
S3, based on vμAnd vΣAnd constructing a local objective function of each stereoscopic grid in the V, calculating a check index delta for quantifying the stereoscopic grid meeting the NDT (normalized difference test) hypothesis degree, and adaptively adjusting the resolution of the grid according to the magnitude of a modulus of the delta.
Step S3 includes the following aspects:
1) NDT parameter v based on stereo gridμAnd vΣAnd constructing a local objective function of each stereoscopic grid in the V. The arguments of the local objective function are the three-dimensional rotational and translational motions and the output is the registration probability score of the target point cloud at the stereo grid and the reference point cloud. If the stereo grid conforms to the NDT "extreme registration" assumption, then when substituting the reference point cloud into the local objective function, the local objective function should converge to the global minimum of the probability score when the rotational and translational motion is 0. Therefore, the present embodiment calculates the index δ from the angle of "extremum convergence point" of the local objective function1Calculating the index delta from the angle of' average probability score2Will delta1And delta2And obtaining the final check index delta after combination.
2) Based on the local objective function, a check index delta of the stereo grid is calculated, wherein the delta represents the degree of the grid meeting the NDT extreme value registration assumption. Calculating delta from the angle of the "extremum convergence point1: first according to vμAnd vΣEstablishing an NDT model of point clouds in the three-dimensional grid, substituting all point cloud data in the grid into the NDT model to obtain a local target function, solving the limit of the local target function, recording the rotation and translation motions corresponding to the minimum value of the local target function as p1, wherein the module value of p1 is the check index delta1。δ1The smaller the size, the more accurate the NDT model is to model the point cloud data, i.e., the more the stereo grid conforms to the NDT assumption.
Calculating delta from the perspective of the "average probability score2: substitution of the rotational and translational movements p1 into the local objective function of the spatial gridCalculating the total probability score of the point cloud data in the grid at the moment, and dividing the total score by the number of the point cloud data to obtain a check index delta2. If the stereo grid conforms to the NDT "extreme registration" assumption, then δ2Is the global minimum, δ, of the local objective function2The smaller the stereo grid, the more compliant the NDT "extreme registration" assumption.
3) The larger the resolution of the stereo grid, the more the point cloud data inside the grid conforms to the normal distribution model in the NDT assumption, but the more computationally intensive the point cloud registration. The resolution of the stereoscopic grid may be adjusted to a suitable size according to the modulus value of the check index δ. Firstly, a threshold value theta is set according to the working environment of the robot, then the magnitude of delta and theta is compared through a module function Norm, and when Norm (delta) > = theta, the fact that the stereo grid does not meet the NDT extreme value registration assumption is shown, and the subsequent point cloud registration does not work. If more useless three-dimensional grids exist, the global objective function cannot be converged, and the point cloud registration performance is reduced. For example, θ =0.1 and Norm (δ) =1.2, where Norm (δ) is much larger than θ, which indicates that the stereo grid corresponding to δ does not meet the NDT assumption.
Therefore, when Norm (δ) > = θ, the original grid needs to be divided into a plurality of smaller grids to improve the resolution, so as to extract more detailed information of the point cloud. Common partitioning methods are: equally dividing, clustering, etc. For example, two cluster centers of a rectangular frame and a triangular frame are obviously present in the three-dimensional grid V2 in fig. 3, and V2 is obviously not in accordance with the NDT assumption, but if V2 can be divided into two small grids by equal division, so that only one cluster is present in each small grid, the divided small grids are more in accordance with the NDT assumption. And after the divided grids are obtained, recalculating the check index delta of the divided grids.
And S4, designing a conversion function, and converting the check index delta into the weight of the three-dimensional grid in the point cloud registration target function. Since each stereo grid in the reference point cloud conforms to the NDT assumption to a higher degree, its characterization of the registration state of the reference point cloud and the target point cloud in the global objective function is more accurate. Thus, by converting the modulus values of δ to weights of the stereoscopic grid in the point cloud registration global objective function through the conversion function, a grid with a high weight will provide more constraint information and a grid with a low weight will provide relatively less constraint information when the global objective function is solved.
Let the output of the transfer function be the weight w of the stereo grid, and the input be the Norm (δ) of δ, and analyze the transfer requirements of Norm (δ) and w as follows: the smaller Norm (delta), the more the grid conforms to the NDT "extreme registration" assumption, and the larger w needs, indicating that Norm (delta) needs to be negatively correlated with w; when Norm (δ) takes the minimum value of 0, w needs to take the maximum value, and when Norm (δ) takes the maximum value, w needs to take the minimum value, which indicates that the transfer function needs to pass through some key point coordinates in addition to the negative correlation.
The present embodiment designs the transfer function according to the following method: first, for comparison and calculation, the value range of the weight w is limited to a minimum value of 0 to a maximum value t. Secondly, when Norm (delta) =0, let w obtain the maximum value t, t is a positive number greater than 0; since the grid needs to be subdivided when Norm (δ) > = θ in S303, w =0, that is, the transfer function needs to pass through the key points (0, t) and (θ, 0), and the magnitudes of θ and t can be set according to the scene requirements, such as θ =0.1, 0.2, etc., t =1,2, etc. Then, a number of key point coordinates satisfying a negative correlation condition, such as (0.1, 0.8), (0.3, 0.4), etc., may be added continuously between 0 and θ. Finally, in order to calculate the corresponding weight w for any input Norm (δ), curve fitting needs to be performed on the coordinates of the above key points to obtain the functional relationship between Norm (δ) and w. Commonly used fitting functions are: a logarithmic function, an exponential function, a polynomial function, etc. in this embodiment, a polynomial function that can obtain an accurate fit is selected, and it should be noted that w is always 0 when Norm (δ) > = θ. As shown in the function curve in fig. 6, the weight w gradually decreases from 1 to 0 when the abscissa Norm (δ) changes from 0 to 0.1, and the weight w is always 0 when Norm (δ) > = 0.1.
S5, constructing a global object function for registration of the weighted target point cloud and the weighted reference point cloud, wherein the specific method comprises the following steps: taking the translation and rotation motion of the target point cloud as variables to be optimized of the global target function, and taking symbolsT represents, each data point in the target point cloud is traversed, three-dimensional transformation is carried out on each data point according to T, a three-dimensional grid to which each data point belongs is calculated according to transformed three-dimensional coordinates, and v of the three-dimensional grid is calculated according toμAnd vΣAnd obtaining an NDT model of the grid, and obtaining the weight w of the grid according to a transfer function. And substituting each data point into the respective NDT model and multiplying by the weight w to obtain the probability score of each data point. And accumulating and summing the probability scores of all the data points in the target point cloud to obtain a global target function of point cloud registration.
The weight in the global objective function enables the function to accurately represent the registration state between the reference point cloud and the target point cloud, when the global objective function is subjected to iterative solution, grids with high weight better accord with NDT hypothesis, more constraint information can be provided, and conversely, grids with low weight can provide relatively less constraint information.
And S6, substituting the target point cloud into the global target function, wherein the smaller the output of the function is, the higher the registration degree of the target point cloud and the reference point cloud is, and the higher the probability of the occurrence of the corresponding rotation and translation motion T is. Therefore, the optimal point cloud registration result can be obtained by iteratively solving the minimum value of the global objective function through methods such as gauss-newton, gradient descent, LM and the like.
The minimum solving process in this embodiment is as follows: substituting the target point cloud, the initial rotation and translation motion T of the target point cloud in the global target function and the weight of the target point cloud into the global target function, and then calculating the optimal motion variable quantity Δ T of the target point cloud and the reference point cloud by using an extremum solving method. When the mode value of the max T is larger than the threshold value of the extremum solving method, it is described that the global function is not converged, the value of the initial motion T needs to be updated to T +. T, the target point cloud is subjected to three-dimensional transformation according to T +. T, the weight of each data point in the target point cloud is recalculated and substituted into the global target function, and the optimal motion variable quantity Δ T is calculated by the extremum solving method again. When the mode value of the margin T is smaller than the threshold value of the extremum solving method, the global objective function is converged, and the point cloud registration result at this time is T +. DELTA.T.
Embodiments of the present invention relate to an "extremum registration" assumption and validation of this assumption in NDT point cloud registration methods. Specifically, the embodiment of the invention designs an index delta for checking the NDT hypothesis rationality from two angles of 'extreme convergence points' and 'average probability scores' of a local objective function, and improves the accuracy of point cloud registration after converting the delta into the weight of a global objective function.
The following is a specific application example of the point cloud registration optimization method based on the verification of the assumption of normal distribution transformation.
In the embodiment, the point cloud data obtained by scanning the urban road environment by the three-dimensional laser radar is adopted, and according to the following implementation mode, the input reference point cloud and the target point cloud can be registered to obtain the continuous running track of the robot.
The overall point cloud data is composed of a plurality of basic points, lines and surfaces in shape, and the more complex the shape is, the more the shape cannot be represented by a single probability distribution model. Therefore, after obtaining the input reference point cloud of the laser radar, the reference point cloud can be divided into continuous three-dimensional grids according to a fixed resolution, as shown in fig. 2. In fig. 2, according to the rule of the right-hand coordinate system, the left-hand diagram shows the top view of the input reference point cloud, the X-axis represents the front of the laser radar, and the Y-axis represents the front left of the radar; the right diagram shows a front view of the reference point cloud, and the Z axis shows the position right above the radar; in the figure, a solid line indicates a boundary of the point cloud data, and a dotted line indicates a boundary when the point cloud data is raster-divided.
The original NDT point cloud registration method assumes that the point clouds in each solid grid conform to a normal distribution. For the point cloud data in this embodiment, because the geometric shape of the scene is relatively complex, there is a phenomenon of "multiple clusters" or "hollows" in some point clouds in the stereo grid, where "multiple clusters" means that the point clouds obviously have equal to or more than two set centers, and "hollows" means that the distribution of the point clouds at the cluster centers is relatively sparse, which may cause a case that the assumption of the original NDT registration method fails, as shown in fig. 3 in particular. In fig. 3, the upper left represents the point cloud distribution within the stereoscopic grid V1, the 8 three-dimensional points in V1 are uniformly distributed within the grid centered on the coordinates [37.6,0.62, -1.51], and the elements [0.160,0.163,0.11] on the diagonal of the covariance matrix represent that the point cloud shape within V1 is close to a cube in space, which approximately coincides with the actual point cloud shape, indicating that V1 conforms to the assumption of the NDT registration method. However, in the bottom left V2, the point cloud within the grid clearly can partition two sets of clusters of rectangular and triangular boxes, which do not fit into a normal distribution centered at [9.8,5.5,2.4], indicating that the assumption of the NDT registration method fails on V2.
If a stereo grid that does not conform to the NDT assumption, such as V2, is added to the global objective function for point cloud registration, the NDT representation of the grid and the actual point cloud distribution do not coincide, resulting in the target point cloud not being registered or directly matched to the reference point cloud in the wrong location. On the contrary, if the check index of the three-dimensional grid to the NDT hypothesis can be calculated, and the importance of the grid corresponding to the check index in the global objective function is distributed according to the check index, the accuracy and stability of point cloud registration can be improved.
To verify whether the "extremum registration" assumption in the original NDT holds, the present embodiment calculates the verification index from two angles of "extremum convergence point" and "average probability score" of the local objective function of the stereoscopic grid, respectively. The point cloud data is obtained by scanning a laser 16-line laser radar, the three-dimensional size S =1m of the three-dimensional grid, and an LM (Levenberg-Marquarelt, Levenberg-Marquardt) iterative algorithm is adopted in an extreme value solving method.
First, a brief introduction is made to the "extremum registration" assumption in the original NDT method to facilitate understanding of the subsequent assumption check indicators, and the main steps of the original NDT method are briefly introduced as follows:
obtaining an input reference point cloud and performing three-dimensional grid division on the input reference point cloud to obtain a three-dimensional grid set V, V = { V =1,v2,...,vnN is the number of the stereoscopic grids. For each grid in V, calculating V when it conforms to normal distribution transformationμAnd vΣWherein v isμMean value of coordinates, v, representing a stereoscopic gridΣA covariance matrix representing the stereoscopic grid. In the hypothesis of the gridOn the premise that the internal point cloud conforms to the normal distribution, constructing a global target function for registering the target point cloud and the reference point cloud as follows:
wherein p represents the rotational and translational motion to be solved, x represents the three-dimensional coordinates of a point in the target point cloud, m represents the number of the target point cloud, Tr represents the motion transformation represented by applying p to x, and Pr represents the basis of vμAnd vΣThe constructed probability distribution model, f (p), represents the probability score of the overall target point cloud. The smaller the F (p), the more the distribution of the target point cloud conforms to the probability distribution model of the solid grid.
Analyzing the original NDT point cloud registration method can obtain: assuming that f (p) is the minimum value, the reference point cloud and the target point cloud are completely registered, when the reference point cloud in the stereoscopic grid is input into the local target function of the grid instead of the target point cloud, since the target point cloud and the reference point cloud replaced at this time are completely the same, the local target function should obtain the minimum value when the rotational and translational motion p is 0. On the contrary, when the "extreme value registration" assumption is not satisfied, the local objective function cannot obtain a minimum value when p is 0. Thus, the check index δ of the NDT "extreme registration" hypothesis can be divided into δ1And delta2Two parts of which delta1The method comprises the following steps: when the local objective function reaches an extreme value convergence point, rotating and translating the module value of the motion p; delta2The method comprises the following steps: the probability of the reference point cloud itself on the local objective function is averaged. The detailed calculation method of the two methods is as follows:
1) check indicator delta1: first according to v of the three-dimensional gridμAnd vΣAnd establishing an NDT model of the point cloud in the grid, substituting the reference point cloud in the grid for the target point cloud, and substituting the target point cloud into the NDT model to obtain a local target function. The arguments of the local objective function are the three-dimensional rotational and translational motions and the output is the registration probability score of the target point cloud at the stereo grid and the reference point cloud. If the stereo grid conforms to the NDT "extreme registration" assumption, thenThe local objective function should converge to a global minimum of the probability score when the rotational and translational motion is 0. Conversely, when the point cloud in the stereoscopic grid is in a situation that does not conform to the NDT assumption, such as "multi-cluster" or "hollow", the rotation and translation motions gradually move away from the zero point along with the iteration of the extremum solving method. Solving the limit of the local target function, recording the rotation and translation motion corresponding to the minimum value of the local target function as p1, and taking the module value of p1 as the verification index delta1。δ1The smaller the size, the more accurate the NDT model is to model the point cloud data, i.e., the more the stereo grid conforms to the NDT assumption.
The method of gauss newton, gradient descent, LM, etc. may be used to calculate the limit of the local objective function, in this embodiment, the LM method is used to calculate the minimum value of the local objective function, and the Norm (p1) method is used to calculate δ1Wherein Norm (p1) represents the modulo of p1, detailed procedure as follows:
assuming that f (p) represents a local objective function of a certain stereo grid, first order taylor expansion is performed on f (p):
wherein, J is the Jacobian matrix, Δ p is the moving distance of each iteration p, and when the local objective function is iteratively approximated to the extremum convergence state, the optimal motion variable amount p is calculated as follows:
and when the LM algorithm is adopted to carry out iterative solution on the above formula, stopping iteration when the iteration number is more than the time threshold or the Δ p is less than the step threshold. At this time delta1Equal to the Norm (p1) of the rotational and translational movement p1, as shown in detail in fig. 4. In fig. 4, point a represents an initial state of iteration of the local objective function F, that is, p =0 at this time, point B represents an extreme convergence state of F, and then a distance between point a and point B is the check indicator δ1。
2) SchoolTest index delta2: will check the index delta1Substituting the corresponding p1 into the local target function F (p) of the reference point cloud in the three-dimensional grid, dividing the calculation result of the function by the number of the point clouds to obtain an average probability score, which is represented by a symbol F1.
If the probability distribution model of the grid is consistent with the actual data of the reference point cloud, F1 is the global minimum of the local objective function F (p), and there are no other solutions smaller than F1 in F (p); on the contrary, if δ1The corresponding p1 is only a local minimum of the local objective function f (p), and the point cloud registration fails when the point cloud distribution and the probability model are inconsistent and f (p) takes a smaller value.
Thus, the check index δ2It can be represented by the size of F1, and its detailed calculation method is as follows:
where τ represents an artificially set parameter for shifting the average probability score of each grid to around 0 for comparison, it is noted that when τ is 0, δ is2Directly equal to the average probability score.
According to the steps, the checking index delta of whether each three-dimensional grid in the reference point cloud conforms to the assumption of 'extreme value registration' can be calculated1And delta2. Index delta of the grid1When the data distribution is smaller than the set threshold value, the data distribution of the point cloud in the grid is consistent with the probability distribution model, and when the index delta of the grid is smaller than the set threshold value, the data distribution is consistent with the probability distribution model2When the local objective function of the grid is smaller than the set threshold value, the local objective function of the grid obtains a minimum value when the rotation and translation motion is 0, and when delta is smaller than the set threshold value1And delta2When the point clouds are all smaller than the set threshold value, the point clouds in the grids are in accordance with the NDT 'extreme value registration' hypothesis. Due to delta1And delta2The value ranges of (a) and (b) are different, and sometimes they cannot be directly put together for comparison, so that they can be scaled according to the specific scene requirements. Such as will delta2Multiply by 0.1 and then sum by1Combined to form the final check index [ delta ]1，0.1δ2]. In the embodiment of the invention, for convenience of calculation, the two are directly combined together, namely delta = [ delta ]1，δ2]。
After the check index delta of each three-dimensional grid in the reference point cloud is calculated, whether the current grid needs to be divided into grids with smaller sizes again or not needs to be judged according to the modulus value of the delta, namely, the resolution of the grids is improved. When the modulus value of delta is larger than a set threshold value, the grid is not in accordance with the NDT assumption, and the subsequent point cloud registration is not acted. If the number of invalid stereo grids is large, the subsequent global objective function cannot be converged, and the point cloud registration performance is reduced.
Since the higher the resolution of the three-dimensional grid, the more the probability distribution model can capture the detailed information of the point cloud data, the smaller the δ is, therefore, the grid which does not meet the assumption can be divided into a plurality of smaller grids by common methods such as equal division or clustering, so as to increase the number of effective three-dimensional grids. For example, for the stereo grid V2 in fig. 3 that does not conform to the NDT assumption, if V2 can be divided into two small grids by dividing equally so that there is only one cluster in each small grid, the divided small grids will conform to the NDT assumption more.
The embodiment of the invention judges whether the grid needs to be subdivided or not by comparing the modulus value of delta with the threshold theta, and the grid is subdivided by adopting an equal division method. The specific method comprises the following steps:
firstly, a threshold theta is artificially set according to a scene of point cloud registration, and the magnitude of delta and theta is compared through a Norm function. When Norm (δ) > = θ, it is stated that the stereo grid does not meet the assumption of "extreme value registration", then the midpoint of the grid in length, width and height is taken to perform segmentation, so as to obtain 8 small grids, and the length, width and height of each small grid are half of those of the original grid. It can be understood that the grid may be cut by taking only the height midpoint of the grid, and the original grid is divided into two upper and lower parts. When Norm (δ) < θ, the stereo grid is said to conform to the "extremum registration" assumption, without re-partitioning. For example, when θ =0.2, δ = [0.1,0.25], since the modulus value of δ is greater than 0.2, then the grid does not meet the NDT assumption and needs to be repartitioned.
In order to enable the global objective function to more accurately represent the registration state of the reference point cloud and the target point cloud, the check index delta needs to be converted into the weight of each three-dimensional grid in the global objective function, when the global objective function is solved, the grid with high weight can provide more constraint information, and the grid with low weight can provide relatively less constraint information, so that the calculated point cloud result is more stable and reliable.
Analyzing the conversion relation between the module value Norm (delta) of the check index and the stereo grid weight w to obtain: the smaller the Norm (δ), the more the grid conforms to the NDT assumption, the larger the w need, the negative correlation of Norm (δ) need and w; norm (δ) is minimum 0, w is the maximum value needed; since the grid needs to be subdivided when Norm (δ) > = θ, w needs to take a minimum value at this time.
Therefore, the transfer function g (δ) can be designed as follows: firstly, limiting the value range of the weight w between a minimum value 0 and a maximum value t, wherein t is a positive number greater than 0, when Norm (δ) =0, w obtains the maximum value t, and when Norm (δ) > = θ, w obtains the minimum value 0, that is, the transfer function g (δ) needs to pass through the key points (0, t) and (θ, 0), and then, continuously adding the key point coordinates meeting the negative correlation condition between 0 and θ, such as: (0.1, 0.8), (0.3, 0.4), etc.
In order to calculate the corresponding weight w for any Norm (δ) of the input, curve fitting needs to be performed on the coordinates of the key points to obtain the functional relationship between the Norm (δ) and w. Commonly used fitting functions are: a logarithmic function, an exponential function, a polynomial function, etc. in this embodiment, a polynomial function g (δ) that can obtain an accurate fit is selected, and it should be noted that when Norm (δ) > = θ, w is always 0.
The specific form of the transfer function g (δ) selected in this embodiment is as follows:
the value range of the weight w is between 0 and t, the coefficient a is smaller than 0 to represent that w and Norm (delta) are in negative correlation, the offset b is used for ensuring that g (delta) can pass through a part of key points, and k is the order grade of the transfer function g (delta).
After setting two key points (0, t) and (θ, 0) that the function g needs to pass through, the coefficients a and b can be obtained by means of curve fitting. θ is used to represent the cutoff threshold for the weight w, and at θ =0.1 and t =1, an example of the transfer function is shown in fig. 5, where w =1 when Norm (δ) =0 and w =0 when Norm (δ) = 0.1; curves k =1,2,3,4 represent values of an index k in the function g, respectively, and the smaller k is, the stricter the requirement of g on δ is, and the faster the weight w drops when δ is the same. The curve step shows that the step response is used directly to filter out δ with a modulus value less than 0.1. For example, in fig. 5, straight lines connecting solid circles indicate a transfer function when k =1, a = -10, and b =1, and in this case, w =0.9 when Norm (δ) =0.1, and w =0.4 when Norm (δ) = 0.6.
After the conversion function is obtained, substituting the check indexes delta of all the three-dimensional grids in the reference point cloud into the conversion function to obtain a weight set W of all the three-dimensional grids in the global objective function. For example, W = {1,0.9,0.8, 0.85., 0.5}, and the number of elements in W is equal to the number of reference point cloud stereo grids.
After a target point cloud which needs to be registered with a reference point cloud at a certain moment is obtained, firstly, a global target function of point cloud registration needs to be constructed, and the specific method is as follows: taking the translation and rotation motion of the target point cloud as variables to be optimized of a global target function, expressing the variables with a symbol T, traversing each data point in the target point cloud, carrying out three-dimensional transformation on each data point according to T, calculating a three-dimensional grid to which each data point belongs according to the transformed three-dimensional coordinates, and calculating the v of the three-dimensional gridμAnd vΣAnd obtaining an NDT model of the grid, and obtaining the weight w of the grid according to a transfer function. And substituting each data point into the respective NDT model and multiplying by the weight w to obtain the probability score of each data point. And accumulating and summing the probability scores of all the data points in the target point cloud to obtain a global target function of point cloud registration. The detailed calculation method is as follows:
where T denotes the rotational and translational movement to be optimized, xkRepresenting the three-dimensional coordinates of a point within the target point cloud, m representing the number of target point clouds, Tr representing the pair xkApplying a motion transformation denoted by T, N being based on vμAnd vΣConstructed NDT model, wkDenotes xkAnd H (T) represents the output of the global objective function, and the smaller the output is, the higher the registration degree of the target point cloud and the reference point cloud is, the higher the probability of the occurrence of the corresponding rotation and translation motion T is. w is akThe function H (T) can accurately represent the registration state between the reference point cloud and the target point cloud, when the function H (T) is subjected to iterative solution, grids with high weights are more in line with NDT assumption, more constraint information can be provided, and conversely, grids with low weights are relatively less constraint information can be provided.
And finally, iteratively solving the minimum value of the global objective function by methods of Gauss Newton, gradient descent, LM and the like to obtain an optimal point cloud registration result. The minimum solving process in this embodiment is as follows: substituting the target point cloud, the initial rotation and translation motion T of the target point cloud in the global target function and the weight of the target point cloud into the global target function, and then calculating the optimal motion variable quantity Δ T of the target point cloud and the reference point cloud by using an extremum solving method. When the mode value of the max T is larger than the threshold value of the extremum solving method, it is described that the global function is not converged, the value of the initial motion T needs to be updated to T +. T, the target point cloud is subjected to three-dimensional transformation according to T +. T, the weight of each data point in the target point cloud is recalculated and substituted into the global target function, and the optimal motion variable quantity Δ T is calculated by the extremum solving method again. When the mode value of the margin T is smaller than the threshold value of the extremum solving method, the global objective function is converged, and the point cloud registration result at this time is T +. DELTA.T.
The following is one point cloud registration embodiment enumerated for the present invention.
Firstly, point cloud data is obtained: a laser 16-line laser radar is arranged right above the robot, the robot is controlled to drive along a road at the speed of 1m/s to the right front, and after a series of operations of straight running, left turning, straight running, turning around, straight running, right turning, straight running and the like, the robot finally returns to the starting point and records point cloud data in driving.
Then testing the expression effect of the original NDT point cloud registration method: setting the grid size S =1m, the maximum iteration number C =20, and the pos iteration step =0.1, and sequentially registering the input point cloud data to obtain the robot trajectory as shown in fig. 6. Wherein the dotted line points represent the real trajectory, and the solid line represents the trajectory calculated by the NDT point cloud registration method. Comparing the solid line and the dotted line points in fig. 6, the point cloud registered track and the real track have obvious deviation in enlarged areas L1 and L2, and there are about 15 deviation points in L1 and about 21 deviation points in L2, and there are 36 deviation points.
Finally, on the premise that the experimental parameters and the original point cloud data are fixed and unchanged, according to the steps in the embodiment of the invention, the assumed check index δ is calculated, and is converted into the weight in global registration by using the conversion function, and after the input point cloud data is sequentially registered, the robot track is obtained as shown in fig. 7 and 8. In fig. 7, k =2 in the conversion function g, and comparing the solid line and the dotted line points in fig. 7, the point cloud registration is still biased at L2, but remains substantially stable at L1, and the biased points at L2 are about 17. In fig. 8, k =4 in the conversion function g, and comparing the solid line and the dotted line points in fig. 8, the point cloud registration has a deviation at L1, but a steady state is completely maintained at L2, and the deviation point at L1 is only 4. Through the embodiment, compared with an original NDT point cloud registration method, the point cloud registration stability and accuracy are obviously improved after the calculation of the hypothesis verification index delta and the weight conversion.
The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (5)
1. A point cloud registration optimization method based on normal distribution transformation hypothesis verification is characterized by comprising the following steps:
(1) acquiring input reference point cloud, and performing three-dimensional grid division on the reference point cloud according to the resolution of a fixed size to obtain a set V, V = { V } of three-dimensional grids1,v2,...,vnN is the number of the stereoscopic grids;
(2) for each stereo grid in the set V, calculating V when the stereo grid conforms to normal distribution transformationμAnd vΣWherein v isμMean value of coordinates, v, representing the point cloud in a three-dimensional gridΣA covariance matrix representing coordinates of the point cloud within the grid;
(3) based on vμAnd vΣCalculating a check index delta for quantifying the coincidence of the three-dimensional grid with the NDT hypothesis degree, and adjusting the resolution of the three-dimensional grid according to the magnitude of a modulus of the delta;
(4) designing a conversion function, and converting the hypothetical check index delta into the weight of the three-dimensional grid in the point cloud registration global objective function through the conversion function;
(5) constructing a global target function of registration of the weighted target point cloud and the reference point cloud based on an NDT model of the reference point cloud three-dimensional grid;
(6) substituting the target point cloud into a global target function for registering the target point cloud and the reference point cloud, and iteratively solving the limit of the function to obtain a point cloud registration result;
the check index delta is calculated by adopting the following method:
based on vμAnd vΣConstructing a local objective function of the three-dimensional grid, and calculating an index delta from an angle of an extreme convergence point of the local objective function1Calculating the index delta from the angle of' average probability score2Will delta1And delta2Obtaining a final check index delta after combination;
delta. the1The method comprises the following steps: first according to vμAnd vΣEstablishing an NDT model of the point cloud in the three-dimensional grid, and substituting all point cloud data in the grid into the point cloudObtaining a local target function from the NDT model, solving the limit of the local target function, recording the rotation and translation motion corresponding to the minimum value as p1, and taking the module value of p1 as the verification index delta1；
Delta. the2The method comprises the following steps: substituting p1 into the local objective function of the three-dimensional grid, calculating the total probability score of the point clouds in the grid, and dividing the total score by the number of the point cloud data to obtain the verification index delta2。
2. The point cloud registration optimization method based on the verification of the assumption of normal distribution transformation as claimed in claim 1, wherein the stereo grid resolution is adjusted according to the following method: firstly, setting a threshold theta according to the working environment of the robot, then comparing the sizes of check indexes delta and theta of the three-dimensional grid through a module function Norm, dividing the original three-dimensional grid into a plurality of small grids through an equal division or clustering method when Norm (delta) > = theta, and recalculating delta of the divided new grid.
3. The point cloud registration optimization method based on the verification of the assumption of normal distribution transformation as claimed in claim 1, wherein in step (4), the output of the conversion function is the weight w of the stereo grid, and the input is the Norm value Norm (δ) of δ, and the conversion function is designed as follows:
1) limiting the value range of the weight w between a minimum value 0 and a maximum value t, wherein t is larger than 0, the conversion function passes through the coordinates (0, t) of the key points and (theta, 0), when Norm (delta) =0, w obtains the maximum value t, and when Norm (delta) > = theta, w obtains the minimum value 0;
2) continuously adding key point coordinates (Norm (delta), w) meeting a negative correlation condition between 0 and theta, wherein the smaller the Norm (delta), the larger the w;
3) and performing curve fitting on the coordinates of the key points to obtain a conversion function.
4. The point cloud registration optimization method based on the verification of the assumption of normal distribution transformation as claimed in claim 1, wherein in the step (5), the weighted global objective function is constructed as follows:
1) taking the translation and rotation motion of the target point cloud as variables to be optimized of a global target function, expressing the variables by using a symbol T, traversing each data point in the target point cloud, carrying out three-dimensional transformation on each data point according to T, and calculating a three-dimensional grid to which each data point belongs according to the transformed three-dimensional coordinates;
2) v according to the three-dimensional gridμAnd vΣObtaining an NDT model of the grid, and obtaining the weight w of the grid according to a transfer function;
3) and substituting each data point into the respective NDT model and multiplying by the weight w to obtain the probability score of each data point, and accumulating and summing the probability scores of all the data points in the target point cloud to obtain the global objective function.
5. The point cloud registration optimization method based on the verification of the assumption of normal distribution transformation as claimed in claim 1, wherein in step (6), the global objective function can be solved iteratively according to the following method:
1) substituting the target point cloud, the initial rotation and translation motion T of the target point cloud in the global target function and the weight of the target point cloud into the global target function;
2) calculating the optimal motion change amount (Δ T) of the target point cloud and the reference point cloud by using an extremum solving method, updating the value of the initial motion T to T + Δ T when the mode value of the Δ T is larger than the threshold value of the extremum solving method, performing three-dimensional transformation on the target point cloud according to T + Δ T, and recalculating the weight of each data point;
3) substituting the data into the global objective function again, and calculating the optimal motion variable T again; when the mode value of the T is smaller than the threshold value, the global objective function is converged, and the point cloud registration result is T +. DELTA.T.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111487497.0A CN113920180B (en) | 2021-12-08 | 2021-12-08 | Point cloud registration optimization method based on normal distribution transformation hypothesis verification |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111487497.0A CN113920180B (en) | 2021-12-08 | 2021-12-08 | Point cloud registration optimization method based on normal distribution transformation hypothesis verification |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113920180A CN113920180A (en) | 2022-01-11 |
| CN113920180B true CN113920180B (en) | 2022-04-01 |
Family
ID=79248772
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111487497.0A Active CN113920180B (en) | 2021-12-08 | 2021-12-08 | Point cloud registration optimization method based on normal distribution transformation hypothesis verification |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113920180B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111079801A (en) * | 2019-11-29 | 2020-04-28 | 上海有个机器人有限公司 | Method, medium, terminal and device for quickly searching closest point based on point cloud matching |
| US11037346B1 (en) * | 2020-04-29 | 2021-06-15 | Nanjing University Of Aeronautics And Astronautics | Multi-station scanning global point cloud registration method based on graph optimization |
| CN113256693A (en) * | 2021-06-04 | 2021-08-13 | 武汉工控仪器仪表有限公司 | Multi-view registration method based on K-means and normal distribution transformation |
| CN113640822A (en) * | 2021-07-07 | 2021-11-12 | 华南理工大学 | High-precision map construction method based on non-map element filtering |
-
2021
- 2021-12-08 CN CN202111487497.0A patent/CN113920180B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111079801A (en) * | 2019-11-29 | 2020-04-28 | 上海有个机器人有限公司 | Method, medium, terminal and device for quickly searching closest point based on point cloud matching |
| US11037346B1 (en) * | 2020-04-29 | 2021-06-15 | Nanjing University Of Aeronautics And Astronautics | Multi-station scanning global point cloud registration method based on graph optimization |
| CN113256693A (en) * | 2021-06-04 | 2021-08-13 | 武汉工控仪器仪表有限公司 | Multi-view registration method based on K-means and normal distribution transformation |
| CN113640822A (en) * | 2021-07-07 | 2021-11-12 | 华南理工大学 | High-precision map construction method based on non-map element filtering |
Non-Patent Citations (1)
| Title |
|---|
| 基于激光点云NDT特征的两步回环检测;柴梦娜等;《激光与红外》;20200120(第01期);19-26 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113920180A (en) | 2022-01-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109410321B (en) | Three-dimensional reconstruction method based on convolutional neural network | |
| CN108871353B (en) | Road network map generation method, system, equipment and storage medium | |
| CN110930495A (en) | Multi-unmanned aerial vehicle cooperation-based ICP point cloud map fusion method, system, device and storage medium | |
| Hoppe et al. | Photogrammetric camera network design for micro aerial vehicles | |
| Mu et al. | Research on SLAM algorithm of mobile robot based on the fusion of 2D LiDAR and depth camera | |
| CN116449384A (en) | Radar inertial tight coupling positioning mapping method based on solid-state laser radar | |
| CN112197773A (en) | Visual and laser positioning mapping method based on plane information | |
| CN115205354A (en) | Phased array laser radar imaging method based on RANSAC and ICP point cloud registration | |
| CN114004894A (en) | Method for determining space relation between laser radar and binocular camera based on three calibration plates | |
| CN109255815A (en) | A kind of object detection and recognition methods based on order spherical harmonic | |
| CN114066773B (en) | Dynamic object removal based on point cloud characteristics and Monte Carlo expansion method | |
| Wei et al. | Optimization of 3-D pose measurement method based on binocular vision | |
| Jiang et al. | Learned local features for structure from motion of uav images: A comparative evaluation | |
| CN117541614B (en) | Space non-cooperative target close-range relative pose tracking method based on improved ICP algorithm | |
| CN113313200B (en) | Point cloud precision matching method based on normal constraint | |
| CN114581519A (en) | Laser autonomous positioning mapping method for quadruped robot in cross-country environment | |
| CN110851978A (en) | Camera position optimization method based on visibility | |
| CN112446952B (en) | Three-dimensional point cloud normal vector generation method and device, electronic equipment and storage medium | |
| CN113808152A (en) | Unmanned aerial vehicle autonomous navigation method based on ORB _ SLAM2 | |
| CN111815684B (en) | Space multivariate feature registration optimization method and device based on unified residual error model | |
| CN117629215A (en) | Chassis charging pile-returning method based on single-line laser radar point cloud registration | |
| CN113920180B (en) | Point cloud registration optimization method based on normal distribution transformation hypothesis verification | |
| CN116704024A (en) | Vehicle-mounted laser point cloud pose graph optimization method and system combining multiclass constraints | |
| CN108595373A (en) | It is a kind of without control DEM method for registering | |
| CN114842074A (en) | Unmanned aerial vehicle image positioning method based on model matching |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |