CN114494368A - Low-overlapping-rate point cloud registration method combining dimensionality reduction projection and feature matching - Google Patents
Low-overlapping-rate point cloud registration method combining dimensionality reduction projection and feature matching Download PDFInfo
- Publication number
- CN114494368A CN114494368A CN202111558350.6A CN202111558350A CN114494368A CN 114494368 A CN114494368 A CN 114494368A CN 202111558350 A CN202111558350 A CN 202111558350A CN 114494368 A CN114494368 A CN 114494368A
- Authority
- CN
- China
- Prior art keywords
- point cloud
- point
- dimensionality reduction
- matching
- low
- 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.)
- Pending
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Collating Specific Patterns (AREA)
Abstract
A low-overlapping-rate point cloud registration method combining dimensionality reduction projection and feature matching comprises the following steps: 1) taking the mean normal vector direction of the point cloud as an initial dimensionality reduction direction; 2) performing direction correction on the initial dimensionality reduction direction to obtain a final dimensionality reduction direction; 3) projecting along the final dimension reduction direction to generate a depth image after dimension reduction; 4) generating a multi-resolution image according to the size of the grid; 5) carrying out two-step screening according to the voting mechanism and the distance invariance of rigid body transformation to obtain reliable matching point pairs; 6) carrying out two-step screening by using a voting mechanism and the distance invariance of rigid body transformation; 7) and solving the rotation matrix for the final target point pair, and performing iterative optimization by using an ICP (inductively coupled plasma) algorithm to obtain a final accurate conversion matrix. The method solves the problems that the low-overlapping-rate point cloud feature matching is easy to generate errors, the iterative search time is long and the like, and improves the registration precision and efficiency.
Description
Technical Field
The invention relates to the technical field of point cloud registration application in three-dimensional vision, in particular to a low-overlapping-rate point cloud registration method combining dimension reduction projection and feature matching.
Background
The purpose of point cloud registration is to solve a transformation matrix of point clouds with different postures under the same coordinate, unify the multi-view scanning point clouds under the same coordinate system by using the transformation matrix, and finally acquire a complete three-dimensional mode and scene. In recent years, point cloud registration is widely researched, but the performance of a traditional point cloud registration algorithm is limited by a good initial position and a large overlapping rate. The overlapping rate is defined as the ratio of the number of points in the overlapping area to the number of points in the complete point cloud, and when the overlapping rate of the source point cloud and the target point cloud is lower than 60%, the overlapping degree of the two point clouds is low. Therefore, a fast and accurate registration algorithm for low overlap ratio point clouds remains a point cloud registration difficulty.
At present, point cloud registration algorithms are numerous and can be roughly divided into two types, namely registration algorithms based on global search and local features.
The registration algorithm based on global search is to search corresponding points from the point cloud to be matched according to the constraint of geometric position relation, calculate a transformation matrix, and perform multiple loop iterations to obtain an optimal transformation matrix. The registration algorithm based on global search is classically an Iterative Closest Point (ICP) algorithm, but the algorithm has a high requirement on the initial position of the Point cloud. Various improved forms such as PLICP, GICP and the like are derived according to the ICP algorithm, the improved algorithms are improved to a certain extent in the aspects of precision, speed, noise resistance and the like of point cloud registration, and low-overlap-rate point cloud registration cannot be well completed; the registration method based on probability statistics is also based on global unit, and utilizes probability density function to estimate point cloud distribution, such as the Normal Distribution Transform (NDT) algorithm proposed by Biber et al; there are other registration algorithms based on global search, such as Random sample consensus (RANSAC) based algorithm, 4-point consistency sets (4 PCS) based algorithm proposed by Mellado et al, and modified D4PCS algorithm, Super4PCS algorithm, etc. However, the registration algorithm based on global search generally has the characteristic of high computational complexity and consumes long time; meanwhile, aiming at the point cloud registration with low overlapping rate, the search of non-overlapping areas does not work on the registration, but increases the calculated amount and reduces the overall registration efficiency.
The registration algorithm based on local features mainly works by defining feature descriptors and establishing the corresponding relation of features among point clouds. Rusu et al propose a Point Feature Histogram (PFH) for counting normal vector included angles of two adjacent points as a descriptor and an improved form thereof, namely a Fast Point Feature Histogram (FPFH); converting point cloud into Spinning Images (SI) characteristic quantity in a grid form, which is proposed by Daniel F.Huber et al; tombari et al combine the ideas of point signatures and point feature histograms to propose a SHOT feature descriptor; andrea et al propose a three-dimensional shape context (3DSC) that describes shape profile features using a histogram in a logarithmic polar coordinate system, which is an extension of the two-dimensional shape context; li and the like cluster the target three-dimensional point cloud to obtain small-scale clustered point cloud with obvious characteristics, so that the registration efficiency and precision are effectively improved; besides, there are also common local feature descriptors such as Scale Invariant Feature Transform (SIFT), Rotation Invariant Feature Transform (RIFT), and the like. The method has the advantages that the method is high in complexity of the point cloud local feature extraction algorithm, and is easy to cause problems such as mismatching, so that the method cannot be directly used for registering the point cloud with a low overlapping rate; meanwhile, local features are easy to be disturbed by disorder and shielding during extraction, namely when the three-dimensional point cloud has disorder and shielding, the registration effect is greatly reduced.
In order to solve the problem, many scholars propose a two-step registration strategy, namely extracting a target point and an overlapping region first, and then iteratively searching in the range of the overlapping region to obtain an optimal solution. Wang et al propose an overlap region extraction method based on region segmentation, which can still complete the extraction of the overlap region under the condition of large difference of point cloud acquisition visual angles, thereby improving the registration efficiency; li and the like provide an improved TrICP algorithm based on contribution factors, the algorithm automatically calculates the overlapping degree, and can reliably, efficiently and automatically register laser and image reconstruction point clouds containing a large amount of noise, partial overlapping and non-homologous; lu and the like divide point cloud into a plurality of regions, firstly perform D4PCS registration on each region, and then enlarge the action of the overlapped region to complete global registration, but the algorithm consumes long time and is not suitable for large-scale point cloud data; combining region blocking and convex optimization by using Zhang Yuan and the like, respectively extracting an overlapping region and optimizing a corresponding relation, and finally finishing registration by using an ICP (inductively coupled plasma) algorithm; although the accuracy of the point cloud registration with the low overlapping rate is improved to a certain extent by the algorithm, the feature point calculation and extraction are carried out in the three-dimensional space, and the algorithm has the problems of high calculation complexity, long time consumption and the like.
Disclosure of Invention
In order to solve the defects in the prior art, the invention aims to provide a low-overlap-rate point cloud registration method combining dimensionality reduction projection and feature matching, which aims at the problem of low-overlap-rate point cloud registration, completes extraction, description and matching of feature points in an image space after dimensionality reduction by combining dimensionality reduction projection and image feature matching, and solves the problems of low-overlap-rate point cloud feature point extraction difficulty, easy occurrence of mismatching, long iteration time consumption and the like.
In order to achieve the above object, the low overlap ratio point cloud registration method combining dimensionality reduction projection and feature matching provided by the invention comprises the following steps:
1) taking the mean normal vector direction of the point cloud as an initial dimensionality reduction direction;
2) performing direction correction on the initial dimensionality reduction direction to obtain a final dimensionality reduction direction;
3) projecting along the final dimension reduction direction to generate a depth image after dimension reduction;
4) generating a multi-resolution image according to the size of the grid;
5) carrying out two-step screening according to the voting mechanism and the distance invariance of rigid body transformation to obtain reliable matching point pairs;
6) carrying out two-step screening by using a voting mechanism and the distance invariance of rigid body transformation;
7) and solving the rotation matrix for the final target point pair, and performing iterative optimization by using an ICP (inductively coupled plasma) algorithm to obtain a final accurate conversion matrix.
Further, the step 1) further comprises the steps of,
solving the normal vector, p, of discrete points by principal component analysisiAnd (3) normally measuring the eigenvector corresponding to the minimum eigenvalue of the matrix:
Eξi＝λiξi，i＝{1，2，3}，
where R is the neighborhood radius, N is the number of points in the neighborhood, piThree-dimensional coordinate vector representing ith point in the domain, p' represents three-dimensional coordinate vector of centroid of the domain, diEuclidean distance from ith point to centroid in representation field, xiiIs the characteristic vector of the matrix E, i is an integer greater than or equal to 1;
piand (2) measuring the eigenvector corresponding to the minimum eigenvalue of the matrix by the point normal method, and recording the eigenvector as n ═ xi，yi，zi) Then, determining the initial projection direction of the dimensionality reduction as the mean normal vector of each point of the point cloud P:
further, the step 2) further comprises the steps of,
uniformly dividing spatial neighborhoods along the initial projection direction, and respectively projecting along a plurality of neighborhood directions to obtain different dimension reduction images;
calculating a hash fingerprint image of the dimension-reduced image by using a hash algorithm, comparing the hamming distances of the dimension-reduced image, regarding the image with the minimum distance value as the image with the highest similarity, and regarding the corresponding direction as the corrected direction;
carrying out rotation transformation on the original point cloud to align the Z-axis direction of the rotated point cloud space coordinate system with the projection direction:
wherein the content of the first and second substances,
further, the step 3) further comprises,
calculate the maximum depth value max within each grid cD(c) Then, define point set P' (c) as the outermost set of points in the grid:
P′(c)＝{p∈P||maxD(c)-pz|＜δ1}；
the gaussian weight interpolation is performed on the Z coordinate of the midpoint in P' (c), and the pixel value f (c) corresponding to the grid is calculated:
wherein the normalization parameter W is expressed as:
the gaussian function g is expressed as:
and (4) carrying out interpolation calculation on the pixel value corresponding to each grid, thereby obtaining the depth image after dimension reduction.
Further, the step 4) further comprises,
establishing a relation between the grid size D and the mean distance D of the point cloud discrete points:
d＝l·D，
wherein l is a coefficient;
the resolution is expressed as:
further, the step 5) further comprises,
carrying out ORB matching on n different image layers by adopting an ORB operator, wherein the corresponding image characteristic point pair of each image layer is expressed as:
matches[i]，i∈[0，n-1]，
wherein, matches [0]]Corresponds to l0The layer with the highest lower resolution;
counting the repetition times of the matching relation in other layers in matches [0], and adding one to the ticket number when the following formula is satisfied:
l0×(matches[0][j].X，matches[0][j].Y)
＝li×(matches[i][k].X，matches[i][k].Y)
wherein, matches [0] [ j ] represents the jth pair of matching points in the 1 st layer, and matches [ i ] [ k ] represents the kth pair of matching points in the (i + 1) th layer.
Further, the step 6) further comprises,
determining a threshold delta for the number of votes based on the number of layers2Is greater than delta2The matching points are regarded as reliable matching points and are mapped to the original point cloud to obtain Matches;
any two point pairs (p) in Matchesi，qi) And (p)j，qj) If the matching point pair is correct, the distance invariance according to rigid body transformation exists a relation dist (p)i，pj)＝dist(qi，qj) (ii) a Optionally selecting two pairs of points, selecting delta3> 0, such that:
the numerator represents the distance difference of the matching points, the denominator is used for reducing the sensitivity of the distance constraint condition to the point cloud scale, and the final matching point pair meeting the distance constraint is determined when the above formula is met.
Further, the step 7) further comprises,
obtaining an initial conversion matrix of the final matching point pair by using a quaternion method, and converting the cloud of the point to be registered to a similar position;
and (4) applying an ICP (inductively coupled plasma) algorithm to carry out iterative computation to obtain an accurate conversion matrix.
To achieve the above object, the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program running on the processor, and the processor executes the computer program to perform the steps of the low-overlap-rate point cloud registration method combining dimensionality reduction projection and feature matching as described above.
To achieve the above object, the present invention further provides a computer-readable storage medium having stored thereon a computer program which, when running, performs the steps of the low-overlap-rate point cloud registration method combining dimensionality reduction projection and feature matching as described above.
Compared with the prior art, the low-overlap-rate point cloud registration method combining the dimensionality reduction projection and the feature matching has the following beneficial effects: aiming at the problem of low-overlap-rate point cloud registration, the extraction, description and matching of feature points are completed in an image space after dimensionality reduction by combining dimensionality reduction projection and image feature matching, so that the problems of low-overlap-rate point cloud feature point extraction difficulty, easiness in occurrence of mismatching, long iteration time consumption and the like are solved, and the registration precision and efficiency and the adaptability to low overlap rate are improved; the velocity advantage is more pronounced with larger point cloud sizes.
Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.
Drawings
The accompanying drawings are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:
FIG. 1 is a flow chart of a low overlap ratio point cloud registration method combining reduced dimension projection and feature matching according to the present invention;
FIG. 2 is a schematic diagram of a point cloud dimensionality reduction according to the present invention;
FIG. 3 is a schematic view of the orientation correction according to the present invention;
FIG. 4 is a diagram illustrating the image effect after dimension reduction of different coefficients according to the present invention;
FIG. 5 is a feature matching and location resolution flow diagram according to the present invention;
FIG. 6 is an experimental plot of directional correction according to the present invention;
FIG. 7 is a hash fingerprint map of a dimension reduced image according to the present invention;
FIG. 8 is a graph of feature matching and screening results according to the present invention;
FIG. 9 is a graph illustrating the comparison of registration results obtained according to the present invention with other registration methods;
FIG. 10 is a schematic of an error curve and a velocity curve according to the present invention.
Detailed Description
The preferred embodiments of the present invention will be described in conjunction with the accompanying drawings, and it will be understood that they are described herein for the purpose of illustration and explanation and not limitation.
Example 1
Fig. 1 is a flowchart of a low-overlap-rate point cloud registration method combining dimension reduction projection and feature matching according to the present invention, and the low-overlap-rate point cloud registration method combining dimension reduction projection and feature matching according to the present invention will be described in detail with reference to fig. 1.
First, in step 101, normal vector information of the point cloud is obtained according to a principal component analysis method, and a mean normal vector is used as an initial projection direction for dimension reduction.
In the embodiment of the invention, the normal vector p of the discrete point is solved by a principal component analysis methodiAnd (3) normally measuring the eigenvector corresponding to the minimum eigenvalue of the matrix:
Eξi＝λiξi，i＝{1，2，3}，
wherein R is the radius of the neighborhood, N is the number of points in the neighborhood, piThree-dimensional coordinate vector representing ith point in the domain, p' represents three-dimensional coordinate vector of centroid of the domain, diEuclidean distance from ith point to centroid in representation field, xiiIs the eigenvector of the matrix E, i is an integer greater than or equal to 1；
piAnd (2) measuring the eigenvector corresponding to the minimum eigenvalue of the matrix by the point normal method, and recording the eigenvector as n ═ xi，yi，zi) Then, determining the initial projection direction of the dimensionality reduction as the mean normal vector of each point of the point cloud P:
in step 102, direction correction is performed by performing a neighborhood partition of the initial projection direction and a hash algorithm.
In the embodiment of the present invention, in order to reduce the dimension of the point cloud overlapping region along the approximate direction to increase the image feature matching points, the initial projection direction obtained in step 101 is corrected:
1) after the initial projection direction is determined, uniformly dividing the spatial neighborhood along the direction, and respectively projecting along a plurality of neighborhood directions to obtain different dimension reduction images;
2) calculating a hash fingerprint image of the image by using a hash algorithm, comparing the hamming distances, and regarding the image with the minimum distance value as the image with the highest similarity, and taking the direction corresponding to the image as the corrected direction;
3) carrying out rotation transformation on the original point cloud to align the Z-axis direction of the rotated point cloud space coordinate system with the projection direction:
wherein the content of the first and second substances,
in step 103, discrete points in the projection grid are interpolated by a gaussian function to obtain pixel values corresponding to the grid.
In the embodiment of the present invention, since the pixel grid may include a plurality of discrete points, and the depth of one of the discrete points is taken as information that the pixel value cannot accurately represent all the points in the area, it is considered that the pixel value is obtained by using a gaussian weight interpolation method.
1) First the maximum depth value max within each grid c is calculatedD(c) Then, define point set P' (c) as the outermost set of points in the grid:
P′(c)＝{p∈P||maxD(c)-pz|＜δ1}
2) gaussian weight interpolation is performed on the Z coordinate of the midpoint in P' (c), and the pixel value f (c) corresponding to the grid is calculated:
wherein the normalization parameter W is expressed as:
the gaussian function g is expressed as:
and (4) carrying out interpolation calculation on the pixel value corresponding to each grid, thereby obtaining the depth image after dimension reduction.
FIG. 2 is a schematic diagram of a point cloud dimension reduction according to the present invention, as shown in FIG. 2, firstly, a projection direction is determined by normal vector information and direction correction of a point cloud, rotation and translation transformation are performed on original point cloud data, and a mean normal vector is aligned with a z-axis direction; then, projecting the point cloud data to the xoy plane, determining the resolution according to the mean distance of the point cloud, and rasterizing; and finally, interpolating the Z coordinate in each grid to obtain a pixel value, and converting the three-dimensional point cloud into a two-dimensional depth map.
Fig. 3 is a schematic diagram of direction correction according to the present invention, as shown in fig. 3, for partially overlapped point clouds, correct matching relations are mainly concentrated in an overlapping region, and for an image space after dimension reduction of such point clouds, these correct matching relations are obtained on the premise that the overlapping regions are projected along approximately the same direction, which is obtained by direction correction in the dimension reduction process of the present invention.
The neighborhood space of the initial direction of the target point cloud is uniformly divided into k parts, and the direction with an included angle theta (30 degrees or 45 degrees can be taken) with the initial direction is taken as a candidate.
As shown in FIG. 3, n1、n2、...、nkTo be uniformly distributed in
In the neighborhood and
normal vector with angle theta, in nkFor example, assuming that its component along the Z axis is Δ Z, then
Then n iskAnd
can solve the conversion relation of
Fig. 4 is a schematic diagram of an image effect obtained by dimension reduction of different coefficients l according to the present invention, as shown in fig. 4, where (a): 1; fig. 4 (b): l is 1.5; fig. 4 (c): l is 2; in fig. 4, (d) l ═ 3; as can be seen from fig. 4, when l ≦ 1.2, some points (outliers) with a pixel value of 0 are unevenly distributed in the image, which is caused by the fact that the grid is small and no projection point falls in the grid, in which case more pseudo feature points are added in the feature extraction; when l ≧ 3, the image has more obvious jaggy at the edge, and fine features are lost, because the grid is larger, and more points are contained in the grid, so that details are lost through interpolation, and extractable feature points are greatly reduced. When l is more than 1.2 and less than 3, the image resolution is more appropriate, and the dimension reduction process can reduce the loss of point information and accurately express point cloud information. In conclusion, multiple groups of l are selected in the interval (1.2,3), and a multi-resolution dimension-reduced image is generated for further feature matching.
In step 104, grids of different sizes are selected according to the mean distance of the point cloud to generate dimension reduction images of different resolutions.
In the embodiment of the invention, point clouds obtained by different devices have different point set scales and different resolutions, and in order to enable the registration method to have self-adaptability to the point cloud scales, the relation between the grid size D and the mean distance D of point cloud discrete points is established in the dimension reduction process:
d＝l·D
the grid size d determines the number of discrete points in the grid, i.e. the resolution of the image after dimensionality reduction, which can be expressed as:
in step 105, the layers are divided according to the resolution, ORB feature matching is respectively carried out between different layers of the layers, and the number of the tickets repeatedly appearing in the matching relationship is counted.
In the embodiment of the invention, after the multi-resolution image is generated in the step 104, in the image feature matching, an ORB operator is adopted in the invention to carry out ORB matching on n different image layers, and the image feature point pair corresponding to each image layer is expressed as matches [ i []，i∈[0，n-1]Wherein matches [0]]Corresponds to l0And (5) the layer with the highest resolution. Statistics matches [0]]The number of times of repetition of the medium matching relation in other layers is increased by one when the following formula is satisfied:
l0×(matches[0][j].X，matches[0][j].Y)
＝li×(matches[i][k].X，matches[i][k].Y)
wherein, matches [0] [ j ] represents the jth pair of matching points in the 1 st layer, and matches [ i ] [ k ] represents the kth pair of matching points in the (i + 1) th layer.
In step 106, two times of screening are performed according to the voting mechanism and the distance invariance of rigid body transformation, so as to obtain a reliable matching relationship.
In the embodiment of the invention, the threshold value delta of the ticket number is determined according to the number of the layers2Is greater than delta2And (4) taking the matching points as reliable matching points, and mapping to the original point cloud to obtain Matches. Any two point pairs (p) in Matchesi，qi) And (p)j，qj) If the matching point pair is correct, then there is a relation dist (p) according to the distance invariance of rigid body transformationi，pj)＝dist(qi，qj) Thus, two pairs of points are selected and the appropriate delta is selected3> 0, such that:
the numerator represents the distance difference of the matching points, the denominator is used for reducing the sensitivity of the distance constraint condition to the point cloud scale, and the final matching point pair meeting the distance constraint is determined when the above formula is met.
In step 107, a quaternion method is applied to the final matching points to find a position conversion matrix.
In the embodiment of the present invention, the final matching point pair is obtained in step 106, and the initial transformation matrix R, T is obtained through the quaternion method, so that the cloud of the point to be registered is transformed to the similar position on this basis.
At step 108, an ICP algorithm is used for iterative calculations to find an accurate solution for registration.
In the embodiment of the invention, after the initial registration by the quaternion method, the point cloud to be registered is converted to the similar position, and the ICP algorithm is applied to the point cloud to be registered, so that the accurate conversion matrix can be obtained by iterative computation with less times.
Fig. 5 is a feature matching and position calculating flow chart according to the present invention, which will be further described with reference to fig. 5.
In step 501, feature matching is performed on the two-dimensional reduced-dimension image: regarding the point cloud to be registered as a layer in the same coefficient descending dimension image, and respectively performing image feature matching under different layers;
in step 502, feature point screening is performed to obtain three-dimensional matching point pairs: counting the repeated occurrence times of the same matching point pair under different image layers, and eliminating the matching point pair with the ticket number smaller than a threshold value as unreliable matching; mapping the reserved matching point pairs back to point clouds to obtain three-dimensional matching point pairs, and screening according to the distance invariance of rigid body transformation to obtain final target point pairs;
at step 503, the transformation matrix is computed: resolving the position relation between target point pairs by using a quaternion method so as to complete the initial registration of the point cloud; and finally, refining and registering by adopting ICP (inductively coupled plasma), and obtaining an accurate result.
FIG. 6 is a diagram of an experiment for correcting direction according to the present invention, as shown in FIG. 6, in which (a) in FIG. 6 is a candidate direction; fig. 6 (b) shows a direction dimension reduction image; fig. 6 (c) shows a direction dimension reduction image; in fig. 6, (d) is a direction dimension reduction image, the difference between the initial direction angles of the cloud to be registered is large, and the images obtained by dimension reduction along the initial direction are not suitable for feature matching and are prone to have more wrong matching relations, so that direction correction is needed.
FIG. 7 is a hash fingerprint of a dimension-reduced image according to the present invention, as shown in FIG. 7, solved by pHash algorithm to obtain q0To qkSimilarity to p. In FIG. 7, (a) to (d) correspond to p and q0、q1、q2The hamming distance is calculated to obtain the highest similarity between (a) and (d) in FIG. 7, so that the q corresponding to (d) is taken2The image is used as a dimension reduction image after Q direction correction.
Fig. 8 is a diagram of a feature matching and screening result according to the present invention, as shown in fig. 8, after determining a dimension reduction direction, changing a value to generate a multi-resolution image, performing feature point matching on each layer by using an ORB algorithm, and screening out a feature point pair with a lower ticket number to obtain an image matching result; and then mapping the matching characteristic points to the point cloud, and carrying out secondary screening according to the distance invariance of rigid body transformation to obtain a matching result as shown in the following graph.
Table 1 gives that the registration method steps are time consuming. The registration algorithm is high in time consumption proportion of feature matching and direction correction, and multiple dimensionality reduction operations of direction correction including point cloud account for independent operations due to the fact that feature matching is conducted through multiple matching of different image layers.
TABLE 1 registration method steps are time consuming
| Step (ii) of | Time consuming (ms) | Percent (%) |
| Dimensionality reduction transformation | 379 | 14.87 |
| Direction correction | 726 | 28.49 |
| Feature matching | 941 | 36.93 |
| Quaternion solution | 95 | 3.72 |
| ICP iteration | 407 | 15.97 |
| Total of | 2548 | 100 |
Fig. 9 is a schematic diagram comparing the registration results obtained by the present invention with other registration methods, as shown in fig. 9, and it can be seen that applying the ICP algorithm directly to the low overlap ratio point cloud, which results in the registration failure, by combining fig. 9 with table 2; the Super4PCS algorithm is not suitable for point cloud registration with low overlapping rate and large scale; the method can obtain better registration effect for the point cloud with the overlapping rate as low as 30%.
TABLE 2 comparison of registration results
Fig. 10 is a schematic diagram of an error curve and a velocity curve according to the present invention, and as shown in fig. 10, the low overlap ratio point cloud registration method combining the dimensionality reduction projection and the feature matching of the present invention is applicable to low overlap ratio point cloud registration, and compared with other enumerated algorithms, the present invention has advantages in both velocity and precision, and the velocity advantage is more significant as the point cloud scale is larger.
In an embodiment of the present invention, there is further provided an electronic device, including a memory and a processor, where the memory stores a computer program running on the processor, and the processor executes the computer program to perform the steps of the memory management method suitable for lsi backend design as described above.
In an embodiment of the present invention, a computer-readable storage medium is further provided, on which a computer program is stored, where the computer program is executed to perform the steps of the memory management method suitable for lsi back-end design as described above.
The memory management method suitable for the large-scale integrated circuit back-end design adopts a virtual memory technology, a memory binary system storage technology and a dirty page marking technology, uniformly realizes the high-efficiency memory management method suitable for the large-scale integrated circuit back-end design, can quickly store and load data, utilizes the memory binary system direct storage to combine with the dirty page marking technology, and can naturally realize the dirty page storage technology to adapt to the future development trend of the distributed clouding technology; meanwhile, cold data can be placed on a hard disk by utilizing a Swap mechanism of the virtual memory file hidden of the linux system so as to break through the limit of the actual physical memory capacity.
Those of ordinary skill in the art will understand that: although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that changes may be made in the embodiments and/or equivalents thereof without departing from the spirit and scope of the invention. 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 (10)
1. A low-overlapping-rate point cloud registration method combining dimensionality reduction projection and feature matching comprises the following steps:
1) taking the mean normal vector direction of the point cloud as an initial dimensionality reduction direction;
2) carrying out direction correction on the initial dimensionality reduction direction to obtain a final dimensionality reduction direction;
3) projecting along the final dimension reduction direction to generate a depth image after dimension reduction;
4) generating a multi-resolution image according to the size of the grid;
5) carrying out two-step screening according to the voting mechanism and the distance invariance of rigid body transformation to obtain reliable matching point pairs;
6) carrying out two-step screening by using a voting mechanism and the distance invariance of rigid body transformation;
7) and solving the rotation matrix for the final target point pair, and performing iterative optimization by using an ICP (inductively coupled plasma) algorithm to obtain a final accurate conversion matrix.
2. The method for low-overlap point cloud registration by combining dimensionality reduction projection and feature matching according to claim 1, wherein the step 1) further comprises,
solving the normal vector, p, of discrete points by principal component analysisiAnd (3) normally measuring the eigenvector corresponding to the minimum eigenvalue of the matrix:
Eξi＝λiξi，i＝{1，2，3}，
where R is the neighborhood radius, N is the number of points in the neighborhood, piThree-dimensional coordinate vector representing ith point in the domain, p' represents three-dimensional coordinate vector of centroid of the domain, diEuclidean distance from ith point to centroid in representation field, ξiI is an integer which is greater than or equal to 1 and is a characteristic vector of the matrix E;
piand (2) measuring the eigenvector corresponding to the minimum eigenvalue of the matrix by the point normal method, and recording the eigenvector as n ═ xi，yi，zi) Then, determining the initial projection direction of the dimensionality reduction as the mean normal vector of each point of the point cloud P:
3. the method for low-overlap point cloud registration by combining dimensionality reduction projection and feature matching according to claim 1, wherein the step 2) further comprises,
uniformly dividing spatial neighborhoods along the initial projection direction, and respectively projecting along a plurality of neighborhood directions to obtain different dimension reduction images;
calculating a hash fingerprint image of the dimension-reduced image by using a hash algorithm, comparing the hamming distances of the dimension-reduced image, regarding the image with the minimum distance value as the image with the highest similarity, and regarding the corresponding direction as the corrected direction;
carrying out rotation transformation on the original point cloud to align the Z-axis direction of the rotated point cloud space coordinate system with the projection direction:
wherein the content of the first and second substances,
4. the method for low-overlap point cloud registration by combining dimensionality reduction projection and feature matching according to claim 1, wherein the step 3) further comprises,
calculate the maximum depth value max within each grid cD(c) Then, define point set P' (c) as the outermost set of points in the grid:
P′(c)＝{p∈P||maxD(c)-pz|＜δ1}；
the gaussian weight interpolation is performed on the Z coordinate of the midpoint in P' (c), and the pixel value f (c) corresponding to the grid is calculated:
wherein the normalization parameter W is expressed as:
the gaussian function g is expressed as:
and (4) carrying out interpolation calculation on the pixel value corresponding to each grid, thereby obtaining the depth image after dimension reduction.
5. The method for low-overlap point cloud registration by combining dimensionality reduction projection and feature matching according to claim 1, wherein the step 4) further comprises,
establishing a relation between the grid size D and the mean distance D of the point cloud discrete points:
d＝l·D，
wherein l is a coefficient;
the resolution is expressed as:
6. the method for low-overlap point cloud registration by combining dimensionality reduction projection and feature matching according to claim 1, wherein the step 5) further comprises,
and (3) carrying out ORB matching on n different image layers by adopting an ORB operator, wherein the corresponding image characteristic point pair of each image layer is represented as:
matches[i]，i∈[0，n-1]，
wherein, matches [0]]Corresponds to l0The layer with the highest lower resolution;
counting the repetition times of the matching relation in other layers in matches [0], and adding one to the ticket number when the following formula is satisfied:
l0×(matches[0][j].X，matches[0][j].Y)
＝li×(matches[i][k].X，matches[i][k].Y)
wherein, matches [0] [ j ] represents the jth pair of matching points in the 1 st layer, and matches [ i ] [ k ] represents the kth pair of matching points in the (i + 1) th layer.
7. The method for registering a point cloud with low overlapping rate by combining dimension reduction projection and feature matching according to claim 1, wherein the step 6) further comprises,
determining a threshold delta for the number of votes based on the number of layers2Is greater than delta2The matching points are regarded as reliable matching points and are mapped to the original point cloud to obtain Matches;
any two point pairs (p) in Matchesi，qi) And (p)j，qj) If it is notIf the matching point pair is correct, there is a relation dist (p) according to the distance invariance of rigid body transformationi，pj)＝dist(qi，qj) (ii) a Optionally selecting two pairs of points, selecting delta3> 0, such that:
the numerator represents the distance difference of the matching points, the denominator is used for reducing the sensitivity of the distance constraint condition to the point cloud scale, and the final matching point pair meeting the distance constraint is determined when the above formula is met.
8. The method for low-overlap point cloud registration by combining dimensionality reduction projection and feature matching according to claim 1, wherein the step 7) further comprises,
obtaining an initial conversion matrix of the final matching point pair by using a quaternion method, and converting the cloud of the point to be registered to a similar position;
and (4) applying an ICP (inductively coupled plasma) algorithm to carry out iterative computation to obtain an accurate conversion matrix.
9. An electronic device, comprising a memory and a processor, wherein the memory stores a computer program running on the processor, and the processor executes the computer program to perform the steps of the method for low-overlap point cloud registration combining dimensionality reduction projection and feature matching according to any one of claims 1 to 8.
10. A computer-readable storage medium, on which a computer program is stored, which when executed performs the steps of the low-overlap-rate point cloud registration method combining dimensionality reduction projection and feature matching of any one of claims 1 to 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111558350.6A CN114494368A (en) | 2021-12-20 | 2021-12-20 | Low-overlapping-rate point cloud registration method combining dimensionality reduction projection and feature matching |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111558350.6A CN114494368A (en) | 2021-12-20 | 2021-12-20 | Low-overlapping-rate point cloud registration method combining dimensionality reduction projection and feature matching |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114494368A true CN114494368A (en) | 2022-05-13 |
Family
ID=81493372
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111558350.6A Pending CN114494368A (en) | 2021-12-20 | 2021-12-20 | Low-overlapping-rate point cloud registration method combining dimensionality reduction projection and feature matching |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114494368A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114972448A (en) * | 2022-05-26 | 2022-08-30 | 合肥工业大学 | ICP algorithm-based dimensionality reduction acceleration point cloud registration method |
| CN116740156A (en) * | 2023-08-10 | 2023-09-12 | 西南交通大学 | Registration method of arbitrary pose construction element based on Gaussian sphere and principal plane distribution |
| CN117961197A (en) * | 2024-04-01 | 2024-05-03 | 贵州大学 | Self-adaptive deviation rectifying method of unmanned turbine blade micropore electric machining unit |
-
2021
- 2021-12-20 CN CN202111558350.6A patent/CN114494368A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114972448A (en) * | 2022-05-26 | 2022-08-30 | 合肥工业大学 | ICP algorithm-based dimensionality reduction acceleration point cloud registration method |
| CN116740156A (en) * | 2023-08-10 | 2023-09-12 | 西南交通大学 | Registration method of arbitrary pose construction element based on Gaussian sphere and principal plane distribution |
| CN116740156B (en) * | 2023-08-10 | 2023-11-03 | 西南交通大学 | Registration method of arbitrary pose construction element based on Gaussian sphere and principal plane distribution |
| CN117961197A (en) * | 2024-04-01 | 2024-05-03 | 贵州大学 | Self-adaptive deviation rectifying method of unmanned turbine blade micropore electric machining unit |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Salti et al. | SHOT: Unique signatures of histograms for surface and texture description | |
| JP6216508B2 (en) | Method for recognition and pose determination of 3D objects in 3D scenes | |
| CN111795704A (en) | Method and device for constructing visual point cloud map | |
| CN114677418B (en) | Registration method based on point cloud feature point extraction | |
| CN102800097B (en) | The visible ray of multi-feature multi-level and infrared image high registration accuracy method | |
| CN114494368A (en) | Low-overlapping-rate point cloud registration method combining dimensionality reduction projection and feature matching | |
| CN108388902B (en) | Composite 3D descriptor construction method combining global framework point and local SHOT characteristics | |
| CN108830888B (en) | Coarse matching method based on improved multi-scale covariance matrix characteristic descriptor | |
| CN102800099B (en) | Multi-feature multi-level visible light and high-spectrum image high-precision registering method | |
| CN101964112B (en) | Adaptive prior shape-based image segmentation method | |
| CN115861397A (en) | Point cloud registration method based on improved FPFH-ICP | |
| CN112861983A (en) | Image matching method, image matching device, electronic equipment and storage medium | |
| CN114358166B (en) | Multi-target positioning method based on self-adaptive k-means clustering | |
| Min et al. | Non-rigid registration for infrared and visible images via gaussian weighted shape context and enhanced affine transformation | |
| CN117237428A (en) | Data registration method, device and medium for three-dimensional point cloud | |
| CN111127667B (en) | Point cloud initial registration method based on region curvature binary descriptor | |
| Song | Local voxelizer: A shape descriptor for surface registration | |
| CN111582293B (en) | Plane geometry consistency detection method, computer device and storage medium | |
| CN114022526A (en) | SAC-IA point cloud registration method based on three-dimensional shape context | |
| Galun et al. | Wide baseline stereo matching with convex bounded distortion constraints | |
| Xu et al. | Point Cloud Registration with Low Overlap Based on Dimension Reduction and Feature Matching | |
| CN117788539B (en) | Point cloud data registration method and system and electronic equipment | |
| Luo et al. | Robust matching algorithm based on SURF | |
| Lian et al. | 3D-SIFT Point Cloud Registration Method Integrating Curvature Information | |
| CN117541614B (en) | Space non-cooperative target close-range relative pose tracking method based on improved ICP algorithm |
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 |