CN114186588A - Real-time hierarchical filtering method - Google Patents
Real-time hierarchical filtering method Download PDFInfo
- Publication number
- CN114186588A CN114186588A CN202111493587.0A CN202111493587A CN114186588A CN 114186588 A CN114186588 A CN 114186588A CN 202111493587 A CN202111493587 A CN 202111493587A CN 114186588 A CN114186588 A CN 114186588A
- Authority
- CN
- China
- Prior art keywords
- point
- points
- point cloud
- filtering
- algorithm
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
Abstract
The invention discloses a real-time hierarchical filtering method, which comprises the steps of collecting output data of a laser radar; performing conditional filtering to divide the point cloud into near points Gg0And non-near point Gu1(ii) a Dividing the near points into ground points G conforming to a plane model by using a random sampling consistency algorithmgAnd non-ground point Gu2(ii) a Merge Gu1And Gu2Performing rapid coarse filtering on the merged point cloud by adopting a curved surface voxel; selectively down-sampling the point cloud after rough filtering and selectively reducing the dimension of principal component analysis; determining DBSCAN algorithm parameters Eps and Nmin(ii) a Carrying out precise filtering by a multi-thread parallel execution DBSCAN algorithm to obtain effective point cloud GE(ii) a And 8: merging effective point clouds GEAnd ground point GgAnd finishing the filtering. The invention solves the problems of difficult parameter setting and difficult speed and precision compromise in the laser radar data filtering process, reduces the algorithm complexity and ensures the algorithm real-time propertyAnd real-time hierarchical filtering based on the point cloud distribution characteristics is realized, and the purpose of real-time high-precision point cloud filtering is achieved.
Description
Technical Field
The invention belongs to the technical field of three-dimensional point cloud processing, and relates to a real-time hierarchical filtering method, in particular to a real-time hierarchical filtering method based on point cloud distribution characteristics.
Background
The quality of the point cloud data determines the accuracy of the front-end matching algorithm, which in turn affects the accuracy of the positioning algorithm. However, some noise points inevitably appear in the point cloud data acquired in the outdoor environment due to factors such as equipment accuracy, operator experience, environmental changes, electromagnetic wave diffraction characteristics, and changes in the surface properties of the measured object. In addition, due to the influence of external interference factors such as sight line obstruction, obstacles and the like, discrete points far away from the main point cloud exist in the data. The invalid point influences the processes of feature extraction, point cloud matching and the like in the positioning algorithm, so point cloud filtering is always one of important research directions in the point cloud research field.
Most of the data of the laser radar sensor exists in a disordered form, and disordered point cloud filtering can be divided into two types according to whether the data are based on a grid model or not. The mesh model based filtering means that acquired disordered point cloud data is converted into a mesh model and mesh feature information is calculated, but the efficiency is low in any step, so that the method is gradually replaced by a non-mesh model based method. The non-grid model filtering has the main idea that all data points are traversed, and the relation between a sampling point and surrounding neighborhood points is directly constructed. The method can be widely applied because the noise reduction process can be simplified, and a mainstream filtering method based on statistics, neighborhood, projection and density is further extended. The filtering method based on statistics adopts a statistical concept, and although a complex mathematical method can increase the detection precision and robustness of noise, a large amount of computing resources are consumed, and the point cloud characteristics cannot be well maintained under some conditions. The filtering method based on the neighborhood mainly utilizes the similarity between sampling points and neighborhood points, wherein the representative bilateral filtering has very low sensitivity to small-scale noise besides large calculation amount; the radius filtering is based on the connectivity analysis and has strong parameter dependence. The projection-based filtering method is similar to bilateral filtering, and the point cloud edge smoothing is realized through the positions of moving points. The method has poor real-time performance and a plurality of limiting conditions, and is not beneficial to engineering application. The density-based filtering method is widely applied due to the advantages of strong identification, high sensitivity and the like, but the efficiency of processing point cloud under a complex scene is greatly reduced. In summary, the traditional point cloud filtering algorithm does not consider the point cloud distribution characteristics, the algorithm adaptability is limited by the used parameters, and meanwhile, the speed and the precision of the algorithm cannot be considered due to different application backgrounds.
Disclosure of Invention
Aiming at the prior art, the invention aims to provide a real-time hierarchical filtering method based on point cloud distribution characteristics, and solve the problems of difficult parameter setting and difficult speed and precision compromise in the laser radar data filtering process.
In order to solve the technical problem, the real-time hierarchical filtering method of the invention comprises the following steps:
step 1: collecting laser radar output data;
step 2: performing conditional filtering to divide the point cloud into near points Gg0And non-near point Gu1；
And step 3: dividing the near points into ground points G conforming to a plane model by using a random sampling consistency algorithmgAnd non-ground point Gu2；
And 4, step 4: merge Gu1And Gu2Performing rapid coarse filtering on the merged point cloud by adopting a curved surface voxel;
and 5: selectively down-sampling the point cloud after rough filtering and selectively reducing the dimension of principal component analysis;
step 6: determining DBSCAN algorithm parameters Eps and Nmin；
And 7: carrying out precise filtering by a multi-thread parallel execution DBSCAN algorithm to obtain effective point cloud GE；
And 8: merge validationPoint cloud GEAnd ground point GgAnd finishing the filtering.
Further, conditional filtering is carried out in the step 2, and the point cloud is divided into near points Gg0And non-near point Gu1The method specifically comprises the following steps: traversing the point cloud, judging whether the three-dimensional coordinates of each point meet the preset range condition, and further separating the near place G meeting the range conditiong0And non-perigee G not meeting the range conditionu1。
Further, in step 3, a random sampling consistency algorithm is applied to divide the near points into ground points G which conform to the plane modelgAnd non-ground point Gu2The method comprises the following steps:
step 3.1: at a near point Gg0Randomly selecting three non-collinear points (x) in the data seti,yi,zi) i is 1,2,3, and is assumed to be an in-office point;
step 3.2: judging whether other points in the point cloud meet a plane model according to a set distance threshold dis _ thre, wherein the plane model is as follows:
Axi+Byi+Czi+D＝0
wherein A, B, C, D is a known plane parameter, and A, B, C is not 0 at the same time.
Computing a three-dimensional spatial midpoint (x)4,y4,z4) The distance to the plane is given by:
if d is less than dis _ thre, marking the point as an in-office point, and recording the number A of in-office points conforming to the plane modelnN represents the nth random selection point, initializing A 00; when A isn>An-1Then, choose AnCorresponding planar model, otherwise, selecting An-1A corresponding planar model;
step 3.3: when the set iteration ending condition is reached, the local interior points of the plane model selected in the step 3.2 form a ground point GgThe rest is marked as the ground point Gu2(ii) a If not, then,and re-randomly selecting three non-collinear points, and returning to the step 3.2.
Further, the step 4 of performing fast coarse filtering on the merged point cloud by using the surface voxels comprises:
traversing the point cloud, converting the three-dimensional coordinate points into spherical coordinates, and constructing a spatial curved surface voxel;
obtaining a curved surface voxel index, constructing a hash table of each point mapped to the voxel index, and finishing a clustering process through the hash table to obtain a clustering label;
and traversing the clustering labels, and removing the class with the point number smaller than the threshold value as the outlier.
Further, converting the three-dimensional coordinate points into spherical coordinates, and constructing a spatial surface voxel specifically comprises:
wherein P is a point converted to a spherical coordinate system, P is a radial distance, theta is an azimuth angle,
is the vertical angle, Δ ρ, Δ θ and
is a unit parameter.
Further, in step 5, the selective down-sampling processing and the principal component analysis selective dimension reduction processing are performed on the local point cloud after coarse filtering, and the processing method includes:
step 5.1: judgment of
Whether the current category is satisfied or not is judged, wherein N represents the number of the current category points, A represents the area of the current category, and epsilon is a given threshold value;
step 5.2: when the current category is established, the current category is down-sampled, and then step 5.3 is executed, otherwise step 5.3 is executed;
step 5.3: calculating the covariance matrix of the current category point cloud and obtaining the corresponding eigenvalue lambda through SVD1,λ2,λ3And a feature vector ε1,ε2,ε3Defining a discriminant formula:
wherein λ is1,λ2,λ3Is an eigenvalue of a covariance matrix, and1＞λ2＞λ3η is a given threshold;
judging whether a discrimination formula is established, and reducing the point cloud dimension to two dimensions when the discrimination formula is established to obtain point cloud data after dimension reduction; otherwise, the point cloud is maintained as it is.
Further, in step 6, DBSCAN algorithm parameters Eps and N are determinedminThe method comprises the following steps:
selecting Nmin：
Nmin2 x D-1, D x dimension;
traversing the current category point cloud, and calculating the Nth distance from each point to the current category point cloudminAnd the distances of the near points are sorted from large to small, the coordinate of the inflection point position is found according to a discriminant, namely the value of the Eps, and the discriminant is as follows:
(Dist[i-1]-Dist[i])/(Dist[i]-Dist[i+1])＞ε
wherein Dist [ k ]]N denotes the k-th point to the N-th point from itminThe distance of the near points, ε, represents the gradient factor between adjacent points.
Further, in step 7, the multi-thread parallel execution DBSCAN algorithm is performed for fine filtering to obtain the effective point cloud GEThe method comprises the following steps: firstly, randomly selecting a sampling point from a point cloud, searching the number of points in the neighborhood of the sampling point in the radius Eps range, and enabling the number of points in the neighborhood to be larger than NminThe point of (2) is marked as a core point; secondly, traversing the data set, finding all density reachable objects of random sampling points to form a point cloud cluster C, and passing the densityConnecting to form a new data set; then, marking points which are in the neighborhood range of the core point but do not belong to the core point in the traversal process as edge points, and marking points which do not belong to the core point or the edge points as noise points; finally, the point cloud is divided into three labels of a core point, an edge point and a noise point, the core point and the edge point are reserved as effective points, the noise point is removed as small-scale noise, and then effective point cloud G is obtainedE。
The invention has the beneficial effects that: because the point cloud presents the distribution characteristic of close density and distant sparsity, the adaptability of the conventional filtering algorithm is reduced by fixing or calculating parameters based on the distance function; the usual filter design is a compromise between speed and accuracy, high accuracy algorithms suffer from poor real-time performance due to computational complexity, and high efficiency algorithms mean the abandonment of details and loss of accuracy. According to the invention, a new point cloud filtering thought is provided through research and performance analysis of the existing point cloud filtering algorithm, aiming at the distribution characteristics of close density and far sparse point cloud, firstly, point cloud filtering is introduced by adopting the idea of dividing the point cloud based on curved surface voxels, the influence of distance on parameters is weakened, voxels with volume related to sensor resolution and distance are constructed according to the point cloud distribution characteristics, and the mapping relation between points and voxels is established, so that the problem of difficult parameter setting is solved, and the rapid coarse filtering of the point cloud is realized on the basis of self-adaption of clustering parameters; secondly, in order to further improve the precision, the local point cloud is finely processed by combining a density clustering-based DBSCAN algorithm, so that the defect of rough clustering of the local point cloud is overcome. In addition, in order to meet the requirements of actual engineering on algorithm calculation amount and real-time performance, the method firstly reduces the number of point clouds by selectively down-sampling aiming at a coarse filtering result, secondly selectively reduces the dimensionality of the point clouds after carrying out principal component analysis on the point clouds, thereby reducing the complexity of the algorithm, and finally ensures the real-time performance of the algorithm based on multi-thread parallel computation, thereby realizing real-time hierarchical filtering based on the distribution characteristics of the point clouds and achieving the aim of real-time high-precision point cloud filtering.
Drawings
FIG. 1 is a schematic diagram of a real-time hierarchical filtering algorithm based on point cloud distribution characteristics according to the present invention;
FIG. 2 is a motion profile based on a public data set;
FIG. 3 is a partial enlarged view of the motion trajectory 1 (public);
FIG. 4 is a partial enlarged view of the motion trajectory 2 (public);
FIG. 5 is an x-direction error map (public) of the point cloud filtering algorithm, the original algorithm and the comparison algorithm of the present invention;
FIG. 6 is a y-direction error diagram (public) of the point cloud filtering algorithm, the original algorithm and the comparison algorithm of the invention;
FIG. 7 is an enlarged view (public) of the position error of the X-direction track end of the point cloud filtering algorithm, the original algorithm and the comparison algorithm of the invention;
FIG. 8 is an enlarged view (public) of the position error of the Y-direction track end of the point cloud filtering algorithm, the original algorithm and the comparison algorithm of the invention;
FIG. 9 is a front end runtime table (public) of the point cloud filtering algorithm, the original algorithm, and the comparison algorithm of the present invention;
FIG. 10 is a table (public) of root mean square error of paths of the point cloud filtering algorithm, the original algorithm and the comparison algorithm of the present invention;
FIG. 11 is a table (public) of path standard deviations of the point cloud filtering algorithm, the original algorithm, and the comparison algorithm of the present invention;
FIG. 12 is a motion profile based on a self-test data set;
FIG. 13 is a partial enlarged view of the motion trajectory 1 (self test);
FIG. 14 is a partial enlarged view of the motion trajectory 2 (self test);
FIG. 15 is a graph of error in the x direction (self-test) of the point cloud filtering algorithm, the original algorithm, and the comparison algorithm of the present invention;
FIG. 16 is a y-direction error diagram (self-test) of the point cloud filtering algorithm, the original algorithm, and the comparison algorithm of the present invention;
FIG. 17 is an enlarged view (self-test) of the position error of the X-direction track end of the point cloud filtering algorithm, the original algorithm and the comparison algorithm;
FIG. 18 is an enlarged view (self-test) of the position error of the Y-direction track end of the point cloud filtering algorithm, the original algorithm and the comparison algorithm of the invention
FIG. 19 is a front end runtime table (self test) of the point cloud filtering algorithm, the original algorithm, and the comparison algorithm of the present invention;
FIG. 20 is a table of root mean square error of the paths of the point cloud filtering algorithm, the original algorithm, and the comparison algorithm (self-test) according to the present invention;
FIG. 21 is a table of path standard deviations (self-test) of the point cloud filtering algorithm, the original algorithm, and the comparison algorithm of the present invention.
Detailed Description
The invention is further described with reference to the following figures and specific embodiments.
With reference to fig. 1, the embodiment of the present invention includes the following steps:
(1) collecting laser radar output data;
(2) performing conditional filtering to divide the point cloud into near points Gg0And non-near point Gu1；
(3) Selecting out ground points G conforming to a plane model from the near points by using a Random Sample consensus (RANSAC)gAnd non-ground point Gu2；
(4) Merge Gu1And Gu2Performing rapid coarse filtering on the merged point cloud based on the curved surface voxels, firstly converting point cloud coordinates and constructing spatial curved surface voxels; secondly, completing a clustering process through a hash table in an accelerating way; finally, the category points are used as the discrimination conditions of the large-scale noise;
(5) traversing the coarse clustering filtering result, and performing downsampling on the point cloud class with overlarge point cloud number or area;
(6) calculating a point cloud covariance matrix and obtaining a characteristic value lambda by decomposition calculation1,λ2,λ3And a feature vector ε1,ε2,ε3Determining whether to perform principal component analysis dimensionality reduction according to a judgment formula, if so, reducing the point cloud dimensionality to two dimensions based on the first and second dimensional feature vectors, and otherwise, keeping the point cloud dimensionality unchanged;
(7) calculating each point to the Nth point therefromminThe close point spacing is ranked from big to small, then the inflection point is found out to be used as the Eps neighborhood radius parameter of the DBSCAN algorithm, and the neighborhood point number parameter N of the DBSCAN algorithm under different dimensions is obtained according to the empirical valuemin. General twoAnd N in three dimensionsminRespectively taking 3 and 5;
(8) performing fine filtering by using a multi-thread parallel execution DBSCAN, reserving core points and edge points as effective points, removing noise points as small-scale noise to obtain effective point cloud GE；
(9) Merging point cloud after fine filtering and ground point GgAnd finishing the filtering.
Wherein conditional filtering is performed to divide the point cloud into near points Gg0And non-near point Gu1The method specifically comprises the following steps:
and traversing the point cloud, judging whether the three-dimensional coordinates of each point meet a preset coordinate range condition, and marking the coordinates as a near place when the coordinates are in the range condition. Thereby separating out the near point Gg0And a large number of spots Gu1。
Wherein, a Random Sample consensus algorithm (Random Sample consensus, RANSAC) is used to select a ground point G conforming to the plane model from the near pointsgAnd non-ground point Gu2The method specifically comprises the following steps:
on the basis of screening out the near points through conditional filtering, a plurality of points are randomly selected from the data set and are assumed as local points, and the selected points are determined by the model characteristics. Since the invention requires fitting the ground, three points (x) are chosen that are non-collineari,yi,zi) i is 1,2, 3. Judging whether other points in the point cloud meet a plane model according to a preset distance threshold dis _ thre, wherein the plane model is as follows:
Axi+Byi+Czi+D＝0
wherein A, B, C, D is a known plane parameter, and A, B, C is not 0 at the same time.
Computing a three-dimensional spatial midpoint (x)4,y4,z4) The distance to the plane is given by:
if d < dis _ thre (dis _ thre is usually about 0.1), mark the point as a local point and record the characterNumber of local points of the synthetic plane model. And replacing the initially selected points for repeated iteration, and discarding the plane model generated in the iterative process if the number of the points in the local is small, and selecting the plane model if the point number is better than that of the existing model. When the iteration times are satisfied or the sum of the distances from all local points to the plane is small enough, the loop is jumped out, and finally the most reasonable result G in the whole iteration process is obtainedgThe rest is marked as the ground point Gu2。
Wherein G is combinedu1And Gu2And performing rapid coarse filtering on the merged point cloud. Firstly, converting point cloud coordinates and constructing a spatial curved surface voxel; secondly, completing a clustering process through a hash table in an accelerating way; finally, the discrimination condition of taking the category point number as the large-scale noise is specifically as follows:
aiming at the problem that the point cloud presents the distribution characteristics of close density, far distance and sparse density, the adaptability of the traditional filtering algorithm can be reduced by fixing or calculating parameters based on a distance function, the method adopts the idea of curved surface voxels to weaken the influence of the distance on the parameters, firstly, voxels related to the resolution and distance of the sensor are constructed, and the mapping relation between the points and the voxels is established; secondly, clustering the voxels where the sampling points are located and surrounding voxels, and combining the voxels to form clusters to obtain clustering labels, namely points of each category; and finally, the category point number is used as a discrimination condition of large-scale noise, and the category with the point number smaller than a threshold value is taken as an outlier to be removed, so that the rapid coarse filtering is realized.
a. Surface voxels:
i.e. a spatial cell containing three-dimensional spherical coordinates. Surface voxel SVi,j,kThe points in the spherical coordinate system contained in the method are as shown in the formula:
wherein P represents a point converted into a spherical coordinate system, ρ represents a radial distance, θ represents an azimuth angle,
representing a vertical angle. Δ ρ, Δ θ and
representing unit parameter size, in relation to sensor resolution.
b. Hash table accelerated coarse filtering:
merge Gu1And Gu2Traversing the point cloud, converting the three-dimensional coordinate point into a spherical coordinate, and combining unit parameters delta rho, delta theta and
and obtaining the index of the voxel on the curved surface, constructing a preliminary hash table of each point mapped to the voxel index, and deleting the voxel without point cloud. After the table is constructed, each point P is accessed, around the target voxel containing the point P
And finding out adjacent points in each voxel, and combining to form a cluster to obtain a cluster label, namely the number of points in each category. And finally, the category points are used as the discrimination conditions of the large-scale noise, the clustering labels are traversed, and the categories with the points smaller than the threshold are taken as outliers to be removed, so that the aim of quickly removing the large-scale noise is fulfilled.
Although the threshold setting process is simplified while the algorithm speed is ensured in the process, the point cloud clustering process is quite rough. In order to further improve the precision, a DBSCAN algorithm capable of identifying noise and realizing arbitrary shape clustering is introduced on the basis to solve the problem of small-scale noise of local point cloud, and finally, coarse-to-fine hierarchical filtering is realized. The DBSCAN algorithm is high in complexity, long in convergence time for large data sample clustering and difficult to meet the requirement of real-time performance. Although the KD tree or octree management data can improve the searching speed of the algorithm, the KD tree-DBSCAN algorithm is difficult to apply in practical engineering, and the acceleration of the KD tree-DBSCAN algorithm is accelerated by adopting main component analysis selectivity reduction and down-sampling.
Traversing the coarse clustering filtering result, and performing downsampling on the point cloud class with the overlarge point cloud number or area specifically comprises the following steps:
point cloud down-sampling is to voxelize the three-dimensional space where the point cloud is located, and if a laser point exists in a voxel, the centroid of a point set in the voxel is used for replacing the point set. However, considering that the position of the point is changed by down-sampling, only the class with the point number and area meeting the conditions is down-sampled before the selective dimension reduction, and the judgment standard is as follows:
in the formula, N represents the number of current category points, A represents the area of current category, and N is generalmaxTake 5000, Amax Take 50, ε takes 1.
Wherein, a point cloud covariance matrix is calculated and an eigenvalue lambda is obtained through SVD decomposition calculation1,λ2,λ3And a feature vector ε1,ε2,ε3And determining whether to perform principal component analysis dimensionality reduction according to a judgment formula, if so, reducing the point cloud dimensionality to two dimensions based on the first and second dimension feature vectors, and otherwise, keeping the point cloud dimensionality unchanged specifically:
assuming that the number of the three-dimensional point cloud data is m, the point cloud data is firstly arranged into a matrix X with 3 rows and m columns, and each row of the matrix X is subjected to zero equalization, that is, each row of data is subjected to subtraction of the current row average value, so that each row of data of the matrix takes zero as the center. This ensures that the average value for each row is zero, which is expressed as:
X＝[p1,p2,...,pm],pi＝[xi,yi,zi]T(i＝1,2,...,m)
the covariance matrix C is obtained by:
let 3 eigenvectors of covariance matrix C be e1,e2,e3Forming a matrix E ═ (E) by columns1,e2,e3) It is possible to obtain:
and Λ is a diagonal matrix, and diagonal elements are eigenvalues corresponding to the eigenvectors of the covariance matrix. Thus, the dimensionality reduction matrix P ═ E is obtainedT。
Due to different point cloud distribution conditions, effective data can be lost by falsely reducing the dimension of the data. For this case, the criteria introduced are as follows:
in the formula, λi(i ═ 1,2,3) represents the eigenvalues of the above covariance matrix and λ1＞λ2＞λ3. That is, when the ratio of the first and second feature values is greater than η, η is generally (0.95,1) and is set to allow principal component analysis dimensionality reduction.
If the eigenvalue of the covariance matrix corresponding to the point set X 'meets the above formula, the first two rows of the dimensionality reduction matrix P are taken to form a matrix P'. And obtaining the point cloud data Y after dimensionality reduction by using the point cloud data Y as P 'X'.
Wherein each point is calculated to be N from itminThe close point spacing is ranked from big to small, then the inflection point is found out to be used as the Eps neighborhood radius parameter of the DBSCAN algorithm, and the neighborhood point number parameter N of the DBSCAN algorithm under different dimensions is obtained according to the empirical valueminThe method specifically comprises the following steps:
after point cloud is input, algorithm parameters are initializedNumbers Eps and NminDue to the uneven density of the point cloud sample, the consistency of the parameters is kept, so that the adaptability of the algorithm is poor, and N is an appropriate parameter for ensuring that each rough filtering result can be obtainedminThe selection criteria, depending on the dimensions, are shown as follows:
Nmin2 x D-1, D x dimension
The neighborhood radius parameter Eps is obtained by self-adaptive calculation according to the point cloud distribution condition, and proper N is selected according to the method for the point clouds with different dimensionsminThen, traversing the current category point cloud, and calculating to obtain a point-distance correlation value, namely calculating the distance from each point to the Nth pointminAnd the distances of the near points are sorted from large to small, and finally, the coordinate of the inflection point position is found according to a discriminant, namely the value of the Eps. The discriminant is as follows:
(Dist[i-1]-Dist[i])/(Dist[i]-Dist[i+1])＞ε
wherein Dist [ k ]]N denotes the k-th point to the N-th point from itminDistance of the near point. Epsilon represents a gradient factor between adjacent points, and is generally 3 or 4;
the multithread parallel execution DBSCAN carries out fine filtering, the core points and the edge points are reserved as effective points, and the noise points are specifically removed as small-scale noise:
considering that the fine filtering process is not influenced mutually, firstly, arranging the coarse filtering results according to the size of the number of points; secondly, reasonably grouping to enable the classes with more points to be different from one group, and finally adopting multi-thread parallel computing. In order to improve the neighborhood searching speed of each point in the algorithm, the KD tree is introduced into the algorithm to improve the operation efficiency. Firstly, randomly selecting a sampling point from a point cloud, searching the number of points in the neighborhood of the sampling point in the radius Eps range, and enabling the number of points in the neighborhood to be larger than NminThe point of (2) is marked as a core point; secondly, traversing the data set, finding out all density reachable objects of random sampling points to form a point cloud cluster C, and forming a new data set through density connection; then, in the traversal process, marking points which are in the neighborhood range of the core point but do not belong to the core point as edge points, and marking points which do not belong to the core point or the edge points as noise points; finally, the point cloud is divided into core points and edgesPoint and noise point. As the DBSCAN does not need to know the category number in advance and is not limited by the shape of the point cloud, the abnormal points can be found well by the characteristics of the DBSCAN, and therefore the purpose of removing small-scale noise is achieved by removing the noise according to the labels.
The invention is based on the public data set and the measured data set, and compares the invention with the statistical filtering and spherical filtering algorithm through two aspects of qualitative analysis and quantitative analysis:
firstly, public data set verification is carried out on the method:
in order to verify the point cloud filtering performance of the invention, data collected by HDL-64E-Velodyne laser radar in a public data set is used, a GPS/IMU combined navigation system provided by the public data set is used as a position reference system, and an ALOAM algorithm is selected to calculate the positioning result of the laser odometer. The whole data collection process is about 470s, the traditional algorithm cannot be normally used due to the fact that the data volume is too large, and data are uniformly subjected to down-sampling before experiments.
FIG. 2 is a public data set based motion profile showing various profiles based on the public data set. The figure shows that the coincidence degree of each algorithm with the GPS is better on the whole performance, no large-scale deviation occurs, and the experiment has effectiveness. Fig. 3 and 4 are partial enlarged views of the motion trajectory, and it can be seen that the addition of the conventional filtering method not only does not improve the accuracy of the positioning algorithm, but also has a negative effect; the method has more adaptability, can use more appropriate parameters in a constantly changing environment and ensure the precision, and optimizes the original algorithm to enable the result to be closer to a reference value.
Fig. 5 and 6 show horizontal position errors, and the results show that the conventional filtering method fluctuates sharply above and below the 0 value, but the results of the present invention are more stable in the whole process compared with the ALOAM original algorithm and the two sets of comparison algorithms. Fig. 7 and 8 are enlarged views of position errors at the ends of the motion trajectories, and it is easy to see that the result obtained by the present invention has the best effect in the Y direction, which is improved by 18.25% compared with the accuracy of the original algorithm, and also has a certain effect in the X direction, although slightly worse than the traditional filtering algorithm, improved by 20.47% compared with the accuracy of the original algorithm.
The above are qualitative analyses, which only help to understand the algorithm effect intuitively, and the error of the path end point is not representative enough. In order to more conveniently and objectively analyze and compare the performance and accuracy of each algorithm, quantitative analysis using the algorithm running time, the root mean square error and the standard deviation is required. Fig. 9 is a comparison table of the operation time of the algorithm, and it can be seen that the application of various methods to the front end of the algorithm can maintain the processing time within 100ms, thereby ensuring the real-time condition of the system operation.
FIGS. 10 and 11 show the public data set path root mean square error comparison and path standard deviation comparison, respectively. It can be obviously seen that the introduction of the traditional algorithm has no effect, but weakens the performance of the algorithm, and further proves that the traditional algorithm lacks adaptability due to the problems of threshold setting and the like, the original accuracy of the algorithm is reduced after effective data is removed, the accuracy and the data stability of the algorithm are improved to a certain extent, the accuracy is improved by 8.816% in the X direction and by 8.869% in the Y direction compared with the original algorithm, and the data stability is improved by 9.170% in the X direction and by 6.155% in the Y direction compared with the original algorithm.
Secondly, the self-test data set verification is carried out on the invention:
in order to further verify the point cloud filtering performance of the invention, data collected by Velodyne-32E laser radar in a self-testing data set is used, a GPS/IMU combined navigation system of GI-7660 type provided by a self-built platform is used as a position reference system, and the ALOAM algorithm is also selected to calculate the positioning result of the laser odometer. The whole data set acquisition process is about 307s, and the data volume is smaller than 64 lines of laser radar, which is more close to the actual situation. According to experimental verification, statistical filtering cannot meet the real-time requirement, and is combined with down-sampling for use.
FIG. 12 is a graph of motion profiles for various scenarios in a self-test dataset. It can be seen that the degree of agreement of each algorithm is reduced compared to the public data set because a large number of moving objects exist in the self-test data set and the lidar data are not distorted and removed, the performance of the original algorithm is reduced, but large-scale deviation does not occur, and the experiment has effectiveness. Fig. 13 and 14 are partial enlarged views of the motion trajectory, and it can be seen that the addition of the conventional filtering method not only does not improve the accuracy of the positioning algorithm, but also has a negative effect; the method has more adaptability, can use more appropriate parameters in a constantly changing environment and ensure the precision, and optimizes the original algorithm to enable the result to be closer to a reference value.
The results of fig. 15 and 16 are horizontal position errors, which show that the algorithm of the present invention has substantially the same trend as the original algorithm and the two sets of comparison algorithms, but the algorithm of the present invention has smaller fluctuation and is maintained at a value around 0 to a greater extent. FIGS. 17 and 18 are enlarged views of the position error at the end of the motion trajectory, and it is easy to see that the results obtained by the present invention have good effect no matter in the X direction or the Y direction, the precision of the X direction is improved by 80.39% compared with the original algorithm, and the precision of the Y direction is improved by 94.56% compared with the original algorithm
The above is a qualitative analysis based on self-test data sets, and it is also desirable to further quantify the performance of the present invention. Fig. 19 is an algorithm running time comparison table, and it can be seen from the figure that the processing time can be maintained within 100ms when various methods are applied to the front end of the algorithm, so as to ensure the real-time condition of the system running, but compared with the using condition of statistical filtering, the algorithm of the present invention does not need to downsample data, and avoids introducing new error information due to the movement of the point cloud position.
Fig. 20 and fig. 21 show the root mean square error comparison result and the standard deviation comparison result of the path of the self-test data set, respectively, and it is obvious that, unlike the public data set, the performance of the algorithm can be improved by introducing the conventional algorithm, which indicates that the environmental adaptability of the conventional filtering algorithm is poor due to the limitation of the parameters, and the same parameters produce different effects in different environments. Compared with the original algorithm, the algorithm of the invention reduces 4.44% in the Y direction in precision, but improves 50.77% in the X direction, and improves 33.76% and 38.23% in data stability in two directions respectively.
Claims (8)
1. A real-time hierarchical filtering method is characterized by comprising the following steps:
step 1: collecting laser radar output data;
step 2: performing conditional filtering to divide the point cloud into near points Gg0And non-near point Gu1；
And step 3: dividing the near points into ground points G conforming to a plane model by using a random sampling consistency algorithmgAnd non-ground point Gu2；
And 4, step 4: merge Gu1And Gu2Performing rapid coarse filtering on the merged point cloud by adopting a curved surface voxel;
and 5: selectively down-sampling the point cloud after rough filtering and selectively reducing the dimension of principal component analysis;
step 6: determining DBSCAN algorithm parameters Eps and Nmin；
And 7: carrying out precise filtering by a multi-thread parallel execution DBSCAN algorithm to obtain effective point cloud GE；
And 8: merging effective point clouds GEAnd ground point GgAnd finishing the filtering.
2. The real-time hierarchical filtering method according to claim 1, wherein: step 2, conditional filtering is carried out, and the point cloud is divided into near points Gg0And non-near point Gu1The method specifically comprises the following steps: traversing the point cloud, judging whether the three-dimensional coordinates of each point meet the preset range condition, and further separating the near place G meeting the range conditiong0And non-perigee G not meeting the range conditionu1。
3. The real-time hierarchical filtering method according to claim 1, wherein: step 3, dividing the near points into ground points G conforming to the plane model by using a random sampling consistency algorithmgAnd non-ground point Gu2The method comprises the following steps:
step 3.1: at a near point Gg0Randomly selecting three non-collinear points (x) in the data seti,yi,zi) i is 1,2,3, and is assumed to be an in-office point;
step 3.2: judging whether other points in the point cloud meet a plane model according to a set distance threshold dis _ thre, wherein the plane model is as follows:
Axi+Byi+Czi+D＝0
wherein A, B, C, D is a known plane parameter, and A, B, C is not 0 at the same time.
Computing a three-dimensional spatial midpoint (x)4,y4,z4) The distance to the plane is given by:
if d is less than dis _ thre, marking the point as an in-office point, and recording the number A of in-office points conforming to the plane modelnN represents the nth random selection point, initializing A00; when A isn>An-1Then, choose AnCorresponding planar model, otherwise, selecting An-1A corresponding planar model;
step 3.3: when the set iteration ending condition is reached, the local interior points of the plane model selected in the step 3.2 form a ground point GgThe rest is marked as the ground point Gu2(ii) a Otherwise, randomly selecting three non-collinear points again, and returning to the step 3.2.
4. The real-time hierarchical filtering method according to claim 1, wherein: step 4, the step of performing rapid coarse filtering on the merged point cloud by adopting the curved surface voxels comprises the following steps:
traversing the point cloud, converting the three-dimensional coordinate points into spherical coordinates, and constructing a spatial curved surface voxel;
obtaining a curved surface voxel index, constructing a hash table of each point mapped to the voxel index, and finishing a clustering process through the hash table to obtain a clustering label;
and traversing the clustering labels, and removing the class with the point number smaller than the threshold value as the outlier.
5. The real-time hierarchical filtering method according to claim 4, wherein: converting the three-dimensional coordinate points into spherical coordinates, and constructing the space surface voxels specifically comprises the following steps:
wherein P is a point converted to a spherical coordinate system, P is a radial distance, theta is an azimuth angle,
is the vertical angle, Δ ρ, Δ θ and
is a unit parameter.
6. The real-time hierarchical filtering method according to claim 1, wherein: and 5, performing selective down-sampling processing and principal component analysis selective dimension reduction processing on the local point cloud after rough filtering, wherein the processing comprises the following steps:
step 5.1: judgment of
Whether the current category is satisfied or not is judged, wherein N represents the number of the current category points, A represents the area of the current category, and epsilon is a given threshold value;
step 5.2: when the current category is established, the current category is down-sampled, and then step 5.3 is executed, otherwise step 5.3 is executed;
step 5.3: calculating the covariance matrix of the current category point cloud and obtaining the corresponding eigenvalue lambda through SVD1,λ2,λ3And a feature vector ε1,ε2,ε3Defining a discriminant formula:
wherein λ is1,λ2,λ3Is an eigenvalue of a covariance matrix, and1＞λ2＞λ3η is a given threshold;
judging whether a discrimination formula is established, and reducing the point cloud dimension to two dimensions when the discrimination formula is established to obtain point cloud data after dimension reduction; otherwise, the point cloud is maintained as it is.
7. The real-time hierarchical filtering method according to claim 1, wherein: step 6, determining the parameters Eps and N of the DBSCAN algorithmminThe method comprises the following steps:
selecting Nmin：
Nmin2 x D-1, D x dimension;
traversing the current category point cloud, and calculating the Nth distance from each point to the current category point cloudminAnd the distances of the near points are sorted from large to small, the coordinate of the inflection point position is found according to a discriminant, namely the value of the Eps, and the discriminant is as follows:
(Dist[i-1]-Dist[i])/(Dist[i]-Dist[i+1])＞ε
wherein Dist [ k ]]N denotes the k-th point to the N-th point from itminThe distance of the near points, ε, represents the gradient factor between adjacent points.
8. The real-time hierarchical filtering method according to claim 1, wherein: 7, carrying out fine filtering by the multi-thread parallel execution DBSCAN algorithm to obtain effective point cloud GEThe method comprises the following steps: firstly, randomly selecting a sampling point from a point cloud, searching the number of points in the neighborhood of the sampling point in the radius Eps range, and enabling the number of points in the neighborhood to be larger than NminThe point of (2) is marked as a core point; secondly, traversing the data set, finding out all density reachable objects of random sampling points to form a point cloud cluster C, and forming a new data set through density connection; then, the points in the neighborhood range of the core point but not belonging to the core point in the traversal process are marked as edge points, and the points will not belong to the core point nor the core pointPoints not belonging to the edge points are marked as noise points; finally, the point cloud is divided into three labels of a core point, an edge point and a noise point, the core point and the edge point are reserved as effective points, the noise point is removed as small-scale noise, and then effective point cloud G is obtainedE。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111493587.0A CN114186588A (en) | 2021-12-08 | 2021-12-08 | Real-time hierarchical filtering method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111493587.0A CN114186588A (en) | 2021-12-08 | 2021-12-08 | Real-time hierarchical filtering method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114186588A true CN114186588A (en) | 2022-03-15 |
Family
ID=80603898
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111493587.0A Pending CN114186588A (en) | 2021-12-08 | 2021-12-08 | Real-time hierarchical filtering method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114186588A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116310360A (en) * | 2023-05-18 | 2023-06-23 | 实德电气集团有限公司 | Reactor surface defect detection method |
-
2021
- 2021-12-08 CN CN202111493587.0A patent/CN114186588A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116310360A (en) * | 2023-05-18 | 2023-06-23 | 实德电气集团有限公司 | Reactor surface defect detection method |
| CN116310360B (en) * | 2023-05-18 | 2023-08-18 | 实德电气集团有限公司 | Reactor surface defect detection method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Ji et al. | A novel simplification method for 3D geometric point cloud based on the importance of point | |
| CN106600622A (en) | Point cloud data partitioning method based on hyper voxels | |
| CN106886980B (en) | Point cloud density enhancement method based on three-dimensional laser radar target identification | |
| CN113628263A (en) | Point cloud registration method based on local curvature and neighbor characteristics thereof | |
| CN116310849B (en) | Tree point cloud monomerization extraction method based on three-dimensional morphological characteristics | |
| CN115861397A (en) | Point cloud registration method based on improved FPFH-ICP | |
| CN112861983A (en) | Image matching method, image matching device, electronic equipment and storage medium | |
| CN115620261A (en) | Vehicle environment sensing method, system, equipment and medium based on multiple sensors | |
| CN115063555A (en) | Method for extracting vehicle-mounted LiDAR point cloud street tree growing in Gaussian distribution area | |
| CN116258857A (en) | Outdoor tree-oriented laser point cloud segmentation and extraction method | |
| CN114186588A (en) | Real-time hierarchical filtering method | |
| Cai et al. | A lightweight feature map creation method for intelligent vehicle localization in urban road environments | |
| CN112581511A (en) | Three-dimensional reconstruction method and system based on approximate vertical scanning point cloud rapid registration | |
| CN112070787A (en) | Aviation three-dimensional point cloud plane segmentation method based on opponent reasoning theory | |
| Wang et al. | Particle filter vehicles tracking by fusing multiple features | |
| Zhang et al. | A road extraction method based on high resolution remote sensing image | |
| Wang et al. | A 64-line Lidar-based road obstacle sensing algorithm for intelligent vehicles | |
| CN114022526A (en) | SAC-IA point cloud registration method based on three-dimensional shape context | |
| CN114463396A (en) | Point cloud registration method using plane shape and topological graph voting | |
| CN112989946A (en) | Method, device and equipment for determining lane line and vehicle | |
| CN117576172B (en) | Registration method and device based on improved key points | |
| Ren et al. | A robust ground segmentation method for vehicle lidar | |
| Yu et al. | Registration method for point clouds of complex rock mass based on dual structure information | |
| CN117541614B (en) | Space non-cooperative target close-range relative pose tracking method based on improved ICP algorithm | |
| CN117788735A (en) | Dynamic point cloud removing method based on grid division |
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 |