CN114494287A

CN114494287A - Long-distance laser radar point cloud data processing method

Info

Publication number: CN114494287A
Application number: CN202111671088.6A
Authority: CN
Inventors: 陈钱; 丁玲慧; 何伟基; 张闻文
Original assignee: Nanjing University of Science and Technology
Current assignee: Nanjing University of Science and Technology
Priority date: 2021-12-31
Filing date: 2021-12-31
Publication date: 2022-05-13

Abstract

The invention discloses a remote laser radar point cloud data processing method, which comprises the steps of scanning a target environment of data to be acquired through an optical machine scanning system, acquiring the photon flight time of each pixel point, and exporting and extracting the acquired data; calculating filter parameters by using a point cloud data segmentation method to complete parameter self-adaptation; performing dimensionality reduction on the data by applying a principal component analysis technology; and (3) carrying out point cloud clustering by using a density-based spatial clustering Denoising (DBSCAN) algorithm, deleting outliers, namely noise points, and finally upgrading the dimension of the obtained two-dimensional point cloud to three-dimensional point cloud data to complete denoising and filtering of the original point cloud data. The invention effectively reduces the complexity of the algorithm while keeping the environmental characteristics.

Description

Long-distance laser radar point cloud data processing method

Technical Field

The invention belongs to the technical field of laser radars, and particularly relates to a method for processing point cloud data of a long-distance laser radar.

Technical Field

A system in which a laser radar detects a distance, an azimuth, an altitude, a speed, an intensity, and the like of a target using laser light as a carrier wave of a signal. The laser is used for detecting the remote non-cooperative target, and the method has the advantages of strong anti-electromagnetic interference capability, good inhibition effect on ground objects and background noise, strong anti-stealth capability, good distance, angle and speed resolution, capability of realizing synchronous acquisition of distance, speed and three-dimensional image information of the target and the like. The laser radar well utilizes the advantages of laser in this aspect, and can obtain higher imaging precision compared with the traditional microwave radar. With the development of laser radars, the demand for weak signal detection is becoming stronger and stronger. When the detection distance is relatively long or the detection distance is influenced by environmental factors such as rain, fog and dust, the energy of the optical echo signal received by the detector is extremely weak, and usually, the optical echo signal only has the magnitude of a plurality of photon energies. The intensity of weak signals is lower than the detection limit of the traditional photoelectric device, so that the traditional photoelectric detector is difficult to meet the requirements of modern application. Against this background, the advent of lidar based on geiger mode avalanche photodiodes effectively addressed this problem. The Gm-APD works at a bias voltage slightly higher than the critical breakdown voltage of the PN junction of the Gm-APD, and an incoming single photon can enable the Gm-APD to generate an avalanche effect, so that the arrival time of the photon is recorded. Compared with other detectors, the Gm-APD has the advantages of small volume, high integration, high reliability and the like. The laser radar can achieve the detection sensitivity of a single photon level and has the advantages of long distance measurement and high imaging speed. The method is mainly applied to the fields of remote laser ranging, underwater target detection, atmospheric pollution measurement, astronomical observation, biological waveguide detection, weak light wavefront sensing, particle physics, fluorescence medical imaging and the like.

With the application of new probe materials and the development of manufacturing techniques, lidar is also continuously improved and developed. Pulse type photon laser radar, pulse modulation type photon laser radar and chirp modulation type photon laser radar appear in sequence. At present, the main research work of the laser radar is mostly focused on the aspects of system structure design and manufacturing and improvement of a detector, the research results of performance analysis and later data processing of the laser radar system are relatively rare, and the research on data processing methods in the later period of the system is relatively less. The photon counting laser radar adopts Gm-APD as a detector, and the laser radar is very easily interfered by noise due to the extremely high sensitivity of the laser radar. The research on the later data processing method of the laser radar can enlarge the application range of the laser radar and obtain more accurate imaging effect.

In the existing point cloud filtering algorithm, a Statistical Outlier Removal (SOR) filter distinguishes signals and noise by calculating the average distance between each point and the nearest K points; the Spatial Frequency (SF) outlier filter constructs a sphere with each point as the center, and removes noise points according to the number of points in the sphere. The statistical outlier removal filter described above cannot identify outlier groups, while the radius outlier removal filter may erroneously remove a large number of useful environmental features. Wang et al propose an adaptive ellipsoid search filter based on an existing radius filter, which takes the center of an ellipsoid as a target point to be detected and identifies noise according to the number of neighboring points of the ellipsoid. However, the algorithm is too complex and the real-time performance is not ideal enough.

Disclosure of Invention

Aiming at the problems, the invention provides a remote laser radar point cloud data processing method, which aims to reduce the dimension based on a Principal Component Analysis (PCA) technology and reduce the noise based on a density space clustering Denoising (DBSCAN) algorithm, and can effectively reduce the algorithm complexity while retaining the environmental characteristics.

The technical scheme for realizing the purpose of the invention is as follows: a long-distance laser radar point cloud data processing method comprises the following steps:

scanning a target environment of data to be acquired by using a remote laser radar point cloud acquisition system, acquiring the photon flight time of each pixel point, and exporting and extracting the acquired data;

carrying out region segmentation on the collected data by applying a point cloud data segmentation method, and calculating a clustering parameter corresponding to each region data;

performing principal component analysis technology on the data of each data block to reduce the dimension of the data;

for the data of each data block, performing point cloud clustering by using a density-based spatial clustering denoising algorithm, deleting outliers, and increasing the dimension of the obtained two-dimensional point cloud to three-dimensional point cloud data;

and splicing the new three-dimensional data obtained by processing each region to obtain complete new point cloud data subjected to filtering and denoising, and completing denoising and filtering of the original point cloud data.

Preferably, the collected point cloud is three-dimensional data composed of the coordinates of a target pixel point and the flight time of the returned photons of the pixel point target.

Preferably, the collected data is segmented by applying point cloud data by a space segmentation method with a cylindrical structure, specifically:

dividing the whole space into a plurality of cylindrical blocks by taking a vertical line of the laser radar sensor as a central axis, wherein the space between every two cylinders is a point cloud dividing area;

in the cylinder model, each region is guaranteed to have the same volume, and the cross-sectional area between every two adjacent cylinders satisfies:

namely, it is

Wherein r is_i(i ═ 1,2, …, t) is the radius of the ith cylinder, r₁The distance from a point on the cylinder to the center of the sensor; h is the height of the cylinder and t is the total number of regions into which the region is divided.

Preferably, the specific process of applying the point cloud data to the collected data to segment by using the spatial segmentation method with the cylindrical structure is as follows:

step 2.1: determining the total area number t of the area segmentation according to the conditions required to be met by the boundary radius of each area, and calculating the radius r of the boundary of the nearest area to the detector₁，r₁From the point furthest from the sensor determines:

step 2.2: according to when the i-th cylinder cross-sectional radius is the first cylinder cross-sectional radius

Double instant

The volumes of all the areas are equal, the boundary radius of each area is calculated, and the area segmentation of the detection space is completed;

step 2.3: dividing the collected original point cloud data into a plurality of data blocks according to a space area equal-volume division principle: calculating the distance between the space point cloud and a vertical line with the laser radar sensor as a central axis by using the flight time of the point cloud in the space: if the distance is between r_i-1And r_iThe point cloud is classified in the ith group of data block;

step 2.4: when denoising and filtering are carried out on the point cloud data of each region, the domain radius of each region is defined as:

Eps_i＝kr_i

wherein Eps_i(i-1, 2, …, t) represents the area radius of the i-th area, r_iIndicating the distance between the point cloud of the area i and the detector, due to

Therefore, the method comprises the following steps:

preferably, the specific method for performing principal component analysis on the data of each data block to perform dimensionality reduction on the data is as follows:

step 3.1: arranging the three-dimensional data points into a matrix of 3 rows and m columns, m being the number of three-dimensional data points, and subtracting the mean value

To ensure the resulting matrixXi each row is centered on zero; the covariance matrix C is obtained by:

step 3.2: arranging the three feature vectors into E ═ E (E)₁,e₂,e₃) The matrix E satisfies E^TCE ═ Λ, where Λ is the diagonal matrix;

obtaining three eigenvectors and corresponding eigenvalues thereof, arranging the eigenvectors into a matrix according to rows of the eigenvalues, taking the first two eigenvectors to generate a matrix P, and constructing a two-dimensional matrix according to the formula:

Y_2×m＝P_2×3×ξ_3×m。

preferably, for the data of each data block, a density-based spatial clustering denoising algorithm is applied to perform point cloud clustering, and a specific method for deleting outliers is as follows:

step 4.1: for each non-access point p in the dataset_iDetermine all its points in the Eps field using the RANGE _ QUERY function and collect these points into the subset N_iWherein the subset N_iThe number of points in is | N_i|；

Step 4.2: if | N_iIf | is less than min pts, then p_iLabeling as outliers; | N_i| is more than or equal to min pts, and the subset N_iAdding all points in (S ← N) to the set S_i) In, minpts is a density threshold parameter;

step 4.3: for subset N_iAll points in the Eps field of the points which are not visited are collected into a subset N_jIn, if | N_j| is more than or equal to min pts, and the subset N_jAdding all points in the set S (S ← S ≧ N ^ N)_i) In (1).

Step 4.4: repeating steps 4.1 to 4.3 until each point in the data is traversed, the data set being divided into k clusters S₁,S₂,…,S_kTherein each cluster S_iThe number of points in (i ═ 1,2, …, k) is | S_iBy mixing | S_iCompare | with a threshold ψ if | S_i|＞Psi, then S_iIs a signal cluster; otherwise it is identified as a noise cluster and removed.

Preferably, the specific formula for raising the obtained two-dimensional point cloud to three-dimensional point cloud data is as follows:

wherein Y' is a point cloud data set obtained after filtering in a two-dimensional space,

is the average value subtracted from the original matrix, and P is the matrix generated by the first two eigenvectors.

Compared with the prior art, the invention has the following remarkable advantages: the method firstly adopts the PCA technology to carry out dimensionality reduction on the original point cloud data, generates two-dimensional data by extracting a first principal component and a second principal component of the original data, has small information loss, can obviously reduce the overall complexity because the principal signal processing is executed on the two-dimensional data, and simultaneously effectively removes noise and retains environmental characteristics; in addition, the invention also provides a method for carrying out self-adaptive clustering on the two-dimensional point cloud before the two-dimensional point cloud is restored to the three-dimensional point cloud, and the accuracy rate of restoring the target point cloud image can be obviously improved by carrying out self-adaptive parameter setting on the two-dimensional space of the two main components; meanwhile, in order to solve the near-far effect caused by the distance difference between the viewpoint and the scanning point, the invention adopts a point cloud data segmentation method to process the point cloud data with different area characteristics in a partition manner, thereby obtaining better noise reduction effect.

Drawings

Fig. 1 is a schematic overall framework diagram of a method for processing remote lidar point cloud data according to the present invention.

Fig. 2 is a schematic diagram of a long-distance lidar point cloud acquisition system according to the present invention.

FIG. 3 is a schematic diagram of the principle of the density-based spatial clustering Denoising (DBSCAN) algorithm provided by the present invention.

FIG. 4 is a flow chart of the density-based spatial clustering Denoising (DBSCAN) algorithm proposed by the present invention.

Fig. 5 is a schematic diagram of a method for segmenting point cloud data of a long-distance lidar according to the present invention.

Detailed Description

The following detailed description of the preferred embodiments of the present invention, taken in conjunction with the accompanying drawings, will make the advantages and features of the invention easier to understand by those skilled in the art, and will thus make the scope of the invention more clearly and clearly defined.

As shown in fig. 1 to 5, a method for processing point cloud data of a long-distance lidar according to an embodiment of the present invention includes:

step 1: and scanning a target environment of data to be acquired by using a remote laser radar point cloud acquisition system, acquiring the flight time of each pixel point, and exporting and extracting the acquired data. The collected point cloud is three-dimensional data formed by the coordinates of a target pixel point and the flight time of return photons of the pixel point target.

Fig. 2 is a schematic diagram of a long-distance lidar point cloud acquisition system according to the present invention. The functional distinction is mainly divided into six units: the device comprises an optical transmitting unit, a beam splitting and expanding unit, a scanning unit, an optical receiving unit and a control unit. The laser comprises two output ends, namely a standard output end and a signal monitoring end, the photoelectric detector receives a signal of the monitoring end and inputs the signal into a 0 channel of TCSPC, and the TCSPC records the signal and marks the signal as starting time. The pulse laser output by the standard output end is collimated by the collimating mirror and then enters the optical beam expanding antenna through a small hole in the middle of the spectroscope, the pulse laser is further focused and shaped by the optical beam expanding antenna and then is incident on the reflecting surface of the quick reflecting mirror, the pulse laser reaches the surface of a target after being reflected by the mirror surface, and the quick reflecting mirror changes a pulse emission light path through biaxial shimmy so as to realize the scanning of the target. The target diffusely reflects the pulse laser back to the original optical path, the echo pulse is reflected by the reflecting surface of the light splitting unit after passing through the fast reflector and the optical beam expanding antenna, and is coupled into the single photon detector through the optical fiber after sequentially passing through the reflector, the filter and the focusing lens. Incident photons trigger the single photon detector to cause an avalanche effect, a response signal is output after quantization and input into the 1 channel of TCSPC, and the TCSPC records the 1 channel photon response time and marks the response time as the cutoff time. The FPGA controller reads data of a channel 0 and a channel 1 of the TCSPC, the upper computer calculates time difference of the two channels to obtain photon flight time data, and target original point cloud data are obtained after subsequent processing.

Step 2: since the point cloud density is strongly correlated with the point cloud spatial distribution, it depends on the distance from the lidar sensor. Thus, the further away from the sensor, the faster the signal obtained by clustering, and the larger the corresponding clustering parameter should be. Therefore, before processing the point cloud data, the point cloud data needs to be subjected to region segmentation, and a clustering parameter corresponding to each region data is calculated. The invention provides a space segmentation method of a cylindrical structure, which is characterized in that the whole space is segmented into a plurality of cylindrical blocks by taking the vertical line of a laser radar sensor as a central axis as shown in figure 5 (left), and the space between every two cylinders is the segmentation area of point cloud. In the cylinder model, to ensure that each region has the same volume, the cross-sectional area (shown in fig. 5 (right)) between each two adjacent cylinders should satisfy:

wherein r is_i(i ═ 1,2, …, t) is the radius of the ith cylinder, i.e. the distance of a point on the cylinder from the center of the sensor; h is the height of the cylinder.

Step 2.1: determining the total area number t of the area segmentation according to the condition that the boundary radius of each area needs to be met by the space segmentation method of the cylindrical structure, and calculating the radius r of the boundary of the area closest to the detector₁，r₁From the point furthest from the sensor determines:

Double instant

And the volumes of all the areas are equal, the boundary radius of each area is calculated, and the area segmentation of the detection space is completed.

Step 2.3: dividing the collected original point cloud data into a plurality of data blocks according to a space area equal-volume division principle: calculating the distance between the space point cloud and a vertical line with the laser radar sensor as a central axis by using the flight time of the point cloud in the space: if the distance is between r_i-1And r_iAnd the point cloud is classified in the ith group of data block.

Step 2.4: when denoising and filtering are carried out on the point cloud data of each region, Eps of each region is defined as:

Eps_i＝kr_i

wherein Eps_i(i-1, 2, …, t) represents the area radius of the i-th area, r_iRepresenting the distance between the point cloud of region i and the detector. Due to the fact that

Therefore, the method comprises the following steps:

and 3, step 3: before using a two-dimensional density-based spatial clustering (DBSCAN) algorithm to cluster point clouds, raw three-dimensional data needs to be preprocessed. After the point cloud data segmentation is completed, a principal component analysis technique is performed on the data of each data block to extract principal components and convert the three-dimensional data into two-dimensional data. The dimensionality reduction technology based on principal component analysis adopted by the invention is a component technology capable of extracting main features of data, and the corresponding feature vector of the minimum feature value is usually related to noise. By extracting the first principal component and the second principal component of the original data and discarding the feature vector related to noise to generate two-dimensional data, the complexity of the algorithm can be reduced while the reconstructed target feature is retained.

Step 3.1: assuming that the number of three-dimensional data points is m, the data points are first arranged in a matrix of 3 rows and m columns. And by subtracting the mean value

To ensure that each row of the resulting matrix ξ is centered at zero, i.e., the average value of each row is zero. The covariance matrix C is then obtained by:

step 3.2: arranging the three feature vectors into E ═ E (E)₁,e₂,e₃) The matrix E needs to satisfy E^TCE ═ Λ, where Λ is the diagonal matrix. The three eigenvectors and their corresponding eigenvalues can be obtained from the above equation. Since the variance contribution rate of the first principal component and the second principal component exceeds 95%, the third principal component contains less information and is deleted, and the first two eigenvectors are reserved. The eigenvectors are arranged into a matrix according to the rows of the eigenvalues, and the first two eigenvectors are taken to generate a matrix P. Constructing a two-dimensional matrix by operating the following equation:

Y_2×m＝P_2×3×ξ_3×m

and 4, step 4: based on the long-distance laser radar point cloud obtaining method and the dimensionality reduction theory, the invention provides a density-based spatial clustering Denoising (DBSCAN) algorithm for filtering two-dimensional point cloud on a plane. Assuming that the point cloud data sample set is D, D ═ x₁,x₂,…,x_n) For sample x_jBelongs to D, and the E domain set of the E domain set is the sum of x in the sample set D_jSet of subsamples with a distance not greater than E, i.e. N_E(x_j)＝{x_i∈D|d(x_i,x_j) More than or equal to E }, the number of the subsample sets is | N_E(x_j)|：

For sample x_je.D if the number of the sample sets contained in the E field set is not less than P_minI.e. | N_E(x_j)|≥P_minThen x_jIs a core object;

for sample x_je.D if the number of sample sets contained in E field set is less than P_minI.e. | N_E(x_j)|＜P_minThen x_jIs a boundary;

for sample x_je.D if the number of sample sets contained in E field set is less than P_minAnd N is_E(x_j) Without a core object, then x_jAre isolated points.

Fig. 3 is a schematic diagram of the DBSCAN algorithm, in which point a is a core point, point B, C is a boundary point, and N is an isolated point. The core points and the boundary points together form signal points, and the isolated points form outliers. The DBSCAN algorithm can find all dense areas of sample points and treat these dense areas as clusters. By counting the number of point clouds within each cluster, noise can be identified by comparison to a threshold. Identifying point cloud clusters having less than a threshold as signal clusters; otherwise it is identified as a noise cluster.

The original three-dimensional point cloud is represented by a set of unordered points:

ξ＝{(x_i,y_i,z_i)^T,i＝1,2,…,m}

wherein (x)_i,y_i,z_i)^T∈R³Coordinate vector, m is the total number of points in the point cloud.

According to the principal component analysis-based dimensionality reduction technology provided by the invention, the original three-dimensional point cloud data can be converted into two-dimensional data. Each two-dimensional point cloud is composed of a first principal component and a second principal component, denoted by f and s, respectively. The point cloud input into the DBSCAN cluster can thus be represented as:

Y＝{(f_i,s_i)^T,i＝1,2,…,m}

wherein (f)_i,s_i)^T∈R³Is a coordinate vector in a two-dimensional space made up of two principal components. Fig. 4 is a flow chart of the density-based spatial clustering (DBSCAN) algorithm proposed by the present invention. Data meansThere is a data set consisting of a first principal component f and a second principal component s. In a two-dimensional plane, the calculation formula of the distance can be expressed as:

step 4.1: for each non-access point p in the dataset_iDetermine all its points in the Eps field using the RANGE _ QUERY function and collect these points into the subset N_iWherein the subset N_iThe number of points in is | N_i|。

Step 4.2: if | N_iIf | is less than min pts, then p_iLabeling as outliers; | N_i| is more than or equal to min pts, and the subset N_iAdding all points in (S ← N) to the set S_i) In (1).

Step 4.3: for subset N_iAll points in the Eps field of the points which are not visited are collected into a subset N_jIn (1). If | N_j| is more than or equal to min pts, and the subset N_jAdding all points in the set S (S ← S ≧ N ^ N)_i) In (1).

Step 4.4: repeating steps 4.1 to 4.3 until each point in the data is traversed, the data set being divided into k clusters S₁,S₂,…,S_k}. Wherein each cluster S_iThe number of points in (i ═ 1,2, …, k) is | S_iBy mixing | S_iCompare | with a threshold ψ if | S_iIf | > psi, S_iIs a signal cluster; otherwise it is identified as a noise cluster and removed.

And 5: and carrying out inverse PCA (principal component analysis) upscaling calculation on the new two-dimensional data obtained after denoising and filtering. The two-dimensional point cloud data Y' obtained by filtering and removing noise may be represented as Y { (f)_i,s_i)^TI is 1,2, …, n, which is the filtered coordinate vector in the two-dimensional space, and n is the number of point clouds remaining after filtering. After removing noise, the two-dimensional data needs to be restored to three-dimensional data, which can be obtained by the following formula:

is the average value subtracted from the original matrix.

Step 6: repeating the step 3 to the step 5: and performing the processes of reducing the dimensions of the point cloud data of all the areas by a principal component analysis technology, denoising by a density-based spatial clustering algorithm and performing inverse PAC (programmable automation controller) dimension increasing calculation, and then splicing the new three-dimensional data obtained by processing each area to obtain complete new point cloud data subjected to filtering and denoising.

Claims

1. A long-distance laser radar point cloud data processing method is characterized by comprising the following steps:

2. The method of claim 1, wherein the collected point cloud is three-dimensional data consisting of coordinates of a pixel point of the target and a time of flight of returned photons from the pixel point target.

3. A remote lidar point cloud data processing method according to claim 1, characterized in that the collected data is segmented using a spatial segmentation method of a cylindrical structure, specifically:

namely that

Wherein r is_i(i-1, 2, …, t) is the radius of the ith cylinder, r₁The distance from a point on the cylinder to the center of the sensor; h is the height of the cylinder and t is the total number of regions into which the region is divided.

4. A long-range lidar point cloud data processing method as claimed in claim 3, wherein the specific process of applying the point cloud data to the collected data by the cylindrical structure space segmentation method is as follows:

step 2.1: determining the total area number t of the area segmentation according to the conditions required to be met by the boundary radius of each area, and calculating the radius r of the boundary of the nearest area to the detector₁，r₁Determined by the point furthest from the sensor:

Double instant

Eps_i＝kr_i

Therefore, the method comprises the following steps:

5. the long-distance lidar point cloud data processing method of claim 1, wherein the specific method for performing principal component analysis on the data of each data block to reduce the dimension of the data is as follows:

To ensureEach row of the obtained matrix xi takes zero as the center; the covariance matrix C is obtained by:

step 3.2: arranging three feature vectors into E ═ E (E)₁,e₂,e₃) The matrix E satisfies E^TCE ═ Λ, where Λ is the diagonal matrix;

Y_2×m＝P_2×3×ξ_3×m。

6. the long-distance lidar point cloud data processing method of claim 1, wherein for the data of each data block, a density-based spatial clustering denoising algorithm is applied to perform point cloud clustering, and the specific method for deleting outliers is as follows:

Step 4.2: if | N_iIf | is less than min pts, then p_iLabeling as outliers; | N_i| is more than or equal to min pts, and the subset N_iAdding all the points in (S ← N) to the set S (S ← N)_i) In, minpts is a density threshold parameter;

step 4.3: for subset N_iThe points which are not visited are collected into a subset N_jIn, if | N_j| is more than or equal to min pts, and the subset N_jAdding all points in the set S (S ← S ≧ N ^ N)_i) In (1).

Step 4.4: repeating steps 4.1 to 4.3 until each point in the data is traversed, the data set being divided into k clusters S₁,S₂,…,S_kTherein each cluster S_iThe number of points in (i ═ 1,2, …, k) is | S_iBy mixing | S_iCompare | with a threshold ψ if | S_iIf | > psi, S_iIs a signal cluster; otherwise it is identified as a noise cluster and removed.

7. The method for processing the point cloud data of the long-distance laser radar as claimed in claim 1, wherein a specific formula for increasing the dimension of the obtained two-dimensional point cloud to the three-dimensional point cloud data is as follows:

is the average value subtracted from the original matrix, and P is the first two eigenvector generating matrix.