A method for stitching point cloud data from a single laser radar

Abstract

The present invention discloses a point cloud data splicing method of a single laser radar, which belongs to the technical field of point cloud data splicing, and is used for point cloud data splicing, wherein a plurality of feature points in the point cloud data are selected as reference points, a dynamic neighborhood is determined by feature information of the reference points, and the projection distance of points other than the reference points in the dynamic neighborhood to the reference points in a specific direction is obtained, and the projection distance is counted in the form of a cumulative histogram as a local feature of the reference point; the chi-square distance between the cumulative histograms of two groups of feature points is calculated, and the matrix transformation between the source point cloud data and the target point cloud data is calculated by the best matching pair, and the splicing of the source point cloud data and the target point cloud data is completed by the matrix transformation. The present invention accurately splices point cloud data at different angles to eliminate blind spots and increase the point cloud density, and can also ensure the rapidity and efficiency of the algorithm, and can meet the real-time requirements of point cloud data processing.

Description

Technical Field

[0001] The invention discloses a point cloud data splicing method of a single laser radar, belonging to the technical field of point cloud data splicing. Background Art

[0002] In many application scenarios, a single LiDAR scan cannot cover the entire scene, resulting in data blind spots, or the density of the acquired point cloud data does not meet the requirements due to the influence of parameters such as the resolution of the LiDAR itself. Therefore, the splicing and fusion of point cloud data has become an important research direction. To solve the above problems, existing technologies usually use multi-sensor data fusion to fuse point cloud data, thereby eliminating data blind spots and increasing point cloud data density, but this will increase costs and increase the complexity of the system. Summary of the invention

[0003] The purpose of the present invention is to provide a point cloud data stitching method of a single laser radar to solve the problem in the prior art that a single laser radar scan cannot provide all-round coverage, resulting in data stitching, and data stitching relies on multiple sensors, resulting in increased costs and complexity.

[0004] A single laser radar point cloud data stitching method includes acquiring point cloud data and filtering and denoising, the point cloud data includes source point cloud data and target point cloud data, selecting multiple feature points in the point cloud data as reference points, determining a dynamic neighborhood through feature information of the reference points, obtaining projection distances of points other than the reference points in the dynamic neighborhood to the reference points in a specific direction, and counting the projection distances in the form of a cumulative histogram as local features of the reference points;

[0005] The chi-square distance between the cumulative histograms of the two groups of feature points is calculated. When the chi-square distance is the smallest, it is considered to be the best matching pair. The two groups of feature points are retained as the best matching pairs and the matching pairs with chi-square distance greater than the threshold are eliminated. The matrix transformation between the source point cloud data and the target point cloud data is calculated through the best matching pairs, and the splicing of the source point cloud data and the target point cloud data is completed through the matrix transformation.

[0006] The feature information includes local average curvature, point cloud density and curvature-based wavelet transform detail coefficients.

[0007] Filter denoising includes the first Points , select distance Recent points as a set of neighborhood points, calculate the point arrive The mean distance of the neighboring points :

[0008] ;

[0009] In the formula, for To other points in the neighborhood point set Distance:

[0010] ;

[0011] The average distance from all points in the point cloud data to the mean of the neighboring points is :

[0012] ;

[0013] The standard deviation is:

[0014] ;

[0015] in, is the number of points in the point cloud data;

[0016] Let the standard deviation be , if the average distance between points satisfy , keep the point, otherwise remove the point.

[0017] Select multiple feature points in the point cloud data, including calculating the minimum bounding box that surrounds the point cloud data based on the point cloud data, and finding the maximum and minimum values ​​of the bounding box in the X, Y, and Z directions , , , , , , the side length of the minimum bounding box is , , :

[0018] ;

[0019] Divide the side length of the minimum bounding box into equal parts in the X, Y, and Z directions , , The minimum bounding box is divided into × × small grids, each of which has a side length of , , :

[0020] ;

[0021] Calculate the centroid of the point cloud data in each small grid :

[0022] ;

[0023] In the formula, It is the first Point cloud data, is the total number of point cloud data in the small grid, and the point closest to the center of gravity in each small grid is used as the feature point .

[0024] The local mean curvature and point cloud density include calculating the local mean curvature and point cloud density in each small grid. The maximum principal direction of the grid is used to construct the covariance matrix through other points in the grid. :

[0025] ;

[0026] right Perform eigenvalue decomposition to obtain eigenvalues , , , > > , the eigenvalue corresponding to the eigenvector is , , , we can conclude The maximum main direction , through the small grid The covariance matrix of the point cloud data calculates the eigenvalues ​​of other points , , , > > , the first The curvature of the point cloud data for:

[0027] ;

[0028] The order of the input curvatures is determined by the order of the maximum principal direction of each point in the grid;

[0029] The average curvature of the points in the small grid is :

[0030] ;

[0031] Point cloud density within a small grid for:

[0032] .

[0033] The curvature-based wavelet transform detail coefficients include: As input signal Perform a wavelet transform:

[0034] ;

[0035] Select DB4 as the wavelet basis. Discrete wavelet transform is performed. DB4 wavelet transform includes two filters: a low-pass filter for extracting the low-frequency part of the signal and a high-pass filter for extracting the high-frequency part of the signal. The input curvature signal is subjected to low-pass approximation filtering and high-pass detail filtering using the DB4 wavelet filter. The approximation coefficient is obtained by low-pass filtering. , high-pass filtering to obtain detail coefficients , and then downsample these two coefficients, taking one data point for each two samples. After the first layer of transformation, the curvature signal for:

[0036] ;

[0037] After the first layer transformation is completed, Perform wavelet transform again to get the approximation coefficient of the second layer and detail factor :

[0038] ;

[0039] Repeat the transformation until the preset number of decomposition levels is reached , the input curvature signal is decomposed into approximation coefficients and detail factor :

[0040] ;

[0041] Take out the detail coefficient part, calculate the sum of the squares of this part of the detail coefficient, and get the wavelet transform detail coefficient based on curvature :

[0042] ;

[0043] In the formula, Indicates Layer transformation.

[0044] Determining the dynamic neighborhood includes:

[0045] ;

[0046] In the formula, is a dynamic neighborhood, is a proportional coefficient used to control the overall scale of the neighborhood, which is determined by the average distance between points in the neighborhood. It is The local point cloud density of the point cloud data.

[0047] Local features include Calculate feature points of The range of Point cloud data arrive Projection distance :

[0048] ;

[0049] After calculating all the points After that, draw a cumulative histogram and separate the minimum and maximum projection distances As the starting point and the end point respectively, the middle is divided into multiple equally spaced intervals, and the number of points in each distance interval with the cumulative projection distance is used to form a Cumulative histogram of projection distance for reference:

[0050] ;

[0051] In the formula, For the The interval corresponding to the point cloud data The number of internal points, Indicates count.

[0052] Chi-square distance include:

[0053] ;

[0054] In the formula, The source point cloud data Point cloud data , The target point cloud data Point cloud data .

[0055] The matrix transformation includes calculating the centroid of the source point cloud data and the target point cloud data:

[0056] ;

[0057] ;

[0058] In the formula, is the centroid of the source point cloud data, It is Source point cloud data, is the centroid of the target point cloud data, It is Target point cloud data, translate each matching pair in the two sets of data from the centroid so that the centroid is aligned to the origin:

[0059] ;

[0060] ;

[0061] In the formula, and After translation and ;

[0062] After translation and The centroids are all located at the origin, and the covariance matrix is ​​calculated :

[0063] ;

[0064] right Perform singular value decomposition:

[0065] ;

[0066] In the formula, and is an orthogonal matrix, is a diagonal matrix containing singular values, the rotation matrix for:

[0067] ;

[0068] Calculate the translation vector :

[0069] ;

[0070] Transform the source point cloud data into the target point cloud data:

[0071] ;

[0072] Complete the stitching of point cloud data.

[0073] Compared with the prior art, the present invention has the following beneficial effects: the present invention accurately splices point cloud data at different angles to eliminate blind spots and increase point cloud density, and can also ensure the fast and efficient operation of the algorithm, thereby meeting the real-time requirements of point cloud data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1This is the schematic diagram of the splicing of Experiment 1;

[0075] Figure 2 The source point cloud data distribution and target point cloud data volume distribution diagram of Experiment 1;

[0076] Figure 3 This is the splicing schematic diagram of Experiment 2;

[0077] Figure 4 The source point cloud data distribution and target point cloud data volume distribution diagram of Experiment 2;

[0078] Figure 5 This is the schematic diagram of the splicing of Experiment 3;

[0079] Figure 6 The source point cloud data distribution and target point cloud data volume distribution diagram of Experiment 3;

[0080] Figure 7 This is the schematic diagram of the splicing of Experiment 4;

[0081] Figure 8 The source point cloud data distribution and target point cloud data volume distribution diagram of Experiment 4;

[0082] Fig. 9 This is the schematic diagram of the splicing of Experiment 5;

[0083] Fig.10 The source point cloud data distribution and target point cloud data volume distribution diagram of Experiment 5;

[0084] Fig.11 This is the filtering and denoising effect diagram of Experiment 3;

[0085] Fig.12 This is the filtering and denoising effect diagram of Experiment 4. DETAILED DESCRIPTION

[0086] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0087] A single laser radar point cloud data stitching method includes acquiring point cloud data and filtering and denoising, the point cloud data includes source point cloud data and target point cloud data, selecting multiple feature points in the point cloud data as reference points, determining a dynamic neighborhood through feature information of the reference points, obtaining projection distances of points other than the reference points in the dynamic neighborhood to the reference points in a specific direction, and counting the projection distances in the form of a cumulative histogram as local features of the reference points;

[0088] The chi-square distance between the cumulative histograms of the two groups of feature points is calculated. When the chi-square distance is the smallest, it is considered to be the best matching pair. The two groups of feature points are retained as the best matching pairs and the matching pairs with chi-square distance greater than the threshold are eliminated. The matrix transformation between the source point cloud data and the target point cloud data is calculated through the best matching pairs, and the splicing of the source point cloud data and the target point cloud data is completed through the matrix transformation.

[0089] The feature information includes local average curvature, point cloud density and curvature-based wavelet transform detail coefficients.

[0090] Filter denoising includes the first Points , select distance Recent points as a set of neighborhood points, calculate the point arrive The mean distance of the neighboring points :

[0091] ;

[0092] In the formula, for To other points in the neighborhood point set Distance:

[0093] ;

[0094] The average distance from all points in the point cloud data to the mean of the neighboring points is :

[0095] ;

[0096] The standard deviation is:

[0097] ;

[0098] in, is the number of points in the point cloud data;

[0099] Let the standard deviation be , if the average distance between points satisfy , keep the point, otherwise remove the point.

[0100] Select multiple feature points in the point cloud data, including calculating the minimum bounding box that surrounds the point cloud data based on the point cloud data, and finding the maximum and minimum values ​​of the bounding box in the X, Y, and Z directions , , , , , , the side length of the minimum bounding box is , , :

[0101] ;

[0102] Divide the side length of the minimum bounding box into equal parts in the X, Y, and Z directions , , The minimum bounding box is divided into × × small grids, each of which has a side length of , , :

[0103] ;

[0104] Calculate the centroid of the point cloud data in each small grid :

[0105] ;

[0106] In the formula, It is the first Point cloud data, is the total number of point cloud data in the small grid, and the point closest to the center of gravity in each small grid is used as the feature point .

[0107] The local mean curvature and point cloud density include calculating the local mean curvature and point cloud density in each small grid. The maximum principal direction of the grid is used to construct the covariance matrix through other points in the grid. :

[0108] ;

[0109] right Perform eigenvalue decomposition to obtain eigenvalues , , , > > , the eigenvalue corresponding to the eigenvector is , , , we can conclude The maximum main direction , through the small grid The covariance matrix of the point cloud data calculates the eigenvalues ​​of other points , , , > > , the first The curvature of the point cloud data for:

[0110] ;

[0111] The order of the input curvatures is determined by the order of the maximum principal direction of each point in the grid;

[0112] The average curvature of the points in the small grid is :

[0113] ;

[0114] Point cloud density within a small grid for:

[0115] .

[0116] The curvature-based wavelet transform detail coefficients include: As input signal Perform a wavelet transform:

[0117] ;

[0118] Select DB4 as the wavelet basis. Discrete wavelet transform is performed. DB4 wavelet transform includes two filters: a low-pass filter for extracting the low-frequency part of the signal and a high-pass filter for extracting the high-frequency part of the signal. The input curvature signal is subjected to low-pass approximation filtering and high-pass detail filtering using the DB4 wavelet filter. The approximation coefficient is obtained by low-pass filtering. , high-pass filtering to obtain detail coefficients , and then downsample these two coefficients, taking one data point for each two samples. After the first layer of transformation, the curvature signal for:

[0119] ;

[0120] After the first layer transformation is completed, Perform wavelet transform again to get the approximation coefficient of the second layer and detail factor :

[0121] ;

[0122] Repeat the transformation until the preset number of decomposition levels is reached , the input curvature signal is decomposed into approximation coefficients and detail factor :

[0123] ;

[0124] Take out the detail coefficient part, calculate the sum of the squares of this part of the detail coefficient, and get the wavelet transform detail coefficient based on curvature :

[0125] ;

[0126] In the formula, Indicates Layer transformation.

[0127] Determining the dynamic neighborhood includes:

[0128] ;

[0129] In the formula, is a dynamic neighborhood, is a proportional coefficient used to control the overall scale of the neighborhood, which is determined by the average distance between points in the neighborhood. It is The local point cloud density of the point cloud data.

[0130] Local features include Calculate feature points of The range of Point cloud data arrive Projection distance :

[0131] ;

[0132] After calculating all the points After that, draw a cumulative histogram and separate the minimum and maximum projection distances As the starting point and the end point respectively, the middle is divided into multiple equally spaced intervals, and the number of points in each distance interval with the cumulative projection distance is used to form a Cumulative histogram of projection distance for reference:

[0133] ;

[0134] In the formula, For the The interval corresponding to the point cloud data The number of internal points, Indicates count.

[0135] Chi-square distance include:

[0136] ;

[0137] In the formula, The source point cloud data Point cloud data , The target point cloud data Point cloud data .

[0138] The matrix transformation includes calculating the centroid of the source point cloud data and the target point cloud data:

[0139] ;

[0140] ;

[0141] In the formula, is the centroid of the source point cloud data, It is Source point cloud data, is the centroid of the target point cloud data, It is Target point cloud data, translate each matching pair in the two sets of data from the centroid so that the centroid is aligned to the origin:

[0142] ;

[0143] ;

[0144] In the formula, and After translation and ;

[0145] After translation and The centroids are all located at the origin, and the covariance matrix is ​​calculated :

[0146] ;

[0147] right Perform singular value decomposition:

[0148] ;

[0149] In the formula, and is an orthogonal matrix, is a diagonal matrix containing singular values, the rotation matrix for:

[0150] ;

[0151] Calculate the translation vector :

[0152] ;

[0153] Transform the source point cloud data into the target point cloud data:

[0154] ;

[0155] Complete the stitching of point cloud data.

[0156] The experimental point cloud data of the present invention includes the open source Dragon data set and the point cloud data collected by the laser radar. It includes indoor point cloud data, storage box point cloud data and chair point cloud data. Experiment 1 is indoor point cloud data, such as Figure 1 As shown in the figure, point cloud data 1 and 2 are indoor point cloud data scanned at different angles (some point cloud data in the two sets of data are different). After splicing, more complete point cloud data can be obtained, where the red data is the original point cloud data and the green data is the target point cloud data. The feature point selection process is as follows Figure 2 As shown, the source point cloud data volume is 112586 (1178 feature points), and the target point cloud data volume is 494827 (1307 feature points).

[0157] Experiment 2 is from the open source Dragon dataset, such as Figure 3 As shown in the figure, the red data is the source point cloud data, and the green data is the target point cloud data. The two sets of point cloud data are the same, but the position angles are different. After splicing, completely overlapping point cloud data can be obtained. The feature point selection process is as follows: Figure 4 As shown, the source point cloud data volume is 29103 (1914 feature points), and the target point cloud data volume is 29103 (1571 feature points).

[0158] Experiment 3 uses LiDAR to scan a storage box to obtain point cloud data, such as Figure 5 As shown in the figure, point cloud data 1 is obtained from one perspective (with occluded parts), and point cloud data 2 is scanned from another perspective (the occluded parts of point cloud data 1 can be scanned). After the splicing algorithm, the complete storage box point cloud data can be obtained. The feature point selection process is as follows Figure 6 As shown, the source point cloud data volume is 115768 (618 feature points), and the target point cloud data volume is 494827 (1002 feature points).

[0159] Experiment 4 is to use LiDAR to scan a chair to obtain point cloud data, such as Figure 7As shown in the figure, point cloud data 1 is obtained from one perspective (with occluded parts, incomplete), and point cloud data 2 is relatively complete point cloud data obtained by scanning. After the splicing algorithm, the complete storage box point cloud data can be obtained. The feature point selection process is as follows Figure 8 As shown, the source point cloud data volume is 76099 (790 feature points), and the target point cloud data volume is 65433 (1605 feature points).

[0160] Experiment 5 uses LiDAR to scan the excavator bucket to obtain point cloud data, such as Fig. 9 As shown in the figure, point cloud data 1 is the complete point cloud data of an empty bucket, and point cloud data 2 is the point cloud data of an excavator bucket full of soil scanned at a certain angle (only part of the point cloud data, the rest of the data is blocked due to the single perspective). The complete point cloud information of the excavator bucket full of soil can be obtained through the stitching algorithm. The feature point selection process is as follows Fig.10 As shown, the source point cloud data volume is 728493 (999 feature points), and the target point cloud data volume is 16682 (845 feature points).

[0161] The present invention also provides the filtering and denoising effects of Experiments 3 and 4. The filtering and denoising effect of Experiment 3 is as follows: Fig.11 As shown in Figure 4, the filtering and denoising effect of Experiment 4 is as follows Fig.12 As shown in the figure, by comparison, it can be seen that the filtering preprocessing makes the edge detail information of the point cloud data smoother, which not only eliminates a large number of noise points and outliers, but also retains the detailed features of the point cloud data, and improves the subsequent point cloud data matching accuracy.

[0162] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features may be replaced by equivalents, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for stitching point cloud data of a single laser radar, characterized in that: The method includes acquiring point cloud data and performing filtering and denoising, wherein the point cloud data includes source point cloud data and target point cloud data, selecting a plurality of feature points in the point cloud data as reference points, determining a dynamic neighborhood through feature information of the reference points, obtaining projection distances of points other than the reference points in the dynamic neighborhood to the reference points in a specific direction, and using a cumulative histogram to count the projection distances as local features of the reference points; The chi-square distance between the cumulative histograms of the two groups of feature points is calculated. When the chi-square distance is the smallest, it is considered to be the best matching pair. The two groups of feature points are retained as the best matching pairs and the matching pairs with chi-square distance greater than the threshold are eliminated. The matrix transformation between the source point cloud data and the target point cloud data is calculated through the best matching pairs, and the splicing of the source point cloud data and the target point cloud data is completed through the matrix transformation.

2. The point cloud data stitching method of a single laser radar according to claim 1, characterized in that: The feature information includes local average curvature, point cloud density and curvature-based wavelet transform detail coefficients.

3. The point cloud data stitching method of a single laser radar according to claim 1, characterized in that: Filter denoising includes the first Points , select distance Recent points as a set of neighborhood points, calculate the point arrive The mean distance of the neighboring points : ； In the formula, for To other points in the neighborhood point set Distance: ； The average distance from all points in the point cloud data to the mean of the neighboring points is : ； The standard deviation is: ； in, is the number of points in the point cloud data; Let the standard deviation be , if the average distance between points satisfy , keep the point, otherwise remove the point.

4. The point cloud data stitching method of a single laser radar according to claim 2, characterized in that: Select multiple feature points in the point cloud data, including calculating the minimum bounding box that surrounds the point cloud data based on the point cloud data, and finding the maximum and minimum values ​​of the bounding box in the X, Y, and Z directions , , , , , , the side length of the minimum bounding box is , , : ； Divide the side length of the minimum bounding box into equal parts in the X, Y, and Z directions , , The minimum bounding box is divided into × × small grids, each of which has a side length of , , : ； Calculate the centroid of the point cloud data in each small grid : ； In the formula, It is the first Point cloud data, is the total number of point cloud data in the small grid, and the point closest to the center of gravity in each small grid is used as the feature point .

5. The point cloud data stitching method of a single laser radar according to claim 4, characterized in that: The local mean curvature and point cloud density include calculating the local mean curvature and point cloud density in each small grid. The maximum principal direction of the grid is used to construct the covariance matrix through other points in the grid. : ； right Perform eigenvalue decomposition to obtain eigenvalues , , , > > , the eigenvalue corresponding to the eigenvector is , , , we can conclude The maximum main direction , through the small grid The covariance matrix of the point cloud data calculates the eigenvalues ​​of other points , , , > > , the first The curvature of the point cloud data for: ； The order of the input curvatures is determined by the order of the maximum principal direction of each point in the grid; The average curvature of the points in the small grid is : ； Point cloud density within a small grid for: 。 6. The point cloud data stitching method of a single laser radar according to claim 5, characterized in that: The curvature-based wavelet transform detail coefficients include: As input signal Perform a wavelet transform: ； Select DB4 as the wavelet basis. Discrete wavelet transform is performed. DB4 wavelet transform includes two filters: a low-pass filter for extracting the low-frequency part of the signal and a high-pass filter for extracting the high-frequency part of the signal. The input curvature signal is subjected to low-pass approximation filtering and high-pass detail filtering using the DB4 wavelet filter. The approximation coefficient is obtained by low-pass filtering. , high-pass filtering to obtain detail coefficients , and then downsample these two coefficients, taking one data point for each two samples. After the first layer of transformation, the curvature signal for: ； After the first layer transformation is completed, Perform wavelet transform again to get the approximation coefficient of the second layer and detail factor : ； Repeat the transformation until the preset number of decomposition levels is reached , the input curvature signal is decomposed into approximation coefficients and detail factor : ； Take out the detail coefficient part, calculate the sum of the squares of this part of the detail coefficient, and get the wavelet transform detail coefficient based on curvature : ； In the formula, Indicates Layer transformation.

7. The point cloud data stitching method of a single laser radar according to claim 6, characterized in that: Determining the dynamic neighborhood includes: ； In the formula, is a dynamic neighborhood, is a proportional coefficient used to control the overall scale of the neighborhood, which is determined by the average distance between points in the neighborhood. It is The local point cloud density of the point cloud data.

8. The point cloud data stitching method of a single laser radar according to claim 7, characterized in that: Local features include Calculate feature points of The range of Point cloud data arrive Projection distance : ； After calculating all the points After that, draw a cumulative histogram and separate the minimum and maximum projection distances As the starting point and the end point respectively, the middle is divided into multiple equally spaced intervals, and the number of points in each distance interval with the cumulative projection distance is used to form a Cumulative histogram of projection distance for reference: ； In the formula, For the The interval corresponding to the point cloud data The number of internal points, Indicates count.

9. The point cloud data stitching method of a single laser radar according to claim 8, characterized in that: Chi-square distance include: ； In the formula, The source point cloud data Point cloud data , The target point cloud data Point cloud data .

10. The point cloud data stitching method of a single laser radar according to claim 9, characterized in that: The matrix transformation includes calculating the centroid of the source point cloud data and the target point cloud data: ； ； In the formula, is the centroid of the source point cloud data, It is Source point cloud data, is the centroid of the target point cloud data, It is Target point cloud data, translate each matching pair in the two sets of data from the centroid so that the centroid is aligned to the origin: ； ； In the formula, and After translation and ; After translation and The centroids are all located at the origin, and the covariance matrix is ​​calculated : ； right Perform singular value decomposition: ； In the formula, and is an orthogonal matrix, is a diagonal matrix containing singular values, the rotation matrix for: ； Calculate the translation vector : ； Transform the source point cloud data into the target point cloud data: ； Complete the stitching of point cloud data.

Images