A laser point cloud registration method combining feature extraction and clustering algorithm

Abstract

The application discloses a laser point cloud registration method combining feature extraction and clustering algorithm, and comprises the following steps: combining the Euclidean distance and the scanning characteristics of a 2D laser radar, performing region segmentation on point cloud data to be registered by using a dynamic threshold; in the divided region, irregular features such as an arc line and a broken line are re-divided, then a RANSAC algorithm is used for linear parameter fitting to extract linear features; the extracted linear features are used to estimate a transformation matrix by using a density clustering algorithm, and coarse registration is performed through rotation and translation; and the point cloud data after coarse registration is subjected to fine registration by using a TrICP (Trimmed Iterative Closest Point) algorithm. The application can effectively solve the shortcoming that a traditional point cloud registration algorithm has a high requirement on an initial position, and can fully improve the accuracy and applicability of point cloud registration.

Description

Technical Field

[0001] This invention belongs to the field of robot autonomous mapping technology, and relates to a point cloud registration method, specifically a laser point cloud registration method that combines feature extraction and clustering algorithms. Background Technology

[0002] Laser point cloud registration is a fundamental step in robot autonomous mapping and global localization using lidar. During map construction, lidar is needed to acquire point cloud data of the surrounding environment. However, due to the limited scanning distance of a single lidar scan and obstacle occlusion, complete point cloud data cannot be obtained in one go. Therefore, point cloud data obtained from multiple consecutive scans at different viewpoints are registered. Essentially, laser point cloud registration involves solving for the rotation matrix and translation vector during the registration process, transforming two or more frames of point cloud data into the same coordinate system.

[0003] The Iterative Closest Point (ICP) algorithm is the most classic and widely used point cloud registration algorithm. However, this algorithm uses iterative transformations to solve the problem during registration, resulting in low registration efficiency. When using this algorithm to register laser point cloud data with large initial position differences, the registration effect is poor, and it is prone to getting trapped in local optima. Therefore, many researchers have attempted to improve the Iterative Closest Point (ICP) registration algorithm to address its various shortcomings. However, when the distance between corresponding point cloud data is large, the registration effect of existing algorithms is still poor. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a laser point cloud registration method that combines feature extraction and clustering algorithms. This method can effectively solve the problem of high requirements for initial position in traditional point cloud registration algorithms and can significantly improve the accuracy and applicability of point cloud registration.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution.

[0006] A laser point cloud registration method combining feature extraction and clustering algorithms includes the following steps:

[0007] Step 1: Use a LiDAR to scan the surrounding environment at an initial position 100cm apart on the horizontal axis and 50cm apart on the vertical axis to obtain the point cloud data to be registered; combine the scanning characteristics of the LiDAR to perform dynamic threshold region segmentation on the point cloud data to be registered.

[0008] Step 2: Within the segmented region, perform feature division, use the random sampling consensus algorithm to fit the line parameters, and extract the line features;

[0009] Step 3: Use density clustering algorithm to estimate the transformation matrix of the extracted straight line features, and perform coarse registration by rotation and translation to obtain the coarsely registered source point cloud data P′;

[0010] Step 4: Perform fine registration on the coarsely registered source point cloud data P′ using the TrICP cropping iterative nearest point algorithm;

[0011] The point cloud data that needs to be registered includes: source point cloud data P = {P i =(X i ,Y i ), i∈[1,2,…n]}, where X i Y i Let x and y be the x and y coordinates of the i-th source point cloud data point, respectively, and n be the number of source point cloud data points; the target point cloud data Q = {Q j =(M j N j ),j=[1,2,…m]}, where M j N j ...

[0012] Specifically, step 1 includes the following steps:

[0013] Step 1.2. Calculate the distance d between each adjacent point cloud data point in both the source and target point cloud data. i d j :

[0014]

[0015]

[0016] Step 1.3. Calculate the dynamic threshold δ for region segmentation in the source point cloud data and the target point cloud data respectively. i δ j It can be seen that the greater the distance between the point cloud data points and the radar transmission point, the smaller the segmentation threshold should be. The specific formula is:

[0017]

[0018]

[0019] Where ρ i and ρ i+1 The laser-radar source point cloud data point P is respectively i and P i+1 The Euclidean distance, γ, is the source point cloud data point P. i and P i+1 The included angle between them; ρ jand ρ j+1 The laser radar reaches the target point cloud data point Q. j and Q j+1 The Euclidean distance, σ is the target point cloud data point Q. j and Q j+1 The included angle between them; in this invention, the value of ξ1 is 5, and ξ1 can be adjusted appropriately when using different models of lidar.

[0020] Further, in step 1: when d i ≥δ i At that time, the source point cloud data point P is considered to be... i Using d as the dividing point, the source point cloud data is divided into regions; when d j ≥δ j When this happens, the target point cloud data point Q is considered to be... j Using the dividing point as the dividing point, the target point cloud data is divided into regions, and finally the source point cloud data and the target point cloud data are each divided into several regions {D}. 1 D 2 ,…,D m The algorithm determines the number of point cloud data points within the segmented region after dividing the source and target point cloud data. If the number of point cloud data points within the region is less than or equal to 3, the region is considered environmental noise and is deleted.

[0021] Specifically, step 2 includes the following steps:

[0022] Step 2.1. In the region D that was segmented in Step 1... i In the middle, select the first point cloud data point k. s (X s Y s ) and the last point cloud data point k n (X n Y n To construct a straight line; where (X) s Y s ), (X n Y n ) are respectively k s k n The coordinates;

[0023] Step 2.2. Calculate region D i The distances from each point in the point cloud data to the line are calculated to obtain a set of all distances. Record the maximum value among them Let its corresponding point be k max (X max Y max );

[0024] Step 2.3. In region Di In the middle, according to endpoint k s and k n The distance between them determines the partition threshold d. T The specific formula is as follows:

[0025]

[0026] Where ρ(<1) is the scaling parameter; the maximum value in the distance set. and the dividing threshold d T Compare; if Then take point cloud data point k max Divide region D into two parts. i Divide into two parts; otherwise, do not divide.

[0027] Following steps 2.1-2.3, the region {D} can be... 1 D 2 ,…,D m The data is further divided to obtain a new point cloud data region {A}. 1 A 2 ,…,A n};

[0028] Step 2.4. Use the random sample consensus algorithm to partition the region A. i Perform line parameter fitting to obtain the line parameters in the source point cloud data. Linear parameters in target point cloud data The straight line extracted from the source point cloud data is The straight line extracted from the target point cloud data is

[0029] Furthermore, the aforementioned random sampling consensus algorithm is used for region A. i The process of fitting line parameters and extracting line features includes:

[0030] 2.4.1. Since any two point cloud data points can determine a straight line y = ax + b, we first proceed sequentially from region A. i Select two points in the equation and solve for the line parameters a and b.

[0031] 2.4.2. Solving for the current region A i The distance from all point cloud data points to the line y = ax + b is calculated. If the distance of a point cloud data point is less than a set threshold, it means that the point cloud data point supports the current parameter. Finally, the number of point cloud data points that support the line parameter is obtained.

[0032] Repeat the above two steps until the selected region A has been traversed.i All point cloud data points are selected, and then the line parameter supported by the most point cloud data points is selected as the final parameter.

[0033] Specifically, in step 3, the extraction of straight-line features is used to estimate the transformation matrix using a density clustering algorithm, and coarse registration is performed through rotation and translation to obtain the coarsely registered source point cloud data P′, which includes the following steps:

[0034] Step 3.1. Using the straight line features extracted in Step 2, calculate the angles of the straight lines in the source point cloud data. Angle with the straight line in the target point cloud data

[0035]

[0036]

[0037] Step 3.2. Place any two By subtracting each pair of elements, we can obtain the angle difference θ. ij ; the one-dimensional angle difference θ ij Convert to two-dimensional variable (θ) ij ,θ ij Clustering is performed using a density-based clustering algorithm, where the average value of the clusters with the highest density is the final rotation angle α. The rotation matrix S is calculated using the following formula for rotation registration:

[0038]

[0039] Step 3.3. Calculate the translation matrix T of the rotated source point cloud data. i =(t x ,t y ):

[0040]

[0041]

[0042] in, For the target point cloud data, the straight line The x and y coordinates of the midpoint Straight lines in the source point cloud data The x and y coordinates of the midpoint;

[0043] Then, the translation matrix T i Clustering is performed using a density clustering algorithm, where the average value of the cluster with the highest density is the translation matrix T. The rotated source point cloud data is then translated by T to complete coarse registration, resulting in coarsely registered source point cloud data P′.

[0044] Specifically, in step 4, the process of finely registering the coarsely registered source point cloud data P′ using the TrICP cropping iterative nearest point algorithm includes the following steps:

[0045] Step 4.1. Find the nearest point in the coarsely registered source point cloud data P′ in the target point cloud data Q to form a point pair, and calculate the Euclidean distance μ between the point pairs. i ;

[0046] Step 4.2. Transfer μ i Sort in ascending order, starting with the first pair and selecting the preceding ones. The root mean square error is calculated for each point pair using the following formula:

[0047]

[0048] The n mentioned above represents the number of point cloud data points in the source point cloud data P′ after coarse registration.

[0049] Step 4.3. If the root mean square error is less than the set threshold or the number of iterations reaches the set value, then terminate the iteration; otherwise, continue to the next step.

[0050] Step 4.4. Using the selection from Step 4.2 For each pair of points, the rotation matrix R and translation vector t are solved using the least squares method.

[0051] Step 4.5. Perform rotation and translation transformations on the coarsely registered source point cloud data P′ using the rotation matrix R and translation vector t. The specific formula is as follows:

[0052] P′=P′×R+t

[0053] Step 4.6. Return to step 4.1 and perform iterative solution.

[0054] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0055] This invention employs a "coarse-fine" registration strategy. First, the point cloud data to be registered is segmented into regions using dynamic thresholds. Then, within these segmented regions, irregular features such as arcs and polygons are further segmented, and a random sampling consensus algorithm is used to fit straight line parameters and extract straight line features. Subsequently, based on these straight line features, density clustering is used to estimate the optimal transformation matrix for coarse registration via rotation and translation. Finally, the point cloud data processed in the above steps is finely registered. Experiments were conducted using LiDAR to collect indoor point cloud data at initial positions with a 100cm difference in horizontal coordinates and a 50cm difference in vertical coordinates. The root mean square error (RMSE) of the ICP algorithm was 6.5334cm, while the RMS error of this invention was only 4.7007cm, effectively improving the accuracy of LiDAR point cloud registration. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating a laser point cloud registration method combining feature extraction and clustering algorithms according to the present invention.

[0057] Figure 2 The point cloud data to be registered is obtained by laser radar scanning in step one of an embodiment of the present invention.

[0058] Figure 3 The straight line feature is extracted in step two of an embodiment of the present invention.

[0059] Figure 4 This is the coarse registration result of the point cloud data after rotation and translation in step two of an embodiment of the present invention.

[0060] Figure 5 The registration result is obtained using the Iterative Closest Point (ICP) algorithm.

[0061] Figure 6 The final point cloud registration result is obtained using the method of this invention. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings.

[0063] like Figure 1 As shown, the present invention provides a laser point cloud registration method combining feature extraction and clustering algorithms, comprising the following steps:

[0064] Step 1: Using a LiDAR scanner, scan the surrounding environment at initial positions 100cm apart on the horizontal axis and 50cm apart on the vertical axis to acquire the source point cloud data P and target point cloud data Q to be registered. Based on the scanning characteristics of the LiDAR, perform dynamic thresholding region segmentation on the point cloud data to be registered.

[0065] Step 2: Within the segmented region, perform feature division, use the random sampling consensus algorithm to fit the line parameters, and extract the line features.

[0066] Step 3: Use density clustering algorithm to estimate the transformation matrix of the extracted straight line features, and perform coarse registration by rotation and translation to obtain the coarsely registered source point cloud data P′.

[0067] Step 4: Perform fine registration on the coarsely registered source point cloud data P′ using the Trimmed Iterative ClosestPoint (TRICP) algorithm.

[0068] Step 1 includes the following steps:

[0069] Step 1.1: First, use a LiDAR to scan the surrounding environment at an initial position 100cm apart in horizontal coordinates and 50cm apart in vertical coordinates to obtain the point cloud data to be registered, such as... Figure 2 As shown, the source point cloud data P = {P} can be obtained. i =(X i ,Y i ), i∈[1,2,…n]}, where X i Y i Let x and y be the x and y coordinates of the i-th source point cloud data point, respectively, and n be the number of source point cloud data points. The target point cloud data Q = {Q...} j =(M j N j ),j=[1,2,…m]}, where M j N j ...

[0070] Step 1.2: Calculate the distance d between each adjacent point cloud data point in both the source and target point cloud data. i d j The specific formula is as follows:

[0071]

[0072]

[0073] Step 1.3: Calculate the dynamic threshold δ for region segmentation in the source point cloud data and the target point cloud data respectively. i δ j It can be seen that the greater the distance between the point cloud data points and the radar transmission point, the smaller the segmentation threshold should be. The specific formula is:

[0074]

[0075]

[0076] Where ρ i and ρ i+1 The laser-radar source point cloud data point P is respectively i and P i+1 The Euclidean distance, γ, is the source point cloud data point P. i and P i+1 The included angle between them; ρ j and ρ j+1 The laser radar reaches the target point cloud data point Q. j and Q j+1 The Euclidean distance, σ is the target point cloud data point Q. j and Q j+1The included angle between them; in this invention, the value of ξ1 is 5, and ξ1 can be adjusted appropriately when using different models of lidar.

[0077] Step 1.4: When d i ≥δ i At that time, the source point cloud data point P is considered to be... i Using d as the dividing point, the source point cloud data is divided into regions; when d j ≥δ j When this happens, the target point cloud data point Q is considered to be... j Using the dividing point as the dividing point, the target point cloud data is divided into regions, and finally the source point cloud data and the target point cloud data are each divided into several regions {D}. 1 D 2 ,…,D m The algorithm then determines the number of point cloud data points within the segmented region after dividing the source and target point cloud data. If the number of point cloud data points within a region is less than or equal to 3, the region is considered environmental noise and is deleted.

[0078] Perform region segmentation on the source point cloud data and target point cloud data according to steps 1.1-1.4.

[0079] Step 1 above is based on the working principle of LiDAR: the greater the distance between the obstacle and the LiDAR, the greater the spacing between adjacent point cloud data points. When using a fixed threshold to segment point cloud data into regions, it may incorrectly segment linear features at different distances, affecting the subsequent registration effect. This invention proposes a dynamic threshold region segmentation method, which enables scientific and reasonable segmentation of point cloud data at different distances, which is beneficial to the subsequent registration process.

[0080] Step 2 includes the following steps:

[0081] Step 2.1: In the region D that was segmented in Step 1 i In the middle, select the first point cloud data point k. s (X s Y s ) and the last point cloud data point k n (X n Y n To construct a straight line, we use (X) as the reference. s Y s ), (X n Y n ) are respectively k s k n The coordinates.

[0082] Step 2.2: Calculate region D i The distances from each point in the point cloud data to the line are calculated to obtain a set of all distances. Record its maximum value Let its corresponding point be k max (X max Y max ).

[0083] Step 2.3: In region D i In the middle, according to endpoint k s and k n The distance between them determines the partition threshold d. T The specific formula is as follows:

[0084]

[0085] Where ρ (<1) is the scaling parameter. The maximum value in the distance set... and the dividing threshold d T Compare. If Then take point cloud data point k max Divide region D into two parts. i It is divided into two parts; otherwise, it is not divided.

[0086] Following steps 2.1-2.3, the region {D} can be... 1 D 2 ,…,D m The data is further divided to obtain a new point cloud data region {A}. 1 A 2 ,…,A n}

[0087] Step 2.4: Use the random sample consensus algorithm to divide the region A. i Perform line parameter fitting to obtain the line parameters from the source point cloud data. Linear parameters in target point cloud data like Figure 3 As shown, the straight line extracted from the source point cloud data is The straight line extracted from the target point cloud data is

[0088] In step 2.4 of this invention, a random sampling consensus algorithm is used to process region A. i The process involves fitting linear parameters and extracting linear features. This includes the following steps:

[0089] 2.4.1: Since any two point cloud data points can determine a straight line y = ax + b, we first sequentially start from region A. i Select two points and solve for the line parameters a and b.

[0090] 2.4.2: Solving for the current region A i The distance from all point cloud data points to the line y = ax + b is calculated. If the distance of a point cloud data point is less than a set threshold, it means that the point cloud data point supports the current parameter. Finally, the number of point cloud data points that support the line parameter is obtained.

[0091] Repeat the above two steps until the selected region A has been traversed. i All point cloud data points are selected, and then the line parameter supported by the most point cloud data points is selected as the final parameter.

[0092] Step 2 proposes a scientific and reasonable method for segmenting the region {D}. 1 D 2 ,…,D m The irregular features such as arcs and broken lines in the data were further segmented. This step can segment irregular features into straight line features for unified subsequent processing, enabling effective feature extraction from point cloud data in complex environments and expanding the application scenarios of laser point cloud registration.

[0093] The linear features extracted in step 3 are used to estimate the transformation matrix using a density clustering algorithm, and coarse registration is performed through rotation and translation, including the following steps:

[0094] Step 3.1: Using the straight line features extracted in Step 2, calculate the angles of the straight lines in the source point cloud data. Angle with the straight line in the target point cloud data The specific formula is as follows:

[0095]

[0096]

[0097] Step 3.2: Place any two By subtracting each pair of elements, we can obtain the angle difference θ. ij The one-dimensional angle difference θ ij Convert to two-dimensional variable (θ) ij ,θ ij Clustering is performed using a density-based clustering algorithm, where the average value of the clusters with the highest density is taken as the final rotation angle α. The rotation matrix S is calculated using the following formula for rotation registration:

[0098]

[0099] Step 3.3: Calculate the translation matrix T of the rotated source point cloud data. i =(t x ,t y The specific formula is:

[0100]

[0101]

[0102] in For the target point cloud data, the straight line The x and y coordinates of the midpoint Straight lines in the source point cloud data The x and y coordinates of the midpoint. Then, the translation matrix T. i Clustering is performed using a density-based clustering algorithm, where the average value of the cluster with the highest density is the translation matrix T. For example... Figure 4 As shown, the rotated source point cloud data is translated by T to complete coarse registration, resulting in coarsely registered source point cloud data P′.

[0103] For the extracted straight line features, step 3 first performs coarse registration by subtracting the angles and using density clustering to solve for the rotation angle. Adaptive density clustering is employed, which is insensitive to noise and exhibits good robustness. Furthermore, it does not require pre-setting the K value, and accurately and conveniently solves for the rotation matrix. Then, clustering is used to calculate the displacement of the midpoints of the lines to obtain the final translation matrix. The rotated source point cloud data is then translated by T, completing the coarse registration simply and effectively.

[0104] Step 4 involves re-registering the coarsely registered source point cloud data P′ using the TrICP (Trimmed Iterative ClosestPoint) algorithm, including the following steps:

[0105] Step 4.1: Find the nearest point in the coarsely registered source point cloud data P′ in the target point cloud data Q to form a point pair, and calculate the Euclidean distance μ between the point pairs. i ;

[0106] Step 4.2: Transfer μ i Sort in ascending order, starting with the first pair and selecting the preceding ones. The root mean square error is calculated for each pair of points (where n is the number of point data points in the source point cloud data P′ after coarse registration). The specific formula is as follows:

[0107]

[0108] Step 4.3: If the root mean square error is less than the set threshold or the number of iterations reaches the set value, then terminate the iteration. Otherwise, continue to the next step;

[0109] Step 4.4: Using the selection from step 4.2 For each pair of points, the rotation matrix R and translation vector t are solved using the least squares method.

[0110] Step 4.5: Perform rotation and translation transformations on the coarsely registered source point cloud data P′ using the rotation matrix R and translation vector t. The specific formula is as follows:

[0111] P′=P′×R+t

[0112] Step 4.6: Return to step 4.1 and perform iterative solution.

[0113] This step sorts the Euclidean distances and uses only a certain proportion of point cloud data points to solve for the transformation matrix, effectively reducing the impact of incorrectly matched points.

[0114] Experiment based on the method of this invention: In the experiment, indoor point cloud data was collected using lidar at initial positions with a horizontal coordinate difference of 100cm and a vertical coordinate difference of 50cm. Registration was compared using the ICP algorithm and the algorithm of this invention. Figure 5 The ICP registration results show a low overlap rate in the point cloud data, with a root mean square error of 6.5334 cm. The registration results of the method of this invention are as follows: Figure 6 As shown, the point cloud data has a high overlap rate and a root mean square error of only 4.7007 cm, which effectively improves the accuracy of laser point cloud registration.

Claims

1. A laser point cloud registration method combining feature extraction and clustering algorithms, characterized in that, Includes the following steps: Step 1: Use a LiDAR to scan the surrounding environment at an initial position 100cm apart on the horizontal axis and 50cm apart on the vertical axis to obtain the point cloud data to be registered; combine the scanning characteristics of the LiDAR to perform dynamic threshold region segmentation on the point cloud data to be registered. Step 2: Within the segmented region, perform feature division, use the random sampling consensus algorithm to fit the line parameters, and extract the line features; Step 3: Use density clustering algorithm to estimate the transformation matrix of the extracted straight line features, and perform coarse registration by rotation and translation to obtain the coarsely registered source point cloud data P′; Step 4: Perform fine registration on the coarsely registered source point cloud data P′ using the TrICP cropping iterative nearest point algorithm; The point cloud data that needs to be registered includes: source point cloud data P = {P i =(X i ,Y i ), i∈[1,2,…n]}, where X i Y i Let x and y be the x and y coordinates of the i-th source point cloud data point, respectively, and n be the number of source point cloud data points; the target point cloud data Q = {Q j =(M j N j ),j=[1,2,…m]}, where M j N j ... In step 3, the extraction of straight-line features is used to estimate the transformation matrix using a density clustering algorithm, and coarse registration is performed through rotation and translation to obtain the coarsely registered source point cloud data P′. This includes the following steps: Step 3.

1. Using the straight line features extracted in Step 2, calculate the angles of the straight lines in the source point cloud data. Angle with the straight line in the target point cloud data Step 3.

2. Place any two By subtracting each pair of elements, we can obtain the angle difference θ. ij ; the one-dimensional angle difference θ ij Convert to two-dimensional variable (θ) ij ,θ ij Clustering is performed using a density-based clustering algorithm, where the average value of the cluster with the highest density is the final rotation angle i. The rotation matrix S is calculated using the following formula for rotation registration: Step 3.

3. Calculate the translation matrix T of the rotated source point cloud data. i =(t x ,t y ): in, For the target point cloud data, the straight line The x and y coordinates of the midpoint Straight lines in the source point cloud data The x and y coordinates of the midpoint; Then, the translation matrix T i Clustering is performed using a density clustering algorithm, where the average value of the cluster with the highest density is the translation matrix T. The rotated source point cloud data is then translated by T to complete coarse registration, resulting in coarsely registered source point cloud data P′.

2. The laser point cloud registration method combining feature extraction and clustering algorithms according to claim 1, characterized in that, Step 1 includes the following steps: Step 1.

2. Calculate the distance d between each adjacent point cloud data point in both the source and target point cloud data. i d j : Step 1.

3. Calculate the dynamic threshold δ for region segmentation in the source point cloud data and the target point cloud data respectively. i δ j It can be seen that the greater the distance between the point cloud data points and the radar transmission point, the smaller the segmentation threshold should be. The specific formula is: Where ρ i and ρ i+1 These are laser-radar source point cloud data points P. i and P i+1 The Euclidean distance, γ, is the source point cloud data point P. i and P i+1 The included angle between them; ρ j and ρ j+1 The laser radar reaches the target point cloud data point Q. j and Q j+1 The Euclidean distance, σ is the target point cloud data point Q. j and Q j+1 The included angle between them; in this invention, the value of ξ1 is 5, and ξ1 can be adjusted appropriately when using different models of lidar.

3. The laser point cloud registration method combining feature extraction and clustering algorithms according to claim 2, characterized in that, In step 1: when d i ≥δ i At that time, the source point cloud data point P is considered to be... i Using d as the dividing point, the source point cloud data is divided into regions; when d j ≥δ j When this happens, the target point cloud data point Q is considered to be... j Using the dividing point as the dividing point, the target point cloud data is divided into regions, and finally the source point cloud data and the target point cloud data are each divided into several regions {D}. 1 D 2 ,…,D m The algorithm determines the number of point cloud data points within the segmented region of the source and target point cloud data. If the number of point cloud data points within the region is less than or equal to 3, the region is considered environmental noise and is deleted.

4. The laser point cloud registration method combining feature extraction and clustering algorithms according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.

1. In the region D that was segmented in Step 1... i In the middle, select the first point cloud data point k. s (X s Y s ) and the last point cloud data point k n (X n Y s To construct a straight line; where (X) s Y s ), (X n Y n ) are respectively k s k n The coordinates; Step 2.

2. Calculate region D i The distances from each point in the point cloud data to the line are calculated to obtain a set of all distances. Record the maximum value among them Let its corresponding point be k max (X max Y max ); Step 2.

3. In region D i In the middle, according to endpoint k s and k n The distance between them determines the partition threshold d. T The specific formula is as follows: Where ρ(<1) is the scaling parameter; the maximum value in the distance set. and the dividing threshold d T Compare; if Then take point cloud data point k max Divide region D into two parts. i Divide into two parts; otherwise, do not divide. Following steps 2.1-2.3, the region {D} can be... 1 D 2 ,…,D m The data is further divided to obtain a new point cloud data region {A}. 1 A 2 ,…,A n }; Step 2.

4. Use the random sample consensus algorithm to partition the region A. i Perform line parameter fitting to obtain the line parameters in the source point cloud data. Linear parameters in target point cloud data The straight line extracted from the source point cloud data is The straight line extracted from the target point cloud data is 5. The laser point cloud registration method combining feature extraction and clustering algorithms according to claim 4, characterized in that, The aforementioned random sampling consensus algorithm is used for region A i The process of fitting line parameters and extracting line features includes: 2.4.

1. Since any two point cloud data points can determine a straight line y = ax + b, we first proceed sequentially from region A. i Select two points in the equation and solve for the line parameters a and b. 2.4.

2. Solving for the current region A i The distance from all point cloud data points to the line y = ax + b is calculated. If the distance of a point cloud data point is less than a set threshold, it means that the point cloud data point supports the current parameter. Finally, the number of point cloud data points that support the line parameter is obtained. Repeat the above two steps until the selected region A has been traversed. i All point cloud data points are selected, and then the line parameter supported by the most point cloud data points is selected as the final parameter.

6. The laser point cloud registration method combining feature extraction and clustering algorithms according to claim 1, characterized in that, In step 4, the coarsely registered source point cloud data P′ is finely registered using the TrICP cropping iterative nearest point algorithm, which includes the following steps: Step 4.

1. Find the nearest point in the coarsely registered source point cloud data P′ in the target point cloud data Q to form a point pair, and calculate the Euclidean distance μ between the point pairs. i ; Step 4.

2. Transfer μ i Sort in ascending order, starting with the first pair and selecting the preceding ones. The root mean square error is calculated for each point pair using the following formula: The n mentioned above represents the number of point cloud data points in the source point cloud data P′ after coarse registration. Step 4.

3. If the root mean square error is less than the set threshold or the number of iterations reaches the set value, then terminate the iteration; otherwise, continue to the next step. Step 4.

4. Using the selection from Step 4.2 For each pair of points, the rotation matrix R and translation vector t are solved using the least squares method. Step 4.

5. Perform rotation and translation transformations on the coarsely registered source point cloud data P′ using the rotation matrix R and translation vector t. The specific formula is as follows: P′=P′×R+t Step 4.

6. Return to step 4.1 and perform iterative solution.

Images