Vehicle-mounted LiDAR point cloud registration method and system based on adaptive segmentation model

Abstract

The present invention discloses a method and system for point cloud registration of an on-board laser radar based on an adaptive segmentation model. The method comprises the following steps: first, using improved voxel filtering and straight-through filtering to simplify the point cloud. Then, the ISS algorithm is used to analyze the point cloud features and extract key points. Next, the 4PCS algorithm is used to calculate the initial pose under the key point constraint to complete the coarse registration. Finally, the KD-tree is used to obtain the key point density features to construct an adaptive segmentation model, and the NDT algorithm is combined to achieve fast and precise registration of the point cloud. The present invention can not only effectively solve the problems of slow speed and low stability of point cloud registration of MEMS laser radar on autonomous driving vehicles, but also has strong robustness to external interference.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent information perception technology and vehicle-mounted laser radar technology, and specifically relates to a vehicle-mounted laser radar point cloud registration method and system based on an adaptive segmentation model. Background Art

[0002] As one of the sensors used for autonomous driving environment perception, lidar is considered the "eyes" of autonomous vehicles due to its high resolution, strong anti-interference capabilities, and high stability. As the vehicle moves, the onboard lidar captures point cloud information from different perspectives of the target. Therefore, it is necessary to stitch the previous and next point cloud frames together to obtain a complete 3D point cloud of the target.

[0003] Point cloud registration involves aligning two input point clouds to maximize their overlap. Its core objective is to determine the spatial rotation and translation matrices between the source and target point clouds. In autonomous driving applications, the frequently changing road conditions and vehicle speeds often result in point cloud data collected by on-board MEMS lidars exhibiting high noise, low overlap, and complex and diverse matching objects, significantly increasing the difficulty of registration. Currently, researchers in the field primarily use deep learning-based and non-deep learning-based methods for point cloud processing. Non-deep learning methods are further categorized into traditional and feature-based methods. Traditional methods include the Iteration Closest Point (ICP) algorithm, generalized ICP (GICP), the Normal Distributions Transform (3D-NDT) algorithm, and the 4-Point Congruent Sets (4PCS) method. Feature-based methods consist of feature description and extraction. Feature description aims to convert high-dimensional feature information into lower-dimensional descriptors that are easier to distinguish and compare. Feature extraction aims to detect and retain highly unique portions of the point cloud. Commonly used feature descriptors include PFH, FPFH, SHOT, 3DSC, CSF, RMF, 2D Line Features, LPPF, etc.; key point extraction methods include 3D-DoG, 3D-Harris, ISS, SIFT, Hough Transform, RANSAC-based plane detection, etc.

[0004] The ICP algorithm offers high registration accuracy, but its application requirements are stringent, requiring a good initial pose, low point cloud noise, and a high overlap rate. While NDT can quickly process large-scale point cloud data, it is slow and difficult to set parameters. 4PCS doesn't require an initial pose, but its registration accuracy is poor and its stability is insufficient, making it prone to registration errors. There are many different feature-based methods, most of which have a narrow scope of application and cannot adapt to a variety of point cloud types. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a vehicle-mounted lidar point cloud registration method and system based on an adaptive segmentation model, thereby solving the problem of slow operation speed in the existing technology.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] The vehicle-mounted laser radar point cloud registration method based on the adaptive segmentation model includes the following steps:

[0008] S1: point cloud reduction using improved voxel filtering and pass-through filter;

[0009] S2: Use the ISS algorithm to analyze the features of the simplified point cloud and extract key points;

[0010] S3: Under the key point constraints, the 4PCS algorithm is used to calculate the initial pose to obtain the coarse registration point cloud;

[0011] S4: Use KD-tree to obtain key point density features, build an adaptive segmentation model, and based on the adaptive segmentation model, combine the NDT algorithm to achieve fine registration of the coarsely registered point cloud.

[0012] Preferably, in S1, the method for simplifying the point cloud using the improved voxel filtering and the straight-through filter includes:

[0013] S11: Get the original point cloud data P = {p i}, i=1,2,…n and target point cloud data Q={q j},j=1,2,…m,p i and q i are points in point clouds P and Q respectively, and n and m are the number of point clouds in P and Q respectively;

[0014] S12: Preprocess the original point cloud data P and the target point cloud data Q using a straight-through filter and a voxel filter with a minimum number of points constraint, and output point clouds P and Q respectively. pre ={p i},i=1,2,…n pre and Q pre ={q j},j=1,2,…mpre , n pre and m pre are the number of point clouds of P and Q after preprocessing.

[0015] Preferably, in S2, the method of using the ISS algorithm to analyze the features of the simplified point cloud and extract key points includes:

[0016] Use the ISS algorithm to analyze the point cloud P pre and the point cloud Q pre Extract key points and output point clouds P key ={p i},i=1,2,…n key and Q key ={q j},j=1,2,…m key , n key and m key The extracted P key and Q key The number of point clouds.

[0017] Preferably, in S3, under the key point constraint condition, the method of using the 4PCS algorithm to calculate the initial pose to obtain the coarse registration point cloud includes:

[0018] The point cloud P is processed by 4PCS algorithm. key and the point cloud Q key Perform coarse registration and output point cloud P coarse ={p i},i=1,2,…n coarse and Q coarse ={q i},i=1,2,…m coarse , n coarse and m coarse After rough registration, P key and Q key The number of point clouds.

[0019] Preferably, in S4, the method of using KD-tree to obtain key point density features and constructing an adaptive segmentation model includes:

[0020] S41: Calculate point cloud P key Any point p i The Euclidean distance L between the points constitutes the set P L ={L1,…,L n-1};

[0021] S42: The set P L The elements in are sorted in descending order;

[0022] S43: Based on the set P sorted in descending order L , calculate the point cloud P according to formula (9) key Any point p i The minimum distance value l i ,

[0023] l i =min(L(p i ,q)),q=1,2,…,n key -1,q≠p i (9)

[0024] Where q is P key Not with p i Any point of coincidence;

[0025] S44: Repeat S41, S42 and S43 until the minimum distance value corresponding to each point is obtained;

[0026] S45: Based on the minimum distance value corresponding to each point, establish a KD-Tree for search and calculate the density of the point cloud P according to formula (10)

[0027]

[0028] S46: Parameter density multiplier M for adjusting block size according to density ρ , building an adaptive segmentation model.

[0029] Preferably, in S4, the method for achieving fine registration of the coarsely registered point cloud in combination with the NDT algorithm includes:

[0030] Using NDT algorithm to calculate P coarse and Q coarse Perform precise registration and output point cloud P fine ={p i},i=1,2,…n fine and Q fine ={q i},i=1,2,…m fine , n finee and m fine After fine registration, P fine and Q fine The number of point clouds.

[0031] The present invention also provides a vehicle-mounted laser radar point cloud registration system based on an adaptive segmentation model, comprising: a simplification module, an extraction module, a coarse registration module and a fine registration module;

[0032] The simplification module is used to simplify the point cloud using improved voxel filtering and pass-through filter;

[0033] The extraction module is used to analyze the features of the simplified point cloud and extract key points using the ISS algorithm;

[0034] The coarse registration module is used to calculate the initial pose using the 4PCS algorithm under the key point constraint conditions to obtain a coarse registration point cloud;

[0035] The fine registration module is used to obtain key point density features using KD-tree, build an adaptive segmentation model, and achieve fine registration of the coarse registration point cloud based on the adaptive segmentation model in combination with the NDT algorithm.

[0036] Preferably, the streamlined module includes: an acquisition unit and a pre-processing unit;

[0037] The acquisition unit is used to acquire the original point cloud data P={p i}, i=1,2,…n and target point cloud data Q={q j},j=1,2,…m,p i and q i are points in point clouds P and Q respectively, and n and m are the number of point clouds in P and Q respectively;

[0038] The pre-processing unit is used to pre-process the original point cloud data P and the target point cloud data Q using a straight-through filter and a voxel filter with a minimum number of points constraint, respectively, and output point clouds P and Q respectively. pre ={p i},i=1,2,…n pre and Q pre ={q j},j=1,2,…m pre , n pre and m pre are the number of point clouds of P and Q after preprocessing.

[0039] Preferably, in the extraction module, the process of using the ISS algorithm to analyze the features of the simplified point cloud and extract key points includes:

[0040] Use the ISS algorithm to analyze the point cloud P pre and the point cloud Q pre Extract key points and output point clouds P key ={p i},i=1,2,…n key and Q key ={q j},j=1,2,…m key , n key and m key The extracted P key and Q key The number of point clouds.

[0041] Preferably, in the coarse registration module, under the key point constraint condition, the process of calculating the initial pose using the 4PCS algorithm to obtain the coarse registration point cloud includes:

[0042] The point cloud P is processed by 4PCS algorithm. key and the point cloud Q key Perform coarse registration and output point cloud P coarse ={p i},i=1,2,…n coarse and Q coarse ={qi},i=1,2,…m coarse , n coarse and m coarse After rough registration, P key and Q key The number of point clouds.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] Compared to traditional denoising algorithms, this invention integrates an adaptive segmentation model with NDT's MEMS-LIDAR point cloud registration algorithm, significantly improving registration speed while minimizing precision loss and enhancing resistance to external interference. Using point cloud density for adaptive parameter setting significantly simplifies the parameter adjustment process. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 It is a flow chart of the vehicle-mounted laser radar point cloud registration method based on the adaptive segmentation model of the present invention;

[0047] Figure 2 It is a schematic diagram of the present invention;

[0048] Figure 3 This is a schematic diagram of voxel filtering with minimum point constraints (the red dot is the origin cloud; the blue dot is the voxel centroid; the green dot is the closest point to the voxel centroid);

[0049] Figure 4 It is a schematic diagram of the present invention constructing and fusing an adaptive segmentation model;

[0050] Figure 5 It is the KDF-NDT flow chart;

[0051] Figure 6This is a visual comparison diagram of the present invention and various methods;

[0052] Figure 7 This is a comparison chart of the registration accuracy of the present invention;

[0053] Figure 8 This is a graph showing changes in registration results using the same type of method;

[0054] Figure 9 It is a comparison diagram of the relative error scatter distribution of each method before and after the four interferences. DETAILED DESCRIPTION

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Example 1

[0058] like Figure 1 As shown, the present invention provides a vehicle-mounted laser radar point cloud registration method based on an adaptive segmentation model, comprising the following steps:

[0059] S1: point cloud reduction using improved voxel filtering and pass-through filter;

[0060] S2: Use the ISS algorithm to analyze the features of the simplified point cloud and extract key points;

[0061] S3: Under the key point constraints, the 4PCS algorithm is used to calculate the initial pose to obtain the coarse registration point cloud;

[0062] S4: Use KD-tree to obtain key point density features, build an adaptive segmentation model, and based on the adaptive segmentation model, combine the NDT algorithm to achieve fine registration of the coarsely registered point cloud.

[0063] In this embodiment, the solution of the present invention refers to Figure 2 As shown, the present invention provides an adaptive segmentation and registration method for vehicle-mounted MEMS lidar point cloud, comprising the following steps:

[0064] S1: Methods for reducing point clouds using improved voxel filtering and pass-through filters include:

[0065] S11: Taking the three scenes of Hesai Technology’s PandaSet dataset as an example, input the original point cloud data P = {p i},i=1,2,…n and target point cloud data Q={qj},j=1,2,…m,p i and q i is a point in point cloud P and Q respectively, n and m are the number of point clouds in P and Q respectively. The number of point clouds is 169171, 148700, and 178386 respectively. Figure 6 .

[0066] S12: Preprocess the point cloud data P and Q obtained in S11 using a straight-through filter and a voxel filter with a minimum number of points constraint, and output point clouds P and Q respectively. pre ={p i},i=1,2,…n pre and Q pre ={q j},j=1,2,…m pre , n pre and m pre are the number of point clouds of P and Q after preprocessing, see Figure 3 .

[0067] The specific process of S12 includes:

[0068] (1-1) Calculate the side length L of the small grid according to formula (1).

[0069]

[0070] Where α is the proportional factor for adjusting the side length of the small grid, L X 、L Y 、L Z are the length, width and height of the small grid respectively, and N represents the total number of point clouds P.

[0071] (1-2) Determine the coordinates of the center of gravity within the grid according to formula (2) and find the point that replaces this grid.

[0072]

[0073] (1-3) According to formula (3), the center of gravity point P b Use the voxel center of gravity closest point P v Alternative.

[0074]

[0075] Where: n min The number of points in each voxel grid must reach the minimum limit; n is the number of points in the voxel grid; (x i ,y i ,zi ) and (x b ,y b ,z b ) are any point p in the grid i and the voxel center point p b 's coordinates.

[0076] S2: The method of using ISS algorithm to analyze the features of the simplified point cloud and extract key points includes: using ISS algorithm to analyze P pre and Q pre Extract key points and output point clouds P key ={p i},i=1,2,…n key and Q key ={q j},j=1,2,…m key , n key and m key The extracted P key and Q key The number of point clouds.

[0077] The specific process of S2 includes:

[0078] (2-1): For each point p in the point cloud P j (x j ,y j ,z j ) Set the search radius r or the minimum number of neighbors k.

[0079] (2-2): With p j Search for all points within the radius r as the center, and calculate their weight W according to the difference in distribution of neighboring points in formula (4) ji , whose values ​​are:

[0080]

[0081] Where p k For neighboring points

[0082] (2-3): At point p j Analyze the features within the range of its neighboring points, and calculate each point p according to formula (5) j The covariance matrix of :

[0083]

[0084] (2-4): Calculate each cov(p j )'s eigenvalue {λ j 1 ,λ j 2 ,λj 3} and eigenvectors Arrange the eigenvalues ​​in descending order.

[0085] (2-5): Set thresholds ε1 and ε2 according to formula (6)

[0086] λ j 2 / λ j 1 ≤ε1,λ j 3 / λ j 2 ≤ε2,0<ε1,ε2<1 (6)

[0087] (2-6): Repeat the above steps until all points are traversed to obtain the key point set P key and Q key .

[0088] S3: Under the key point constraint condition, the method of using 4PCS algorithm to calculate the initial pose to obtain the rough registration point cloud includes: using 4PCS algorithm to calculate P key and Q key Perform coarse registration and output point cloud P coarse ={p i},i=1,2,…n coarse and Q coarse ={q i},i=1,2,…m coarse , n coarse and m coarse After rough registration, P key and Q key The number of point clouds.

[0089] The specific process of S3 includes:

[0090] (3-1): In P key Select 4 points a, b, c, and d, and let line ac intersect line bd at point e.

[0091] (3-2): Calculate the two affine invariant ratios R1 and R2 according to formula (7):

[0092]

[0093] (3-3): Calculate the two intersection points that meet the two affine ratios according to formula 8:

[0094] e1=i1+R1(j1-i1), e2=i1+R2(j1-i1) (8)

[0095] Where e1, e2 are possible intersection points; i1, i2, j1, j2 are Q key Four points in

[0096] (3-4): Eliminate point sets with large differences in the paired point sets.

[0097] (3-5): Perform spatial transformation according to the optimal transformation matrix.

[0098] In this embodiment, the ISS key point extraction algorithm in S2 and the 4PCS algorithm in S3 are combined to form the I4PCS algorithm.

[0099] S4: Methods for constructing adaptive segmentation models include: Figure 4 、 Figure 5 .

[0100] (4-1): Calculate the point cloud P key Any point p i The Euclidean distance L between the points constitutes the set P L ={L1,…,L n-1}

[0101] (4-2): P L Sort the elements in descending order

[0102] (4-3): Calculate the point cloud P according to formula (9) key Any point p i The minimum distance value l i for:

[0103] l i =min(L(p i ,q)),q=1,2,…,n key -1,q≠p i (9)

[0104] Where q is P key Not with p i Any point of overlap.

[0105] (4-4): Repeat the above steps until the l corresponding to each point is obtained i

[0106] (4-5): Establish a KD-Tree for search and use the KD-tree algorithm to search for P i The nearest neighbor point, K is the number of neighborhood points, let the neighborhood point set be N(p i )={p ik ∈p j}, where p ik For p iThe density of the point cloud P is calculated according to formula (10) based on the K neighborhood points of

[0107]

[0108] (4-6): Parameter density multiplier M that adjusts block size according to density ρ

[0109] S4: The method of combining NDT algorithm to realize the fine registration of coarse registration point cloud includes: using NDT algorithm to P coarse and Q coarse Perform coarse registration and output point cloud P fine ={p i},i=1,2,…n fine and Q fine ={q i},i=1,2,…m fine , n finee and m fine After fine registration, P fine and Q fine The number of point clouds.

[0110] The specific process of S4 includes:

[0111] (5-1): Point cloud P coarse The space occupied by the adaptive segmentation model built by S4 is divided into blocks, and Q coarse Put it into the corresponding block.

[0112] (5-2): Calculate the mean vector and covariance matrix of the probability density function of the block according to formula (11)

[0113]

[0114] Where m is the number of point clouds in a single block; is the scanning point of the target point cloud in the grid; T is the transformation matrix.

[0115] (5-3): Initialize or update the transformation parameters and re-search the corresponding block.

[0116] (5-4): Calculate the probability distribution function of the block according to formula (12).

[0117]

[0118] Where: D is the vector dimension, which is 3; is an origin cloud scan point.

[0119] (5-5): Embed the adaptive segmentation model into the objective function according to formula (13) and minimize it.

[0120]

[0121] Where: Indicates passing Go to Transform Point k is the point cloud P fine The amount of model constructed in max It is point cloud P fine The volume of the largest bounding box; k ρ It is the point cloud block {r1,…,r m}, r m is a point within the block.

[0122] (5-6): Use Newton's method to solve the optimal transformation, calculate and update the Jacobian matrix, Hessian matrix and Newton step size.

[0123] (5-7): Finally, update the Newton step size and determine whether the update amount converges or reaches the number of iterations. If so, jump out, otherwise repeat S4.

[0124] In this embodiment, the adaptive segmentation model in S4 is integrated with the NDT algorithm to form the KDF-NDT algorithm.

[0125] The source point cloud and the target point cloud are registered according to the above method, and the algorithm parameters are shown in Table 1.

[0126] Table 1

[0127]

[0128] S5: Output transformed point cloud P fine and Q fine .

[0129] S6: Finally, calculate the evaluation index and output the algorithm visualization diagram. Figure 6 、 Figure 7 .

[0130] S7: Change the algorithms for the coarse and fine registration stages and calculate the evaluation indicators. Figure 8 .

[0131] S8: Add 4 types of interference respectively: remove 10% of the local point cloud (IF1), remove 20% of the global point cloud (IF2), add 20% of the local noise (IF3), and add 50% of the global noise (IF4), and calculate the evaluation index. Figure 9 .

[0132] In the Hesai Technology PandaSet dataset, the comparison results of the registration accuracy of ICP, GICP, NDT, NDT+ICP, the method in reference

[31] and the method in reference

[32] with the method of the present invention are as follows: Figure 7 As shown. Among them, the reference

[31] is Xu G, Pang Y, Bai Z, Wang Y, Lu ZA Fast Point Clouds Registration Algorithm for Laser Scanners. Applied Sciences. 2021; 11(8): 3426. The reference

[32] is Baek J, Park J, Cho S, Lee C. 3D Global Localization in the Underground Mine Environment Using Mobile LiDAR Mapping and Point Cloud Registration. Sensors. 2022; 22(8): 2873.

[0133] In the Hesai Technology PandaSet dataset, the registration time and whether the registration is successful are shown in Table 2 for ICP, GICP, 4PCS, NDT, NDT+ICP, the method in reference

[31] , the method in reference

[32] and the method of the present invention.

[0134] Table 2

[0135]

[0136] From Table 2 and Figure 7 It can be seen that the registration algorithm integrating the adaptive segmentation model and NDT improves the registration efficiency by 70%-80% in the three PandaSet scenarios. Therefore, the proposed method has a significant advantage in registration efficiency while ensuring registration accuracy, thus proving the feasibility of the present invention.

[0137] Depend on Figure 8 It can be seen that the I4PCS algorithm and KDF-NDT significantly improve the stability of the algorithm.

[0138] Depend on Figure 9 It can be seen that under the four interferences, the relative error of the proposed method is small, indicating that the anti-interference ability is strong.

[0139] Example 2

[0140] The present invention also provides a vehicle-mounted laser radar point cloud registration system based on an adaptive segmentation model, comprising: a simplification module, an extraction module, a coarse registration module and a fine registration module;

[0141] The reduction module is used to reduce point clouds using improved voxel filtering and pass-through filters;

[0142] The extraction module is used to analyze the features of the simplified point cloud and extract key points using the ISS algorithm;

[0143] The coarse registration module is used to calculate the initial pose using the 4PCS algorithm under key point constraints to obtain the coarse registration point cloud;

[0144] The fine registration module is used to obtain key point density features using KD-tree, build an adaptive segmentation model, and achieve fine registration of the coarsely registered point cloud based on the adaptive segmentation model and combined with the NDT algorithm.

[0145] In this embodiment, the streamlined module includes: an acquisition unit and a pre-processing unit;

[0146] The acquisition unit is used to obtain the original point cloud data P = {p i}, i=1,2,…n and target point cloud data Q={q j},j=1,2,…m,p i and q i are points in point clouds P and Q respectively, and n and m are the number of point clouds in P and Q respectively;

[0147] The preprocessing unit is used to preprocess the original point cloud data P and the target point cloud data Q using a straight-through filter and a voxel filter with a minimum number of points constraint, respectively, and output point clouds P and Q respectively. pre ={p i},i=1,2,…n pre and Q pre ={q j},j=1,2,…m pre , n pre and m pre are the number of point clouds of P and Q after preprocessing.

[0148] In this embodiment, in the extraction module, the process of using the ISS algorithm to analyze the features of the simplified point cloud and extract key points includes:

[0149] Use the ISS algorithm to analyze the point cloud P pre and the point cloud Q pre Extract key points and output point clouds P key ={p i},i=1,2,…n key and Q key ={q j},j=1,2,…m key , n key and m key The extracted P key and Qkey The number of point clouds.

[0150] In this embodiment, in the coarse registration module, under the key point constraint condition, the process of calculating the initial pose using the 4PCS algorithm to obtain the coarse registration point cloud includes:

[0151] The point cloud P is processed by 4PCS algorithm. key and the point cloud Q key Perform coarse registration and output point cloud P coarse ={p i},i=1,2,…n coarse and Q coarse ={q i},i=1,2,…m coarse , n coarse and m coarse After rough registration, P key and Q key The number of point clouds.

[0152] In this embodiment, in the fine registration module, the process of using KD-tree to obtain key point density features and constructing an adaptive segmentation model includes:

[0153] S41: Calculate point cloud P key Any point p i The Euclidean distance L between the points constitutes the set P L ={L1,…,L n-1};

[0154] S42: The set P L The elements in are sorted in descending order;

[0155] S43: Based on the set P sorted in descending order L , calculate the point cloud P according to formula (9) key Any point p i The minimum distance value l i ,

[0156] l i =min(L(p i ,q)),q=1,2,…,n key -1,q≠p i (9)

[0157] Where q is P key Not with p i Any point of coincidence;

[0158] S44: Repeat S41, S42 and S43 until the minimum distance value corresponding to each point is obtained;

[0159] S45: Based on the minimum distance value corresponding to each point, establish a KD-Tree for search and calculate the density of the point cloud P according to formula (10)

[0160]

[0161] S46: Parameter density multiplier M for adjusting block size according to density ρ , building an adaptive segmentation model.

[0162] In this embodiment, in the fine registration module, a method for implementing fine registration of the coarse registration point cloud in combination with the NDT algorithm includes:

[0163] Using NDT algorithm to calculate P coarse and Q coarse Perform precise registration and output point cloud P fine ={p i},i=1,2,…n fine and Q fine ={q i},i=1,2,…m fine , n finee and m fine After fine registration, P fine and Q fine The number of point clouds.

[0164] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A vehicle-mounted lidar point cloud registration method based on an adaptive segmentation model, characterized in that: The following steps are involved: S1: using improved voxel filtering and straight-through filtering to simplify the point cloud, wherein the point cloud is point cloud data collected by a lidar; S2: Use the ISS algorithm to analyze the features of the simplified point cloud and extract key points; S3: Under the key point constraints, the 4PCS algorithm is used to calculate the initial pose to obtain the coarse registration point cloud; S4: Use KD-tree to obtain key point density features, build an adaptive segmentation model, and based on the adaptive segmentation model, combine the NDT algorithm to achieve fine registration of the coarsely registered point cloud; In S4, the method of using KD-tree to obtain key point density features and constructing an adaptive segmentation model includes: S41: Calculate point cloud P key Any point p i The Euclidean distance L between the points constitutes the set P L ={L1,…,L n-1 }; S42: The set P L The elements in are sorted in descending order; S43: Based on the set P sorted in descending order L , calculate the point cloud P according to formula (9) key Any point p i The minimum distance value l i , （9） Where q is P key Not with p i Any point of coincidence; S44: Repeat S41, S42 and S43 until the minimum distance value corresponding to each point is obtained; S45: Based on the minimum distance value corresponding to each point, establish a KD-Tree for search and calculate the density of the point cloud P according to formula (10) （10） S46: Parameter density multiplier for adjusting block size according to density , build an adaptive segmentation model; The method of realizing the fine registration of the rough registration point cloud by combining the NDT algorithm includes: using the NDT algorithm to coarse and Q coarse Perform coarse registration and output point cloud P fine ={p i },i=1,2,…n fine and Q fine ={q i },i=1,2,…m fine , n finee and m fine After fine registration, P fine and Q fine The number of point clouds; The specific process of S4 includes: (5-1): Point cloud P coarse The space occupied by the adaptive segmentation model built by S4 is divided into blocks, and Q coarse Invest in the corresponding block; (5-2): Calculate the mean vector and covariance matrix of the probability density function of the block according to formula (11) （11） Where m is the number of point clouds in a single block; is the scanning point of the target point cloud in the grid; T is the transformation matrix; (5-3): Initialize or update the transformation parameters and re-search the corresponding block; (5-4): Calculate the probability distribution function of the block according to formula (12); （12） Where: D is the vector dimension, which is 3; is an origin cloud scan point; (5-5): Embed the adaptive segmentation model into the objective function according to formula (13) and minimize it; （13） Where: V max It is point cloud P fine The volume of the largest bounding box; It is the point cloud block {r1,…,r m }, r m is the point in the block; (5-6): Use Newton's method to solve the optimal transformation, calculate and update the Jacobian matrix, Hessian matrix and Newton step size; (5-7): Finally, update the Newton step size and determine whether the update amount converges or reaches the number of iterations. If so, jump out, otherwise repeat S4.

2. The vehicle-mounted laser radar point cloud registration method based on the adaptive segmentation model according to claim 1 is characterized in that: In S1, the method for simplifying point clouds using improved voxel filtering and pass-through filtering includes: S11: Get the original point cloud data P={p i },i=1,2,…n and target point cloud data Q={q j },j=1,2,…m,p i and q i are points in point clouds P and Q respectively, and m are the number of point clouds of P and Q respectively; S12: The original point cloud data P and the target point cloud data Preprocessing is performed using a straight-through filter and a voxel filter with a minimum number of points constraint, and the point clouds P are output respectively. pre ={p i },i=1,2,…n pre and Q pre ={q j },j=1,2,…m pre , n pre and m pre are the number of point clouds of P and Q after preprocessing.

3. The vehicle-mounted laser radar point cloud registration method based on the adaptive segmentation model according to claim 2 is characterized in that: In S2, the method of using the ISS algorithm to analyze the features of the simplified point cloud and extract key points includes: Use the ISS algorithm to analyze the point cloud P pre and the point cloud Q pre Extract key points and output point clouds P key ={p i },i=1,2,…n key and Q key ={q j },j=1,2,…m key , n key and m key The extracted P key and Q key The number of point clouds.

4. The vehicle-mounted laser radar point cloud registration method based on the adaptive segmentation model according to claim 3 is characterized in that: In S3, under the key point constraint condition, the method of using the 4PCS algorithm to calculate the initial pose to obtain the coarse registration point cloud includes: The point cloud P is processed by 4PCS algorithm. key and the point cloud Q key Perform coarse registration and output point cloud P coarse ={p i },i=1,2,…n coarse and Q coarse ={q i },i=1,2,…m coarse , n coarse and m coarse After rough registration, P key and Q key The number of point clouds.

5. The vehicle-mounted lidar point cloud registration system based on the adaptive segmentation model is characterized by: include: Simplification module, extraction module, coarse registration module and fine registration module; The simplification module is used to simplify the point cloud using improved voxel filtering and pass-through filter; The extraction module is used to analyze the features of the simplified point cloud and extract key points using the ISS algorithm; The coarse registration module is used to calculate the initial pose using the 4PCS algorithm under the key point constraint conditions to obtain a coarse registration point cloud; The fine registration module is used to obtain key point density features using KD-tree, build an adaptive segmentation model, and achieve fine registration of the coarse registration point cloud based on the adaptive segmentation model in combination with the NDT algorithm; The method of realizing the fine registration of the rough registration point cloud by combining the NDT algorithm includes: using the NDT algorithm to coarse and Q coarse Perform coarse registration and output point cloud P fine ={p i },i=1,2,…n fine and Q fine ={q i },i=1,2,…m fine , n finee and m fine After fine registration, P fine and Q fine The number of point clouds; The specific process of S4 includes: (5-1): Point cloud P coarse The space occupied by the adaptive segmentation model built by S4 is divided into blocks, and Q coarse Invest in the corresponding block; (5-2): Calculate the mean vector and covariance matrix of the probability density function of the block according to formula (11) （11） Where m is the number of point clouds in a single block; is the scanning point of the target point cloud in the grid; T is the transformation matrix; (5-3): Initialize or update the transformation parameters and re-search the corresponding block; (5-4): Calculate the probability distribution function of the block according to formula (12); （12） Where: D is the vector dimension, which is 3; is an origin cloud scan point; (5-5): Embed the adaptive segmentation model into the objective function according to formula (13) and minimize it; （13） Where: V max It is point cloud P fine The volume of the largest bounding box; It is the point cloud block {r1,…,r m }, r m is a point within the block; (5-6): Use Newton's method to solve the optimal transformation, calculate and update the Jacobian matrix, Hessian matrix and Newton step size; (5-7): Finally, update the Newton step size and determine whether the update amount converges or reaches the number of iterations. If so, jump out, otherwise repeat S4.

6. The vehicle-mounted laser radar point cloud registration system based on the adaptive segmentation model according to claim 5 is characterized in that: The streamlined module includes: an acquisition unit and a pre-processing unit; The acquisition unit is used to acquire the original point cloud data P={p i },i=1,2,…n and target point cloud data Q={q j },j=1,2,…m,p i and q i are points in point clouds P and Q respectively, and m are the number of point clouds of P and Q respectively; The pre-processing unit is used to combine the original point cloud data P and the target point cloud data Preprocessing is performed using a straight-through filter and a voxel filter with a minimum number of points constraint, and the point clouds P are output respectively. pre ={p i },i=1,2,…n pre and Q pre ={q j },j=1,2,…m pre , n pre and m pre are the number of point clouds of P and Q after preprocessing.

7. The vehicle-mounted laser radar point cloud registration system based on the adaptive segmentation model according to claim 6 is characterized in that: In the extraction module, the process of using the ISS algorithm to analyze the features of the simplified point cloud and extract key points includes: Use the ISS algorithm to analyze the point cloud P pre and the point cloud Q pre Extract key points and output point clouds P key ={p i },i=1,2,…n key and Q key ={q j },j=1,2,…m key , n key and m key The extracted P key and Q key The number of point clouds.

8. The vehicle-mounted laser radar point cloud registration system based on the adaptive segmentation model according to claim 7 is characterized in that: In the coarse registration module, under the key point constraint condition, the process of calculating the initial pose using the 4PCS algorithm to obtain the coarse registration point cloud includes: The point cloud P is processed by 4PCS algorithm. key and the point cloud Q key Perform coarse registration and output point cloud P coarse ={p i },i=1,2,…n coarse and Q coarse ={q i },i=1,2,…m coarse , n coarse and m coarse After rough registration, P key and Q key The number of point clouds.

Images