3D laser point cloud feature matching method assisted by geometric and semantic information

Abstract

The present invention proposes a three-dimensional laser point cloud feature matching method assisted by geometric and semantic information. The present invention achieves high-precision registration of laser point clouds by combining geometric and semantic information to construct constraints, effectively overcoming the traditional method's dependence on motion prior information and sensitivity to high mismatching rates. Semantic information is used to achieve building-to-building matching, ground-to-ground matching, and tree-to-tree matching, effectively filtering dynamic noise and reducing the dimensionality of the search space, thereby improving matching efficiency and accuracy. False matches are roughly eliminated through geometric and semantic constraints, and are finely eliminated through precise mathematical modeling, significantly reducing the complexity of problem solving. The invention can achieve high-precision three-dimensional laser point cloud registration more reliably and stably, and can still achieve accurate matching under high mismatching rates, solving many complex scene problems, and has better applicability.

Description

Technical Field

[0001] This invention belongs to the field of surveying, mapping, remote sensing, and laser 3D reconstruction technology, and specifically relates to a 3D laser point cloud feature matching method assisted by geometric and semantic information. By combining geometric and semantic information to construct constraints, highly reliable 3D registration of laser point cloud data is achieved. Background Art

[0002] Simultaneous Localization and Mapping (SLAM) technology involves using sensors to obtain scene observations (such as images, point clouds, and IMU data) while an unmanned system is moving in an unknown environment. This technology then uses the spatiotemporal correlation of these observations to recover the vehicle's position and posture (abbreviated as pose), while simultaneously constructing a map model of the surrounding environment. Compared to visual SLAM, laser SLAM (referred to herein as 3D laser) offers advantages such as high mapping accuracy, excellent stability, immunity to changes in lighting and viewing angle, and no scale drift, making it a popular choice for large-scale, real-world applications.

[0003] Laser SLAM uses two approaches for data association: feature matching (such as the LOAM algorithm) and point set matching (such as the ICP algorithm). Feature-based methods are popular due to their high efficiency. However, these methods are highly dependent on initial values and are prone to falling into local extrema. Mainstream approaches all assume that the vehicle is in a state of local uniform or slow motion to avoid local extrema, which leads to the problem of prior assumption dependence and is unsuitable for complex motion scenarios. Image matching uses descriptors to construct matching relationships between feature points, independent of any motion priors. Similarly, methods based on 3D feature descriptors can effectively address the problem of prior assumption dependence. Similar to image feature matching, 3D feature matching also involves the following steps: First, a feature detector extracts highly salient keypoints from the point cloud frame. Then, based on the feature descriptors, the local surface characteristics are analyzed and each keypoint is encoded into a compact feature vector to improve the distinguishability of the feature points. Next, a matching score is calculated for the feature vectors, and a one-to-one correspondence is established using nearest neighbor distance ratios or chi-square tests. Finally, a robust estimation method is used to eliminate false matches and estimate the pose. However, due to the problems of disordered organization, uneven density, lack of texture, structural occlusion, etc. in point cloud data, the significance of three-dimensional feature descriptors is greatly reduced, and the error matching points can reach 90%, resulting in the problem of correct matches being "swamped", which hinders its application in laser SLAM. Summary of the Invention

[0004] To address the problems and shortcomings of existing technologies, this paper proposes a 3D laser point cloud feature matching method that utilizes geometric and semantic information. This method uses semantic information to effectively filter dynamic targets (such as cars and pedestrians) and achieves classified matching of feature points (e.g., matching ground points only with ground points). It also uses topological constraints to fully utilize the spatial relationships between matching points, improving matching reliability. This method effectively reduces the matching search space and noise interference, effectively accommodating situations with high mismatch rates, and achieving highly reliable and high-precision feature-level matching of 3D laser point clouds.

[0005] Traditional methods usually use geometric models to eliminate false matches, but ignore the importance of semantic information and topological constraints. Semantic information can effectively filter dynamic targets (such as cars, pedestrians, etc.) and realize classified matching of feature points (such as ground points only match ground points). Topological constraints can make full use of the spatial relationship of matching points to improve matching reliability. Based on the above analysis, the present invention proposes a highly reliable three-dimensional feature matching method assisted by geometric and semantic constraints. First, the Transformer is introduced to design the feature detection and feature description network structure to improve the feature repetition rate and significance; then, the semantic and topological information is used to realize noise filtering and classified matching; finally, a false match elimination model is constructed based on the maximum consensus theory.

[0006] To achieve the above objectives, the present invention provides a method for matching three-dimensional laser point clouds assisted by geometric and semantic information, which mainly includes the following steps:

[0007] Step 1: Design a feature detection and feature description network structure based on context information, construct a matching loss function, and obtain an initial 3D feature descriptor set;

[0008] Step 2: Based on the feature points and feature descriptions obtained in step 1, semantic information is introduced to assist in the removal of dynamic noise points and classification matching, and a set of classification matching points is obtained;

[0009] Step 3: Based on the classification matching results, a topological graph structure network is constructed, and the supporting line voting strategy is used to eliminate matching pairs that do not meet the geometric constraints;

[0010] Step 4: Based on the rough elimination results, a 3D feature matching maximum consensus model is constructed and solved based on the boundary conflict effect to obtain the final accurate matching point set.

[0011] Furthermore, the specific implementation of step 1 includes the following sub-steps:

[0012] First, construct a feature detection network: select PointMLP as the backbone network of FPN to generate candidate feature points, and introduce the self-attention mechanism of Transformer to encode the context information of feature points, using the probabilistic Chamfer function As loss, an unsupervised learning method is used to obtain a set of three-dimensional feature points;

[0013] Then, a feature description network is constructed: the CCN module and the Transformer module are integrated to form a feature description network, the local reference frame of the feature points is calculated, and rotation invariance is achieved; hierarchical geometric information is encoded by CCN, global information is encoded by Transformer, and LRN is used for normalization to obtain the feature descriptor vector.

[0014] Furthermore, the feature detection network uses the probabilistic Chamfer function As the loss, given a set of candidate features and The mathematical formula is,

[0015]

[0016] Where, express Any point The distance in is the best matching distance (||·|| is the l2 norm), express Any point Medium x i The shortest distance between points, express Any point middle y j The shortest distance between points; x and y are and The characteristic points in σ ij and σ ji Describes the significant uncertainty of the best match, which can be estimated by feature description or the discreteness of three-dimensional space, σ ij express Midpoint and The matching similarity uncertainty factor of the corresponding points in , σ ji It means The point in The matching uncertainty of the corresponding points in .

[0017] Furthermore, the feature description network adopts the circle loss function To improve the flexibility and convergence characteristics of network optimization, its mathematical formula is:

[0018]

[0019] Where, Ω P and Ω N Represent the positive and negative sample sets respectively; and are the weights of positive and negative samples respectively; and are the feature distances of positive and negative samples respectively; Δ P and Δ N Represent the expected value of the feature distance of the positive and negative sample sets respectively.

[0020] Furthermore, the specific implementation of step 2 includes the following sub-steps:

[0021] First, the laser point cloud frame is projected to generate a two-dimensional grid depth map. The depth map is used as the input value of RangeNet to obtain a semantic segmentation map. The feature points obtained in step 1) are projected onto the semantic segmentation map to obtain the semantic label corresponding to each feature point.

[0022] Secondly, potential dynamic noise points are identified and removed based on semantic labels, including feature points located on dynamic objects such as cars, pedestrians, and bicycles. Non-dynamic feature points are then classified based on semantic labels.

[0023] Finally, descriptor matching is performed on feature points of the same category in the two frames of point clouds, including matching of building points with building points, ground points with ground points, and tree points with tree points. By calculating the Euclidean distance between the description vectors and finding the optimal match through nearest neighbor search, a one-to-one correspondence is constructed to achieve classification matching and obtain a set of classified matching points.

[0024] Furthermore, the specific implementation of step 3 is as follows:

[0025] The initial matching set is obtained by semantic classification matching After, among them and It is the feature set and The set after removing dynamic noise points, x and y represent the corresponding matching points; for each feature point x i ,and The remaining points x j ,j≠i forms a connecting line Formed with x i The radiating topology network is centered, and the connecting line on the radiating topology network is point x i The support line of the laser point cloud has the real world scale, and the support line length is a rotation and translation invariant. Therefore, when the support line The endpoint (x i ,y i ) and (x j ,y j ) are all correctly matched, there is When there is a mismatch in the endpoints, the equal length constraint of the support line does not hold. In other words, if there is and (xj , y j ) is the correct match, that is cast a vote for it. The higher the number of votes at the matching point, the higher the probability that it is the correct same-name point. Eliminate the matching points with low votes and zero votes. Finally, take the top P% of the matching pairs with the highest votes as the set of matching points after rough elimination.

[0026] Furthermore, the maximum consensus model in step 4 is as follows;

[0027] The core idea of the maximum consensus model is to seek the optimal transformation model by maximizing the inlier consensus set. The 3D feature matching mathematical model based on this idea is

[0028]

[0029] In the formula, is the inlier consensus set, represents the size of the inlier set, (R, t) are the rotation and translation pose parameters, is in step 3) the refined subset after the support line voting analysis, ε is the noise threshold, and x and y represent the corresponding matching points. In the case of a high false matching ratio, the direct model solution of this problem is very difficult. The present invention converts it into a series of sub-problems and then solves them based on the boundary conflict effect. Specifically, the formula (3) is equivalently transformed into

[0030]

[0031] In the formula, p k is the maximum consensus set of the following sub-problem (P k ),

[0032]

[0033] Sub-problem (P k ) is to find k all the matching points in k that conform to the model (R k , t k ) while (x ), y

[0034] Furthermore, the boundary conflict effect in step 4 is as follows;

[0035] Assume the lower bound value l of the optimal solution of the given formula (3) and the upper bound value u of the optimal solution p k of the sub-problem. If u < l, then, (x k , y k)It must be a wrong match. RANSAC is used to determine the lower bound value l, and the upper bound value u is determined by the extended matching matrix method. The wrong matches are eliminated according to the upper and lower bound conflicts.

[0036] Furthermore, the specific implementation method of step 4 is as follows;

[0037] First, based on the idea of maximizing the inlier consensus set on the basis of rough elimination, a maximum consensus mathematical model for three-dimensional feature matching is constructed, that is, formula (3), and it is converted into a series of sub-problems, that is, formula (4), and each matching point corresponds to a sub-problem (P k );

[0038] Secondly, use the RANSAC algorithm to obtain an approximate solution of formula (3), obtain an inlier set, and take the number of matching pairs included in the inlier set as the lower bound l of the optimal solution of formula (3);

[0039] Thirdly, use the inlier set obtained by RANSAC and (x k , y k ) to construct an extended matching matrix, and obtain the upper bound value u of the optimal solution p k ) of the sub-problem (P k ) through the extended matching matrix. If u < l, then (x k , y k ) must be a wrong match and is eliminated;

[0040] Finally, after performing the above boundary conflict effect detection on each matching point, a refined eliminated matching set is obtained, and then the RANSAC method is used to obtain the final reliable matching point set and the rigid body transformation model they satisfy. In addition, each matching point has a semantic label attribute.

[0041] Furthermore, in step 3, the matching pairs with the top 30% of the vote counts are selected as the matching point set after rough elimination.

[0042] The present invention has the following positive effects:

[0043] 1) The present invention proposes a three-dimensional laser point cloud feature matching method assisted by geometric and semantic information. By jointly constructing constraints with geometric and semantic information, high-reliable three-dimensional registration of laser point cloud data is achieved, overcoming the dependence of traditional methods on motion prior information and the sensitivity to a high wrong matching ratio.

[0044] 2) The present invention proposes a classification matching strategy. This strategy uses semantic information to achieve building-to-building matching, ground-to-ground matching, and tree-to-tree matching. This strategy can effectively filter dynamic noise points and achieve dimensionality reduction of the search space, improving the matching efficiency and accuracy.

[0045] 3) The present invention provides a coarse-to-fine three-dimensional feature matching concept, which firstly performs coarse elimination of mismatches through geometric and semantic constraints, and then achieves fine elimination of mismatches through precise mathematical modeling, thereby reducing the complexity of direct solution of the problem.

[0046] This invention enables more reliable and stable high-precision 3D laser point cloud registration. The descriptor-based 3D feature matching method is independent of prior motion information, offering improved applicability. Furthermore, this method provides a feature matching solution for situations with high mismatch rates, enabling it to address a variety of complex scenarios. This invention has broad application prospects in fields such as laser SLAM, 3D urban modeling, military navigation, and high-precision mapping. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Flowchart of the present invention.

[0048] Figure 2 Schematic diagram of the 3D feature detection and feature description network framework.

[0049] Figure 3 Schematic diagram of semantic-assisted classification matching strategy.

[0050] Figure 4 Schematic diagram of geometric topology constraints. DETAILED DESCRIPTION

[0051] The solution of the present invention is further described in detail below with reference to the accompanying drawings.

[0052] like Figure 1 As shown, the geometric and semantic information-assisted 3D laser point cloud feature matching method includes the following steps:

[0053] Step 1), such as Figure 2 As shown in the figure, the feature detection and feature description network structure is designed based on the context information, the loss function is selected, and a set of feature point descriptors is generated.

[0054] First, PointMLP is selected as the backbone network of FPN to generate candidate feature points, and the self-attention mechanism of Transformer is introduced to encode the context information of the feature points. As the loss, an unsupervised learning method is used to obtain a set of three-dimensional feature points.

[0055] Then, the CCN module and the Transformer module are integrated to form a feature description network, calculating the local reference frame of the feature points and achieving rotation invariance. CCN encodes hierarchical geometric information, while the Transformer encodes global information, and LRN is used for normalization to obtain the feature descriptor vector.

[0056] Specifically, to address the problem that traditional CNN feature description networks have a small receptive field and tend to overlook the geometric structural characteristics of laser point clouds, we introduce the Transformer module and the Cylindrical Convolution Network (CCN) module to design a feature detection and description network framework that integrates contextual information and local structural characteristics. The Transformer's self-attention mechanism effectively captures contextual information, while the multi-scale design of the Columnar Convolution Network (CCN) captures hierarchical geometric information. The feature detection network uses an unsupervised learning approach and consists of an FPN (Feature Proposal Network) and a Transformer. The FPN adopts a PointMLP backbone to generate candidate feature points. The feature description network consists of the CCN and Transformer, with the LRN performing l2 normalization. First, the local reference frame (LRF) is used to achieve rotational invariance of features. The CCN encodes hierarchical geometric information and combines it with the contextual information from the Transformer to perform feature description of local point cloud patches.

[0057] Feature detection network using probabilistic Chamfer function As loss, given a set of candidate features and The mathematical formula is,

[0058]

[0059] Where, express Any point The distance in is the best matching distance (||·|| is the l2 norm), express Any point Medium x i The shortest distance between points, express Any point middle y j The shortest distance between points; x and y are and The characteristic points in σ ij and σ ji Describes the significant uncertainty of the best match, which can be estimated by feature description or the discreteness of three-dimensional space, σ ij express Midpoint and The matching similarity uncertainty factor of the corresponding points in , σ ji It means The point in The probabilistic Chamfer loss takes into account the significant uncertainty of features and can suppress candidate feature points on flat surfaces.

[0060] The feature description network uses a circle loss function To improve the flexibility and convergence characteristics of network optimization, its mathematical formula is:

[0061]

[0062] Where, Ω P and Ω N Represent the positive and negative sample sets respectively; and are the weights of positive and negative samples respectively; and are the feature distances of positive and negative samples respectively; Δ P and Δ N Represent the expected value of the feature distance of the positive and negative sample sets respectively.

[0063] Step 2), such as Figure 3 As shown, based on the feature extraction and description in step 1), semantic information is introduced to assist in the removal of dynamic noise points and classification matching, and a set of classification matching points is obtained.

[0064] The motion state of a dynamic target is inconsistent with the carrier, which may cause multiple motion models to exist in the matching set. These dynamic points are interference points for inter-frame matching. In extreme cases, when dynamic matching points dominate, the inter-frame matching method will inevitably fail. Semantic information can effectively eliminate the interference of dynamic targets, and at the same time can achieve classification matching, reduce the search space of descriptor matching, and improve the efficiency of the matching algorithm. For example, assuming the candidate feature set and Both feature points, brute force matching requires calculation Sub-descriptor distance; if semantic information is used Divided into z categories (assuming equal division), we only need to calculate The distance is reduced, and efficiency is increased by a factor of z. At the same time, due to the reduction in the search space, the possibility of matching ambiguity is reduced, and the matching accuracy rate is also improved. The specific idea is to project the point cloud frame into a depth map, use RangeNet to quickly extract semantic information, discard dynamic target points such as cars and pedestrians, and then classify the feature points, and only perform descriptor matching on feature points of the same category.

[0065] First, the laser point cloud frame is projected to generate a 2D grid depth map. This depth map is used as the input value of RangeNet to obtain a semantic segmentation map. The feature points obtained in step 1) are projected onto the semantic segmentation map to obtain the semantic label corresponding to each feature point.

[0066] Secondly, potential dynamic noise points are identified and eliminated based on semantic labels, such as feature points located on dynamic objects such as cars, pedestrians, and bicycles, and non-dynamic feature points are classified according to semantic labels.

[0067] Finally, descriptor matching is performed on feature points of the same category in the two point clouds, such as building points, ground points, and tree points. By calculating the Euclidean distance between the descriptor vectors and finding the optimal match through nearest neighbor search, a one-to-one correspondence is established, achieving classification matching and obtaining a set of classified matching points.

[0068] Step 3) constructs a geometric topological graph structure and adopts a support line voting strategy to preliminarily eliminate erroneous matches in the classified matching point set to obtain a roughly eliminated matching point set.

[0069] First, in the semantic classification matching results On this basis, for each feature point x i ,and The remaining points x j ,j≠i forms a connecting line Construct a topological network, where the connecting lines on the network are points x i The support line of each feature point x i Can form Support lines. and It is the feature set and The set after removing dynamic noise points, x and y represent the corresponding matching points.

[0070] Secondly, for each feature point x i , calculate the length of its support line Similarly, calculation The corresponding matching support line Length like (ε is the correct matching threshold), then it is considered that (x i ,y i ) and (x j ,y j ) may be a correct match, i.e. Voted for it.

[0071] Finally, count each matching pair (x i ,yi ) are sorted from high to low based on the number of votes. Generally speaking, matches with more votes are more reliable than those with fewer votes. Figure 4 As shown, (x i ,y i ) has 4 votes, (x g ,y g ) has 1 vote, so (x i ,y i ) is more likely to be a correct match. The top 30% of the matching pairs are taken as the set of matching points after rough elimination.

[0072] Step 4) Build a maximum consensus mathematical model for 3D feature matching, remove false matches through boundary conflict effects, and use RANSAC to obtain the final reliable matching point set.

[0073] Maximum consensus model: Its core idea is to find the optimal transformation model by maximizing the inlier consensus set. The mathematical model of three-dimensional feature matching based on this idea is:

[0074]

[0075] Where, is an interior point consistent set, represents the size of the internal point set, (R, t) is the rotation and translation pose parameter, For step 3) The simplified subset after the support line voting analysis, ε is the noise threshold, x and y represent the corresponding matching points. In the case of high mismatch ratio, it is very difficult to solve the problem with a direct model. This paper converts it into a series of sub-problems and then solves them based on the boundary conflict effect. Specifically, formula (3) is equivalently converted to,

[0076]

[0077] Where p k For the following sub-problem (P k ),

[0078]

[0079] Subproblem (P k ) is in (x k ,y k ) must satisfy the transformation model (R k ,t k ) and find All matching points that conform to the model.

[0080] Boundary conflict effect: Assume the lower bound value l of the optimal solution of the given formula (3) and the optimal solution p of the subproblem and the upper bound value u of the optimal solution p of the subproblem k . If u < l, then (x k , y k ) must be a wrong match. The present invention uses RANSAC to determine the lower bound value l, uses the extended matching matrix method to determine the upper bound value u, and eliminates wrong matches according to the upper and lower bound conflicts

[0081] . First, based on the idea of maximizing the inlier consensus set on the basis of rough elimination, a maximum consensus mathematical model (formula (3)) for three-dimensional feature matching is constructed and converted into a series of subproblems (formula (4)), and each matching point corresponds to a subproblem (P k ).

[0082] . Second, use the RANSAC algorithm to obtain an approximate solution of formula (3), obtain an inlier set, and take the number of matching pairs included in the inlier set as the lower bound l of the optimal solution of formula (3);

[0083] . Third, use the inlier set obtained by RANSAC and (x k , y k ) to construct an extended matching matrix, and obtain the upper bound value u of the optimal solution p of the subproblem (P k ) through the extended matching matrix. If u < l, then (x k , y k ) must be a wrong match and be eliminated k .

[0084] . Finally, after performing the above boundary conflict effect detection on each matching point, obtain the matching set after fine elimination, and then use the RANSAC method to obtain the final reliable matching point set and the rigid body transformation model they satisfy. In addition, each matching point has a semantic label attribute

[0085] . The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention

Claims

1. A three-dimensional laser point cloud feature matching method assisted by geometric and semantic information, characterized in that: The following steps are involved: Step 1: Design a feature detection and feature description network structure based on context information, construct a matching loss function, and obtain an initial 3D feature descriptor set; The specific implementation of step 1 includes the following sub-steps: First, construct a feature detection network: select PointMLP as the backbone network of FPN to generate candidate feature points, and introduce the self-attention mechanism of Transformer to encode the context information of feature points, using the probabilistic Chamfer function As loss, an unsupervised learning method is used to obtain a set of three-dimensional feature points; Then, a feature description network is constructed: the CCN module (i.e., Cylindrical convolution network) and the Transformer module are integrated to form a feature description network. The local reference frame of the feature points is calculated to achieve rotation invariance. The hierarchical geometric information is encoded by CCN, the global information is encoded by Transformer, and the feature descriptor vector is obtained by normalization using LRN. Feature detection network using probabilistic Chamfer function As the loss, given a set of candidate features and , The mathematical formula is, (1) Where, express Any point middle The shortest distance between points, express Any point middle The shortest distance between points, for l 2 norm; and They are and The feature points in and Describe the significant uncertainty of the best match, estimated by the feature description or the discreteness of the three-dimensional space, express Midpoint and The matching similarity uncertainty of the corresponding points in It means The point in Uncertainty factors in matching corresponding points; Step 2: Based on the feature points and feature descriptions obtained in step 1, semantic information is introduced to assist in the removal of dynamic noise points and classification matching, and a set of classification matching points is obtained; Step 3: Based on the classification matching results, a topological graph structure network is constructed, and the supporting line voting strategy is used to eliminate matching pairs that do not meet the geometric constraints; Step 4: Based on the rough elimination results, a 3D feature matching maximum consensus model is constructed and solved based on the boundary conflict effect to obtain the final accurate matching point set.

2. The geometric and semantic information-assisted 3D laser point cloud feature matching method according to claim 1, characterized in that: The feature description network uses a circular loss function To improve the flexibility and convergence characteristics of network optimization, its mathematical formula is: (2) Where, and Represent the positive and negative sample sets respectively; and are the weights of positive and negative samples respectively; and are the feature distances of positive and negative samples respectively; and Represent the expected value of the feature distance between the positive and negative sample sets respectively.

3. The geometric and semantic information-assisted 3D laser point cloud feature matching method according to claim 1, characterized in that: The specific implementation of step 2 includes the following sub-steps: First, the laser point cloud frame is projected to generate a two-dimensional grid depth map. The depth map is used as the input value of RangeNet to obtain a semantic segmentation map. The feature points obtained in step 1) are projected onto the semantic segmentation map to obtain the semantic label corresponding to each feature point. Secondly, potential dynamic noise points are identified and removed based on semantic labels, including feature points located on dynamic objects such as cars, pedestrians, and bicycles. Non-dynamic feature points are then classified based on semantic labels. Finally, descriptor matching is performed on feature points of the same category in the two frames of point clouds, including matching of building points with building points, ground points with ground points, and tree points with tree points. By calculating the Euclidean distance between the description vectors and finding the optimal match through nearest neighbor search, a one-to-one correspondence is constructed to achieve classification matching and obtain a set of classified matching points.

4. The geometric and semantic information-assisted 3D laser point cloud feature matching method according to claim 1, characterized in that: The specific implementation of step 3 is as follows: The initial matching set is obtained by semantic classification matching After, among them and It is the feature set and The set after removing dynamic noise points, and Indicates the corresponding matching points; for each feature point ,and Remaining points Forming connecting lines , forming a The radiating topology network is centered, and the connecting lines on the radiating topology network are points The support line of the laser point cloud has the real world scale, and the support line length is a rotation and translation invariant. Therefore, when the support line Endpoint and When both are correctly matched, ; When there is a mismatch in the endpoints, the equal length constraint of the support line does not hold. In other words, if there is , and For a correct match, A vote is cast for it. The higher the number of votes for a matching point, the higher the probability that it is the correct homonymous point. Matching points with low or zero votes are eliminated, and finally the top P% matching pairs are obtained as the set of matching points after rough elimination.

5. The geometric and semantic information-assisted 3D laser point cloud feature matching method according to claim 1, characterized in that: The maximum consensus model in step 4 is as follows; The core idea of the maximum consensus model is to seek the optimal transformation model by maximizing the consistent set of interior points. The mathematical model of three-dimensional feature matching based on this idea is: (3) Where, is an interior point consistent set, represents the size of the inner point set, are the rotation and translation pose parameters, For step 3 The streamlined subset after support line voting analysis, is the noise threshold, and Indicates the corresponding matching point; in the case of high mismatch ratio, the direct model solution of this problem is very difficult, so it is converted into a series of sub-problems, and then solved based on the boundary conflict effect. Specifically, Formula (3) is equivalently converted to, (4) Where, For the following sub-problems P k The maximum consistent set of ( P k ) Subproblems P k is Satisfy transformation model At the same time, find All matching points that conform to the model.

6. The method for 3D laser point cloud feature matching assisted by geometric and semantic information according to claim 5, characterized in that: The boundary conflict effects in step 4 are as follows; Assume that the optimal solution of formula (3) is The lower bound of Optimal solution to the sum subproblem The previous value ,if ,So, It must be a wrong match, and RANSAC is used to determine the next value , determine the previous value by extending the matching matrix method , eliminate incorrect matches based on upper and lower bound conflicts.

7. The method for 3D laser point cloud feature matching assisted by geometric and semantic information according to claim 6, characterized in that: The specific implementation of step 4 is as follows; First, based on the rough elimination, the maximum consensus mathematical model of three-dimensional feature matching is constructed based on the idea of maximizing the consistent set of internal points, namely formula (3), and it is converted into a series of sub-problems, namely formula (4). Each matching point corresponds to a sub-problem P k ; Secondly, the RANSAC algorithm is used to find an approximate solution of formula (3) and obtain an inlier set. The number of matching pairs contained in the inlier set is used as the lower bound of the optimal solution of formula (3). ; Again, the inlier set obtained by RANSAC and Construct an extended matching matrix and solve subproblems through the extended matching matrix P k Optimal solution The previous value ,if ,So, It must be a wrong match and should be eliminated; Finally, after performing the above-mentioned boundary conflict effect detection on each matching point, a matching set after fine elimination is obtained. The RANSAC method is then used to obtain the final set of reliable matching points and the rigid body transformation model they satisfy. In addition, each matching point has a semantic label attribute.

8. The method for 3D laser point cloud feature matching assisted by geometric and semantic information according to claim 4, characterized in that: In step 3, the top 30% matching pairs with the highest number of votes are selected as the matching point set after rough elimination.

Images