Three-dimensional laser radar rotation scanning mapping method based on limited space and medium

Abstract

The invention discloses a three-dimensional laser radar rotary scanning mapping method based on a limited space and a medium. The method comprises the following steps: S10, assembling a hardware structure of a laser radar scanner for an underground well chamber space; s20, performing edge point extraction on the point cloud data, fitting a straight line, a plane and a cylinder, and constructing an environment constraint through multiple geometric features; s30, constructing an optimization objective function of multiple geometric features, defining error measurement of point cloud registration, solving a transformation matrix between point clouds by using a micro RANSAC point cloud registration method, and optimizing point cloud registration precision by using a gradient descent algorithm; and S40, providing a point cloud key frame construction strategy, carrying out global point cloud map fusion, and carrying out visualization through a remote program. In conclusion, the defects in the prior art are overcome, all angles of operation can be covered through the design of the rotary laser radar scanning head, the necessary movement angles of the acquisition equipment are reduced, and the device has high social use value and application prospects.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional lidar rotational scanning mapping technology, and more particularly to a three-dimensional lidar rotational scanning mapping method and medium based on limited space. Background Art

[0002] Existing handheld 3D mapping devices are limited by the cost and weight of LiDAR, typically equipped with only a single LiDAR unit. Due to the limited field of view of the LiDAR, operators need to frequently move the device to obtain complete scene data during the mapping process, which not only increases operational complexity but also seriously affects mapping efficiency. Furthermore, existing Simultaneous Localization and Mapping (SLAM) algorithms mainly rely on the spatial geometric information of point clouds for mapping. This method works effectively in environments with rich geometric structures, but in environments with sparse geometric features or many repetitive structures (such as tunnels, underground spaces, etc.), point cloud registration is prone to failure, leading to incomplete mapping results or severe cumulative errors, resulting in poor system robustness. For example, Chinese Patent Publication No. CN202111040513.1 discloses a point cloud registration method based on local curvature and its nearest neighbor features, using the ICP iterative nearest point algorithm for point cloud registration. This method is prone to mapping drift problems caused by mismatched point clouds in degraded environments.

[0003] To address the aforementioned problems, this invention proposes a highly efficient intensity-enhanced 3D laser mapping system. The invention employs a three-feature collaborative constraint model (line, plane, and cylinder) and introduces a differentiable RANSAC point cloud registration method to construct a gradient descent optimization objective function, thereby improving the accuracy and convergence speed of point cloud registration. Summary of the Invention

[0004] To address the problems mentioned in the background section, this invention provides a three-dimensional lidar rotation scanning mapping method and medium based on limited space.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A three-dimensional lidar rotation scanning mapping method based on finite space includes the following steps:

[0007] S10. Hardware structure for assembling a lidar scanner in underground well chamber space;

[0008] S20. Extract edge points from point cloud data, fit lines, planes and cylinders, and construct environmental constraints with multiple geometric features;

[0009] S30. Construct an optimization objective function with multiple geometric features, define an error metric for point cloud registration, solve for the transformation matrix between point clouds using the differentiable RANSAC point cloud registration method, and optimize the point cloud registration accuracy using the gradient descent algorithm.

[0010] S40. A point cloud keyframe construction strategy is proposed to perform global point cloud map fusion and visualize it through a remote program.

[0011] Preferably, the hardware structure includes a lidar scanning head and a data processing device host, and the two are arranged vertically.

[0012] Preferably, the lidar scanning head includes a lidar perpendicular to the data processing device host and a horizontally mounted gimbal motor; the data processing device host includes a core board, a battery, and a hardware driver board, which are stacked vertically; a working handle is installed on top of the data processing device host with reference to the center of gravity; and a signal antenna for data transmission is installed at the tail.

[0013] Preferably, the specific steps for fitting the line, plane, and cylinder in step S30 are as follows:

[0014] A lidar point cloud consists of multiple scan lines, each forming a line of point cloud image. The scan lines are divided into segments, assuming p... ij It is the j-th point on the i-th scan line, S ij Including p in the i-th row ij From the surrounding points, we can obtain:

[0015]

[0016] If the c-value of the laser scanning point is higher than 95% or greater than five times the median c, m If the c-value is large, then the point is considered an edge point; if the edge points are close to each other, only the point with the largest c-value is retained, these points are formed into a set E, and the scan line is divided into segments, and the points between the endpoints are formed into a set F.

[0017] Starting from the first scan line of the lidar, for a point p0∈E belonging to the i-th scan line, first find the nearest point p1∈E in the (i+1)-th scan line, and use p0 and p1 to calculate a line l. Then find the nearest point p2∈E of p1 in the (i+2)-th scan line. If the distance between p2 and line l is less than a certain distance threshold, use p0, p1 and p2 to fit a new line. This process is repeated until no new points can be added.

[0018] Using the same method, given a line segment P on the i-th scan line, for each point in P, first find the k nearest points that belong to F but are not on the i-th scan line, and use the RANSAC algorithm to fit the plane to these points; if the average distance from the interior points to the plane is greater than a threshold or the number of interior points is less than a threshold, then use the RANSAC algorithm to fit the cylindrical surface to these points. If the number of interior points fitted into the cylindrical surface is large and the average distance is small, then assign these points the cylindrical feature; finally, repeat this step until no points can be added to the cluster.

[0019] Suppose we let m represent a line segment, a plane, or a cylinder, m j Represents the j-th feature, and in pose X i The place is observed, and a set of observations is obtained. Record Will The homogeneous coordinates of all points in the middle are stacked together to obtain a The matrix.

[0020] Preferably, the step S20 of constructing environmental constraints using multiple geometric features specifically includes:

[0021] Suppose the least squares problem is The LM algorithm uses x k+1 =x k +Δ updates the solution in the k-th iteration, where Δ is a solution to the following equation:

[0022] (J T J+λI)Δ=-J T f(x)

[0023] Where J is x k The Jacobian matrix, λ, is a parameter of the LM algorithm, adjusted in each iteration to ensure... The value decreases.

[0024] Preferably, step S30 specifically includes:

[0025] S301: Extract three different types of geometric features: lines, planes, and cylinders, define the fitting error for each geometric feature, and construct a comprehensive objective function of the transformation matrix through the error measure of the geometric features to optimize the registration of point clouds.

[0026] S302: Using the differentiable RANSAC point cloud registration method, calculate the minimum registration error of the comprehensive optimization objective function that includes geometric features of lines, planes, and cylinders;

[0027] S303: By iteratively calculating the gradient descent optimization algorithm, the minimum inter-frame point cloud feature residual is obtained, and the accurate pose estimate of the current frame is obtained.

[0028] Preferably, step 301 specifically includes:

[0029] The basic goal of point cloud registration is to transform the source point cloud P using the transformation matrix T. s With the target point cloud P t Alignment:

[0030] T(P s ) = RP s +t

[0031] Where R is a 3*3 rotation matrix and t is a 3*1 translation vector;

[0032] To introduce geometric features from the source point cloud P s and target point cloud P t Extract three different types of geometric features: lines, planes, and cylinders, and define the fitting error for each geometric feature:

[0033] For a straight line feature, its parameter form is L: r = r0 + λd, where r0 is a point on the line, d is the direction vector of the line, and λ is a variable parameter. For a point cloud P s One point p i Distance d to the line i That is, the straight line error d i satisfy:

[0034] For planar features, let the general equation of the plane satisfy: Ax + By + Cz + D = 0, where A, B, and C are the components of the plane normal vector, and D is the offset of the plane. Let the point cloud P... s One point p i =(x i ,y i ,z i Error d of the distance to the plane i satisfy:

[0035] The general equation of a cylinder is: (x - x0) 2 +(y-y0) 2 =r 2 Where (x0, y0) is the center of the cylinder, r is the radius, and the calculation point is p. i =(x i ,y i ,z i Radial error of the distance between the cylinder and the cylinder

[0036] By using the error metric of geometric features, a comprehensive objective function for the transformation matrix is ​​constructed to optimize the registration of point clouds:

[0037]

[0038] Where, N l ,N p ,N c Point clouds representing the features of lines, planes, and cylinders, respectively. These represent the corresponding errors, and α, β, and γ are their respective dynamic weighting coefficients.

[0039] Preferably, step 302 specifically includes: the gradient of the objective function with respect to the rotation matrix R can be obtained by summing the derivatives of each error term; the gradients of the error terms for the line, plane, and cylinder with respect to the rotation matrix are:

[0040]

[0041] Similarly, the gradient of the translation vector t is also calculated using the derivative of the error term with respect to t:

[0042]

[0043] Preferably, step 303 specifically includes: optimizing the objective function E(T) with respect to the rotation matrix R and translation vector t using gradient descent, with the update rule being:

[0044]

[0045] Where η is the learning rate, which controls the step size of each iteration.

[0046] Preferably, step 40 specifically includes: using a point cloud keyframe construction strategy for the pose estimation of the current frame, and creating a new key point cloud sub-graph when the following three conditions are met:

[0047] 1) More than 20% of the points in the current frame are not being tracked;

[0048] 2) The rotation angle between the current frame and the last keyframe is greater than 5°;

[0049] 3) The distance between the current frame and the last keyframe is greater than a threshold, which is 0.3m by default;

[0050] New local geometric features (plane, line, and cylinder features) are detected in keyframes and attempted to match them with global features. Unmatched features are added as global features, and the accuracy of new matches is verified by geometric consistency check (GCC).

[0051] Next, the poses of all keyframes are jointly optimized together with all detected global geometric features to minimize the distance error from the point to the model. In the local adjustment, only the sliding window optimization algorithm is used to optimize the keyframes and their observed features within the window, while the poses of the keyframes outside the sliding window remain fixed. In the global adjustment process, feature errors caused by false detections are detected and corrected.

[0052] During all frame traversals, local geometric features (planes, lines, cylinders) are matched with global features using the aforementioned differentiable RANSAC algorithm to ensure the accuracy of the association between point clouds;

[0053] After pose matching is completed for all point clouds and their feature descriptions, all point cloud data are fused to obtain a global point cloud map.

[0054] The entire point cloud mapping process involves transmitting the current frame point cloud data and global point cloud data to the mobile point cloud display software via the device's Wi-Fi module. To achieve real-time point cloud transmission and optimal display on the mobile device, the point cloud data is thinned using a random downsampling method.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] 1. The overall structure of the lidar data acquisition unit is relatively compact, effectively addressing operational scenarios in confined spaces. Simultaneously, the rotating lidar scanning head design covers the entire operational angle, reducing the necessary movement angle of the acquisition equipment and increasing the depth of operation in limited spaces. Furthermore, compared to fixed lidar scanning structures, the hardware structure design of this invention has a smaller blind zone radius and greater spatial coverage, capable of mapping spaces with a minimum diameter of 30cm.

[0057] 2. This invention extracts edge points from point cloud data and fits the geometric features of lines, planes, and cylinders. It fully considers the mapping scenario of underground well chambers, increasing the diversity and universality of features in point cloud registration and avoiding performance degradation when a single feature is missing. Since the geometric features of lines, planes, and cylinders may exhibit unbounded variable attributes, the LM algorithm is used to solve the least squares problem that arises during the fitting process, accelerating the convergence speed of geometric feature fitting calculations and improving stability.

[0058] 3. The algorithm constructs a comprehensive objective function based on multiple geometric feature transformation matrices and uses the differentiable RANSAC algorithm for point cloud registration. The optimization process no longer relies on randomly selected sample points, but obtains the optimal registration transformation by minimizing the global error. Simultaneously, the gradient descent optimization algorithm is combined, which not only improves the accuracy of point cloud registration but also reduces computation time and enhances the robustness of registration. Compared with the single-feature ICP point cloud registration method, it effectively avoids the mapping failure problem in degraded scenarios and accelerates the convergence speed of point cloud registration.

[0059] 4. In the process of pose estimation and global point cloud construction, a dynamic keyframe generation and feature matching mechanism is designed. The selection of keyframes is not only based on time and distance, but also dynamically adjusted according to the point cloud coverage of the current frame and the success rate of matching. This avoids the algorithm from updating keyframes too frequently, while ensuring that keyframe optimization is performed when necessary, thus ensuring the accuracy of data association and the continuity of the map. Existing point cloud keyframe updates based on fixed time / distance thresholds are prone to redundancy or omissions in a limited space. The multi-condition triggering mechanism designed in this invention improves the accuracy of data association and the continuity of the map.

[0060] In summary, this invention overcomes the shortcomings of existing technologies. By designing a rotating lidar scanning head, it can cover the entire angle of the operation, reducing the necessary movement angle of the acquisition equipment, and has high social value and application prospects. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a flowchart of the three-dimensional lidar rotation scanning mapping method based on limited space according to the present invention.

[0063] Figure 2 Point cloud mapping method built for fast-lio (open source solution) Figure 1 ;

[0064] Figure 3 The point cloud map established by the mapping method proposed in this invention Figure 1 ;

[0065] Figure 4 Point cloud mapping method built for fast-lio (open source solution) Figure 2 ;

[0066] Figure 5 The point cloud map established by the mapping method proposed in this invention Figure 2 . Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] Example 1

[0069] This embodiment discloses a motor-driven, finite-space-based three-dimensional lidar rotation scanning mapping method, including the following steps:

[0070] S10. Hardware structure for assembling a lidar scanner in underground well chamber space;

[0071] S20. Extract edge points from point cloud data, fit lines, planes and cylinders, and construct environmental constraints with multiple geometric features;

[0072] S30. Construct an optimization objective function with multiple geometric features, define an error metric for point cloud registration, solve for the transformation matrix between point clouds using the differentiable RANSAC point cloud registration method, and optimize the point cloud registration accuracy using the gradient descent algorithm.

[0073] S40. A point cloud keyframe construction strategy is proposed to perform global point cloud map fusion and visualize it through a remote program.

[0074] In one specific embodiment, step S10 is as follows:

[0075] For the limited operating scenarios of underground well chambers, the overall hardware includes a lidar scanning head and a data processing equipment host, and the two are arranged vertically, that is, the lidar scanning head and the data processing equipment host are on the same axis.

[0076] The lidar scanning head includes a LivoxMid-360 lidar perpendicular to the main unit of the data processing equipment and a horizontally mounted gimbal motor. Under the limited working space of the acquisition equipment, it ensures that the lidar's field of view effectively covers all angles and has a smaller blind zone radius, enabling effective operation in limited spaces with a radius of more than 10cm.

[0077] The data processing equipment main unit consists of an RK3588 core board, a battery, and a hardware driver board. These three components are stacked vertically to reduce the cross-sectional area of ​​the main unit, preventing obstruction of the lidar's field of view while ensuring the equipment can operate deep into underground well chambers. Additionally, a working handle is mounted on top of the main unit, with the center of gravity as a reference point, and a signal antenna for data transmission is installed at the rear.

[0078] In one specific embodiment, step S20, which involves extracting edge points from the point cloud data, includes the following steps:

[0079] A lidar point cloud consists of multiple scan lines, each forming a line of point cloud image. The scan lines are then divided into segments. Assume p... ij It is the j-th point on the i-th scan line, S ij Including p in the i-th row ij From the surrounding points, we can obtain:

[0080]

[0081] If the c-value of the laser scanning point is higher than 95% or greater than five times the median c, m If the c-value is large, then the point is considered an edge point. If the edge points are close to each other, only the point with the largest c-value is retained. These points are formed into a set E, and the scan line is divided into segments, with the points between the endpoints forming a set F.

[0082] Starting with the first scan line of the LiDAR, for a point p0∈E belonging to the i-th scan line, first find the nearest point p1∈E in the (i+1)-th scan line. Calculate a line l using p0 and p1. Then find the nearest point p2∈E to p1 in the (i+2)-th scan line. If the distance between p2 and line l is less than a certain distance threshold, fit a new line using p0, p1, and p2. This process is repeated until no more new points can be added.

[0083] Using the same method, given a line segment P on the i-th scan line, for each point in P, first find the k nearest points belonging to F but not on the i-th scan line, and use the RANSAC algorithm to fit the plane to these points. If the average distance from the interior points to the plane is greater than a threshold or the number of interior points is less than a threshold, then use the RANSAC algorithm to fit a cylindrical surface to these points. If the number of interior points fitted to the cylindrical surface is large and the average distance is small, then these points are assigned cylindrical features. Finally, repeat this step until no more points can be added to the cluster.

[0084] Suppose we let m represent a line segment, a plane, or a cylinder, m j Represents the j-th feature, and in pose X i The location is observed. A set of observations is obtained. Record Will The homogeneous coordinates of all points in the middle are stacked together to obtain a The matrix.

[0085] However, since lines, planes, and cylinders may be unbounded variables, It can become infinitely large, leading to large-scale least squares problems, which are addressed using the Levenberg-Marquardt (LM) algorithm.

[0086] Suppose the least squares problem is The LM algorithm uses x k+1 =x k +Δ updates the solution in the k-th iteration, where Δ is a solution to the following equation:

[0087] (J T J+λI)Δ=-J T f(x)

[0088] Where J is x k The Jacobian matrix, λ, is a parameter of the LM algorithm, adjusted in each iteration to ensure... The value decreases.

[0089] This yields linear, planar, and cylindrical geometric descriptions of local point cloud data, enhancing the environmental constraints of the point cloud.

[0090] In one specific embodiment, the multi-geometric feature differentiable RANSAC point cloud registration step in step S30 is as follows:

[0091] The basic goal of point cloud registration is to transform the source point cloud P using the transformation matrix T. s With the target point cloud P t Alignment:

[0092] T(P s ) = RP s +t

[0093] Where R is a 3x3 rotation matrix and t is a 3x1 translation vector.

[0094] To introduce geometric features from the source point cloud P s and target point cloud P t It extracts three different types of geometric features: lines, planes, and cylinders, and defines the fitting error for each geometric feature.

[0095] For a straight line feature, its parameter form is L: r = r0 + λd, where r0 is a point on the line, d is the direction vector of the line, and λ is a variable parameter. For a point cloud P s One point p iDistance d to the line i That is, the straight line error d i satisfy:

[0096] For planar features, let the general equation of the plane satisfy: Ax + By + Cz + D = 0, where A, B, and C are the components of the plane's normal vector, and D is the plane's offset. Point cloud P s One point p i =(x i ,y i ,z i Error d of the distance to the plane i satisfy:

[0097] Similarly, the general equation of a cylinder is: (x - x0) 2 +(y-y0) 2 =r 2 Where (x0, y0) is the center of the cylinder, r is the radius, and the calculation point is p. i =(x i ,y i ,z i Radial error of the distance between the cylinder and the cylinder

[0098] By using the error metric of geometric features, a comprehensive objective function of the transformation matrix is ​​constructed to optimize the registration of point clouds.

[0099]

[0100] Where, N l ,N p ,N c Point clouds representing the features of lines, planes, and cylinders, respectively. These represent the corresponding errors, and α, β, and γ are their respective dynamic weighting coefficients.

[0101] Next, the differentiable RANSAC point cloud registration method is used to calculate the minimum registration error of the comprehensive optimization objective function that includes geometric features of lines, planes, and cylinders.

[0102] The gradient of the objective function with respect to the rotation matrix R can be obtained by summing the derivatives of each error term. The gradients of the error terms for the line, plane, and cylinder with respect to the rotation matrix are:

[0103]

[0104] Similarly, the gradient of the translation vector t is also calculated using the derivative of the error term with respect to t:

[0105]

[0106] Finally, the gradient of the objective function E(T) with respect to the rotation matrix R and the translation vector t is calculated, and the gradient descent method is used for optimization. The update rule is as follows:

[0107]

[0108] Where η is the learning rate, controlling the step size of each iteration. Through iterative calculation using the gradient descent optimization algorithm, the minimized inter-frame point cloud feature residuals are obtained, leading to a precise pose estimate for the current frame.

[0109] In one specific embodiment, the point cloud keyframe construction strategy and the global point cloud map fusion and display steps in step S40 are as follows:

[0110] The pose estimation of the current frame uses a point cloud keyframe construction strategy, which creates a new key point cloud subgraph when the following three conditions are met:

[0111] 1) More than 20% of the points in the current frame are not being tracked;

[0112] 2) The rotation angle between the current frame and the last keyframe is greater than 5°;

[0113] 3) The distance between the current frame and the last keyframe is greater than a threshold, which is 0.3m by default.

[0114] New local geometric features (plane, line, and cylinder features) are detected in keyframes and attempts are made to match them with global features. Unmatched features are added as global features, and the accuracy of the new matches is verified through a geometric consistency check (GCC).

[0115] Next, the poses of all keyframes are jointly optimized along with all detected global geometric features to minimize the distance error from points to the model. In local adjustments, only the sliding window optimization algorithm is used to optimize the keyframes and their observed features within the window, while the poses of keyframes outside the sliding window remain fixed. During global adjustments, feature errors caused by false detections are detected and corrected.

[0116] During all frame traversals, local geometric features (planes, lines, cylinders) are matched with global features using the aforementioned differentiable RANSAC algorithm to ensure the accuracy of the association between point clouds.

[0117] After pose matching is completed for all point clouds and their feature descriptions, all point cloud data are fused to obtain a global point cloud map.

[0118] The entire point cloud mapping process involves transmitting the current frame's point cloud data and global point cloud data to the mobile device's point cloud display software via the device's Wi-Fi module. To achieve real-time point cloud transmission and optimal display quality on the mobile device, the point cloud data is thinned using a random downsampling method.

[0119] In another aspect, the present invention also includes a three-dimensional lidar rotating scanning mapping equipment based on a limited space, comprising a three-dimensional lidar scanner, a scanning mapping system, and software running on a mobile device.

[0120] In another aspect, the present invention also includes a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method described above.

[0121] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the three-dimensional lidar rotation scanning mapping methods based on finite space in the above embodiments.

[0122] In the above embodiments, the following instructions can be implemented using a computer storage medium:

[0123] 1. Turn on the scanner and click the "Start Mapping" button in the remote client software to send sensor and mapping algorithm start commands to the scanner's built-in system;

[0124] 2. After starting the mapping task, the scanning first extracts point cloud features, and then uses differentiable RANSAC to perform local point cloud registration;

[0125] 3. Simultaneously, the scanner continuously transmits the downsampled point cloud to the client software via Wi-Fi, and the point cloud can be dragged and zoomed on the display interface;

[0126] 4. Finally, perform global point cloud map fusion. Click the "End" button in the remote client software to stop scanning and save the global point cloud to a pcd file.

Claims

1. A three-dimensional lidar rotational scanning mapping method based on finite space, characterized in that, Includes the following steps: S10. Hardware structure for assembling a lidar scanner in underground well chamber space; S20. Extract edge points from point cloud data, fit lines, planes and cylinders, and construct environmental constraints with multiple geometric features; S30. Construct an optimization objective function with multiple geometric features, define an error metric for point cloud registration, solve for the transformation matrix between point clouds using the differentiable RANSAC point cloud registration method, and optimize the point cloud registration accuracy using the gradient descent algorithm. S40. A point cloud keyframe construction strategy is proposed to perform global point cloud map fusion and visualize it through a remote program.

2. The three-dimensional lidar rotation scanning mapping method based on finite space according to claim 1, characterized in that: The hardware structure includes a lidar scanning head and a data processing device host, and the two are arranged vertically.

3. The three-dimensional lidar rotation scanning mapping method based on finite space according to claim 2, characterized in that: The lidar scanning head includes a lidar perpendicular to the data processing device host and a horizontally mounted gimbal motor; the data processing device host includes a core board, a battery, and a hardware driver board, which are stacked one on top of the other. A working handle is installed on top of the data processing device host with the center of gravity as a reference, and a signal antenna for data transmission is installed at the rear.

4. The three-dimensional lidar rotation scanning mapping method based on finite space according to claim 1, characterized in that: The specific steps for fitting the line, plane, and cylinder in step S30 are as follows: A lidar point cloud consists of multiple scan lines, each forming a line of point cloud image. The scan lines are divided into segments, assuming p... ij It is the j-th point on the i-th scan line, S ij Including p in the i-th row ij From the surrounding points, we can obtain: If the c-value of the laser scanning point is higher than 95% or greater than five times the median c, m If the c-value is large, then the point is considered an edge point; if the edge points are close to each other, only the point with the largest c-value is retained, these points are formed into a set E, and the scan line is divided into segments, and the points between the endpoints are formed into a set F. Starting from the first scan line of the lidar, for a point p0∈E belonging to the i-th scan line, first find the nearest point p1∈E in the (i+1)-th scan line, and use p0 and p1 to calculate a line l. Then find the nearest point p2∈E of p1 in the (i+2)-th scan line. If the distance between p2 and line l is less than a certain distance threshold, use p0, p1 and p2 to fit a new line. This process is repeated until no new points can be added. Using the same method, given a line segment P on the i-th scan line, for each point in P, first find the k nearest points that belong to F but are not on the i-th scan line, and use the RANSAC algorithm to fit the plane to these points; if the average distance from the interior points to the plane is greater than a threshold or the number of interior points is less than a threshold, then use the RANSAC algorithm to fit the cylindrical surface to these points; if the number of interior points fitted into the cylindrical surface is large and the average distance is small, then assign these points the cylindrical feature; finally, repeat this step until no points can be added to the cluster. Suppose we let m represent a line segment, a plane, or a cylinder, m j Let j represent the j-th feature, and in pose X i The place is observed, and a set of observations is obtained. Record Will The homogeneous coordinates of all points in the middle are stacked together to obtain a The matrix.

5. The three-dimensional lidar rotation scanning mapping method based on finite space according to claim 3, characterized in that: The step S20, which involves constructing environmental constraints using multiple geometric features, specifically includes: Suppose the least squares problem is The LM algorithm uses x k+1 =x k +Δ updates the solution in the k-th iteration, where Δ is a solution to the following equation: (J T J+λI)Δ=-J T f(x) Where J is x k The Jacobian matrix, λ, is a parameter of the LM algorithm, adjusted in each iteration to ensure... The value decreases.

6. The three-dimensional lidar rotation scanning mapping method based on finite space according to claim 5, characterized in that: Step S30 specifically includes: S301: Extract three different types of geometric features: lines, planes, and cylinders, define the fitting error for each geometric feature, and construct a comprehensive objective function of the transformation matrix through the error measure of the geometric features to optimize the registration of point clouds. S302: Using the differentiable RANSAC point cloud registration method, calculate the minimum registration error of the comprehensive optimization objective function that includes geometric features of lines, planes, and cylinders; S303: By iteratively calculating the gradient descent optimization algorithm, the minimum inter-frame point cloud feature residual is obtained, and the accurate pose estimate of the current frame is obtained.

7. The three-dimensional lidar rotation scanning mapping method based on finite space according to claim 6, characterized in that: Step 301 specifically includes: The basic goal of point cloud registration is to transform the source point cloud P using the transformation matrix T. s With the target point cloud P t Alignment: T(P s )=RP s +t Where R is a 3*3 rotation matrix and t is a 3*1 translation vector; To introduce geometric features from the source point cloud P s and target point cloud P t Extract three different types of geometric features: lines, planes, and cylinders, and define the fitting error for each geometric feature: For a straight line feature, its parameter form is L: r = r0 + λd, where r0 is a point on the line, d is the direction vector of the line, and λ is a variable parameter. For a point cloud P s One point p i Distance d to the line i That is, the straight line error d i satisfy: For planar features, let the general equation of the plane satisfy: Ax + By + Cz + D = 0, where A, B, and C are the components of the plane normal vector, and D is the offset of the plane. Let the point cloud P... s One point p i =(x i ,y i ,z i Error d of the distance to the plane i satisfy: The general equation of a cylinder is: (x - x0) 2 +(y-y0) 2 =r 2 Where (x0, y0) is the center of the cylinder, r is the radius, and the calculation point is p. i =(x i ,y i ,z i Radial error of the distance between the cylinder and the cylinder By using the error metric of geometric features, a comprehensive objective function for the transformation matrix is ​​constructed to optimize the registration of point clouds: Where, N l N p N c Point clouds representing the features of lines, planes, and cylinders, respectively. These represent the corresponding errors, and α, β, and γ are their respective dynamic weighting coefficients.

8. The three-dimensional lidar rotation scanning mapping method based on finite space according to claim 7, characterized in that: Step 302 specifically includes: the gradient of the objective function with respect to the rotation matrix R can be obtained by summing the derivatives of each error term. The gradients of the error terms for the line, plane, and cylinder with respect to the rotation matrix are: Similarly, the gradient of the translation vector t is also calculated using the derivative of the error term with respect to t: Step 303 specifically includes: optimizing the objective function E(T) with respect to the rotation matrix R and translation vector t using gradient descent, with the update rule being: Where η is the learning rate, which controls the step size of each iteration; Step 40 specifically includes: using a point cloud keyframe construction strategy for the pose estimation of the current frame, and creating a new key point cloud sub-graph when the following three conditions are met: 1) More than 20% of the points in the current frame are not being tracked; 2) The rotation angle between the current frame and the last keyframe is greater than 5°; 3) The distance between the current frame and the last keyframe is greater than a threshold, which is 0.3m by default; New local geometric features (plane, line, and cylinder features) are detected in keyframes and attempted to match them with global features. Unmatched features are added as global features, and the accuracy of new matches is verified by geometric consistency check (GCC). Next, the poses of all keyframes are jointly optimized together with all detected global geometric features to minimize the distance error from the point to the model. In the local adjustment, only the sliding window optimization algorithm is used to optimize the keyframes and their observed features within the window, while the poses of the keyframes outside the sliding window remain fixed. In the global adjustment process, feature errors caused by false detections are detected and corrected. During all frame traversals, local geometric features (planes, lines, cylinders) are matched with global features using the aforementioned differentiable RANSAC algorithm to ensure the accuracy of the association between point clouds; After pose matching is completed for all point clouds and their feature descriptions, all point cloud data are fused to obtain a global point cloud map. The entire point cloud mapping process involves transmitting the current frame point cloud data and global point cloud data to the mobile point cloud display software via the device's Wi-Fi module. To achieve real-time point cloud transmission and optimal display on the mobile device, the point cloud data is thinned using a random downsampling method.

9. A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method as claimed in any one of claims 1 to 8.

Images