Laser radar mobile robot adaptive mapping method and system

Abstract

The application discloses a laser radar mobile robot self-adaptive mapping method and system, which comprises the following steps: obtaining a point cloud data set scanned by a laser radar for a target environment, the point cloud data set comprising N frames of data and IMU measured reference poses between two frames, each frame of data comprising a time stamp and original point cloud data; fusing the first frame of preprocessed point cloud data into a blank global map; obtaining corresponding feature points by performing feature extraction on each frame of preprocessed point cloud data; matching each feature point of the Nth frame with each feature point of the nearest effective frame to obtain a feature point matching pair set; performing pose transformation calculation on the set by using a pose optimization algorithm to obtain a pose estimation result; comparing the pose estimation result with the reference pose; analyzing whether the Nth frame of point cloud data is effective; if the Nth frame of point cloud data is effective, fusing the Nth frame of point cloud data into the global map; taking n as 2-N; obtaining a global point cloud map after the analysis; removing noise points and outliers from the global point cloud map; and projecting the identified obstacle point cloud cluster to a two-dimensional plane to construct a two-dimensional grid map.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of robot technology, in particular to a laser radar mobile robot adaptive mapping method and system, which is suitable for the field of robot environment perception and planning navigation. BACKGROUND

[0002] With the development of robot technology, the research of mobile robots has been paid more and more attention in recent years and has been applied to many industrial and military scenes. While being applied more and more widely, the industry has also put forward higher requirements for the perception and exploration ability of mobile robots in unknown environments. In an unknown environment, how to obtain the position and surrounding information of itself, and expand the perception range of the environment for better planning and navigation of the mobile robot has become a very important direction in the field of mobile robots. The current perception scheme of mobile robots in unknown environments mainly includes two types: a camera-based vision scheme and a point cloud mapping scheme based on a laser radar. The camera-based vision scheme can provide strong surrounding environment perception ability close to human vision, but it is subject to environmental factors such as light, which cannot guarantee stable environmental perception ability; while the point cloud mapping scheme based on a laser radar can obtain a large range of surrounding environment information and self-position information, and is not sensitive to environmental factors. The existing scheme often starts from a specific task and selects different mapping algorithms according to different environments. This method makes the environmental perception effect of the mobile robot strongly related to the environment, which has great limitations. If the environment changes, the algorithm needs to be re-set, otherwise it cannot guarantee good environmental perception effect. At the same time, the current market has a step separation from mapping to two-dimensional grid map generation. This inspires us to develop an adaptive and integrated scheme from point cloud mapping to two-dimensional grid map generation, which can maintain good environmental perception ability in different environments and save the time of two-dimensional grid map generation, which is of great significance to improve the perception and adaptability of the robot. SUMMARY

[0003] The present application provides a laser radar mobile robot adaptive mapping method and system to solve the problems and deficiencies of the prior art.

[0004] The present application solves the above technical problems by the following technical solutions:

[0005] The present application provides a laser radar mobile robot adaptive mapping method, which is characterized by comprising:

[0006] S1, obtaining a point cloud data set collected by scanning with a laser radar for a certain target environment Point cloud data set The two frames of poses between which the N frames of data in time sequence and IMU measurement are obtained as a reference, each frame of data comprising a time stamp and original point cloud data with three-dimensional coordinate information, and N is a positive integer;

[0007] S2, preprocessing each frame of original point cloud data to obtain corresponding preprocessed point cloud data;

[0008] S3, constructing a blank global map, and fusing the first frame of preprocessed point cloud data into the global map;

[0009] S4, extracting features from each frame of preprocessed point cloud data to obtain corresponding point cloud feature points;

[0010] S5, for the nth frame:

[0011] performing similarity matching on each point cloud feature point p cur of the nth frame and each point cloud feature point p pre of the nearest valid frame before to obtain a feature point matching pair set

[0012] adopting a pose optimization algorithm to perform pose transformation calculation on the feature point matching pair set to obtain a pose estimation result, comparing the pose estimation result with the reference pose, and analyzing whether the nth frame of point cloud data is valid, when the difference between the two is less than a preset threshold, the nth frame of point cloud data is fused and updated into the global map, otherwise, the global map is not updated;

[0013] n is sequentially taken as 2-N, and after the Nth frame is analyzed, a final global point cloud map is obtained;

[0014] S6, performing downsampling processing and adaptive density filtering processing on the global point cloud map, eliminating noise points and outliers, identifying potential obstacle point cloud clusters, projecting the obstacle point cloud clusters to a two-dimensional plane, constructing a two-dimensional grid map, and optimizing the two-dimensional grid map to obtain an optimized two-dimensional grid map.

[0015] The application also provides a laser radar mobile robot adaptive mapping system, characterized by comprising a laser radar, an environment mapping algorithm controller and a two-dimensional grid map converter;

[0016] The laser radar is used for acquiring a point cloud data set collected by scanning a target environment with a laser radar The two frames of poses between which the N frames of data in time sequence and IMU measurement are obtained as a reference, each frame of data comprising a time stamp and original point cloud data with three-dimensional coordinate information, and N is a positive integer;

[0017] The environment mapping algorithm controller is used to preprocess each frame of raw point cloud data to obtain the corresponding preprocessed point cloud data, construct a blank global map, merge the first frame of preprocessed point cloud data into the global map, and extract features from each frame of preprocessed point cloud data to obtain the corresponding point cloud feature points.

[0018] The environment mapping algorithm controller is also used for the nth frame:

[0019] The feature points p of each point cloud in the nth frame cur The feature points p of each point cloud in the most recent valid frame pre Perform similarity matching to obtain a set of feature point matching pairs between the two frames.

[0020] Using pose optimization algorithms to match feature point pairs Perform pose transformation calculations to obtain pose estimation results, compare the pose estimation results with the reference pose, and analyze whether the point cloud data of the nth frame is valid. If the difference between the two is less than a preset threshold, the point cloud data of the nth frame is valid and then the point cloud data of the nth frame is fused and updated to the global map; otherwise, it is invalid and the global map is not updated.

[0021] n is taken as 2-N. After the Nth frame is analyzed, the final global point cloud map is obtained.

[0022] The two-dimensional raster map converter is used to perform downsampling and adaptive density filtering on the global point cloud map, remove noise points and outliers, identify potential obstacle point cloud clusters, project the obstacle point cloud clusters onto a two-dimensional plane to construct a two-dimensional raster map, and optimize the two-dimensional raster map to obtain an optimized two-dimensional raster map.

[0023] The positive and progressive effects of this invention are as follows:

[0024] This invention realizes point cloud mapping based on LiDAR, which can acquire environmental data in real time using LiDAR and accurately construct point cloud maps. The innovation lies in the accurate construction and generation of point cloud maps, which provides a solid foundation for obtaining accurate two-dimensional raster maps in the future.

[0025] This invention realizes point cloud data optimization and raster map generation. It can filter and optimize existing point cloud data and generate highly available raster maps, providing a good data foundation for subsequent operations such as path planning.

[0026] In summary, this invention provides a comprehensive environmental perception solution, which effectively supports the perception and planning of intelligent mobile robots in unknown environments, and has broad application prospects and technical value. Attached Figure Description

[0027] Figure 1 A flow chart of a laser radar mobile robot adaptive mapping method according to an embodiment of the present application.

[0028] Figure 2 A block diagram of a laser radar mobile robot adaptive mapping system according to an embodiment of the present application. DETAILED DESCRIPTION

[0029] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0030] As shown in Figure 1 The present embodiment provides a laser radar mobile robot adaptive mapping method, which enables the robot to rapidly perceive and establish a two-dimensional grid map in an unknown environment, comprising the following steps:

[0031] Step 101: obtaining a point cloud data set collected by laser radar scanning for a target environment (such as a lawn environment) Point cloud data set including N frames of data in time sequence and a pose between two frames as a reference (i.e. reference pose, including a rotation transformation as a reference and a translation component as a reference) measured by an IMU (i.e. inertial sensor), each frame of data including a time stamp and original point cloud data with three-dimensional coordinate information, N is a positive integer, and the laser radar is a pseudo multi-line laser radar.

[0032] Step 102: pre-processing each frame of original point cloud data to obtain corresponding pre-processed point cloud data.

[0033] In this step, each frame of original point cloud data is pre-processed, including removing noise points, downsampling and converting to a unified coordinate system, to generate corresponding pre-processed point cloud data.

[0034] Step 103: constructing a blank global map and fusing the first frame of pre-processed point cloud data into the global map. In this method, the first frame of pre-processed point cloud data is default effective point cloud data.

[0035] Step 104: extracting features from each frame of pre-processed point cloud data to obtain corresponding point cloud feature points.

[0036] In this step, feature extraction is performed on each frame of preprocessed point cloud data to obtain the corresponding point cloud feature points. The point cloud feature points are cluster centers based on the point cloud density distribution and geometric structure. The cluster centers are obtained by performing clustering operations on the preprocessed point cloud data to divide it into multiple clusters and extracting the geometric center point of each cluster.

[0037] Step 105, for the nth frame:

[0038] The feature points p of each point cloud in the nth frame cur The feature points p of each point cloud in the most recent valid frame pre Perform similarity matching to obtain a set of feature point matching pairs between the two frames.

[0039] Using pose optimization algorithms to match feature point pairs The pose transformation calculation is performed to obtain the pose estimation result. The pose estimation result is compared with the reference pose to analyze whether the point cloud data of the nth frame is valid. If the difference between the two is less than a preset threshold, the point cloud data of the nth frame is valid and then the point cloud data of the nth frame is fused and updated into the global map. Otherwise, if it is invalid, the global map is not updated.

[0040] n is taken as 2-N. After the Nth frame is analyzed, the final global point cloud map is obtained.

[0041] For example: when the feature points p of the point cloud in the second frame cur The feature points p of each point cloud in frame 1 pre Perform similarity matching to obtain a set of feature point matching pairs between the two frames. Using pose optimization algorithms to match feature point pairs Pose transformation calculations are performed to obtain pose estimation results. These results are compared with the baseline pose to analyze the validity of the second frame point cloud data. If the second frame point cloud data is valid, it is fused and updated into the global map. Then, when the feature points p of the third frame point cloud are... cur Compared with the feature points p of each point cloud in the second frame pre Similarity matching is performed...; if the point cloud data of the second frame is invalid, the point cloud data of the second frame will not be merged and updated into the global map, that is, the global map will not be updated using the point cloud data of the second frame. Since the point cloud data of the second frame is invalid, the feature points p of each point cloud in the third frame should be used next. cur The feature points p of each point cloud in frame 1 pre Perform similarity matching...

[0042] In step 105, the feature points p of each point cloud in the nth frame are... cur The feature points p of each point cloud in the most recent valid frame preSimilarity matching calculation is performed, when the similarity calculated between the a-th point cloud feature point of the n-th frame and the b-th point cloud feature point of the nearest effective frame reaches or exceeds a preset similarity, it is determined that the two point cloud feature points are matched as a feature point matching pair, and based on this, a set of feature point matching pairs of the two frames is obtained

[0043] A matching relationship is determined by calculating the similarity measure between feature points, and when the similarity reaches a set value, it is considered to be matched. For example, the Euclidean distance is used to calculate the similarity between the point cloud feature point p cur and the point cloud feature point p pre , and the matching relationship is determined according to the distance size. The specific solving problem is:

[0044] The Euclidean distance between the point cloud feature point p cur and the point cloud feature point p pre is:

[0045]

[0046] wherein, and respectively represent the a-th feature point in the n-th frame and the b-th feature point in the nearest effective frame before.

[0047] In step 105, the pose optimization algorithm is used for the n-th frame to calculate the pose transformation of the feature point matching pair set to obtain the pose estimation result:

[0048] Each feature point matching pair in the feature point matching pair set is given a weight based on the local density of the point cloud.

[0049] The calculation method of the weight is:

[0050]

[0051] wherein, W i is the weight of the i-th feature point matching pair in the feature point matching pair set , p is the point cloud feature point of the n-th frame in the i-th feature point matching pair, is the point density around p , σ is a decay coefficient for controlling the decay amplitude (the larger σ is, the slower the weight of low density decreases), and exp(·) is an exponential function.

[0052] A weighted registration error loss function Loss is constructed based on each feature point matching pair and its weight.

[0053] The calculation formula of the weighted registration error loss function Loss is:

[0054]

[0055] wherein, is the point cloud feature point of the nearest valid frame in the ith feature point matching pair, R is a rotation transformation, and t is a translation component.

[0056] The weighted registration error loss function Loss is taken as an optimization target for the rotation transformation and the translation component respectively, a rotation transformation loss function Loss(R) and a translation component loss function Loss(t) are constructed, and the least square method is used for optimization. Due to the priority relationship between the rotation transformation and the translation component (the rotation transformation solution is not affected by the translation component), the optimal rotation transformation R is solved first * , and then the optimal translation component t is solved * .

[0057] The calculation formula of the rotation transformation loss function Loss(R) is:

[0058]

[0059] The calculation formula of the translation component loss function Loss(t) is:

[0060]

[0061] The optimal rotation transformation is compared with the reference rotation transformation, and the optimal translation component is compared with the reference translation component. When the difference between the optimal rotation transformation and the reference rotation transformation is less than a first preset threshold, and the difference between the optimal translation component and the reference translation component is less than a second preset threshold, it is analyzed that the nth frame of point cloud data is valid, and the nth frame of point cloud data is fused and updated into the global map, otherwise it is analyzed that the nth frame of point cloud data is invalid, and the global map is not updated.

[0062] Step 106, performing down-sampling processing and adaptive density filtering processing on the global point cloud map, eliminating noise points and outliers, identifying potential obstacle point cloud clusters, and projecting the obstacle point cloud clusters to a two-dimensional plane to construct a two-dimensional grid map, and performing optimization processing on the two-dimensional grid map to obtain an optimized two-dimensional grid map.

[0063] The specific implementation process of this step is: for the global point cloud map, performing down-sampling processing and adaptive density filtering processing, eliminating noise points and outliers, and generating target point cloud data.

[0064] Based on the target point cloud data, potential obstacle point cloud clusters are identified through a K-Means clustering algorithm.

[0065] The identified obstacle point cloud clusters are projected onto an XY plane to generate two-dimensional point cloud data.

[0066] The two-dimensional grid map is generated according to two-dimensional point cloud data, and optimization processing is performed on the two-dimensional grid map, including removing isolated occupied grids, smoothing the map edge, and a morphological operation algorithm, to obtain an optimized two-dimensional grid map.

[0067] The adaptive density filtering processing includes:

[0068] Based on the preset search radius, the number of neighbor points of each point cloud point in the global point cloud map within the preset search radius range is counted.

[0069] Based on the number of neighbor points, a density-aware window weight function is used to dynamically determine the weighting coefficient of the corresponding point cloud point.

[0070] The calculation formula of the density-aware window weight function is:

[0071]

[0072] Where w(p) is the weighting coefficient of the point cloud point p, c is the preset search radius, k is the weight of the filtering window, and d(p) is the number of neighbor points within the preset search radius around the point cloud point p.

[0073] Within the preset search radius, the statistical characteristics of each point cloud point are calculated based on the local statistical characteristic calculation formula, and statistical filtering is performed using the statistical characteristics of each point cloud point to obtain a filtered global point cloud map.

[0074] The local statistical characteristic calculation formula is:

[0075]

[0076] Where F(p) is the local statistical characteristic value of the point cloud point p, R(p) is a circle with the point cloud point p as the center and the preset search radius c, q is a neighbor point of the point cloud point p within R(p), N R(p) is the number of point cloud points in R(p), w(q) is the weighting coefficient of the point cloud point q, and d(q) is the number of neighbor points within the preset search radius of the point cloud point q.

[0077] The distance mean between the current point cloud point in the filtered global point cloud map and all neighbor points within its preset search radius range is calculated, and a dynamic distance threshold is generated based on the distance mean.

[0078] Based on the dynamic distance threshold, it is determined whether the current point cloud point is an outlier. If the distance between the current point cloud point and any neighbor point within the preset search radius range is greater than the dynamic distance threshold, the current point cloud point is an outlier, and the current point cloud point is removed. Otherwise, the current point cloud point is not an outlier, and the current point cloud point is retained.

[0079] The dynamic distance threshold calculation formula is:

[0080] T = a * mean(|p' - q'|) q’∈R(p’)

[0081] Wherein, T is the dynamic distance threshold of the point cloud point p' in the filtered global point cloud map, a is the outlier coefficient, |p'-q'| is the distance absolute value of the point cloud point q' to the point cloud point p' in the filtered global point cloud map, mean(·) is the average function, R(p') is a circle with the point cloud point p' as the center and a preset search radius c, q'∈R(p') is the neighbor point of the point cloud point p' in R(p').

[0082] In addition, in step 101, the point cloud data set Also includes the initial coordinates of the laser radar when collecting any frame of original point cloud data and the initial orientation of the laser radar.

[0083] After updating the global map and not updating the global map in step 105: based on the initial coordinates and the initial orientation of the laser radar, it is detected whether there is a closed loop, i.e. the robot returns to the area that has been visited before, if the closed loop is detected, a closed loop optimization algorithm based on the beam adjustment method is used to eliminate the cumulative error. Wherein, detecting whether there is a closed loop and eliminating the cumulative error when the closed loop is present is the prior art, which will not be repeated here.

[0084] As Figure 2 shown, the application also provides a laser radar mobile robot adaptive mapping system, comprising a laser radar, an environment mapping algorithm controller and a two-dimensional grid map converter; the environment mapping algorithm controller and the two-dimensional grid map converter are integrated on an edge computing industrial computer, and the edge computing industrial computer is installed on the mobile robot.

[0085] The laser radar is used to obtain a point cloud data set The point cloud data set includes N frames of data in time sequence and the pose between two frames as a reference obtained by IMU measurement, each frame of data includes a time stamp and original point cloud data with three-dimensional coordinate information, and N is a positive integer.

[0086] The environment mapping algorithm controller is used to pre-process each frame of original point cloud data to obtain corresponding pre-processed point cloud data, construct a blank global map, fuse the first frame of pre-processed point cloud data into the global map, and extract features from each frame of pre-processed point cloud data to obtain corresponding point cloud feature points.

[0087] The environment mapping algorithm controller is also used for the nth frame: similarity matching between each point cloud feature point p cur of the nth frame and each point cloud feature point p pre of the previous nearest valid frame to obtain a feature point matching pair set of the two frames Adopting a pose optimization algorithm to a feature point matching pair set A pose transformation calculation is performed to obtain a pose estimation result, the pose estimation result is compared with a reference pose, and it is analyzed whether the nth frame of point cloud data is effective, when the difference between the two is less than a preset threshold, the nth frame of point cloud data is fused and updated into the global map, otherwise, the global map is not updated; n is taken as 2-N in turn, after the Nth frame is analyzed, the final global point cloud map is obtained.

[0088] The two-dimensional grid map converter is used for performing down-sampling processing and adaptive density filtering processing on the global point cloud map, eliminating noise points and outliers, identifying potential obstacle point cloud clusters, and projecting the obstacle point cloud clusters to a two-dimensional plane to construct a two-dimensional grid map, and performing optimization processing on the two-dimensional grid map to obtain an optimized two-dimensional grid map.

[0089] The application makes full use of the point cloud information obtained by laser radar scanning of the target environment, cooperates with an efficient and adaptable point cloud map construction algorithm, obtains a point cloud map, and performs short-time point cloud filtering and two-dimensional projection, realizes integrated generation from point cloud data to a two-dimensional grid map, and has generalization, can be applied to various robot carriers and different environments, and improves the perception ability of the robot. The application realizes low-cost radar to obtain high adaptability and excellent environment mapping effect through laser radar and point cloud mapping algorithm for point cloud map construction. The grid map constructor based on down-sampling filtering and high-dimensional plane projection realizes integrated generation from a point cloud map to a two-dimensional grid map, brings higher mapping efficiency to the system, and the adjustable construction method makes the system have stronger adjustability and expandability.

[0090] Although the specific embodiments of the application are described above, those skilled in the art should understand that these are only illustrative, the protection scope of the application is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of the application, and these changes and modifications all fall within the protection scope of the application.

Claims

1. An adaptive mapping method for a lidar-based mobile robot, characterized in that, include: S1. Obtain the point cloud dataset collected by LiDAR scanning for a specific target environment. Point cloud dataset It includes N frames of data in time sequence and the pose between two frames measured by IMU as a reference. Each frame of data includes a timestamp and raw point cloud data with three-dimensional coordinate information. N is a positive integer. S2. Preprocess each frame of raw point cloud data to obtain the corresponding preprocessed point cloud data. S3. Construct a blank global map and merge the preprocessed point cloud data of the first frame into the global map; S4. Perform feature extraction on each frame of preprocessed point cloud data to obtain the corresponding point cloud feature points. S5, For the nth frame: The feature points p of each point cloud in the nth frame cur The feature points p of each point cloud in the most recent valid frame pre Perform similarity matching to obtain a set of feature point matching pairs between the two frames. Using pose optimization algorithms to match feature point pairs The pose transformation calculation is performed to obtain the pose estimation result. The pose estimation result is compared with the reference pose to analyze whether the point cloud data of the nth frame is valid. If the difference between the two is less than the preset threshold, the point cloud data of the nth frame is valid and then the point cloud data of the nth frame is fused and updated to the global map. Otherwise, if it is invalid, the global map is not updated. n is taken as 2-N. After the Nth frame is analyzed, the final global point cloud map is obtained. S6. Perform downsampling and adaptive density filtering on the global point cloud map to remove noise points and outliers, identify potential obstacle point cloud clusters, project the obstacle point cloud clusters onto a two-dimensional plane to construct a two-dimensional raster map, optimize the two-dimensional raster map to obtain the optimized two-dimensional raster map.

2. The adaptive mapping method for a LiDAR mobile robot as described in claim 1, characterized in that, In S4, feature extraction is performed on each frame of preprocessed point cloud data to obtain the corresponding point cloud feature points. The point cloud feature points are cluster centers based on the point cloud density distribution and geometric structure. The cluster centers are obtained by performing clustering operations on the preprocessed point cloud data, dividing it into multiple clusters, and extracting the geometric center point of each cluster.

3. The adaptive mapping method for a LiDAR mobile robot as described in claim 1, characterized in that, In S2, each frame of raw point cloud data is preprocessed, including removing noise points, downsampling, and transforming to a unified coordinate system to generate the corresponding preprocessed point cloud data.

4. The adaptive mapping method for a LiDAR mobile robot as described in claim 1, characterized in that, In S5, the feature points p of each point cloud in the nth frame are... cur The feature points p of each point cloud in the most recent valid frame pre Similarity matching calculations are performed. When the similarity calculated between the a-th point cloud feature point in the n-th frame and the b-th point cloud feature point in the most recent valid frame reaches or exceeds a preset similarity, these two point cloud feature points are confirmed to match as a feature point matching pair. Based on this, a set of feature point matching pairs between two frames is obtained.

5. The adaptive mapping method for a LiDAR mobile robot as described in claim 1, characterized in that, In S5, a pose optimization algorithm is used to match the set of feature points for the nth frame. Perform pose transformation calculations to obtain pose estimation results: Matching pairs of feature points Each feature point matching pair is assigned a weight based on the local density of the point cloud; A weighted registration error loss function Loss is constructed based on each feature point matching pair and its weight. The weighted registration error loss function Loss is optimized with rotation transformation and translation components as optimization objectives. The rotation transformation loss function Loss(R) and the translation component loss function Loss(t) are constructed and optimized using the least squares method. The optimal rotation transformation and the optimal translation component are calculated respectively. In S1, the pose used as a reference includes the rotation transformation and translation components used as a reference. The optimal rotation transformation is compared with the reference rotation transformation, and the optimal translation component is compared with the reference translation component. When the difference between the optimal rotation transformation and the reference rotation transformation is less than the first preset threshold, and the difference between the optimal translation component and the reference translation component is less than the second preset threshold, the point cloud data of the nth frame is analyzed as valid, and the point cloud data of the nth frame is fused and updated into the global map; otherwise, the point cloud data of the nth frame is analyzed as invalid, and the global map is not updated.

6. The adaptive mapping method for a LiDAR mobile robot as described in claim 5, characterized in that, Weight calculation method: Among them, W i For the set of feature point matching pairs The weight of the matching pair of the i-th feature point in the middle. To match the point cloud feature points of the nth frame in the pair for the i-th feature point, for The surrounding point density, σ is the attenuation coefficient used to control the attenuation amplitude, and exp(·) is the exponential function; The formula for calculating the weighted registration error loss function Loss is as follows: in, For the i-th feature point, match the point cloud feature points of the nearest valid frame in the matching pair, where R is the rotation transformation and t is the translation component; The formula for calculating the rotation transformation loss function Loss(R) is as follows: The formula for calculating the translation component loss function Loss(t) is as follows: Where R* is the optimal rotation transformation and the optimal translation component is denoted as t*.

7. The adaptive mapping method for a LiDAR mobile robot as described in claim 1, characterized in that, In S1, point cloud dataset It also includes the initial coordinates and initial orientation of the lidar when acquiring any frame of raw point cloud data; After updating the global map and not updating the global map in S5: Based on the initial coordinates and orientation of the LiDAR, the system detects whether a closed loop exists, i.e., the robot has returned to a previously visited area. If a closed loop is detected, a closed loop optimization algorithm based on the bundle adjustment method is used to eliminate the accumulated error.

8. The adaptive mapping method for a LiDAR mobile robot as described in claim 1, characterized in that, S6 include: For the global point cloud map, downsampling and adaptive density filtering are performed to remove noise points and outliers, generating target point cloud data; Based on target point cloud data, potential obstacle point cloud clusters are identified using the K-Means clustering algorithm; The identified obstacle point cloud clusters are projected onto the XY plane to generate two-dimensional point cloud data. A two-dimensional raster map is generated from two-dimensional point cloud data. The two-dimensional raster map is then optimized, including removing isolated occupied graticules, smoothing map edges, and applying morphological operation algorithms, to obtain an optimized two-dimensional raster map.

9. The adaptive mapping method for a LiDAR mobile robot as described in claim 8, characterized in that, Adaptive density filtering includes: Based on the preset search radius, count the number of neighboring points of each point cloud point in the global point cloud map within the preset search radius. Based on the number of neighboring points, a density-aware window weighting function is used to dynamically determine the weighting coefficients of the corresponding point cloud points; Within a preset search radius, the statistical features of each point cloud point are calculated based on the local statistical feature calculation formula. The statistical features of each point cloud point are then used for statistical filtering to obtain a filtered global point cloud map. Calculate the average distance between the current point in the filtered global point cloud map and all its neighboring points within a preset search radius, and generate a dynamic distance threshold based on this average distance. The current point cloud point is determined based on a dynamic distance threshold. If the distance between the current point cloud point and any neighboring point within the preset search radius is greater than the dynamic distance threshold, then the current point cloud point is an outlier and is removed. Otherwise, the current point cloud point is not an outlier and is retained.

10. The adaptive mapping method for a LiDAR mobile robot as described in claim 9, characterized in that, The formula for calculating the density-aware window weight function is: Where w(p) is the weighting coefficient of point p in the point cloud, c is the preset search radius, k is the weight of the filter window, and d(p) is the number of neighboring points within the preset search radius around point p in the point cloud. The formula for calculating local statistical characteristics is: Where F(p) is the local statistical feature value of point p in the point cloud, R(p) is a circle with point p as the center and a preset search radius of c, q∈R(p) are the neighbor points of point p in R(p), and N R(p) R(p) represents the number of points in the point cloud, w(q) represents the weighting coefficient of point q, and d(q) represents the number of neighboring points of point q within the preset search radius. The formula for calculating the dynamic distance threshold is: T=α·mean(|p’-q’|) q’∈R(p’) Where T is the dynamic distance threshold of point p' in the filtered global point cloud map, α is the outlier coefficient, |p'-q'| is the absolute distance from point q' to point p' in the filtered global point cloud map, mean(·) is the average value function, R(p') is a circle with point p' as the center and a preset search radius of c, and q'∈R(p') are the neighbor points of point p' in R(p').

11. A lidar mobile robot adaptive mapping system, characterized in that, This includes LiDAR, environmental mapping algorithm controller, and 2D raster map converter; The lidar is used to acquire point cloud datasets collected by lidar scanning of a specific target environment. Point cloud dataset It includes N frames of data in time sequence and the pose between two frames measured by IMU as a reference. Each frame of data includes a timestamp and raw point cloud data with three-dimensional coordinate information. N is a positive integer. The environment mapping algorithm controller is used to preprocess each frame of raw point cloud data to obtain the corresponding preprocessed point cloud data, construct a blank global map, merge the first frame of preprocessed point cloud data into the global map, and extract features from each frame of preprocessed point cloud data to obtain the corresponding point cloud feature points. The environment mapping algorithm controller is also used for the nth frame: The feature points p of each point cloud in the nth frame cur The feature points p of each point cloud in the most recent valid frame pre Perform similarity matching to obtain a set of feature point matching pairs between the two frames. Using pose optimization algorithms to match feature point pairs The pose transformation calculation is performed to obtain the pose estimation result. The pose estimation result is compared with the reference pose to analyze whether the point cloud data of the nth frame is valid. If the difference between the two is less than the preset threshold, the point cloud data of the nth frame is valid and then the point cloud data of the nth frame is fused and updated to the global map. Otherwise, if it is invalid, the global map is not updated. n is taken as 2-N. After the Nth frame is analyzed, the final global point cloud map is obtained. The two-dimensional raster map converter is used to perform downsampling and adaptive density filtering on the global point cloud map, remove noise points and outliers, identify potential obstacle point cloud clusters, project the obstacle point cloud clusters onto a two-dimensional plane to construct a two-dimensional raster map, and optimize the two-dimensional raster map to obtain an optimized two-dimensional raster map.

Images