Machine-based collaborative mapping method in unknown dynamic environments

Abstract

The present invention discloses a method for machine-based collaborative mapping in an unknown dynamic environment. The present invention combines the perception advantages of the wide field of view of aerial intelligent perception equipment and the complete details of ground intelligent perception equipment. First, the three-dimensional point cloud maps constructed by the two are pre-processed by dynamic point cloud filtering, noise point cloud filtering, ceiling and ground point cloud filtering, and out-of-field outlier point cloud group filtering to obtain a high-quality static point cloud map; then, the SAC‑IA algorithm based on the FPFH feature is used for coarse alignment, and then the ICP algorithm is used for fine alignment to improve the matching accuracy while preventing local convergence. The method of the present invention is 8% more accurate than the mapping accuracy of aerial intelligent perception equipment alone, and 51% more accurate than the mapping accuracy of ground intelligent perception equipment alone. It significantly reduces the mapping error, improves the perception accuracy, improves the system robustness, and achieves better perception effects in small and medium-sized indoor complex environments.

Description

Technical Field

[0001] The present invention relates to the field of intelligent perception technology, and in particular to a machine-based collaborative mapping method in an unknown dynamic environment. Background Art

[0002] With the development of the indoor service robot industry, perception and planning technologies have become more mature, and society has put forward more requirements for robot navigation systems used in small and medium-sized indoor scenes. For general indoor robot navigation systems, there are mainly three modules: perception module, positioning module and planning module. Among them, the perception module is a module that intelligent robots use sensors to identify the surrounding environment, sound and obtain their own state information. It is also the basis for the robot to make subsequent plans and decisions. To obtain the perception information of the surrounding environment, it is necessary to use sensors such as single and binocular cameras, lidar, inertial measurement units, odometers, etc. When the robot builds a map in real time in an unknown environment, it is necessary to build a point cloud map because there is no global information in advance. The common solution for obtaining point cloud information and building a point cloud map is SLAM (Simultaneous Localization and Mapping). SLAM generally estimates the inter-frame pose state of the sensor and the position of the map point through front-end tracking methods such as feature point detection or optical flow method. After the back-end optimization receives the pose information measured by the front-end, it uses EKF (Extended Kalman Filter) filter or nonlinear optimization method to perform motion correction and pose optimization to obtain an odometer and map with smaller errors. Although a single robot can currently complete relatively complex tasks, due to problems such as limited hardware performance and single sensors, independent robots are limited in the tasks they can complete, have low efficiency, and low precision, and are unable to adapt to more needs of today's society. Therefore, multi-agent collaboration has become a key research direction.

[0003] Accurate map construction plays a very important role in the subsequent positioning accuracy and planning effectiveness of intelligent robots. Therefore, it is very necessary to adopt a multi-agent collaborative perception solution to fuse the information of multiple sensors and reduce the accumulation of errors. The focus of multi-agent collaborative research is to enable independent intelligent agents to carry more sensors, so that they can share information and allocate tasks with each other, complete complex control tasks or solve complex problems through consultation, coordination, and collaboration, so as to achieve stable and efficient collaborative work effects. At present, multi-intelligent robot collaboration has been applied in various scenarios such as national defense, fire fighting, logistics, detection, and rescue. Robot clusters such as air-to-ground collaboration, air-to-air collaboration, and ground-to-ground collaboration have been developed, which greatly improves work efficiency. The present invention conducts collaborative mapping for indoor small and medium-sized environments, which are characterized by obvious edge geometry, complex internal structure, high obstacle density, and susceptibility to noise and dynamic obstacles. For the traditional method of multi-ground intelligent body collaboration, the system field of view is limited and the mapping efficiency is low. Therefore, the present invention equips the ground-side intelligent body with a collaborative aerial intelligent perception infrastructure platform. The aerial intelligent perception infrastructure platform is equipped with sensors such as cameras and radars, and has a wide field of view. It can speed up the mapping rate and provide more details for the mapping of the ground-side intelligent body, reducing the problems of low accuracy and positioning drift during its navigation. However, the traditional collaborative point cloud registration scheme ignores the processing of dynamic obstacles. In small and medium-sized indoor environments, there are problems of high obstacle density and many dynamic obstacles, which will cause the constructed point cloud map to have more ghosts, affecting visualization and subsequent point cloud registration.

[0004] In summary, the traditional single-agent scheme has limited perception capabilities and low mapping accuracy. The traditional multi-agent collaborative method has limited perception capabilities for small and medium-sized complex dynamic indoor environments. Therefore, there is an urgent need for a new mapping method suitable for small and medium-sized complex dynamic indoor environments. Summary of the invention

[0005] In view of this, the present invention provides a machine-based collaborative mapping method in an unknown dynamic environment, which adopts the joint perception of the machine-based platform, and filters and matches the perception point cloud to achieve accurate mapping of indoor small and medium-sized complex environments with high obstacle density and many dynamic obstacles.

[0006] The machine-based collaborative mapping method in an unknown dynamic environment of the present invention includes:

[0007] Step 1: Use ground intelligent sensing equipment and aerial intelligent sensing equipment to perform three-dimensional perception of the environment, respectively, to obtain ground perception 3D point cloud and aerial perception 3D point cloud;

[0008] Step 2, preprocessing the 3D point cloud obtained in step 1, including filtering out dynamic point clouds, filtering out noise point clouds, filtering out ceiling and ground point clouds, and filtering out outlier point cloud clusters outside the field;

[0009] The improved ERASOR algorithm is used to filter out dynamic point clouds from the 3D point cloud map obtained in step 1; the point cloud in the grid bin is described in a distributed manner in the z direction, divided into multiple height intervals along the height z axis, and the number of point clouds in each height interval is counted respectively; the ratio of the number of point clouds in the corresponding height intervals of the local map and the prior map is calculated; if the ratio of the number of point clouds in a certain height interval is greater than the set value A, it is considered that there are dynamic obstacles in the height interval, and the point cloud in the height interval is deleted;

[0010] Step 3, coarsely matching the preprocessed ground-based 3D point cloud and the aerial-based 3D point cloud;

[0011] Step 4: Perform fine registration based on the coarse registration in step 3 to obtain the transformation relationship between the ground-perceived 3D point cloud and the aerial-perceived 3D point cloud; based on the transformation relationship, convert the two point clouds to the same coordinate system for point cloud fusion to generate a machine-based fusion global map.

[0012] Preferably, in step 1, the ground intelligent sensing mobile device and the aerial intelligent sensing device deploy lidar or camera, and use three-dimensional lidar SLAM or visual SLAM algorithm to achieve three-dimensional perception of the environment; or deploy depth camera and use depth map to achieve three-dimensional perception of the environment.

[0013] Preferably, in the step 2 of dynamic point cloud filtering, the set value A is 0.1-0.3.

[0014] Preferably, in the step 2 of filtering the dynamic point cloud, first mark all the height intervals where dynamic obstacles exist, then perform ground fitting, and finally retain the ground point cloud when deleting the point cloud in the height interval where dynamic obstacles exist. In this way, the ground point cloud can be retained when deleting the dynamic point cloud in the height interval where the ground is located, so that when the ceiling and ground point clouds are deleted later, the ground fitting accuracy is higher and the deletion is more accurate.

[0015] Preferably, in step 2, a voxel filter, a statistical filter or a Gaussian filter is used to filter out noise from the point cloud. A voxel filter is preferred, because the voxel filter can achieve down sampling without destroying the geometric structure of the point cloud itself, is easy to implement, and can effectively filter out noise from the point cloud.

[0016] Preferably, in step 2, the ceiling and ground point clouds are filtered out using methods such as random sampling consensus algorithm (RANSAC), iterative densification TIN model PTD (progressive TIN densification) or cloth simulation filter CSF (Cloth Simulation Filter). RANSAC is preferred because the RANSAC algorithm has good robustness and a better model can be obtained as the number of iterations increases.

[0017] Preferably, in step 2, clustering algorithms, region growing segmentation algorithms, etc. are used to segment the point cloud to remove outlier point cloud clusters. Among them, the Euclidean clustering algorithm is preferably used for segmentation. The principle of the Euclidean clustering algorithm is simple and intuitive. The KD-Tree nearest neighbor query algorithm is used to accelerate the Euclidean clustering, and it has a higher operating efficiency when processing smaller-scale data sets.

[0018] Preferably, in step 3, the SAC-IA algorithm based on FPFH features, the random sampling consensus algorithm (RANSAC), the congruent four-point set algorithm (4PCS) and other algorithms are used for point cloud coarse registration. The SAC-IA algorithm based on FPFH features is preferred, which has high computational efficiency and high robustness, has a large tolerance to interference from outliers, and can be applied to different types of point cloud registration.

[0019] Preferably, in step 4, the iterative closest point algorithm ICP, the normal distribution algorithm NDT, the registration algorithm based on the feature descriptor such as SHOT, etc. are used for precise registration. ICP is preferred, as the algorithm has a simple principle, high registration accuracy, and relatively fast registration speed. However, the ICP algorithm has high requirements on the initial pose of the registration point cloud, otherwise it is easy to fall into the local optimum. Therefore, the SAC-IA coarse registration algorithm is selected to provide a relatively accurate initial pose relationship for ICP, which can effectively solve this problem.

[0020] Preferably, when the iterative closest point algorithm ICP is used for precise registration, the important parameter settings in ICP are shown in Table 1:

[0021] Table 1

[0022]

[0023] After the above parameter adjustment, the source point cloud and the target point cloud were successfully matched, and the error after matching was reduced to about 3 cm in the three dimensions of x, y, and z.

[0024] Beneficial effects:

[0025] The present invention targets unknown dynamic scenes in small and medium-sized indoor areas, combines the perception advantages of wide field of view of aerial intelligent perception equipment and complete details of ground intelligent perception equipment, processes and fuses the three-dimensional point cloud maps built by the two respectively, and pre-processes the three-dimensional point cloud before fusion to improve the robustness of perception, including dynamic point cloud filtering, noise point cloud filtering, ceiling and ground point cloud filtering, and out-of-field outlier point cloud group filtering, to avoid interference of point cloud noise, dynamic obstacle point cloud, ground point cloud and ceiling point cloud on subsequent matching; after pre-processing, the interference points in the original point cloud map are removed, the geometric structure of the point cloud map is strengthened, the visibility of the point cloud is improved, and the accuracy and success rate of subsequent registration can be improved. In the process of dynamic point cloud filtering, the improved ERASOR algorithm is used to perform a distributed description of the point cloud in the grid bin in the z direction, and dynamic obstacles are deleted by comparison according to the height interval, which is more suitable for filtering dynamic obstacle trajectories under the cover of static objects in small and complex indoor scenes; when fusing three-dimensional point clouds, coarse registration is performed first to avoid the subsequent fine registration from falling into local convergence. The machine-based collaborative mapping method and system of the present invention has an 8% improvement in accuracy compared to the independent mapping of aerial intelligent sensing devices, and a 51% improvement in accuracy compared to the independent mapping of ground intelligent sensing devices. It gives full play to the advantages of both, optimizes the mapping effect, significantly reduces the mapping error, improves the perception accuracy, enhances the system robustness, and achieves better perception effects in small and medium-sized indoor complex environments. The framework of the present invention can be applied in many scenarios, such as collaborative perception and navigation of service robots in restaurants and hotels, and has very good research prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A flowchart of the method provided by the present invention;

[0027] Figure 2 The flowchart of the improved dynamic obstacle filtering algorithm using the method provided by the present invention;

[0028] Figure 3 A point cloud registration flow chart using the method provided by the present invention;

[0029] Figure 4 A point cloud map collected by an aerial intelligent perception infrastructure platform using the method provided by the present invention;

[0030] Figure 5 A point cloud map collected by a ground intelligent perception robot using the method provided by the present invention;

[0031] Figure 6 It is a global point cloud map fused by the method provided by the present invention. DETAILED DESCRIPTION

[0032] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0033] The present invention provides a machine-based collaborative mapping method in an unknown dynamic environment, which can realize robust and accurate environmental perception of machine-based collaboration in small and medium-sized indoor scenes. Among them, "machine" refers to a ground intelligent perception robot, and "base" refers to an aerial intelligent perception infrastructure platform. The present invention uses an aerial intelligent perception infrastructure platform and a ground intelligent perception robot to be equipped with three-dimensional laser radars respectively, and uses a SLAM algorithm to realize three-dimensional information perception of an unknown environment, and constructs their own point cloud maps respectively. Since the aerial intelligent perception infrastructure platform has a wide field of view but rough mapping details, while the ground intelligent perception robot has a slow mapping and limited field of view but complete mapping details. Therefore, the present invention uses point cloud processing and fusion algorithms to filter and filter out dynamic obstacles on the point cloud maps constructed by the two, and then aligns and fuses them, so as to achieve the coordinate unification of the machine-based point cloud map and the improvement of the map detail area, and finally generate a global map with a small error.

[0034] The mapping method of the present invention is divided into three parts: point cloud map acquisition, point cloud map processing and machine-based point cloud map fusion. The method block diagram is shown in Figure 1 As shown, the specific steps are as follows:

[0035] Step 1: Point cloud map collection:

[0036] The present invention builds a machine-based collaborative mapping system in an unknown dynamic environment, and uses ground intelligent sensing equipment and aerial intelligent sensing equipment to perform three-dimensional perception of the environment, respectively, to obtain ground perception 3D point cloud and aerial perception 3D point cloud. Among them, the ground intelligent sensing mobile device and the aerial intelligent sensing device can deploy laser radar or camera, and use three-dimensional laser radar SLAM or visual SLAM algorithm to achieve three-dimensional perception of the environment; they can also deploy depth cameras and use depth maps to achieve three-dimensional perception of the environment; or other devices can be used for three-dimensional perception.

[0037] In this embodiment, a laser radar is deployed on the machine-based platform, and a three-dimensional laser radar SLAM algorithm is used to perceive the environment in three dimensions, and the perception information is stored in the form of a 3D point cloud map for subsequent fusion processing. Specifically, the machine-based intelligent platform is equipped with a Velodyne 16-line laser radar, an Intel NUC11PHKi7C processor, and a three-dimensional SLAM algorithm A-LOAM is deployed to achieve three-dimensional perception of the surrounding environment. The aerial intelligent perception infrastructure platform uses an aerial track that can move in a straight line within a fixed range, and the ground intelligent perception robot uses a Mecanum wheel car, and the radar is fixed to the chassis of the smart car for point cloud collection.

[0038] Step 2: Dynamic point cloud filtering: Since dynamic obstacles will cause "ghosts" in the point cloud map, which not only reduces the visualization of the point cloud, but also greatly interferes with the subsequent point cloud registration and fusion, the dynamic obstacle filtering algorithm ERASOR based on grid occupancy difference can be used to process the point cloud map. This algorithm calls the point cloud map with dynamic points obtained by conventional SLAM means a prior map. The prior map is centered on the position of each frame of the radar and divides the local map into a certain radius. , the original radar point cloud of each frame in the map is recorded as , the algorithm is and Divide the grid (bin) and use the two-dimensional descriptor R-POD to describe the distribution of point clouds in the bin. and The distribution of point clouds of bins at the same position in The height difference of the point cloud in the bin is much smaller than The height difference of the point cloud in the corresponding bin, the point cloud in the bin will be judged as a dynamic object.

[0039] Since the 2D descriptor of the ERASOR algorithm only takes the simple vertical height difference of the point cloud to identify dynamic obstacles, it is more suitable for large open scenes, but it cannot effectively handle point cloud noise, dynamic obstacle trajectory filtering under static object occlusion, etc. Therefore, for small and complex indoor scenes, the present invention improves the ERASOR algorithm. The algorithm flow chart is as follows: Figure 2 The algorithm steps are:

[0040] (1) First, according to the experimental scene scale, The point cloud in the range of -1m to 3m in the longitudinal direction is divided into grid bins along the azimuth and radial directions.

[0041] (2) The point cloud in the bin is partitioned in the z-axis direction and the number of point clouds in each height interval is counted.

[0042] In this embodiment, the space from -1m to +3m is divided into intervals of (-1m, 0m), (0m, 1m), (1m, 2m), and (2m, 3m), and the number of point clouds in each interval in the corresponding bins of the prior map and the local map is counted respectively, as shown in formulas (1) and (2), where represent Moment, The height of the bin is The number of point clouds in the interval;

[0043] (1)

[0044] (2)

[0045] (3) If there is an interval in the corresponding bin that satisfies the condition shown in formula (3), that is, the ratio of the number of point clouds in the corresponding interval of the local map to the number of point clouds in the corresponding interval of the prior map is less than the set threshold A, it is determined that there is a dynamic obstacle in the corresponding height area of ​​the bin (since our goal is to optimize the global map, the situation where the number of point clouds in the prior map in the same bin is greater than the number of point clouds in the local map is not considered). The set threshold A is generally 0.1~0.3. In this embodiment, A is 0.2. In formula (3) is the local map at time t The height of the bin is The number of point clouds in the interval, is the local map at time t The height of the bin is The number of point clouds in the interval.

[0046] (3)

[0047] (4) Delete the point cloud of the height interval corresponding to the bin where there are dynamic obstacles.

[0048] When the height interval to be deleted includes the ground, directly deleting all point clouds in the height interval will delete the ground point cloud together, resulting in potholes in the point cloud map; in order to improve the accuracy of ground fitting when filtering out the ceiling and ground point clouds later, the ground can be fitted first in this step, and the points above the ground plane can be removed as dynamic points, retaining the ground point cloud. Specifically, the height intervals that meet condition (3) can be marked first, and after traversing and querying all bins in the area of ​​interest, the ground can be fitted according to the marked dynamic area through PCA (Principle Component Analysis) and the dynamic point cloud in the marked interval other than the ground can be deleted.

[0049] By converting the descriptor of the ERASOR algorithm from a single height difference to a distribution description, the impact of point cloud noise can be reduced, the filtering effect of dynamic obstacles can be improved, and it is more suitable for small and medium-sized indoor environments.

[0050] Step 3: Point cloud filtering: The static point cloud map obtained in step 2 has poor visibility and contains interfering noise data, which will affect the subsequent registration results. Therefore, it is necessary to filter the point cloud to remove the noise data and solve the problems of large point cloud data volume and uneven point cloud data density. Voxel filtering, statistical filtering, Gaussian filtering, etc. can be used to filter the point cloud.

[0051] This embodiment uses a voxel filter to filter the point cloud, divides the space according to the set voxel grid size, and uses the centroid of all points in the grid to replace all points in the grid to remove noise data, solve the problem of large amount of point cloud data being difficult to process and uneven density of point cloud data, and improve the visibility of the map and subsequent matching effects.

[0052] Step 4: Ground segmentation: The ground point cloud and ceiling point cloud will interfere with subsequent matching, so they need to be removed to obtain a static point cloud map with a high degree of visualization. RANSAC, PTD, CSF and other algorithms can be used to segment the image, extract the ceiling and ground point clouds, and then filter out the ceiling and ground point clouds.

[0053] This embodiment performs ground segmentation processing on point clouds based on the random sampling consensus algorithm (RANSAC). First, according to the model characteristics, some points in the data set are selected as inliers, and the mathematical model of these points is calculated. According to the set threshold, other points in the data set are calculated to see whether they meet the mathematical model. Points less than the threshold are classified as inliers; then repeated iterations are performed. If the iterative model has more inliers, it means that the iterative model is better than the existing model. Then the iterative model is used to replace the existing model and the iteration continues. Otherwise, it will be eliminated due to too few inliers; until the exit condition of the iteration is met (a certain number of iterations is reached or the number of inliers meets the set value), the loop is exited.

[0054] After removing the ceiling and ground point clouds, a static point cloud map with a higher degree of visualization can be obtained.

[0055] Step 5, specific point cloud area extraction: Due to the scanning characteristics of the laser radar, there will be a phenomenon of outlier point cloud clustering outside the site in the point cloud map. At the same time, since the small and medium-sized indoor environments targeted by the present invention have obvious advantages in boundary geometric features, it is necessary to extract specific areas from the point cloud map, remove outlier point cloud clusters outside the site, and improve the subsequent registration accuracy. Clustering algorithms, regional growth segmentation algorithms, etc. can be used to segment the point cloud to remove outlier point cloud clusters outside the site. This embodiment uses the Euclidean clustering algorithm to remove outlier point cloud clusters outside the site. The Euclidean clustering algorithm uses the Euclidean distance to measure the degree of intimacy between points. According to the Euclidean distance, the point cloud less than the threshold is regarded as a cluster. In order to accelerate the clustering algorithm, a neighbor query algorithm based on KD-Tree is used to divide the space into two parts. The data points in the space are placed on the left and right subtrees of the binary tree respectively. Through a recursive processing method, the subtrees of each data point are processed, and the KD-Tree is successfully constructed. After the KD-Tree is constructed, the search algorithm can be executed to query the nearest neighbor points and finally remove the outlier point cloud clusters.

[0056] Step 6: Rough registration of point cloud map:

[0057] In order to avoid the fine registration being limited to local convergence, a coarse registration is added before the fine registration. The coarse registration can be performed based on the SAC-IA algorithm based on the FPFH feature, the random sampling consensus algorithm (RANSAC), the congruent four-point set algorithm (4PCS) and other algorithms. In this embodiment, the SAC-IA coarse registration algorithm based on the FPFH feature is added before the fine registration. The overall process flow chart of the point cloud registration is as follows: Figure 3 In the SAC-IA rough registration algorithm based on FPFH features, by finding corresponding points with similar FPFH features in two point clouds, the rigid body transformation matrix between the corresponding points is calculated. The transformation matrix minimizes the error function. The specific operation is as follows:

[0058] (1) From the point cloud to be registered n sampling points are selected from the dataset. In order to ensure that the sampled points have different FPFH characteristics, the distance between the two sampling points should be greater than the pre-given minimum distance threshold d.

[0059] (2) In the target point cloud Search and Point Cloud The sampling points in the sample have one or more points with similar FPFH characteristics, and a point is randomly selected from these similar points as In the target point cloud The corresponding points in ;

[0060] (3) Calculate the rigid body transformation matrix between the corresponding points, and then judge the performance of the current registration transformation by solving the "distance error sum" function after the corresponding point transformation. The distance error sum function here is expressed using the Huber penalty function, which is expressed as:

[0061] (4)

[0062] In formula (4) is a predetermined value, For the The final result of the above process is to find a set of optimal transformations among all the transformations to minimize the value of the error function. However, the matrix obtained by this algorithm is not accurate, so it is used as a coarse registration algorithm to provide initial alignment for subsequent fine registration and prevent local convergence of the ICP algorithm.

[0063] Step 7: Fine registration of point cloud map: Based on the rough registration, the iterative closest point algorithm ICP, the normal distribution algorithm NDT, the registration algorithm based on feature descriptors such as SHOT, etc. are used for fine registration.

[0064] This embodiment uses the iterative closest point algorithm (ICP) for precise registration. The best matching result of the machine-based point cloud is obtained through iteration. The final output result is the rotation matrix and translation vector of the two point clouds, as shown in formula (5):

[0065] (5)

[0066] In the above formula is the target point cloud, is the source point cloud, and The rotation matrix and translation vector of the two point clouds are solved respectively. After the transformation relationship is obtained, the two point clouds are transformed into the same coordinate system for point cloud fusion to generate a machine-based fusion global map with small error. The specific process is as follows:

[0067] (1) Initially register the point cloud (source point cloud after coordinate transformation) and , as the initial point set for precise registration;

[0068] (2) Source point cloud Every point in , in the target point cloud Find the closest corresponding point in , as the corresponding point of the point in the target point cloud, forming an initial corresponding point pair;

[0069] (3) Use the direction vector threshold to eliminate incorrect corresponding point pairs;

[0070] (4) Calculate the rotation matrix and translation vectors , so that the mean square error between corresponding point sets is minimized;

[0071] (5) Setting thresholds and the maximum number of iterations , apply the rigid body transformation obtained in the previous step to the source point cloud , get the new point cloud ,calculate and If the iterative error is less than the threshold Or the current iteration number is greater than , the iteration ends, otherwise the initial registration point set is updated to and , and continue to repeat the above steps until the convergence condition is met.

[0072] For different point cloud data, when the ICP algorithm is used for matching, the algorithm parameters have a great influence on the matching results. The ICP matching parameters in this embodiment are shown in Table 1.

[0073] Table 1 Important parameter settings in the ICP algorithm

[0074]

[0075] Figure 4 The figure shows the point cloud map collected by the aerial intelligent perception infrastructure platform before point cloud processing. Figure 5 The figure shows the point cloud map collected by the ground intelligent perception robot before point cloud processing. Figure 6 The point cloud map after machine-based fusion is shown.

[0076] The method provided by the present invention builds maps through two intelligent platforms, which improves the perception efficiency compared with the single intelligent platform map building, and at the same time avoids the problem of missing map details to a large extent. The present invention is verified by simulation experiments and physical experiments, and the error comparison before and after the point cloud map fusion is obtained, as shown in Table 2 to Table 4.

[0077] Table 2 Mapping error of aerial intelligent perception infrastructure platform

[0078]

[0079] Table 3 Mapping error of ground intelligent perception robot

[0080]

[0081] Table 4 Machine-based fusion mapping error

[0082]

[0083] Through experimental comparison, it is verified that the accuracy of fusion mapping is improved compared with that of single-agent mapping, and the visibility is improved; and due to the viewing angle limitations of lidar scanning, the ground intelligent perception robot has a poor reconstruction effect on the environment, especially after filtering, and many feature points are lost; in terms of detail reconstruction, the detail perception ability of the aerial intelligent perception infrastructure platform is poor, but the point cloud fusion solution provided by this machine-based collaborative system can solve the above problems, which is of great significance to improving the system's perception of the environment, and the solution is highly scalable. By increasing the number of "machine" and "base" agents, more efficient and accurate exploration can be achieved.

[0084] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A machine-based collaborative mapping method in an unknown dynamic environment, characterized in that: include: Step 1: Use ground intelligent sensing equipment and aerial intelligent sensing equipment to perform three-dimensional perception of the environment, respectively, to obtain ground perception 3D point cloud and aerial perception 3D point cloud; Step 2, preprocessing the 3D point cloud obtained in step 1, including filtering out dynamic point clouds, filtering out noise point clouds, filtering out ceiling and ground point clouds, and filtering out outlier point cloud clusters outside the field; The improved ERASOR algorithm is used to filter out dynamic point clouds from the 3D point cloud map obtained in step 1; the point clouds in the grid bin are divided into multiple height intervals along the height z axis, and the number of point clouds in each height interval is counted respectively; the ratio of the number of point clouds in the corresponding height intervals of the local map and the prior map is calculated; if the ratio of the number of point clouds in a certain height interval is greater than the set value A, it is considered that there are dynamic obstacles in the height interval, and the point clouds in the height interval are deleted; Step 3, coarsely matching the preprocessed ground-based 3D point cloud and the aerial-based 3D point cloud; Step 4: Perform fine registration based on the coarse registration in step 3 to obtain the transformation relationship between the ground-perceived 3D point cloud and the aerial-perceived 3D point cloud; based on the transformation relationship, convert the two point clouds to the same coordinate system for point cloud fusion to generate a machine-based fusion global map.

2. The method according to claim 1, characterized in that In step 1, the ground intelligent sensing mobile device and the aerial intelligent sensing device deploy lidar or camera, and use three-dimensional lidar SLAM or visual SLAM algorithm to achieve three-dimensional perception of the environment; or deploy depth camera and use depth map to achieve three-dimensional perception of the environment.

3. The method according to claim 1, characterized in that In the step 2 of dynamic point cloud filtering, the set value A is 0.1-0.

3.

4. The method according to claim 1 or 3, characterized in that In the step 2 of filtering the dynamic point cloud, firstly, all the height intervals with dynamic obstacles are marked, then ground fitting is performed, and finally, when the point cloud of the height interval with dynamic obstacles is deleted, the ground point cloud is retained.

5. The method according to claim 1, characterized in that In the step 2, a voxel filter, a statistical filter or a Gaussian filter is used to filter out noise from the point cloud.

6. The method according to claim 1, characterized in that In the step 2, RANSAC, PTD or CSF is used to segment the ground and ceiling, and then the ceiling and ground point clouds are filtered out.

7. The method according to claim 1, characterized in that In step 2, a clustering algorithm or a region growing segmentation algorithm is used to segment the point cloud to remove outlier point cloud clusters.

8. The method according to claim 1, characterized in that In the step 3, the SAC-IA algorithm based on FPFH features, the random sampling consistency algorithm or the congruent four-point set algorithm is used to perform rough registration of the point cloud.

9. The method according to claim 1 or 8, characterized in that In the step 4, the iterative closest point algorithm ICP, the normal distribution algorithm NDT or the registration algorithm based on the SHOT feature descriptor is used for precise registration.

10. The method according to claim 9, characterized in that In step 4, when the iterative closest point algorithm ICP is used for precise registration, the important parameters of the iterative closest point algorithm ICP are set as follows: Set the maximum number of iterations to 2; set the maximum tolerance of the mean of the Euclidean distances of the point pairs before and after two iterations to 1; set the maximum tolerance of the transformation matrix before and after two iterations to 1e-7; set the Euclidean distance of the maximum corresponding point to 20.

Images