Heterogeneous multi-robot air-ground collaborative material distribution method

Abstract

The invention discloses a heterogeneous multi-robot air-ground collaborative material distribution method, and belongs to the technical field of multi-unmanned system task allocation. Comprising the following steps: determining an air-ground collaborative minimum coverage area unit, carrying out partitioning on the air-ground collaborative minimum coverage area unit, and selecting a collection and distribution center for material distribution and external placement points; carrying out optimal closed-loop sequence solution on a delivery route between the delivery point and the material delivery area to obtain a shortest path with road information; incremental updating and multi-source information fusion mapping are carried out on a delivery area through real-time collaborative environment perception of the unmanned aerial vehicle and the unmanned vehicle; performing optimal path planning based on the distribution task; based on the planned optimal path, a distribution task is completed; and through communication of the Internet of Things, a return route is planned, and a next distribution task is prepared. The unmanned vehicle and the unmanned aerial vehicle are combined, air-ground cooperative three-dimensional autonomous material distribution is achieved, and complementary advantages of large carrying capacity and long cruising ability of the unmanned vehicle and high maneuverability and three-dimensional distribution of the unmanned aerial vehicle are achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of multi-unmanned system task allocation, and specifically relates to a heterogeneous multi-robot air-to-ground collaborative material distribution method. Background Art

[0002] The development of unmanned delivery is of great significance to logistics and distribution. In the process of earthquake relief, flood rescue, fire rescue, and medical emergency, unmanned delivery can also help ensure emergency supplies.

[0003] However, in the existing technology, the use of single robot delivery has become the mainstream of the industry. Compared with single robots, the heterogeneous multi-robot air-ground collaborative material delivery system combines the advantages of unmanned vehicles and drones to achieve air-ground collaborative three-dimensional autonomous material delivery. It can realize the complementary advantages of unmanned vehicles' large carrying capacity and long endurance and drones' high mobility and three-dimensional delivery, and has powerful functions and advantages. Summary of the invention

[0004] The present invention proposes a heterogeneous multi-robot air-to-ground collaborative material distribution method. First, a multi-level and multi-constrained clustering algorithm is constructed based on a hybrid K-means algorithm to select the locations of distribution centers and delivery points. The shortest path starting from the delivery point, reaching the target delivery area (each community) and returning to the delivery point is solved by using an ant colony algorithm and a Dijkstra algorithm. An incremental map with deep sensor coupling and a multi-source information fusion method are constructed. Three-dimensional multi-point task allocation and task allocation in a multi-target obstacle environment are performed. According to the corresponding distribution needs, goods are transported to the designated target floor to realize transportation functions at different heights.

[0005] To achieve the above object, the present invention provides the following solution: a heterogeneous multi-robot air-to-ground collaborative material distribution method, comprising the following steps:

[0006] Step 1: Determine the minimum coverage area unit of air-ground collaboration, divide the coverage area unit into zones, select the distribution center for material distribution and the external location of the delivery point; solve the optimal closed-loop sequence of the delivery route between the delivery point and the material distribution area to obtain the shortest path with road information;

[0007] Step 2: Through the real-time collaborative environmental perception of unmanned aerial vehicles and unmanned vehicles, incremental updates of the distribution area and multi-source information fusion mapping are carried out;

[0008] Step 3: Plan the optimal route based on the delivery task; complete the delivery task based on the planned optimal route; and plan the return route through IoT communication to prepare for the next delivery task.

[0009] Further preferably, in step 1, the method for determining the number of distribution centers and delivery points includes:

[0010] Obtain the geographic location of all cells in the zone, cluster all cells from the perspective of spatial location, consider the material supply efficiency and supply capacity, determine the number of distribution centers in each zone, the optimal location of the distribution center, and the optimal supply range of the distribution center;

[0011] Within the optimal supply range of the distribution center, clustering is performed taking into account spatial convenience and population density to obtain the number of delivery points and the optimal location.

[0012] Further preferably, in step 1, the method for obtaining the shortest path includes:

[0013] Connect the distribution center and the delivery point through Euclidean distance;

[0014] The ant colony algorithm is used to solve the closed-loop sequence starting from the delivery point, reaching each community in the partition and returning to the delivery point; and the optimal sequence is obtained by solving the traveling salesman problem;

[0015] Considering the impact of road information on material distribution, taking distance as the optimization index, the Dijksrta algorithm is used to solve the shortest path considering road information.

[0016] Further preferably, in step 2, multi-source information fusion mapping includes: visual odometer, lidar odometer and dense map construction.

[0017] Further preferably, the method for constructing the visual odometer includes:

[0018] The image frames output by the camera are obtained, image preprocessing and feature extraction are performed on the image frames, ORB feature points and BRIEF feature descriptors are obtained, different image frames are matched according to the metric distance between the BRIEF feature descriptors, and the three-dimensional information of the ORB feature points is calculated; the pose estimation between the image frames is obtained using the PNP method.

[0019] Further preferably, the method for constructing the laser radar odometer includes:

[0020] Construct a depth map maintained by two parts: first, convert the transformation matrix obtained by calibration of the lidar point cloud to the image; second, align the pose estimated by the visual odometry; finally, output the 30Hz motion estimate of the high-frequency image to the lidar odometry; perform inter-frame matching on the lidar point cloud, and then optimize it using the 30Hz frame-to-frame motion estimate obtained from the visual odometry;

[0021] The estimation process includes: LiDAR point cloud distortion correction and frame-to-map registration, and finally outputs a pose with the same frequency as the image frame rate.

[0022] Further preferably, the method for constructing a dense point cloud map includes:

[0023] Match the same point in different images, and build a dense point cloud reconstruction framework based on MVS using patch-based multi-view stereo vision. While the loop detection of motion estimation is in progress, the map is corrected to complete the closed-loop fusion. The point cloud is gridded, and the grid hole repair algorithm is used to smooth the missing parts of the point cloud information. Then, the dense point cloud is discretized, the vector field is defined, and the surfel reconstruction method is used for surface reconstruction to form a visual dense map.

[0024] Further preferably, the method for performing optimal path planning in step three includes:

[0025] After determining the target location, global planning is performed to search for a collision-free shortest path based on the perceived global cost map; local estimation planning obtains the optimal path through the perception information of the visual odometer and lidar odometer combined with the shortest path.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] The present invention constructs a multi-level, multi-constrained clustering algorithm based on the K-means algorithm, and proposes a method for selecting a distribution center and a delivery point. The closed-loop sequence starting from the delivery point, reaching the target distribution area (each community) and returning to the delivery point is solved by the ant colony algorithm, and the optimal sequence is obtained by solving the traveling salesman problem (TSP), and the shortest path with road information is solved by the Dijkstra algorithm.

[0028] Unmanned vehicles build maps on the ground by carrying sensors such as laser radar, IMU and GPS; drones are equipped with GPS, CCD and altimeters to perceive the environment, and use the large-scale environmental perception brought by the aerial perspective to build maps in the air. Using point cloud SLAM detection and recognition technology, establish and improve the theory and method of large-scale real-time semantic-level map construction and incremental dynamic update. Unmanned vehicles and drones work together to build maps, and multi-level raster maps at different heights can be created.

[0029] Construct a multi-level clustering area division algorithm to realize three-dimensional multi-point task allocation and task allocation in multi-target obstacle environments. For task allocation in multi-target obstacle environments, set a new obstacle avoidance distance indicator, segment complex tasks, and design the optimal feedback of task allocation indicators. Using the optimal task allocation algorithm, construct a motion planning model by combining environmental constraints with motion constraints. Combine the random landmark method (PRM) with the ant colony algorithm (ACS) to construct an improved hybrid PRM-ACS hybrid algorithm to solve the problem of solving the optimal TSP path in complex obstacle environments.

[0030] The present invention adopts an omnidirectional mobile vehicle to carry multiple quad-rotor aircraft, combining the advantages of unmanned vehicles and drones to achieve air-ground coordinated three-dimensional autonomous material distribution. It can realize the complementary advantages of the large carrying capacity and long endurance of unmanned vehicles and the high maneuverability and three-dimensional distribution of drones, reduce the system's performance requirements for a single robot, and thus reduce the manufacturing cost of the robot system. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0032] Figure 1 This is a geographical coordinate distribution map of a distribution center and delivery points in a certain urban area according to an embodiment of the present invention;

[0033] Figure 2 It is a directed graph of the upstream, midstream and downstream materials distribution in a certain urban area when only the Euclidean distance is considered in the embodiment of the present invention;

[0034] Figure 3 It is the optimal path with road information solved by the Dijkstra algorithm in the embodiment of the present invention;

[0035] Figure 4 The three-dimensional point cloud map and the two-dimensional grid map constructed by the embodiment of the present invention;

[0036] Figure 5 This is a working principle diagram of incremental update and multi-source information fusion mapping in an embodiment of the present invention;

[0037] Figure 6 is a flowchart of task allocation of a multi-unmanned system according to an embodiment of the present invention;

[0038] Figure 7 It is a technical roadmap for the full coverage of the target area by multi-level multi-machine collaboration according to an embodiment of the present invention;

[0039] Figure 8 It is a schematic diagram of the working process of an embodiment of the present invention;

[0040] Fig. 9 It is a schematic diagram of the overall flow of an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Embodiment 1:

[0044] This embodiment provides a heterogeneous multi-robot air-to-ground collaborative material distribution method, which is applied to an unmanned vehicle equipped with an unmanned aerial vehicle. Specifically, the method includes the following steps:

[0045] Step 1: Determine the minimum coverage area unit of air-ground collaboration, divide the coverage area units into zones, select the distribution center for material distribution and the external location of the delivery points; solve the optimal closed-loop sequence of the delivery route between the delivery point and the material distribution area to obtain the shortest path with road information.

[0046] In this embodiment, considering the spatial geographical location constraints and population density distribution, a multi-level, multi-constrained clustering algorithm is constructed based on the K-means algorithm to select the location of distribution centers and delivery points. Taking the urban area as the basic unit, the optimal location of the distribution center and delivery point is studied. The geographical location of all the cells in the partition is obtained, and all the cells are clustered from the perspective of spatial location. Considering the material supply efficiency and supply capacity, the number of distribution centers in each partition, the optimal location of the distribution center, and the optimal supply range of the distribution center are determined; within the optimal supply range of the distribution center, spatial convenience and population density are considered for clustering, and the number and optimal location of delivery points are obtained.

[0047] Taking into account the actual situation of material supply efficiency and supply capacity, the optimal number of distribution centers in each urban area is 6 to 9. Figure 1 Taking Kuancheng District of Changchun City as an example, there are 9 distribution points in total. After clustering, the optimal location of the distribution point and the optimal community range supplied by the distribution point are obtained, and the farthest supply distance l is taken as max Considering the transportation capacity, the transportation distance between the distribution center and a single delivery point should be no more than 3.5km. The jurisdiction of each distribution center r is as follows:

[0048]

[0049] When the farthest supply distance of the distribution center is l max >3.5km, select 0.8lmax As for the jurisdiction of this distribution center, there are very few communities that are beyond its jurisdiction at this time, but they still receive material supplies from the distribution center and become the "enclaves" of the distribution center. Figure 1 The ★ in the middle left is the optimal distribution center location obtained by the algorithm. The red circle is the optimal jurisdiction of each distribution center. The coordinate points of the same color and shape scattered in the red circle represent the distribution of communities supplied by the distribution center. Very few communities that are far away from the ★ are scattered outside the red circle and become "enclaves."

[0050] In the urban area supplied by the distribution center, considering the population density factor, T is defined as the supply threshold of the distribution point, which represents the maximum number of residents allowed to be supplied by a single distribution point. For extremely large residential groups (the number of community population m>T), such communities set up n distribution points separately (n≥1), where n=m / T, if m>T. Cluster the other communities and select the optimal distribution point; the number of distribution points set up within the supply range of a distribution center is K i for:

[0051]

[0052] The service radius of a single delivery point is selected according to the actual situation. In this embodiment, 1.2 km is selected. The communities of individual principle groups select material delivery points based on the principle of proximity. Taking Kuancheng District of Changchun City as an example, Figure 1 The right side shows the optimal material delivery strategy planned for this embodiment. The “▲” in the figure is a schematic diagram of the geographical coordinate distribution of delivery points. A large-scale delivery point (capable of serving 6,309 people) is selected. The delivery point service radius is 1.2 km, which can meet the needs of all residents. The blue circle in the figure is the service range of each delivery point.

[0053] After determining the location of the distribution center and the delivery point, the material flow and allocation from the distribution center to the delivery point to the community is realized. Solving the optimal shortest path starting from the delivery point, reaching the material distribution area (each community) and returning to the delivery point is divided into two steps. First, considering only the Euclidean distance, the ant colony algorithm is used to solve the closed-loop sequence starting from the delivery point, reaching the target distribution area (each community) and returning to the delivery point. The optimal sequence is obtained by solving the traveling salesman problem (TSP); Figure 2As shown, the green dotted line is the Euclidean straight-line distance from the distribution center to the delivery point. The blue circle is the coverage area of ​​the delivery point, and the cells within this area need to be optimally covered. Due to actual road restrictions, it is necessary to consider the impact of road information on material distribution. According to the sequence solved by the ant colony algorithm, when a vehicle travels from one cell to another, it is necessary to follow the road network information to make the vehicle travel along the road and reach another cell with the optimal trajectory. The distance is used as the optimization indicator, and the Dijkstra algorithm is used to solve the shortest path from one cell to another that takes into account road information. All landmarks within a radius of 3km in the jurisdiction of the delivery point are imported, and an adjacency matrix is ​​constructed based on the road network information. The path with road information is solved based on the Dijkstra algorithm, as shown below: Figure 3 As shown, this ensures more accurate calculation of road resources and enables maximum allocation of resources.

[0054] Step 2: Through real-time collaborative environmental perception of unmanned aerial vehicles and unmanned vehicles, incremental updates of the distribution area and multi-source information fusion mapping are carried out.

[0055] When the UAV and the UGV arrive at the target delivery area, they conduct real-time collaborative environmental perception through the sensors and other equipment they carry. The UAV carries GPS, CCD, altimeter and other equipment to perceive the environment, and uses the large-scale environmental perception capabilities brought by the aerial perspective to build maps in the air. The UGV carries sensors such as LiDAR, IMU and GPS to build maps on the ground. Using point cloud SLAM detection and recognition technology, we establish and improve the theory and method of large-scale real-time semantic-level map construction and incremental dynamic update. The map construction results are as follows: Figure 4 shown.

[0056] Combination Figure 5 In this embodiment, multi-source information fusion mapping includes: visual odometry, lidar odometry and dense map construction. Motion estimation and positioning are mainly the fusion of visual odometry and lidar odometry. First, the visual odometry is used to estimate the motion at a speed of 30Hz. Then the laser odometry uses the motion estimation of the visual odometry to optimize the motion estimation at a low frequency of 5Hz to remove the point cloud motion distortion caused by mileage drift. Then loop detection and correction are performed to eliminate the accumulated error, and map construction and pose estimation are realized. The two are fused. The visual odometry solves the pose estimation of fast motion, and the lidar can ensure low drift and robustness under poor lighting conditions. Finally, the PMVS algorithm and other processing are used to achieve dense map construction.

[0057] Specifically, the method of constructing a visual odometry includes:

[0058] Obtain the image frames output by the camera, perform image preprocessing and feature extraction on the image frames, obtain ORB feature points and BRIEF feature descriptors, match different image frames according to the metric distance between the BRIEF feature descriptors, and calculate the three-dimensional information of the ORB feature points; use the PNP method to obtain the pose estimation between the image frames. In order to solve the problem of low accuracy of depth estimation of visual feature points, the point cloud frames provided by the laser radar are used to perform depth assistance on feature point observation, thereby improving the pose estimation accuracy of the visual odometer. In the process of calculating the three-dimensional information of the feature points, the depth obtained by the laser point cloud is used to perform depth correction on the feature points.

[0059] The method for constructing the laser radar odometer includes:

[0060] Through scale estimation, the depth of the extracted features can be obtained, and through continuous inter-frame matching, the high-frequency motion estimation of the image can be obtained. Construct a two-part dimensional depth map. One is to convert the transformation matrix obtained by calibration of the lidar point cloud to the image; the other is to align the pose estimated by the visual odometry. Finally, the 30Hz motion estimation of the high-frequency image is output to the lidar odometry. Inter-frame matching is performed on the lidar point cloud, and then the 30Hz frame-to-frame motion estimation obtained from the visual odometry is used for optimization. The estimation process includes: lidar point cloud distortion correction and frame-to-map registration, and the frequency of the final output pose is the same as the frame rate of the image. Constructing a two-part dimensional depth map obtained by lidar point cloud conversion and visual odometry estimated pose alignment can improve the accuracy of special extraction, assist target positioning and recognition, adapt to complex scenes, and facilitate subsequent processing and analysis. The high-frequency image motion estimation is output to the lidar odometry to optimize the point cloud inter-frame matching, and finally output the pose consistent with the image frame rate.

[0061] Methods for building dense point cloud maps include:

[0062] The same point of different images is matched pixel by pixel using the polar line search, block matching, motion estimation, and frame state estimation techniques of multi-view stereo vision (MVS). Based on MVS, the dense point cloud reconstruction framework is built using patch-based multi-view stereo vision (PMVS). While the loop detection of motion estimation is being performed, the map is also corrected to complete the closed-loop fusion. Compared with the sparse point cloud that can only be used for navigation and positioning, the map at this time has added functions such as three-dimensional reconstruction and human-computer interaction. In order to improve the physical model properties of the point cloud, the point cloud is gridded, and the grid hole repair algorithm is used to smooth the missing part of the point cloud information. Then the dense point cloud is discretized, the vector field is defined, and the surface is reconstructed using the surfel reconstruction method. Finally, a visual dense point cloud map is formed, and the autonomous return is automatically performed after the map construction is completed.

[0063] Step 3: Combination Figure 6 and Figure 7 , based on the delivery task, the optimal path is planned; based on the planned optimal path, the delivery task is completed; and through IoT communication, the return route is planned to prepare for the next delivery task.

[0064] In this embodiment, the method for performing optimal path planning includes:

[0065] After determining the target location, global planning is first performed to search for a collision-free shortest path (a series of discrete coordinate points) based on the perceived global cost map; local estimation planning obtains the optimal path by combining the perceived information of the visual odometer and the lidar odometer with the shortest path. Global planning uses reinforcement learning to calculate the global benefit optimal planning trajectory. Local obstacle avoidance uses Move_base. global_planner_node is the core planning node for global planning. Figure 8 The figure shows the actual global planning path. Based on the odometer information, the lidar perception information, and the shortest path given by the global path planning, the optimal speed control amount for the unmanned vehicle to avoid collision with obstacles is calculated. The target equation for trajectory execution time is established through the TimeElasticBand algorithm. Under the premise of maintaining a safe distance from obstacles and complying with dynamic constraints during movement, the actual running trajectory of the robot is optimized, and the executable speed of the robot is finally obtained.

[0066] The TimeElasticBand algorithm process is as follows:

[0067] First, define the posture of the unmanned vehicle:

[0068] χ a =(x i ,y i ,β i ) T ∈R 2 ×S 1 , (3)

[0069] In the formula, (x i ,y i ,β i ) represents the position and posture of the unmanned vehicle in the map coordinate system, which is called configuration.

[0070] In the space, the configuration sequence is recorded as follows:

[0071]

[0072] The time interval between two configurations is ΔT i, which indicates the time required for the unmanned vehicle to move from one configuration to another, and records the time series:

[0073] τ={ΔT i} i＝0,1,...,n , (5)

[0074] Combining configuration and time series yields:

[0075] B:=(Q,τ). (6)

[0076] The optimal path point, that is, the optimal Q, is obtained through weighted multi-objective optimization:

[0077]

[0078] In the formula, B * is the optimal result, f k (B) represents the objective function considering various constraints; γ is the weight of the objective function. The objective function enables the robot to obtain the fastest path, and the configuration points on the path are evenly separated in time, rather than the traditional shortest path in space.

[0079] Finally, after the drone has completed the delivery, it plans the return route through IoT communication, returns to the unmanned vehicle, and prepares for the next delivery target. Each drone repeats this cycle. Fig. 9 .

[0080] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A heterogeneous multi-robot air-to-ground collaborative material distribution method, the distribution method is applied to an unmanned vehicle equipped with an unmanned aerial vehicle, characterized in that: The following steps are involved: Step 1: Determine the minimum coverage area unit of air-ground collaboration, divide the coverage area unit into zones, select the distribution center for material distribution and the external placement of the delivery point; Solve the optimal closed-loop sequence of the distribution route between the delivery point and the material distribution area to obtain the shortest path with road information; Step 2: Through the real-time collaborative environmental perception of unmanned aerial vehicles and unmanned vehicles, incremental updates of the distribution area and multi-source information fusion mapping are carried out; Step 3: Based on the delivery task, plan the optimal route; based on the planned optimal route, complete the delivery task; And through IoT communication, it plans the return route and prepares for the next delivery mission.

2. According to claim 1, a heterogeneous multi-robot air-to-ground collaborative material distribution method is characterized in that: In step 1, the method for determining the number of distribution centers and delivery points includes: Obtain the geographic location of all cells in the zone, cluster all cells from the perspective of spatial location, consider the material supply efficiency and supply capacity, determine the number of distribution centers in each zone, the optimal location of the distribution center, and the optimal supply range of the distribution center; Within the optimal supply range of the distribution center, clustering is performed taking into account spatial convenience and population density to obtain the number of delivery points and the optimal location.

3. According to claim 1, a heterogeneous multi-robot air-to-ground collaborative material distribution method is characterized in that: In step 1, the method for obtaining the shortest path includes: Connect the distribution center and the delivery point through Euclidean distance; The ant colony algorithm is used to solve the closed-loop sequence starting from the delivery point, reaching each community in the partition and returning to the delivery point; and the optimal sequence is obtained by solving the traveling salesman problem; Considering the impact of road information on material distribution, taking distance as the optimization index, the Dijksrta algorithm is used to solve the shortest path considering road information.

4. According to claim 1, a heterogeneous multi-robot air-to-ground collaborative material distribution method is characterized in that: In step 2, multi-source information fusion mapping includes: visual odometry, lidar odometry and dense map construction.

5. According to claim 4, a heterogeneous multi-robot air-to-ground collaborative material distribution method is characterized in that: The method for constructing the visual odometry includes: The image frames output by the camera are obtained, image preprocessing and feature extraction are performed on the image frames, ORB feature points and BRIEF feature descriptors are obtained, different image frames are matched according to the metric distance between the BRIEF feature descriptors, and the three-dimensional information of the ORB feature points is calculated; the pose estimation between the image frames is obtained using the PNP method.

6. According to claim 4, a heterogeneous multi-robot air-to-ground collaborative material distribution method is characterized in that: The method for constructing the laser radar odometer includes: Construct a depth map maintained by two parts: first, convert the transformation matrix obtained by calibration of the lidar point cloud to the image; second, align the pose estimated by the visual odometry; finally, output the 30Hz motion estimate of the high-frequency image to the lidar odometry; perform inter-frame matching on the lidar point cloud, and then optimize it using the 30Hz frame-to-frame motion estimate obtained from the visual odometry; The estimation process includes: LiDAR point cloud distortion correction and frame-to-map registration, and finally outputs a pose with the same frequency as the image frame rate.

7. According to claim 4, a heterogeneous multi-robot air-to-ground collaborative material distribution method is characterized in that: Methods for building dense point cloud maps include: Match the same point in different images, and build a dense point cloud reconstruction framework based on MVS using patch-based multi-view stereo vision. While the loop detection of motion estimation is in progress, the map is corrected to complete the closed-loop fusion. The point cloud is gridded, and the grid hole repair algorithm is used to smooth the missing parts of the point cloud information. Then, the dense point cloud is discretized, the vector field is defined, and the surfel reconstruction method is used for surface reconstruction. Finally, a visual dense point cloud map is formed.

8. According to claim 4, a heterogeneous multi-robot air-to-ground collaborative material distribution method is characterized in that: The method for optimal path planning in step 3 includes: After determining the target location, global planning is performed to search for a collision-free shortest path based on the perceived global cost map; local estimation planning obtains the optimal path through the perception information of the visual odometer and lidar odometer combined with the shortest path.

Images