Unmanned aerial vehicle autonomous exploration method for large-scale unknown scene

Abstract

The invention relates to an unmanned aerial vehicle autonomous exploration method for a large-scale unknown scene. The method comprises the following steps: establishing a voxel map by using a binocular map obtained in the flight process of an unmanned aerial vehicle, depth data in the binocular map and IMU data; performing sparse representation on the large-scale unknown space to obtain clustering topology nodes of a plurality of divide-and-conquer spaces; planning a global coverage path based on the clustering topology nodes of the plurality of divide-and-conquer spaces and the voxel graph; and performing local path planning according to the exploration viewpoint and the global coverage path, and performing trajectory optimization on the local path. According to the invention, complete and efficient exploration of the unmanned aerial vehicle in a large-scale unknown scene can be realized.

Description

Technical Field

[0001] The present invention relates to the field of autonomous search technology for unmanned aerial vehicles, and in particular to an autonomous exploration method for unmanned aerial vehicles facing large-scale unknown scenes. Background Art

[0002] Drones have broad application value in disaster relief, mine exploration, factory inspection and other fields due to their high flexibility. How to improve the efficiency of autonomous exploration in large-scale unknown scenes has attracted extensive attention from researchers. At present, the field of autonomous exploration of drones mainly includes boundary-based, sampling-based and hybrid methods. Usually, the part between the explored free area and the unexplored area is established as the exploration boundary. During autonomous exploration, the drone moves towards the boundary and continuously updates the boundary.

[0003] Among them, most boundary-based methods adopt a greedy strategy, that is, to select the next exploration target based on the minimum motion cost or the maximum information gain, which leads to the local optimal exploration in large-scale unknown scenes, causing the drone to miss some areas and have to revisit the exploration, which greatly reduces the exploration efficiency. Sampling-based methods mainly include randomly sampling viewpoints and determining the best viewpoint to observe more unknown spaces through utility functions. Among them, the next best viewpoint planning is the earliest widely used method. It generates a fast-expanding random tree through random sampling and calculates the branch with the largest information gain to guide the robot movement. However, this method requires a lot of calculations to evaluate the viewpoint gain, and does not fully consider the global information in large-scale unknown scenes. In addition, the sampling-based method lacks directionality, is prone to generate a large number of redundant viewpoints, and may misestimate the exploration cost of certain areas, affecting the real-time and effectiveness of exploration.

[0004] It can be seen that both boundary-based methods and sampling-based methods are susceptible to local optimality, resulting in incomplete exploration and moving back and forth and revisiting the exploration area, which greatly affects the exploration efficiency. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method for autonomous exploration of unmanned aerial vehicles (UAVs) in large-scale unknown scenes, which can realize complete and efficient exploration of UAVs in large-scale unknown scenes.

[0006] The technical solution adopted by the present invention to solve the technical problem is: to provide a drone autonomous exploration method for large-scale unknown scenes, comprising the following steps:

[0007] S1, using the binocular image, depth data in the binocular image and IMU data obtained during the flight of the UAV to establish a voxel map;

[0008] S2, sparsely represent the large-scale unknown space and obtain several cluster topological nodes of the divide-and-conquer space;

[0009] S3, planning the global coverage path based on cluster topological nodes and voxel graphs of several divide-and-conquer spaces;

[0010] S4, performs local path planning based on the explored viewpoints and the global coverage path, and optimizes the trajectory of the local path.

[0011] The step S2 performs a sparse representation on the large-scale unknown space to obtain several clustered topological nodes of the divide-and-conquer space, specifically: completing the first sparse representation of the large-scale unknown space by exploring the decomposition and topological representation of the space to obtain a large number of topological nodes, dividing the topological nodes into corresponding divide-and-conquer spaces based on the divide-and-conquer idea, and clustering the topological nodes in each divide-and-conquer space to achieve a second sparse representation to obtain several clustered topological nodes of the divide-and-conquer space.

[0012] The first sparse representation of a large-scale unknown space is completed by exploring the decomposition and topological representation of the space, and a large number of topological nodes are obtained, including:

[0013] (1) Decompose the large-scale unknown space into hexagonal grids. When decomposing, take the coordinate (0,0) as the origin and the length E L and width E W They are the boundaries of the X-axis and Y-axis of the large-scale unknown space, respectively. The hexagonal grid is arranged along the X-axis and Y-axis directions. In the process of generating the grid, each column of hexagons is staggered and offset along the Y-axis direction;

[0014] (2) Performing a topological representation on the decomposed large-scale unknown space, and performing a first sparse representation of the large-scale unknown space using a topological graph. During the first sparse representation, topological nodes are used to represent the area occupied by the hexagonal grid, and topological edges are used to represent the passability of the drone.

[0015] The topological nodes are divided into corresponding divide-and-conquer spaces based on the divide-and-conquer idea, and the topological nodes in each divide-and-conquer space are clustered to achieve a second sparse representation, thereby obtaining clustered topological nodes in several divide-and-conquer spaces, specifically including:

[0016] (1) Based on the divide-and-conquer idea, the topological nodes are divided into corresponding divide-and-conquer spaces according to their spatial distribution, thus realizing the grouping of topological nodes;

[0017] (2) The k-mean clustering algorithm is used to cluster the topological nodes in each divide-and-conquer space to obtain the clustered topological nodes of each divide-and-conquer space.

[0018] The step S3 of planning the global coverage path based on the cluster topology nodes and voxel graphs of the plurality of divide-and-conquer spaces specifically includes:

[0019] (1) A method combining depth-first search and backtracking strategy is used to traverse the cluster topology nodes to obtain a traversal path, and the midpoint of the two cluster topology nodes connected by the traversal path is calculated, and the midpoint is used as the connection point for sequential exploration between the divide-and-conquer spaces;

[0020] (2) Decomposing the grid unit where the UAV is located, decomposing the unconnected areas according to the voxel occupancy information in the voxel map, taking the center of each area as the topological node and the connectivity between the areas as the edge, incrementally constructing a connected graph to obtain the free area and the unknown area;

[0021] (3) Taking the current location of the UAV as the starting point, all the topological nodes of the free area and all the topological nodes of the unknown area are taken as exploration traversal points, and the connection points of the partition space adjacent to the location of the UAV are taken as the destination to plan the global coverage path.

[0022] The method of combining depth-first search with backtracking strategy is used to traverse the cluster topology nodes and plan the traversal path. Specifically, the method is as follows: taking the current position of the drone as the starting point, taking the Euclidean distance between the cluster topology nodes as the traversal cost, starting from the starting node, recursively traversing all unvisited adjacent cluster topology nodes, and recording the path cost; when a cluster topology node is recursively searched, if it has been traversed, backtracking to the previous cluster topology node and trying other paths; when all cluster topology nodes have been visited, the recursion ends and the traversal path is returned.

[0023] The method of incrementally constructing a connectivity graph by taking the centers of each region as topological nodes and the connectivity between regions as edges specifically includes:

[0024] For the edges in the free area, the A* search algorithm is used to search for paths in two adjacent unit areas. If a path connecting the topological nodes in the two unit areas can be planned, the topological nodes in the two adjacent unit areas are connected, and the path length is defined as the weight of the connected edge.

[0025] For partially explored cells, the cells are divided, and the centers of known sub-cells or local explored sub-cells are used to represent free topological nodes, and the centers of unknown sub-cells are used to represent unknown topological nodes; free topological nodes are clustered to represent the central nodes of known areas in the cells, and unknown topological nodes in unknown sub-cells are retained to guide exploration; the A* search algorithm is used to obtain the connecting edges connecting adjacent free topological nodes and unknown topological nodes, and the path length is multiplied by a constant penalty factor to represent the weight of the connecting edge;

[0026] For the edges in the unknown area, it is assumed that two adjacent unknown units or sub-units are connected. A* search is used to obtain the path length between topological nodes between adjacent unknown units or sub-units, and the path length is multiplied by a constant penalty factor to represent the weight of the edge.

[0027] The step S4 performs local path planning according to the exploration viewpoint and the global coverage path, and performs trajectory optimization on the local path, specifically including:

[0028] (1) Determine the UAV’s exploration viewpoint through exploration boundary extraction and viewpoint sampling, and advance toward the unknown area along the global coverage path;

[0029] (2) Taking the current position of the UAV as the starting point and the UAV exploration viewpoint as the traversal point, the local exploration destination is determined according to the traversal order of the topological nodes in the global coverage path and the distance between the topological nodes in the path and the current position of the UAV. The greedy asymmetric traveling salesman problem is used to solve the local exploration path.

[0030] (3) According to the motion constraints of the UAV, the trajectory optimization algorithm based on B-spline curve is used to optimize the local exploration path to obtain the local flight trajectory of the UAV.

[0031] The method of determining the drone exploration viewpoint through exploration boundary extraction and viewpoint sampling is specifically as follows: the area between the explored free area and the unknown area is established as the exploration boundary, and then a sampling method is used based on the exploration boundary to obtain representative viewpoints with a coverage rate exceeding a threshold; for the same grid unit, the average pose of all representative viewpoints is calculated and used as the drone exploration viewpoint.

[0032] The technical solution adopted by the present invention to solve its technical problem is: providing a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned drone autonomous exploration method for large-scale unknown scenes are implemented.

[0033] Beneficial Effects

[0034] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art: the present invention decomposes the exploration space into a uniform hexagonal grid in the sparse representation of the large-scale unknown space, further divides the grid units of the local exploration into equal parts and extracts the center of each area as a topological node. Compared with the existing method, the spatial decomposition method can achieve a higher resolution representation of the local exploration environment, which is conducive to guiding the UAV to achieve complete exploration. At the same time, for the large number of topological nodes generated by the sparse representation of the large-scale unknown scene, the present invention adopts the divide-and-conquer idea to manage, thereby converting the traversal planning problem of a large number of topological nodes into multiple identical sub-problems, avoiding the problem of storing and traversing a large number of topological nodes compared to the existing method, and improving the path planning efficiency. Finally, in the local planning, the exploration viewpoints are traversed and explored considering the guidance of the global coverage path, which solves the local optimal problem existing in autonomous exploration, avoids redundant revisits in the exploration process, and is conducive to improving the exploration efficiency. In summary, the present invention can achieve complete and efficient exploration of UAVs in large-scale unknown scenes through the above process. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a system block diagram of a drone autonomous exploration method for large-scale unknown scenes according to a first embodiment of the present invention;

[0036] Figure 2 is a schematic diagram of a sparse representation of a large-scale unknown space in the first embodiment of the present invention;

[0037] Figure 3 is a connected graph constructed incrementally in the first embodiment of the present invention;

[0038] Figure 4 is a schematic diagram of a global coverage path in a first embodiment of the present invention;

[0039] Figure 5 It is a schematic diagram of a local exploration path in the first embodiment of the present invention. DETAILED DESCRIPTION

[0040] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0041] The first embodiment of the present invention relates to an autonomous exploration method of a drone for a large-scale unknown scene. The system framework of the method is as follows: Figure 1As shown, it mainly includes the drone body system, voxel mapping module, sparse representation module of large-scale unknown space, global coverage path planning module, local path planning and trajectory optimization module. The autonomous exploration method of drones for large-scale unknown scenes in this embodiment specifically includes:

[0042] S1: The binocular image, depth data in the binocular image, and IMU data obtained during the flight of the drone. The drone main body system in this embodiment adopts a quad-rotor structure, which is equipped with a binocular camera and an IMU sensor. The VINS_Fusion algorithm can calculate the drone odometer in real time based on the acquired binocular image and IMU data, and then transmit it together with the depth data in the binocular image to the voxel mapping module. In addition, the drone main body system uses PX4 as a flight controller to convert the received optimized trajectory into a motor control signal, thereby achieving access to each exploration viewpoint.

[0043] S2: The voxel mapping module takes the odometry and depth data as input to obtain a voxel map. The TSDF algorithm is used in the voxel map to calculate the distance from each voxel to the nearest surface to represent the environment, and the distance value is stored in the voxel grid, where a positive value indicates that the voxel is outside the surface and a negative value indicates that the voxel is on the surface. In the application, only voxels near the surface are stored, and the space utilization is optimized through sparse storage method.

[0044] S3: Sparsely represent the large-scale unknown space and obtain several clustered topological nodes of the divide-and-conquer space. In this step, the first sparse representation of the large-scale unknown space is completed by exploring the decomposition and topological representation of the space, and a large number of topological nodes are obtained. Based on the divide-and-conquer idea, the topological nodes are divided into corresponding divide-and-conquer spaces, and the topological nodes in each divide-and-conquer space are clustered to achieve the second sparse representation, and several clustered topological nodes of the divide-and-conquer space are obtained.

[0045] In order to plan the global coverage guidance path, it is necessary to represent the exploration space. The sparse representation of the large-scale unknown space in this implementation is as follows: Figure 2 As shown, the following steps are included:

[0046] (1) Decompose the large-scale unknown space into hexagonal grids. When decomposing, take the coordinate (0,0) as the origin and the length E L and width E W The X-axis and Y-axis boundaries of the large-scale unknown space are used to determine the range of hexagonal grid generation. The basic geometric parameters of the hexagonal grid include side length S, height H, and width W. The relationship between the three is W = 2S. To form a honeycomb-like distribution, the hexagonal grid is arranged along the X-axis and Y-axis directions, with the X-axis coordinate interval being 0.75W and the Y-axis coordinate interval being H. In the process of generating the grid, each column of hexagons is staggered along the Y-axis direction, where the offset y offset According to the current X-axis coordinate value x c Determine, the expression is:

[0047]

[0048] Wherein, x0 is the horizontal coordinate value of the initial center point of the hexagonal grid, and mod represents the modulo operation.

[0049] (2) Perform a topological representation of the decomposed large-scale unknown space, and use the topological graph to perform the first layer of sparse representation of the large-scale exploration space. In the first sparse representation, the topological nodes are used to represent the area occupied by the hexagonal grid, and the topological edges are used to represent the passability of the drone. Since the exploration space is an unknown scene, it is initially assumed that the nodes are passable. Assuming that the initial center point coordinates of the hexagonal grid are (x0, y0), the center coordinates of other hexagonal grids can be calculated according to the arrangement rules of the hexagonal grid. The expression is:

[0050]

[0051] (3) In order to facilitate the subsequent transformation of the traversal planning of a large number of topological nodes into multiple identical planning problems of a small number of topological nodes, the topological nodes are divided into corresponding divide-and-conquer spaces according to their spatial distribution based on the divide-and-conquer idea to achieve the grouping of topological nodes. In the division process, the average coordinates (x a ,y a ), and the center line passing through the coordinate and along the x-axis and y-axis is used as the dividing boundary. Taking the division of topological nodes into four partition spaces as an example, suppose the coordinates of a topological node are (x i ,y i ), can be divided according to the following conditions: if x i ≥x a And y i ≥y a , then the topological node is divided into the first divide-and-conquer space; if x i ≤x a And y i ≥y a , then the topological node is divided into the second partition space; if x i ≥x a And y i ≤y a , then the topological node is divided into the third partition space; if x i ≥x aAnd y i ≤y a , then the topological node is divided into the fourth subspace. If the number of topological nodes in the subspace is still large, each partition space is recursively divided until the number of topological nodes meets the requirements of real-time path planning.

[0052] (4) To facilitate the subsequent traversal planning of each divided space, the divided space needs to be sparsely represented. The k-mean clustering algorithm is used to cluster the topological nodes in each divided space to obtain the clustered topological nodes of each divided space. During clustering, the number of clustering nodes can be adjusted according to the scale of the topological nodes in the divided space. To facilitate the representation of the subsequent global coverage path planning, this implementation method sets the number of clustering topological nodes of the divided space to 1.

[0053] S4: Global coverage path planning based on clustering topological nodes and voxel graphs in several divide-and-conquer spaces.

[0054] In this step, in the global coverage path planning, the first layer of global coverage path planning is first performed, that is, the exploration order of each partition space is obtained by traversing the cluster topology nodes; then the second layer planning is performed on the topology nodes in the partition space in sequence ( Figure 1 In the figure, the solid-line box dsp-1 is taken as an example for the second-level planning. After completing the exploration of this subspace, the planning of dsp-2 to dsp-n is carried out in sequence). During the exploration process, a connectivity graph is constructed according to the local area where the drone is located, and it is aggregated with the second-level planned path to form the final global coverage guidance path.

[0055] The global coverage path planning in this step includes:

[0056] (1) In order to obtain the optimal traversal order of each divide-and-conquer space, a method combining depth-first search and backtracking strategy is used to traverse the cluster topology nodes to obtain the traversal path. The midpoint of the two cluster topology nodes connected by the traversal path is calculated and used as the connection point for sequential exploration between the divide-and-conquer spaces.

[0057] Among them, a method combining depth-first search and backtracking strategy is used to traverse the cluster topology nodes to obtain the traversal path: taking the current position of the drone as the starting point, the Euclidean distance between the topological nodes as the traversal cost, starting from the starting node, recursively traverse all unvisited adjacent cluster nodes, and record the path cost; when a cluster node is recursively searched, if it has been traversed, backtrack to the previous cluster node and try other paths; when all cluster nodes have been visited, the recursion ends and returns the shortest path.

[0058] (2) In the sparse representation of large-scale unknown space, obstacles are not considered to ideally decompose the exploration space. In actual exploration, in order to plan an effective global coverage path, the connectivity between topological nodes needs to be updated in combination with the perception data. First, the grid unit where the drone is located is decomposed, and the unconnected areas are decomposed according to the voxel occupancy information in the voxel map. The center of each area is used as the topological node V, and the connectivity between the areas is used as the edge E. The connectivity graph is incrementally constructed, as shown in Figure 3 shown.

[0059] Using the incremental connected graph construction process to realize two connected graphs, namely the known free graph G f =(V f ,E f ) and the unknown graph G u =(V u ,E u ) are constructed to represent the free area and the unknown area respectively, and the connecting edge E is used p To achieve the connection between the two, the overall space can be used G = (V f ∪V u ,E f ∪E u ∪E p )express.

[0060] In the process of incremental connectivity graph construction, this implementation method finely decomposes the local exploration unit into sub-units, that is, the hexagonal grid is evenly divided into six triangular grids, and the triangular grid is basically consistent with the drone's perception range, which helps to guide the drone to achieve fine exploration. For the explored unit or partially explored sub-unit, the average position of all voxels in the free area is calculated to construct the free topological node set V f , extract the center of the area of ​​the unexplored unit or the remaining sub-unit of the partial exploration to construct the unknown topological node set V u If and only if the adjacent units / subunits are free or unknown, they are considered to be connected. Incrementally constructing connected edges using topological nodes includes the following process:

[0061] For free graph G f The edge E f , in unit γ i Free topological node z in m and the adjacent unit γ j Free topological node z in n The A* search algorithm is used to search for paths in two adjacent unit areas. If a path connecting the topological nodes in the two unit areas can be planned, then node z m and node z n They are connected, and the path length is defined as the weight of the connected edge.

[0062] For partially explored cells, the cell is first finely divided, and the center of the known sub-cell or the center of the locally explored sub-cell is used to represent the free topological node, and the center of the unknown sub-cell is used to represent the unknown topological node. The free topological nodes are clustered to represent the central nodes of the known area in the cell, such as z o , and the unknown topological nodes in the unknown subunits are used to guide the detailed exploration. The A* search algorithm is used to obtain the connecting edges E connecting the adjacent free topological nodes and the unknown topological nodes. p Considering local unknowns, the path length is multiplied by a constant penalty factor a p To represent the weight of the connected edge.

[0063] For unknown graph G u The edge E u Since the scene structure in space is unknown, the default is that between two adjacent unknown units (such as topological node z r and z s ) or between subunits (e.g. topological nodes z p and z q ) is connected. A* search is used to obtain the path length between topological nodes between adjacent unknown units or subunits. Considering the uncertainty of the unknown space, the path length is multiplied by a constant penalty factor b p To represent the edge weight.

[0064] After incrementally constructing the connectivity graph, if the connectivity graph G contains multiple non-intersecting components, it indicates that there are inaccessible areas in the environment. In this case, the relevant topological nodes are directly deleted to exclude them from the exploration plan.

[0065] (3) The global coverage path is planned according to the exploration traversal order and connectivity graph of the divide-and-conquer space. Specifically, the current position of the UAV is taken as the starting point, all the topological nodes of the free area and all the topological nodes of the unknown area are taken as the exploration traversal points, and the connection points of the divide-and-conquer space adjacent to the UAV's location are taken as the destination to plan the global coverage path.

[0066] The traversal path of the topological nodes in the divide-and-conquer space without obstacles is as follows: Figure 4 As shown in (a) of , the actual exploration needs to combine the connectivity graph to plan the global coverage path. The center of the explored area in the unit is used as the node for transition to the unknown area. The local global coverage path considering connectivity is as follows Figure 4As shown in (b) in the figure. In the above process, it is necessary to solve the traversal problem of each topological node, which can be summarized as an asymmetric traveling salesman problem. In order to minimize the exploration cost of the drone, a cost matrix is ​​constructed by evaluating the traversal time between each topological node, and the Lin-Hernighan heuristic algorithm is used to solve it, so as to obtain the global coverage path. At the same time, when the drone reaches a circular area with a radius of 2 meters and a connecting topological node as the center, it is considered to have reached the destination. At this time, according to the connectivity graph, the drone is guided to enter the next divide-and-conquer space with the central node of the local exploration unit as the destination. Figure 4 As shown in (a) in the figure, the global path planning of the partition space is then explored according to the above process. The above process is iterated until the path planning of all partition spaces is completed. In the actual exploration, if the path to the next partition space cannot be found according to the connectivity graph, the traversal path of the cluster topology node is replanned.

[0067] In the asymmetric traveling salesman problem, the cost matrix C is calculated cp First, a hybrid path planning method based on voxels and topology graphs is used to search for the path from the current position of the UAV p c Arrival at topological node z n The path P cn For short-distance search, when ||p c -z n ||<d thr When ||p c -z n ||≥d thr When , the A* algorithm is used to search for a path on the connected graph closest to the two points. If there are unknown voxels / unknown topological nodes in the above path search, they are regarded as free voxels / free topological nodes, but the searched path length needs to be multiplied by the penalty term a p To represent the uncertainty of unknown voxels / unknown topological nodes. The above method is conducive to efficient path search and speeds up the calculation of the cost matrix.

[0068] Taking into account the UAV kinematics, higher costs should be assigned to longer and more tortuous paths. For the target position p c and z n The search path P between cn ={p c ,z0,z1,...,z n}, and estimate its traversal time in a piecewise manner. i The segment z i z i+1 , use the linear uniform acceleration motion model to approximate the traversal time t i :

[0069]

[0070] where v m and a m are the maximum velocity and acceleration. is the initial velocity v i Along segment z i z i+1 The projection scalar. At the starting point v of the drone i =0, during the flight of the drone v i The expression is:

[0071]

[0072] Using the above formula to calculate p c and z n The traversal time between The traversal time between all pairs of target locations constitutes the cost matrix C cp .

[0073] S5: In local path planning and trajectory optimization, local path planning is performed based on the exploration viewpoint and the global coverage path, and the trajectory of the local path is optimized to improve flight smoothness.

[0074] In local path planning and trajectory optimization, the global coverage path information is combined to explore the surrounding boundaries in a flexible and globally optimized order to avoid falling into local optimal exploration. The specific steps include the following:

[0075] (1) The UAV’s exploration viewpoints are determined through exploration boundary extraction and viewpoint sampling, and the UAV advances into the unknown area along the global coverage path to avoid revisit exploration.

[0076] In the exploration boundary extraction, the area between the explored free area and the unknown area is established as the exploration boundary, and then the representative viewpoints with coverage exceeding the threshold are obtained based on the exploration boundary using the sampling method. For the same grid unit, the average pose of all representative viewpoints is calculated and used as the drone exploration viewpoint. Since it comprehensively considers the distribution of the exploration boundary and provides an effective position and orientation for guiding the drone exploration, the exploration viewpoint is used as the traversal point of local exploration.

[0077] When extracting the above exploration boundary, the region growing algorithm is used to cluster the area between the explored free space and its neighboring unknown space to establish the exploration boundary. At the same time, when the voxel map is updated with the latest sensor data, the exploration boundary is also updated, and the exploration boundary where the occupancy state changes, that is, the unknown space is determined to be free or occupied, is removed. At the same time, a new space boundary is searched, and the principal component analysis algorithm is used to divide the larger boundary cluster into two uniform boundaries. When sampling the above viewpoints, the viewpoints are uniformly sampled around the exploration boundary, and each candidate viewpoint is evaluated. In the sampling direction within the sensor field of view, the number of unknown voxels that intersect with the maximum measurement range of the sensor is counted, and the sampled viewpoints are arranged in descending order, and the viewpoint with the most covered boundary voxels is selected as the representative viewpoint. This method is conducive to ensuring that the representative viewpoint can not only contain a large number of boundary voxels, but also a certain number of unknown voxels, thereby improving the exploration capability of unknown environments.

[0078] (2) When planning a local path, the current position of the drone is taken as the starting point, and the drone's exploration viewpoint is taken as the traversal point. The local exploration destination is determined according to the traversal order of the topological nodes in the global coverage path and the distance between the topological nodes in the path and the current position of the drone. Since there are fewer traversal nodes in the local area, the greedy asymmetric traveling salesman problem can be used to obtain the local exploration path, such as Figure 5 As shown in (a) in .

[0079] (3) In order to improve the smoothness of the UAV flight, the trajectory optimization algorithm based on the B-spline curve is used to optimize the local exploration path according to the motion constraints of the UAV to obtain the local flight trajectory of the UAV, such as Figure 5 As shown in (b) in .

[0080] It is not difficult to find that the present invention decomposes the exploration space into a uniform hexagonal grid in the sparse representation of the large-scale unknown space, further divides the grid units of the local exploration into equal parts and extracts the center of each area as a topological node. Compared with the existing method, this space decomposition method can achieve a higher resolution representation of the local exploration environment, which is conducive to guiding the UAV to achieve complete exploration. At the same time, for the large number of topological nodes generated by the sparse representation of the large-scale unknown scene, the present invention adopts the divide-and-conquer idea to manage, thereby converting the traversal planning problem of a large number of topological nodes into multiple identical sub-problems, which avoids the problem of storing and traversing a large number of topological nodes compared to the existing method, and improves the path planning efficiency. Finally, in the local planning, the exploration viewpoints are traversed and explored considering the guidance of the global coverage path, which solves the local optimal problem existing in autonomous exploration, avoids redundant revisits in the exploration process, and is conducive to improving the exploration efficiency. In summary, the present invention can achieve complete and efficient exploration of UAVs in large-scale unknown scenes through the above process.

[0081] A second embodiment of the present invention relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-mentioned method for autonomous exploration of unmanned aerial vehicles for large-scale unknown scenes.

[0082] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) that contain computer-usable program code.

[0083] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0084] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction method, which is implemented in the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0085] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0086] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A drone autonomous exploration method for large-scale unknown scenes, characterized in that: The following steps are involved: S1, using the binocular image, depth data in the binocular image and IMU data obtained during the flight of the UAV to establish a voxel map; S2, sparsely represent the large-scale unknown space and obtain several cluster topological nodes of the divide-and-conquer space; S3, planning the global coverage path based on cluster topological nodes and voxel graphs of several divide-and-conquer spaces; S4, performs local path planning based on the explored viewpoints and the global coverage path, and optimizes the trajectory of the local path.

2. The autonomous exploration method of unmanned aerial vehicles for large-scale unknown scenes according to claim 1 is characterized in that: The step S2 performs a sparse representation on the large-scale unknown space to obtain several clustered topological nodes of the divide-and-conquer space, specifically: completing the first sparse representation of the large-scale unknown space by exploring the decomposition and topological representation of the space to obtain a large number of topological nodes, dividing the topological nodes into corresponding divide-and-conquer spaces based on the divide-and-conquer idea, and clustering the topological nodes in each divide-and-conquer space to achieve a second sparse representation to obtain several clustered topological nodes of the divide-and-conquer space.

3. The autonomous exploration method of unmanned aerial vehicles for large-scale unknown scenes according to claim 2 is characterized in that: The first sparse representation of a large-scale unknown space is completed by exploring the decomposition and topological representation of the space, and a large number of topological nodes are obtained, including: (1) Decompose the large-scale unknown space into hexagonal grids. When decomposing, take the coordinate (0,0) as the origin and the length E L and width E W The X-axis and Y-axis boundaries of the large-scale unknown space are arranged along the X-axis and Y-axis directions. In the process of generating the grid, each column of hexagons is staggered and offset along the Y-axis direction; (2) Performing a topological representation on the decomposed large-scale unknown space, and performing a first sparse representation of the large-scale unknown space using a topological graph. During the first sparse representation, topological nodes are used to represent the area occupied by the hexagonal grid, and topological edges are used to represent the passability of the drone.

4. The autonomous exploration method of unmanned aerial vehicles for large-scale unknown scenes according to claim 2 is characterized in that: The topological nodes are divided into corresponding divide-and-conquer spaces based on the divide-and-conquer idea, and the topological nodes in each divide-and-conquer space are clustered to achieve a second sparse representation, thereby obtaining clustered topological nodes in several divide-and-conquer spaces, specifically including: (1) Based on the divide-and-conquer idea, the topological nodes are divided into corresponding divide-and-conquer spaces according to their spatial distribution, thus realizing the grouping of topological nodes; (2) The k-mean clustering algorithm is used to cluster the topological nodes in each divide-and-conquer space to obtain the clustered topological nodes of each divide-and-conquer space.

5. The autonomous exploration method of unmanned aerial vehicles for large-scale unknown scenes according to claim 1 is characterized in that: The step S3 of planning the global coverage path based on the cluster topology nodes and voxel graphs of the plurality of divide-and-conquer spaces specifically includes: (1) A method combining depth-first search and backtracking strategy is used to traverse the cluster topology nodes to obtain a traversal path, and the midpoint of the two cluster topology nodes connected by the traversal path is calculated, and the midpoint is used as the connection point for sequential exploration between the divide-and-conquer spaces; (2) Decomposing the grid unit where the UAV is located, decomposing the unconnected areas according to the voxel occupancy information in the voxel map, taking the center of each area as the topological node and the connectivity between the areas as the edge, incrementally constructing a connected graph to obtain the free area and the unknown area; (3) Taking the current location of the UAV as the starting point, all the topological nodes of the free area and all the topological nodes of the unknown area are taken as exploration traversal points, and the connection points of the partition space adjacent to the location of the UAV are taken as the destination to plan the global coverage path.

6. The autonomous exploration method of unmanned aerial vehicles for large-scale unknown scenes according to claim 5 is characterized in that: The method of combining depth-first search with backtracking strategy is used to traverse the cluster topology nodes and plan the traversal path. Specifically, the method is as follows: taking the current position of the drone as the starting point, taking the Euclidean distance between the cluster topology nodes as the traversal cost, starting from the starting node, recursively traversing all unvisited adjacent cluster topology nodes, and recording the path cost; when a cluster topology node is recursively searched, if it has been traversed, backtracking to the previous cluster topology node and trying other paths; when all cluster topology nodes have been visited, the recursion ends and the traversal path is returned.

7. The autonomous exploration method of unmanned aerial vehicle for large-scale unknown scenes according to claim 5 is characterized in that: The method of incrementally constructing a connectivity graph by taking the centers of each region as topological nodes and the connectivity between regions as edges specifically includes: For the edges in the free area, the A* search algorithm is used to search for paths in two adjacent unit areas. If a path connecting the topological nodes in the two unit areas can be planned, the topological nodes in the two adjacent unit areas are connected, and the path length is defined as the weight of the connected edge. For partially explored cells, the cells are divided, and the centers of known sub-cells or local explored sub-cells are used to represent free topological nodes, and the centers of unknown sub-cells are used to represent unknown topological nodes; free topological nodes are clustered to represent the central nodes of known areas in the cells, and unknown topological nodes in unknown sub-cells are retained to guide exploration; the A* search algorithm is used to obtain the connecting edges connecting adjacent free topological nodes and unknown topological nodes, and the path length is multiplied by a constant penalty factor to represent the weight of the connecting edge; For the edges in the unknown area, it is assumed that two adjacent unknown units or sub-units are connected. A* search is used to obtain the path length between topological nodes between adjacent unknown units or sub-units, and the path length is multiplied by a constant penalty factor to represent the weight of the edge.

8. The autonomous exploration method of unmanned aerial vehicles for large-scale unknown scenes according to claim 1 is characterized in that: The step S4 performs local path planning according to the exploration viewpoint and the global coverage path, and performs trajectory optimization on the local path, specifically including: (1) Determine the UAV’s exploration viewpoint through exploration boundary extraction and viewpoint sampling, and advance toward the unknown area along the global coverage path; (2) Taking the current position of the UAV as the starting point and the UAV exploration viewpoint as the traversal point, the local exploration destination is determined according to the traversal order of the topological nodes in the global coverage path and the distance between the topological nodes in the path and the current position of the UAV. The greedy asymmetric traveling salesman problem is used to solve the local exploration path. (3) According to the motion constraints of the UAV, the trajectory optimization algorithm based on B-spline curve is used to optimize the local exploration path to obtain the local flight trajectory of the UAV.

9. The autonomous exploration method of unmanned aerial vehicles for large-scale unknown scenes according to claim 8, characterized in that: The method of determining the drone exploration viewpoint through exploration boundary extraction and viewpoint sampling is specifically as follows: the area between the explored free area and the unknown area is established as the exploration boundary, and then a sampling method is used based on the exploration boundary to obtain representative viewpoints with a coverage rate exceeding a threshold; for the same grid unit, the average pose of all representative viewpoints is calculated and used as the drone exploration viewpoint.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the autonomous exploration method of a drone for large-scale unknown scenes as described in any one of claims 1 to 9 are implemented.

Images