Autonomous exploration method, device and equipment of small unmanned aerial vehicle, and storage medium

Abstract

The invention relates to the technical field of image processing, and discloses an autonomous exploration method, device and equipment of a small unmanned aerial vehicle, and a storage medium. The method comprises the following steps: performing three-dimensional rasterization processing on a closed pipeline environment to obtain a three-dimensional grid map represented by voxels; performing classification marking processing on the three-dimensional grid map to obtain a classification map containing idle voxels, occupied voxels and unknown voxels; performing local exploration planning processing on the classification map to obtain a local unmanned aerial vehicle exploration path; performing global exploration planning processing on the classification map to obtain a global unmanned aerial vehicle exploration path; and performing real-time execution and dynamic adjustment processing on the local unmanned aerial vehicle exploration path and the global unmanned aerial vehicle exploration path to obtain closed pipeline exploration data. According to the invention, the autonomous exploration efficiency and accuracy of the small unmanned aerial vehicle are improved.

Description

Technical Field

[0001] The present application relates to the field of image processing, and in particular to an autonomous exploration method, device, equipment and storage medium for a small unmanned aerial vehicle. Background Art

[0002] Closed pipeline systems play a vital role in modern industrial and urban infrastructure, transporting a variety of resources such as water, oil, and natural gas. As pipeline systems age, regular inspection and maintenance become increasingly important to prevent leaks, corrosion, and other potential safety hazards. Traditional pipeline inspection methods include manual inspection, pipeline crawlers, and wired camera systems. However, these methods are often inefficient, costly, and have limited application in complex or dangerous environments. In recent years, the rapid development of drone technology has provided new possibilities for pipeline inspection. With their flexibility and maneuverability, small drones are able to enter small spaces that are difficult for humans to reach, providing innovative solutions for the survey and inspection of the internal environment of pipelines.

[0003] However, the existing drone exploration technology in pipelines still has many limitations. First, most systems rely on preset exploration paths, lack autonomous decision-making capabilities, and cannot effectively cope with complex and changeable pipeline environments. Secondly, due to factors such as the small internal space of the pipeline and electromagnetic signal shielding, the communication and positioning of drones face severe challenges, making it difficult to achieve real-time data transmission and precise navigation. In addition, due to energy constraints, the exploration range and duration of existing systems are often severely restricted, making it difficult to complete a comprehensive inspection of large-scale pipeline networks. Finally, existing technologies have limited capabilities in environmental perception and three-dimensional reconstruction, making it difficult to provide high-precision and high-reliability information on the internal structure and status of pipelines. Summary of the invention

[0004] The present application provides a method, device, equipment and storage medium for autonomous exploration of a small unmanned aerial vehicle, which are used to improve the efficiency and accuracy of autonomous exploration of a small unmanned aerial vehicle.

[0005] In the first aspect, the present application provides an autonomous exploration method for a small unmanned aerial vehicle, and the autonomous exploration method for the small unmanned aerial vehicle includes: performing three-dimensional rasterization processing on a closed pipeline environment to obtain a three-dimensional grid map represented by voxels; performing classification and marking processing on the three-dimensional grid map to obtain a classified map containing three types of voxels: idle, occupied and unknown; performing local exploration and planning processing on the classified map to obtain a local unmanned aerial vehicle exploration path; performing global exploration and planning processing on the classified map to obtain a global unmanned aerial vehicle exploration path; performing real-time execution and dynamic adjustment processing on the local unmanned aerial vehicle exploration path and the global unmanned aerial vehicle exploration path to obtain closed pipeline exploration data.

[0006] In a second aspect, the present application provides an autonomous exploration device for a small unmanned aerial vehicle, the autonomous exploration device for the small unmanned aerial vehicle comprising:

[0007] A processing module, used for performing three-dimensional rasterization processing on the closed pipeline environment to obtain a three-dimensional raster map represented by voxels;

[0008] A labeling module, used for classifying and labeling the three-dimensional grid map to obtain a classified map containing three types of voxels: idle, occupied and unknown;

[0009] An exploration module, used to perform local exploration planning processing on the classification map to obtain a local UAV exploration path;

[0010] A planning module, used to perform global exploration planning processing on the classification map to obtain a global UAV exploration path;

[0011] The execution module is used to perform real-time execution and dynamic adjustment processing on the local UAV exploration path and the global UAV exploration path to obtain closed pipeline exploration data.

[0012] A third aspect of the present application provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the above-mentioned autonomous exploration method of a small unmanned aerial vehicle.

[0013] In the technical solution provided by the present application, a three-dimensional grid map represented by voxels is obtained by three-dimensional grid processing of the closed pipeline environment, which provides an accurate and efficient environment representation method for the drone. This representation method can not only accurately reflect the geometric structure of the pipeline, but also effectively support the subsequent path planning and obstacle avoidance algorithm, thereby improving the safety and efficiency of exploration. Secondly, the three-dimensional grid map is classified and marked to obtain a classified map containing three types of voxels: idle, occupied and unknown, so that the drone can clearly distinguish between the explored area, obstacles and areas to be explored. This classification method provides an important basis for subsequent exploration decisions and helps drones make more intelligent exploration choices in complex pipeline environments. Furthermore, by performing local exploration planning on the classified map, a local drone exploration path is obtained, so that the drone can perform efficient and detailed exploration in the local area. This local exploration strategy can make full use of the perception ability of the drone to ensure comprehensive coverage of the current area while avoiding falling into the local optimum. In addition, the classified map is globally explored and planned to obtain a global drone exploration path, so that the drone can make exploration decisions on a larger scale. This global planning capability enables the drone to make more reasonable long-term decisions when facing complex pipeline networks, avoid repeated exploration, and improve overall exploration efficiency. Finally, by real-time execution and dynamic adjustment of local drone exploration paths and global drone exploration paths, a complete closed pipeline exploration result is obtained, making the entire exploration process highly flexible and adaptable. This dynamic adjustment mechanism enables the drone to respond to environmental changes in a timely manner and optimize the exploration strategy, thereby maximizing the exploration effect under limited energy conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0015] Figure 1 A schematic diagram of an embodiment of an autonomous exploration method for a small UAV in an embodiment of the present application;

[0016] Figure 2 This is a schematic diagram of an embodiment of an autonomous exploration device for a small UAV in an embodiment of the present application. DETAILED DESCRIPTION

[0017] Embodiments of the present application provide an autonomous exploration method, device, equipment and storage medium for a small unmanned aerial vehicle. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0018] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the autonomous exploration method of a small drone in the embodiment of the present application includes:

[0019] Step S101, performing three-dimensional rasterization processing on the closed pipeline environment to obtain a three-dimensional raster map represented by voxels;

[0020] Step S102, classifying and marking the three-dimensional grid map to obtain a classified map containing three types of voxels: idle, occupied, and unknown;

[0021] Step S103: Perform local exploration planning on the classified map to obtain a local UAV exploration path;

[0022] Step S104: Perform global exploration planning on the classified map to obtain a global UAV exploration path;

[0023] Step S105: perform real-time execution and dynamic adjustment processing on the local UAV exploration path and the global UAV exploration path to obtain closed pipeline exploration data.

[0024] It is understandable that the execution subject of the present application may be an autonomous exploration device of a small UAV, or a terminal or a server, which is not limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0025] Specifically, in the local exploration planning layer, given an occupancy map and a local sub-map near the current position of the drone, the planning goal is to find a possible path (avoiding collisions with obstacles in the map) to guide the drone toward an unknown area, and this path should maximize the benefits of exploration. When no possible path can be found in the local exploration planning (the benefits of exploration remain unchanged or are less than a certain threshold), the local exploration phase is considered complete, and the global exploration planning layer is turned on. The global planning mode uses the latest map representation and the current drone position to decide a collision-free path to move the drone toward the frontier of the unknown area. The feasible path in this global problem also takes into account the remaining battery power of the drone. When the environment is completely explored or the battery power reaches the upper limit, the drone needs to plan a feasible path to return to the starting point. The specific functions of the above two modes are summarized as follows.

[0026] The local planning mode first performs boundary detection in the local map space around the current UAV position, and applies the region growing method to cluster the obtained boundary grid to obtain multiple mutually unconnected boundaries, which are often distributed in different pipeline branches. For each detected boundary, the above-mentioned unobstructed sphere generation method is applied to sample a series of candidate sphere centers at the boundary, and the maximum radius of the sphere that does not include the occupied grid is calculated. The corresponding sphere center and radius constitute the unobstructed sphere of this boundary. It should be noted that, unlike the above method, only one unobstructed sphere with the largest radius is generated at each boundary. The advantage of the unobstructed sphere is that there is no need to use a computationally expensive ray tracing algorithm when calculating the field of view coverage, which is suitable for small UAVs with limited onboard resources.

[0027] After generating an unobstructed sphere of the boundary, in order to generate the target viewpoint, firstly, candidate viewpoints are uniformly sampled on the surface of the sphere. The target viewpoint should satisfy the following conditions: 1) the viewpoint is in free space; 2) the viewpoint has the largest number of boundary grids covered by the sensor field of view. Then, for each target viewpoint of the boundary, the A* algorithm is used to generate the shortest path from the current position of the drone to the boundary viewpoint. In the set of all paths facing the target viewpoint, a path consisting of a series of vertices is given. At this time, an information gain can be calculated for each vertex to measure the space expected to be explored. In order to calculate the information gain, for each vertex, the number of unknown voxels in the sensor field of view is calculated. The exploration gain can be obtained by summing the information gains of all vertices. Finally, an optimal local exploration path can be obtained.

[0028] The global planning mode ensures two key behaviors in the exploration task: first, after completing the exploration of the unknown area locally, the global planning mode can re-plan the position of the platform to further explore the unknown and continuously expand the map; second, when the time budget of the drone reaches the critical point, it will return autonomously. In order to perform these behaviors, the planner maintains a set of unreached target viewpoints, which contains the frontier vertices that potentially point to the unexplored part of the environment. The global explorer is to detect and determine the best frontier vertex. The global exploration planner is to generate the next moving target point based on the existing map and track the target point. The environment map in the initial state only includes the perception information within the field of view of the sensor. The candidate target point is found at the frontier of the map, and the evaluation index is designed to determine the optimal target point. If a vertex is found to have a gain greater than the set threshold, the vertex is marked as the frontier. In each local planning step, the frontier vertex is extracted from the local graph, and the frontier within the radius of any existing frontier in the global graph is deleted. If there are still reachable unknown points in the environment, as the boundary between the known area and the unknown area, the frontier continues to extend to the unknown area as the drone moves, providing updated target points for the next search. This process continues until there are no unknown points in the space or the unknown points are unreachable, at which point the autonomous exploration is complete.

[0029] The key issue of autonomous exploration in this part is how to quickly detect the map frontier, and how to evaluate the exploration value of candidate map frontiers and decide on the best target frontier point.

[0030] a) Map frontier point detection

[0031] The system perceives and obtains a three-dimensional map. For the convenience of storage and processing, it is efficiently stored as a voxel map in the form of an octree. First, a cluster analysis is performed on the three-dimensional point cloud map, and it is divided into clusters according to certain distance constraints. Then the voxels corresponding to the clusters are marked as "occupied". At the same time, there is a path between the currently unknown voxels of the unmanned platform and the voxels marked as "occupied", and the voxels on the path are marked as "unoccupied". Other voxels in the space are marked as "unknown". The boundary voxels between known voxels and unknown voxels are the frontier points.

[0032] The obtained frontier points can constitute multiple areas. When the front edge area is smaller than or equal to the size of the unmanned platform, the unmanned platform cannot pass or its collision risk increases, and it is not passable, so the front edge is removed; when the front edge area is larger than the size of the unmanned platform, the front edge is retained.

[0033] b) Target frontier evaluation

[0034] This application sets two frontier evaluation parameters to select a better target frontier point, namely volume gain and frontier distance cost.

[0035] A ball with a radius of r is used to represent the size of the frontier area. The number of unknown map voxels contained in the ball represents the amount of new information added in this exploration, which is represented by I. In order to select the optimal frontier point, the distance cost from the current unknown to all frontier voxels is calculated. The cost value from the starting position to the frontier voxel stored in the voxel map is recorded as V x,y,z The frontier evaluation function composed of map information increment and frontier distance cost is expressed as E = η·I + γ·V x,y,z The voxel with the largest value of the frontier evaluation function is taken as the best frontier point, which is used as the target, and the path from the current position to the best frontier point is obtained using path planning technology.

[0036] Specifically, the autonomous exploration method of small drones first performs three-dimensional rasterization processing on the closed pipeline environment to obtain a three-dimensional raster map represented by voxels. The continuous three-dimensional space is discretized into a cube grid with a fixed resolution using voxel representation. Each voxel represents a small cubic unit in the space with a fixed side length. For example, if the voxel side length is set to 0.1 meter, a pipeline space of 10 meters × 5 meters × 3 meters will be divided into 100 × 50 × 30, a total of 150,000 voxels. Next, the three-dimensional grid map is classified and labeled to obtain a classified map containing three types of voxels: idle, occupied, and unknown. The state of each voxel is determined by analyzing drone sensor data, such as point cloud data from a lidar or depth camera. Idle voxels represent passable spaces without obstacles, occupied voxels represent impassable spaces with obstacles, and unknown voxels represent areas that have not been explored. Assume that in the initial state, the voxels within 1 meter around the drone are marked as known (idle or occupied), and the remaining voxels are marked as unknown.

[0037] Then, the classified map is processed for local exploration planning to obtain the local UAV exploration path. First, boundary detection is performed in the local sub-map near the current position of the UAV to identify the junction of the known space and the unknown space. These boundary regions are clustered using the region growing method to obtain multiple disconnected boundary regions. For example, in a bifurcated pipeline, 2-3 different boundary regions may be obtained. For each boundary region, a maximum radius unobstructed sphere is generated, with the center of the sphere located on the boundary and the radius being the distance from the center of the sphere to the nearest obstacle. Candidate viewpoints are uniformly sampled on the surface of the sphere, and the number of unknown voxels that each viewpoint can cover is evaluated. The viewpoint that covers the most unknown voxels is selected as the target viewpoint. The A* algorithm is used to plan the path from the current position to each target viewpoint, and the exploration benefit of each path (the ratio of the number of unknown voxels covered to the path length) is calculated, and the path with the highest benefit is selected as the local exploration path. The classified map is processed for global exploration planning to obtain the global UAV exploration path. First, the frontier vertex, that is, the junction of the known space and the unknown space, is detected in the global map. The exploration value of each frontier vertex is evaluated, and the number of unknown voxels in the spherical area centered on the point is calculated as the information gain, and the distance from the current position to the point is calculated as the cost. The information gain and distance cost are weighted to obtain the evaluation function value of each frontier vertex. The frontier vertex with the highest evaluation function value is selected as the global exploration target, and the path planning algorithm is used to generate a global exploration path to reach the target.

[0038] Finally, the local UAV exploration path and the global UAV exploration path are executed and dynamically adjusted in real time to obtain the closed pipeline exploration data. The UAV first executes the local exploration path and continuously updates the classification map. When the local exploration gain drops below the preset threshold, it switches to the global exploration path. During the whole process, the remaining power of the UAV is continuously monitored. When the power is lower than the safety threshold, an optimal path is planned to return to the starting point. The whole exploration process is continuously cycled until the exploration of the entire closed pipeline is completed or the battery is insufficient and needs to return.

[0039] For example, suppose that in a complex underground pipe network, a drone starts exploring from the entrance. Initially, only the space within 1 meter around the entrance is marked as known. The drone first performs local exploration, flying in the pipe at the entrance and gradually expanding the known area. When the efficiency of local exploration decreases, it switches to global exploration mode. At this time, it may be found that there are multiple branches in the pipeline, and the global planning selects the most promising unknown branch as the next exploration target. The drone flies to the branch and repeats the process of local exploration and global exploration. During the exploration process, the drone continuously updates the 3D grid map and records the shape, size and distribution of the pipeline. Finally, the drone draws a detailed 3D map of the entire underground pipe network, including all main roads and branches.

[0040] In the embodiment of the present application, by performing three-dimensional rasterization processing on the closed pipeline environment, a three-dimensional raster map represented by voxels is obtained, which provides an accurate and efficient environment representation method for the drone. This representation method can not only accurately reflect the geometric structure of the pipeline, but also effectively support subsequent path planning and obstacle avoidance algorithms, thereby improving the safety and efficiency of exploration. Secondly, the three-dimensional raster map is classified and marked to obtain a classified map containing three types of voxels: idle, occupied and unknown, so that the drone can clearly distinguish between explored areas, obstacles and areas to be explored. This classification method provides an important basis for subsequent exploration decisions and helps drones make more intelligent exploration choices in complex pipeline environments. Furthermore, by performing local exploration planning processing on the classified map, a local drone exploration path is obtained, so that the drone can perform efficient and detailed exploration in the local area. This local exploration strategy can make full use of the perception ability of the drone, ensure comprehensive coverage of the current area, and avoid falling into the local optimum. In addition, the classified map is globally explored and planned to obtain a global drone exploration path, so that the drone can make exploration decisions on a larger scale. This global planning capability enables the drone to make more reasonable long-term decisions when facing complex pipeline networks, avoid repeated exploration, and improve overall exploration efficiency. Finally, by real-time execution and dynamic adjustment of local drone exploration paths and global drone exploration paths, a complete closed pipeline exploration result is obtained, making the entire exploration process highly flexible and adaptable. This dynamic adjustment mechanism enables the drone to respond to environmental changes in a timely manner and optimize the exploration strategy, thereby maximizing the exploration effect under limited energy conditions.

[0041] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0042] (1) The closed pipeline environment is discretized by voxel representation and a fixed grid resolution is set. The grid resolution is used to ensure that each voxel has a fixed side length.

[0043] (2) A three-dimensional grid map data structure is constructed for the closed pipeline environment after the grid resolution is set, and a three-dimensional grid map represented by voxels is obtained.

[0044] Specifically, voxel representation is a three-dimensional space discretization technology that divides the continuous three-dimensional space into a series of regular cubic units, each unit is called a voxel. This method can effectively convert complex closed pipeline environments into discrete data structures that can be processed by computers. When performing discretization processing, it is first necessary to set a fixed grid resolution. The grid resolution determines the size of each voxel and directly affects the accuracy and computational complexity of the three-dimensional grid map. A smaller grid resolution can provide a more detailed environmental representation, but it will increase the computational and storage burden; a larger grid resolution can reduce the amount of computation, but may lose some detailed information. Therefore, it is necessary to select an appropriate grid resolution based on the size of the drone, the accuracy of the sensor, and the mission requirements.

[0045] For example, assuming that the dimensions of a typical closed pipeline environment are 100 meters long, 5 meters wide, and 3 meters high, considering the size of a small drone (usually around 0.5 meters) and the complexity of the internal structure of the pipeline, 0.1 meters can be selected as the grid resolution. This means that the entire environment will be divided into 1000×50×30 voxels, a total of 1.5 million voxels. This resolution can capture most of the details inside the pipeline without placing too much burden on the computing resources of the small drone.

[0046] After setting the grid resolution, we need to build the data structure of the three-dimensional grid map. Considering the characteristics of the closed pipeline environment, octree is an efficient data structure choice. Octree is an extension of the tree data structure in three-dimensional space, and each node has exactly eight child nodes. The use of octree can dynamically adjust the storage precision according to the complexity of the environment, while maintaining high-precision representation, significantly reducing storage space and query time. The process of building an octree is as follows: First, the entire closed pipeline environment is regarded as a large cube as the root node of the octree. Then, according to the occupancy in the environment, the space is recursively divided into eight sub-cubes. If a sub-cube is completely occupied or completely empty, the division stops; otherwise, the division continues until the preset minimum voxel size (i.e., grid resolution) is reached.

[0047] In practical applications, since closed pipe environments are usually long and thin structures, most of the space is either empty (traversable area) or occupied (pipe wall), so many branches of the octree can stop dividing at a higher level, thereby greatly reducing the storage space requirements. For example, in the above 100-meter-long pipe environment, assuming that the thickness of the pipe wall is 0.2 meters, the octree nodes at the outermost layer can mark most of the space as completely empty or completely occupied, and only the area inside the pipe and near the pipe wall needs to be further subdivided. In order to further optimize storage and query efficiency, a hash table can be used to store the leaf nodes of the octree. The coordinates of each voxel can be converted into a unique hash key through simple mathematical operations, thereby achieving fast insertion and query operations. This method is particularly suitable for pipe environments because most areas of the pipe are either empty or occupied, and the use of hash tables can effectively store only those voxels that contain useful information.

[0048] During the construction process, it is also necessary to consider how to represent the state of each voxel. Typically, three states are used: idle (used to represent traversable areas), occupied (used to represent obstacles or pipe walls), and unknown (used to represent unexplored areas). These states can be represented by simple enumeration types or integer values, further saving storage space. Through the above steps, the closed pipeline environment is successfully converted into a three-dimensional grid map represented by voxels. This map not only accurately represents the geometric structure of the pipeline, but also provides an efficient data interface for subsequent path planning and exploration algorithms. For example, when performing collision detection, you only need to check the voxel state around the drone's location; when performing path planning, various search algorithms can be directly applied in this discretized space.

[0049] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0050] (1) Determine the state of each voxel in the three-dimensional grid map and mark the voxels without obstacles as idle;

[0051] (2) Voxels containing obstacles are marked as occupied and voxels that have not been explored are marked as unknown.

[0052] Specifically, the original three-dimensional grid map is converted into a classified map containing three types of voxels: idle, occupied, and unknown, providing a basis for subsequent path planning and exploration strategies. The state judgment process is first based on the sensor data carried by the drone. Small drones are usually equipped with ranging devices such as lidar, depth cameras, or ultrasonic sensors. Taking lidar as an example, it measures the distance information of the surrounding environment by emitting laser beams and receiving reflected signals. Each scan of the lidar generates a set of point cloud data, each of which contains its coordinates and reflection intensity in three-dimensional space. The point cloud data first needs to be preprocessed, including denoising and downsampling. The denoising process uses a statistical outlier filter to calculate the average distance from each point to its nearest neighbor. If the distance exceeds a certain multiple of the global average distance plus or minus the standard deviation, the point is considered a noise point and deleted. Downsampling uses a voxel grid filter to divide the point cloud into a series of small cubes, and the centroid of all points in each cube is used to represent the points in the cube, thereby reducing the amount of data. Next, the preprocessed point cloud data is mapped to the three-dimensional grid map. For each point, the voxel coordinates are calculated and the voxel is marked as occupied.

[0053] Then, a ray tracing algorithm is used to process the voxels between the current position of the drone and each observation point. The ray tracing algorithm simulates the propagation path of the laser beam and marks all voxels along the path as idle until it encounters a voxel marked as occupied. This ensures that the space between the drone and the obstacle is correctly marked as a traversable area. For voxels that have not yet been marked as idle or occupied, they are marked as unknown. These unknown voxels represent areas that have not yet been explored or observed by the drone and are the main targets of subsequent exploration tasks. In practical applications, considering the measurement errors of sensors and the dynamic changes of the environment, a probabilistic model is usually used to represent the state of voxels. Each voxel is no longer simply marked as idle, occupied, or unknown, but is assigned an occupancy probability value. Initially, the occupancy probability of all voxels is set to 0.5, indicating a completely unknown state. Each observation updates the occupancy probability of the relevant voxels, and through multiple observations and updates, the state of the voxels gradually converges. When the probability value exceeds a certain threshold (such as 0.8), the voxel is marked as occupied; when the probability value is lower than a certain threshold (such as 0.2), the voxel is marked as idle; the voxels in between remain in an unknown state.

[0054] For example, suppose a small drone is exploring a section of a cylindrical pipe with a diameter of 1 meter. The drone is equipped with a 360-degree lidar that can obtain point cloud data of 1,000 points per scan. Setting the voxel size to 0.05 meters, the pipe cross section contains approximately 314 voxels. In one scan, the lidar may identify 290 points that fall on the pipe wall, and the voxels corresponding to these points are marked as occupied. The voxels on the ray path from the drone to these points (about 2,000) are marked as free. The remaining voxels (which may be several thousand, depending on the length of the pipe) remain unknown. As the drone moves through the pipe, it continuously updates this classification map. For example, when the drone encounters a T-fork, the new unknown area is added to the map, and the status of the area that has been explored is further confirmed. If there are some attached equipment (such as valves or sensors) in the pipe, these objects will be accurately marked as occupied after multiple observations.

[0055] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0056] (1) Based on the classification map, the local sub-map near the current position of the UAV is subjected to boundary detection to obtain local boundary information;

[0057] (2) Performing region growing clustering on the local boundary information to obtain multiple disconnected boundary regions;

[0058] (3) generating a maximum radius unobstructed sphere for each boundary area to obtain a boundary sphere set;

[0059] (4) uniformly sampling the surface of each bounding sphere in the bounding sphere set to obtain a set of candidate viewpoints;

[0060] (5) Evaluate and select the candidate viewpoint set to obtain the target viewpoint;

[0061] (6) Use the A* algorithm to generate paths from the current position of the drone to each target viewpoint to obtain a set of candidate paths;

[0062] (7) Calculate the exploration benefit of the candidate path set to obtain the exploration benefit value of each path;

[0063] (8) Compare and select the exploration benefit value of each path to obtain the local UAV exploration path.

[0064] Specifically, a series of complex data processing and decision-making processes are performed based on the classification map. First, the local sub-map near the current position of the drone is subjected to boundary detection processing to obtain local boundary information. This process involves scanning the voxels in the classification map to identify the junction between the known space (free voxels) and the unknown space (unknown voxels). Specifically, if there is an unknown voxel among the 26 neighboring voxels of a free voxel, the free voxel is marked as a boundary voxel. Next, the obtained local boundary information is clustered by the region growing method to obtain multiple disconnected boundary regions. The region growing method is a segmentation method that groups adjacent pixels or voxels based on similarity criteria. In this scheme, starting from a boundary voxel, its adjacent voxels are checked. If they are also boundary voxels, they are added to the current region, and the adjacent voxels of the newly added voxels are recursively checked. This process continues until there are no new adjacent boundary voxels to be added, at which time the clustering of a boundary region is completed. Then a new starting point is selected from the remaining unclassified boundary voxels, and the above process is repeated until all boundary voxels are classified.

[0065] Each boundary region is processed by generating an unobstructed sphere with the maximum radius to obtain a set of boundary spheres. The concept of unobstructed spheres originates from visibility calculation in computer graphics and is used to generate high-quality observation points in this scheme. For each boundary region, one of the voxels is selected as the center of the sphere, and then the radius is gradually increased until the sphere touches an obstacle (occupied voxel) or reaches the preset maximum radius. This process ensures that the interior of the sphere is completely unobstructed, thereby ensuring that the boundary where the center of the sphere is located can be observed unobstructed from any point on the surface of the sphere. In the obtained set of boundary spheres, the surface of each boundary sphere is uniformly sampled to obtain a set of candidate viewpoints. Uniform sampling can be achieved through a spherical coordinate system, that is, uniform division is performed on the azimuth and polar angle, and then the spherical coordinates are converted to Cartesian coordinates. The sampling density can be adjusted according to the maneuverability of the drone and the characteristics of the sensor. For example, 100 uniformly distributed points can be generated on the spherical surface, which constitute a set of candidate viewpoints.

[0066] The candidate viewpoint set is evaluated and selected to obtain the target viewpoint. The evaluation criteria include factors such as the number of unknown voxels that the viewpoint can observe and the distance from the viewpoint to the current position of the drone. A weight can be assigned to each factor, and the comprehensive score can be calculated to select the viewpoint with the highest score as the target viewpoint. The A* algorithm is used to generate a path from the current position of the drone to each target viewpoint to obtain a set of candidate paths. The A* algorithm is a heuristic search algorithm that combines the shortest path search of the Dijkstra algorithm with heuristic estimation to speed up the search process. In a three-dimensional grid map, the A* algorithm needs to define an appropriate heuristic function, usually using Euclidean distance or Manhattan distance as a heuristic estimate.

[0067] The exploration benefit calculation process is performed on the candidate path set to obtain the exploration benefit value of each path. The exploration benefit can be defined as the ratio of the number of unknown voxels that can be observed at the end point of the path (target viewpoint) to the path length. This definition balances the two aspects of information acquisition and energy consumption. Finally, the exploration benefit value of each path is compared and selected to obtain the local UAV exploration path. The path with the highest exploration benefit is selected as the final local exploration path.

[0068] For example, suppose a small drone is exploring a complex pipe system. At a certain moment, the drone is located at a T-intersection, where the main pipe has been explored, but the two branches have not yet been explored. The local subgraph size is set to 5m×5m×5m, and the voxel size is 0.1m. Boundary detection identifies two main boundary regions, corresponding to the two unexplored branch entrances. The maximum unobstructed sphere radius generated for the first boundary region is 0.8m, and for the second is 0.6m. 50 points are uniformly sampled on the surface of each of the two spheres, resulting in a total of 100 candidate viewpoints. After evaluation, the five best viewpoints are selected, located at different locations of the two branches. The A* algorithm is used to generate paths from the current position to these five viewpoints, with path lengths of 2.3m, 2.7m, 3.1m, 2.9m, and 3.5m respectively. The exploration benefit of each path is calculated, assuming that they can observe 1200, 1500, 1800, 1600, and 2000 unknown voxels respectively. Then their exploration gains are 522, 556, 581, 552, and 571 (the number of unknown voxels divided by the path length). Finally, the path with an exploration gain of 581 is selected as the local exploration path to guide the drone into the branch where more unknown areas can be observed.

[0069] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0070] (1) Based on the classification map, the frontier vertex detection process is performed on the three-dimensional grid map to obtain a frontier vertex set;

[0071] (2) Perform exploration value evaluation on the frontier vertex set to obtain the evaluation function value of each frontier vertex;

[0072] (3) Compare and select the evaluation function values ​​to obtain the best target frontier point;

[0073] (4) The path planning algorithm is used to generate a path from the current position to the optimal target frontier point to obtain the global UAV exploration path.

[0074] Specifically, the frontier vertex detection process is performed on the three-dimensional grid map to obtain a frontier vertex set. The frontier vertex is the intersection point between the known space and the unknown space, representing a potential exploration opportunity. The detection process involves scanning the entire classified map to identify the intersection of free voxels and unknown voxels. Specifically, if there is an unknown voxel among the 26 neighboring voxels of a free voxel, the free voxel is marked as a frontier vertex. This process produces a frontier vertex set, each of which is a potential exploration target. Next, the frontier vertex set is processed for exploration value evaluation to obtain the evaluation function value of each frontier vertex. The evaluation function comprehensively considers two factors: information gain and exploration cost. Information gain is estimated by calculating the number of unknown voxels in a spherical area centered on the frontier vertex. The radius of the spherical area can be set according to the perception capability of the drone, for example, to 2 meters. The exploration cost takes into account the difficulty of reaching the frontier vertex from the current position of the drone, usually expressed as Euclidean distance. The evaluation function combines these two factors and gives each factor an appropriate weight. For example, the evaluation function can be expressed as information gain divided by the square root of the exploration cost, which encourages exploration of information-rich areas while avoiding selecting targets that are too far away.

[0075] The evaluation function values ​​are compared and selected to obtain the best target frontier point. It is simple and direct. Just traverse the evaluation function values ​​of all frontier vertices and select the vertex with the largest value as the best target frontier point. This best target frontier point represents the most valuable exploration direction in the current situation. Finally, the path planning algorithm is used to generate a path from the current position to the best target frontier point to obtain the global drone exploration path. The path planning algorithm needs to find a safe and efficient path in three-dimensional space. Considering the particularity of the closed pipeline environment, the RRT (Rapid Exploration Random Tree) algorithm is a suitable choice. The RRT algorithm can effectively handle complex three-dimensional environments by randomly sampling in the configuration space and gradually expanding the tree structure.

[0076] In practical applications, this process might proceed as follows: Suppose a small drone is exploring a complex underground pipe network system. The system contains multiple branches and pipes of different diameters. The drone is equipped with a 360-degree laser scanner that continuously updates the 3D grid map during flight. The map resolution is set to 0.1 meters, that is, each voxel represents a space of 10 cm × 10 cm × 10 cm.

[0077] At a certain moment, the drone is located in an explored main pipe with a length of 50 meters and a diameter of 2 meters. Through frontier vertex detection, 300 frontier vertices are identified, mainly distributed at the end of the main pipe and at the entrances of several unexplored branch pipes. The exploration value of these 300 frontier vertices is evaluated, and the evaluation function value of each vertex is calculated. For example, for a frontier vertex A located at the end of the main pipe, a 2-meter radius sphere centered on it contains about 33,500 unknown voxels (considering the shape of the pipe, the actual number is less than the number of voxels in the complete sphere). Point A is 40 meters away from the current position of the drone. Therefore, the evaluation function value of point A may be 33,500 / √40≈5,297. For a frontier vertex B located near the entrance of the branch pipe, its 2-meter radius sphere contains about 25,000 unknown voxels (due to the smaller diameter of the branch pipe) and is only 10 meters away from the current position of the drone. The evaluation function value of point B is 25,000 / √10≈7,906.

[0078] After comparing the evaluation function values ​​of all frontier vertices, the point with the highest value is selected as the best target frontier point. In this example, point B is likely to be selected as the best target. Next, the RRT algorithm is used to plan a path from the current position of the drone to point B. The algorithm randomly samples in the 3D grid map, and each sample is checked to see if it falls within the free space. If so, it tries to connect it to the nearest tree node. This process is repeated until the tree is expanded to the vicinity of the target point. The resulting path may contain about 100 waypoints, defining the trajectory that the drone should follow.

[0079] This global exploration process is repeated continuously. Each time the drone reaches a target frontier point, it updates the map information, re-performs frontier detection and evaluation, and selects a new target.

[0080] In a specific embodiment, a front vertex detection process is performed on the three-dimensional grid map to obtain a front vertex set, including

[0081] (1) Perform cluster analysis on the three-dimensional point cloud in the classification map to obtain a set of occupied voxels;

[0082] (2) Perform path marking processing from the current position of the UAV to the occupied voxel set to obtain the unoccupied voxel set;

[0083] (3) performing unknown labeling processing on the remaining voxels in the classification map except the occupied voxel set and the unoccupied voxel set to obtain an unknown voxel set;

[0084] (4) Identify and process the boundary between the unoccupied voxel set and the unknown voxel set to obtain the frontier vertex set.

[0085] Specifically, the three-dimensional point cloud in the classification map is clustered and analyzed to obtain a set of occupied voxels. Cluster analysis is implemented using DBSCAN (density-based spatial clustering algorithm), which can effectively process irregular point cloud data. The core idea of ​​the DBSCAN algorithm is to find density-connected areas in the point cloud and divide these areas into different clusters. In a closed pipeline environment, these clusters usually correspond to structures such as pipeline walls, valves, flanges, etc. The algorithm needs to set two key parameters: ε (neighborhood radius) and MinPts (minimum number of points). For the pipeline environment, ε can be set to 1.5 times the voxel size, and MinPts can be set to 3-5 points. This can effectively identify the pipeline structure as a set of occupied voxels. Next, the path marking process is performed between the current position of the drone and the occupied voxel set to obtain a set of unoccupied voxels. This step uses an improved flood filling algorithm. Starting from the current position of the drone, the algorithm gradually checks the adjacent voxels. If the adjacent voxel does not belong to the occupied voxel set, it is marked as unoccupied and added to the queue to be processed. This process continues until the queue is empty. In this way, all voxels that can be reached from the current position of the drone are marked as unoccupied, forming an unoccupied voxel set. Then, the remaining voxels in the classified map except the occupied voxel set and the unoccupied voxel set are processed as unknown tags to obtain the unknown voxel set. This step is a simple set operation. The voxel set of the entire map is subtracted from the occupied voxel set and the unoccupied voxel set, and the remaining is the unknown voxel set. These unknown voxels represent areas that have not been explored or observed yet, and are the main targets of subsequent exploration.

[0086] Finally, the boundary between the unoccupied voxel set and the unknown voxel set is identified to obtain the frontier vertex set. This step uses an edge detection algorithm. Specifically, each voxel in the unoccupied voxel set is traversed to check its 26 neighboring voxels (in three-dimensional space, each voxel has a maximum of 26 neighboring voxels). If there is an unknown voxel in the neighboring voxels of an unoccupied voxel, the unoccupied voxel is marked as a frontier vertex. These frontier vertices constitute the frontier vertex set, which represents the boundary between the known space and the unknown space and is a potential target point for exploration.

[0087] For example, suppose a small drone is exploring a complex industrial pipe system. The system contains pipes with diameters varying from 0.5m to 2m, as well as some valves and flanges. The drone is equipped with a 360-degree laser scanner with a resolution of 0.05m. At a certain point, the drone has explored a 50m section of the main pipe and discovered two branches. First, DBSCAN clustering is performed on the acquired point cloud data. ε=0.075m (1.5 times the voxel size of 0.05m) and MinPts=4 are set. The algorithm identifies multiple clusters, including a large pipe wall cluster (containing about 100,000 points) and several smaller clusters (each containing 500-1000 points, corresponding to valves and flanges). These clusters are labeled as occupied voxels, totaling about 22,000 occupied voxels.

[0088] Next, flood filling is performed starting from the current position of the drone (assumed to be in the center of the main pipe). The algorithm expands layer by layer, marking about 150,000 unoccupied voxels, which constitute the space where the drone can fly freely. Then, the total number of voxels in the entire map (assumed to be 500,000) is subtracted from the number of occupied and unoccupied voxels, resulting in about 328,000 unknown voxels. These unknown voxels are mainly distributed in the two unexplored branch pipes and the far end of the main pipe. Finally, the boundaries of unoccupied and unknown voxels are detected. The algorithm traverses 150,000 unoccupied voxels and checks the 26 neighbors of each voxel. If an unoccupied voxel has at least one unknown neighbor, it is marked as a frontier vertex. This process identified about 5,000 frontier vertices, mainly distributed at the entrances of the two branch pipes and the far end of the main pipe.

[0089] In a specific embodiment, the exploration value evaluation process is performed on the frontier vertex set to obtain the evaluation function value of each frontier vertex, including:

[0090] (1) Perform spherical region delineation on the frontier vertices currently being processed to obtain a spherical region containing unknown voxels;

[0091] (2) Count the unknown voxels in the spherical region to obtain the information gain value of the current frontier vertex;

[0092] (3) Calculate the cost of the path from the starting position of the UAV to the current frontier vertex to obtain the distance cost value of the current frontier vertex;

[0093] (4) Perform weighted combination processing on the information gain value and distance cost value of the current frontier vertex to obtain the evaluation function value of the current frontier vertex;

[0094] (5) Summarize the evaluation function values ​​of all frontier vertices to obtain the evaluation function value of each frontier vertex in the frontier vertex set.

[0095] Specifically, the spherical region is delineated for the frontier vertex currently being processed to obtain a spherical region containing unknown voxels. This process uses the sphere equation to define the spherical region, with the frontier vertex as the center of the sphere and the radius set to the effective perception range of the drone sensor. For example, if the effective detection distance of the lidar equipped with the drone is 5 meters, the radius of the sphere is set to 5 meters. In the three-dimensional grid map, all voxels within the sphere are traversed, each voxel is determined to be inside the sphere, and the voxels within the sphere are marked. Next, the unknown voxels within the spherical region are counted to obtain the information gain value of the current frontier vertex. This process traverses all voxels within the sphere and counts the number of voxels in unknown states. The information gain value is directly expressed by the number of unknown voxels, because the more unknown voxels, the more new information that may be obtained by exploring the area.

[0096] Then, the cost calculation process is performed on the path from the starting position of the drone to the current frontier vertex to obtain the distance cost value of the current frontier vertex. Here, the A* algorithm is used to calculate the shortest path. The A* algorithm is performed on a three-dimensional grid map, with the current position of the drone as the starting point and the frontier vertex as the end point, considering occupied voxels as obstacles. The algorithm's heuristic function uses the Euclidean distance, and the cost function considers the actual path length. The final path length is the distance cost value.

[0097] The information gain value and distance cost value of the current frontier vertex are weighted and combined to obtain the evaluation function value of the current frontier vertex. The evaluation function balances the information gain and distance cost, usually in the form of information gain divided by distance cost. This encourages exploration of information-rich areas while avoiding the selection of targets that are too far away. The evaluation function can be adjusted according to specific task requirements. For example, when the battery is sufficient, more emphasis can be placed on information gain, and when the battery is insufficient, more emphasis can be placed on distance cost. Finally, the evaluation function values ​​of all frontier vertices are summarized to obtain the evaluation function value of each frontier vertex in the frontier vertex set. This process stores the evaluation function value of each frontier vertex calculated previously in an array or list to facilitate the subsequent selection of the optimal exploration target.

[0098] For example, suppose a small drone is exploring a complex industrial pipe system. The system contains multiple pipes of different diameters and some valves. The drone is equipped with a 360-degree lidar with an effective detection range of 5 meters. At a certain moment, the drone is located at the center of the main pipe with coordinates (10,10,5) and identifies five frontier vertices, located at (15,10,5), (10,15,5), (10,10,10), (5,10,5) and (10,5,5).

[0099] For the first frontier vertex (15,10,5), first define a spherical area with a radius of 5 meters and centered on it. In this spherical area, it is found that there are 2000 unknown voxels, so the information gain value is 2000. Then use the A* algorithm to calculate the path from the current position of the drone (10,10,5) to the frontier vertex, and the shortest path length is 5 meters, which is the distance cost. Assuming that the evaluation function is defined as information gain divided by distance cost, the evaluation function value of this frontier vertex is 2000 / 5=400.

[0100] Perform similar calculations for other frontier vertices:

[0101] (10,15,5): 1800 unknown voxels, distance 6 meters, evaluation function value 300

[0102] (10,10,10): 2200 unknown voxels, distance 5 meters, evaluation function value 440

[0103] (5,10,5): 1500 unknown voxels, distance 5 meters, evaluation function value 300

[0104] (10,5,5): 1600 unknown voxels, distance 5 meters, evaluation function value 320

[0105] Finally, we get a list of 5 evaluation function values: [400, 300, 440, 300, 320]. These values ​​reflect the potential value of each frontier vertex as the next exploration target. In this example, the frontier vertex with coordinates (10, 10, 10) has the highest evaluation function value of 440, indicating that it is the most valuable exploration target. This may be because it is located in the upper branch of the pipeline, which can provide more information about the vertical structure of the pipeline system.

[0106] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0107] Continue to execute the following loop process until the exploration of the entire closed pipe is completed or the return conditions are met:

[0108] (1) Execute local drone exploration paths and update classification maps in real time;

[0109] (2) Calculate the current local exploration benefit value. When the local exploration benefit value remains unchanged or is less than the preset threshold, trigger the global exploration mode;

[0110] (3) In the global exploration mode, the global exploration planning is re-performed based on the updated classification map to obtain an updated global UAV exploration path;

[0111] (4) Execute the updated global drone exploration path to find new unknown areas;

[0112] (5) While performing the exploration mission, the remaining power of the UAV is monitored in real time to obtain power status information;

[0113] (6) When the power status information is lower than a preset threshold or the exploration of the entire closed pipeline is completed, the following steps are performed: generating a return instruction;

[0114] (7) Based on the currently updated classification map, the return path is planned and processed to obtain the optimal return path from the current position of the UAV to the starting point, and the optimal return path is executed to complete the exploration mission.

[0115] Specifically, a local drone exploration path is executed, and the classification map is updated in real time. When executing the local exploration path, the drone moves along the predetermined path and uses the onboard sensors (such as lidar or depth camera) to continuously obtain data about the surrounding environment. These data are processed in real time and used to update the three-dimensional grid map, marking the newly observed space as idle or occupied, thereby gradually expanding the scope of the known area. While performing local exploration, the algorithm continuously calculates the benefit value of the current local exploration. The benefit value is usually defined as the number of newly discovered unknown voxels per unit time. For example, if 1,000 new unknown voxels are discovered within 10 seconds, the benefit value is 100 / second. When this benefit value remains unchanged or is less than the preset threshold, the global exploration mode is triggered. The setting of the threshold needs to be adjusted according to the specific pipeline environment and drone performance, for example, it can be set to 10 / second.

[0116] After triggering the global exploration mode, the global exploration planning is re-performed based on the updated classification map to obtain the updated global drone exploration path. This process first performs frontier detection to identify the boundaries of known and unknown spaces. Each frontier point is then evaluated to calculate its information gain (the number of unknown voxels that may be observed) and arrival cost (the distance from the current position to that point). The evaluation function takes these two factors into consideration and selects the optimal frontier point as the target of global exploration. A path planning algorithm (such as RRT*) is used to generate a path from the current position to the target point, which is the updated global drone exploration path. The updated global drone exploration path is executed to find new unknown areas. In this process, the drone may enter previously unexplored pipeline branches or areas, thereby significantly increasing the range of the known area. At the same time, the drone continues to use sensors to acquire data and update the classification map.

[0117] Throughout the exploration process, the drone continuously monitors the remaining power in real time to obtain power status information. This is usually achieved by reading data from the battery management system. When the power status information is lower than the preset threshold (for example, 20% power remaining), or when the exploration of the entire closed pipeline is completed (that is, there are no longer unknown areas in the classification map), the return program will be triggered. The return program first generates a return instruction, and then plans the return path based on the currently updated classification map. This process uses the A* algorithm to plan an optimal path from the current position of the drone to the starting point in the known free space. The optimal path is usually the shortest safe path that avoids all known obstacles. After planning, the drone executes this optimal return path to complete the exploration mission.

[0118] For example, let's assume that a small drone is exploring a complex industrial pipeline system. The system contains a main pipeline (2 meters in diameter and 100 meters in length) and multiple branch pipelines (1 meter in diameter and varying lengths). The drone is equipped with a 360-degree lidar with a scanning range of 5 meters and a battery capacity of 5000mAh.

[0119] Initially, the drone started exploring from the entrance of the main pipeline. It first performed local exploration, zigzagging in the main pipeline, and updated the status of about 1,000 voxels per second. In the first 10 seconds, an average of 900 new unknown voxels were discovered per second. As the exploration progressed, by the 100th second, the number of newly discovered unknown voxels per second dropped to 50. At this point, the local exploration gain value (50 / second) was lower than the preset threshold (100 / second), triggering the global exploration mode. The global planning algorithm detected three main frontier areas, located at the far end of the main pipeline and two larger branch entrances. After evaluation, the algorithm selected a branch entrance that was closer as the next exploration target and generated a 20-meter-long path.

[0120] The drone executes this new global path and enters the branch pipeline. In the process, it quickly discovers a large number of new unknown areas, and the exploration benefits are increased again. This process is repeated, and the drone switches between the main pipeline and each branch, gradually building a three-dimensional map of the entire pipeline network. At the same time, the power monitoring system continues to track the battery status. When the power drops from the initial 5000mAh to 1000mAh (20% remaining), the return program is triggered. At this point, the drone has explored about 80% of the pipeline system. The return program uses the A* algorithm to plan a 90-meter path in the known free space, avoiding all known obstacles and narrow branch pipelines. The drone returns safely to the starting point along this path, successfully completing the exploration mission.

[0121] The above describes the autonomous exploration method of the small UAV in the embodiment of the present application. The following describes the autonomous exploration device of the small UAV in the embodiment of the present application. Figure 2 In one embodiment of the present application, an autonomous exploration device for a small UAV includes:

[0122] The processing module 201 is used to perform three-dimensional rasterization processing on the closed pipeline environment to obtain a three-dimensional raster map represented by voxels;

[0123] The marking module 202 is used to classify and mark the three-dimensional grid map to obtain a classification map containing three types of voxels: idle, occupied and unknown;

[0124] An exploration module 203 is used to perform local exploration planning processing on the classification map to obtain a local UAV exploration path;

[0125] A planning module 204 is used to perform global exploration planning processing on the classification map to obtain a global UAV exploration path;

[0126] The execution module 205 is used to perform real-time execution and dynamic adjustment processing on the local UAV exploration path and the global UAV exploration path to obtain closed pipeline exploration data.

[0127] Through the cooperation of the above components, the three-dimensional grid map represented by voxels is obtained by three-dimensional gridding of the closed pipeline environment, which provides an accurate and efficient environment representation method for the UAV. This representation method can not only accurately reflect the geometric structure of the pipeline, but also effectively support the subsequent path planning and obstacle avoidance algorithms, thereby improving the safety and efficiency of exploration. Secondly, the three-dimensional grid map is classified and marked to obtain a classified map containing three types of voxels: idle, occupied and unknown, so that the UAV can clearly distinguish between the explored area, obstacles and areas to be explored. This classification method provides an important basis for subsequent exploration decisions and helps the UAV to make more intelligent exploration choices in complex pipeline environments. Furthermore, by performing local exploration planning on the classified map, the local UAV exploration path is obtained, so that the UAV can conduct efficient and detailed exploration in the local area. This local exploration strategy can make full use of the perception ability of the UAV to ensure full coverage of the current area while avoiding falling into the local optimum. In addition, the classified map is globally explored and planned to obtain the global UAV exploration path, so that the UAV can make exploration decisions on a larger scale. This global planning capability enables the drone to make more reasonable long-term decisions when facing complex pipeline networks, avoid repeated exploration, and improve overall exploration efficiency. Finally, by real-time execution and dynamic adjustment of local drone exploration paths and global drone exploration paths, a complete closed pipeline exploration result is obtained, making the entire exploration process highly flexible and adaptable. This dynamic adjustment mechanism enables the drone to respond to environmental changes in a timely manner and optimize the exploration strategy, thereby maximizing the exploration effect under limited energy conditions.

[0128] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the autonomous exploration method of the small unmanned aerial vehicle.

[0129] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0130] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0131] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An autonomous exploration method for a small unmanned aerial vehicle, characterized in that: The autonomous exploration method of the small UAV includes: Perform three-dimensional rasterization on the closed pipeline environment to obtain a three-dimensional raster map represented by voxels; Performing classification and labeling processing on the three-dimensional grid map to obtain a classification map containing three types of voxels: idle, occupied, and unknown; Performing local exploration planning processing on the classified map to obtain a local UAV exploration path; Performing global exploration planning processing on the classified map to obtain a global UAV exploration path; The local UAV exploration path and the global UAV exploration path are executed and dynamically adjusted in real time to obtain closed pipeline exploration data.

2. The method according to claim 1, characterized in that The step of performing three-dimensional rasterization processing on the closed pipeline environment to obtain a three-dimensional raster map represented by voxels comprises: The closed pipeline environment is discretized by a voxel representation method, and a fixed grid resolution is set, wherein the grid resolution is used to ensure that each voxel has a fixed side length; A three-dimensional grid map data structure is constructed for the closed pipeline environment after the grid resolution is set, and a three-dimensional grid map represented by voxels is obtained.

3. The method according to claim 1, characterized in that The step of performing classification and labeling processing on the three-dimensional grid map to obtain a classification map containing three types of voxels: idle, occupied and unknown includes: The state of each voxel in the three-dimensional grid map is judged, and the voxels without obstacles are marked as idle; Voxels containing obstacles are marked as occupied, and voxels that have not been explored are marked as unknown.

4. The method according to claim 1, characterized in that The step of performing local exploration planning processing on the classification map to obtain a local UAV exploration path includes: Based on the classification map, a boundary detection process is performed on a local sub-map near the current position of the UAV to obtain local boundary information; Performing region growing clustering processing on the local boundary information to obtain a plurality of disconnected boundary regions; Perform a maximum radius unobstructed sphere generation process on each boundary area to obtain a boundary sphere set; Perform uniform sampling on the surface of each bounding sphere in the bounding sphere set to obtain a candidate viewpoint set; Performing evaluation and selection processing on the candidate viewpoint set to obtain a target viewpoint; Use the A* algorithm to generate paths from the current position of the drone to each target viewpoint to obtain a set of candidate paths; Performing exploration benefit calculation processing on the candidate path set to obtain an exploration benefit value of each path; The exploration benefit value of each path is compared and selected to obtain the local UAV exploration path.

5. The method according to claim 1, characterized in that The step of performing global exploration planning processing on the classification map to obtain a global UAV exploration path includes: Based on the classification map, performing frontier vertex detection processing on the three-dimensional grid map to obtain a frontier vertex set; Performing exploration value evaluation processing on the frontier vertex set to obtain an evaluation function value of each frontier vertex; Comparing and selecting the evaluation function values ​​to obtain the best target frontier point; The path planning algorithm is used to generate a path from the current position to the optimal target frontier point to obtain the global UAV exploration path.

6. The method according to claim 5, characterized in that Performing frontier vertex detection processing on the three-dimensional grid map to obtain a frontier vertex set includes: Performing cluster analysis on the three-dimensional point cloud in the classification map to obtain an occupied voxel set; Performing a path marking process between the current position of the UAV and the occupied voxel set to obtain an unoccupied voxel set; Performing unknown labeling processing on remaining voxels in the classification map except the occupied voxel set and the unoccupied voxel set to obtain an unknown voxel set; The boundary between the unoccupied voxel set and the unknown voxel set is identified to obtain the front vertex set.

7. The method according to claim 6, characterized in that Perform exploration value evaluation on the frontier vertex set to obtain an evaluation function value of each frontier vertex, including: Perform spherical region delineation processing on the frontier vertex currently being processed to obtain a spherical region containing unknown voxels; Counting the unknown voxels in the spherical region to obtain the information gain value of the current frontier vertex; Perform cost calculation on the path from the starting position of the drone to the current frontier vertex to obtain the distance cost value of the current frontier vertex; Performing weighted combination processing on the information gain value and the distance cost value of the current frontier vertex to obtain an evaluation function value of the current frontier vertex; The evaluation function values ​​of all frontier vertices are aggregated to obtain the evaluation function value of each frontier vertex in the frontier vertex set.

8. The method according to claim 1, characterized in that The step of performing real-time execution and dynamic adjustment processing on the local UAV exploration path and the global UAV exploration path to obtain complete closed pipeline exploration data includes: Continue to execute the following loop process until the exploration of the entire closed pipe is completed or the return conditions are met: Executing the local UAV exploration path and updating the classification map in real time; Calculate the current local exploration benefit value. When the local exploration benefit value remains unchanged or is less than the preset threshold, trigger the global exploration mode. In the global exploration mode, re-performing global exploration planning based on the updated classification map to obtain an updated global drone exploration path; Execute the updated global drone exploration path to find new unknown areas; While performing the exploration mission, the remaining battery power of the drone is monitored in real time to obtain the battery status information; When the power status information is lower than a preset threshold or the exploration of the entire closed pipeline is completed, the following steps are performed: generating a return instruction; Based on the currently updated classification map, the return path is planned and processed to obtain the optimal return path from the current position of the drone to the starting point, and the optimal return path is executed to complete the exploration mission.

9. An autonomous exploration device for a small unmanned aerial vehicle, characterized in that: The autonomous exploration device of the small UAV includes: A processing module, used for performing three-dimensional rasterization processing on the closed pipeline environment to obtain a three-dimensional raster map represented by voxels; A labeling module, used for classifying and labeling the three-dimensional grid map to obtain a classified map containing three types of voxels: idle, occupied and unknown; An exploration module, used to perform local exploration planning processing on the classification map to obtain a local UAV exploration path; A planning module, used to perform global exploration planning processing on the classification map to obtain a global UAV exploration path; The execution module is used to perform real-time execution and dynamic adjustment processing on the local UAV exploration path and the global UAV exploration path to obtain closed pipeline exploration data.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the autonomous exploration method of the small unmanned aerial vehicle as described in any one of claims 1-8 is implemented.

Images