Unmanned aerial vehicle autonomous exploration method for indoor and underground unknown spaces

Abstract

The invention relates to an unmanned aerial vehicle autonomous exploration method for indoor / underground unknown space. The method comprises the following steps: S1, acquiring a pose of an unmanned aerial vehicle and a depth image provided by a depth camera, and constructing an ESDF environment map; s2, performing boundary search in the constructed environment map, and generating high-quality candidate viewpoints by adopting a mixed viewpoint method based on an existing boundary; s3, formulating a more reasonable boundary exploration sequence by adopting a novel boundary priority strategy and an isolated small region detection method; s4, recording a global path of each time by adopting a historical perception strategy, and calculating the motion smoothing cost to obtain an expected global exploration path; and S5, according to path planning, an LKH algorithm is adopted to solve the traversal problem of a global boundary viewpoint, an optimal traversal sequence is obtained, and an A * algorithm is utilized to search a feasible local path from the current position of the unmanned aerial vehicle to a first candidate viewpoint in a map. And S6, adopting a third-order uniform B spline curve, and introducing a penalty function to optimize the trajectory to obtain an expected trajectory. And S7, the expected flight path is sent to a flight controller of the unmanned aerial vehicle, and the unmanned aerial vehicle flies according to the flight path by applying a PD controller.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicles (UAVs), and in particular relates to an autonomous exploration method of UAVs for unknown indoor / underground spaces. Background Art

[0002] With the rapid development of drone technology, its applications have gradually expanded from traditional outdoor scenarios to include indoor and underground environments, such as post-disaster rescue, pipeline inspection, mine exploration, security monitoring, warehouse management, building surveying and mapping, and urban infrastructure inspection. These applications share common characteristics: complex structures, narrow passageways, limited lighting conditions, and often a lack of GPS signals. This requires drones to rely on autonomous perception and intelligent decision-making to complete their exploration missions. In post-disaster rescue scenarios, such as earthquakes or mining disasters, traditional search methods are inefficient. Autonomous drones, however, can quickly enter ruins or collapsed buildings, detect survivors, and create environmental maps, improving rescue efficiency. In pipeline inspection and mine exploration, drones can penetrate narrow spaces difficult for humans to reach and detect safety hazards such as cracks and gas leaks. In smart warehousing and building surveying, drones can be used for high-precision autonomous navigation and target recognition, improving the efficiency of logistics management and architectural design. Due to the environmental characteristics and complexity of these applications, traditional GPS-based navigation methods are difficult to apply. Therefore, research on autonomous drone exploration technology in indoor and underground environments can not only improve the environmental adaptability of drones but also promote their application in more high-value scenarios.

[0003] Existing exploration methods are mainly divided into boundary-based, sampling-based and machine learning-based methods. Boundary-based methods aim to cover the exploration range by identifying and exploring the boundaries between known and unknown areas. Although boundary clustering is very effective, it requires a lot of computation, which prolongs the planning time and reduces the exploration speed and efficiency. The core of the sampling-based method adopted in the greedy strategy is to select the nearest candidate target or the target that provides the maximum information gain as the subsequent target position. This method often ignores other nearby candidate targets, resulting in inefficient back-and-forth maneuvers and hesitant flight behavior, which ultimately reduces the overall exploration efficiency. Learning-based methods effectively utilize semantic information in the environment, but face challenges in practical implementation due to high computational requirements and poor real-time performance.

[0004] Current autonomous exploration algorithms have relatively high exploration efficiency in indoor and underground spaces, but they still have some limitations. Expensive candidate viewpoint ray casting detection leads to high computing resource consumption and low exploration speed during flight; inaccurate small boundary detection and chaotic exploration sequence cause drones to ignore small areas, resulting in inefficient back-and-forth movement; high-frequency replanning and heuristic calculations lead to hesitant flight behavior of drones, and exploration efficiency needs to be improved urgently. Summary of the Invention

[0005] To solve the above technical problems, the patent of this invention provides an autonomous exploration method for UAVs in indoor / underground unknown spaces. By adopting a hybrid viewpoint sampling method, high-quality candidate viewpoints are generated. At the same time, a boundary priority strategy and a small area detection method are combined to construct a more reasonable and orderly exploration sequence, reducing the inefficient back-and-forth movement and hesitation behavior of the UAV during flight, and improving exploration efficiency.

[0006] In a first aspect, the present invention provides a method for autonomous exploration of indoor / underground unknown spaces by a drone, comprising the following steps:

[0007] S1: Obtain the UAV’s pose and depth image provided by the depth camera to construct the ESDF environment map;

[0008] S2: Perform boundary search in the constructed environment map and generate high-quality candidate viewpoints based on the existing boundaries using a hybrid viewpoint method;

[0009] S3: Adopt boundary priority strategy and isolated small area detection method to formulate a more reasonable boundary exploration order;

[0010] S4: Using a history perception strategy, record each global path, calculate the motion smoothing cost, and obtain the desired global exploration path;

[0011] S5: Path planning uses the LKH algorithm to solve the traversal problem of the global boundary viewpoints, obtains the optimal traversal order, and uses the A* algorithm to search the map for a feasible local path from the current position of the drone to the first candidate viewpoint.

[0012] S6: Use the third-order uniform B-spline curve and introduce the penalty function to optimize the trajectory and obtain the desired trajectory.

[0013] S7: Send the desired flight trajectory to the UAV's flight controller, and the UAV uses the PD controller to fly according to the flight trajectory.

[0014] Furthermore, S1 specifically includes:

[0015] Using the drone's binocular RGB images from its depth camera and the flight controller's IMU as input, the vinsmono visual SLAM algorithm is used to determine the drone's position and pose. Point cloud data is obtained from the drone's depth images, converted to a unified coordinate system based on the drone's current position and pose, and voxel filtering is performed on the point cloud data. An ESDF map is constructed from the point cloud data to determine the drone's distance to obstacles.

[0016] S2 specifically includes:

[0017] The drone searches for boundaries in an unknown map. A boundary is defined as the intersection of known and unknown areas. The found clusters are then segmented, with boundaries with excessively large voxels split into two smaller boundaries. During the exploration process, region growing is used to expand new boundaries. Collision detection is used to dynamically eliminate already explored boundaries.

[0018] First, a collision-free sphere is established with the centroid of the target boundary cluster as the center of the circle and the distance from the centroid of the target boundary cluster to the nearest occupied voxel as the radius. The radius of the collision-free sphere is determined as:

[0019]

[0020] In this case, where p s is the position of the occupied voxel closest to the centroid of the target boundary cluster, denotes the location of the centroid of the target bounding cluster, and r represents the radius of the collision-free sphere. Given that the sphere is collision-free and convex, any line segment drawn from a point inside or on the sphere to a boundary point remains unobstructed. Therefore, all bounding voxels within the sphere are observable from a viewpoint located inside or on the sphere (i.e., a collision-free viewpoint). By applying a uniform sampling technique to the sphere's surface, high-quality viewpoints can be generated efficiently without requiring computationally expensive raycasts for each individual bounding voxel. Furthermore, because the map is constructed using a Euclidean Signed Distance Field (ESDF), the nearest occupied voxel can be easily identified, ensuring efficient generation of collision-free spheres even in large-scale environments.

[0021] Once the collision-free sphere is established, the method uses a spherical coordinate system to evenly distribute sampled viewpoints on its surface. Compared with viewpoints located at the center of the sphere, surface viewpoints provide a larger observation range and greater coverage, especially for the map frontier clusters (boundary clusters) located therein. After generating collision-free viewpoints, a heuristic viewpoint evaluation function is defined that combines the number of map frontier units covered during the flight (the number of visible boundary voxels) with the energy cost to determine the optimal viewpoint. The evaluation function for heuristic viewpoint selection is formulated based on the number of visible boundary voxels and the associated energy consumption, and is expressed by the following equation:

[0022]

[0023] represents the number of visible boundary voxels, represents energy consumption, w v , w en represents the weight constant, and Cv represents the cost of the heuristic viewpoint.

[0024] Due to the sphere's wide radius and extensive LiDAR scanning range, all viewpoints on the sphere's surface can effectively scan the boundary. However, in complex environments, when the viewpoint is directly below the detection area, drones equipped with narrow FOV sensors often fail to scan the upper boundary during exploration, resulting in inefficient boundary circling. To address this issue, we introduce a combined method for generating viewpoints. A frontier cube (bounding cube) is constructed based on the maximum and minimum values of the boundary in the x, y, and z directions. For each boundary, a determination is made as to whether the viewpoint lies within the cube. If the viewpoint lies within the cube, it is considered a near-boundary point. In this case, viewpoints generated by the collision-free sphere may miss the boundary in the z direction, reducing exploration efficiency. Furthermore, we check whether the sphere radius is below a predefined threshold. If this condition is met, the frontier is classified as a small frontier (frontier cluster). When a viewpoint is a near-boundary point or a small frontier, uniform sampling is performed to determine the optimal viewpoint for the boundary. Viewpoints are generated by densely sampling along the radius dimension in cylindrical coordinates, ensuring greater coverage. Our method constructs a collision-free sphere in space, where an extended viewing distance from the viewpoint to the boundary enhances coverage while facilitating flexible transitions. Uniform sampling is performed when the viewpoint is inside the bounding cube or when the sphere radius is small to determine the optimal viewpoint at the boundary. A hybrid viewpoint generation method combining spherical viewpoint generation with uniformly sampled viewpoint generation significantly reduces computation time, prevents loops, and improves exploration efficiency.

[0025] S3 specifically includes:

[0026] When exploring unknown spaces, quadrotors face high costs in thoroughly exploring their boundaries due to spatial characteristics and the sensor's FOV. The drone does not complete its exploration of the unknown area when it first reaches the boundary after a long flight, which causes the drone to repeatedly explore the boundary at a higher cost after covering other areas. To address this challenge, we introduce a boundary-first exploration approach combined with an innovative boundary cost calculation method, designed to reduce the high costs associated with boundary exploration. The first step involves determining the minimum distance from the boundary mean position (boundary mean point) within the FIS (boundary cluster setting) to the boundary area:

[0027]

[0028] 、 They represent the distances of the kth average boundary point from the border area in the x and y directions respectively;

[0029] The boundary cost is expressed as:

[0030]

[0031]

[0032] in, Boundary costs, The y-coordinate of the viewpoint, The x-coordinate of the viewpoint, The maximum value of the distance between the viewpoint and the boundary area in the y direction, The minimum distance between the viewpoint and the boundary area in the y direction, The maximum value of the distance between the viewpoint and the boundary area in the x direction, They represent the minimum value of the viewpoint from the boundary area in the x direction;

[0033] In order to implement the small area priority principle, we designed a simple, efficient and lightweight detection method. For each frontier cluster (boundary cluster), we connect the cluster edge points at both ends of the frontier in the x and y directions, and define them as frontier parallel vectors. Then, the boundary average point is moved 1 meter perpendicular to the boundary parallel vector, and the generated point is called the small area detection point. Starting from the boundary average point and the small area detection point, we perform detection in the positive and negative directions of the x and y axes, within a radius of 1 meter and 1.5 meters, respectively, with a step size of 0.1 meters. We determine whether the detection point is an occupied voxel or close to 1 meter. If a voxel that is inhabited or nearby is detected in both the x and y directions, that is, an occupied voxel, we classify the boundary cluster into an independent small area. The cost of an independent small area is expressed as:

[0034]

[0035] in, is the detection radius of the viewpoint, and the distance from the viewpoint to the occupied voxels is measured using express, and represents the predetermined fixed cost, The cost of an independent small area.

[0036] As mentioned above, accurately selecting the next exploration viewpoint is crucial for optimizing the entire exploration process. In this study, a cost matrix is defined to solve the ATSP (Asymmetric Traveling Salesman Problem) and determine the visit order, considering the main frontier clusters as possible targets.

[0037] Taking into account path time constraints, motion consistency, edge boundaries, and independent small area constraints, the cost formula from the current position to each potential goal is:

[0038]

[0039]

[0040] Among them, M tsp (0, k) represents the cost matrix from the current position to the best viewpoint in the kth boundary cluster, t lb Represents the path time constraint cost; V0 represents the best viewpoint of the current boundary cluster, V k represents the best viewpoint within the kth boundary cluster, w c represents the weight of the smoothing cost, C motion represents the motion consistency cost, w bound represents the weight of the boundary cost, C bound (k) represents the kth boundary cost, w s represents the weight of the isolated small area, C s (k) represents the cost of the isolated small region of the kth boundary cluster, M tsp (k1, k2) represents the cost matrix of the best viewpoint of the boundary cluster from k1 to k2, V k1 represents the best viewpoint to the k1th boundary cluster, V k2 represents the best viewpoint to the k2th boundary cluster, N cls Indicates the number of boundary clusters.

[0041] S4 specifically includes:

[0042] The proposed method stores each global path and generates a new global planning path based on the previously explored sequence during replanning. Because computation is limited to the updated regions, leaving other regions unaffected, updating the historical perception path incurs only a minimal increase in computational time, even with the additional computation. To align the exploration order of newly generated viewpoint clusters with the historical perception path, the historical viewpoint clusters closest to the new cluster (the new viewpoint cluster) are identified and the distance between them is calculated. New clusters are then ranked based on their proximity to the nearest center in the historical perception path. In regions with minimal updates (regions with minimal updates), newly formed viewpoint clusters exhibit only minor modifications. New cluster centers whose distance to the nearest center in the historical perception path changes by less than a preset threshold, d, are typically considered part of the same bounding cluster as the historical perception path, referred to as the reference cluster. The drone's position is also considered part of this reference cluster. To visit clusters within a new viewpoint cluster, they must be connected to the reference cluster to form a continuous path. Adjacent reference clusters are selected as start and end points, the clusters to be visited between them are identified, the exploration order is determined, and the shortest path is calculated. By connecting multiple shortest paths between reference clusters, a new historical perception path is obtained. Finally, the motion consistency cost between the current drone position and the first point (starting point) of the new historical perception path is calculated. The historical perception path cost is calculated as follows:

[0043]

[0044] C represents the global path cost, represents the historical path cost, represents the motion smoothing cost, w dir Represents the weight constant.

[0045] S5 specifically includes:

[0046] The optimal viewpoint traversal problem is solved as a traveling salesman problem, and the LKH algorithm is used to solve the approximate optimal traversal order.

[0047] S6 specifically includes:

[0048] By utilizing the convex hull property of uniform B-splines, the gradient information of the Euclidean distance field and dynamic constraints are fused together, converging in a very short time to generate smooth, safe, and dynamically feasible trajectories.

[0049] A b-spline is a piecewise polynomial whose degree Decide, control points and a knot vector , where the b-spline trajectory is expressed in time parameterization, ,in For uniform B-splines, each knot spans Have the same value .

[0050] , first Normalized to . We can then use the matrix representation to find the position:

[0051]

[0052] Since the four-rotor dynamics is a differential plane, we output for the plane Planning trajectory ,in for Uniformity Spline control points, is the knot span. We find a balance between smoothness and total trajectory time, while satisfying safety and dynamic feasibility. Spline curve boundary state constraints. It can be expressed as the following optimization problem:

[0053]

[0054] is the smoothing cost:

[0055]

[0056] and is a penalty that ensures safety and dynamic feasibility, given the following function:

[0057]

[0058] Collision cost Expressed as the repulsive force exerted by the obstacle on each control point, The value of is:

[0059]

[0060] and Penalize infeasible velocities and accelerations:

[0061]

[0062] By using the convex hull property of B-spline, the feasibility of the algorithm is effectively guaranteed. The 0th to 2nd order derivatives are set at the beginning of to achieve smooth motion. The final 0th order derivative is also set by the Viewpoint. In the implementation, we use three spline, so the associated cost is:

[0063]

[0064] In a second aspect, the present invention provides a drone autonomous exploration system for indoor / underground unknown spaces, comprising:

[0065] Mapping module: The drone obtains the binocular RGB information of the depth camera and the flight control IMU data, obtains the current pose, and uses the depth point cloud extracted from the depth image to construct the ESDF map.

[0066] Boundary search and viewpoint generation module: The boundary module performs boundary search in the environment map and proposes a hybrid viewpoint generation method, combined with cube detection, to generate high-quality candidate viewpoints.

[0067] Path Planning Module: This module determines the order of boundary exploration based on the boundary priority principle and small area detection and telemetry. It then proposes a history perception strategy based on smooth correction to determine the final global planning path and uses the LKH algorithm to solve the boundary viewpoint traversal problem.

[0068] Trajectory optimization module: The trajectory optimization module uses a cubic uniform B-spline curve and introduces a penalty function to optimize the trajectory to obtain the desired trajectory.

[0069] Flight control module: The flight control module inputs the desired trajectory into the PD controller, uses the speed control method, and inputs the output result into Mavros to realize the flight control of the UAV, and the UAV flies according to the desired trajectory.

[0070] In a third aspect, the present invention provides a drone equipped with the above-mentioned drone autonomous exploration system for indoor and underground unknown spaces.

[0071] In a fourth aspect, the present invention provides an execution device comprising a memory and a processor, wherein the memory stores program instructions to be executed by the processor, and the processor calls the program instructions to execute the steps of any one of the above-mentioned methods for autonomous exploration of indoor and underground unknown spaces by unmanned aerial vehicles.

[0072] In a fifth aspect, the present invention provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions enable the computer to execute the steps of the autonomous exploration method of unmanned aerial vehicle for indoor and underground unknown spaces as described in any one of the above claims.

[0073] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0074] The present invention aims to construct a global map. Based on environmental priors, a hybrid viewpoint sampling method that fuses collision-free sphere viewpoint generation and cube detection is proposed to reduce the high computational cost of viewpoint generation while generating high-quality candidate viewpoints. A boundary priority strategy and a small area detection method are proposed to construct a more reasonable and orderly exploration sequence. At the same time, small area detection can avoid missing isolated small areas during the exploration process. A historical perception trajectory based on smooth correction is introduced to reduce the hesitation of the drone during flight and make the trajectory smoother. The exploration problem is modeled as a traveling salesman problem, and the LKH algorithm is used to find the optimal boundary traversal sequence to improve exploration efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0076] Figure 2 Generate a simulated view for a collision-free sphere viewpoint of an embodiment of the present invention;

[0077] Figure 3 generating a simulated view for a bounding cube viewpoint of an embodiment of the present invention;

[0078] Figure 4 A three-dimensional model diagram of an embodiment of the present invention;

[0079] Figure 5 A planning path diagram for an embodiment of the present invention;

[0080] Figure 6 This is a diagram of a real machine experiment of an embodiment of the present invention. DETAILED DESCRIPTION

[0081] The following will describe the technical solution of the present invention in detail with reference to the accompanying drawings and specific implementation schemes. Figure 1 As shown. It includes the following steps:

[0082] S1: Obtain the UAV’s pose and depth image provided by the depth camera to construct the ESDF environment map;

[0083] S2: Perform boundary search in the constructed environment map and generate high-quality candidate viewpoints based on the existing boundaries using a hybrid viewpoint method;

[0084] S3: Adopt a novel boundary priority strategy and isolated small area detection method to formulate a more reasonable boundary exploration order;

[0085] S4: Using a history perception strategy, record each global path, calculate the motion smoothing cost, and obtain the desired global exploration path;

[0086] S5: Path planning uses the LKH algorithm to solve the traversal problem of the global boundary viewpoints, obtains the optimal traversal order, and uses the A* algorithm to search the map for a feasible local path from the current position of the drone to the first candidate viewpoint.

[0087] S6: Use the third-order uniform B-spline curve and introduce the penalty function to optimize the trajectory and obtain the desired trajectory.

[0088] S7: Send the desired flight trajectory to the UAV's flight controller, and the UAV uses the PD controller to fly according to the flight trajectory.

[0089] The S1 step is specifically as follows:

[0090] Using the drone's binocular RGB images from its depth camera and the flight controller's IMU as input, the vinsmono visual SLAM algorithm is used to determine the drone's position and pose. Point cloud data is obtained from the drone's depth images, converted to a unified coordinate system based on the drone's current position and pose, and voxel filtering is performed on the point cloud data. An ESDF map is constructed from the point cloud data to determine the drone's distance to obstacles.

[0091] The S2 step is specifically as follows:

[0092] The drone searches for boundaries in an unknown map. A boundary is defined as the intersection of known and unknown areas. The found clusters are then segmented, with boundaries with excessively large voxels split into two smaller boundaries. During the exploration process, region growing is used to expand new boundaries. Collision detection is used to dynamically eliminate already explored boundaries.

[0093] First, if Figure 2 As shown, a collision-free sphere is established. The collision-free sphere is defined by a circle, where the center represents the center of mass of the target boundary cluster. The radius of the collision-free sphere is determined as:

[0094]

[0095] In this case, where p s is the position of the occupied voxel closest to the center of mass, Represents the location of the centroid of the target boundary cluster. Given that spheres are collision-free and convex, any line segment drawn from a point inside or on the sphere to a boundary point remains unobstructed. Therefore, all bounding voxels within the sphere are observable from a viewpoint located inside or on the sphere. By applying a uniform sampling technique to the sphere's surface, high-quality viewpoints can be generated efficiently without requiring computationally expensive raycasts for each individual bounding voxel. Furthermore, because the map is constructed using a Euclidean Signed Distance Field (ESDF), the nearest occupied voxel can be easily identified, ensuring efficient sphere generation even in large-scale environments.

[0096] Once a collision-free sphere is established, the method uses a spherical coordinate system to evenly distribute sampled viewpoints on its surface. Compared to viewpoints located at the center of the sphere, surface viewpoints provide a larger observation range and greater coverage, especially for frontier clusters located within it. After generating collision-free viewpoints, a heuristic viewpoint evaluation function is defined that combines the number of frontier units covered during the flight with the energy cost to determine the optimal viewpoint. The evaluation function for heuristic viewpoint selection is formulated based on the number of visible boundary elements and the associated energy consumption, and is expressed by the following equation:

[0097]

[0098] represents the number of visible boundary voxels, represents energy consumption, w v , w en represents the weight constant, and Cv represents the cost of the heuristic viewpoint.

[0099] Initially, the number of frontier pixels visible from each viewpoint is calculated using the sensor FOV estimation. Next, the drone's current position and the unknown area at each viewpoint are considered as the starting and ending states, respectively, without considering obstacles. The path the drone follows to reach each viewpoint is then calculated, and the energy consumed in transitioning from the initial state to the final state is defined as energy dissipation.

[0100] Assuming that the initial and the final state , using the Pontryagin minimum principle to determine the optimal input and reduce energy dissipation.

[0101] Energy cost is defined as:

[0102]

[0103] Where T represents the initial state To the final state transition time.

[0104] The objective function is derived by applying the minimum principle:

[0105]

[0106] in is the value of the final state cost, is the maximum value of the intermediate state.

[0107] The Hamiltonian function and co-state function based on the objective function are expressed as follows:

[0108]

[0109] The optimal cost function value representing energy consumption is obtained by integration:

[0110]

[0111] The heuristic function values are calculated for all candidate viewpoints, and the viewpoint that provides the largest gain is selected as the best choice.

[0112] Due to the wide radius of the sphere and the wide LiDAR scanning range, all viewpoints on the sphere surface can effectively scan the boundary. However, in complex environments, drones equipped with narrow FOV sensors often fail to scan the upper boundary during exploration, resulting in inefficient boundary circling. To address this issue, we introduce a combined method for generating viewpoints. Figure 3 As shown in , the Frontier cube is built based on the maximum and minimum values of x, y, and z, and for each boundary, it is determined whether the viewpoint is inside the cube. If the viewpoint is inside the cube, it is considered a near-boundary point. In this case, the viewpoint generated by the collision-free sphere may miss the boundary in the z direction, thereby reducing the exploration efficiency. In addition, we also check whether the sphere radius is below a predefined threshold, such as Figure 2 As shown. If this condition is met, the frontier is classified as a small frontier. Viewpoints are generated by performing dense sampling along the radius dimension in a cylindrical coordinate system, ensuring greater coverage. Our method constructs a collision-free sphere in space, where the extended observation distance from the viewpoint to the boundary enhances coverage while promoting flexible transitions. When the viewpoint is inside the bounding cube or the sphere radius is small, computation time is significantly reduced, loops are prevented, and exploration efficiency is improved.

[0113] The S3 step is specifically as follows:

[0114] When exploring unknown spaces, quadrotors face high costs in thoroughly exploring their boundaries due to spatial characteristics and the sensor's FOV. The drone does not complete its exploration of the unknown area when it first reaches the boundary after a long flight, which causes the drone to repeatedly explore the boundary at a higher cost after covering other areas. To address this challenge, we introduce a boundary-first exploration approach combined with an innovative boundary cost calculation method, designed to reduce the high costs associated with boundary exploration. The first step involves determining the minimum distance from the average position within the FIS set to the boundary area:

[0115]

[0116] The boundary cost is expressed as:

[0117]

[0118]

[0119] In order to implement the small area priority principle, we designed a simple, efficient and lightweight detection method. For each frontier cluster, we connect the cluster edge points at both ends of the frontier in the x and y directions, and define them as frontier parallel vectors. Then, the boundary average point is moved 1 meter perpendicular to the boundary parallel vector, and the generated point is called the small area detection point. Starting from the frontier average point and the small area detection point, we perform detection in the positive and negative directions of the x and y axes, within a radius of 1 meter and 1.5 meters, respectively, with a step size of 0.1 meters. We determine whether the detection point is an occupied voxel or close to 1 meter. If inhabited or nearby voxels are detected in both the x and y directions, we classify the boundary cluster into an independent small area. The cost of an independent small area is expressed as:

[0120]

[0121] As mentioned above, accurately selecting the next exploration viewpoint is crucial for optimizing the entire exploration process. In this study, a cost matrix is defined to solve the ATSP and determine the visiting order, considering the main frontier clusters as possible targets.

[0122] Taking into account path time constraints, motion consistency, edge boundaries, and independent small area constraints, the cost formula from the current position to each potential goal is:

[0123]

[0124]

[0125] The present invention uses distance time cost as the main basis for selecting the frontier sequence, and the path time constraint cost is calculated as follows:

[0126]

[0127] in, represents the best viewpoint within the boundary cluster, Indicates its position and the related yaw angle is expressed by , and the maximum allowable linear velocity and angular velocity of the UAV are respectively and given.

[0128] In order to maintain the consistency of the motion direction, a motion consistency cost is introduced to penalize the frontier cluster that deviates significantly from the current trajectory. The motion consistency cost is defined as:

[0129]

[0130] The S4 step is specifically as follows:

[0131] The proposed method stores each global path and generates a new global planning path based on the previously explored sequence during replanning. Because computation is limited to the updated region, leaving other regions unaffected, updating the historical perception path incurs only minimal computational time, even with the additional computation. To align the exploration order of newly generated viewpoint clusters with the historical perception path, the historical viewpoint clusters closest to the new cluster are identified and the distance between them is calculated. New clusters are then ranked based on their proximity to the nearest center in the historical perception path. Newly formed viewpoint clusters in regions with minimal updates exhibit only minor modifications. Centers with a distance change of less than d are typically considered to be part of the same bounding cluster as the historical perception path, referred to as the reference cluster. The drone's position is also considered part of this reference cluster. To visit clusters within a new viewpoint cluster, they must be connected to the reference cluster to form a continuous path. Adjacent reference clusters are selected as the starting and ending points, the clusters to be visited between them are identified, the exploration order is determined, and the shortest path is calculated. By connecting multiple shortest paths between reference clusters, a new historical perception path is generated. Finally, the motion consistency cost between the current drone position and the first point of the new historical perception path is calculated to prevent a large change in motion direction that would reduce exploration efficiency. The historical perception path is calculated as follows:

[0132]

[0133] The S5 step is specifically as follows:

[0134] The optimal viewpoint traversal problem is solved as a traveling salesman problem, and the LKH algorithm is used to solve the approximate optimal traversal order.

[0135] The S6 step is specifically as follows:

[0136] By utilizing the convex hull property of uniform B-splines, the gradient information of the Euclidean distance field and dynamic constraints are fused together, converging in a very short time to generate smooth, safe, and dynamically feasible trajectories.

[0137] A b-spline is a piecewise polynomial whose degree Decide, control points and a knot vector , where the b-spline trajectory is expressed in time parameterization, ,in For uniform B-splines, each knot spans Have the same value . , first Normalized to . We can then use the matrix representation to find the position:

[0138]

[0139] Since the four-rotor dynamics is a differential plane, we output for the plane Planning trajectory ,in for Uniformity Spline control points, is the knot span. We find a balance between smoothness and total trajectory time, while satisfying safety and dynamic feasibility. Spline curve boundary state constraints. It can be expressed as the following optimization problem:

[0140]

[0141] is the smoothing cost:

[0142]

[0143] and is a penalty that ensures safety and dynamic feasibility, given the following function:

[0144]

[0145] Collision cost Expressed as the repulsive force exerted by the obstacle on each control point, The value of is:

[0146]

[0147] and Penalize infeasible velocities and accelerations:

[0148]

[0149] By using the convex hull property of B-spline, the feasibility of the algorithm is effectively guaranteed. The 0th to 2nd order derivatives are set at the beginning of to achieve smooth motion. The final 0th order derivative is also set by the Viewpoint. In the implementation, we use three spline, so the associated cost is:

[0150]

[0151] This embodiment provides a drone autonomous exploration system for indoor / underground unknown spaces, including:

[0152] Mapping module: The drone obtains the binocular RGB information of the depth camera and the flight control IMU data, obtains the current pose, and uses the depth point cloud extracted from the depth image to construct the ESDF map.

[0153] Boundary search and viewpoint generation module: The boundary module performs boundary search in the environment map and proposes a hybrid viewpoint generation method, combined with cube detection, to generate high-quality candidate viewpoints.

[0154] Path Planning Module: This module determines the order of boundary exploration based on the boundary priority principle and small area detection and telemetry. It then proposes a history perception strategy based on smooth correction to determine the final global planning path and uses the LKH algorithm to solve the boundary viewpoint traversal problem.

[0155] Trajectory optimization module: The trajectory optimization module uses a cubic uniform B-spline curve and introduces a penalty function to optimize the trajectory to obtain the desired trajectory.

[0156] Flight control module: The flight control module inputs the desired trajectory into the PD controller, uses the speed control method, and inputs the output result into Mavros to realize the flight control of the UAV, and the UAV flies according to the desired trajectory.

[0157] It should be noted that the system embodiments provided by the present invention are not only used to implement the methods in the above-mentioned method embodiments, but also used to implement the methods in other method embodiments provided by the present invention. The only difference lies in the setting of corresponding functional modules, and the principles thereof are basically the same as the principles of the above-mentioned system embodiments provided by the present invention. As long as those skilled in the art refer to the specific technical solutions in other method embodiments on the basis of the above-mentioned system embodiments, obtain corresponding technical means and technical solutions composed of these technical means by combining technical features, and on the premise of ensuring the practicality of the technical solutions, improve the modules in the above-mentioned system embodiments to obtain corresponding system class embodiments for implementing the methods in other method class embodiments.

[0158] Based on the same inventive concept as the aforementioned embodiment, an embodiment of the present invention also provides an execution device, including a memory and a processor, wherein the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of the aforementioned method for autonomous exploration of indoor and underground unknown spaces by unmanned aerial vehicles.

[0159] Based on the same inventive concept as the aforementioned embodiment, an embodiment of the present invention also provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions enable the computer to execute the steps of the described method for autonomous exploration of indoor and underground unknown spaces by unmanned aerial vehicles.

[0160] It should be understood that parts not elaborated in detail in this specification belong to the prior art.

[0161] Although this document frequently uses terms such as drones and ground base stations, the use of other terms is not excluded. These terms are used solely to more conveniently describe the essence of the present invention, and interpreting them as any additional limitations would be contrary to the spirit of the present invention.

[0162] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of protection of the patent of the present invention. Under the guidance of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested by the present invention shall be based on the attached claims.

Claims

1. A method for autonomous exploration of indoor and underground unknown spaces by drones, characterized in that: The following steps are involved: Obtain the drone's position and depth image provided by the depth camera to build an environment map; Perform boundary search in the constructed environment map, and use the spherical viewpoint generation method based on the boundary and combine it with cube detection to generate high-quality candidate viewpoints; Formulate a reasonable boundary exploration order based on boundary priority strategy and small area detection; Adopting a historical perception strategy, recording each global path, calculating the motion smoothing cost, and obtaining the desired global exploration path; Path planning: solving the traversal problem of the global boundary viewpoints, obtaining the optimal traversal order, and searching the map for a feasible local path from the current position of the drone to the first candidate viewpoint; A third-order uniform B-spline curve is used and a penalty function is introduced to optimize the trajectory to obtain the desired flight trajectory. Control the drone to fly according to the desired flight trajectory.

2. The autonomous drone exploration method for indoor and underground unknown spaces according to claim 1 is characterized in that: Perform boundary search in the constructed environment map, use the spherical viewpoint generation method combined with cube detection to generate high-quality candidate viewpoints, including the following steps: Create a collision-free sphere with the center of mass of the target boundary cluster as the center of the circle. The radius of the collision-free sphere is , where p s is the position of the occupied voxel closest to the center of mass, represents the location of the centroid of the target boundary cluster; A spherical coordinate system is used to generate uniformly distributed collision-free sampling viewpoints on the surface of the sphere, and a heuristic viewpoint evaluation function is formulated based on the visible boundary voxels and the associated energy loss to determine the optimal viewpoint of the boundary; A bounding cube is established based on the maximum and minimum values of the boundary in the three directions of x, y, and z in the global coordinate system. If the viewpoint is within the bounding cube, it is considered a near-boundary point. When the radius of the sphere is lower than a preset threshold, the boundary is classified as a small frontier cluster. When the viewpoint is close to the boundary or a small frontier cluster, the optimal viewpoint of the boundary is determined by uniform sampling.

3. The autonomous drone exploration method for indoor and underground unknown spaces according to claim 2 is characterized in that: The heuristic viewpoint evaluation function is expressed by the following equation: ； represents the number of visible boundary voxels, represents energy consumption, w v , w en represents the weight constant, C v Represents the cost of the heuristic viewpoint.

4. The autonomous drone exploration method for indoor and underground unknown spaces according to claim 1, characterized in that: A boundary priority strategy is used to prioritize the exploration of border areas. At the same time, the boundary clusters are classified into independent small areas, small area detection is performed, and a reasonable boundary exploration sequence is formulated, including the following steps: Determine the minimum distance from the average boundary point within the boundary cluster to the border area and calculate the boundary cost; For each boundary cluster, the cluster edge points at both ends of the boundary are connected in the x and y directions, which are defined as boundary parallel vectors. The boundary average point is moved a certain distance perpendicular to the boundary parallel vector to generate a small area detection point. Starting from the boundary average point and the small area detection point, detection is performed in the x and y directions with a certain step size to determine whether the detection point is an occupied voxel. If occupied voxels are detected in both the x and y directions, the boundary cluster is classified into independent small regions and the cost of the independent small regions is calculated; Considering the boundary clusters as potential targets, considering the path time constraint, motion consistency, edge boundary and independent small area constraints, the cost from the current position to each potential target is calculated to obtain the cost matrix; Solve the asymmetric traveling salesman problem and determine the order of frontier exploration based on the cost matrix.

5. The autonomous drone exploration method for indoor and underground unknown spaces according to claim 4 is characterized in that: The cost of the independent small area is expressed as: ； in, is the detection radius of the viewpoint, and the distance from the viewpoint to the occupied voxels is measured using express, and represents the predetermined fixed cost, The cost of an independent small area; The cost formula from the current position to each potential goal is: ； ； Among them, M tsp (0, k) represents the cost matrix from the current position to the best viewpoint in the kth boundary cluster, t lb Represents the path time constraint cost; V0 represents the best viewpoint of the current boundary cluster, V k represents the best viewpoint within the kth boundary cluster, w c represents the weight of the motion consistency cost, C motion represents the motion consistency cost, w bound represents the weight of the boundary cost, C bound (k) represents the kth boundary cost, w s represents the weight of the isolated small area, C s (k) represents the cost of the isolated small region of the kth boundary cluster, M tsp (k1, k2) represents the cost matrix of the best viewpoint of the boundary cluster from k1 to k2, V k1 represents the best viewpoint of the k1th boundary cluster, V k2 represents the best viewpoint of the k2th boundary cluster, N cls Indicates the number of boundary clusters.

6. The autonomous drone exploration method for indoor and underground unknown spaces according to claim 1 is characterized in that: Using a historical perception strategy, we record each global path, calculate the motion smoothing cost, and obtain the desired global exploration path, including the following steps: Determine the historical viewpoint clusters close to the new viewpoint cluster and calculate the distance between them; Arrange the new viewpoint clusters according to their proximity to the nearest center in the historical perception path, where the center with a distance less than d is used as the reference cluster, and the drone's position is considered as part of the reference cluster; Connect the clusters within the visited new viewpoint cluster to the reference cluster to form a continuous path; Taking adjacent reference clusters as the starting and ending points, identify the clusters that need to be visited between them, determine the exploration order, and calculate the shortest path; by connecting multiple shortest paths between reference clusters, a new history perception path is obtained; Calculate the motion consistency cost between the current UAV position and the first point of the new historical perception path, calculate the motion smoothing cost, and obtain the desired global exploration path.

7. An autonomous drone exploration system for indoor and underground unknown spaces, characterized by: include: Mapping module: The drone obtains the binocular RGB information of the depth camera and the flight control IMU data, obtains the current posture, and constructs the ESDF map using the depth point cloud extracted from the depth image; Boundary search and viewpoint generation module: The boundary module performs boundary search in the environment map and proposes a hybrid viewpoint generation method, combined with cube detection, to generate high-quality candidate viewpoints; Path Planning Module: This module determines the order of boundary exploration based on the boundary priority principle and small area detection and telemetry. It then proposes a history perception strategy based on smooth correction to determine the final global planning path and uses the LKH algorithm to solve the boundary viewpoint traversal problem. Trajectory optimization module: The trajectory optimization module uses a cubic uniform B-spline curve and introduces a penalty function to optimize the trajectory to obtain the desired trajectory; Flight control module: used to control the drone to fly according to the desired flight trajectory.

8. A drone, characterized in that: The device is equipped with the UAV autonomous exploration system for indoor and underground unknown spaces as described in claim 7.

9. An execution device comprising a memory and a processor, wherein the memory stores program instructions executed by the processor, wherein: The processor calls program instructions to execute the steps of the autonomous exploration method of a drone for indoor and underground unknown spaces according to any one of claims 1 to 6.

10. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium stores computer instructions, which enable the computer to execute the steps of the drone autonomous exploration method for indoor and underground unknown spaces as described in any one of claims 1 to 6.

Images