Air-ground autonomous collaborative inspection method and device

Abstract

The present invention relates to an air-ground autonomous collaborative inspection method and device. According to the lidar and visible light image data obtained by the air-ground collaborative equipment, an inspection environment map is dynamically generated, and the inspection behavior is autonomously decided in real time based on the geometric and semantic information of the environment, and the inspection path is generated and executed, so as to realize air-ground autonomous collaborative inspection. The present invention combines the sensing and behavior capabilities of air-ground equipment, and through dynamic mapping and autonomous decision-making, realizes fully autonomous air-ground collaborative inspection of the inspection area. At the same time, according to multi-modal sensing data, key points of inspection target facilities are extracted and inspected, improving the pertinence and efficiency of the inspection system. The invention is applied to fields such as major infrastructure inspection and power inspection, and significantly improves the autonomy and accuracy of unmanned inspection through autonomous sensing, decision-making and collaboration.

Description

Technical Field

[0001] The present invention relates to the field of unmanned inspection, and particularly to a method and device for autonomous collaborative inspection between air and ground. This technology is mainly applied to fields such as infrastructure inspection and power inspection. Background Art

[0002] With the rapid development of automation and intelligent technologies, unmanned operation has become the development trend of inspection tasks. It has become particularly crucial to improve the efficiency, accuracy, and autonomy of air-ground collaborative inspection. Air-ground collaborative inspection is widely used in major infrastructure inspection, power inspection, farmland monitoring, and emergency rescue. Especially in the rapid inspection of complex and wide-area scenarios, it has the advantages of high efficiency, low cost, and high safety compared with traditional manual inspection methods.

[0003] However, existing inspection systems rely on a single unmanned platform and face challenges in terms of efficiency, stability, and endurance, making it difficult to achieve efficient cooperation and precise inspection in complex and changing environments. On the one hand, the drone inspection system is limited by its endurance, payload, and wind resistance, and the inspection efficiency and stability are restricted. On the other hand, the ground robot inspection system has high positioning and navigation accuracy and endurance, but its passing ability is limited by terrain, making it difficult to fully obtain inspection target information.

[0004] In the face of the above challenges, the technology of autonomous and efficient air-ground collaborative inspection has emerged. By integrating the advantages of drones and unmanned vehicles and through efficient air-ground cooperation and autonomous planning, the autonomy, efficiency, and accuracy of the inspection system in wide-area complex inspection scenarios are improved. This technology has broad application prospects in multiple fields such as major infrastructure inspection, power inspection, farmland monitoring, and emergency rescue. Summary of the Invention

[0005] In view of the limitations of the existing technology, the present invention proposes a method and device for autonomous and efficient air-ground collaborative inspection. This technology realizes more efficient, stable, and flexible autonomous air-ground collaborative inspection by fusing data from three different types of sensors collected by the device, namely visible light image data, LiDAR (Light Detection and Ranging) point cloud data, and inertial measurement unit (IMU) measurement information.

[0006] The autonomous and efficient air-ground collaborative inspection method designed by the present invention includes the following steps:

[0007] Generate a three-dimensional voxel map using the air-ground equipment data of the inspection area;

[0008] Generate a candidate target set, including the key points of the inspection facilities identified from the three-dimensional voxel map and the sampling viewpoints extracted from the map boundary;

[0009] Define a gain function. By evaluating the gains of each target in the candidate target set, the most valuable inspection target is allocated to each device in the air-ground cluster in real time, and an inspection path is dynamically generated. The gain function includes:

[0010] Geometric gain calculation, including calculating the volume of the front cluster and the path length to the optimal viewing point of the front cluster;

[0011] Semantic gain calculation, adding gains to points with rich semantic information, including structural key points, power facility insulators, etc.;

[0012] Multi-robot collaborative gain, guiding the robot away from the positions and target points of other robots in the cluster and tending to move towards the target point assigned to itself;

[0013] Terrain analysis, iteratively maximizing the probability of reaching the inspection target, dynamically making decisions on the current action strategy, generating a motion trajectory, and controlling the robot to move towards the target point stably and efficiently. Further, obtain the visible light image data, lidar data, and IMU pose information of the air-ground devices in the inspection area, and use the multi-robot SLAM method to uniformly transform the lidar point cloud of the air-ground devices into the map coordinate system, incrementally generate the point cloud map of the inspection area, and then incrementally generate the three-dimensional voxel map.

[0014] Further, when generating the three-dimensional voxel map, use the ray casting method. According to the obstacle point cloud data obtained by the air-ground device, integrate the new obstacle information into the three-dimensional voxel map, and mark all newly detected voxels as free or occupied; the Euclidean signed distance field (ESDF) value between each voxel and the nearest obstacle is recorded in the three-dimensional voxel map. For the occupied state, the ESDF value is set to 0, and for the free state, the distance from it to the nearest obstacle is recalculated and set as the ESDF value.

[0015] The specific description of the front search and clustering is as follows: First, during each update iteration, the frontiers within the current frame's perception range are deleted, and the breadth-first search (BFS) is used to detect and cluster new boundaries. This strategy avoids the need to check the entire map and precisely compare the state of each voxel during each iteration, both of which consume a large amount of computing power. Subsequently, frontiers larger than the threshold are recursively split into frontiers with moderate volumes through principal component analysis (PCA) to achieve more refined target extraction.

[0016] The specific description of the viewing point sampling method is as follows: The viewing point set is the candidate positions and directions that cover the frontiers using a specific sensor, serving as the current specific inspection targets. It is a set of viewing points uniformly sampled in cylindrical coordinates around the center of the front cluster. According to the gain function For the center of the front cluster A set of uniformly sampled viewing points in cylindrical coordinates. According to the gain function Evaluate and select the best viewpoints as the covering front for the specific poses.

[0017] Furthermore, the key points of the inspection facilities are extracted in the 3D voxel map by the UPKD method.

[0018] Furthermore, the real-time autonomous planning of the aerial-ground collaborative inspection path and behavior. Define a gain function to identify the most valuable inspection targets, and allocate inspection targets to each device in the aerial-ground cluster in real time, dynamically identify the most valuable exploration targets to promote an efficient collaborative autonomous inspection process. The gain function is a gain function considering geometry, semantics, and collaboration. Considering topological reachability and the semantic information of inspection key points, the aerial-ground collaborative decision-making inspection behavior pattern is used to evaluate and select the optimal viewpoints. According to the environmental voxel map, considering the passing capabilities of aerial and ground devices, the inspection path is dynamically generated.

[0019] The calculation process of the gain function is as follows:

[0020]

[0021] represents the set of front clusters, represents the state of the robot, represents the set of states of all robots, is the information gain of the viewpoint, is the path length; in terms of collaboration, captures the influence of other robots. In terms of inspection semantics, evaluates the inspection semantic value of the target point, and the key points have higher scores; , , , are the corresponding positive constant weights.

[0022] Even further, the specific calculation of the geometric gain is as follows:

[0023]

[0024]

[0025] Among them, is the information gain of the viewpoint, is the path length, represents the viewpoint calculated by the ray casting method, which is the volume of the front cluster that a specific sensor can cover, represents the current position of the robot going to the viewpoint 3D A-star path

[0026] The multi-robot collaborative gain is specifically defined as:

[0027]

[0028] Among them, represents the set of viewpoints, represents the set of current target points of all robots, and are the attractive potential field and the repulsive potential field respectively, represents robot 's current position, represents its target point, represents the target points of robots in the open space cluster except , represents the corresponding position set, and represent the normal constant weight values of the corresponding items. Through this item, the robot can be guided to stay away from the positions and target points of other robots and tend to go to the target point assigned to itself.

[0029] Furthermore, terrain analysis and trajectory generation navigate the unmanned cluster to the target location safely and efficiently. By iteratively maximizing the probability of reaching the target viewpoint , dynamically decide the current action strategy , generate linear velocity and angular velocity commands, and control the robot to move to the target point stably and efficiently. The specific calculation is as follows:

[0030]

[0031] Use Monte Carlo sampling to estimate the probability density. Specifically, generate from samples , and calculate the probability of reaching the target point according to the unoccluded ratio of the Monte Carlo sampling path:

[0032]

[0033] Among them represents the currently calculated state, represents the corresponding intermediate state, represents the path.

[0034] According to the Monte Carlo algorithm, the success probability can be obtained through path sampling, that is, sample n paths for the state , and calculate the success arrival probability respectively. The mathematical expectation of it is the state , Arrival probability

[0035]

[0036] Occlusion function Characterize the path of the occluded situation, indicates that the path is not occluded, otherwise

[0037]

[0038] Therefore, the state Arrival probability Can be calculated as:

[0039] .

[0040] Based on the same inventive concept, the present invention also discloses an air-ground autonomous collaborative inspection device, including a number of unmanned vehicles and unmanned aerial vehicles networked through MESH. Both the unmanned vehicles and unmanned aerial vehicles are equipped with multi-light sensors, including laser, visible light, and infrared bands. The networked unmanned vehicles and unmanned aerial vehicles perform air-ground autonomous collaborative inspection by an air-ground autonomous collaborative inspection method.

[0041] Based on the same inventive concept, the present invention also discloses a computer-readable medium, on which a computer program is stored. When the program is executed by a processor, it implements the air-ground autonomous collaborative inspection method.

[0042] The present invention designs an air-ground autonomous and efficient collaborative inspection device, which mainly includes two unmanned platforms. The first unmanned platform is an unmanned aerial vehicle. A quadrotor unmanned aerial vehicle is built, which has operation capabilities such as autonomous flight, edge computing, and performing task actions. Develop an integrated multi-light sensor unmanned aerial vehicle payload, including laser and visible light. The second unmanned platform is a mobile hangar, based on an unmanned vehicle module, which can perform high-precision positioning and mapping, and autonomous path planning in an indoor scene without GNSS conditions, realizing efficient exploration of the indoor space. At the same time, it has an unmanned aerial vehicle takeoff and landing platform and a charging module to improve the flexibility and endurance of the unmanned aerial vehicle. Through MESH self-networking, a combined system of unmanned aerial vehicles and unmanned vehicles is formed to collaboratively perform sensing and mapping, accurately collect multi-source data. The data output by the payload includes information such as visible light and the corresponding position, and the data is synchronously transmitted to the inspection platform in real time, laying a solid equipment and data foundation for the inspection of major facilities.

[0043] The advantages of the present invention are as follows:

[0044] Design an air-ground collaborative autonomous and efficient inspection method to assist in the high-efficiency, unmanned, and automated inspection tasks through air-ground collaboration. For example, the complex conditions of mountainous terrains pose significant challenges to the efficiency and safety of traditional manual inspections for infrastructure and power facilities. However, the air-ground cross-domain inspection method composed of drones and unmanned vehicles can greatly improve the inspection efficiency and safety in challenging scenarios.

[0045] The present invention integrates the sensing capabilities, passing capabilities, and task execution capabilities of drones and unmanned vehicles. Through efficient autonomous sensing, mapping, and planning algorithms, it not only overcomes the limitations of individual devices but also provides highly accurate and reliable autonomous inspection solutions in various complex environments. The present invention can completely autonomously complete the air-ground collaborative inspection process without a pre-defined inspection path or pre-defined inspection sub-areas. First, it can autonomously sense and map to draw a map of the unknown inspection area to obtain complete environmental information. Second, it can autonomously adjust and dynamically plan the inspection targets of each device in real time based on the obtained local environmental geometric and semantic information and the status of the air-ground devices themselves, and can focus on the key points of the inspection target facilities. Finally, the path optimization module can autonomously plan the path to the current target point, generate corresponding control instructions, and ensure that the air-ground devices complete the inspection tasks autonomously, efficiently, and safely.

[0046] The air-ground autonomous collaborative and efficient inspection device designed by the present invention focuses on major problems such as the precise positioning and autonomous flight of drones in complex inspection environments, breaks through the key technical bottlenecks of multi-modal panoramic perception of unmanned clusters, joint pose state estimation of multi-sensors, multi-aircraft air-ground collaborative positioning and mapping, and autonomous exploration of unmanned systems for task characteristics, couples multiple environmental perception means to recognize indoor spaces, realizes unmanned, autonomous, and collaborative positioning and mapping, establishes a multi-level environmental situation semantic knowledge graph, and forms an intelligent mapping and remote sensing method and technology system for unmanned clusters in complex indoor and outdoor spaces to assist inspection personnel in working efficiently. Brief Description of the Drawings

[0047] Figure 1 Flowchart of the air-ground autonomous collaborative and efficient inspection method.

[0048] Figure 2 Flowchart of key inspection target extraction and viewpoint sampling.

[0049] Figure 3 Flowchart of gain function calculation.

[0050] Figure 4 Schematic diagram of the air-ground autonomous collaborative and efficient inspection device.

[0051] Figure 5 Schematic diagram of air-ground autonomous collaborative and efficient inspection of a substation. Detailed Embodiments

[0052] The technical solution of the present invention will be described below in conjunction with the accompanying drawings and embodiments.

[0053] Embodiment 1

[0054] Taking the power scenario as an example, refer to Figure 1 and Figure 5 The method for autonomous collaborative inspection of air and ground disclosed in this embodiment includes the following steps:

[0055] Step 1: Obtain visible light image data, lidar data, and IMU pose information related to the positioning target in the air-ground collaborative device cluster.

[0056] Step 2: Dynamically generate an inspection environment map. According to the environmental data obtained by the air-ground collaborative devices, incrementally generate a point cloud map of the inspection area, and then incrementally generate a three-dimensional voxel map. First, according to the environmental laser point clouds obtained by the air-ground devices at different positions and different perspectives, use the multi-robot SLAM method to uniformly transform them into the map coordinate system, thereby generating a comprehensive point cloud map of the inspection environment. Subsequently, calculate the three-dimensional voxel map from the collaboratively generated point cloud map, that is, divide the environment into uniform three-dimensional grids, and each grid is a voxel. Using the ray casting method, according to the obstacle point cloud data obtained by the air-ground devices, integrate the newly added obstacle information into the three-dimensional voxel map, and mark all newly detected voxels as free or occupied states. There are three states of voxels in the voxel map: unknown, free, and occupied, and the two non-unknown states are called known. In a dynamic environment, when incrementally constructing the map, the algorithm will add or remove obstacles in real time. Subsequently, update the additional Euclidean Signed Distance Functions (ESDF) accordingly. Each voxel in the three-dimensional voxel map records the distance to the nearest obstacle as the ESDF value. The ESDF value update process is as follows. Each voxel whose state is updated to occupied or free will be added to two queues named insertQueue and deleteQueue respectively. For the elements in insertQueue, set the ESDF to 0. For each voxel in deleteQueue, re-search and calculate its distance to the nearest obstacle as value.

[0057] Step 3: Extract key points of inspection facilities and sample boundary viewpoints. As shown in the appendix Figure 2As shown, this key inspection target extraction method aims to achieve two balanced tasks: one is to autonomously obtain the 3D map of the inspection area, and the other is to inspect key semantic targets. During the acquisition of the 3D map, frontier clusters are generated through the clustering of boundary voxels to guide the aerial and ground equipment to autonomously explore unknown areas and construct the environmental map of the inspection area in an incremental manner. In terms of the inspection of key semantic targets, based on visible light images and lidar data, the UPKD method is used to extract the key points of inspection facilities and prioritize the processing of these key points. Combining these two aspects of goals forms a candidate target set, which includes 1) viewpoints sampled around the frontier clusters and 2) semantic key points of inspection facilities.

[0058] Step 3.1, during the acquisition of the 3D map, frontier clusters are generated through the search and clustering of boundary voxels. These frontier clusters represent the boundary between the free space and the unknown space, where the free space represents the set of voxels that are marked as known and unoccupied, and can effectively guide the aerial and ground collaborative equipment to autonomously explore the uncovered areas, such as drones and unmanned vehicles. Specifically, when the aerial and ground equipment performs the inspection task, the system first identifies the boundary of the unknown space and forms frontier clusters by spatially clustering the boundary voxels. Subsequently, a cylindrical coordinate system is constructed with the cluster center, and viewpoints are sampled within a certain radius as candidate target points. In each iteration, only the voxels within the current frame's perception range are updated. This strategy significantly reduces the consumption of computing resources by avoiding a comprehensive inspection of the entire map M and a precise comparison of the status of each voxel one by one in each iteration. Only the changing areas need to be updated, skipping the redundant processing of the entire map, thereby improving the system's operation efficiency and reducing the computing overhead.

[0059] Step 3.2, the UPKD method is divided into two processing stages. In the first stage, the UPKD network processes the point cloud to generate candidate inspection points, which consists of two main parts: a data normalization module and an unsupervised key point detection network (UKD-Net). The data normalization module compresses information according to the symmetric structure of the tower to reduce the instability of key point detection. The UKD-Net contains a point transformer layer that uses the self-attention mechanism to extract features from the point cloud. In the second stage, a convex optimization strategy is used to filter and obtain the inspection points.

[0060] Step 4, the paths and behaviors of aerial and ground collaborative inspection are autonomously planned by the algorithm and adjusted in real time. To achieve aerial and ground collaborative inspection, the algorithm assigns inspection targets to each device in the aerial and ground cluster in real time. The aerial and ground equipment selects the most valuable inspection target at present in real time according to its own position, the target position, the path length, and the importance of the target semantic information. At the same time, in order to achieve effective collaboration, the positions and targets of other devices are considered, and the equipment preferentially selects inspection targets in other areas to promote an efficient decentralized autonomous inspection process.

[0061] To achieve this goal, a gain function is defined, which comprehensively considers geometric features, semantic information, and multi-robot collaboration benefits. The gain function G can identify the most valuable inspection targets. During the planning process, the system not only takes into account geometric accessibility but also combines the semantic information of inspection key points to jointly decide the inspection modes of aerial and ground devices. In addition, by evaluating the gains of different viewpoints, the system can select the best observation positions and postures. Based on the voxel map of the environment and combined with the passing capabilities of aerial and ground devices, the system dynamically generates inspection paths. The calculation process of the gain function is as follows Figure 3 as shown

[0062]

[0063] represents the set of frontier clusters, represents the set of states of all robots, represents the state of the robot. is the information gain of the viewpoint, is the path length. In terms of collaboration, captures the influence of other robots. In terms of inspection semantics, evaluates the inspection semantic value of the target point, and key points have higher scores. , , , are the corresponding positive constant weights.

[0064] Step 4.1, Geometric gain calculation. Geometric gain usually considers the volume of the frontier clusters and the path length to the optimal viewpoint of the frontier clusters. The specific calculation is as follows:

[0065]

[0066]

[0067] where represents the viewpoint is the volume of the frontier clusters that can be covered by a specific sensor calculated according to the ray casting method. represents the 3D A-star path from the current position of the robot to the viewpoint . The A-star algorithm is a heuristic algorithm for graph search, widely used in path planning and graph traversal, which can efficiently find the shortest path from the starting point to the target point. 3D A-star refers to the version of the A-star algorithm in three-dimensional space.

[0068] Step 4.2, Semantic gain calculation. The semantic gain takes into account the semantic value of the target points, such as the key points of the building structures of major infrastructure, insulators of power facilities, etc., and attaches gains to the points rich in semantic information.

[0069] Step 4.3, Multi-robot collaborative gain. If we consider the robots , given the positions and targets of the other robots in the team, this item is defined as:

[0070]

[0071] where represents the set of viewpoints, represents the set of current target points of all robots, and are the attractive potential field and the repulsive potential field respectively, represents the robot 's current position, represents its target point, represents the target points of the robots in the open space cluster except , represents the corresponding position set. Through this item, the robot can be guided to stay away from the positions and target points of other robots and tend to move towards the target point assigned to itself, achieving effective autonomous inspection of the air-ground collaboration. In addition, and represent the normal constant weight values of the corresponding items.

[0072] Step 5, Through terrain analysis and trajectory generation, ensure that the UAV cluster can navigate to the target position safely and efficiently. The system iteratively optimizes the current action strategy to maximize the probability of reaching the target viewpoints , thereby generating linear velocity and angular velocity commands to control the robot to move steadily and efficiently towards the target point.

[0073] According to the conditional probability formula, the success probability can be calculated as the integral of the product of conditional probabilities , where represents the currently calculated state, represents the corresponding intermediate state,

[0074]

[0075] According to the Monte Carlo algorithm, the above probability can be obtained through path sampling, that is, sampling n paths for the state , and calculating the success probability of reaching respectively. The mathematical expectation of which is the arrival probability of the state

[0076]

[0077] Occlusion function Characterization path of the occluded situation, indicates that the path is not occluded, otherwise

[0078]

[0079] Therefore, the state of the arrival probability can be calculated as:

[0080] .

[0081] Embodiment 2

[0082] Based on the same inventive concept, this embodiment discloses an air-ground autonomous collaborative inspection device, as shown in the appendix Figure 4 shown, including a number of unmanned vehicles and drones networked through MESH. Both the unmanned vehicles and drones are equipped with multi-light sensors, including laser, visible light, and infrared bands; the networked unmanned vehicles and drones perform air-ground autonomous collaborative inspection by the air-ground autonomous collaborative inspection method.

[0083] Since the device introduced in Embodiment 2 of the present invention is the device used to implement the air-ground autonomous collaborative inspection method in Embodiment 1 of the present invention, based on the method introduced in Embodiment 1 of the present invention, those skilled in the art can understand the specific structure and deformation of the electronic device, so it will not be elaborated here. Any electronic device used in the method of Embodiment 1 of the present invention belongs to the scope of protection of the present invention.

[0084] Embodiment 3

[0085] Based on the same inventive concept, the present invention also provides a computer-readable medium, on which a computer program is stored, and when the program is executed by a processor, it implements the method described in Embodiment 1.

[0086] Since the device introduced in Embodiment 3 of the present invention is the computer-readable medium used to implement the air-ground autonomous collaborative inspection method in Embodiment 1 of the present invention, based on the method introduced in Embodiment 1 of the present invention, those skilled in the art can understand the specific structure and deformation of the electronic device, so it will not be elaborated here. Any electronic device used in the method of Embodiment 1 of the present invention belongs to the scope of protection of the present invention.

[0087] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Those skilled in the art to which the present invention pertains may make various modifications or supplements to the described specific embodiments or use similar means for substitution, but will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. An air-ground autonomous collaborative inspection method, characterized in that: The following steps are involved: Generate a 3D voxel map using the open space equipment data in the inspection area; Generate a candidate target set, including inspection facility key points identified from the three-dimensional voxel map and sampled viewpoints extracted from the map boundary; Define a gain function, and by evaluating the gain of each target in the candidate target set, assign the most valuable inspection target to each device in the air-ground cluster in real time, and dynamically generate an inspection path. The gain function includes: Geometric gain calculation, including calculating the volume of a frontier cluster and the path length to the optimal viewpoint of the frontier cluster, the frontier cluster being generated by searching and clustering boundary voxels; Semantic gain calculation, adding gain to points with rich semantic information, including key points of structures and insulators of power facilities; The multi-machine collaborative gain is defined as follows: Among them, ξ c represents a set of viewpoints, Represents the current target point set of all robots, U a and U r are the attractive potential field and the repulsive potential field respectively, x i Indicates the current position of robot i, represents its target point, represents the robot target point other than i in the open space cluster, represents the corresponding position set, k a and κ r represents the positive constant weight of the corresponding item, through which the robot can be guided away from the positions and target points of other robots and tend to go to the target point assigned to itself; The gain function calculation process is as follows: G(C,S i )=ω area G area (x) C,best )-oh dis J dis (p i ,x C,best )-oh other J other (S) +oh semantic G semantic (x) C,best ) C represents the frontier cluster set, p i Indicates the current position, S i represents the state of the robot, S represents the state set of all robots, ξ C,best represents the viewpoint, G area is the information gain of the viewpoint, J dis is the path length; in multi-machine cooperative gain, J other Capturing the influence of other robots, in semantic gain, G semantic Evaluate the inspection semantic value of target points, with higher scores for key structural points and insulators of power facilities; area ,ω dis ,ω other ,ω semantic is the corresponding positive constant weight; Analyze terrain, calculate the maximum probability of reaching the inspection target, dynamically decide the current action strategy, generate the motion trajectory, and control the robot to the target point.

2. The air-ground autonomous collaborative inspection method according to claim 1 is characterized by: Obtain visible light image data, lidar data and IMU pose information of air-ground equipment in the inspection area, and use the multi-machine SLAM method to uniformly convert the laser point cloud of air-ground equipment into the map coordinate system, incrementally generate a point cloud map of the inspection area, and then incrementally generate a three-dimensional voxel map.

3. The air-ground autonomous collaborative inspection method according to claim 2 is characterized in that: When generating a three-dimensional voxel map, the ray casting method is used to integrate the newly added obstacle information into the three-dimensional voxel map according to the obstacle point cloud data obtained by the air-ground equipment, and all newly detected voxels are marked as free or occupied; the Euclidean signed distance field ESDF value between each voxel and the nearest obstacle is recorded in the three-dimensional voxel map. In the occupied state, the ESDF value is set to 0, and in the free state, the distance to the nearest obstacle is re-searched and calculated as the ESDF value.

4. The air-ground autonomous collaborative inspection method according to claim 1 is characterized in that: The key points of the inspection facilities are extracted from the three-dimensional voxel map using the UPKD method.

5. The air-ground autonomous collaborative inspection method according to claim 1 is characterized by: The sampling viewpoint generates a frontier cluster by searching and clustering the boundary voxels of the three-dimensional voxel map, constructs a cylindrical coordinate system with the cluster center, and obtains the sampling viewpoint within a set radius.

6. The air-ground autonomous collaborative inspection method according to claim 1 is characterized by: The geometric gain calculation is specifically as follows: G area (x) C,best )=V raycast (x) C,best ) J dis (p i ,x C,best )=∑path(p i ,x C,best ) Among them, G area is the information gain of the viewpoint, J dis is the path length, V raycast Indicates the viewpoint ξ C,best The frontier cluster volume that a particular sensor can cover, calculated by the ray casting method, path(p i ,ξ C,best ) represents the robot’s current position p i Go to viewpoint C,best 3D A-star path.

7. The air-ground autonomous collaborative inspection method according to claim 1 is characterized by: The specific process of generating motion trajectory is: G Calculated as the conditional probability product P G (s f )P(s f ∣s s ), where s s Indicates the current calculation status, s f represents the corresponding intermediate state, According to the Monte Carlo algorithm, the success probability can be obtained by path sampling, that is, for state s s Sample n paths θ i , calculate the probability of successful arrival P respectively G (θ i ), whose mathematical expectation is the state s s The arrival probability P G (s s ) Occlusion function c(θ i ) represents the path θ i The occlusion situation, c(θ i )=1 means the path is not blocked, otherwise c(θ i )=0 Therefore, the state s s The arrival probability P G (s s ) can be calculated as:

8. An air-ground autonomous collaborative inspection device, characterized in that: It includes several unmanned vehicles and drones networked through MESH, and the unmanned vehicles and drones are equipped with multi-light sensors, including laser, visible light, and infrared bands; the networked unmanned vehicles and drones perform air-ground autonomous collaborative inspection by the air-ground autonomous collaborative inspection method described in any one of claims 1-7.

9. A computer readable medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, the air-ground autonomous collaborative inspection method as described in any one of claims 1-7 is implemented.

Images