Unmanned aerial vehicle autonomous target searching method based on viewpoint scoring and dynamic replanning

Abstract

The invention discloses an unmanned aerial vehicle autonomous target searching method based on viewpoint scoring and dynamic re-planning, which comprises the following steps: 1) constructing an occupied grid map in real time through an RGBD camera of an unmanned aerial vehicle, and dividing a space into an observable area and an unobserved area according to distance information; 2) selecting an optimal observation viewpoint; 3) clustering the selected observation viewpoints through a collision-free viewpoint clustering method, and clustering the observation viewpoints in the step 2) into a plurality of collision-free subsets through obstacle detection; 4) planning a global path which starts from the current position and passes through all viewpoint clustering centers through an unmanned aerial vehicle dynamic re-planning method; and 5) obtaining the pixel coordinates of the target in the RGB image, obtaining the three-dimensional coordinates of the target, displaying the searched target in the Rviz, and marking the global coordinates of the target.. The method solves the problems of inconsistent path planning of the unmanned aerial vehicle and incomplete search area, thereby comprehensively improving the success rate and efficiency of target search of the unmanned aerial vehicle.

Description

Technical Field

[0001] The present invention relates to autonomous target search for unmanned aerial vehicles (UAVs), and in particular to an autonomous target search method for UAVs based on viewpoint scoring and dynamic replanning. Background Art

[0002] With the development of the low-altitude economy, the application of unmanned aerial vehicles (UAVs) in a variety of fields, including disaster relief, industrial inspection, and mine safety, is expanding. In these tasks, drones can effectively replace humans in exploring completely unknown and dangerous environments. However, simple spatial exploration is often not the ultimate goal of research, but rather a key step in achieving more complex tasks. The value of exploration itself lies in supporting other technologies. Through deep integration with technologies such as semantic understanding, object detection, and environmental interaction, robots can be given broader capabilities. In these applications, a typical capability of UAVs is autonomous target search in unknown environments. UAVs must map the unknown environment to identify and locate targets of interest while avoiding obstacles. Traditional autonomous exploration methods are simply combined with target recognition technologies (such as YOLO) for autonomous target search in unknown and complex environments. While theoretically achieving both environmental coverage and target detection is possible, the design goals and operating mechanisms of the two differ fundamentally. The core goal of autonomous exploration is to efficiently explore unknown space and reduce unperceived areas. It is often assumed that targets or key information are more likely to appear at the unexplored boundary, while paying less attention to the perceived occupied area. Target recognition, on the other hand, requires the drone to pause, adjust its posture, or adjust its viewing angle to perceive the target. This difference in goal orientation leads to decision conflicts. If exploration is prioritized, the target occupying the area may be missed; if detection is prioritized, the exploration progress may stagnate and the environment coverage may be incomplete.

[0003] The following defects still exist in the UAV target search scenario:

[0004] Search strategies lack specificity: Traditional autonomous exploration focuses primarily on generating exploration viewpoints at the boundaries, but does not specifically generate viewpoints for occupied areas, let alone plan viewpoints for occupied areas, resulting in a lack of goal orientation in the exploration process.

[0005] Simplified viewpoint selection strategy: Traditional viewpoint extraction uses a random extraction method without considering the target distribution characteristics, resulting in low search efficiency and missed target detection.

[0006] Inconsistent Path Planning: Traditional methods replan the global path after updating the environmental map. This frequent path update causes the drone to frequently change flight direction, resulting in "path inconsistency" and a large number of repeated paths. This reduces the drone's observation time in certain key areas, reducing search success rate and efficiency. Summary of the Invention

[0007] The present invention aims to address the problems of decision-making conflicts, simplistic viewpoint selection strategies, and inconsistent path planning caused by the simple combination of traditional UAV autonomous exploration and target detection. This method provides a method for autonomous target search for UAVs based on viewpoint scoring and dynamic replanning. This method resolves target-directed conflicts by generating search viewpoints primarily within the occupied area. High-quality search viewpoints are selected by comprehensively considering multiple factors, including information gain, detection viewing angle, inspection distance, and detection height. Collision-free viewpoint clustering and a dynamic replanning strategy address inconsistent UAV path planning and incomplete search areas, thereby comprehensively improving the success rate and efficiency of UAV target search.

[0008] The technical solution for achieving the purpose of the present invention is:

[0009] A method for autonomous target search for an unmanned aerial vehicle (UAV) based on viewpoint scoring and dynamic replanning, the method comprising the following steps:

[0010] 1) Using the drone's RGBD camera, an occupancy grid map is constructed in real time. The space perceived by the RGBD camera is divided into observable and unobservable areas based on distance information. Observation viewpoints are generated within the observable area to avoid missed detection or misidentification of targets due to observation distances exceeding the accuracy of the RGBD camera.

[0011] 2) Score the generated observation viewpoints by considering information gain, viewing angle, distance, and height, and select the best observation viewpoint;

[0012] 3) Clustering the observation viewpoints selected in step 2) using a collision-free viewpoint clustering method. Clustering the optimal observation viewpoints selected in step 2) into multiple collision-free subsets through obstacle detection to enhance viewpoint collaboration and improve adaptability to dynamic environments, reduce ineffective UAV maneuvers, and improve the efficiency and success rate of target search.

[0013] 4) A global path is planned from the current position through the center of all viewpoint clusters through the UAV dynamic replanning method. This enhances the consistency of the UAV path planning, avoids the UAV frequently changing flight direction and misses targets hidden in complex terrain, and improves target search.

[0014] 5) Use the Yolo-Fastestv2 target recognition algorithm to identify targets in the observable area, obtain the pixel coordinates of the target in the RGB image, combine the drone pose and camera coordinates to derive the target's three-dimensional coordinates, and display the searched target in Rviz, while marking the global coordinates of the searched target.

[0015] The occupancy grid map in step 1) specifically refers to collecting RGB images and depth information of the target area through an RGBD camera, mapping the depth information to the space observed by the RGBD camera and discretizing it into grid cells, where each grid cell is marked as occupied (obstacle) or free (passable), thereby constructing a grid map representing environmental obstacles and free space.

[0016] The process of space division in step 1) is as follows: the RGBD camera obtains the depth information of the surrounding environment in real time, and divides the space observed by the RGBD camera into free space and occupy space Then, the voxels that have been mapped as occupied are projected into the camera coordinate system, and the closest observation distance of the voxels that are within the camera field of view (FOV) and not blocked by any occupied space is calculated. The closest observation distance is greater than the maximum observation distance. The voxels are marked as unobserved areas , the closest observation distance is less than the maximum observation distance The voxels are marked as observable areas .

[0017] The process of generating observation viewpoints in step 1) is as follows: uniformly sampling candidate viewpoints in a cylindrical coordinate system near the occupied space, optimizing the yaw angle to maximize the RGBD camera coverage, and selecting the viewpoint with the highest coverage.

[0018] The viewpoint scoring process in step 2) includes scoring all viewpoints in each cluster using four criteria: information gain, viewing angle, height, and distance:

[0019] Information gain: The information gain obtained by the drone at each viewpoint is calculated by the number of unobserved boundaries of the RGBD camera. Decision, that is

[0020]

[0021] in Indicates the score impact factor, through Control the influence of information gain on the score;

[0022] Perspective: To avoid the drone missing targets due to extreme perspectives, the average normal of each obstacle's occupied grid cells is calculated. , according to the obstacle center To Viewpoint Vector The average normal of the grid cells occupied by each obstacle The viewing angle is scored based on the closeness of:

[0023]

[0024] Distance: Use the Gaussian function to measure the deviation between the viewpoint distance and the optimal distance to dynamically adjust the score. Control the sensitivity of distance deviation to adapt to different task requirements, using the adjustment factor of distance effect To control the influence of distance factors on the score;

[0025]

[0026] in is the optimal observation distance, is the distance from the viewpoint to the obstacle;

[0027] Height: Use Gaussian function to measure the deviation between the viewpoint height and the optimal height to dynamically adjust the score. Control the sensitivity of altitude deviation to suit different mission requirements, using an altitude impact adjustment factor To control the influence of height on the score;

[0028]

[0029] in is the best observation altitude, is the height from the viewpoint to the ground;

[0030] Final score for each viewpoint It is calculated as follows:

[0031]

[0032] The process of the collision-free viewpoint clustering method in step 3) is as follows: the current position of the drone is used as the starting point of the first cluster, and clustering is completed step by step. In each clustering process, the center of the current cluster is used as the starting point, and viewpoints that are not blocked by obstacles and are within the specified radius R are added to the current cluster through obstacle collision detection. The cluster center is dynamically updated to the average position of all viewpoints in the current cluster. After the current clustering is completed, the nearest point is selected from the remaining viewpoints as the starting point of the new cluster, and the above clustering process is repeated until all viewpoints are assigned to the cluster.

[0033] The obstacle detection in step 3) is as follows: connecting the generated different viewpoints. If the line between the two viewpoints is not a straight line, it means that there is an obstacle blocking the view.

[0034] The steps of step 4) the UAV dynamic replanning method are as follows:

[0035] 4.1) In each global search path planning, the improved traveling salesman problem (ATSP) algorithm is used to generate a historical shortest path consisting of the centers of all current viewpoint clusters. ,The UAV first searches in the order of the historical shortest path, where, represents the nth viewpoint in the path. The UAV visits each viewpoint in the path in turn and scans the observable area of ​​the viewpoint;

[0036] 4.2) As the path is executed, the environment map will be continuously updated. Once the drone detects a new area during flight , these new perspectives will be Add to the list of tasks to be tested , but the order of the original path will be retained, and the viewpoint search in the original path will be completed first;

[0037] 4.3) After the historical path is completed, the drone will dynamically re-search the newly discovered area;

[0038] 4.4) After the drone has searched the newly discovered area, it will continue to use the return viewpoint as the starting point to generate a new round of historical shortest paths. Repeat steps 4.1) to 4.3) until all areas have been searched.

[0039] The steps of calculating the target three-dimensional coordinates in step 5) are as follows:

[0040] 5.1) Obtaining target pixel coordinates and depth information: Using an RGBD camera to capture the RGB image and depth information of the target area;

[0041] Use target detection algorithm to obtain target pixel coordinates and depth value ;

[0042] 5.2) Convert pixel coordinates to camera coordinate system: use camera intrinsic parameters, i.e. focal length , and main point , The two-dimensional pixel coordinates and depth value Convert to 3D coordinates in the camera coordinate system :

[0043]

[0044]

[0045]

[0046] Output the three-dimensional coordinates of the camera coordinate system:

[0047]

[0048] 5.3) Camera coordinate system to UAV coordinate system: using fixed rotation matrix Convert camera coordinate system coordinates to UAV coordinate system coordinates , rotation matrix represents the installation posture of the camera relative to the UAV coordinate system:

[0049]

[0050] Rotation matrix:

[0051]

[0052] Output UAV coordinate system three-dimensional coordinates:

[0053] ;

[0054] 5.4) UAV coordinate system to world coordinate system: using UAV 3x3 pose rotation matrix and translation vector Convert UAV coordinate system coordinates to world coordinate system coordinates , where the pose rotation matrix represents the rotation relationship of the UAV relative to the world coordinate system, and the translation vector is in meters;

[0055]

[0056] Where: is the translation vector of the UAV relative to the world coordinate system, in meters, multiplied by 1000 to convert to millimeters;

[0057] Output world coordinate system three-dimensional coordinates, unit: millimeter:

[0058]

[0059] 5.5) Real-time display of Yolo-Fastestv2 target recognition algorithm in observable area in visualization tool Rviz, and world coordinate system three-dimensional coordinates of the target identified by the algorithm, to ensure the visualization and accuracy of the search results.

[0060] This technical solution solves the target-oriented conflict by generating search viewpoints mainly in the occupied area, selects high-quality search viewpoints by comprehensively considering multiple factors such as information gain, detection viewing angle, inspection distance and detection height, and solves the problems of inconsistent UAV path planning and incomplete search area through collision-free viewpoint clustering and dynamic replanning strategy, thereby comprehensively improving the success rate and efficiency of UAV target search. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 is a system block diagram of an embodiment;

[0062] Figure 2 This is a flowchart of real-time update of the drone observation area in the embodiment;

[0063] Figure 3 This is a schematic diagram of real-time update of the drone observation area in the embodiment;

[0064] Figure 4 This is a flowchart of the drone target search viewpoint scoring in the embodiment;

[0065] Figure 5 This is a flowchart of collision-free viewpoint clustering in an embodiment;

[0066] Figure 6 Schematic diagram of non-collision viewpoint clustering in the embodiment;

[0067] Figure 7 This is a flowchart of the dynamic replanning of the drone in the embodiment;

[0068] Figure 8 Schematic diagram of dynamic replanning of a drone in an embodiment. DETAILED DESCRIPTION

[0069] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the present invention is not limited thereto.

[0070] Example:

[0071] Reference Figure 1 、 Figure 2 , an autonomous target search method for a UAV based on viewpoint scoring and dynamic replanning, the method comprising the following steps:

[0072] 1) Using the drone's RGBD camera, an occupancy grid map is constructed in real time. The space perceived by the RGBD camera is divided into observable and unobservable areas based on distance information. Observation viewpoints are generated within the observable area to avoid missed detection or misidentification of targets due to observation distances exceeding the accuracy of the RGBD camera.

[0073] 2) Score the generated observation viewpoints by considering information gain, viewing angle, distance, and height, and select the best observation viewpoint;

[0074] 3) Clustering the observation viewpoints selected in step 2) using a collision-free viewpoint clustering method. Clustering the optimal observation viewpoints selected in step 2) into multiple collision-free subsets through obstacle detection to enhance viewpoint collaboration and improve adaptability to dynamic environments, reduce ineffective UAV maneuvers, and improve the efficiency and success rate of target search.

[0075] 4) A global path is planned from the current position through the center of all viewpoint clusters through the UAV dynamic replanning method. This enhances the consistency of the UAV path planning, avoids the UAV frequently changing flight direction and misses targets hidden in complex terrain, and improves target search.

[0076] 5) Use the Yolo-Fastestv2 target recognition algorithm to identify targets in the observable area, obtain the pixel coordinates of the target in the RGB image, combine the drone pose and camera coordinates to derive the target's three-dimensional coordinates, and display the searched target in Rviz, while marking the global coordinates of the searched target.

[0077] The occupancy grid map in step 1) specifically refers to collecting RGB images and depth information of the target area through an RGBD camera, mapping the depth information to the space observed by the RGBD camera and discretizing it into grid units, namely voxels, where each grid unit is marked as occupied (obstacle) or free (passable), thereby constructing a grid map representing environmental obstacles and free space.

[0078] like Figure 3 As shown in the figure, the process of space division in step 1) is as follows: the RGBD camera obtains the depth information of the surrounding environment in real time, and divides the space observed by the RGBD camera into free space and occupy space Then, the voxels that have been mapped as occupied are projected into the camera coordinate system, and the closest observation distance of the voxels that are within the camera field of view (FOV) and not blocked by any occupied space is calculated. The closest observation distance is greater than the maximum observation distance. The voxels are marked as unobserved areas , the closest observation distance is less than the maximum observation distance The voxels are marked as observable areas .

[0079] The process of generating observation viewpoints in step 1) is as follows: uniformly sampling candidate viewpoints in a cylindrical coordinate system near the occupied space, optimizing the yaw angle to maximize the RGBD camera coverage, and selecting the viewpoint with the highest coverage.

[0080] like Figure 4As shown, the process of viewpoint scoring in step 2) includes scoring all viewpoints in each cluster based on the four criteria of information gain, viewing angle, height, and distance:

[0081] Information gain: The information gain obtained by the drone at each viewpoint is calculated by the number of unobserved boundaries of the RGBD camera. Decision, that is

[0082]

[0083] in Indicates the score impact factor, through Control the influence of information gain on the score;

[0084] Perspective: To avoid the drone missing targets due to extreme perspectives, the average normal of each obstacle's occupied grid cells is calculated. , according to the obstacle center To Viewpoint Vector The average normal of the grid cells occupied by each obstacle The viewing angle is scored based on the closeness of:

[0085]

[0086] Distance: To ensure that the distance between the drone and the obstacle is kept within a reasonable range and that the object is not missed due to being too small, a Gaussian function is used to measure the deviation between the viewpoint distance and the optimal distance to dynamically adjust the score. Control the sensitivity of distance deviation to adapt to different task requirements, using the adjustment factor of distance effect To control the influence of distance factors on the score;

[0087]

[0088] in is the optimal observation distance, is the distance from the viewpoint to the obstacle;

[0089] Height: Avoid the drone being too high or too low, and miss the low or high targets. Like the distance score, the Gaussian function is used to measure the deviation between the viewpoint height and the optimal height to dynamically adjust the score. Control the sensitivity of altitude deviation to suit different mission requirements, using an altitude impact adjustment factor To control the influence of height on the score;

[0090]

[0091] in is the best observation height, is the height of the viewpoint to the ground;

[0092] the final score of each viewpoint is calculated as follows:

[0093]

[0094] As Figure 5 shown, the process of the collision-free viewpoint clustering method in step 3) is as follows: taking the current position of the UAV as the starting point of the first cluster, iteratively completing the clustering step by step, in each clustering process, taking the center of the current cluster as the starting point, through obstacle collision detection, the viewpoints without obstacle occlusion and within the specified radius R are added to the current cluster, and the cluster center is dynamically updated to the average position of all viewpoints in the current cluster, after the completion of the current cluster, the nearest point from the remaining viewpoints is selected as the starting point of the new cluster, and the above clustering process is repeated until all viewpoints are assigned to the cluster, as shown in Figure 6 .

[0095] The obstacle detection in step 3) is as follows: connecting the generated different viewpoints, if the connection between two viewpoints is not a straight line, it means that there is an obstacle occlusion.

[0096] As Figure 7 shown, the steps of the UAV dynamic re-planning method in step 4) are as follows:

[0097] 4.1) In each global search path planning, an improved Traveling Salesman Problem (ATSP) algorithm is used to generate a historical shortest path composed of all current viewpoint cluster centers , the UAV first searches according to the order of the historical shortest path, wherein, represents the nth viewpoint in the path, the UAV visits each viewpoint in the path in turn, and scans the observable area of the viewpoint;

[0098] 4.2) As the path is executed, the environment map is continuously updated, and once the UAV detects new areas during flight, these new viewpoints will be added to the task list to be detected , but the original path order will be preserved, and the search of the viewpoints in the original path will be completed first;

[0099] 4.3) After the execution of the historical path, the UAV will dynamically re-search the newly discovered areas;

[0100] 4.4) After the UAV searches the newly discovered areas, it will continue to generate a new round of historical shortest path Repeat steps 4.1) to 4.3) until all regions are searched. Figure 8 shown.

[0101] The steps of calculating the target three-dimensional coordinates in step 5) are as follows:

[0102] 5.1) Obtaining target pixel coordinates and depth information: Using an RGBD camera to capture the RGB image and depth information of the target area;

[0103] Use target detection algorithm to obtain target pixel coordinates and depth value ;

[0104] 5.2) Convert pixel coordinates to camera coordinate system: use camera intrinsic parameters, i.e. focal length , and main point , The two-dimensional pixel coordinates and depth value Convert to 3D coordinates in the camera coordinate system :

[0105]

[0106]

[0107]

[0108] Output the three-dimensional coordinates of the camera coordinate system:

[0109]

[0110] 5.3) Camera coordinate system to drone coordinate system: using a fixed rotation matrix The camera coordinate system coordinates Convert to drone coordinate system coordinates , the rotation matrix Indicates the installation posture of the camera relative to the drone coordinate system:

[0111]

[0112] Rotation matrix:

[0113]

[0114] Output the three-dimensional coordinates of the drone coordinate system:

[0115] ;

[0116] 5.4) UAV coordinate system to world coordinate system: Use the UAV 3×3 attitude rotation matrix and translation vectors The coordinates of the drone coordinate system Convert to world coordinates , where the attitude rotation matrix Represents the rotation relationship of the drone relative to the world coordinate system, translation vector The unit is meter;

[0117]

[0118] in: is the translation vector of the drone relative to the world coordinate system, The unit is meter, multiply by 1000 to convert to millimeter;

[0119] Output the world coordinate system 3D coordinates in millimeters:

[0120]

[0121] 5.5) Display the targets and world coordinates of the targets identified by the Yolo-Fastestv2 target recognition algorithm in the observable area in real time in the visualization tool Rviz to ensure the visualization and accuracy of the search results.

Claims

1. A method for autonomous target search for UAVs based on viewpoint scoring and dynamic replanning, characterized in that: The method comprises the following steps: 1) Using the drone's RGBD camera, an occupancy grid map is constructed in real time. The space perceived by the RGBD camera is divided into observable and unobservable areas based on distance information. Observation viewpoints are generated within the observable area to avoid missed detection or misidentification of targets due to observation distances exceeding the accuracy of the RGBD camera. 2) Score the generated observation viewpoints by considering information gain, viewing angle, distance, and height, and select the best observation viewpoint; 3) Clustering the observation viewpoints selected in step 2) using a collision-free viewpoint clustering method, and clustering the best observation viewpoints selected in step 2) into multiple collision-free subsets through obstacle detection; 4) A global path is planned from the current position through the center of all viewpoint clusters through the UAV dynamic replanning method. This enhances the consistency of the UAV path planning, avoids the UAV frequently changing flight direction and misses targets hidden in complex terrain, and improves the target search success rate. 5) Use the Yolo-Fastestv2 target recognition algorithm to identify targets in the observable area, obtain the pixel coordinates of the target in the RGB image, combine the drone pose and camera coordinates to derive the target's three-dimensional coordinates, and display the searched target in Rviz, while marking the global coordinates of the searched target.

2. The method for autonomous target search for unmanned aerial vehicles based on viewpoint scoring and dynamic replanning according to claim 1, characterized in that: The occupancy grid map in step 1) specifically refers to collecting RGB images and depth information of the target area through an RGBD camera, mapping the depth information to the space observed by the RGBD camera and discretizing it into grid units, namely voxels, where each grid unit is marked as occupied or free, thereby constructing a grid map representing environmental obstacles and free space.

3. The method for autonomous target search for unmanned aerial vehicles based on viewpoint scoring and dynamic replanning according to claim 2, characterized in that: The process of space division in step 1) is as follows: the RGBD camera obtains the depth information of the surrounding environment in real time, and divides the space observed by the RGBD camera into free space and occupy space Then, the voxels that have been mapped as occupied are projected into the camera coordinate system, and the closest observation distance of the voxels that are within the camera field of view and not blocked by any occupied space is calculated. The closest observation distance is greater than the maximum observation distance. The voxels are marked as unobserved areas , the closest observation distance is less than the maximum observation distance The voxels are marked as observable areas .

4. The method for autonomous target search for unmanned aerial vehicles based on viewpoint scoring and dynamic replanning according to claim 3, characterized in that: The process of generating observation viewpoints in step 1) is as follows: uniformly sampling candidate viewpoints in a cylindrical coordinate system near the occupied space, optimizing the yaw angle to maximize the RGBD camera coverage, and selecting the viewpoint with the highest coverage.

5. The method for autonomous target search for unmanned aerial vehicles based on viewpoint scoring and dynamic replanning according to claim 1, characterized in that: The viewpoint scoring process in step 2) includes scoring all viewpoints in each cluster using four criteria: information gain, viewing angle, height, and distance: Information gain: The information gain obtained by the drone at each viewing angle is calculated by the number of unobserved boundaries of the RGBD camera. Decision, namely: in Indicates the score impact factor, through Control the influence of information gain on the score; Perspective: To avoid the drone missing targets due to extreme perspectives, the average normal of each obstacle's occupied grid cells is calculated. , according to the obstacle center To Viewpoint Vector The average normal of the grid cells occupied by each obstacle The viewing angle is scored based on the closeness of: Distance: Use the Gaussian function to measure the deviation between the viewpoint distance and the optimal distance to dynamically adjust the score. Control the sensitivity of distance deviation to adapt to different task requirements, using the adjustment factor of distance effect To control the influence of distance factors on the score; in is the optimal observation distance, is the distance from the viewpoint to the obstacle; Height: Use Gaussian function to measure the deviation between the viewpoint height and the optimal height to dynamically adjust the score. Control the sensitivity of altitude deviation to suit different mission requirements, using an altitude impact adjustment factor To control the influence of height on the score; in is the best observation altitude, is the height from the viewpoint to the ground; Final score for each viewpoint It is calculated as follows:

6. The method for autonomous target search for unmanned aerial vehicles based on viewpoint scoring and dynamic replanning according to claim 1, characterized in that: The process of the collision-free viewpoint clustering method in step 3) is as follows: the current position of the drone is used as the starting point of the first cluster, and clustering is completed step by step. In each clustering process, the center of the current cluster is used as the starting point, and the viewpoints that are not blocked by obstacles and are within the specified radius are selected through obstacle collision detection. The viewpoints are added to the current cluster, and the cluster center is dynamically updated to the average position of all viewpoints in the current cluster. After the current clustering is completed, the nearest point is selected from the remaining viewpoints as the starting point of the new cluster, and the above clustering process is repeated until all viewpoints are assigned to the cluster.

7. The method for autonomous target search for unmanned aerial vehicles based on viewpoint scoring and dynamic replanning according to claim 1, characterized in that: The obstacle detection in step 3) is as follows: connecting the generated different viewpoints. If the line between the two viewpoints is not a straight line, it means that there is an obstacle blocking the view.

8. The method for autonomous target search for unmanned aerial vehicles based on viewpoint scoring and dynamic replanning according to claim 1, characterized in that: The steps of step 4) the UAV dynamic replanning method are as follows: 4.1) In each global search path planning, use the improved traveling salesman problem algorithm to generate a historical shortest path consisting of all current viewpoint cluster centers ,The UAV first searches in the order of the historical shortest path, where, represents the nth viewpoint in the path. The UAV visits each viewpoint in the path in turn and scans the observable area of ​​the viewpoint; 4.2) As the path is executed, the environment map will be continuously updated. Once the drone detects a new area during flight , these new perspectives will be Add to the list of tasks to be tested , but the order of the original path will be retained, and the viewpoint search in the original path will be completed first; 4.3) After the historical path is completed, the drone will dynamically re-search the newly discovered area; 4.4) After the drone has searched the newly discovered area, it will continue to use the return viewpoint as the starting point to generate a new round of historical shortest paths. Repeat the process from step 4.1) to step 4.3) until all areas have been searched.

9. The method for autonomous target search for unmanned aerial vehicles based on viewpoint scoring and dynamic replanning according to claim 1, characterized in that: The steps of calculating the target three-dimensional coordinates in step 5) are as follows: 5.1) Obtaining target pixel coordinates and depth information: Using an RGBD camera to capture the RGB image and depth information of the target area; Use target detection algorithm to obtain target pixel coordinates and depth value ; 5.2) Convert pixel coordinates to camera coordinate system: use camera intrinsic parameters, i.e. focal length , and main point , The two-dimensional pixel coordinates and depth value Convert to 3D coordinates in the camera coordinate system : Output the three-dimensional coordinates of the camera coordinate system: 5.3) Camera coordinate system to drone coordinate system: using a fixed rotation matrix The camera coordinate system coordinates Convert to drone coordinate system coordinates , the rotation matrix Indicates the installation posture of the camera relative to the drone coordinate system: Rotation matrix: Output the three-dimensional coordinates of the drone coordinate system: ； 5.4) UAV coordinate system to world coordinate system: Use the UAV 3×3 attitude rotation matrix and translation vectors The coordinates of the drone coordinate system Convert to world coordinates , where the attitude rotation matrix Represents the rotation relationship of the drone relative to the world coordinate system, translation vector The unit is meter; in: is the translation vector of the drone relative to the world coordinate system, The unit is meter, multiply by 1000 to convert to millimeter; Output the world coordinate system 3D coordinates in millimeters: 5.5) Display the targets and their three-dimensional world coordinates identified by the Yolo-Fastestv2 target recognition algorithm in the observable area in real time in the visualization tool Rviz.

Images