Unmanned aerial vehicle autonomous target searching method, system and device based on hierarchical decision

Abstract

The invention discloses an unmanned aerial vehicle autonomous target search method, system and device based on hierarchical decision, which combine global exploration and local search to improve the search efficiency of an unmanned aerial vehicle in a complex unknown environment. The method comprises the following steps: in a global exploration layer, constructing an incremental environmental map, and modeling an access sequence of to-be-searched and explored areas as an asymmetric traveling salesman problem; a Bi-RRT algorithm is utilized to generate a connection path in a free space, the visibility cost of the path is evaluated, and an access sequence with the minimum total cost is selected; in the local search layer, a hierarchical Gaussian process is used for fitting and evaluating the search value of the target area, and a priority target area is determined; generating a plurality of candidate viewpoints for the local target area, and selecting an optimal viewpoint as a local target search point; and navigating to the target search point and identifying the specific target article in combination with the positioning information and the target detection model. The method has efficient search and environment adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles (UAVs), and in particular to a method, system and device for autonomous target search of UAVs based on hierarchical decision-making. Background Art

[0002] With the development of intelligent robotics, drones, as a key offshoot, have demonstrated robust autonomous capabilities in areas such as environmental monitoring and emergency rescue. Compared to traditional manual methods, drones offer advantages such as high maneuverability, rapid response, and three-dimensional perception, making them particularly well-suited for target search tasks in high-risk, complex scenarios. Improving drones' autonomous search capabilities in unknown environments has become a key research area. The key lies in accurately and realistically perceiving and constructing the environment in accordance with the mission requirements of target search, while also achieving efficient and stable autonomous flight and path planning within that environment.

[0003] However, in practical applications, autonomous drone search still faces problems such as incomplete environmental perception, redundant path planning, and delayed target recognition. Existing papers and patents usually combine methods such as semantic mapping, full-area coverage planning, and boundary-based greedy strategies for analysis. However, in the absence of prior information, these strategies focus more on improving the high-precision maps constructed in the autonomous exploration field, and lack the ability to recognize and search for target objects in real time. Summary of the Invention

[0004] The present invention provides a method, system and device for autonomous target search of unmanned aerial vehicles based on hierarchical decision-making, aiming to solve the problems of difficult balance between search and exploration in complex unknown environments, insufficient utilization of environmental features, and a large amount of time wasted on blind exploration and map construction.

[0005] The purpose of the present invention is to achieve the following technical solutions: a method, system and device for autonomous target search of unmanned aerial vehicles based on hierarchical decision-making. The specific technical solutions are as follows:

[0006] An autonomous target search system for UAVs based on hierarchical decision-making includes: an environment perception module, a global exploration module, a local search module, a motion control module, and a target detection module.

[0007] The environmental perception module is equipped with a multimodal sensor to collect environmental depth information and visual data in real time to build an incremental three-dimensional map;

[0008] The global exploration module uses an improved Bi-RRT algorithm to generate a global path and optimizes the region visit sequence through an asymmetric traveling salesman problem;

[0009] The local search module includes a hierarchical Gaussian process modeling unit and a viewpoint optimization unit to achieve value assessment of the target area and calculation of the optimal observation pose;

[0010] The motion control module integrates a position controller and an attitude controller to ensure that the UAV accurately performs three-dimensional posture adjustment;

[0011] The target recognition module deploys a lightweight deep learning model to achieve real-time target detection and multimodal verification;

[0012] A method for autonomous target search of UAV based on hierarchical decision-making includes the following steps:

[0013] Step 1: In the global exploration layer, an incremental environment map is constructed, and the search and exploration areas are modeled as an asymmetric traveling salesman problem to optimize the visit order of the target areas through global path planning;

[0014] Step 2: Use the Bi-RRT algorithm to generate feasible paths connecting each target area in the current free space. By evaluating the visibility cost of the path, the access sequence with the minimum total cost is selected as the planned output path;

[0015] Step 3: In the local search layer, the search value of the target area is evaluated using hierarchical Gaussian process fitting, and the priority search area is selected based on the depth uncertainty of the environmental information;

[0016] Step 4: Generate multiple candidate viewpoints for the local search area, comprehensively consider the occlusion cost and rotation cost of perception angles such as yaw, pitch, and roll, and select the optimal viewpoint as the local target search point;

[0017] Step 5: Combine the positioning information with the target detection model to navigate the drone to the local target search point, and use the real-time target detection model to identify and confirm the specific target object and its location.

[0018] Furthermore, the implementation method of step 1 includes:

[0019] Step 101: By fusing information from multiple sensors such as binocular vision and an inertial measurement unit, the current positioning information and visual field perception image of the drone are obtained, and the depth information is converted into obstacle positions in a three-dimensional world coordinate system.

[0020] Step 102: construct and update an incremental three-dimensional occupancy grid map of the environment by ray casting, dividing the free space and obstacle areas of the known area;

[0021] Step 103: extracting a to-be-searched area and uncovered candidate exploration areas from the map, wherein the to-be-searched area is within the known area, and the candidate exploration area is the boundary area between the known and unknown areas. The geometric centers of the boundary points of the candidate exploration area and the local search points of the to-be-searched area are used as nodes in the traveling salesman problem.

[0022] Step 104: Construct an asymmetric cost matrix based on the node information. The cost matrix does not satisfy symmetry and the first column is zero, that is, the case of returning to the starting point is not considered. The items of the cost matrix include path length cost, path visibility cost, and angle cost;

[0023] Furthermore, the step 2 uses the Bi-RRT algorithm to connect the feasible paths of each target area, and the specific implementation method is as follows:

[0024] (1) Starting from p init and the endpoint p goal Initialize two search trees T for the root nodes a and T e , the search tree includes a node set N and an edge set E; generate a sampling point p in the free space rand ;

[0025] (2) In the current expansion tree, determine the distance p rand The nearest node p near ;

[0026] p near =argmin‖pp rand ‖2, p∈N

[0027] (3) Generate candidate nodes p with a fixed step size λ new ;

[0028] p new =p near +λ(p near -p near ) / ‖p near -p near ‖2

[0029] (4) For the path segment (p near ,p new ) to perform collision detection. If there is no collision, p new and edge (p near ,p new ) join the current tree;

[0030] (5) Calculate the Euclidean distance between the latest nodes of the two trees. If it is less than the connection threshold, the two trees are considered connected and a complete path Γ is generated.

[0031] (6) Alternately select a tree as the current expansion tree;

[0032] (7) Repeat steps (2) to (7) until the path is connected or the maximum number of iterations is reached, and calculate the path length as the cost.

[0033] Furthermore, the visibility cost of the path in step 2 is calculated based on the obtained path Γ, and the voxel block index V(i, j, k) of each path point (x, y, z) is calculated, where r is the voxel block resolution, and discrete sampling along the path is used to obtain the set of all voxel blocks on the path segment {V1, V2…V m}, for each voxel V i , extract the unknown voxel set U it contains, where the unknown voxel refers to the voxel whose occupancy probability value is within the default unknown threshold range, and count the total number of non-repeated unknown voxels E in all voxel blocks passed by the path ij =|∪U(V i )|, as the global visibility evaluation cost of the path segment.

[0034] Furthermore, the cost matrix calculation in step 104 is based on the cost terms calculated in step 2 to construct a cost matrix C, which comprehensively considers the following factors:

[0035] (1) The cost of position change is proportional to the path length L;

[0036] (2) Speed ​​change cost, which is related to the angle Δθ between the current position speed direction and the target position direction;

[0037] (3) yaw angle change cost, which depends on the difference Δψ between the starting yaw angle and the target yaw angle;

[0038] (4) Global visibility evaluation cost, based on the number of unknown voxels E on the global path;

[0039] (5) Cost matrix C ij The formula for calculating the cost item in the i-th row and j-th column of is:

[0040]

[0041] Among them, v max is the maximum speed, ω dir is the weight coefficient of velocity direction change, is the maximum yaw rate;

[0042] (6) Based on the cost matrix C, the traveling salesman problem algorithm is used to determine the optimal exploration sequence, which maximizes the information acquisition of the unknown environment while satisfying the motion constraints.

[0043] Furthermore, the implementation method of the hierarchical Gaussian process in step three includes:

[0044] (1) Design Gaussian kernel function

[0045] In the Gaussian process, the kernel function k(x, x′) is the core component of the model. It defines the similarity measure between any two points in the input space, evaluates whether the region has structural features, and uses a dynamic weight mechanism to balance local and global deep features. The kernel function expression is:

[0046] k(x,x′)=ω gp k RBF (x,x′)+(1-ω gp )ρ gp

[0047] ρ gp =n occ / s 3

[0048] k RBF (x,x′)=exp(-‖xx′‖2 / 2l 2 )

[0049] Among them, k RBF is the RBF kernel function, l is the length scale parameter, which determines the smoothness of the function change, ρ gp Reflects the density of occupied points within the voxel block, n occ is the number of voxels in the voxel block, s is the side length of the voxel block, ω gp Dynamic adjustment through Sigmoid function:

[0050] ω gp =1 / (1+exp(k(ρ gp -τ gp ))

[0051] τ gp Represents the critical judgment threshold of the voxel block occupancy density, k is the difference weight coefficient, when the occupied points are sparse, ω gp Approaching 1, the kernel function is dominated by the standard RBF kernel, highlighting the local depth smoothness; when the occupied points are dense, ω gp Approaching 0, the kernel function focuses on the global depth density distribution and adapts to complex spatial structures.

[0052] (2) Hierarchical Gaussian Process Regression

[0053] Gaussian process regression based on kernel function k(x,x′) realizes multi-level modeling of obstacle distribution within voxel blocks. First, the kernel matrix K of the training dataset is constructed:

[0054] K=k(X,X)+σ n I 2

[0055] Where k(X,X) is the covariance matrix based on the dynamic weight kernel function, X is the input matrix, i.e., the voxel coordinate point, σn I 2 is the noise term. To improve computational efficiency, Cholesky decomposition is used to decompose the kernel matrix K into the product of the lower triangular matrix L and its transpose:

[0056] K=LL T

[0057] Quickly solve for the weight vector using forward-backward substitution:

[0058] α gp =L -T (L -1 y))

[0059] Weight vector α gp It is the core of the posterior distribution calculation, combining the information of the observed data y and the kernel matrix. For the point x to be predicted, calculate its covariance vector k(x,X) with the training point, the posterior mean μ(x) and the variance σ 2 (x) are:

[0060] μ(x)=k(x,X) T α gp

[0061] σ 2 (x)=k(x,x)-v T v

[0062] Where v = L -1 k(x,X) is an intermediate variable. The predicted mean μ(x) provides the voxel depth estimate, and the variance σ 2 (x) Quantify uncertainty.

[0063] (3) Layered fusion and obstacle determination

[0064] The voxel block is layered vertically, and a Gaussian process is fitted to the center point of each layer to obtain s sets of prediction results. Free space points are assigned a value of 0, and obstacle points are assigned a preset depth value. The mean and variance of each layer are combined to output the final prediction of the voxel block. If the variance is below the threshold, it is determined to be a uniform obstacle such as a wall; otherwise, it is marked as a potential target area.

[0065] (4) Dynamic clustering and regional optimization

[0066] By calculating the cluster center of the obstacle voxel block, the candidate point set of the local search area is determined; based on the clustering results and hierarchical variance evaluation, the local target area boundary that meets the threshold condition is determined. For the potential target voxel block, the adaptive density clustering algorithm is used to generate the candidate area. First, the voxel block V is calculated. i The geometric center c i :

[0067]

[0068] Initialize cluster radius r cluster When introducing dynamic adjustment mechanism:

[0069]

[0070] where α cluster is the density coefficient, the default value is 1.2, U is the volume of the voxel block, N cluster is the number of voxel blocks in the current cluster. j , if its center c j With the current cluster center c i The Euclidean distance is less than or equal to r cluster , then classify them into the same cluster C k , and update the cluster center by using the mean shift method, iterating until all voxel blocks are classified.

[0071]

[0072] The default setting is the number of voxel blocks in the cluster threshold N thr =10, if C k The number of voxel blocks in the cluster is greater than N thr , it is determined to be an oversized cluster, and the density adjustment coefficient is reduced. The cluster radius is recalculated based on the new parameters, and adaptive clustering is recursively performed within the oversized cluster to generate a subcluster set {C k1 ,C k2} is added to the cluster to form a new voxel block cluster, namely the target area V t And so on.

[0073] Furthermore, the implementation method of step 4 includes:

[0074] (1) Using Gaussian process regression to fit the voxel depth in step 3, the target region V is determined by adaptive density clustering. t Constructing a local search area;

[0075] (2) Establish a spherical coordinate system at the geometric center of the target area and construct a candidate viewpoint set VP containing position and three-dimensional posture angle i , i depends on the number of candidate viewpoint sets, where each viewpoint vp i,j =(p i,j ,ψ i,j ,θ i,j ,φ i,j ), containing the three-dimensional coordinates p i,j , yaw angle ψ i,j , pitch angle θ i,j and roll angle φ i,j, realize three-dimensional posture design;

[0076] (3) Calculate the 3D observation quality of each candidate viewpoint, including the pitch angle optimization evaluation I pitch Quantifying vertical observation effect and roll angle optimization evaluation I roll Improve the field of view edge coverage capability and multi-angle collaborative gain I synergy Evaluate the combined effect of the three perspectives and achieve comprehensive evaluation through weighted summation;

[0077] (4) Dynamically adjust the angle weight according to the environmental characteristics. When there are rich vertical structures, the optimization of pitch angle is emphasized. When there are narrow passages, the optimization of roll angle is emphasized. When there is open environment, the optimization is balanced.

[0078] (5) Considering the dynamic characteristics of the UAV, safety constraints are imposed on the pitch and roll angles, including maximum angle limits and angle change rate limits, to ensure flight stability;

[0079] (6) Fusion of three-dimensional angle evaluation indicators, and selection of the best viewpoint vp through weighted optimization optimal , while satisfying all angle constraints;

[0080] (7) The UAV is positioned according to the optimal spatial position and three-dimensional attitude angle (p optimal ,ψ optimal ,θ optimal ,φ optimal ) for precise positioning, and achieve all-round and efficient observation of local target areas.

[0081] Furthermore, the implementation method of step five is that based on the target posture information provided by the optimal spatial position and three-dimensional attitude angle determined in step four, the UAV autonomously navigates to the target area through the fusion positioning system and trajectory generation and optimization algorithm, and a monocular camera is installed on the front axis of the UAV body. During the process, a lightweight target detection model is run in real time to identify the preset target objects, and a multi-frame detection fusion and dynamic confidence evaluation mechanism are used to ensure the recognition accuracy. When the target is detected, its precise three-dimensional coordinates are recorded and the recognition result is verified.

[0082] The present invention also discloses a UAV autonomous target search device based on hierarchical decision-making, including a UAV, an onboard computer, a ground station, a positioning system and a perception camera. The onboard computer stores executable code, and when the one or more processors execute the executable code, they are used to implement the steps of the UAV autonomous target search method. The perception camera is used for identification and mapping, and the positioning system provides posture information of the UAV throughout the mission.

[0083] Beneficial effects of the present invention:

[0084] (1) The method designed in the present invention can improve the search efficiency while reducing the overall task execution time, reduce the ineffective round-trip movement of the UAV in the local area, and improve the effectiveness and continuity of path execution.

[0085] (2) The method designed in the present invention has good environmental adaptability and is suitable for complex environments with different unknown structures. It has strong generalization ability and can meet the search task requirements of multiple scenarios and multiple platforms.

[0086] (3) The present invention has strong safety, and the drone is always restricted to fly within the passable area and maintain a safe distance.

[0087] (4) The present invention has significant advantages in terms of computational efficiency and hardware configuration. It adopts a lightweight target detection model and an efficient hierarchical path planning algorithm, with low computational overhead and fast response speed. It supports operation under conditions with limited onboard computing resources and does not require a lidar sensor, thereby reducing the weight of the drone. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 This is a schematic diagram of the framework of the present invention;

[0089] Figure 2 A diagram of a system architecture applicable to an embodiment of the present invention;

[0090] Figure 3 This is a flow chart of an embodiment of the present invention;

[0091] Figure 4 It is a schematic diagram of the task space;

[0092] Figure 5 Schematic diagram for estimating visible unknown voxels in the path;

[0093] Figure 6 Schematic diagram of depth modeling for hierarchical Gaussian process fitting;

[0094] Figure 7 This is a frame-by-frame breakdown of the target search effect in the actual machine experiment. DETAILED DESCRIPTION

[0095] The specific embodiments of the present invention are further described in detail below with reference to the accompanying drawings.

[0096] like Figure 1 As shown, the autonomous target search method of a UAV based on hierarchical decision-making of the present invention includes the following steps:

[0097] Step 1: In the global exploration layer, an incremental environment map is constructed, and the search and exploration areas are modeled as an asymmetric traveling salesman problem to optimize the visit order of the target areas through global path planning;

[0098] Step 2: Use the Bi-RRT algorithm to generate feasible paths connecting each target area in the current free space. By evaluating the visibility cost of the path, the access sequence with the minimum total cost is selected as the planned output path;

[0099] Step 3: In the local search layer, the search value of the target area is evaluated using hierarchical Gaussian process fitting, and the priority search area is selected based on the depth uncertainty of the environmental information;

[0100] Step 4: Generate multiple candidate viewpoints for the local search area, comprehensively consider the occlusion cost and rotation cost of perception angles such as yaw, pitch, and roll, and select the optimal viewpoint as the local target search point;

[0101] Step 5: Combine the positioning information with the target detection model to navigate the drone to the local target search point, and use the real-time target detection model to identify and confirm the specific target object and its location.

[0102] Take the quadrotor drone as the experimental target to perform autonomous search tasks, such as Figure 2 and Figure 3 As shown, the task implementation and design steps are:

[0103] Step 1: In the global exploration layer, first construct an incremental environmental map and update the distribution of unknown and known areas in real time. The environmental map uses a voxel representation method to discretize the three-dimensional space into a series of voxel units of equal size, and further aggregate the voxel units into voxel block units. Each voxel is marked as known idle, known occupied, or unknown based on the sensor observation information. Based on the environmental map, the area to be searched and explored is modeled as an asymmetric traveling salesman problem, in which each candidate target area is regarded as a visit node. By solving the asymmetric traveling salesman problem, the global optimal visit order is planned to reduce the energy consumption and search time of the drone. For example Figure 4 As shown in Figure 2, the voxelized environment map and task space partitioning demonstrate the main functions of the global exploration layer, which uses the boundaries and viewpoints in the map as areas to be visited for subsequent calculation of the optimal visit order.

[0104] Step 2: For the target area access sequence determined by the global layer, this step uses the Bi-RRT algorithm to generate a feasible path connecting each target area in the current free space. The Bi-RRT algorithm grows random trees from the starting point and the end point at the same time, which accelerates the path search process. During the path generation process, the algorithm comprehensively evaluates the following costs: visibility cost: calculates the ratio of the total number of unknown voxels in the field of view of each point on the path to all voxel blocks; angle cost: evaluates the turning angle range from the current position to the candidate point; path length cost: evaluates the total length of the path and its impact on the energy consumption of the drone. By weighted fusion of these cost indicators, the access sequence with the smallest total cost is selected as the planned output path. As Figure 5As shown, a method for evaluating the visibility cost is presented, which calculates the situation of all voxel blocks passed on the path, rather than the number of unknown voxels within the field of view of a single position.

[0105] Step 3: In the local search layer, a depth sensor is first used to acquire a depth image of the scene. The depth features of typical indoor obstacles, such as shelves and corridors, are extracted and their depth histogram distribution characteristics are analyzed. The environmental space is discretized into voxel blocks, each of which records obstacle occupancy information. Voxel blocks containing obstacles are also labeled. Considering the vertical layering of indoor environments, a hierarchical modeling approach is used to divide the voxel blocks vertically into several layers, and Gaussian process modeling is performed independently on each layer.

[0106] In kernel function design, a dynamic weighted RBF kernel function is used, whose expression comprehensively considers local depth smoothness and global density distribution characteristics. A sigmoid function is used to dynamically adjust the length scale parameter, emphasizing local smoothness when the voxel block occupancy points are sparse and focusing on global density distribution when the voxel block occupies dense points. A Gaussian process regression model is constructed based on this kernel function, using Cholesky decomposition to improve the computational efficiency of the kernel matrix, and a forward-backward substitution method to quickly solve the weight vector.

[0107] In the hierarchical modeling process, Figure 6 As shown, a Gaussian process is independently fitted to the center point of each layer, resulting in multiple sets of prediction results. By fusing the mean and variance of each layer, a final prediction is made for the voxel block: those with a variance below a threshold are identified as uniform obstacles (such as walls); otherwise, they are marked as potential target areas. During the region optimization phase, an adaptive density clustering algorithm is used to generate candidate regions. The cluster radius is dynamically adjusted based on the voxel block density, and the cluster center is iteratively updated using the mean shift method. For oversized clusters, recursive sub-clustering is performed by reducing the density coefficient, ultimately forming a voxel block cluster of the target area.

[0108] Step 4: First, based on the voxel depth fitting results of Gaussian process regression, adaptive density clustering is used to determine the target area. A spherical coordinate system is established at the geometric center of the area, generating a set of candidate viewpoints containing three-dimensional coordinates and attitude angles, namely yaw, pitch, and roll. For each viewpoint, the effect of pitch on vertical observation, the improvement of roll on edge coverage, and the synergistic gain of the three angles are evaluated separately, and a comprehensive quality score is obtained through weighted summation. The weight distribution is dynamically adjusted according to environmental characteristics: vertical structures focus on pitch, narrow passages focus on roll, and open areas are optimized evenly. At the same time, the dynamic constraints of the drone are considered, limiting the maximum pitch angle, roll angle, and angular velocity. Finally, the viewpoint with the best overall quality is selected through constrained weighted optimization, guiding the drone to perform observations in the optimal posture to ensure that the target area coverage rate reaches more than 80%.

[0109] Step 5: Based on the target pose information provided by the optimal spatial position and three-dimensional attitude angle determined in step 4, the UAV autonomously navigates to the target area by integrating the positioning system and trajectory generation and optimization algorithm. A monocular camera is installed on the front axis of the UAV body. During the process, the lightweight target detection YOLOv5s model runs in real time to identify preset target objects, such as common indoor beverage bottles and other characteristic objects. Multi-frame detection fusion and dynamic confidence assessment mechanism are used to ensure recognition accuracy. When the target is detected, its precise three-dimensional coordinates are recorded and the recognition result is verified.

[0110] The autonomous target search system of a UAV based on hierarchical decision-making of the present invention comprises: an environment perception module, a global exploration module, a local search module, a motion control module, and a target detection module.

[0111] The environmental perception module is equipped with a multimodal sensor to collect environmental depth information and visual data in real time;

[0112] The global exploration module constructs an incremental three-dimensional map, generates a global path using an improved Bi-RRT algorithm, and optimizes the target area visit sequence through an asymmetric traveling salesman problem;

[0113] The local search module includes hierarchical Gaussian process modeling and viewpoint optimization to achieve value assessment of the target area and calculation of the optimal observation pose;

[0114] The motion control module integrates a position controller and an attitude controller to ensure that the UAV accurately performs three-dimensional posture adjustment;

[0115] The target recognition module deploys a lightweight deep learning model to achieve real-time target detection and recognition verification.

[0116] Example 1:

[0117] like Figure 7 As shown in the figure, the monocular camera is placed on the front axis of the drone, and its direction is consistent with the flight control direction of the drone. The drone flies near the obstacle where the target object is stored. Through the local search algorithm, the target area is found. Through the analysis of 5 frames of pictures, the drone's head is obviously turned and turned to the side where the object is stored, that is, the side with complex depth changes. At the same time, the target detection result is presented. The target detection frame is clearly marked in the figure, and the specific position of the beverage bottle is found. The entire trajectory revolves around the obstacle. The field of view image presents clear image quality only when the target detection result is obtained. Rapid movement and detection are performed at other times, meeting the needs of rapid target search of the drone.

[0118] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. An autonomous target search system for unmanned aerial vehicles based on hierarchical decision-making, comprising: The environment perception module, global exploration module, local search module, motion control module, and target detection module are characterized by: The environmental perception module is equipped with a multimodal sensor to collect environmental depth information and visual data in real time; The global exploration module constructs an incremental three-dimensional map, generates a global path using an improved Bi-RRT algorithm, and optimizes the target area visit sequence through an asymmetric traveling salesman problem; The local search module includes hierarchical Gaussian process modeling and viewpoint optimization to achieve value assessment of the target area and calculation of the optimal observation pose; The motion control module integrates a position controller and an attitude controller to ensure that the UAV accurately performs three-dimensional posture adjustment; The target recognition module deploys a lightweight deep learning model to achieve real-time target detection and recognition verification.

2. A UAV autonomous target search method based on hierarchical decision making, characterized by The steps include: Step 1: In the global exploration layer, an incremental environment map is constructed, and the areas to be searched and explored are modeled as an asymmetric traveling salesman problem to optimize the order of visiting target areas through global path planning. Step 2: Use the Bi-RRT algorithm to generate feasible paths connecting each target area in the current free space. By evaluating the visibility cost of the path, the access sequence with the minimum total cost is selected as the planned output path; Step 3: In the local search layer, the search value of the target area is evaluated using hierarchical Gaussian process fitting, and the priority search area is selected based on the depth uncertainty of the environmental information; Step 4: Generate multiple candidate viewpoints for the local search area, comprehensively consider the occlusion cost and rotation cost of perception angles such as yaw, pitch, and roll, and select the optimal viewpoint as the local target search point; Step 5: Combine the positioning information with the target detection model to navigate the drone to the local target search point, and use the real-time target detection model to identify and confirm the specific target object and its location.

3. The autonomous target search method for unmanned aerial vehicle based on hierarchical decision making according to claim 2, characterized in that: The step 1 specifically includes the following steps: Step 101: By fusing information from multiple sensors such as binocular vision and an inertial measurement unit, the current positioning information and visual field perception image of the drone are obtained, and the depth information is converted into obstacle positions in a three-dimensional world coordinate system. Step 102: construct and update an incremental three-dimensional occupancy grid map of the environment by ray casting, dividing the free space and obstacle areas of the known area; Step 103: extracting a to-be-searched area and uncovered candidate exploration areas from the map, wherein the to-be-searched area is within the known area, and the candidate exploration area is the boundary area between the known and unknown areas. The geometric centers of the boundary points of the candidate exploration area and the local search points of the to-be-searched area are used as nodes in the traveling salesman problem. Step 104: Construct an asymmetric cost matrix based on the node information. The cost matrix does not satisfy symmetry and the first column is zero, that is, the case of returning to the starting point is not considered. The items of the cost matrix include path length cost, path visibility cost and angle cost.

4. The autonomous target search method for unmanned aerial vehicles based on hierarchical decision-making according to claim 2, characterized in that: The second step uses the Bi-RRT algorithm to connect the feasible paths of each target area. The specific implementation method is as follows: (1) Starting from p init and the endpoint p goal Initialize two search trees T for the root nodes a and T e , the search tree includes a node set N and an edge set E; (2) Generate sampling point p in free space rand ; (3) In the current expansion tree, determine the distance p rand The nearest node p near ; p near =argmin‖p-p rand ‖2,p∈N (4) Generate candidate nodes p with a fixed step size λ new ; p new =p near +λ(p near -p near ) / ‖p near -p near ‖2 (5) For the path segment (p near ,p new ) to perform collision detection. If there is no collision, p new and edge (p near ,p new ) join the current tree; (6) Calculate the Euclidean distance between the latest nodes of the two trees. If it is less than the connection threshold, the two trees are considered connected and a complete path Γ is generated. (7) Alternately select a tree as the current expansion tree; Repeat steps (2) to (7) until the connection is achieved or the maximum number of iterations is reached, and the path length is calculated as the cost; the visibility cost of the path in step 2 is specifically calculated based on the generated complete path Γ, and for each path point (x, y, z), the voxel block index V(i, j, k) where it is located is calculated, where r is the voxel block resolution, and discrete sampling along the path is used to obtain the set of all voxel blocks on the path segment {V1, V2…V m }, for each voxel V i , extract the unknown voxel set U it contains, where the unknown voxel refers to the voxel whose occupancy probability value is within the default unknown threshold range, and count the total number of non-repeated unknown voxels E in all voxel blocks passed by the path ij =|∪U(V i )|, as the global visibility evaluation cost of the path segment.

5. The autonomous target search method for unmanned aerial vehicle based on hierarchical decision-making according to claim 4, characterized in that: The cost matrix calculation in step 104 is based on the cost terms to construct a cost matrix C, which comprehensively considers the following factors: (1) The cost of position change is proportional to the path length L; (2) Speed ​​change cost, which is related to the angle Δθ between the current position speed direction and the target position direction; (3) yaw angle change cost, which depends on the difference Δψ between the starting yaw angle and the target yaw angle; (4) Global visibility evaluation cost, based on the number of unknown voxels E on the global path; (5) Cost matrix C ij The formula for calculating the cost item in the i-th row and j-th column of is: Among them, v max is the maximum speed, ω dir is the weight coefficient of velocity direction change, is the maximum yaw rate; (6) Based on the cost matrix C, the traveling salesman problem algorithm is used to determine the optimal exploration sequence, which maximizes the information acquisition of the unknown environment while satisfying the motion constraints.

6. The autonomous target search method for unmanned aerial vehicles based on hierarchical decision-making according to claim 2, characterized in that: The depth feature modeling of the hierarchical Gaussian process fitting in step three is to obtain a depth image of the scene, extract common indoor obstacle types, including shelves and corridors, and analyze their depth features and depth histogram distribution, divide the environmental space into a voxel block structure, each voxel block records obstacle occupancy information, mark the voxel blocks where obstacles exist, and for each voxel block containing an obstacle, considering the structured characteristics of the obstacle, apply hierarchical modeling, divide the voxel block into n layers in the vertical direction, independently extract Gaussian process data for each layer, and construct a hierarchical Gaussian process model based on the inter-layer relationship, so that the data of each layer can represent the characteristic distribution of obstacles at that height level.

7. The autonomous target search method for unmanned aerial vehicle based on hierarchical decision making according to claim 6, characterized in that: The implementation method of the hierarchical Gaussian process includes: (1) Design Gaussian kernel function In the Gaussian process, the kernel function k(x,x ′ ) is the core component of the model, which defines the similarity measure between any two points in the input space, evaluates whether the region has structural features, and uses a dynamic weight mechanism to balance local and global depth features. The kernel function expression is: k(x,x ′ )=ω gp k RBF (x,x ′ )+(1-ω gp )r gp r gp =n occ / s 3 k RBF (x,x ′ )=exp(-‖x-x ′ ‖2 / 2l 2 ) Among them, k RBF is the RBF kernel function, l is the length scale parameter, which determines the smoothness of the function change, ρ gp Reflects the density of occupied points within the voxel block, n occ is the number of voxels in the voxel block, s is the side length of the voxel block, ω gp Dynamic adjustment through Sigmoid function: oh gp =1 / (1+exp(k(ρ gp -t gp )) τ gp Represents the critical judgment threshold of the voxel block occupancy density, k is the difference weight coefficient, when the occupied points are sparse, ω gp Approaching 1, the kernel function is dominated by the standard RBF kernel, highlighting the local depth smoothness; when the occupied points are dense, ω gp Approaching 0, the kernel function focuses on the global depth density distribution and adapts to complex spatial structures; (2) Hierarchical Gaussian Process Regression Based on the kernel function k(x,x ′ ) realizes the multi-level modeling of obstacle distribution within the voxel block; first, the kernel matrix K of the training dataset is constructed: K=k(X,X)+σ n I 2 Where k(X,X) is the covariance matrix based on the dynamic weight kernel function, X is the input matrix, i.e., the voxel coordinate point, σ n I 2 is the noise term. To improve computational efficiency, Cholesky decomposition is used to decompose the kernel matrix K into the product of the lower triangular matrix L and its transpose: K=LL T Solve for the weight vector using forward-backward substitution: α gp =L -T (L -1 y)) Weight vector α gp It is the core of the posterior distribution calculation, combining the information of the observed data y and the kernel matrix. For the point x to be predicted, calculate its covariance vector k(x,X) with the training point, the posterior mean μ(x) and the variance σ 2 (x) are: μ(x)=k(x,X) T a gp σ 2 (x)=k(x,x)-v T v Where v = L -1 k(x,X) is an intermediate variable. The predicted mean μ(x) provides the voxel depth estimate, and the variance σ 2 (x) quantify uncertainty; (3) Layered fusion and obstacle determination The voxel block is layered vertically, and the center point of each layer is independently fitted with a Gaussian process to obtain s groups of prediction results. Free space points are assigned a value of 0, and obstacle points are assigned a preset depth value. The final prediction of the voxel block is output after fusing the mean and variance of each layer. If the variance is lower than the threshold, it is determined to be a uniform obstacle such as a wall; otherwise, it is marked as a potential target area. (4) Dynamic clustering and regional optimization By calculating the cluster center of the obstacle voxel block, the candidate point set of the local search area is determined; based on the clustering results and hierarchical variance evaluation, the local target area boundary that meets the threshold condition is determined; for the potential target voxel block, the adaptive density clustering algorithm is used to generate the candidate area; first calculate each voxel block V i The geometric center c i : Initialize cluster radius r cluster When introducing a dynamic adjustment mechanism: where α cluster is the density coefficient, the default value is 1.2, U is the volume of the voxel block, N cluster is the number of voxel blocks in the current cluster. j , if its center c j With the current cluster center c i The Euclidean distance is less than or equal to r cluster , then classify them into the same cluster C k , and use the mean shift method to update the cluster center, iterate until all voxel blocks are classified; update the cluster center by mean shift iteration The default setting is the number of voxel blocks in the cluster threshold N thr =10, if C k The number of voxel blocks in the cluster is greater than N thr , it is determined to be an oversized cluster, and the density adjustment coefficient is reduced. The cluster radius is recalculated based on the new parameters, and adaptive clustering is recursively performed within the oversized cluster to generate a subcluster set {C k1 ,C k2 } is added to the cluster to form a new voxel block cluster, namely the target area V t , and so on.

8. The autonomous target search method for unmanned aerial vehicles based on hierarchical decision-making according to claim 2, characterized in that: The implementation method of step 4 includes: (1) Gaussian process regression is used to fit the voxel depth, and the target region V is determined by adaptive density clustering. t Constitute a local search area; (2) Establish a spherical coordinate system at the geometric center of the target area and construct a candidate viewpoint set VP containing position and three-dimensional posture angle i , i depends on the number of candidate viewpoint sets, where each viewpoint vp i,j =(p i,j ,ψ i,j ,θ i,j ,φ i,j ), containing the three-dimensional coordinates p i,j , yaw angle ψ i,j , pitch angle θ i,j and roll angle φ i,j , realize three-dimensional posture design; (3) Calculate the 3D observation quality of each candidate viewpoint, including the pitch angle optimization evaluation I pitch Quantifying vertical observation effect and roll angle optimization evaluation I roll Improve the field of view edge coverage capability and multi-angle collaborative gain I synergy Evaluate the combined effect of the three perspectives and achieve comprehensive evaluation through weighted summation; (4) Dynamically adjust the angle weight according to the environmental characteristics. When there are rich vertical structures, the optimization of pitch angle is emphasized. When there are narrow passages, the optimization of roll angle is emphasized. When there is open environment, the optimization is balanced. (5) Considering the dynamic characteristics of the UAV, safety constraints are imposed on the pitch and roll angles, including maximum angle limits and angle change rate limits, to ensure flight stability; (6) Fusion of three-dimensional angle evaluation indicators, and selection of the best viewpoint vp through weighted optimization optimal , while satisfying all angle constraints; (7) The UAV is positioned according to the optimal spatial position and three-dimensional attitude angle (p optimal ,ψ optimal ,θ optimal ,φ optimal ) for precise positioning, and achieve all-round and efficient observation of local target areas.

9. The method for autonomous target search of a UAV based on hierarchical decision making according to claim 8, characterized in that: The implementation method of step five includes: based on the target posture information provided by the optimal spatial position and three-dimensional attitude angle determined in step four, the UAV autonomously navigates to the target area through the fusion positioning system and trajectory generation and optimization algorithm, and carries a monocular camera on the front axis of the UAV body. During the process, a lightweight target detection model is run in real time to identify the preset target object, and multi-frame detection fusion and dynamic confidence assessment mechanism are used to ensure recognition accuracy. When the target is detected, its precise three-dimensional coordinates are recorded and the recognition result is verified.

10. An autonomous target search device for unmanned aerial vehicles based on hierarchical decision making, characterized in that: The system comprises a drone, an onboard computer, a ground station, a positioning system and a perception camera. The onboard computer stores executable code. When one or more processors execute the executable code, the system is used to implement the steps of the drone autonomous target search method described in any one of claims 2 to 9. The perception camera is used for identification and mapping, and the positioning system provides posture information of the drone throughout the mission.

Images