A three-dimensional continuous monitoring method for UAV swarm targets with adaptive range adjustment

Abstract

The present invention relates to a method for three-dimensional continuous monitoring of drone swarm targets with adaptive range adjustment, belonging to the technical field of drone target monitoring. This method addresses the problem of three-dimensional continuous monitoring of a single drone swarm based on visual detection. It utilizes a weighted graph structure to model the swarm targets. Horizontally, flight decisions are made by finding the center position. In altitude, adaptive flight altitude decisions are made by utilizing an offline dataset combined with an offline reinforcement learning algorithm, taking into account drone energy consumption and visual system accuracy. This method can adaptively adjust the drone's detection range, effectively improving the monitoring performance of a single drone.

Description

Technical Field

[0001] The present invention relates to a three-dimensional continuous monitoring method for drone swarm targets with adaptive range adjustment, and specifically to a method that utilizes graph structure modeling and an offline reinforcement learning algorithm to adaptively adjust the drone detection range and improve target monitoring performance, belonging to the technical field of drone target monitoring. Background Art

[0002] With the development of drone technology, drones have broad application prospects in a variety of fields, including logistics, security, and military. In recent years, the rapid development of artificial intelligence (AI) has provided new research approaches for drone target monitoring. Currently, drone target monitoring technology has become one of the hottest research areas at the intersection of drones and AI.

[0003] Continuous drone monitoring of targets is a key research area in drone intelligent control. Continuous tracking tasks require the monitored target to remain within the drone's field of view, enabling it to obtain information such as the target's location at all times, providing accurate perception results for subsequent tasks. Continuous drone monitoring missions are widely used. In wildlife conservation, drones are used to continuously monitor animal populations, study animal migration patterns, and protect endangered wildlife from poaching. In adversarial environments, drones can be used to monitor various maneuvering targets, facilitating appropriate decision-making. Therefore, continuous drone monitoring missions are widely used in many aspects of society and have significant social benefits.

[0004] While research on drone-based target monitoring methods has made some progress, several challenges remain. First, existing technologies fail to fully explore the maximum monitoring capabilities of a single drone, specifically how to adaptively adjust its detection range. The monitoring performance of a single drone is the foundation and prerequisite for multi-drone collaboration. However, existing methods for group target monitoring primarily utilize multiple drones to collaboratively monitor a single target, or assign each drone only one target. Second, existing technologies overlook the group characteristics of targets, such as fleets, convoys, and animal populations, whose movements often exhibit similar patterns and trends. Furthermore, existing technologies fail to fully incorporate sensor characteristics. For example, in adversarial environments, drones' onboard radar systems are typically silent to prevent their locations from being exposed and potentially attacked. Therefore, full utilization of visual sensors is necessary, requiring adaptive adjustment of various parameters to enable the drone to achieve maximum monitoring and decision-making capabilities within three-dimensional space. Summary of the Invention

[0005] The present invention addresses the shortcomings and deficiencies of existing technologies, specifically the inability to adaptively adjust the detection range when a single drone monitors multiple targets, resulting in poor monitoring performance. Furthermore, it addresses the issues of neglecting the characteristics of a group of targets and the relationship between the characteristics of the drone's visual detection system and altitude in adversarial environments. To effectively improve the monitoring performance of a single drone, a method for three-dimensional continuous monitoring of drone swarms with adaptive range adjustment is creatively proposed.

[0006] First, the relevant concepts are explained.

[0007] 1. Continuous Monitoring

[0008] It means that after the drone detects the target, it will track the target so that all targets are always kept within the drone's field of view, so as to grasp the target's movement trend and provide accurate target information for subsequent tasks.

[0009] 2. Adaptive range

[0010] This refers to the drone's adaptive adjustment of its monitoring range, thereby varying the number of targets covered. The drone's visual perception system typically stitches together image information from multiple cameras, so the monitoring range can be adjusted based on internal parameters such as the camera's focal length. When these internal parameters reach their maximum threshold, the monitoring range, and thus the monitoring capability, can be altered by adjusting the distance between the camera and the target.

[0011] 3. Group Target

[0012] When multiple targets have similar motion trends, such as when the target is far away from the drone, the multiple targets can be regarded as a group. At this time, the movement within the entire group can be represented by the target with the highest degree of centrality within the group.

[0013] The present invention is implemented by adopting the following technical solutions.

[0014] A method for three-dimensional continuous monitoring of a drone swarm target with adaptive range adjustment includes the following steps:

[0015] Step 1: Set the drone's onboard perception system to radar silence and use only visual sensors for target detection. The target detection area of ​​the drone is obtained by stitching together multiple visual sensors to obtain group target information.

[0016] Among them, the radar is set to silent mode to prevent the opponent from interfering with the target detection results through electronic interference and other means, or attacking and destroying drones in a confrontational environment.

[0017] The multi-visual sensor stitching is achieved by stitching images from multiple cameras to obtain a larger field of view, meet the detection requirements of group targets, and improve the detection capability of a single drone.

[0018] Step 2: Model the acquired group target information as a weighted fully connected graph, calculate the centrality of each target in the weighted graph, and obtain the center position of the group target. This position is used as the horizontal flight position of the drone or the center position of the drone's visual detection system.

[0019] Specifically, step 2 may include the following steps:

[0020] Step 2.1: Use group target information to build a weighted fully connected graph.

[0021] The target state obtained by the visual sensor is the position information of the camera (drone) as the reference point State = [x i ,y i ], where x i 、y i Indicates the horizontal and vertical coordinates of the i-th target position.

[0022] A fully connected graph G with weights is established by taking a finite set V(G) of all detected targets as vertices and the lines E(G) between the vertices. target =(V,E). Using the shortest distance between each target as the weight, establish the connection matrix A:

[0023]

[0024] Among them, a ij represents the distance between target i and target j.

[0025] Step 2.2: Find the center position of the group target.

[0026] Through the adjacency matrix of the weighted graph, we know that the elements in the i-th row of the matrix represent the shortest distance between the i-th target and other targets respectively. Therefore, the inverse of the sum of the elements in each row of the adjacency matrix is ​​used. i Indicates the centrality of the target in the entire target group:

[0027]

[0028] Among them, n represents the total number of targets, a ij Represents the shortest distance between target value i and target value j. C i The larger the value, the higher the centrality of the target. The target with the highest centrality is selected as the center of the group target.

[0029] Step 2.3: Make horizontal trajectory decisions for the drone.

[0030] In the horizontal direction, the center position of the group target is used as the flight position of the drone or the center position of the visual detection system. On this basis, the position of the target at the center position is used as the observation information:

[0031] Z=[x′ i ,y′ i ]

[0032] Where Z is the observation vector from the visual detection system, x′ i , y′ i are the horizontal and vertical coordinates of the target at the center. The extended Kalman filter method is used to predict the position of the target and obtain a smooth UAV flight trajectory in the horizontal position.

[0033] Step 3: Make the UAV’s adaptive flight altitude decision.

[0034] Based on the adjacency matrix of the weighted graph constructed from the group targets and the location of the central target, the maximum value of the row containing the central target in the adjacency matrix is ​​found. Using the relationship between this value, the radius of the visual inspection system, and the flight altitude, an offline reinforcement learning method is used to adaptively determine the flight altitude that maximizes the drone's monitoring performance.

[0035] Among them, the relationship between the radius of the visual inspection system and the flight altitude means that when the changes in parameters affecting the field of view, such as the focal length of the camera in the visual inspection system, have reached their thresholds, the distance between the camera and the object can be adaptively adjusted to change the field of view.

[0036] Specifically, step 3 may include the following steps:

[0037] Step 3.1: Collect offline datasets for the flight mission.

[0038] By collecting target state and altitude data of manned aircraft equipped with the same visual detection system when performing similar tasks, the offline dataset is used as input for training the offline reinforcement learning method.

[0039] Step 3.2: Set up the offline reinforcement learning elements in flight altitude decision.

[0040] The status includes the distance of the farthest target from the central target in the group, the flight altitude of the drone, and the detection radius of the visual detection system, as follows:

[0041] State=[d max ,h uav ,r reg ]

[0042] Among them, State represents the state of the drone, d max Indicates the distance between the farthest target and the central target, h uav Indicates the flight altitude of the UAV, r reg Indicates the detection range radius of the drone.

[0043] Reward settings include the drone's field of view utilization, field of view clarity, and altitude change rate, as follows:

[0044] Reward=w r (r reg -d max / r reg ) -1 +w h (h t,uav -h min ) -1 +w t (h t,uav -h t-1,uav ) -1

[0045] Among them, r reg is the detection radius of the visual inspection system, d max is the maximum distance from the center target, (r reg -d max / r reg ) -1 It represents the utilization rate of the UAV’s field of view, which aims to monitor the target with the clearest possible vision; r is the proportional factor that controls the utilization of vision in the total reward; h t,uav is the real-time flight altitude of the UAV at time t, h min is the minimum flight altitude of the UAV, (h uav -h min ) -1 In consideration of the energy consumption and visual resolution of the drone, the drone's altitude should be as low as possible while meeting the monitoring requirements of all targets. h is the proportional factor of the minimum height of the drone in the total reward; h t-1,uav is the flight altitude of the UAV at time t-1, (h t,uav -h t-1,uav ) -1 In order to save energy and smooth the flight path, the drone should avoid frequent changes in altitude. t It is the proportional factor that controls the change in the drone's altitude in the total reward.

[0046] Step 3.3: Make the drone's altitude decision.

[0047] Based on the offline dataset and the set state and reward, the training model is obtained by training with the offline reinforcement learning algorithm.

[0048] After the model is deployed, the optimal flight altitude is obtained adaptively each time a decision is made.

[0049] Step 4: Continuously monitor the task for exception handling.

[0050] When the UAV is at the maximum detection range of the visual detection system, if the maximum value of the row where the central target in the adjacency matrix is ​​located exceeds the detection radius at this time, it is considered that the target has exceeded the monitoring capability of the current UAV, and other UAVs are requested to cooperate or report to the command center for processing.

[0051] The maximum value of the detection range refers to the detection range of the visual detection system when the UAV has reached the maximum flight altitude, that is, if the UAV flies above this altitude, the target detection accuracy will be affected.

[0052] At this point, from step 1 to step 4, based on the group targets detected by the visual inspection system, a weighted fully connected graph is modeled, horizontal direction decisions are made based on the center position estimation of the graph, and height direction decisions are made based on offline reinforcement learning training, thus achieving three-dimensional continuous monitoring of the group targets.

[0053] Beneficial effects

[0054] Compared with the prior art, the method of the present invention has the following advantages:

[0055] This method addresses the problem of continuous three-dimensional monitoring of swarms of targets by a single drone based on visual detection in adversarial environments. It utilizes a weighted graph structure to model the swarm. Horizontally, flight decisions are made by finding the center position. In altitude, it leverages an offline dataset combined with an offline reinforcement learning algorithm to achieve adaptive flight altitude decisions, taking into account the drone's energy consumption and visual system accuracy. This allows for adaptive adjustment of the drone's detection range, improving the monitoring performance of individual drones. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Flowchart of the method of the present invention. DETAILED DESCRIPTION

[0057] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0058] like Figure 1 As shown, a method for three-dimensional continuous monitoring of UAV swarm targets with adaptive range includes the following steps:

[0059] Step A: Detection system settings;

[0060] Specifically in this embodiment, the details are the same as step 1 in the invention content;

[0061] Step B: Modeling group target information;

[0062] Specifically in this embodiment, the target is used as a vertex and the distance between the targets is used as the weight to establish a weighted fully connected graph, which is the same as step 2.1 of the invention;

[0063] Step C: Calculate the center position of the group target;

[0064] Specifically in this embodiment, the details are the same as step 2.2 of the invention;

[0065] Step D: Generate horizontal UAV trajectory decision;

[0066] Specifically in this embodiment, the details are the same as step 2.3 of the invention;

[0067] Step E: Collect flight offline dataset:

[0068] Specifically in this embodiment, it is the same as step 3.1 of the invention;

[0069] Step F: UAV flight altitude adaptive decision;

[0070] Specifically in this embodiment, it is the same as step 3.2 and step 3.3 of the invention;

[0071] Step G: Continuously monitor the exception handling in the task;

[0072] Specifically in this embodiment, it is the same as step 4 in the invention content.

[0073] Example

[0074] This example uses the scenario of "a drone equipped with multiple cameras continuously monitoring eight moving targets in a hostile environment. All targets have the same motion trend and can be treated as a group target. The visual detection system's detection range is adjusted based on the distance between the targets and the drone." This example uses a specific example to illustrate the adaptive range, three-dimensional, continuous drone swarm target monitoring method of the present invention. The specific steps of the method are detailed below.

[0075] A three-dimensional continuous monitoring method for UAV swarm targets with adaptive range, such as Figure 1 As shown, the following steps are included:

[0076] Step A: Detection system settings;

[0077] Specifically in this embodiment, a circular detection field of view with a radius of R is obtained by stitching together multiple cameras carried by the drone;

[0078] Step B: Modeling group target information;

[0079] Specifically in this embodiment, a weighted graph is established with 8 targets as vertices and the distances between targets as weights, and an 8×8 adjacency matrix is ​​obtained;

[0080] Step C: Calculate the center position of the group target;

[0081] Specifically in this embodiment, based on the adjacency matrix, the target centrality is calculated to obtain a 1×8 vector, and the target with the largest centrality is selected as the center position;

[0082] Step D: Generate horizontal UAV trajectory decision;

[0083] Specifically in this embodiment, the extended Kalman filter algorithm is used to track the target at the center position to obtain a horizontal trajectory.

[0084] Step E: Collect flight offline dataset:

[0085] Specifically in this embodiment, the same visual detection system is used, with the pilot completing the target monitoring task and recording the flight data;

[0086] Step F: UAV flight altitude adaptive decision;

[0087] Specifically in this embodiment, the target state and reward are set, and the flight altitude of the drone is obtained using an offline reinforcement learning algorithm;

[0088] Step G: Handling abnormal targets in continuous monitoring tasks;

[0089] Specifically in this embodiment, when the farthest target exceeds the maximum range of the target detection system, the command system is reported or the companion drones are informed for collaborative monitoring.

Claims

1. A three-dimensional continuous monitoring method for drone swarm targets with adaptive range adjustment, characterized in that: The following steps are involved: Step 1: Set the drone's onboard perception system to radar silence and use only visual sensors for target detection. The target detection area of ​​the drone is obtained by stitching together multiple visual sensors to obtain group target information. Step 2: Model the obtained group target information as a weighted fully connected graph, calculate the centrality of each target in the weighted graph, and obtain the central position of the group target; the shortest distance between targets is used as the weight; This position is used as the horizontal flight position of the UAV or the center position of the UAV's visual detection system; Step 3: Make the UAV’s adaptive flight altitude decision; Based on the adjacency matrix of the weighted graph created by the group targets and the location of the central target, the maximum value of the row containing the central target in the adjacency matrix is ​​found. The relationship between this value and the radius of the visual inspection system and the flight altitude is used, and an offline reinforcement learning method is used to adaptively obtain the flight altitude that maximizes the drone's monitoring performance. The relationship between the radius of the visual inspection system and the flight altitude refers to the ability to adaptively adjust the distance between the camera and the object to change the field of view when the change in the parameters affecting the field of view in the visual inspection system has reached its threshold. The parameters include the focal length of the camera. Step 4: Continuously monitor the exception handling of the task; When the drone is at the maximum detection range of the visual detection system, if the maximum value of the row where the central target in the adjacency matrix is ​​located exceeds the detection radius at that time, it is considered that the target has exceeded the monitoring capability of the current drone, and other drones are requested to cooperate, or the command center is reported for processing; The maximum detection range refers to the detection range of the visual detection system when the drone has reached its maximum flight altitude. If the drone flies above this altitude, the target detection accuracy will be affected. Based on the group targets detected by the visual inspection system, weighted fully connected graph modeling is performed, horizontal direction decisions are made according to the center position estimation of the graph, and height direction decisions are made based on offline reinforcement learning training, thus achieving three-dimensional continuous monitoring of the group targets.

2. The method for three-dimensional continuous monitoring of a drone swarm with adaptive range adjustment according to claim 1, characterized in that: Step 2 includes the following steps: Step 2.1: Use group target information to build a weighted fully connected graph; The target state obtained by the visual sensor is the position information State=[x i ,y i ], where x i 、y i Indicates the horizontal and vertical coordinates of the i-th target position; A fully connected graph G with weights is established by taking a finite set V(G) of all detected targets as vertices and the lines E(G) between the vertices. target =(V,E); using the shortest distance between each target as the weight, establish the adjacency matrix A: Among them, a ij represents the distance between target i and target j; Step 2.2: Find the center position of the group target; The elements in the i-th row of the matrix represent the shortest distance between the i-th target and other targets, and the inverse of the sum of the elements in each row of the adjacency matrix is ​​C. i Indicates the centrality of the target in the entire target group: Among them, n represents the total number of targets, a ij represents the shortest distance between target value i and target value j; C i The larger the value, the higher the centrality of the target; the target with the highest centrality is selected as the center of the group target; Step 2.3: Make horizontal trajectory decisions for the UAV; In the horizontal direction, the center position of the group target is used as the flight position of the drone or the center position of the visual detection system, and the position of the target at the center position is used as the observation information: Z=[x′ i ,y′ i ] Where Z is the observation vector from the visual detection system, x′ i , y′ i are the horizontal and vertical coordinates of the target at the center position respectively; the extended Kalman filter method is used to predict the position of the target and obtain a smooth UAV flight trajectory in the horizontal position.

3. The method for three-dimensional continuous monitoring of a drone swarm with adaptive range adjustment according to claim 1, characterized in that: Step 3 includes the following steps: Step 3.1: Collect offline datasets of flight missions; By collecting target state and altitude data from manned aircraft equipped with the same visual detection system performing similar tasks, the data is used as an offline dataset for input into the training of offline reinforcement learning methods. Step 3.2: Set up the offline reinforcement learning elements in the flight altitude decision; The status includes the distance of the farthest target from the central target in the group, the flight altitude of the drone, and the detection radius of the visual detection system, as follows: State=[d max ,h uav ,r reg ] Among them, State represents the state of the drone, d max Indicates the distance between the farthest target and the central target, h uav Indicates the flight altitude of the UAV, r reg Indicates the detection range radius of the drone; Reward settings include the drone's field of view utilization, field of view clarity, and altitude change rate, as follows: Reward=w r (r reg -d max / r reg ) -1 +w h (h t,uav -h min ) -1 +w t (h t,uav -h t-1,uav ) -1 Among them, r reg is the detection radius of the visual inspection system, d max is the maximum distance from the center target, (r reg -d max / r reg ) -1 It represents the utilization rate of the UAV’s field of view, which aims to monitor the target with the clearest possible vision; r is the proportional factor that controls the utilization of vision in the total reward; h t,uav is the real-time flight altitude of the UAV at time t, h min is the minimum flight altitude of the UAV, (h uav -h min ) -1 In consideration of the energy consumption and visual resolution of the drone, the drone's altitude should be as low as possible while meeting the monitoring requirements of all targets. h is the proportional factor of the minimum height of the drone in the total reward; h t-1,uav is the flight altitude of the UAV at time t-1, (h t,uav -h t-1,uav ) -1 In order to save energy and smooth the flight path, the drone should avoid frequent changes in altitude. t is the proportional factor that controls the change in the drone's altitude in the total reward; Step 3.3: Make the UAV’s altitude decision; Based on the offline dataset and the set state and reward, the offline reinforcement learning algorithm is used for training to obtain a training model; After the model is deployed, the optimal flight altitude is obtained adaptively each time a decision is made.

Images