Coordinated multi-target hunting method of drone swarm based on probabilistic graphical model

Abstract

The present invention relates to a drone cluster collaborative multi-target capture method based on a probability graph model, comprising: establishing a hierarchical probability graph model, and determining model nodes and node states; obtaining scene parameters in a capture scene, and initializing scene parameters and state values ​​of model nodes; individual drones detect and perceive the surrounding environment to obtain individual perception information, determine the current capture stage based on the individual perception information, and determine the expected node state according to the capture stage; calculate the conditional probability distribution of the expected node state, and update the node probability parameters of the model nodes; make decisions on actions performed by individual drones based on the expected node states and node probability parameters; the individual drones execute the output actions and update their positions, the individual drones detect and perceive the surrounding environment again, and determine the current capture stage based on the newly obtained individual perception information, and if it is the capture stage, the capture is successful.

Description

Technical Field

[0001] The present invention relates to the field of drone cluster collaborative target capture, and in particular to a drone cluster collaborative multi-target capture method based on a probabilistic graph model. Background Art

[0002] In recent years, the collaboration of UAV swarm systems has attracted extensive attention due to their swarm intelligence characteristics. The application range has expanded from aggregation and search to cluster formation and collaborative hunting, especially in adversarial and dynamic uncertain environments, where collaborative hunting has become a key research area. The main challenge is to optimize the collaborative strategy between individual UAV swarms while ensuring their safety in the face of uncertainties such as target movement and environmental obstacles. Traditional centralized control methods are limited by their computational complexity and communication delay, making them difficult to scale to larger systems. In contrast, distributed algorithms allow autonomous decision-making through local perception, reduce the computational load, and are more suitable for UAV swarm systems. However, existing hunting methods still have shortcomings such as formation configuration constraints, weak adaptability, large communication requirements, and poor interpretability, which reduce their effectiveness in multi-target scenarios in dynamic uncertain environments and communication denial conditions. Therefore, it is still a challenge to achieve fully distributed multi-target hunting based on local perception information without relying on communication and predefined formation configurations. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a drone cluster collaborative multi-target capture method based on a probabilistic graphical model.

[0004] To achieve the above-mentioned purpose, the present invention provides a drone cluster collaborative multi-target capture method based on a probabilistic graphical model, comprising:

[0005] S1. Establishing a hierarchical probability graph model for multi-target collaborative roundup action decision-making, and determining the model nodes and node states in the hierarchical probability graph model;

[0006] S2. Obtaining scene parameters for the drone cluster capture scene, and initializing the scene parameters and the state values ​​of the model nodes;

[0007] S3. The individual drones in the drone cluster detect and sense the surrounding environment to obtain individual perception information, and determine the current capture stage based on the individual perception information, and determine the desired node state according to the capture stage; wherein the capture stage includes: a search stage, a siege stage, and a capture stage;

[0008] S4. Calculate the conditional probability distribution of the desired node state based on the individual perception information, and update the node probability parameters of the model nodes in the hierarchical probability graph model;

[0009] S5. Based on the desired node state and the node probability parameter, the action performed by the drone individual is decided and output, wherein the action that maximizes the probability of the drone individual reaching the desired node state value is selected as the output;

[0010] S6. The individual drone executes the output action and updates its position. After the position update, the individual drone re-detects and perceives the surrounding environment, and determines the current capture stage based on the newly acquired individual perception information. If the capture stage is the capture stage, the capture is successful, otherwise, steps S3 to S6 are re-executed.

[0011] According to one aspect of the present invention, in step S1, in the step of establishing a hierarchical probability graph model for multi-target collaborative encirclement action decision-making and determining the model nodes and node states in the hierarchical probability graph model, the hierarchical probability graph model includes: a decision layer for deciding and outputting individual drone actions, a perception layer for the individual drone to detect and perceive the surrounding environment to obtain individual perception information;

[0012] The model nodes include: a decision action node set in the decision layer, and a prey-seeking node, a collision avoidance node, a target density node, a neighbor density node, and an attack situation node set in the perception layer; wherein the decision action node is a parent node, and the prey-seeking node, the collision avoidance node, the target density node, the neighbor density node, and the attack situation node are all child nodes.

[0013] According to one aspect of the present invention, in step S1, in the step of determining the model nodes and node states in the hierarchical probabilistic graphical model, the model nodes and the node states are respectively expressed as:

[0014] Decision action node, which is represented by , the number of states is 32, and the state value is Action 1 ~ Action 32 ;

[0015] Find prey nodes, which are represented by , the number of states is 2, and the state value is yes / no;

[0016] The collision avoidance node is represented as , the number of states is 2, and the state value is yes / no;

[0017] The target density node is represented by , the number of states is 2, and the state value is good / bad;

[0018] Neighbor density node, which is expressed as , the number of states is 2, and the state value is good / bad;

[0019] Attack situation node, which is represented by , the number of states is 2, and the state value is good / bad;

[0020] in, Action 1 ~ Action 32 Represents the optional actions in the individual action space of the individual drone;

[0021] The individual action space is expressed as:

[0022] ;

[0023] ;

[0024] in, is the acceleration value of the drone, Indicates the direction of the acceleration of the drone individual, is the variable coefficient, is the maximum acceleration value of the drone individual, is a set of natural numbers, is the variable coefficient, is the sign number, used to indicate the A drone individual.

[0025] According to one aspect of the present invention, in step S3, in the step of detecting and sensing the surrounding environment to obtain individual perception information of the individual drones in the drone cluster, the individual perception information includes: companion information, target information and obstacle information;

[0026] In step S3, the current capture phase is determined based on the individual perception information, and the desired node state is determined according to the capture phase. When the drone individual does not perceive the target, it is determined that the drone individual is in the search phase, and the task of the drone individual in the search phase is to search for prey. The corresponding desired node state is and ;

[0027] When the drone senses the target, it is determined that the drone is in the encirclement stage, and the drone's task in the encirclement stage is to surround the target. The corresponding desired node state is , and ;

[0028] When the drone senses the target and determines that the target is surrounded, the drone is in the capture phase, and the drone's task in the capture phase is to capture the target. The corresponding desired node state is and .

[0029] According to one aspect of the present invention, in step S3, when the drone individual senses the target, it is determined whether the drone individual is in the encirclement stage based on the Apollonius circle, wherein the Apollonius circle is constructed based on the drone individual and the target;

[0030] When the drone individual senses that the target is surrounded by three or more Apollonius circles, and the Apollonius circles surrounding the target are sequentially tangent or intersecting, it is determined that the target is successfully surrounded, that is, the drone individual is in the encirclement stage.

[0031] According to one aspect of the present invention, in step S3, the step of constructing the Apollonius circle based on the drone individual and the target includes:

[0032] The perceived target is the first fixed point , taking the drone as the second fixed point ;

[0033] Get the first fixed point and the second fixed point The circular trajectory points between , where the circular trajectory point With the first fixed point The first distance between the points and the circular trajectory With the second fixed point The ratio of the second distance between ,and ;

[0034] Based on the circular trajectory points Construct the encirclement of the second fixed point Circle of Apollonius.

[0035] According to one aspect of the present invention, in step S4, in the step of calculating the conditional probability distribution of the expected node state value based on the individual perception information, the state probability distribution of the decision action node is expressed as:

[0036] ;

[0037] in, represents the number of actions in the action space of the drone individual, that is, the number of states of the decision action node, Indicates the node j status, Indicates the state of the node, subscript j For the flag bit;

[0038] The state probability distribution of the prey-seeking node is expressed as:

[0039] ;

[0040] in, Indicates the first state of the node, represents the total value of pheromones within the sensing range, and , Indicates the current moment, Represents a virtual drone At the present moment The sensing range of the virtual drone By drone individual Execute decision action Action j When generated, Indicates the range of action of the pheromone point. Represents the intersection of the perception range and the pheromone action range. represents pheromone points, Indicates the pheromone point at the current time The value of is expressed as: , Indicates the time when the pheromone point is generated. Indicates the number of signs, indicating the individual drones;

[0041] The state probability distribution of the collision avoidance node is expressed as:

[0042] ;

[0043] ;

[0044] in, Indicates The set of entities perceived by each drone, Represents the predicted perceived entity , Represents a virtual drone Entities with predictions The distance between Represents individual drones and the perceived entity The safe distance between To avoid collision distance, the perceived entity Includes obstacles and other drones that are being pursued ;

[0045] The state probability distribution of the target density node is expressed as:

[0046] ;

[0047] in, Represents a virtual drone The relative target density value is used to estimate the spatial distribution of surrounding targets. Represents individual drones Expected density value relative to the target;

[0048] The state probability distribution of the neighbor density node is expressed as:

[0049] ;

[0050] in, Represents a virtual drone Relative density values ​​relative to neighbors, used to estimate individual drones R i The spatial distribution of neighbors, Represents individual drones The expected density value relative to the neighbors, Indicates the flag bit, used to indicate neighbors;

[0051] The state probability distribution of the attack situation node is expressed as:

[0052] ;

[0053] in, Represents individual drones The target of the capture, Represents a virtual drone With the target of the capture The distance between.

[0054] According to one aspect of the present invention, in step S5, in the step of selecting the action that maximizes the probability of the drone individual reaching the desired node state value as the output, the joint probability of each model node is calculated, and according to the desired node state under the different divided round-up stages, the maximum value of the joint probability of each model node reaching the desired node state is obtained to obtain the action with the maximum probability; wherein the joint probability is expressed as:

[0055] ;

[0056] in, Represents a child node At a given parent node The conditional probability under the state, that is, the node probability parameter, n Indicates the number of child nodes.

[0057] Beneficial effects of the present invention:

[0058] According to one solution of the present invention, the distributed collaborative multi-target capture method of drone swarm proposed by the present invention combines the probabilistic graph model and smoothed particle fluid dynamics, so that the drone swarm can effectively capture highly maneuverable escaping targets in a dynamic environment. The method can operate in an unknown environment with communication denial and dynamic obstacles, and does not require target allocation and is not restricted by the formation configuration. By combining local perception with real-time probability estimation, high adaptability and scalability are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A flowchart of a method for cooperative multi-target capture by drone swarms based on a probabilistic graphical model according to an embodiment of the present invention;

[0060] Figure 2 A flowchart of a drone cluster collaborative multi-target capture method based on a probabilistic graphical model according to an embodiment of the present invention;

[0061] Figure 3 A hierarchical probability graph model diagram of an embodiment of the present invention;

[0062] Figure 4 An individual motion space diagram of an individual drone according to an embodiment of the present invention;

[0063] Figure 5 An individual motion space diagram of an individual drone according to an embodiment of the present invention;

[0064] Figure 6 A schematic diagram of an Apollonius circle according to an embodiment of the present invention;

[0065] Figure 7 A schematic diagram of a scene in which a target is surrounded in one embodiment of the present invention, wherein (a) represents the target being surrounded by multiple Apollonius circles in a tangent state. (b) shows the target under the state of multiple Apollonius circles intersecting The encircling scene;

[0066] Figure 8 A diagram of a simulation experiment environment of Example 1 of an implementation manner of the present invention;

[0067] Fig. 9This is a diagram of experimental results of a simulation experiment of Example 1 of an implementation manner of the present invention;

[0068] Fig.10 In the simulation experiment of Example 1 of an implementation mode of the present invention The key frame display diagram when

[0069] Fig.11 In the simulation experiment of Example 1 of an implementation mode of the present invention The key frame display diagram when

[0070] Fig.12 In the simulation experiment of Example 1 of an implementation mode of the present invention The key frame display diagram when

[0071] Fig.13 In the simulation experiment of Example 1 of an implementation mode of the present invention The key frame display diagram when

[0072] Fig.14 In the simulation experiment of Example 1 of an implementation mode of the present invention The key frame display diagram when

[0073] Fig.15 In the simulation experiment of Example 1 of an implementation mode of the present invention The key frame display diagram when

[0074] Fig.16 In the simulation experiment of Example 1 of an implementation mode of the present invention The key frame display diagram when

[0075] Fig.17 In the simulation experiment of Example 1 of an implementation mode of the present invention The key frame display diagram when

[0076] Fig.18 A diagram of a physical experiment environment of Example 2 of an implementation mode of the present invention;

[0077] Fig.19 A diagram of a physical experiment result of an initial position in a physical experiment of Example 2 of an implementation manner of the present invention;

[0078] Fig. 20 In the physical experiment of Example 2 of one embodiment of the present invention The actual experimental results of searching for the target;

[0079] Fig.21 In the physical experiment of Example 2 of one embodiment of the present invention The actual experimental results of capturing the target at the same time;

[0080] Fig. 22 In the physical experiment of Example 2 of one embodiment of the present invention The actual experimental result diagram of the target No. 2 being captured;

[0081] Fig.23 In the physical experiment of Example 2 of one embodiment of the present invention The actual experimental result diagram of continuing to capture the remaining targets;

[0082] Fig.24 In the physical experiment of Example 2 of one embodiment of the invention The actual experimental result diagram of target No. 1 being captured. DETAILED DESCRIPTION

[0083] The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments cannot be described one by one here, but the embodiments of the present invention are not therefore limited to the following embodiments.

[0084] Combination Figure 1 and Figure 2 As shown, according to one embodiment of the present invention, a drone cluster collaborative multi-target capture method based on a probabilistic graphical model of the present invention includes:

[0085] S1. Establish a hierarchical probability graph model for multi-target collaborative roundup action decision-making, and determine the model nodes and node states in the hierarchical probability graph model;

[0086] S2. Obtaining scene parameters in the capture scene for the drone cluster, and initializing the scene parameters and state values ​​of the model nodes;

[0087] S3. The individual drones in the drone cluster detect and perceive the surrounding environment to obtain individual perception information, and determine the current capture stage based on the individual perception information, and determine the desired node state according to the capture stage; wherein the capture stage includes: search stage, encirclement stage and capture stage;

[0088] S4. Calculate the conditional probability distribution of the expected node state value based on the individual perception information, and update the node probability parameters of the model nodes in the hierarchical probability graph model;

[0089] S5. Based on the desired node state and node probability parameters, the actions performed by the individual drones are decided and output, wherein the action that maximizes the probability of the individual drone reaching the desired node state value is selected as the output;

[0090] S6. The individual drone executes the output action and updates its position. After the position is updated, the individual drone re-detects and perceives the surrounding environment, and determines the current capture stage based on the newly acquired individual perception information. If the capture stage is the capture stage, the capture is successful, otherwise, steps S3 to S6 are re-executed.

[0091] According to an embodiment of the present invention, in step S1, a hierarchical probability graph model for multi-target collaborative roundup action decision-making is established, and the model nodes and node states in the hierarchical probability graph model are determined. Figure 3 The hierarchical probability graph model diagram includes: a decision layer for outputting individual actions of drones, and a perception layer for drones to detect and perceive the surrounding environment to obtain individual perception information; wherein the model nodes include: decision action nodes set in the decision layer (see Figure 3 Node settings in the decision layer), set up prey-seeking nodes, collision avoidance nodes, target density nodes, neighbor density nodes, and attack situation nodes in the perception layer (see Figure 3 Node settings of the perception layer in the middle); among them, the decision action node is the parent node, and the prey search node, collision avoidance node, target density node, neighbor density node, and attack situation node are all child nodes.

[0092] In this embodiment, the specific parameters of the model nodes and node states can be found in Table 1;

[0093] Table 1

[0094]

[0095] in, Action 1 ~ Action 32 Represents the optional actions in the individual action space of the drone individual (see Figure 4 );

[0096] In this embodiment, the individual action space is expressed as:

[0097] ;

[0098] ;

[0099] in, is the acceleration value of the drone, Indicates the direction of the acceleration of the drone individual, is the variable coefficient, is the maximum acceleration value of the drone individual, is a set of natural numbers, is the variable coefficient, is the sign number, used to indicate the A drone individual.

[0100] According to one embodiment of the present invention, in step S2, scene parameters in the capture scene for the drone cluster are obtained, and the scene parameters and the state values ​​of the model nodes are initialized. The scene parameters include: site size, number of drone individuals, number of targets, individual maximum speed, target maximum speed, individual maximum acceleration, and target maximum acceleration; wherein corresponding initialization settings can be performed according to specific actual values.

[0101] According to an embodiment of the present invention, in step S3, individual drones in the drone cluster detect and sense the surrounding environment to obtain individual perception information, and the individual perception information includes: companion information, target information and obstacle information; further, in step S3, the current capture stage is determined based on the individual perception information, and the desired node state is determined according to the capture stage, mainly based on the companion information and target information in the individual perception information to achieve the current capture stage. For details, see Figure 5 The individual action space diagram of the drone individual shows the relationship between the search phase, the encirclement phase and the capture phase. When the drone individual does not perceive the target (i.e., the target is lost), the drone individual is judged to be in the search phase (see Figure 5 ), and the task of the drone individual in the search phase is to find prey (i.e., escape target), and the corresponding desired node state is and ; Therefore, the node state expected by the model node means that the action decided by the individual drone should enable the individual drone to search for the target as much as possible while avoiding collision to ensure its own safety.

[0102] For further information, see Figure 5 , when the drone individual senses the target (i.e. discovers the target), it is determined that the drone individual is in the encirclement stage, and the drone individual’s task in the encirclement stage is to surround the target. The corresponding desired node state is , and ; Therefore, the meaning of the node state expected by the model node is: the action decided by the drone individual should make the target density and companion density perceived by the drone individual as close to their respective expected values ​​as possible, while avoiding collisions to ensure its own safety. This will enable the individual to approach the target and form an encirclement situation on the target while ensuring its own safety.

[0103] For further information, see Figure 5, when the drone individual senses the target and determines that the target is surrounded (i.e., surrounds the target), the drone individual is in the capture phase (i.e., the attack phase of approaching and capturing the target), and the drone individual’s task in the capture phase is to capture the target. The corresponding desired node state is and ; Therefore, the meaning of the node state expected by the model node is: the action decided by the drone individual should make the drone individual approach and capture the target as much as possible while avoiding collision to ensure its own safety.

[0104] From the above, we can see that the roundup phase and the corresponding expected node state (i.e., node expected state) can be summarized as shown in Table 2;

[0105] Table 2

[0106]

[0107] According to an embodiment of the present invention, in step S3, when the drone individual senses the target, it is determined whether the drone individual is in the encirclement stage based on the Apollonius circle, wherein the Apollonius circle is constructed based on the drone individual and the target (see Figure 6 ); Specifically, when the drone perceives that the target is surrounded by three or more Apollonius circles, and the Apollonius circles surrounding the target are tangent in sequence (see Figure 7 (a)) or intersect (see Figure 7 In (b), these Apollonius circles cover the escapeable area around the target, and the target is then judged to be successfully surrounded, that is, the individual drone is in the encirclement stage. Of course, it can be understood that no matter which direction the target moves, it will be intercepted by at least one individual drone, thus achieving successful encirclement.

[0108] Furthermore, to facilitate understanding of this scheme, the concept of the Apollonius circle adopted by it is briefly described, where the Apollonius circle is a set of points on a plane, where the ratio of the distance to two fixed points is a constant . As attached Figure 6 As shown, when the second fixed point and the first fixed point are the two fixed points of the Apollonius circle, points The locus of is the Apollonius circle, that is, the point is the circular trajectory point; among them, the circular trajectory point The first distance from the first fixed point E and the circular locus point With the second fixed point The ratio of the second distance between (in ). Furthermore, when it is applied in the present invention, the perceived target is taken as the first fixed point , taking the drone as the second fixed point ; Then we can get the first fixed point and the second fixed point The circular trajectory points between , where the circular trajectory point With the first fixed point The first distance between the points and the circular trajectory With the second fixed point The ratio of the second distance between ,and ; Based on circular trajectory points Construct encircling the second fixed point In this embodiment, the ratio of the first distance to the second distance can be achieved by using the speed ratio of the pursuing drone individual and the target, that is, the speed ratio is a constant Therefore, the pursuing drone and the target can reach the Apollonius circle at the same time. point, thereby achieving the interception of the target. Using this feature as the theoretical basis for successfully encircling a target, when the drone individual perceives that the target is surrounded by three (or more) Apollonius circles (formed by the drone individual and the target), and these Apollonius circles are tangent or intersecting in sequence, so that these Apollonius circles cover the escapeable area around the target, the target can be determined to be successfully surrounded. This shows that no matter which direction the target moves, it will be intercepted by at least one pursuing drone individual, thus achieving a successful encirclement. The scene of successful encirclement is shown in the attached figure. Figure 7 shown.

[0109] According to an embodiment of the present invention, in step S4, in the step of calculating the conditional probability distribution of the expected node state value based on the individual perception information, the state probability distribution of the decision action node is expressed as:

[0110] ;

[0111] in, represents the number of actions in the action space of the drone individual, that is, the number of states of the decision action node, Indicates the node j status, Indicates the state of the node, subscript j For the flag bit;

[0112] The method of obtaining it is as follows:

[0113] For decision action nodes For example, since the decision action node is the root node, it only has child nodes but no parent node, so the probability of this node belongs to the prior probability, which can generally be given based on expert experience. When there is no external input, it is considered that the probability distribution of this node follows a uniform distribution, that is, the probability of each state is equal. Therefore, the state probability distribution of the decision action node is expressed as:

[0114] .

[0115] Furthermore, since the decision action node It is the root node of the model, and its child nodes have only one common parent node. , so its child nodes , , , and The conditional probability of the root node According to the conditional independence assumption of the hierarchical probabilistic graph model, that is, each node is independent of all other non-parent nodes, when a given parent node When the state of is taken, the conditional probabilities of its child nodes are independent of each other. For all state values ​​of , the conditional probabilities of the remaining nodes are calculated independently, thereby obtaining the conditional probability distribution of the nodes of the entire hierarchical probabilistic graph model.

[0116] In this embodiment, the state probability distribution of the prey-seeking node is expressed as:

[0117] ;

[0118] in, Indicates the first state of the node, represents the total value of pheromones within the sensing range, and , Indicates the current moment, Represents a virtual drone At the present moment The sensing range of the virtual drone By drone individual Execute decision action Action j When generated, Indicates the range of action of the pheromone point. Represents the intersection of the perception range and the pheromone action range. represents pheromone points, Indicates the pheromone point at the current time The value of is expressed as: , Indicates the time when the pheromone point is generated. Indicates the number of signs, indicating the individual drones;

[0119] The method of obtaining it is as follows:

[0120] The prey-seeking node solves the problem of target location uncertainty by combining the concept of biological pheromones to help the tracker (i.e., individual drones) locate the target. Unknown areas should be explored extensively to maximize the chance of finding the target. However, since the location of the target changes dynamically, if the individual drones are not considered As for the timeliness of the exploration range, there is a situation where the target cannot be found even after exploring the entire scene.

[0121] In order to solve the above problems, the present invention introduces the probability of biopheromone to mark individual drones. By taking advantage of the pheromone decay characteristics over time, the drones are given In this embodiment, for each drone individual A virtual pheromone map is constructed. Each move will leave a pheromone point at the corresponding position. , and record the time when the pheromone point is generated as , the pheromone point Scope of action Defined as a drone individual exist The perception range at the moment. Furthermore, the law of the decay of pheromone value over time is expressed as:

[0122] ;

[0123] Therefore, in the drone individual Within the perception range, the smaller the total pheromone value, the lower the possibility that the area has been explored, and the more likely it is to find the target.

[0124] Furthermore, the state probability distribution of the prey-seeking node can be expressed as:

[0125] ;

[0126] in, The larger the value, the more likely it is that the individual drone has explored the area, and the lower the probability of detecting the target.

[0127] In this embodiment, the state probability distribution of the collision avoidance node is expressed as:

[0128] ;

[0129] ;

[0130] in, Indicates The set of entities perceived by each drone, Represents the predicted perceived entity , Represents a virtual drone Entities with predictions The distance between Represents individual drones and the perceived entity The safe distance between To avoid collision distance;

[0131] The method of obtaining it is as follows:

[0132] The collision avoidance node is used to avoid collisions between individual drones. The probability of collision is estimated by predicting the position of the perceived entity (including obstacles and other drones that are chasing targets) and combining it with the current position of the individual drone. In this embodiment, the principle of least squares method is used to analyze the observed historical trajectory of the perceived entity to predict its future position. The next predicted position of the perceived entity can be expressed as:

[0133] ;

[0134] in, represents the location of the perceived entity, It indicates the change in the position of the perceived entity. The subscript c is a flag used to indicate the perceived entity.

[0135] Therefore, the state probability distribution of the collision avoidance node is expressed as:

[0136] ;

[0137] ;

[0138] Among them, when the virtual drone Entities with predictions When the distance between the two is greater than the collision avoidance distance, it is considered that there will be no collision between the two, and the probability of successful collision avoidance is 1; when the virtual drone Entities with predictions When the distance between the virtual drones is less than the safe distance, the probability of collision avoidance is considered to be 0; Entities with predictions When the distance between them is between the collision avoidance distance and the safety distance, the probability of successful collision avoidance is expressed as It can be seen that the closer the pursuing drone is to the perceived entity, the lower the probability of successful collision avoidance.

[0139] In this embodiment, the state probability distribution of the target density node is expressed as:

[0140] ;

[0141] in, Represents a virtual drone The relative target density value is used to estimate the spatial distribution of surrounding targets. Represents individual drones Expected density value relative to the target;

[0142] The method of obtaining it is as follows:

[0143] The target density node evaluates the interaction between the individual drones being pursued and the density field formed by the target. In this embodiment, the probability of robot density is introduced, and a density-based pursuer interaction mechanism is proposed. Traditional tracking models usually rely on absolute distance information for control. Compared with traditional tracking models, the present invention designs a density-based distributed control model. This control model utilizes the density and density field of the robot and adopts negative feedback control to regulate the interaction between the robot and the target. Therefore, the state probability distribution of the target density node can be expressed as:

[0144] ;

[0145] in, represents the reference density of the expected distribution of the individual drones controlling the pursuit relative to the target. The above formula is used as a density adjustment function to drive the individual drones Towards the desired density Move nearby.

[0146] In this embodiment, the calculation of the density value involved is as follows:

[0147] The smoothing function in the Smoothed Particle Hydrodynamics (SPH) method is introduced to assist in calculating the density value of the individual; the density calculation function is designed as:

[0148] ;

[0149] ;

[0150] ;

[0151] in, Represents an intermediate variable, indicating and The ratio of is the input of the density function, Represents nearby neighbor individuals and current drone individuals The distance between For drone individuals The perception radius, For drone individuals The relative spatial distribution of other individuals perceived by an individual can be converted into a density value using the density function. Represents individual drones No. l Therefore, individual drones The density value can be expressed as:

[0152] ;

[0153] in, represents the density value of the individual itself, Represents the perceived other individuals (neighboring individuals ), Represents the focal individual i Other individuals perceived (neighboring individuals )'s collection.

[0154] In this embodiment, the state probability distribution of the neighbor density node is expressed as:

[0155] ;

[0156] in, Represents a virtual drone Relative density values ​​relative to neighbors, used to estimate individual drones R i The spatial distribution of neighbors, Represents individual drones The expected density value relative to the neighbors, Indicates the flag bit, used to indicate neighbors;

[0157] The method of obtaining it is as follows:

[0158] The neighbor density node evaluates the interaction between the pursuer and the density field formed by its neighboring nodes. Then the state probability distribution of the neighbor density node can be expressed as:

[0159] ;

[0160] in, is the reference density, used to control the individual drones in pursuit Relative to its neighbors (i.e. other adjacent drone individuals ) is the expected distribution of . Therefore, the above formula converts the individual drones being pursued into The density of the drone was compared with the reference density. When the density exceeded the reference density, the individual drones in pursuit Mutually exclusive; when below the reference density, the pursuing drone individuals Mutual attraction. Furthermore, this negative feedback self-regulation ensures that the individual drones that are chasing A suitable spatial distribution density can be maintained.

[0161] In this embodiment, the state probability distribution of the attack situation node is expressed as:

[0162] ;

[0163] in, Represents individual drones The target of the capture, Represents a virtual drone With the target of the capture The distance between

[0164] The method of obtaining it is as follows:

[0165] The attack situation node is used to describe the individual drones being pursued. The impact of the behavior of the target on the attack posture. Usually, the individual drones that are chasing The closer to the target, the better the attack posture, and the greater the probability of successfully attacking the target. Therefore, the probability of a good or bad attack posture is related to the individual number of the pursuing drone. It is negatively correlated with the distance between the target and the state. Then the state probability distribution of the attack situation node is expressed as:

[0166] .

[0167] Based on the above formula, we can know that the individual drones that are being pursued The probability that the attack posture relative to the target is good is negatively correlated with the distance between the two. The farther the distance, the smaller the probability that the attack posture is good, and vice versa.

[0168] Therefore, the calculation based on the conditional probability distribution can be used to update the node probability parameters of the model nodes in the hierarchical probability graph model, that is, the probability of the node state expected by each model node can be updated.

[0169] According to an embodiment of the present invention, in step S5, in the step of selecting the action that maximizes the probability of the drone individual reaching the desired node state value as the output, the selected action can maximize the probability of the node reaching the desired state; specifically, the joint probability of each model node is calculated, and according to the desired node state under the different roundup stages (see Table 2), the maximum value of the joint probability that each model node reaches the desired node state is obtained to obtain the action that maximizes the probability; in this embodiment, according to the conditional independence assumption and chain rule of the probabilistic graphical model, the joint probability distribution can be expressed as the product of the conditional probability of each node. Therefore, the joint probability is expressed as:

[0170] ;

[0171] in, Represents a child node At a given parent node The conditional probability under the state, that is, the node probability parameter, n Indicates the number of child nodes. For the constructed hierarchical probabilistic graph model, each child node has one and only one parent node. Therefore, given each state of the parent node, the joint conditional probability of the remaining nodes can be calculated. Take the maximum value of the joint conditional probability and find the value that makes The action with the largest value is output as the decision action.

[0172] In order to further illustrate this scheme, simulation and physical examples are given to illustrate it.

[0173] Simulation experiment

[0174] (1) Experimental environment

[0175] Figure 8 The experimental setup shown consists of n autonomous drones chasing each other in a two-dimensional space of L × L units. m Static obstacles are represented by black entities, and dynamic obstacles are represented by gray. The perception radius of each drone is represented by a blue circle, and the perception radius of each target is represented by a red circle.

[0176] (2) Experimental results

[0177] The simulation test was conducted in an environment with both dynamic and static obstacles, with a size of 200×200, including 6 pursuers and 2 dynamic targets with escape capabilities. The targets were given high maneuverability. The maximum moving speed of the targets was 2m / s and the maximum acceleration was 1m / s. 2 ; The maximum speed of the pursuer is 1.7m / s, and the maximum acceleration is 1m / s 2 .

[0178] The simulation results are shown in the attached Figures 9 to 17 As shown in the figure, it can be seen that in a dynamic obstacle environment, the pursuer can make effective decisions to successfully avoid obstacles while capturing highly maneuverable targets. It is worth noting that the pursuer's trajectory shows obvious curvature near obstacles, indicating an effective obstacle avoidance effect and confirming the effectiveness of this method.

[0179] Physical experiment

[0180] (1) Experimental environment

[0181] like Fig.18 As shown in the figure, the experiment was conducted in a 5.2m×10m area, and the task of the six pursuers was to capture two highly maneuverable non-cooperative targets. The maximum speed of the pursuers was 0.2m / s, and the maximum acceleration was 0.2m / s 2 ; The maximum speed of the target is 0.3m / s, and the maximum acceleration is 0.3m / s 2 The experimental scene includes three static obstacles and two dynamic obstacles represented by other drones. To ensure that the target's initial position is out of the pursuer's detection range, the target and pursuer are placed at both ends of the area. Obstacles are placed in the center to increase the complexity of the environment, while dynamic obstacles increase uncertainty, making the roundup task more challenging.

[0182] (2) Experimental results

[0183] Figures 19 to 24 Some key frames of the multi-target capture experiment results in a real environment are shown. The initial state, intermediate state and final state of the system are shown in turn. Figures 19 to 24 A top-down thumbnail is provided in the lower right corner to show the scene more clearly. The red solid circle represents the target, the blue solid circle represents the pursuer, and the gray solid circle represents the dynamic obstacle. Static obstacles are represented by black, and the gray curve represents the motion trajectory. It can be seen that the pursuer can effectively avoid environmental obstacles and successfully capture all targets in 129.8 seconds, verifying the effectiveness of the proposed method.

[0184] In summary, the present invention is aimed at the problem of collaborative target hunting of drone clusters, and proposes a collaborative target hunting method for self-organizing drone clusters based on a hierarchical probabilistic graphical model and SPH theory, which can effectively pursue multiple agile, non-cooperative goals in communication-constrained and dynamically complex environments. The method eliminates the dependence on global communication by utilizing a local perception and autonomous decision-making framework, while the SPH theory optimizes the distribution of drones by adjusting the "density" between nodes. The method integrates a task-specific "perception-decision" PGM model, enabling drones to autonomously adjust their hunting strategies. The staged division of tasks and the "dispersed encirclement" strategy further improve the efficiency of hunting. Numerical simulations and actual experiments have demonstrated the effectiveness of this method in handling multi-target tracking tasks.

[0185] Therefore, the method proposed in the present invention can be effectively applied to the coordinated target capture of drone clusters in communication-constrained and dynamic complex environments, and has strong practical value.

[0186] The above contents are merely examples of specific solutions of the present invention. For devices and structures not described in detail therein, it should be understood that they can be implemented by adopting general devices and general methods available in the art.

[0187] The above is only one solution of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A drone cluster collaborative multi-target capture method based on a probabilistic graphical model, characterized in that: include: S1. Establishing a hierarchical probability graph model for multi-target collaborative encirclement action decision-making, and determining the model nodes and node states in the hierarchical probability graph model; wherein the hierarchical probability graph model includes: a decision layer for making decisions and outputting individual drone actions, and a perception layer for the individual drone to detect and perceive the surrounding environment to obtain individual perception information; The model nodes include: a decision action node set in the decision layer, and a prey-seeking node, a collision avoidance node, a target density node, a neighbor density node, and an attack situation node set in the perception layer; wherein the decision action node is a parent node, and the prey-seeking node, the collision avoidance node, the target density node, the neighbor density node, and the attack situation node are all child nodes; S2. Obtaining scene parameters for the drone cluster capture scene, and initializing the scene parameters and the state values ​​of the model nodes; S3. The individual drones in the drone cluster detect and sense the surrounding environment to obtain individual perception information, and determine the current capture stage based on the individual perception information, and determine the desired node state according to the capture stage; wherein the capture stage includes: a search stage, a siege stage, and a capture stage; S4. Calculate the conditional probability distribution of the desired node state based on the individual perception information, and update the node probability parameters of the model nodes in the hierarchical probability graph model; S5. Based on the desired node state and the node probability parameter, the action performed by the drone individual is decided and output, wherein the action that maximizes the probability of the drone individual reaching the desired node state value is selected as the output; S6. The individual drone executes the output action and updates its position. After the position update, the individual drone re-detects and perceives the surrounding environment, and determines the current capture stage based on the newly acquired individual perception information. If the capture stage is the capture stage, the capture is successful, otherwise, steps S3 to S6 are re-executed.

2. The method for cooperative multi-target capture by drone swarm based on probabilistic graphical model according to claim 1 is characterized in that: In step S1, in the step of determining the model nodes and node states in the hierarchical probability graph model, the model nodes and the node states are respectively expressed as: Decision action node, which is represented by , the number of states is 32, and the state value is Action 1~ Action 32 ; Find prey nodes, which are represented by , the number of states is 2, and the state value is yes / no; The collision avoidance node is represented as , the number of states is 2, and the state value is yes / no; The target density node is represented by , the number of states is 2, and the state value is good / bad; Neighbor density node, which is expressed as , the number of states is 2, and the state value is good / bad; Attack situation node, which is represented by , the number of states is 2, and the state value is good / bad; in, Action 1~ Action 32 Represents the optional actions in the individual action space of the individual drone; The individual action space is expressed as: in, is the acceleration value of the drone, Indicates the direction of the acceleration of the drone individual, is the variable coefficient, is the maximum acceleration value of the drone individual, is a set of natural numbers, is the variable coefficient, is the sign number, used to indicate the A drone individual.

3. The method for cooperative multi-target capture by drone swarm based on probabilistic graphical model according to claim 2 is characterized in that: In step S3, the individual drones in the drone cluster detect and perceive the surrounding environment to obtain individual perception information, and the individual perception information includes: companion information, target information and obstacle information; In step S3, the current capture phase is determined based on the individual perception information, and the desired node state is determined according to the capture phase. When the drone individual does not perceive the target, it is determined that the drone individual is in the search phase, and the task of the drone individual in the search phase is to search for prey. The corresponding desired node state is and ; When the drone senses the target, it is determined that the drone is in the encirclement stage, and the drone's task in the encirclement stage is to surround the target. The corresponding desired node state is , and ; When the drone senses the target and determines that the target is surrounded, the drone is in the capture phase, and the drone's task in the capture phase is to capture the target. The corresponding desired node state is and .

4. The method for cooperative multi-target capture by drone swarm based on probabilistic graphical model according to claim 3 is characterized in that: In step S3, when the drone individual senses the target, it is determined whether the drone individual is in the encirclement stage based on the Apollonius circle, wherein the Apollonius circle is constructed based on the drone individual and the target; When the drone individual senses that the target is surrounded by three or more Apollonius circles, and the Apollonius circles surrounding the target are sequentially tangent or intersecting, it is determined that the target is successfully surrounded, that is, the drone individual is in the encirclement stage.

5. The method for cooperative multi-target capture by drone swarm based on probabilistic graphical model according to claim 4 is characterized in that: In step S3, the steps of constructing the Apollonius circle based on the drone individual and the target include: The perceived target is the first fixed point , taking the drone as the second fixed point ; Get the first fixed point and the second fixed point The circular trajectory points between , where the circular trajectory point With the first fixed point The first distance between the points and the circular trajectory With the second fixed point The ratio of the second distance between ,and ; Based on the circular trajectory points Construct the encirclement of the second fixed point Circle of Apollonius.

6. The method for cooperative multi-target capture by drone swarm based on probabilistic graphical model according to claim 5 is characterized in that: In step S4, in the step of calculating the conditional probability distribution of the expected node state value based on the individual perception information, the state probability distribution of the decision action node is expressed as: in, represents the number of actions in the action space of the drone individual, that is, the number of states of the decision action node, Indicates the node status, Indicates the state of the node, subscript For the flag bit; The state probability distribution of the prey-seeking node is expressed as: in, Indicates the first state of the node, represents the total value of pheromones within the sensing range, and , Indicates the current moment, Represents a virtual drone At the present moment The sensing range of the virtual drone By drone individual Execute decision action Action j When generated, Indicates the range of action of the pheromone point. Represents the intersection of the perception range and the pheromone action range. represents pheromone points, Indicates the pheromone point at the current time The value of is expressed as: , Indicates the time when the pheromone point is generated. Indicates the number of signs, indicating the individual drones; The state probability distribution of the collision avoidance node is expressed as: in, Indicates The set of entities perceived by each drone, Represents the predicted perceived entity , Represents a virtual drone Entities with predictions The distance between Represents individual drones and the perceived entity The safe distance between To avoid collision distance, the perceived entity Includes obstacles and other drones that are being pursued ; The state probability distribution of the target density node is expressed as: in, Represents a virtual drone The relative target density value is used to estimate the spatial distribution of surrounding targets. Represents individual drones Expected density value relative to the target; The state probability distribution of the neighbor density node is expressed as: in, Represents a virtual drone Relative density values ​​relative to neighbors, used to estimate individual drones R i The spatial distribution of neighbors, Represents individual drones The expected density value relative to the neighbors, Indicates the flag bit, used to indicate neighbors; The state probability distribution of the attack situation node is expressed as: in, Represents individual drones The target of the capture, Represents a virtual drone With the target of the capture The distance between.

7. The method for cooperative multi-target capture by drone swarm based on probabilistic graphical model according to claim 6 is characterized in that: In step S5, in the step of selecting the action that maximizes the probability of the drone individual reaching the desired node state value as the output, the joint probability of each model node is calculated, and according to the desired node state under the different roundup stages, the maximum value of the joint probability that each model node reaches the desired node state is obtained to obtain the action that maximizes the probability; wherein the joint probability is expressed as: in, Represents a child node At a given parent node The conditional probability under the state, that is, the node probability parameter, n Indicates the number of child nodes.

Images