A collaborative decision-making method for drone swarms based on multi-dimensional decision fusion

Abstract

The present invention discloses a multi-dimensional decision-making fusion UAV cluster collaborative decision-making method, the method comprising: step 1: dividing the forest scene into regions, each region is assigned a detection and attack integrated UAV group for inspection; step 2: using the detection and attack integrated UAV group for inspection, and equipped with two external sensors, camera and laser radar, for detection; step 3: based on the target contour and position information, the detection information of different sensors of the UAV is registered, and then based on the registered detection data, the detection information of multiple UAVs is registered to construct a cluster information package; step 4: the detection information collected by different UAVs and different sensors is fused at different levels, and finally the information for identifying the target is obtained; step 5: establishing a UAV decision model and making a decision. The present invention makes decisions by fusing the information data collected by multiple agents in different time and space at different levels, and completes the task more efficiently and accurately.

Description

Technical Field

[0001] The invention relates to a multi-dimensional decision-fusion unmanned aerial vehicle (UAV) cluster collaborative decision-making method, belonging to the technical field of decision-fusion and UAV collaboration. Background Art

[0002] In recent years, drone swarms have developed rapidly. They have been deployed and applied in daily life and on military battlefields. Drone swarms use multiple drones to perform tasks together and achieve goals, which greatly improves efficiency. However, existing drone swarms have obvious defects and deficiencies in performing tasks in complex scenarios. For target detection in complex scenarios, first of all, the changes in lighting conditions, obstructions and climate conditions in complex environments make target detection tasks extremely difficult. Drone swarms need to be able to accurately identify targets in the case of insufficient light, obstructions, and shaking perspectives. However, existing target detection algorithms are often not robust enough to these complex situations, resulting in inaccurate detection results or even missed detections. Secondly, the dynamic changes and mobility of targets increase the difficulty of target detection. In dynamic scenes, targets may move or change direction at any time, and drone swarms are required to have real-time update detection strategies to ensure that targets are continuously and effectively monitored. Finally, the perspective problem of multiple drones also has a certain impact on target detection. Since different drones are in different positions, the perspectives of observing targets are also different, which may lead to differences in the information collected by drones, thereby affecting the collaborative execution efficiency of target detection tasks.

[0003] In addition, for collaborative execution of tasks, existing technologies make it difficult to quickly adjust the cluster's action strategy due to the dynamic nature and quantity changes of targets, and thus cannot ensure more accurate tracking and attacking of targets. Summary of the invention

[0004] The purpose of the present invention is to address the defects and shortcomings of the above-mentioned prior art and propose a drone cluster collaborative decision-making method with multi-dimensional decision fusion. By fusing the information collected by each drone in different time and space at different levels, the drone cluster can make consistent decisions to cope with dynamic changes in complex scenarios.

[0005] The technical solution adopted by the present invention to solve the technical problem is: a multi-dimensional decision-making fusion UAV cluster collaborative decision-making method, the method comprising the following steps:

[0006] Step 1: Divide the forest scene into regions. The drone cluster is a reconnaissance and strike integrated drone cluster. Due to the concealment of the target, each region is assigned a reconnaissance and strike integrated drone group for patrol;

[0007] Step 2: First, each area is inspected using a reconnaissance and strike integrated drone group. The reconnaissance and strike integrated drone is equipped with two external sensors, a camera and a laser radar, to detect and obtain image data information and laser radar data information;

[0008] Step 3: Based on the target contour and position information, the detection information of different sensors of the UAV is registered, and then the detection information of multiple UAVs is registered based on the registered detection data, and a cluster information package is constructed based on the fused information;

[0009] Step 4: Based on the registered detection information, the detection information collected by different drones and different sensors is fused at different levels to finally obtain the information for identifying the target;

[0010] Step 5: Establish a UAV decision model based on the constructed cluster information package and the reconnaissance and strike UAV's own status information, and make a decision.

[0011] Furthermore, step 2 includes: the targets are divided into two types: threatening targets and non-threatening targets. After the drone cluster searches for the target, the image data collected by the drone camera is processed, and the information G i as follows:

[0012] G i = {G i (Tgt j )},j=1,2,L,n

[0013] in,

[0014]

[0015] Among them, i represents the serial number of the drone, G i (Tgt j ) represents the distance between drone i and target Tgt j Image recognition information, For the i-th UAV image target recognition target Tgt j Category, The i-th UAV identifies the target Tgt through image target recognition j The confidence level of identifying a target as a threat or a non-threat target, For the i-th UAV image target recognition target Tgt j The profile information, UAV i uses image target recognition to identify the target Tgt j The confidence level of identifying a threatening target or a non-threatening target, where n is the number of targets in the field of view;

[0016] Similarly, after the point cloud data obtained by the drone laser radar scanning is processed, the information contained in L i as follows:

[0017] L i ={L i (Tgt j )},j=1,2,L,n

[0018] in,

[0019]

[0020] Among them, i represents the serial number of the drone, L i (Tgt j ) represents the distance between drone i and target Tgt j Point cloud recognition information, Point cloud target recognition target Tgt for the i-th UAV j Category, The i-th UAV identifies the target Tgt through point cloud target recognition j The confidence level of identifying a target as a threat or a non-threat target, Point cloud target recognition target Tgt for the i-th UAV j The profile information, UAV i uses point cloud target recognition to identify the target Tgt j The confidence level of identifying a threatening target or a non-threatening target, where n is the number of targets in the field of view.

[0021] Furthermore, in step 3, each contour vector in the image recognition information is traversed, and the vectors with the smallest Euclidean distance to each contour vector in the point cloud recognition information are taken out to form a vector pair:

[0022]

[0023] in, is the jth contour vector of the UAV i image recognition target, is the kth contour vector of the target identified by the drone i point cloud, and n is the number of targets identified by the drone;

[0024] The image information corresponding to the vectors that make up the vector pair and the point cloud information form an information pair, and the information package of the drone i after self-registration is obtained.

[0025]

[0026] in,

[0027]

[0028] in, Identify target Tgt for UAV i j Information, G i (Tgt j )、L i (Tgt j ) are respectively the image recognition targets Tgt of UAV i j Information and drone i point cloud recognition target Tgt j information, n is the number of targets identified by the drone;

[0029] For multiple drones, obtain the location information of each drone's own identified target and compare it with the location information of the other drones' identified targets, and select the closest location to form a location pair:

[0030]

[0031] Among them, Pos i (Tgt j ) identifies the target Tgt for UAV i j Location information, Pos k (Tgt l ) is the target Tgt for UAV k l The location information of the target, j and l are the target serial numbers, and n is the number of targets identified by the drone;

[0032] The targets corresponding to the position information of the position pair are regarded as the same target, and the target registration is finally completed, and the information package dp after multi-UAV target registration is obtained cluster :

[0033]

[0034] Among them, m is the number of drones in the drone cluster.

[0035] Furthermore, in step 4, due to the changes in terrain, vegetation type and tree height in the forest environment, the vegetation coverage density presents differentiated distribution characteristics. This difference directly affects the penetration and distribution of light, resulting in significant differences in light intensity inside the forest. The light intensity can be described as follows:

[0036]

[0037] Among them, I i is the light intensity at the location of UAV i, I 0 is the light intensity when it is not blocked, k I is the attenuation coefficient, which indicates the attenuation degree of light intensity per unit vegetation coverage density, k I >0,λ(xi ,y i ) is the vegetation coverage at the location of UAV i;

[0038] Similarly, due to the different vegetation coverage at different locations in the forest, the ambient humidity at different locations is different, which can be expressed as follows:

[0039] H i =H 0 ×(1+k H ×λ(x i ,y i ))

[0040] Among them, H i is the ambient humidity of the location of UAV i, H 0 is the average humidity of the environment, k H is the humidity quantification coefficient, which is used to quantify the impact of vegetation coverage on humidity, k H >0,λ(x i ,y i ) is the vegetation coverage at the location of UAV i;

[0041] Vegetation coverage also affects wind speed. The effect of vegetation coverage on wind speed can be expressed as follows:

[0042]

[0043] Among them, V i is the wind speed at the location of UAV i, V 0 is the wind speed when there is no obstruction, k V is the wind speed quantification coefficient, which is used to quantify the impact of vegetation coverage on wind speed, k V >0,λ(x i ,y i ) is the vegetation coverage at the location of UAV i;

[0044] In summary, due to the influence of light intensity, ambient humidity and wind speed, the Tgt of a target in the field of view j For example, based on the recognition information after registration, UAV i will identify the target Tgt j The confidence levels of targets identified as threatening and non-threatening targets are:

[0045]

[0046] conf i,nthr (Tgt j )=1-conf i,thr (Tgt j )

[0047] Among them, confi,thr (Tgt j ) is the target Tgt for drone i j The comprehensive confidence level of the target identified as a threat, conf i,nthr (Tgt j ) is the target Tgt for drone i j The overall confidence level of a target identified as non-threatening, I i is the light intensity at the location of UAV i, I 0 is the light intensity when it is not blocked, H i is the humidity at the location of drone i, H 0 is the average humidity of the environment, H max is the maximum humidity of the environment, V i is the wind speed at the location of UAV i, V c is the wind speed threshold for interfering with drones, Identify the target Tgt through the image for the drone i j is the confidence level of the threat target, Identify target Tgt for UAV i point cloud j The confidence level of the threat target;

[0048] Since the relative positions of each UAV and the target are different, the observation factor VA of each UAV is first calculated i :

[0049]

[0050] Among them, μ i , μ(Tgt j ) are respectively UAV i and target Tgt j The unit direction vector, σ is the standard deviation of the control attenuation, λ(x i ,y i ), λ max (x i ,y i ) are the vegetation coverage and the maximum vegetation coverage of the location of UAV i, respectively, α VA , β VA It is used to adjust the effect of viewing angle and field of view occlusion on the observation factors of each drone, and has:

[0051] α V +β V =1

[0052] Based on the observation factors of each drone, the Tgt of each drone to the target j The recognition confidence of each UAV is processed to obtain the confidence conf used for fusion decision making. i, ' thr (Tgt j):

[0053]

[0054] Among them, conf i, ' thr (Tgt j ) is the target Tgt considered by drone i after processing j There is a threatening confidence;

[0055] Obtain the recognition confidence of each drone after processing. According to the confidence, divide the targets into consensus clusters and non-consensus clusters. The serial numbers of the drones in the cluster are as follows:

[0056]

[0057] UID con UUID ncon ={1,2,L,m}

[0058] in,

[0059]

[0060] Among them, UID con 、UID ncon are the numbers of the drones in the consensus cluster and the non-consensus cluster, respectively. m is the number of drones in the drone cluster. c represents the possible value of confidence. ε c is the length of the confidence interval, and:

[0061]

[0062] Tgt of the target for the drone in the cluster j The confidence of target recognition is integrated as follows:

[0063]

[0064] in,

[0065]

[0066] α conf +β conf =1,α conf <β conf

[0067] in, Indicates that cluster target recognition considers the target Tgt j is the confidence level of the threat target, conf i, ' thr (Tgt j) is the comprehensive confidence of UAV i identifying the target as a threatening target, m represents the number of UAVs constituting the cluster, and m con is the number of drones in the consensus cluster, c represents the possible value of confidence, α conf , β conf is the weight coefficient, which is used to adjust the ratio of the front and back parts, τ min , τ max They are the segmentation thresholds of the multi-UAV fusion confidence formula;

[0068] Obtain the cluster's identification attributes of the target and combine them with the surrounding environmental factors to establish the threshold conf of the target posing an actual threat thr :

[0069] conf thr =g 1 (n)×g 2 (φ j ,a j )×conf 0

[0070] in,

[0071]

[0072] Among them, conf 0 is the basic confidence threshold, g 1 (n) is the target quantity impact factor, k n is the target number quantification coefficient, which is used to quantify the impact of the number of targets on the threat confidence level. n is the number of targets, n c is the target quantity impact threshold, g 2 (φ j ,a j ) is the target trajectory function, φ j Used to measure target Tgt j The direction of travel, a j Target Tgt j The acceleration, ε a is the target acceleration threshold, k a is the target acceleration quantized value, and 0<k a <1,dis cluster (Tgt j |ts k )、dis cluster (Tgt j |ts k+1 ) represent the target Tgt j With cluster in ts k Moment and ts k+1 The distance of time;

[0073] According to the threshold constructed above, the target category and confidence level determined after the group fusion decision can be finally obtained:

[0074]

[0075] Among them, cls cluster (Tgt j ) is the cluster consensus determination target Tgt j The type of target, thr, nthr respectively indicates the target type is a threatening target and a non-threatening target;

[0076] The same steps are performed on each target in the field of view, and finally the categories and corresponding confidences of all targets are obtained.

[0077] Furthermore, in step 5, based on the information obtained by fusion in step 4, the target group is first divided into two parts, a threat target group and a non-threat target group:

[0078] Tgt thr ={Tgt j |cls cluster (Tgt j )=thr},j=1,2,Ln 1

[0079] Tgt nthr ={Tgt l |cls cluster (Tgt l )=nthr},l=1,2,Ln 2

[0080] The quantitative relationship is as follows:

[0081] n 1 +n 2 =n

[0082] Among them, Tgt thr , Tgt nthr are threat target group and non-threat target group respectively, n 1 、n 2 , n are the number of threat targets, non-threat targets, and all targets respectively;

[0083] The drone decision model is established from the following two aspects:

[0084] Support phase: The drone cluster first performs a self-check to compare the number of drones in the cluster with the number of targets. If the following conditions are met:

[0085] m<n 1

[0086] That is, if the number of drones is less than the number of threatening targets in the target group, a coordination request is sent to the remaining drones. When drone i receives the request, it obtains the two-dimensional position of each drone cluster that issued the coordination request. The cluster position is defined as follows:

[0087]

[0088] Among them are:

[0089]

[0090] in, is the two-dimensional position of the k-th drone cluster, is the horizontal coordinate of the kth UAV cluster, is the ordinate of the kth drone cluster, m is the number of drones in the cluster, x l ,y l They represent the horizontal and vertical coordinates of each drone in the cluster, K is the number of drone clusters that issue coordination requests, and m is the number of drones in the drone cluster;

[0091] Drone i decides whether to go to the collaboration through its own decision-making. First, it reads its own image information G i With point cloud information L i , we can get the following profit function and its constraints:

[0092]

[0093] in,

[0094] ω 1 +ω 2 +ω 3 =1

[0095]

[0096] Among them, d i,k is the distance between UAV i and cluster k, d max is the maximum distance among all the drones in the cluster that issued the request received by drone i, ρ i,k is the tree density in the direction of the line connecting the position of UAV i and the position of cluster k, ρ max is the maximum value of tree density in the connection direction of all the drone clusters that issued requests received by drone i, is the number of threatening targets among the targets found by drone cluster k, ω 1 ,ω 2 ,ω 3 is the influencing factor, which represents the importance of each part and is used to adjust the proportion of each part. Contains the types of all targets detected by drone i through image detection and point cloud detection respectively;

[0097] After the above calculations, each drone obtains the benefits of assisting each drone cluster. After comparison, the drone cluster that assists to obtain the maximum benefit is selected first. When the drone cluster that issues the request is included in mn 1 After the drone is launched, no assistance request will be issued;

[0098] Tracking / strike phase: For the drone swarm, its action space is as follows:

[0099] A cluster ={Attack,Track}

[0100] Among them, A cluster Actions that can be performed by drone clusters, including attack and tracking;

[0101] The following decision-making methods are used for the drone swarm after the support replenishment is completed:

[0102]

[0103] Among them are:

[0104]

[0105] Among them, C is the comprehensive score, which is used to decide the action, M(Tgt j ) is a moving target counting function, which is used to determine whether the target is moving. Pos(Tgt j |ts k )、Pos(Tgt j |ts k+1 ) respectively represent in ts k ,ts k+1 Time target Tgt j The position of n is the number of targets, n 1 is the number of threatened targets, A cluster is the cluster action decision, C attack is the scoring threshold and C attack ∈(0,1);

[0106] When the cluster action decision is track, the following decisions are made for the target group:

[0107]

[0108] Among them, UID track The number of the drone that performs the tracking mission. The serial number is UID track The drone, dis i(Tgt j |Tgt j ∈Tgt thr ) is the UAV i and the threat target Tgt j The distance between thr is the threat target group, m is the number of drones in the drone cluster, n 1 is the number of threat targets. The above process assigns a nearest UAV to each threat target to form a body-tracking task pair. Single-unit tracking is performed, and then the drones and targets that have formed a task pair no longer participate in the allocation, and the remaining unassigned drones stay and wait. For the drones that are performing single-unit tracking, the decision formula is continuously used to calculate the comprehensive score and make a decision;

[0109] Similarly, when the cluster action decision is attack, the following decisions are made for the target group:

[0110]

[0111]

[0112] Among them, UID attack The number of the drone that carried out the strike mission, The serial number is UID attack The drone, dis i (Tgt j |Tgt j ∈Tgt thr ) is the UAV i and the threat target Tgt j The distance between thr is the threat target group, m is the number of drones in the drone cluster, n 1 is the number of threat targets. The above process is to assign a UAV closest to each threat target to form a body-strike mission pair. Conduct a single strike, and then the drone and target that have formed a task pair will no longer participate in the allocation;

[0113] If there are any unassigned drones, they will assist the nearest strike drone in attacking according to the principle of proximity until all targets in the area are eliminated.

[0114] Beneficial effects:

[0115] 1. The present invention mainly includes information fusion between different sensors of a single UAV based on environmental factors such as lighting conditions, wind speed, humidity, etc. in forest scenes, and further considers the viewing angle factors of different UAVs to perform multi-UAV information fusion. At the same time, the attributes of the target group are considered to set the detection thresholds under different situations, which can more efficiently and accurately complete the detection of targets in complex scenes.

[0116] 2. The present invention takes into account the impact of target group identification and environmental factors on drone cluster decision-making and establishes a drone cluster decision-making model. In the face of different situations, the drone cluster can quickly make corresponding decisions, ensuring the efficient execution of drone cluster tasks in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0117] Figure 1 This is a flow chart of a multi-dimensional decision-fusion UAV cluster collaborative decision-making method of the present invention.

[0118] Figure 2 This is the information fusion diagram between the drone sensors of the present invention.

[0119] Figure 3 This is an information fusion diagram between drones of the present invention. DETAILED DESCRIPTION

[0120] The invention will be further described in detail below in conjunction with the accompanying drawings.

[0121] like Figure 1 As shown, the present invention proposes a multi-dimensional decision-making fusion UAV cluster collaborative decision-making method, and the specific operation steps of the method are as follows:

[0122] Step 1: Divide the forest scene into regions. The drone cluster is a reconnaissance and strike integrated drone cluster. Due to the concealment of the target, each region is assigned a reconnaissance and strike integrated drone group for patrol;

[0123] Step 2: First, use the integrated reconnaissance and strike drone group to patrol each area. The integrated reconnaissance and strike drone is equipped with two external sensors, camera and laser radar, to detect and obtain its image data information and laser radar data information;

[0124] Targets are divided into two types: threatening targets and non-threatening targets. After the drone cluster searches for the target, it processes the image data collected by the drone camera. The information Gi included is as follows:

[0125] G i = {G i (Tgt j )},j=1,2,L,n

[0126] in,

[0127]

[0128] Among them, i represents the serial number of the drone, G i (Tgt j ) represents the distance between drone i and target Tgt j Image recognition information, For the i-th UAV image target recognition target Tgt j Category, The i-th UAV identifies the target Tgt through image target recognition j The confidence level of identifying a target as a threat or a non-threat target, For the i-th UAV image target recognition target Tgt j The profile information, UAV i uses image target recognition to identify the target Tgt j The confidence level of identifying a threatening target or a non-threatening target, where n is the number of targets in the field of view;

[0129] Similarly, after processing, the point cloud data obtained by the drone laser radar scanning contains the following information Li:

[0130] L i ={L i (Tgt j )},j=1,2,L,n

[0131] in,

[0132]

[0133] Among them, i represents the serial number of the drone, L i (Tgt j ) represents the distance between drone i and target Tgt j Point cloud recognition information, Point cloud target recognition target Tgt for the i-th UAV j Category, The i-th UAV identifies the target Tgt through point cloud target recognition j The confidence level of identifying a target as a threat or a non-threat target, Point cloud target recognition target Tgt for the i-th UAV j The profile information, UAV i uses point cloud target recognition to identify the target Tgt j The confidence level of identifying a threatening target or a non-threatening target, where n is the number of targets in the field of view;

[0134] Step 3: Based on the target contour and position information, the detection information of different sensors of the UAV is registered. Further, the detection information of multiple UAVs is registered based on the registered detection data, and a cluster information package is constructed based on the fused information;

[0135] Traverse each contour vector in the image recognition information, and take out the vector with the smallest Euclidean distance to each contour vector in the point cloud recognition information to form a vector pair:

[0136]

[0137] in, is the jth contour vector of the UAV i image recognition target, is the kth contour vector of the target identified by the drone i point cloud, and n is the number of targets identified by the drone;

[0138] The image information corresponding to the vectors that make up the vector pair and the point cloud information form an information pair, and the information package of the drone i after self-registration is obtained.

[0139]

[0140] in,

[0141]

[0142] in, Identify target Tgt for UAV i j Information, G i (Tgt j )、L i (Tgt j ) are respectively the image recognition targets Tgt of UAV i j Information and drone i point cloud recognition target Tgt j information, n is the number of targets identified by the drone;

[0143] For multiple drones, obtain the location information of each drone's own identified target and compare it with the location information of the other drones' identified targets, and select the closest location to form a location pair:

[0144]

[0145]

[0146] Among them, Pos i (Tgt j ) identifies the target Tgt for UAV i j Location information, Pos k (Tgt l) is the target Tgt for UAV k l The location information of the target, j and l are the target serial numbers, and n is the number of targets identified by the drone;

[0147] The targets corresponding to the position information of the position pair are regarded as the same target, and the target registration is finally completed, and the information package dp after multi-UAV target registration is obtained cluster :

[0148]

[0149] Where m is the number of drones in the drone cluster;

[0150] Step 4: Based on the registered detection information, the detection information collected by different drones and different sensors is fused at different levels to finally obtain the information for identifying the target;

[0151] Due to the changes in terrain, vegetation type and tree height in the forest environment, the vegetation coverage density presents differentiated distribution characteristics. This difference directly affects the penetration and distribution of light, resulting in significant differences in light intensity inside the forest. The light intensity can be described as follows:

[0152]

[0153] Among them, I i is the light intensity at the location of UAV i, I 0 is the light intensity when it is not blocked, k I is the attenuation coefficient, which indicates the attenuation degree of light intensity per unit vegetation coverage density, k I >0,λ(x i ,y i ) is the vegetation coverage at the location of UAV i;

[0154] Similarly, due to the different vegetation coverage at different locations in the forest, the ambient humidity at different locations is different, which can be expressed as follows:

[0155] H i =H 0 ×(1+k H ×λ(x i ,y i ))

[0156] Among them, H i is the ambient humidity of the location of UAV i, H 0 is the average humidity of the environment, k H is the humidity quantification coefficient, which is used to quantify the impact of vegetation coverage on humidity, k H >0,λ(x i ,yi ) is the vegetation coverage at the location of UAV i;

[0157] Vegetation coverage also affects wind speed. The effect of vegetation coverage on wind speed can be expressed as follows:

[0158]

[0159] Among them, V i is the wind speed at the location of UAV i, V 0 is the wind speed when there is no obstruction, k V is the wind speed quantification coefficient, which is used to quantify the impact of vegetation coverage on wind speed, k V >0,λ(x i ,y i ) is the vegetation coverage at the location of UAV i;

[0160] In summary, due to the influence of light intensity, ambient humidity and wind speed, the Tgt of a target in the field of view j For example, based on the recognition information after registration, UAV i will identify the target Tgt j The confidence levels of targets identified as threatening and non-threatening targets are:

[0161]

[0162] conf i,nthr (Tgt j )=1-conf i,thr (Tgt j )

[0163] Among them, conf i,thr (Tgt j ) is the target Tgt for drone i j The comprehensive confidence level of the target identified as a threat, conf i,nthr (Tgt j ) is the target Tgt for drone i j The overall confidence level of a target identified as non-threatening, I i is the light intensity at the location of UAV i, I 0 is the light intensity when it is not blocked, H i is the humidity at the location of drone i, H 0 is the average humidity of the environment, H max is the maximum humidity of the environment, V i is the wind speed at the location of UAV i, V c is the wind speed threshold for interfering with drones, Identify the target Tgt through the image for the drone i j is the confidence level of the threat target, Identify target Tgt for UAV i point cloud j The confidence level of the threat target;

[0164] Since the relative positions of each UAV and the target are different, the observation factor VA of each UAV is first calculated i :

[0165]

[0166] Among them, μ i , μ(Tgt j ) are respectively UAV i and target Tgt j The unit direction vector, σ is the standard deviation of the control attenuation, λ(x i ,y i ), λ max (x i ,y i ) are the vegetation coverage and the maximum vegetation coverage of the location of UAV i, respectively, α VA , β VA It is used to adjust the effect of viewing angle and field of view occlusion on the observation factors of each drone, and has:

[0167] α V +β V =1

[0168] Based on the observation factors of each drone, the Tgt of each drone to the target j The recognition confidence of each UAV is processed to obtain the confidence conf used for fusion decision making. i, ' thr (Tgt j ):

[0169]

[0170] Among them, conf i, ' thr (Tgt j ) is the target Tgt considered by drone i after processing j There is a threatening confidence;

[0171] Obtain the recognition confidence of each drone after processing. According to the confidence, divide the targets into consensus clusters and non-consensus clusters. The serial numbers of the drones in the cluster are as follows:

[0172]

[0173] UID con UUID ncon ={1,2,L,m}

[0174] in,

[0175]

[0176] Among them, UID con 、UID ncon are the numbers of the drones in the consensus cluster and the non-consensus cluster, respectively. m is the number of drones in the drone cluster. c represents the possible value of confidence. ε c is the length of the confidence interval, and:

[0177]

[0178] Tgt of the target for the drone in the cluster j The confidence of target recognition is integrated as follows:

[0179]

[0180] in,

[0181]

[0182] α conf +β conf =1,α conf <β conf

[0183] in, Indicates that cluster target recognition considers the target Tgt j is the confidence level of the threat target, conf i, ' thr (Tgt j ) is the comprehensive confidence of UAV i identifying the target as a threatening target, m represents the number of UAVs constituting the cluster, and m con is the number of drones in the consensus cluster, c represents the possible value of confidence, α conf , β conf is the weight coefficient, which is used to adjust the ratio of the front and back parts, τ min , τ max They are the segmentation thresholds of the multi-UAV fusion confidence formula;

[0184] Obtain the cluster's identification attributes of the target and combine them with the surrounding environmental factors to establish the threshold conf of the target posing an actual threat thr :

[0185] conf thr =g 1 (n)×g 2 (φ j ,a j )×conf 0

[0186] in,

[0187]

[0188] Among them, conf 0 is the basic confidence threshold, g 1 (n) is the target quantity impact factor, k n is the target number quantification coefficient, which is used to quantify the impact of the number of targets on the threat confidence level. n is the number of targets, n c is the target quantity impact threshold, g 2 (φ j ,a j ) is the target trajectory function, φ j Used to measure target Tgt j The direction of travel, a j Target Tgt j The acceleration, ε a is the target acceleration threshold, k a is the target acceleration quantized value, and 0<k a <1,dis cluster (Tgt j |ts k )、dis cluster (Tgt j |ts k+1 ) represent the target Tgt j With cluster in ts k Moment and ts k+1 The distance of time;

[0189] According to the threshold constructed above, the target category and confidence level determined after the group fusion decision can be finally obtained:

[0190]

[0191] Among them, cls cluster (Tgt j ) is the cluster consensus determination target Tgt j The type of target, thr, nthr respectively indicates the target type is a threatening target and a non-threatening target;

[0192] The same steps are performed on each target in the field of view, and finally the categories and corresponding confidences of all targets are obtained.

[0193] Step 5: Establish a UAV decision model based on the constructed cluster information package and the state information of the reconnaissance and attack UAV, and make a decision;

[0194] Based on the information obtained by fusion in step 4 of the present invention, the target group is first divided into two parts, a threat target group and a non-threat target group:

[0195] Tgt thr ={Tgt j |cls cluster (Tgt j )=thr},j=1,2,Ln 1

[0196] Tgt nthr ={Tgt l |cls cluster (Tgt l )=nthr},l=1,2,Ln 2

[0197] The quantitative relationship is as follows:

[0198] n 1 +n 2 =n

[0199] Among them, Tgt thr , Tgt nthr are threat target group and non-threat target group respectively, n 1 、n 2 , n are the number of threat targets, non-threat targets, and all targets respectively;

[0200] Furthermore, the UAV decision model is established from the following two aspects:

[0201] (1) Support phase

[0202] The drone cluster first performs a self-check to compare the number of drones in the cluster with the target number. If:

[0203] m<n 1

[0204] That is, if the number of drones is less than the number of threatening targets in the target group, a coordination request is sent to the remaining drones. When drone i receives the request, it obtains the two-dimensional position of each drone cluster that issued the coordination request. The cluster position is defined as follows:

[0205]

[0206] Among them are:

[0207]

[0208] in, is the two-dimensional position of the k-th drone cluster, is the horizontal coordinate of the kth UAV cluster, is the ordinate of the kth drone cluster, m is the number of drones in the cluster, x l ,y lThey represent the horizontal and vertical coordinates of each drone in the cluster, K is the number of drone clusters that issue coordination requests, and m is the number of drones in the drone cluster;

[0209] Drone i decides whether to go to the collaboration through its own decision-making. First, it reads its own image information G i With point cloud information L i , we can get the following profit function and its constraints:

[0210]

[0211] in,

[0212] ω 1 +ω 2 +ω 3 =1

[0213]

[0214] In the above formula, d i,k is the distance between UAV i and cluster k, d max is the maximum distance among all the drones in the cluster that issued the request received by drone i, ρ i,k is the tree density in the direction of the line connecting the position of UAV i and the position of cluster k, ρ max is the maximum value of tree density in the connection direction of all the drone clusters that issued requests received by drone i, is the number of threatening targets among the targets found by drone cluster k, ω 1 ,ω 2 ,ω 3 is the influencing factor, which represents the importance of each part and is used to adjust the proportion of each part. Contains the types of all targets detected by drone i through image detection and point cloud detection respectively;

[0215] After the above calculations, each drone obtains the benefits of assisting each drone cluster. After comparison, the drone cluster that assists to obtain the maximum benefit is selected first. When the drone cluster that issues the request is included in mn 1 After the drone is launched, no assistance request will be issued;

[0216] (2) Tracking / strike phase

[0217] For a drone swarm, its action space is as follows:

[0218] A cluster ={Attack,Track}

[0219] The above formula shows that A clusterActions that can be performed by drone clusters, including attack and tracking;

[0220] The following decision-making methods are used for the drone swarm after the support replenishment is completed:

[0221]

[0222] Among them are:

[0223]

[0224] Among them, C is the comprehensive score, which is used to decide the action, M(Tgt j ) is a moving target counting function, which is used to determine whether the target is moving. Pos(Tgt j |ts k )、Pos(Tgt j |ts k+1 ) respectively represent in ts k ,ts k+1 Time target Tgt j The position of n is the number of targets, n 1 is the number of threatened targets, A cluster is the cluster action decision, C attack is the scoring threshold and C attack ∈(0,1);

[0225] When the cluster action decision is track, the following decisions are made for the target group:

[0226]

[0227] Among them, UID track The number of the drone that performs the tracking mission. The serial number is UID track The drone, dis i (Tgt j |Tgt j ∈Tgt thr ) is the UAV i and the threat target Tgt j The distance between thr is the threat target group, m is the number of drones in the drone cluster, n 1 is the number of threat targets. The above process assigns a nearest UAV to each threat target to form a body-tracking task pair. Single-unit tracking is performed, and then the drones and targets that have formed a task pair no longer participate in the allocation, and the remaining unassigned drones stay and wait. For the drones that are performing single-unit tracking, the decision formula is continuously used to calculate the comprehensive score and make a decision;

[0228] Similarly, when the cluster action decision is attack, the following decisions are made for the target group:

[0229]

[0230] Among them, UID attack The number of the drone that carried out the strike mission, The serial number is UID attack The drone, dis i (Tgt j |Tgt j ∈Tgt thr ) is the UAV i and the threat target Tgt j The distance between thr is the threat target group, m is the number of drones in the drone cluster, n 1 is the number of threat targets. The above process is to assign a UAV closest to each threat target to form a body-strike mission pair. Conduct a single strike, and then the drone and target that have formed a task pair will no longer participate in the allocation;

[0231] If there are any unassigned drones, they will assist the nearest strike drone in attacking according to the principle of proximity until all targets in the area are eliminated.

[0232] It should be pointed out that the description of the above embodiments is only used to help understand the method of the present application and its core idea. For ordinary technicians in this technical field, several improvements and modifications can be made to the present application without departing from the principles of the present application. These improvements and modifications are also within the scope of protection of the claims of the present application.

Claims

1. A multi-dimensional decision-making fusion method for drone swarm collaborative decision-making, characterized in that: The method comprises the following steps: Step 1: Divide the forest scene into regions. The drone cluster is a reconnaissance and strike integrated drone cluster. Due to the concealment of the target, each region is assigned a reconnaissance and strike integrated drone group for patrol; Step 2: First, each area is inspected using a reconnaissance and strike integrated drone group. The reconnaissance and strike integrated drone is equipped with two external sensors, a camera and a laser radar, to detect and obtain image data information and laser radar data information; Step 3: Based on the target contour and position information, the detection information of different sensors of the drone is registered, and then the detection information of multiple drones is registered based on the registered detection data. Based on the fused information, a cluster information package is constructed, and each contour vector in the image recognition information is traversed. The vector with the smallest Euclidean distance to each contour vector in the image recognition information is taken out to form a vector pair: in, is the jth contour vector of the target in the image recognition of UAV i, is the kth contour vector of the target identified by the drone i point cloud, and n is the number of targets identified by the drone; The image information corresponding to the vectors that make up the vector pair and the point cloud information form an information pair, and the information package of the drone i after self-registration is obtained. in, in, Identify target Tgt for UAV i j Information, G i (Tgt j )、L i (Tgt j ) are respectively the image recognition targets Tgt of UAV i j Information and drone i point cloud recognition target Tgt j information, n is the number of targets identified by the drone; For multiple drones, obtain the location information of each drone's own identified target and compare it with the location information of the other drones' identified targets, and select the closest location to form a location pair: Among them, Pos i (Tgt j ) identifies the target Tgt for UAV i j Location information, Pos k (Tgt l ) is the target Tgt for UAV k l The location information of the target, j and l are the target serial numbers, and n is the number of targets identified by the drone; The targets corresponding to the position information of the position pair are regarded as the same target, and the target registration is finally completed, and the information package dp after multi-UAV target registration is obtained cluster : Where m is the number of drones in the drone cluster; Step 4: Based on the registered detection information, the detection information collected by different drones and different sensors is fused at different levels to finally obtain the information for identifying the target. Due to the changes in terrain, vegetation type and tree height in the forest environment, the vegetation coverage density presents differentiated distribution characteristics. This difference directly affects the penetration and distribution of light, resulting in significant differences in light intensity inside the forest. The light intensity is described as follows: Among them, I i is the light intensity at the location of UAV i, I0 is the light intensity when it is not blocked, k I is the attenuation coefficient, which indicates the attenuation degree of light intensity per unit vegetation coverage density, k I >0,λ(x i ,y i ) is the vegetation coverage at the location of UAV i; Similarly, due to the different vegetation coverage at different locations in the forest, the ambient humidity at different locations is different. The ambient humidity is expressed as follows: H i =H0×(1+k H ×λ(x i ,y i )) Among them, H i is the ambient humidity of the location of UAV i, H0 is the average ambient humidity, k H is the humidity quantification coefficient, which is used to quantify the impact of vegetation coverage on humidity, k H >0,λ(x i ,y i ) is the vegetation coverage at the location of UAV i; Vegetation coverage also affects wind speed. The impact of vegetation coverage on wind speed is shown as follows: Among them, V i is the wind speed at the location of UAV i, V0 is the wind speed when it is not blocked, k V is the wind speed quantification coefficient, which is used to quantify the impact of vegetation coverage on wind speed, k V >0,λ(x i ,y i ) is the vegetation coverage at the location of UAV i; Due to the influence of light intensity, ambient humidity and wind speed, the target Tgt in the field of view j For example, based on the recognition information after registration, UAV i will identify the target Tgt j The confidence levels of targets identified as threatening and non-threatening targets are: conf i,nthr (Tgt j )=1-conf i,thr (Tgt j ) Among them, conf i,thr (Tgt j ) is the target Tgt for drone i j The comprehensive confidence level of the target identified as a threat, conf i,nthr (Tgt j ) is the target Tgt for drone i j The overall confidence level of a target identified as non-threatening, I i is the light intensity at the location of UAV i, I0 is the light intensity when it is not blocked, H i is the humidity at the location of UAV i, H0 is the average humidity of the environment, H max is the maximum humidity of the environment, V i is the wind speed at the location of UAV i, V c is the wind speed threshold for interfering with drones, Identify the target Tgt through the image for the drone i j is the confidence level of the threat target, Identify target Tgt for UAV i point cloud j The confidence level of the threat target; Since the relative positions of each UAV and the target are different, the observation factor VA of each UAV is first calculated i : Among them, μ i , μ(Tgt j ) are respectively UAV i and target Tgt j The unit direction vector, σ is the standard deviation of the control attenuation, λ(x i ,y i ), λ max (x i ,y i ) are the vegetation coverage and the maximum vegetation coverage of the location of UAV i, respectively, α VA , β VA It is used to adjust the effect of viewing angle and field of view occlusion on the observation factors of each drone, and has: α V +β V ＝1 Based on the observation factors of each drone, the Tgt of each drone to the target j The recognition confidence of each UAV is processed to obtain the confidence conf used for fusion decision making i, ' thr (Tgt j ): Among them, conf i, ' thr (Tgt j ) is the target Tgt considered by drone i after processing j There is a threatening confidence; Obtain the recognition confidence of each drone after processing. According to the confidence, divide the targets into consensus clusters and non-consensus clusters. The serial numbers of the drones in the cluster are as follows: ID con ∪UID ncon ={1,2,...,m} in, Among them, UID con 、UID ncon are the numbers of the drones in the consensus cluster and the non-consensus cluster, respectively. m is the number of drones in the drone cluster. c represents the value of confidence. ε c is the length of the confidence interval, and: Tgt of the target for the drone in the cluster j The confidence of target recognition is integrated as follows: in, α conf +β conf ＝1,α conf ＜β conf in, Indicates that cluster target recognition considers the target Tgt j is the confidence of the threat target, conf′ i,thr (Tgt j ) is the comprehensive confidence of UAV i identifying the target as a threatening target, m represents the number of UAVs constituting the cluster, and m con is the number of drones in the consensus cluster, c represents the value of confidence, α conf , β conf is the weight coefficient, which is used to adjust the ratio of the front and back parts, τ min , τ max They are the segmentation thresholds of the multi-UAV fusion confidence formula; Obtain the cluster's identification attributes of the target and combine them with the surrounding environmental factors to establish the threshold conf of the target posing an actual threat thr : conf thr =g1(n)×g2(φ j ,a j )×conf0 in, Among them, conf0 is the basic confidence threshold, g1(n) is the target number influencing factor, k n is the target number quantification coefficient, which is used to quantify the impact of the number of targets on the threat confidence level. n is the number of targets, n c is the target quantity impact threshold, g2(φ j ,a j ) is the target trajectory function, φ j Used to measure target Tgt j The direction of travel, a j Target Tgt j The acceleration, ε a is the target acceleration threshold, k a is the target acceleration quantized value, and 0<k a <1,dis cluster (Tgt j |ts k )、dis cluster (Tgt j |ts k+1 ) represent the target Tgt j With cluster in ts k Moment and ts k+1 The distance of time; According to the threshold constructed above, the target category and confidence level determined after group fusion decision are finally obtained: Among them, cls cluster (Tgt j ) is the cluster consensus determination target Tgt j The type of target, thr, nthr respectively indicates the target type is a threatening target and a non-threatening target; The same steps are performed on each target in the field of view, and finally the categories and corresponding confidences of all targets are obtained; Step 5: Based on the constructed cluster information package and the state information of the integrated reconnaissance and attack UAV, a UAV decision model is established and a decision is made. Based on the information obtained by fusion in the above step 4, the target group is first divided into two parts, the threat target group and the non-threat target group: Tgt thr ={Tgt j |cls cluster (Tgt j )=thr},j=1,2,…n1 Tgt nthr ={Tgt l |cls cluster (Tgt l )=nthr},l=1,2,…n2 The quantitative relationship is as follows: n1+n2=n Among them, Tgt thr , Tgt nthr are threat target group and non-threat target group respectively, n1, n2, and n are the number of threat targets, non-threat targets, and all targets respectively; The drone decision model is established from the following two aspects: Support phase: The drone cluster first performs a self-check to compare the number of drones in the cluster with the number of targets. If the following conditions are met: m<n1 That is, if the number of drones is less than the number of threatening targets in the target group, a coordination request is sent to the remaining drones. When drone i receives the request, it obtains the two-dimensional position of each drone cluster that issued the coordination request. The cluster position is defined as follows: Among them are: in, is the two-dimensional position of the kth drone cluster, is the horizontal coordinate of the kth UAV cluster, is the ordinate of the kth drone cluster, m is the number of drones in the cluster, x l ,y l They represent the horizontal and vertical coordinates of each drone in the cluster, K is the number of drone clusters that issue coordination requests, and m is the number of drones in the drone cluster; Drone i decides whether to go to the collaboration through its own decision-making. First, it reads its own image information G i With point cloud information L i , then the following profit function and its constraints are obtained: in, ω1+ω2+ω3=1 Among them, d i,k is the distance between UAV i and cluster k, d max is the maximum distance among all the drones in the cluster that issued the request received by drone i, ρ i,k is the tree density in the direction of the line connecting the position of UAV i and the position of cluster k, ρ max is the maximum value of tree density in the connection direction of all the drone clusters that issued requests received by drone i, is the number of threatening targets among the targets found by drone cluster k, ω1, ω2, and ω3 are influencing factors, which represent the importance of each part and are used to adjust the proportion of each part. Contains the types of all targets detected by drone i through image detection and point cloud detection respectively; After the above calculations, each drone obtains the benefits of assisting each drone cluster. After comparison, the drone cluster that obtains the greatest benefits is selected first. When the drone cluster that issues the request includes m-n1 drones, it will no longer issue assistance requests. Tracking / strike phase: For the drone swarm, its action space is as follows: A cluster ={Attack,Track} Among them, A cluster Actions that can be performed by drone clusters, including attack and tracking; The following decision-making methods are used for the drone swarm after the support replenishment is completed: Among them are: Among them, C is the comprehensive score, which is used to decide the action, M(Tgt j ) is a moving target counting function, which is used to determine whether the target is moving. Pos(Tgt j |ts k )、Pos(Tgt j |ts k+1 ) respectively represent in ts k ,ts k+1 Time target Tgt j The position of n is the number of targets, n1 is the number of threatening targets, A cluster is the cluster action decision, C attack is the scoring threshold and C attack ∈(0,1); When the cluster action decision is track, the following decisions are made for the target group: Among them, UID track The number of the drone that performs the tracking mission. The serial number is UID track drones, For UAV i and threatening target Tgt j The distance between thr is the threat target group, m is the number of drones in the drone cluster, and n1 is the number of threat targets. The above process assigns a drone closest to each threat target to form a body-tracking task pair. Single-unit tracking is performed, and then the drones and targets that have formed a task pair no longer participate in the allocation, and the remaining unassigned drones stay and wait. For the drones that are performing single-unit tracking, the decision formula is continuously used to calculate the comprehensive score and make a decision; Similarly, when the cluster action decision is attack, the following decisions are made for the target group: Among them, UID attack The number of the drone that carried out the strike mission, The serial number is UID attack The drone, dis i (Tgt j |Tgt j ∈Tgt thr ) is the UAV i and the threat target Tgt j The distance between thr is the threat target group, m is the number of drones in the drone cluster, and n1 is the number of threat targets. The above process assigns a drone closest to each threat target to form a body-strike mission pair. Conduct a single strike, and then the drone and target that have formed a task pair will no longer participate in the allocation; If there are any unassigned drones, they will assist the nearest strike drone in attacking according to the principle of proximity until all targets in the area are eliminated.

2. The multi-dimensional decision-making fusion UAV cluster collaborative decision-making method according to claim 1 is characterized in that: The step 2 includes: the targets are divided into two types: threatening targets and non-threatening targets. After the drone cluster searches for the target, the image data collected by the drone camera is processed, and the information G i as follows: G i ={G i (Tgt j )},j=1,2,…,n in, Among them, i represents the serial number of the drone, G i (Tgt j ) represents the distance between drone i and target Tgt j Image recognition information, For the i-th UAV image target recognition target Tgt j Category, The i-th UAV identifies the target Tgt through image target recognition j The confidence level of identifying a target as a threat or a non-threat target, For the i-th UAV image target recognition target Tgt j The profile information, UAV i uses image target recognition to identify the target Tgt j The confidence level of identifying a threatening target or a non-threatening target, where n is the number of targets in the field of view; Similarly, after the point cloud data obtained by the drone laser radar scanning is processed, the information contained in L i as follows: THE i ={L i (Tgt j )},j=1,2,…,n… in, Among them, i represents the serial number of the drone, L i (Tgt j ) represents the distance between drone i and target Tgt j Point cloud recognition information, Point cloud target recognition target Tgt for the i-th UAV j Category, The i-th UAV identifies the target Tgt through point cloud target recognition j The confidence level of identifying a target as a threat or a non-threat target, Point cloud target recognition target Tgt for the i-th UAV j The profile information, UAV i uses point cloud target recognition to identify the target Tgt j The confidence level of identifying a threatening target or a non-threatening target, where n is the number of targets in the field of view.

Images