Air-ground cooperative path planning method for multi-type target reconnaissance

Abstract

The invention discloses an air-ground cooperative path planning method for multi-type target reconnaissance, and the method comprises the steps: obtaining the position information of a base and all targets, carrying out the clustering, obtaining a plurality of target sub-regions, taking each clustering center as a virtual point corresponding to the target sub-region, and enabling each target sub-region to have at least one target; generating a ground vehicle path based on the coordinates of the base and each virtual point, and optimizing the ground vehicle path; feature points of various types of reconnaissance targets are obtained, unmanned aerial vehicle paths of various target sub-regions are obtained based on improved adaptive large neighborhood search, each unmanned aerial vehicle path takes a virtual point of the target sub-region as a starting point and an ending point of the unmanned aerial vehicle path, and the feature points of the targets are connected; and integrating the ground vehicle path and the unmanned aerial vehicle path of each target sub-region to complete air-ground cooperative path planning. The method is applied to the field of path planning, full-coverage reconnaissance of various types of targets such as points, lines and planes can be efficiently completed, and the charging frequency of the unmanned aerial vehicle is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of path planning, and in particular to an air-ground collaborative path planning method for multi-type target reconnaissance. Background Art

[0002] In recent years, unmanned aerial vehicles (UAVs) have become a vital tool in reconnaissance due to their rapid deployment, high maneuverability, and ability to operate in complex and hazardous environments. Equipped with advanced sensors, UAVs can capture real-time data such as images, videos, and thermal maps, playing a key role in disaster relief, infrastructure inspection, and environmental monitoring. However, in scenarios requiring large-scale, multi-target reconnaissance, such as disaster relief, the limited flight range of civilian UAVs leads to frequent refueling, significantly lengthening mission cycles and reducing operational efficiency. To overcome these limitations, collaborative operations between ground vehicles and UAVs have shown significant advantages. In this mode, a ground vehicle acts as a mobile base station, carrying a UAV close to the target area and then releasing it to perform the mission. This allows the UAV's energy to be focused on reconnaissance operations, significantly improving energy efficiency. Furthermore, the vehicle can replace or recharge the UAV's battery while maneuvering, complementing the UAV's rapid coverage capabilities.

[0003] Although the advantages of collaboration between ground vehicles and drones are significant, their path planning faces new challenges. First, due to the limitation of battery life, drones need to frequently return to vehicles to recharge. Second, to ensure a safe landing, vehicles must reach the predetermined rendezvous point before the battery runs out. This dual time and space collaboration requirement greatly increases the complexity of the problem. In addition, existing research results mainly focus on the problem of air-ground collaborative path planning in logistics and distribution scenarios, and rarely pay attention to target reconnaissance scenarios. In target reconnaissance scenarios, the collaborative reconnaissance of multiple types of targets is often involved. For example, in earthquake rescue, drones need to simultaneously reconnaissance point targets (damaged buildings), line targets (roads and bridges), and surface targets (survivor search and rescue areas), while all targets in logistics and distribution can be regarded as point targets. Line targets need to consider the flight direction, while surface targets allow drones to enter from any boundary point, and the shortest scanning path under the constraint of limited battery life needs to be considered, making this type of problem more complex. Summary of the Invention

[0004] In response to the above-mentioned deficiencies in the existing technology, the present invention provides an air-ground collaborative path planning method for multi-type target reconnaissance, which can efficiently complete full-coverage reconnaissance of multi-type targets such as points, lines, and surfaces, and reduce the number of times the drone needs to be charged.

[0005] To achieve the above object, the present invention provides an air-ground collaborative path planning method for multi-type target reconnaissance, comprising the following steps:

[0006] Step 1: Model multiple types of targets and the reconnaissance process, define the feature points corresponding to point targets, line targets, and area targets, and build an air-ground collaborative path planning model to minimize the total driving distance of the air-ground collaborative system;

[0007] Step 2: Obtain the location information of the base and all targets, and cluster them based on the center coordinates of all targets to obtain several target sub-regions. The center of each cluster is used as the virtual point corresponding to the target sub-region, wherein each target sub-region has at least one target.

[0008] Step 3, generating and optimizing a ground vehicle path based on the coordinates of the base and each of the virtual points;

[0009] Step 4: Obtain the drone paths of each target sub-area based on the improved adaptive large neighborhood search, wherein each drone path uses the virtual point of the target sub-area as the starting point and end point of the drone path and connects the characteristic points of each target;

[0010] Step 5: Integrate the ground vehicle path and the UAV path of each target sub-area to complete the air-ground collaborative path planning for multi-type target reconnaissance.

[0011] Compared with the prior art, the present invention has the following beneficial technical effects:

[0012] 1. This invention takes target type differences into account during the UAV-ground vehicle collaborative reconnaissance path planning process. Compared with traditional target reconnaissance methods that use only UAVs, the assistance of ground vehicles can help UAVs more efficiently complete full-coverage reconnaissance of multiple target types, such as points, lines, and surfaces, while reducing the number of UAV recharges.

[0013] 2. The present invention decomposes air-ground collaborative path planning into two parts: ground vehicle path planning and UAV trajectory planning, and uses clustering combined with heuristic operators to realize ground vehicle path planning. At the same time, an improved adaptive large neighborhood search algorithm is adopted for human-machine trajectory planning. It first obtains a high-quality initial solution through a population-based initial solution generation mechanism, and uses solution destruction and repair operators that consider multi-type target characteristics for iterative optimization, and allows a multi-level acceptance criterion for occasional solution degradation to escape local optimality, which has good robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0015] Figure 1 Flowchart of an air-ground collaborative path planning method for multi-type target reconnaissance according to an embodiment of the present invention;

[0016] Figure 2 Schematic diagram of the GU-MTR problem in an embodiment of the present invention;

[0017] Figure 3 Schematic diagram of a point target in an embodiment of the present invention;

[0018] Figure 4 Schematic diagram of a midline target according to an embodiment of the present invention;

[0019] Figure 5 Schematic diagram of same-side point pairs of a surface target according to an embodiment of the present invention, where: Figure 5 (a) is a schematic diagram of the first combination of import and export points on the same side. Figure 5 (b) is a schematic diagram of the second combination of import and export points on the same side. Figure 5 (c) is a schematic diagram of the third combination of entry and exit points on the same side. Figure 5 (d) is a schematic diagram of the fourth combination of entry and exit points for the same-side point pair;

[0020] Figure 6 Schematic diagram of diagonal point pairs of a surface target in an embodiment of the present invention, where: Figure 6 (a) is a schematic diagram of the first combination of import and export points for diagonal point pairs. Figure 6 (b) is a schematic diagram of the second combination of import and export points for diagonal point pairs. Figure 6 (c) is a schematic diagram of the third combination of import and export points of diagonal point pairs. Figure 6 (d) is a schematic diagram of the fourth combination of entry and exit points for diagonal point pairs;

[0021] Figure 7 A schematic diagram illustrating the generation of a UAV path within a target sub-area according to an embodiment of the present invention;

[0022] Figure 8 Schematic diagram of four types of destruction operators in an embodiment of the present invention;

[0023] Figure 9 Schematic diagram of three repair operators in an embodiment of the present invention;

[0024] Figure 10 Schematic diagram of convergence curves of Examples A26 and A29 in the embodiment of the present invention, wherein: Figure 10 (a) is the convergence curve diagram of Example A26, Figure 10 (b) is a convergence curve diagram of Example A29;

[0025] Figure 11Schematic diagram of algorithm running time of Examples A1-A30 in an embodiment of the present invention;

[0026] Figure 12 This is a diagram showing the algorithm differences in an embodiment of the present invention, where: Figure 12 (a) is the algorithm Gap curve diagram based on examples A19-A30, Figure 12 (b) is the result of 30 runs of the algorithm for example A25;

[0027] Figure 13 This is a schematic diagram showing the impact of the target ratio on the problem cost in an embodiment of the present invention, where: Figure 13 (a) is the driving distance curve of problem instances B1-B7 and C1-C7. Figure 13 (b) Schematic diagram of the number of drones required for problem instances B1-B7 and C1-C7;

[0028] Figure 14 Schematic diagram of the effect of δ on the KDHA method in an embodiment of the present invention, wherein: Figure 14 (a) is a schematic diagram of the impact of Example A5, Figure 14 (b) is a schematic diagram showing the impact of Example A30;

[0029] Figure 15 Schematic diagram of the effect of the initial solution set size on the KDHA method in an embodiment of the present invention, wherein: Figure 15 (a) is a schematic diagram of the impact of Example A12, Figure 15 (b) is a schematic diagram of the impact of Example A27.

[0030] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0032] In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the fact that ordinary technicians in this field can implement it. When the combination of technical solutions is mutually contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0033] like Figure 1 The present embodiment discloses an air-ground collaborative path planning method for multi-type target reconnaissance, which mainly includes the following steps:

[0034] Step 1: Model multiple types of targets and the reconnaissance process, define feature points corresponding to point targets, line targets, and area targets, and build an air-ground collaborative path planning model to minimize the total driving distance of the air-ground collaborative system. Target types include point targets, line targets, and area targets.

[0035] Step 2: Obtain the location information of the base and all targets, and cluster them based on the center coordinates of all targets to obtain several target sub-regions. The center of each cluster is used as the virtual point corresponding to the target sub-region, where each target sub-region has at least one target.

[0036] Step 3: Generate and optimize the ground vehicle path based on the coordinates of the base and each virtual point;

[0037] Step 4: Obtain the drone paths of each target sub-area based on the improved adaptive large neighborhood search, wherein each drone path uses the virtual point of the target sub-area as the starting point and end point of the drone path and connects the characteristic points of each target;

[0038] Step 5: Integrate the ground vehicle path and the UAV path of each target sub-area to complete the air-ground collaborative path planning for multi-type target reconnaissance.

[0039] In the air-ground collaborative path planning problem for multi-type target reconnaissance (GU-MTR problem), a ground vehicle carries a swarm of drones to perform reconnaissance missions for three types of targets: point, line, and surface. The ground vehicle and the drone depart from the same base and return to the base after completing the reconnaissance of all targets. Unlike existing studies that require ground vehicles to release drones at the mission point, the method in this embodiment allows drones to be deployed from virtual launch points (which do not need to coincide with the target point). Its rationality is based on two points: First, line targets and surface targets themselves do not have discrete launch positions, and it is difficult and unnecessary to forcibly define launch points. Setting virtual launch points in adjacent multi-target areas can improve operational efficiency and reduce costs. Second, in actual operations, some point targets may not be directly reached by ground vehicles. The setting of virtual points can enhance the flexibility of drone deployment.

[0040] During the reconnaissance process, when the ground vehicle arrives at the virtual point, each drone is launched from the ground vehicle in turn and visits the adjacent targets along the optimized coverage path to conduct reconnaissance, ensuring that each target is fully surveyed and visited only once. When the drone completes all reconnaissance missions within the endurance range, it will be recovered and recharged by the ground vehicle at the same virtual point. Each drone path includes an inter-target path (the sequence of visited targets) and an intra-target path (the coverage trajectory of line / surface targets). Figure 2 is a schematic diagram of the GU-MTR problem in this embodiment, Figure 2Center: Red dots represent virtual points, solid lines represent ground vehicle routes, and dashed lines represent drone flight paths. Entry points (feature points) of line / surface targets are marked with triangles, while corresponding exit points (feature points) are marked with diagonal vertices or square symbols.

[0041] In addition, to facilitate modeling, this embodiment also makes the following reasonable assumptions: the drones are homogeneous and equipped with the same sensors; the drones fly at a constant speed, and the energy consumption of take-off and landing can be ignored; the ground vehicles travel at a constant speed and have a cruising range sufficient to support the scale of the operation; the risk of mid-air collision when releasing / recovering multiple drones at the same time can be ignored.

[0042] In this embodiment, the modeling of multiple types of targets and the detection process is specifically as follows:

[0043] refer to Figure 3 For point targets: if the UAV is hovering directly above the center point of the target and the entire target is within the UAV's FOV (sensor field of view), the target is defined as a point target; the detection process of a point target is when the UAV arrives directly above the point target, where the position information of the point target is the coordinates of its center point.

[0044] refer to Figure 4 For line targets: If the target's length exceeds the drone's FOV but its width is within it, the target is defined as a line target. The detection process for a line target is to fly along its long side from the center point of its short side to the center point of its opposite short side. The two short side centers are the characteristic points of the line target, and the length of the long side (i.e., the distance between the two characteristic points) is the reconnaissance distance. Therefore, the drone's entry and exit points must be determined for line targets to determine the reconnaissance direction. The position of a line target is the coordinates of the midpoint of the line connecting its two characteristic points.

[0045] For surface targets: If the length and width of the target are beyond the FOV range of the drone, the target is defined as a surface target. To facilitate the calculation of the reconnaissance path, the surface target can be approximated as a rectangle. To reduce the energy consumption of frequent turning, a round-trip scanning trajectory along the long side of the rectangle is used for reconnaissance, where s represents the entry point, e represents the exit point, and the length of the dotted line represents the reconnaissance distance. According to the length of the short side, two types of entry and exit point pairs can be obtained, which are divided into same-side point pairs ( Figure 5 shown) and diagonal pairs ( Figure 6 As shown in Figure 2 ). Based on the four vertices of the surface target, four reconnaissance schemes (i.e., four combinations of entry and exit points) can be derived for each case. Although these schemes have the same coverage trajectory, they will result in differences in the transfer distances between consecutive reconnaissance targets, thus affecting the vehicle's total range. Therefore, for surface target reconnaissance, it is necessary to determine a reasonable drone entry point, which will also determine the corresponding exit point. Among them, the four points on the short side of the rectangle of the surface target with a distance of FOV from the four vertices are the feature points of the surface target, and the position information of the surface target is the coordinates of the center of its rectangle.

[0046] Specifically, each target can be represented as a tuple {Type t ,Point t}, where Type t =0,1,2 represent point target, line target, and surface target respectively. t The feature point set representing the target. For a point target, the feature point is its center point (x, y). For a line target, the center points of the two short sides are the two feature points, i.e. (x w ,y w ) and (x e ,y e For a surface target, the four feature points are the four points on the short side with a distance of FOV from the four vertices, that is, (x wn ,y wn ),(x ws ,y ws ),(x es ,y es ),(x en ,y en ). Therefore, Point t It can be expressed as:

[0047]

[0048] The length of the drone's reconnaissance path for a surface target for:

[0049]

[0050] L=max(||(x wn -x ws ),(y wn -y ws )||2,||(x en -x wn ),(y en -y wn )||2)

[0051] W=min(||(x wn -x ws ),(y wn -y ws )||2,||(x en -x wn ),(y en -y wn )||2)

[0052] Where d is the radius of the UAV sensor's field of view, L is the length of the long side of the surface target, and W is the length of the short side of the surface target.

[0053] In this embodiment, the total travel distance of the air-ground collaborative system includes: the transfer distance between the ground vehicle and the base, the flight distance between the virtual point and the feature point, and the coverage and reconnaissance distance of the drone to the opposite target. That is, the air-ground collaborative path planning model is:

[0054]

[0055] Among them, f is the objective function, V P is a set of virtual points and bases, is the distance traveled by the ground vehicle from point m to point n, K is the number of all available UAV flights in the feasible solution, is the set of feature points in the kth UAV path, is the range of the UAV from point i to point j, is the target set visited by the k-th UAV, is the length of the UAV's coverage reconnaissance path for the surface target t; x mn is a 0-1 variable. When the ground vehicle travels from point m to point n, x mn =1, otherwise x mn =0;y ij is a 0-1 variable. When the drone flies from point i to point j, y ij =1, otherwise y ij =0.

[0056] In the specific implementation process, the constraints of the air-ground collaborative path planning model include:

[0057] The constraints used to limit the reconnaissance range of a single UAV to no more than its maximum range are:

[0058]

[0059] Where E is the maximum range of a single UAV;

[0060] The constraint that requires the ground vehicle to visit each virtual point exactly once is:

[0061]

[0062] The constraints used to ensure that each feature point is visited only once by one drone are:

[0063]

[0064] The constraints used to ensure that each reconnaissance target must be visited by the drone exactly once are:

[0065]

[0066] Among them, T is the target set, z ij is a 0-1 variable. When the drone flies from target i to target j, z ij =1, otherwise z ij = 0; the constraint for limiting the ground vehicle path to start and end at the base is:

[0067]

[0068] Among them, V is the set of feature points, virtual points, and bases;

[0069] The constraint that requires each drone’s flight path to start from a virtual point is:

[0070]

[0071] The constraints used to maintain the flow balance of each UAV at the virtual point are:

[0072]

[0073] The constraints used to prohibit drones from flying directly between the entry and exit points of the area target are:

[0074]

[0075] in, For the target t∈T A ∪T L The set of entry points, For the target t∈T A ∪T L The set of exit points

[0076] The constraint used to enforce that the drone route must contain the connecting segment between two feature points of the line target is:

[0077]

[0078] Among them, T A is the surface target set, T L A collection of line targets.

[0079] In the ground vehicle path planning stage of this embodiment, first, the positions of the virtual points launched by the unmanned aerial vehicle (UAV) are determined, and then the optimized ground vehicle routes traversing these virtual points and the base are calculated. Specifically, the virtual points are generated by K-means clustering, and its key parameters include the distance from the target point to the base and the maximum flight range of the UAV. The obtained clustering centers are used as the virtual points. This method can effectively locate the midpoint near multiple targets, thereby maximizing the energy utilization rate of the UAV. Then, an initial ground vehicle route is generated by randomly arranging the order of the virtual points and the base, and heuristic operators such as 2-opt and 3-opt are used for optimization. The specific implementation process includes the following steps:

[0080] Step 301: Taking the base as the starting and ending points of the ground vehicle, and taking each virtual point as the passing point of the ground vehicle, randomly generate an initial vehicle path v and its path length len;

[0081] Step 302: Obtain a random variable rand1 between 0 and 1, and determine whether rand1 < 0.5 holds:

[0082] If so, update the vehicle path v based on the 2-opt operator to obtain the current vehicle path v1 and its path length len1. In the 2-opt operator, randomly select two nodes and reverse the sub-path between them to obtain the path v1;

[0083] Otherwise, update the vehicle path v based on the 3-opt operator to obtain the current vehicle path v1 and its path length len1. In the 3-opt operator, by removing the edges of three non-adjacent nodes, evaluate all possible reconnection schemes, and adopt the configuration with the minimum cost to form the vehicle path v1;

[0084] Step 303: Determine whether len1 < len holds:

[0085] If so, after setting v = v1 and len = len1, perform Step ۳۰۴;

[0086] ]>Otherwise, perform Step ۳۰۴;

[0087] Step 304: Determine whether the iteration termination condition is satisfied, for example, determine whether the maximum number of iterations is reached: '

[0088] If so, take the current vehicle path v as the ground vehicle path and output it;

[0089] Otherwise, return to Step ۳۰۲.

[0090] During the UAV path planning phase for each target sub-area, the goal is to obtain multi-UAV coordinated routes within each sub-area divided by virtual points. To this end, this embodiment proposes an improved adaptive large neighborhood search method, the core of which includes a population-based initial solution strategy, a problem-specific destruction-repair operator, and a multi-level acceptance criterion. In the specific implementation process, the UAV paths for each target sub-area obtained based on the adaptive large neighborhood search specifically include:

[0091] Step 401: For the current target sub-area, generate the initial solution set R of its drone path. The initial solution set R contains N R solutions, each solution contains a path sequence s, a drone allocation u and a cost c;

[0092] Step 402: Select the solution with the minimum cost from the initial solution set R as the current solution x0, and initialize the global optimal solution x best is the current solution, that is, x best =x0, and let the iteration parameter I = 1;

[0093] Step 403: Select the removal operator and repair operator with the largest weight in the removal operator set and the repair operator set according to the roulette algorithm, and adjust the path sequence s0 in the current solution x0 according to the selected removal operator and repair operator to obtain a new solution x0. i ;

[0094] Step 404, determine the new solution x i Cost x i Is it less than the cost c0 of the current solution x0?

[0095] If so, let x0 = x i Then, proceed to step 405;

[0096] Otherwise, proceed to step 406;

[0097] Step 405: Is the cost c0 of the current solution x0 less than the optimal solution x best The cost of c best :

[0098] If so, let x best = x0, proceed to step 407;

[0099] Otherwise, proceed to step 407;

[0100] Step 406: Determine the new solution x i and the optimal solution x best Is the deviation δ within the tolerance range? For example, the new solution x i and the optimal solution x best The cost difference is less than the set threshold:

[0101] If so, let x0 = xi Then, proceed to step 407;

[0102] Otherwise, proceed to step 407;

[0103] Step 407: Determine whether the iteration termination condition is met:

[0104] If so, output the current optimal solution x best The drone path as the current target sub-area;

[0105] Otherwise, after updating the weights of the removal operator and the repair operator, the process returns to step 403 .

[0106] In the process of solving the UAV path, the quality of the initial solution has an important influence on the convergence speed, so its generation method is crucial. Currently, typical methods for generating initial solutions can be divided into two categories: constructive methods and iterative methods. The constructive method can quickly generate solutions but may sacrifice quality, while the iterative method improves the quality of the solution through repeated optimization, but it increases the calculation time. In order to achieve an effective balance, this embodiment adopts a population-based initial solution strategy to efficiently generate each candidate solution in the population through a specific construction method, and acceptable quality can be obtained without multiple iterations. In addition, unlike the traditional construction method that produces the same solution each time, the method of this embodiment integrates random factors to generate a diverse set of initial solutions, and then selects the optimal solution from them as the initial solution. This strategy avoids the defects of limited iterative optimization while maintaining high-quality solutions.

[0107] refer to Figure 7 , assuming that the target set contains targets 1 to 8, where 3, 5, 6, and 7 are point targets, 2 and 8 are line targets, and 1 and 4 are surface targets. The characteristic points of the line targets are a and b, and the characteristic points of the surface targets are a, b, c, and d. For a solution in the initial solution set R, the generation process is:

[0108] Randomly generate an ordered sequence of all targets in the current target sub-region (e.g. Figure 7 [3, 1, 2, 4, 6, 5, 8, 7]), visit the target in the ordered sequence;

[0109] Identify the target type and determine its feature points: For line targets, randomly select one from two possible direction point pairs, for example, select [a, b] instead of [b, a], with a as the entry point; for area targets, randomly select one from four possible entry and exit point combinations ([a, c], [c, a], [b, d], [d, b]), for example, select [a, c], with a as the entry point and c as the exit point;

[0110] Check whether the path remains feasible after adding the target: If it exceeds the maximum range of the drone, terminate the current path at the virtual point (for example, the path of drone 1 is constructed as 0→3→1a~1c→2a→2b→0, which means starting from the virtual point, reconnoitering targets 3, 1, and 2 in sequence, entering target 1 from point a and leaving from point c, entering reconnoitering from point a of target 2, leaving from point b, and finally returning to the virtual point), and construct a new path with the current target as the first target visited after the virtual point; otherwise, add the target to the current path;

[0111] The above process is iterated until all targets are added to the valid UAV paths, and a solution of the initial solution set R is obtained.

[0112] In this embodiment, the removal operator set includes:

[0113] Massive destruction operator: randomly select two paths in the path sequence s, delete all targets in them and add them to the unassigned target pool;

[0114] Random destruction operator: traverses all paths in the path sequence s, randomly removing one target each time, until the specified number of targets are removed to the unassigned target pool;

[0115] Association destruction operator: First, randomly select a target from the path sequence s and remove it. Then, calculate the association between the remaining targets and the removed targets, sort them by association, and remove the targets with the strongest association in turn, until the specified number of targets are removed to the unassigned target pool.

[0116] Maximum cost destruction operator: When traversing the path sequence s, remove each target in turn and calculate the change in travel distance after removal. Select the target that reduces the travel distance the most and remove it. Repeat this process until the specified number of targets are removed to the unassigned target pool.

[0117] Assume that the random destruction operator, the associated destruction operator, and the maximum cost destruction operator need to remove four targets. Figure 8 Schematic diagrams of four types of destruction operators are shown.

[0118] In this embodiment, the repair operator set includes:

[0119] Greedy repair operator: For each target in the unassigned target pool, insert it into every possible position of the path, record the position that results in the smallest increase in travel distance, and then insert the target at the best position determined by the greedy algorithm;

[0120] Random repair operator: randomly selects a target from the pool of unassigned targets and inserts it into any available position that satisfies the constraints;

[0121] Sequential repair operator: sorts the targets in the unassigned target pool according to the order in which they were added to the pool, and processes each target in sequence, i.e., it traverses each position of the existing drone path, and once a position that meets the constraints is found, the target is directly inserted into that position;

[0122] Greedy repair operator, random repair operator and sequential repair operator are used to restore the path destroyed by random destruction operator. The repair process is as follows: Figure 9 shown.

[0123] In the process of drone path planning, this embodiment periodically updates the weights of the destruction-repair operator pairs based on an adaptive mechanism. When applied, the selection of the destruction-repair operator pairs adopts a roulette mechanism, and the probability of their selection is proportional to the current weight of each operator pair. Since the effectiveness of operator pairs in different cluster sub-regions varies, the weight update is determined by the global solution evaluation (rather than local performance). This periodic weight adjustment can effectively avoid the selection bias caused by over-emphasizing the superior performance of specific operator pairs in individual sub-regions. Therefore, in step 407 of this embodiment, the weights of the removal operator and the repair operator are updated as follows:

[0124] ω=(1-N ω )ω0+(N ω ·s ω ) / (V ω ·u ω )

[0125] Among them, ω is the weight after the removal operator or repair operator is updated, ω0 is the weight before the removal operator or repair operator is updated, N ω is the reaction factor used to control the weight update speed (N ω The larger the value, the faster the update response). ω is a normalization factor used to adjust the range of rewards and punishments to keep them within a reasonable range. This parameter is a global configuration parameter; ω The number of times the operator is removed or repaired, s ω is the score of removing or repairing the operator.

[0126] At the beginning of the iteration, the weights and scores of each removal operator and each repair operator are the same. During each iteration, the scores are increased step by step based on the performance of the removal operator and / or repair operator, and the decision on whether to accept the new solution is made as follows:

[0127] If the new solution obtained after destruction / repair is better than the current optimal solution, the corresponding removal operator and repair operator score σ best , and accept the new solution as the current solution and the optimal solution;

[0128] If the new solution obtained after destruction / repair is inferior to the current optimal solution, but better than the current solution, then the corresponding removal operator and repair operator score σ improve , and accept the new solution as the current solution;

[0129] If the new solution obtained after destruction / repair is worse than the current solution, but the degree of worseness than the current optimal solution does not exceed the tolerance threshold δ, then the corresponding removal operator and repair operator score σ accept-1 , and accept the new solution as the current solution;

[0130] If the new solution obtained after destruction / repair is inferior to the current solution and has not been updated to the current solution, the corresponding removal operator and repair operator score σ accept-2 ;

[0131] Among them, σ best >σ improve >σ accept-1 >σ accept-2 .

[0132] The following further illustrates the air-ground collaborative path planning method for multi-type target reconnaissance (hereinafter referred to as KDHA) in this embodiment with reference to specific examples.

[0133] First, the effectiveness of the proposed method is verified by comparing the target reconnaissance results with and without the two-stage strategy. Then, based on a carefully designed benchmark test case, a comparative experiment is conducted with four representative methods to demonstrate the superiority of KDHA. Finally, the key parameters (δ and N) are respectively R ) and problem characteristic parameters (proportion of line targets and proportion of area targets) were used for sensitivity testing.

[0134] To address the lack of readily available benchmark examples for the GU-MTR problem, we constructed 30 multi-type target reconnaissance instances, A1-A30, based on the CVRP problem A series dataset from the Solomon benchmark library. As shown in Table 1, these instances have varying spatial distributions and heterogeneous numbers of targets (points, lines, and surfaces). All randomly generated line and surface targets meet the maximum flight range constraint of a single UAV (100 units). Each instance is configured with one ground vehicle and five UAVs for target reconnaissance.

[0135] To further study the impact of problem feature parameters, two additional sets of instances are generated based on the A-n80-k10 instance of the Solomon dataset (Table 2):

[0136] B1-B7: used to study the impact of line targets. The number of area targets was fixed at 15, and the proportion of line targets was increased from 10% to 70% (in steps of 10%). The rest were point targets.

[0137] C1-C7: used to analyze the role of area targets. The number of line targets is fixed at 15, and the proportion of area targets is increased from 10% to 70% (step size 10%). The rest are point targets.

[0138] In order to determine the robustness of KDHA, sensitivity analysis was performed on two key parameters:

[0139] δ: tested in the range [0,50] (step size 10) on instances A5 / A30;

[0140] N R : Evaluated on instances A12 / A27 for the range [20,100] (step 20).

[0141] Table 1. Problem benchmark examples

[0142]

[0143]

[0144] Table 2 Sensitivity analysis examples

[0145]

[0146] To validate the effectiveness of KDHA's population-based initial solution generation, problem-oriented damage repair operator, and multi-level acceptance criteria, four heuristic algorithms were adaptively improved based on the problem characteristics as comparative approaches for UAV path planning. These included: Adaptive Large Neighborhood Search (ALNS), Hybrid Genetic Simulated Annealing (HGSA), Adaptive Hybrid Particle Swarm Optimization with Local Search (AHPSO), and Adaptive Differential Evolution (DEA). These four algorithms, named KH-ALNS, KH-HGSA, KH-AHPSO, and KH-DEA, all integrated ground vehicle path heuristics for target clustering and path generation. The parameters of all the compared algorithms were fully optimized through preliminary experiments.

[0147] The specific parameter settings are as follows: the reaction factor of KDHA and KH-ALNS is 1, the normalization factor is 0.5, the population size of the other three methods (KH-HGSA / KH-AHPSO / KH-DEA) and the initial solution set size of KDHA are all set to 100, the crossover rate and mutation rate of KH-HGSA are 0.6 and 0.2, respectively, KH-DEA adopts a random crossover rate with a mean of 0.75, and the mutation rate is adaptively adjusted with the number of iterations, the inertia weight of KH-AHPSO is 0.8, the cognitive factor and social factor are both 2, and the maximum particle speed is 2.

[0148] To verify the effectiveness and necessity of a two-stage approach for solving the GU-MTR problem, a comparative experiment was conducted between two-stage and non-two-stage approaches based on Case A8. The drone-only model is unable to reconnaissance some distant targets, and each drone only covers a very small number of targets per round trip, resulting in extremely low reconnaissance efficiency. In contrast, the air-ground collaborative model effectively addresses these issues, optimizing drone resource utilization while significantly improving reconnaissance efficiency.

[0149] All methods were run independently 30 times, with 200 iterations per run. The average driving distance (Cost) and computation time for each problem instance were compared and analyzed. The best results are highlighted in bold in Table 3.

[0150] Table 3 Comparison results of KDHA with other algorithms based on examples A1-A30

[0151]

[0152] The results in Table 3 show that the computational time of KDHA and KH-ALNS is significantly lower than that of the other three methods. In particular, KDHA can continuously obtain the optimal or suboptimal solution (such as A2, A12, A14 and A18 instances) in a short time and produce the optimal driving distance results in 20 problem instances (especially the large target set scenario), while KH-HGSA performs better in other cases. These results can be seen from Figure 10 Typical example convergence curve and Figure 11 Verification of the algorithm runtime for the A1-A30 examples.

[0153] For scenarios with a small number of targets (such as A3 and A5), KH-HGSA quickly generates diversified solutions by using cross-mutation operations in a limited solution space, and effectively escapes from the local optimum in combination with the SA mechanism. At this time, the limited number of targets leads to a significant reduction in the target density of each sub-region after clustering, resulting in fewer drone paths generated in the second stage, making it difficult for the adaptive operator that depends on the complexity of the solution space to play a role. The mechanism advantages of KDHA and KH-ALNS are greatly weakened. On the contrary, in complex scenarios with many targets, such as Figure 10 As shown by the convergence curves A26 and A29 in Figure 3, KDHA and KH-ALNS, through their adaptive mechanisms, can better utilize information from the search process, dynamically adjust operator probabilities, and appropriately accept suboptimal solutions to avoid local optima. The population-based initialization strategy enables KDHA to obtain better initial solutions than KH-ALNS. Furthermore, the infeasible solution repair mechanism enhances the ability to explore the solution space. This explains the generally superior performance of KDHA over KH-ALNS in Table 3.

[0154] From the perspective of execution time, the adaptive mechanism of KDHA and KH-ALNS reduces the time spent on invalid or inefficient operations, allowing the algorithm to focus on potential search directions more quickly and thus shorten the execution time. Since only a single initial solution is generated and no infeasible solution repair mechanism is adopted, KH-ALNS has a shorter running time than KDHA. KH-HGSA, KH-AHPSO and KH-DEA need to optimize all population individuals in each iteration, which usually takes more time. In addition, Figure 11 It can be seen that as the number of targets increases and the complexity of distribution increases, the difficulty of problem solving increases, resulting in an overall upward trend in the running time of all algorithms.

[0155] In order to compare the performance of KDHA with KH-ALNS, KH-HGSA, KH-AHPSO and KH-DEA, 12 examples (A19-A30) were selected for analysis from the two dimensions of the gap value and coefficient of variation (CV). The gap value describes the difference between the KDHA and other algorithms using the KDHA as the benchmark algorithm, and the coefficient of variation reflects the robustness of the algorithm by the ratio of the standard deviation of the driving distance of each example to the average value. Table 4 and Figure 12 (a) shows the performance gap between KDHA and other algorithms in terms of driving distance. The results clearly show that KDHA produces solutions that significantly outperform KH-HGSA, KH-AHPSO, and KH-DEA, with only KH-ALNS occasionally approaching their performance. This advantage stems from several key factors: Unlike KH-HGSA, KH-AHPSO, and KH-DEA, which rely on multiple parameters that are difficult to optimize for complex multi-target reconnaissance scenarios, KDHA possesses a more robust adaptive search mechanism. Furthermore, while these algorithms are limited by local search efficiency (especially in multi-target structured path optimization), KDHA and KH-ALNS extensively search the solution space to adjust the neighborhood structure through an effective destroy-repair operation. This process not only improves the overall target sequence but also optimizes the entry and exit point selection for linear / surface targets. Furthermore, KDHA's infeasible solution repair mechanism ensures a more effective balance between global exploration and local exploitation, further enhancing its optimization capabilities.

[0156] Table 4 clearly shows that in most instances, KDHA's CV value for driving distance is significantly lower than that of other algorithms, confirming its superior performance in stability, robustness, and reliability. To provide a more intuitive comparison of performance, the results of 30 independent runs are shown using the A25 instance as a representative example: Figure 12(b) shows that KDHA and KH-ALNS have the smallest performance fluctuations and the most stable travel distances. KH-HGSA exhibits moderate performance with significant data dispersion, while KH-DEA exhibits the largest distance fluctuations and the worst performance. It is noteworthy that KDHA not only consistently maintains the shortest travel distance but also achieves the highest stability, a clear advantage over the other methods. KH-DEA and KH-AHPSO, on the other hand, consistently produce higher distance values ​​with greater fluctuations.

[0157] Table 5 Comparison of Gap value and CV of KDHA and other algorithms

[0158]

[0159] For each problem instance, the KDHA algorithm is run 30 times, and the maximum number of drones actually used in each cluster sub-region during the target reconnaissance process is used as the number of drones required to complete the full coverage reconnaissance of the instance. Figure 13 The trend of driving distance and number of drones changing with target ratio is shown. Figure 13 A notable phenomenon can be observed: the driving distance and drone requirements for cases C1-C7 are significantly higher than those for cases B1-B7 (except C1), and the cost growth trend is even more pronounced than for cases B1-B7. This phenomenon is primarily attributed to the higher proportion of area targets in C1-C7, which significantly increases the driving distance and drone requirements for target coverage reconnaissance. In contrast, the change in the proportion of line targets in B1-B7 has a relatively small impact on drone requirements, as the coverage driving distance is equivalent to the distance between two point targets. This difference clearly demonstrates that the geometric characteristics of area targets require more resource investment to ensure full coverage reconnaissance. Furthermore, it was found that the number of drones remained constant when the proportion of the two types of targets increased from 30% to 70%, demonstrating that the two-stage approach effectively optimizes drone resource utilization in actual reconnaissance.

[0160] from Figure 14 It can be observed that the impact of the tolerance threshold δ on the KDHA algorithm varies significantly with problem size: for small-scale problems with few targets, increasing the tolerance threshold has limited improvement in solution quality, but when it exceeds 30, computation time increases significantly. However, for large-scale problems with many targets, a higher tolerance threshold not only significantly improves solution quality but also unexpectedly reduces computation time. Therefore, the tolerance threshold setting needs to be adjusted based on the problem size. For small-scale problems, a conservative value is recommended to maintain computational efficiency, while for large-scale problems, a relaxed tolerance threshold is suitable to achieve both improved solution quality and computational performance.

[0161] from Figure 15As can be seen, when the initial solution set size increases to 100, the driving distance for different problem instances only shows a limited decrease. However, for instances with a large number of targets, the KDHA algorithm exhibits a clear trade-off between computation time and performance. While solution quality may improve as the initial solution set size increases, computation time increases significantly. This characteristic is particularly pronounced in complex, large-scale instances.

[0162] The above description is only a preferred embodiment of the present invention and does not limit the scope of protection of the present invention. All equivalent structural transformations made by using the contents of the present invention description and drawings under the inventive concept of the present invention, or direct / indirect application in other related technical fields are included in the scope of protection of the present invention.

Claims

1. A method for air-ground collaborative path planning for multi-type target reconnaissance, characterized by: The steps include: Step 1: Model multiple types of targets and the reconnaissance process, define the feature points corresponding to point targets, line targets, and area targets, and build an air-ground collaborative path planning model to minimize the total driving distance of the air-ground collaborative system; Step 2: Obtain the location information of the base and all targets, and cluster them based on the center coordinates of all targets to obtain several target sub-regions. The center of each cluster is used as the virtual point corresponding to the target sub-region, wherein each target sub-region has at least one target. Step 3: generating and optimizing a ground vehicle path based on the coordinates of the base and each of the virtual points; Step 4: Obtain the drone paths of each target sub-area based on the improved adaptive large neighborhood search, wherein each drone path uses the virtual point of the target sub-area as the starting point and end point of the drone path and connects the characteristic points of each target; Step 5: Integrate the ground vehicle path and the UAV path of each target sub-area to complete the air-ground collaborative path planning for multi-type target reconnaissance.

2. The air-ground collaborative path planning method for multi-type target reconnaissance according to claim 1 is characterized in that: In step 1, the modeling of multiple types of targets and the detection process is specifically as follows: For point targets: If the UAV is hovering directly above the target's center point and the entire target is within the UAV's FOV, the target is defined as a point target. The detection process for a point target is when the UAV arrives directly above the point target, where the point target's location information is the coordinates of its center point. For line targets: If the target length exceeds the drone's FOV but the width is within the FOV, the target is defined as a line target. The detection process for line targets is to fly along the long side from the center point of the short side to the center point of the opposite short side. That is, the two short side centers are the characteristic points of the line target, where the position information of the line target is the coordinates of the midpoint of the line connecting the two characteristic points. For area targets: If the length and width of the target are both outside the FOV of the drone, the target is defined as an area target. The detection process of an area target is to scan along the long side of the rectangle of the area target. That is, the four points on the short side of the rectangle of the area target that are at a distance of FOV from the four vertices are the feature points of the area target. The position information of the area target is the coordinates of the center of its rectangle. Therefore, each target can be represented as a tuple {Type t ,Point t }, where Type t =0,1,2 represent point target, line target, and surface target respectively. t A set of feature points representing the target.

3. The air-ground collaborative path planning method for multi-type target reconnaissance according to claim 2 is characterized in that: In step 1, the total travel distance of the air-ground collaborative system includes the transfer distance between the ground vehicle and the base, the flight distance between the virtual point and the feature point, and the coverage and reconnaissance distance of the drone to the target. That is, the air-ground collaborative path planning model is: Where f is the objective function, V P is a set of virtual points and bases, is the distance traveled by the ground vehicle from point m to point n, K is the number of all available UAV flights in the feasible solution, is the set of feature points in the kth UAV path, is the range of the UAV from point i to point j, is the target set visited by the k-th UAV, is the length of the UAV’s coverage reconnaissance path for the surface target t; x mn is a 0-1 variable. When the ground vehicle travels from point m to point n, x mn =1, otherwise x mn =0; y ij is a 0-1 variable. When the drone flies from point i to point j, y ij =1, otherwise y ij =0.

4. The air-ground collaborative path planning method for multi-type target reconnaissance according to claim 1, 2 or 3, characterized in that: In step 3, the process of generating and optimizing the ground vehicle path is as follows: Step 301: Using the base as the starting point and end point of the ground vehicle and each virtual point as the waypoint of the ground vehicle, an initial vehicle path v and its path length len are randomly generated; Step 302: Get a random variable rand1 between 0 and 1, and determine whether rand1 < 0.5 holds: If so, the vehicle path v is updated based on the 2-opt operator to obtain the current vehicle path v1 and its path length len1; Otherwise, the vehicle path v is updated based on the 3-opt operator to obtain the current vehicle path v1 and its path length len1; Step 303, determine whether len1 < len holds: If so, after setting v = v1 and len = len1, proceed to Step 304; Otherwise, proceed to Step 304; Step 304, determine whether the iteration termination condition is satisfied: If so, take the current vehicle path v as the ground vehicle path and output it; Otherwise, return to Step 302.

5. The air-ground collaborative path planning method for multi-type target reconnaissance according to claim 1, 2 or 3, characterized in that: In Step 4, the specific process of obtaining the UAV paths of each target sub-region based on adaptive large neighborhood search includes: Step 401: For the current target sub-area, generate the initial solution set R of its drone path. The initial solution set R includes N R solutions, each solution contains a path sequence s, a drone allocation u, and a cost c; Step 402: Select the solution with the minimum cost from the initial solution set R as the current solution x0, and initialize the global optimal solution x best is the current solution, that is, x best =x0, and let the iteration parameter I = 1; Step 403: Select the removal operator and repair operator with the largest weight in the removal operator set and the repair operator set according to the roulette algorithm, and adjust the path sequence s0 in the current solution x0 according to the selected removal operator and repair operator to obtain a new solution x0. i ; Step 404, determine the new solution x i Cost x i Is it less than the cost c0 of the current solution x0? If so, let x0 = x i Then, proceed to step 405; Otherwise, proceed to Step 406; Step 405: Is the cost c0 of the current solution x0 less than the optimal solution x best The cost of c best : If so, let x best = x0, proceed to step 407; Otherwise, proceed to Step 407; Step 406: Determine the new solution x i and the optimal solution x best Is the deviation δ within the tolerance range? If so, let x0 = x i Then, proceed to step 407; Otherwise, proceed to Step 407; Step 407, determine whether the iteration termination condition is satisfied: If so, output the current optimal solution x best The drone path for the currently mentioned target sub-area; Otherwise, update the weights of the removal operator and the repair operator, and then return to Step 403.

6. The air-ground collaborative path planning method for multi-type target reconnaissance according to claim 5 is characterized in that: In Step 401, the generation process of a solution in the initial solution set R is as follows: Randomly generate an ordered sequence of all targets in the current target sub-region, and visit the targets in order of the ordered sequence; Identify the target type and determine its feature points: for line targets, randomly select one from two possible direction point pairs; for surface targets, randomly select one from four possible combinations of entry and exit point pairs; Check whether the path remains feasible after adding the target: if it exceeds the maximum flight range of the UAV, terminate the current path at a virtual point, and construct a new path with the current target as the first target to be visited after the virtual point; Otherwise, add the target to the current path; Iterate the above process until all targets are added to the effective UAV path to obtain a solution in the initial solution set R.

7. The air-ground collaborative path planning method for multi-type target reconnaissance according to claim 5 is characterized in that: The removal operator set includes: Large-scale destruction operator: Randomly select two paths in the path sequence s, delete all targets in them and add them to the unassigned target pool; Random destruction operator: Traverse all paths in the path sequence s, randomly remove one target each time until a specified number of targets are removed and added to the unassigned target pool; Associated destruction operator: First randomly select a target from the path sequence s and remove it, then calculate the association degree between the remaining targets and the removed target, sort them by the association degree, and then remove the most strongly associated targets in turn until a specified number of targets are removed and added to the unassigned target pool; Maximum cost destruction operator: When traversing the path sequence s, remove each target in turn and calculate the change in the travel distance after removal, select the target that causes the maximum reduction in the travel distance for removal, and repeat this process until a specified number of targets are removed and added to the unassigned target pool.

8. The air-ground collaborative path planning method for multi-type target reconnaissance according to claim 5 is characterized in that: The repair operator set includes: Greedy repair operator: For each target in the unassigned target pool, insert it into every possible position in the path, record the position that causes the minimum increase in the travel distance, and then insert the target at the best position determined by the greedy algorithm; Random repair operator: Randomly select a target from the unassigned target pool and insert it into any available position that satisfies the constraints; Sequential repair operator: Sort the targets in the unassigned target pool according to the order in which they are added to the unassigned target pool, and process each target in order, that is, traverse each position of the existing UAV path, and once a position that satisfies the constraints is found, directly insert the target into that position.

9. The air-ground collaborative path planning method for multi-type target reconnaissance according to claim 5 is characterized in that: In Step 407, the specific process of updating the weights of the removal operator and the repair operator is as follows: ω=(1-N ω )ω0+(N ω ·s ω ) / (V ω ·u ω ) Among them, ω is the weight after the removal operator or repair operator is updated, ω0 is the weight before the removal operator or repair operator is updated, N ω is the reaction factor used to control the weight update speed, N ω is the normalization factor, u ω The number of times the operator is removed or repaired, s ω The score for removing or repairing an operator; At the beginning of the iteration, the weights and scores of each removal operator and each repair operator are the same. During each iteration, the scores are increased step by step based on the performance of the removal operator and / or repair operator, and the decision on whether to accept the new solution is made as follows: If the new solution obtained after destruction / repair is better than the current optimal solution, the corresponding removal operator and repair operator score σ best , and accept the new solution as the current solution and the optimal solution; If the new solution obtained after destruction / repair is inferior to the current optimal solution, but better than the current solution, then the corresponding removal operator and repair operator score σ improve , and accept the new solution as the current solution; If the new solution obtained after destruction / repair is worse than the current solution, but the degree of worseness than the current optimal solution does not exceed the tolerance threshold δ, then the corresponding removal operator and repair operator score σ accept-1 , and accept the new solution as the current solution; If the new solution obtained after destruction / repair is inferior to the current solution and has not been updated to the current solution, the corresponding removal operator and repair operator score σ accept-2 ; Among them, p best >s improve >s accept-1 >s accept-2 。

Images