Multi-unmanned aerial vehicle cooperative search and coverage optimization method based on improved loyd algorithm

Abstract

The invention discloses a multi-unmanned aerial vehicle cooperative search and coverage optimization method based on an improved Lloyd algorithm, and is suitable for a target search task in a dynamic environment. Firstly, target distribution is perceived and estimated in real time by using a probability hypothesis density filter, and a target probability map is dynamically updated, so that the response capability and search precision of a system to target state change are improved; secondly, a multi-graph fusion mechanism is adopted, a target probability map, an uncertainty map and a search pheromone map are subjected to dynamic weighted integration, priority sharing and updating of key area information are achieved, and the cooperation efficiency and the communication resource utilization rate among multiple unmanned aerial vehicles are improved; and finally, based on an improved Lloyd algorithm, introducing a multi-step prediction mechanism and a direction-guided centroid updating strategy, optimizing search path generation, and preventing the path from being converged to a low-value region. Through the method of the invention, the system can realize adaptive and efficient search and coverage control in a complex dynamic environment, and stability, real-time performance and overall robustness of task execution of multiple unmanned aerial vehicles are significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of multi-agent cooperative control of unmanned aerial vehicles (UAVs) and relates to a coverage optimization method based on multi-step prediction and time utility map, which is suitable for target search and coverage tasks in dynamic environments. Background Technology

[0002] With the rapid development of artificial intelligence, robotics, and unmanned aerial vehicle (UAV) systems, multi-agent collaborative systems have been widely applied in disaster relief, environmental monitoring, and border patrol. Multi-UAV systems, through collaborative work, can significantly improve the efficiency of target search and coverage tasks. However, these systems still face many key technical challenges that need to be addressed in dynamic mission scenarios.

[0003] First, insufficient adaptability to dynamic environments is a major problem currently faced by multi-UAV systems. Traditional distributed control algorithms are typically based on static weights or region allocation, making it difficult to respond quickly to dynamic changes in the target. Especially when target distribution changes frequently or environmental information is uncertain, existing methods often lead to low coverage efficiency, uneven resource allocation, and even excessive overlap or task conflicts between agents. The root cause of these problems is that the Bayesian update mechanism of traditional methods usually assumes that the target state is discrete and invariant, making it difficult to update the target probability in real time in dynamic environments.

[0004] Secondly, the burden of communication bandwidth is also a key factor restricting the performance of multi-agent systems. In multi-UAV collaborative missions, agents need to share mission information and environmental status in real time. Frequent information transmission can lead to communication bandwidth overload, reducing the system's real-time performance and reliability. Existing information transmission mechanisms often use full or fixed-weight information synchronization methods, which cannot effectively filter low-priority or outdated information, further exacerbating the waste of communication resources.

[0005] Furthermore, the global optimization performance of multi-agent systems still needs improvement. Current algorithms are often limited to optimizing local regions and lack a global perspective, making it difficult to balance global efficiency and local accuracy in coverage control. For example, in the traditional Lloyd algorithm, due to the limitations of the centroid adjustment mechanism, agents are prone to centroid oscillations or convergence stagnation in complex regions or under non-uniform weight distributions, reducing the overall task efficiency of the system. Summary of the Invention

[0006] To address the issues of path planning redundancy, insufficient coverage, heavy communication burden, and inadequate global optimization in multi-UAV cooperative control under dynamic environments, this invention provides a coverage optimization method based on dynamic region partitioning and multi-step predictive path planning, specifically as follows: Figure 1As shown, sub-tasks are assigned by dividing regions, and path planning is performed within each region. At the same time, triggering conditions are set to realize the dynamic division of regions.

[0007] The specific technical solution of this invention is as follows:

[0008] A coverage optimization method based on dynamic region partitioning and multi-step predictive path planning includes the following steps:

[0009] Step 1: Improved Bayesian Update Process

[0010] An improved Bayesian update strategy, combined with a probability hypothesis density filter, enables continuous construction and real-time updating of the target probability map. In each mission round, the UAV predicts and corrects the target state based on current sensor observations, updating the target's probability distribution in space accordingly. To enhance the modeling capability for target location uncertainty, a range expansion mechanism is introduced in the update process. This not only updates the target point itself but also performs a weighted expansion update of its surrounding area, ensuring spatial continuity of the target's influence range and improving the map's stability and robustness. Through continuous iterative perception and updating, this method enables the system to have continuous responsiveness, providing reliable probabilistic information support for subsequent path optimization and collaborative search, ensuring the UAV maintains high sensitivity and adaptability to target states in dynamic environments.

[0011] Step 2: Selective data transmission based on time utility map

[0012] To reduce communication overhead, this invention constructs a time utility map and prioritizes information from each region based on real-time assessment. By setting a time utility threshold, data from high-priority regions is filtered for transmission, ensuring priority updates of important information and avoiding the transmission of outdated data. This selective data transmission mechanism can significantly reduce communication bandwidth usage and optimize the real-time performance and resource utilization of unmanned swarm systems.

[0013] Step 3: Dynamic region partitioning based on Venotu-Lloyd

[0014] To achieve a reasonable allocation of task partitioning and search areas among UAVs, this invention employs a dynamic region partitioning strategy based on Venn diagrams and the Lloyd algorithm. Firstly, as... Figure 3 As shown, based on the location of each agent, a Venn diagram is used to divide the search area into multiple smaller regions. Each agent is responsible for its nearest region. Then, a time-triggered signal is waited for, and when the region division requirement is met again, the region is re-divided based on the current location of the drone.

[0015] Step 4: Dynamic Path Optimization Based on Multi-Step Prediction

[0016] After the region is divided, each UAV performs path optimization within its assigned search area. The Lloyd algorithm, based on the Veno diagram, first calculates the weighted centroid of each Veno diagram region, then iteratively moves the UAV to the weighted centroid of its corresponding Veno diagram region, repeatedly updating the region division and centroid position, ultimately achieving uniform spatial segmentation and optimal coverage. In generating paths using weighted centroids, the UAV considers not only the target probability map but also uses an uncertainty map and a pheromone map to calculate the weighted centroid. The aim is to converge the weighted centroid to unsearched areas with high uncertainty and the potential presence of targets. Specifically, the uncertainty map reflects the observation confidence level within the current region, guiding the UAV to prioritize areas with insufficient perception or ambiguous information; the pheromone map records historical search trajectories and reduces the possibility of repeated visits through a time decay mechanism, thus avoiding resource waste. By integrating these three types of map information into a composite weighting basis for path planning, the system can dynamically generate search paths for future multi-step paths, effectively avoiding path redundancy, reducing turning frequency, and improving coverage continuity.

[0017] This invention employs a multi-step prediction mechanism to guide the path gradually towards the high-probability target region, thereby improving search performance and coverage efficiency. Specifically, as follows... Figure 4 As shown, the UAV starts with the current centroid of the Vinograph region, then calculates the first predicted centroid within a circle centered at that centroid and with a radius of r1. Similarly, it continues to calculate the second predicted centroid within a circle centered at the first predicted centroid and with a radius of r2, and so on, allowing for multiple predictions. It's important to note that the prediction radius gradually decreases towards convergence. Furthermore, the total number of prediction steps is continuously adjusted according to the search progress. In the early stages of the search, the number of prediction steps is small, and even additional predicted centroid calculations may not be necessary. In the later stages of the search, as the fragmentation of unsearched regions reduces search efficiency, the number of prediction steps will continuously increase to prevent the centroid from converging into unsearched regions. However, considering computational efficiency, an upper limit on the number of prediction steps is set.

[0018] Step 5: Task Execution and Dynamic Updates

[0019] While the UAVs perform search missions, the system continuously updates the target probability map, uncertainty map, and pheromone map in real time. Information fusion and transmission operate routinely, ensuring timely environmental perception and up-to-date map data. If the Venn map region re-division is not triggered, the system skips the task reassignment process, and each UAV independently performs dynamic path optimization based on multi-step prediction using the latest probability map, continuously adjusting its search direction to approach high-probability areas and improve responsiveness and path foresight. The pheromone map is used to record historical trajectories and decays over time to prevent duplicate overwriting. The entire execution process maintains high adaptability while ensuring coordination and consistency, ensuring that the system continuously maintains task efficiency and global robustness in dynamic environments.

[0020] Furthermore, in step 1, to effectively represent the spatiotemporal distribution of targets in the search space within a dynamic environment, a probabilistic hypothesis density filter is used to model the multi-target state. The probabilistic hypothesis density represents the expected number of targets per unit space; it is not direct density information and cannot be directly used for path planning and region weight allocation. Therefore, to achieve probabilistic map-driven path optimization, after each round of perception, the system normalizes the current probabilistic hypothesis density output in a discrete space, thereby forming a target probability map that can be used for decision-making.

[0021] Furthermore, in step 2, during the distributed information fusion process, a consensus protocol is used to verify the local maps shared by each UAV. When conflicting updates occur in the same area, the system prioritizes retaining data from more recent sources with higher confidence levels, and ensures the stability and consistency of the fused map within its neighborhood through a local consistency update strategy.

[0022] Furthermore, in step 2, the search pheromone map employs a time decay mechanism, meaning that the weight of historical search trajectories gradually decreases over time to prevent the UAV from continuously avoiding already searched areas. This mechanism also retains a moderate reference to historical paths, allowing the UAV to reassess and cover previous paths in cases of dynamic target changes or task reassignment, improving adaptability to dynamic targets and area reuse efficiency. This method effectively balances the comprehensiveness and exploratory nature of the search, avoiding the formation of long-term path blind spots.

[0023] Furthermore, in step 2, a weighted fusion strategy based on a time utility model is introduced during the map fusion process. This model comprehensively considers the observation latency, detection frequency, and update confidence of each region, dynamically calculating the fusion weight of each region in the map. The system can thus adaptively select high-priority regions for map synthesis and updating, ensuring that the target probability map maintains high freshness and usability globally, thereby improving the overall response efficiency of the search task.

[0024] Furthermore, during the multi-UAV collaborative search process described in step 3, each UAV exchanges its locally constructed target probability map through a distributed information sharing mechanism and evaluates the fusion priority of received information in conjunction with a time utility map. Through this mechanism, the system can dynamically schedule communication resources, allowing UAVs to prioritize receiving high-value, high-time-sensitivity information, avoiding communication redundancy and improving overall collaboration efficiency. Based on this, UAVs can continuously update local maps and dynamically adjust search strategies during mission execution, achieving seamless collaboration.

[0025] Furthermore, in the region partitioning process described in step 3, this invention employs a joint partitioning strategy based on Venn diagrams and the Lloyd algorithm. The search space is first generated as a Venn diagram region based on the current location information of each UAV. Then, by introducing target probability density as a region weighting factor, the centroid of the Venn diagram region is iteratively adjusted to achieve dynamic reallocation of the search space. This strategy ensures that high-target-probability regions are prioritized for task allocation, improving resource utilization efficiency, reducing redundant coverage by multiple UAVs in low-target-density regions, and enhancing the overall system performance.

[0026] Furthermore, in the path planning process of step 4, the UAV employs a prediction-enhanced Lloyd algorithm to model and proactively adjust the future target distribution. The prediction mechanism, based on historical probability evolution trends, infers target hotspots over multiple future time steps and dynamically corrects the path direction in conjunction with the current search status.

[0027] Furthermore, in step 4, to avoid frequent sharp turns or unnatural movement in the path, the centroid calculation in the Lloyd algorithm introduces a directional weighting factor: when the UAV calculates the direction of movement to the next centroid, it simultaneously considers the angle weight between its current flight orientation and the expected target direction, thereby smoothing the turning trajectory and improving the executability and continuity of the path.

[0028] Furthermore, to avoid divergent behavior in the multi-step path prediction process in step 4, this invention introduces a path convergence constraint mechanism. Centroid prediction stops and the current path is executed when any of the following conditions are met: 1) If the centroid variation of the path calculated in consecutive steps is lower than a set threshold during the prediction process, the prediction is considered stable, and subsequent steps are terminated, directly executing the current path; otherwise, the prediction radius is dynamically reduced and centroid prediction continues; 2) A predetermined upper limit for the number of prediction steps is reached, where the specific upper limit threshold adapts to different search stages. This mechanism improves the stability and reliability of path planning.

[0029] Furthermore, in the predictive centroid path planning process described in step 4, the target path generated by the UAV is no longer limited to its initially assigned Venn diagram region. Since the path optimization process preferentially guides the UAV towards local or global probability maxima, its final centroid position may cross the initial region boundary and enter the region of an adjacent UAV.

[0030] In summary, this invention effectively improves the global search performance of multi-UAV systems in dynamic environments by combining the improved Lloyd algorithm with multi-step predictive path optimization and integrating multi-source map information for task division and collaborative control, significantly enhancing the success rate and execution efficiency of target search. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of a collaborative search framework based on region division and multi-part predictive path planning.

[0032] Figure 2 This is a flowchart of a distributed collaborative search method for unmanned aerial vehicles (UAVs).

[0033] Figure 3 It is a method for dividing the region of a Venn diagram based on Euclidean distance.

[0034] Figure 4 This is a schematic diagram of a prediction-enhanced Lloyd algorithm. Detailed Implementation

[0035] The present invention will now be described in detail with reference to the accompanying drawings.

[0036] This invention proposes a multi-UAV cooperative search and coverage optimization method based on an improved Lloyd algorithm. It updates the target probability map through a probability hypothesis density filter, and combines information fusion mechanisms and multi-step prediction path optimization to improve the search efficiency and target coverage capability of UAVs in dynamic environments. The flowchart is as follows: Figure 2 As shown, the specific embodiments of the present invention will be described in detail below.

[0037] Step 1: Target probability map update based on probability hypothesis density filter. In multi-target detection tasks, the probability hypothesis density represents the expected distribution of the number of targets in space. To convert it to a standard probability density function, normalization can be performed:

[0038]

[0039] Where v(x) represents the probability hypothesis density value at position x. P PHD (x) represents in the subspace The probability of the in-place target appearing at point x satisfies the normalization constraint:

[0040]

[0041] Search area Divided into sizes N i ×N j Grid G xy Each grid cell g xy Represents a fixed spatial region whose boundary is defined by coordinates [x] i,min ,x i,max ] and [y j,min ,y j,max [Confirmed. Target probability assigned to the nearest grid center:]

[0042]

[0043] Among them, (x c ,y c ) is the grid g xy The center of , ||·||2 represents the Euclidean distance.

[0044] Traditional Bayesian updates assume that the observations are perfectly reliable, but in real-world environments, sensor data contains noise and uncertainty. Therefore, a probability hypothesis density filter is used as a dynamic confidence weighting factor to correct the target probability update.

[0045]

[0046] Among them, P xy (k) represents the prior probability of the target at time k. d For the detection probability, P f P represents the false alarm rate. PHD (x) serves as a weighting factor, improving the dynamic adaptability of probability updates.

[0047] To account for measurement uncertainties, the update range is not limited to the target point, but rather extended based on the covariance matrix of the target location:

[0048]

[0049] Wherein, the covariance matrix of the target position (x0, y0) This reflects uncertainty. The Gaussian weighting ensures that greater update weights are assigned to regions closer to the target point, improving the robustness of detection.

[0050] To control the spatial diffusion range of updates, the system introduces an effective support boundary constraint mechanism, performing update operations only within regions that satisfy the following elliptical constraint conditions:

[0051]

[0052] Here, n is a constant, typically taking the value of 2 or 3, corresponding to a Gaussian distribution confidence range of approximately 95% or 99.7%. This boundary condition effectively limits the spatial propagation range of the update calculation, avoids the meaningless diffusion of probability weights over large-scale regions, and ensures the locality and convergence efficiency of the update calculation.

[0053] Step 2: Information Fusion and Time Utility Map. This invention employs a multi-layered information fusion mechanism, combining a target probability map, an uncertainty map, and a search pheromone map:

[0054] w xy (k)=αP xy (k)+βU xy (k)+γS xy (k)

[0055] Among them, P xy (k) is the target probability map, output by the probability hypothesis density filter. U xy (k) is an uncertainty map, calculated based on Shannon entropy:

[0056] U xy (k)=-P xy (k)log2(P xy (k))-(1-P xy (k))log2(1-P xy (k))S xy (k) is a pheromone map with a decay property over time, and its decay mechanism is as follows: |

[0057] S xy (k+1)=S xy (k)·e -λt

[0058] Where t is the time step and λ is the decay factor, ensuring that historical search information gradually diminishes, enabling the UAV to adapt to dynamically changing environments. The weighting coefficients α, β, and γ of the three types of layers are dynamically set according to the task type, target dynamism, and resource constraints, used to adjust the focus of the fusion map at different stages.

[0059] Step 3: Distributed collaborative search and information sharing. The drones achieve information fusion through consensus:

[0060]

[0061] Among them, w ii (k) and w ij (k) Dynamically adjust based on the time utility map:

[0062]

[0063] Where, τ xy (k)=e -λ(k-t(x,y)) As a time-use decay factor, it ensures that drones use the latest data.

[0064] Step 4: Region Division and Event-Triggered Update Mechanism

[0065] To achieve efficient collaborative search among multiple UAVs in dynamic environments, this invention incorporates a region division stage before path planning to clearly define the task responsibilities of each UAV. The search space among UAVs is divided using a Venn diagram, where each sub-region... For a given UAV i, the region is dynamically determined based on its current position and the target probability distribution. However, frequent redrawing of the Vennogram region significantly increases the computational burden on the system, especially in large-scale multi-UAV systems. Therefore, this invention designs an event-triggered region update mechanism that redraws the Vennogram region only when necessary, thereby reducing computational overhead while maintaining system flexibility.

[0066] This mechanism introduces a trigger function Γ k A comprehensive assessment of the magnitude of current environmental changes and system operational stability is necessary.

[0067]

[0068] in, It represents the global change in the target probability map between two iterations, reflecting changes in target movement or perceived uncertainty; This represents the average offset of all intelligent points, reflecting whether the system has entered a stable state; Γ is the change normalization factor, and ∈ is the centroid shift threshold, which makes the indicators comparable; w1+w2=1 is the dynamic weighting coefficient, used to adjust the relative importance of environmental changes and system stability. k When ≤γ, the system triggers a Vinograph repartition update, where γ is the set trigger threshold. This strategy implements adaptive updates under two key conditions:

[0069] • When the target distribution changes drastically, i.e. When the size is large, the system actively adjusts the region division;

[0070] When the system tends to stabilize, i.e., Δc k When the size is small, it can be replanned in advance.

[0071] This mechanism avoids unnecessary repetitive partitioning calculations and can adaptively adjust regions according to the dynamic nature of tasks, significantly reducing the computational burden while improving the system's response speed to environmental changes.

[0072] Step 5: Improve the path optimization of the Lloyd algorithm. When performing Venn diagram region partitioning and centroid calculation, a composite weight function w′ integrating multi-source information is introduced. xy Used to guide the centroid to shift towards higher value directions:

[0073] w′ xy =w xy ·(1+α dir ·cosθ)

[0074] Among them, w xy The basic weights represent a combination of target probability, uncertainty, and pheromone. The vector pointing from the current position to the target position. With heading The angle between them; Indicates the consistency between the current action direction and the target direction; α dir This is the directional adjustment coefficient, which controls the effect of path orientation on centroid offset. This directional guidance mechanism can suppress frequent turning, ensure path continuity and directional stability, and improve overall motion efficiency.

[0075] Path prediction employs a sliding window iterative Lloyd update method. Let UAV i be currently at time k, and its path planning be based on a predicted depth N. p Generate a set of future centroid sequences of length:

[0076]

[0077] Among them, the The predicted centroid is generated according to the following recursive rule:

[0078]

[0079] Where Γ(c,r) represents the location of the local weighted centroid calculated by combining the weighted map within a radius of c and step size r; For the first The search radius of the step satisfies the decreasing constraint:

[0080]

[0081] Final path based on predicted centroid The ultimate goal is to predict depth N based on the search progress. p Adaptive change, defining the total search space region as The covered area is The non-coverage rate is:

[0082]

[0083] Let N be the maximum number of prediction steps. max The minimum is N min Then the number of prediction steps can be adaptively set as follows:

[0084]

[0085] Where, ρ k ∈[0,1], the proportion of uncovered areas; a larger value indicates an earlier search phase; β>0, adjusts the non-linearity of step growth, ultimately allowing the number of steps to automatically increase in the later stages of the search. It's important to note that when the maximum prediction depth N is reached... p Previously, if the change in the centroid of the path, calculated based on consecutive steps, was lower than a preset threshold... The system will consider the prediction stable, stop predicting further, and execute the current path. This mechanism allows the drone to make large-scale directional predictions in the early stages and automatically refine the path resolution as it approaches the target area, thereby improving positioning accuracy.

Claims

1. A coverage optimization method based on dynamic region partitioning and multi-step predictive path planning, characterized in that... Includes the following steps: Step 1: Improved Bayesian Update Process An improved Bayesian update strategy, combined with a probability hypothesis density filter, enables continuous construction and real-time updating of the target probability map. In each round of mission execution, the UAV predicts and corrects the target state based on current sensor observations, and updates the target's probability distribution in space accordingly. To enhance the modeling ability for target position uncertainty, a range expansion mechanism is introduced in the update process, which not only updates the target point itself but also performs weighted expansion updates on its surrounding area, making the target's influence range spatially continuous and improving the stability and robustness of the map. Through continuous iterative perception and updating, this method enables the system to have continuous response capabilities, providing reliable probabilistic information support for subsequent path optimization and collaborative search, ensuring that the UAV maintains high sensitivity and adaptability to the target state in dynamic environments. Step 2: Selective data transmission based on time utility map To reduce communication burden, a time utility map was constructed, and the priority of information in each region was evaluated in real time. By setting a time utility threshold, data from high-priority regions was filtered for transmission to ensure that important information was updated first and to avoid the transmission of outdated data. This selective data transmission mechanism can significantly reduce the use of communication bandwidth and optimize the real-time performance and resource utilization of the unmanned swarm system. Step 3: Dynamic region partitioning based on Venotu-Lloyd To achieve a reasonable allocation of task partitioning and search areas among UAVs, a dynamic region partitioning strategy based on the Venn diagram and Lloyd's algorithm is adopted. First, based on the location of each agent, the Venn diagram is divided into regions, and the search area is divided into multiple smaller regions; each agent is responsible for its nearest region; Then wait for the time-triggered signal, and when the requirements for area division are met again, the area is re-divided based on the current position of the drone. Step 4: Dynamic Path Optimization Based on Multi-Step Prediction After the region is divided, each UAV performs path optimization within its assigned search area. The Lloyd algorithm, based on the Veno map, first calculates the weighted centroid of each Veno map region, then iteratively moves the UAV to the weighted centroid of its corresponding Veno map region, repeatedly updating the region division and centroid position, ultimately achieving uniform spatial segmentation and optimal coverage. In the process of generating paths using the weighted centroid, the UAV not only considers the target probability map but also uses the uncertainty map and pheromone map to calculate the weighted centroid. The aim is to converge the weighted centroid to unsearched areas with high uncertainty and the possibility of targets. Specifically, the uncertainty map reflects the observation confidence level within the current region, guiding the UAV to prioritize areas with insufficient perception or ambiguous information. The pheromone map records historical search trajectories and reduces the possibility of repeated visits through a time decay mechanism, thereby avoiding resource waste. By integrating these three types of map information into a composite weight basis for path planning, the system can dynamically generate search paths for multiple future steps, effectively avoiding path redundancy, reducing turning frequency, and improving coverage continuity. The algorithm employs a multi-step prediction mechanism to guide the path towards the high-probability target area, thereby improving search performance and coverage efficiency. The UAV starts with the current centroid of the Vinograph region and calculates the first predicted centroid within a circle centered on this centroid with a radius of r1. Similarly, it continues to calculate the second predicted centroid within a circle centered on the first predicted centroid with a radius of r2, and so on, allowing for multiple predictions. The prediction radius gradually decreases as the search progresses. Furthermore, the total number of prediction steps is continuously adjusted based on the search progress. In the early stages of the search, fewer prediction steps are needed, and even additional predicted centroid calculations may not be required. In the later stages of the search, as the fragmentation of unsearched areas reduces search efficiency, the number of prediction steps continuously increases to prevent the centroid from converging into unsearched areas. However, considering computational efficiency, an upper limit on the number of prediction steps is set. Step 5: Task Execution and Dynamic Updates While the UAVs perform search missions, the system continuously updates the target probability map, uncertainty map, and pheromone map in real time. Information fusion and transmission operate routinely to ensure timely environmental perception and up-to-date map data. If the Vinyasa map region re-division is not triggered, the system skips the task reassignment process, and each UAV independently performs dynamic path optimization based on multi-step prediction using the latest probability map, continuously adjusting its search direction to approach high-probability areas and improve responsiveness and path foresight. The pheromone map is used to record historical trajectories and decays over time to prevent repeated overwriting. The entire execution process maintains high adaptability while ensuring coordination and consistency, ensuring that the system continuously maintains task efficiency and global robustness in dynamic environments.

2. The coverage optimization method based on dynamic region division and multi-step predictive path planning according to claim 1, characterized in that, In step 2, during the distributed information fusion process, a consensus protocol is used to verify the local map shared by each UAV. When there are conflicting updates in the same area, the system prioritizes retaining data with more recent source time and higher confidence, and ensures the stability and consistency of the fused map within the neighborhood through a local consistency update strategy.

3. The coverage optimization method based on dynamic region partitioning and multi-step predictive path planning according to claim 1, characterized in that, In step 2, the search pheromone map adopts a time decay mechanism, that is, as time goes by, the weight of historical search trajectories gradually weakens, so as to avoid the drone from continuously avoiding the searched area.

4. The coverage optimization method based on dynamic region partitioning and multi-step predictive path planning according to claim 1, characterized in that, In step 2, a weighted fusion strategy based on a time utility model is introduced during the map fusion process. This model comprehensively considers the observation delay, detection frequency and update confidence of each region, and dynamically calculates the fusion weight of each region in the map.

5. The coverage optimization method based on dynamic region partitioning and multi-step predictive path planning according to claim 1, characterized in that, In step 3, during the multi-UAV collaborative search process, each UAV exchanges locally constructed target probability maps through a distributed information sharing mechanism and evaluates the fusion priority of received information in conjunction with a time utility map; during the region division process, a joint division strategy based on Veno diagrams and Lloyd's algorithm is adopted. The search space is first generated by creating a Venn diagram region based on the current location information of each UAV. Then, by introducing the target probability density as a weighting factor for the region, the centroid of the Venn diagram region is iteratively adjusted to achieve dynamic redistribution of the search space.

6. The coverage optimization method based on dynamic region partitioning and multi-step predictive path planning according to claim 1, characterized in that, In the path planning process in step 4, the UAV uses the prediction-enhanced Lloyd algorithm to model and make forward-looking adjustments to the future target distribution. The prediction mechanism is based on historical probability evolution trends to infer the target point area in multiple future time steps and dynamically corrects the path direction in combination with the current search status.

7. The coverage optimization method based on dynamic region division and multi-step predictive path planning according to claim 1, characterized in that, In step 4, to avoid frequent sharp turns or unnatural movement in the path, the centroid calculation in the Lloyd algorithm introduces a directional weighting factor: when the UAV calculates the direction of movement to the next centroid, it also considers the angle weight between its current flight orientation and the expected target direction, thereby smoothing the turning trajectory and improving the executability and continuity of the path.

8. The coverage optimization method based on dynamic region partitioning and multi-step predictive path planning according to claim 1, characterized in that, To avoid divergent behavior in the multi-step path prediction process in step 4, a path convergence constraint mechanism is introduced. Centroid prediction stops and the current path is executed when any of the following conditions are met: 1) If the centroid variation of the path calculated in consecutive steps is lower than a set threshold, the prediction is considered stable, and subsequent steps are terminated, directly executing the current path; otherwise, the prediction radius is dynamically reduced and centroid prediction continues; 2) The predetermined upper limit of prediction steps is reached, where the specific upper limit threshold adapts to different search stages. This mechanism improves the stability and reliability of path planning.

9. The coverage optimization method based on dynamic region partitioning and multi-step predictive path planning according to claim 1, characterized in that, In step 4, during the predictive centroid path planning process, the target path generated by the UAV is no longer limited to its initially assigned Venn diagram region; since the path optimization process prioritizes guiding the UAV to move towards local or global probability maxima, its final centroid position may cross the initial region boundary and enter the region of an adjacent UAV.

10. The coverage optimization method based on dynamic region partitioning and multi-step predictive path planning according to claim 1, characterized in that, Step 1: Update the target probability map based on the probability hypothesis density filter; In multi-target detection tasks, the probability hypothesis density represents the expected distribution of the number of targets in space; To convert it to a standard probability density function, normalization can be performed: Where v(x) represents the probability hypothesis density value at position x; P PHD (x) represents in the subspace The probability of the in-place target appearing at point x satisfies the normalization constraint: Search Area Divided into sizes N i ×N j Grid G xy Each grid cell g xy Represents a fixed spatial region whose boundary is defined by coordinates [x] i,min ,x i,max ] and [y j,min ,y j,max [Confirmed; target probability assigned to the nearest grid center:] Among them, (x c ,y c ) is the grid g xy The center of , ||·||2 represents the Euclidean distance; Traditional Bayesian updates assume that the observations are completely reliable, but in real-world environments, sensor data contains noise and uncertainty. Therefore, a probability hypothesis density filter is used as a dynamic confidence weighting factor to correct the target probability update. Among them, P xy (k) represents the prior probability of the target at time k; P d For the detection probability, P f P represents the false alarm rate. PHD (x) serves as a weighting factor, improving the dynamic adaptability of probability updates; To account for measurement uncertainties, the update range is not limited to the target point, but rather extended based on the covariance matrix of the target location: Wherein, the covariance matrix of the target position (x0, y0) This reflects uncertainty; the Gaussian weighting ensures that greater update weights are assigned in regions closer to the target point, improving the robustness of detection. To control the spatial diffusion range of updates, the system introduces an effective support boundary constraint mechanism, performing update operations only within regions that satisfy the following elliptical constraint conditions: Where n is a constant, usually taking the value of 2 or 3, corresponding to a Gaussian distribution confidence range of about 95% or 99.7%; this boundary condition can effectively limit the spatial propagation range of update calculations, avoid meaningless diffusion of probability weights in large-scale regions, and ensure the locality and convergence efficiency of update calculations. Step 2: Information Fusion and Time Utility Map; This invention employs a multi-layered information fusion mechanism, combining a target probability map, an uncertainty map, and a search pheromone map: w xy (k)=αP xy (k)+βU xy (k)+γS xy (k) Among them, P xy (k) is the target probability map, output by the probability hypothesis density filter; U xy (k) is an uncertainty map, calculated based on Shannon entropy: U xy (k)=-P xy (k)log2(P xy (k))-(1-P xy (k))log2(1-P xy (k)) S xy (k) is a pheromone map, which has the property of decaying over time. Its decay mechanism is as follows: S xy (k+1)=S xy (k)·e -λt Where t is the time step and λ is the decay factor, ensuring that historical search information gradually weakens, enabling the UAV to adapt to the dynamically changing environment; the weighting coefficients α, β, and γ of the three types of layers are dynamically set according to the task type, the dynamic degree of the target, and resource constraints, and are used to adjust the focus of the fusion map at different stages. Step 3: Distributed collaborative search and information sharing; UAVs perform information fusion through consensus: Among them, w ii (k) and w ij (k) Dynamically adjust based on the time utility map: Where, τ xy (k)=e -λ(k-t(x,y)) As a time-use decay factor, it ensures that the drone uses the latest data; Step 4: Region Division and Event-Triggered Update Mechanism To achieve efficient collaborative search among multiple UAVs in dynamic environments, this invention incorporates a region division stage before path planning to clarify the task responsibilities of each UAV; the search space between UAVs is divided using a Venn diagram, where each sub-region... For a given UAV i, the location is dynamically determined based on its current position and the target probability distribution. However, frequent redrawing of the Venn diagram region can significantly increase the computational burden of the system, especially in large-scale multi-UAV systems. To address this, this invention designs an event-triggered region update mechanism that redraws the Venn diagram region only when necessary, thereby reducing computational overhead while ensuring system flexibility. This mechanism introduces a trigger function Γ k A comprehensive assessment of the magnitude of current environmental changes and system operational stability is necessary. in, It represents the global change in the target probability map between two iterations, reflecting changes in target movement or perceived uncertainty; This represents the average offset of all intelligent points, reflecting whether the system has entered a stable state; Γ is the change normalization factor, ∈ is the centroid shift threshold, and the two make the indicators comparable; w1+w2=1 is the dynamic weighting coefficient, used to adjust the relative importance of environmental changes and system stability; when Γ k When ≤γ, the system triggers a Vinograph repartition update, where γ is the set trigger threshold; this strategy implements adaptive updates under two key conditions: When the target distribution changes drastically, i.e. When the size is large, the system actively adjusts the region division; When the system tends to stabilize, i.e., Δc k When the size is small, it can be replanned in advance; This mechanism avoids unnecessary repetitive partitioning calculations and can adaptively adjust regions according to the dynamic nature of tasks, significantly reducing the computational burden and improving the system's response speed to environmental changes. Step 5: Improve the path optimization of the Lloyd algorithm; when performing the region partitioning and centroid calculation of the Venn diagram, introduce a composite weight function w′ that integrates multi-source information. xy Used to guide the centroid to shift towards higher value directions: w′ xy =w xy ·(1+a dir ·cosθ) Among them, w xy The basic weights represent a combination of target probability, uncertainty, and pheromone. The vector pointing from the current position to the target position. With heading The angle between them; Indicates the consistency between the current action direction and the target direction; α dir The directional adjustment coefficient controls the influence of the path orientation on the centroid offset; this directional guidance mechanism can suppress frequent turning phenomena, ensure path continuity and directional stability, and improve overall motion efficiency. Path prediction employs a sliding window iterative Lloyd update method; assuming UAV i is currently at time k, its path planning uses a predicted depth N. p Generate a set of future centroid sequences of length: Among them, the The predicted centroid is generated according to the following recursive rule: Where Γ(c,r) represents the location of the local weighted centroid calculated by combining the weight map within a radius of c and step size r; For the first The search radius of the step satisfies the decreasing constraint: Final path based on predicted centroid The ultimate goal is to predict depth N based on the search progress. p Adaptive change, defining the total search space region as The covered area is The non-coverage rate is: Let N be the maximum number of prediction steps. max The minimum is N min Then the number of prediction steps can be adaptively set as follows: Where, ρ k ∈[0,1], the proportion of uncovered areas; a larger value indicates an earlier search phase; β>0, adjusts the non-linearity of step growth, ultimately allowing the number of steps to automatically increase in the later stages of the search. When the maximum prediction depth N is reached... p Previously, if the change in the centroid of a path calculated with consecutive steps was less than a preset threshold θ, the prediction would be considered stable, and the prediction would be stopped and the current path would be executed.

Images