A method, system, device and medium for cooperative search of a lost UAV cluster

Abstract

The application discloses a kind of out-of-contact unmanned aerial vehicle cluster cooperative search method, system, equipment and medium, it is related to unmanned aerial vehicle technical field. First by ground control station according to out-of-contact unmanned aerial vehicle information and last position, initial search area is divided into multiple sub-regions, and each search unmanned aerial vehicle is allocated to by search cluster through decentralized distributed auction mechanism; Each unmanned aerial vehicle searches in distribution area, collects multimodal data and environment adaptive dynamic weighted fusion, and real-time calculates comprehensive confidence score;When comprehensive confidence score is over threshold value, determine to find machine and select leading perception mode;Discovery machine generates and real-time updates probability heat map, simultaneously broadcasts strong clue alarm;Adjacent unmanned aerial vehicle autonomously decides to participate in cooperative confirmation;After confirming target, the rest unmanned aerial vehicle switches to debris search mode, returns data and generates visual debris distribution atlas. Both improve multimodal fusion accuracy and environmental adaptability, and realize distributed autonomous cooperation and whole-process closed loop, improve search efficiency and success rate.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and in particular to a method, system, device, and medium for collaborative search of a swarm of missing UAVs. Background Technology

[0002] Rapidly and accurately locating missing drones is crucial for recovering valuable payloads, analyzing accident causes, preventing potential safety hazards, and minimizing property damage. However, current search technologies for missing drones still have significant shortcomings, resulting in low search efficiency and unreliable success rates. Existing methods mainly rely on two approaches: first, the last known GPS coordinate method, where searchers use the GPS coordinates last transmitted by the drone before it went missing to conduct a visual search; second, the manual radio direction finding method, where, assuming the drone's battery is not completely damaged and it continues to emit radio frequency signals, technicians use handheld radio direction finding equipment, rotating the directional antenna to find the direction with the strongest signal strength, gradually approaching the target. However, both methods have fundamental limitations, namely weak multi-source sensing fusion capabilities, lack of autonomous collaborative decision-making capabilities, poor environmental adaptability, and low search efficiency, making it difficult to meet the practical needs of efficient and accurate searches in complex environments. These limitations are specifically reflected in the following aspects.

[0003] First, the sensing methods are limited and the ability to fuse multi-source data is weak. Traditional search methods rely excessively on single GPS coordinate information or continuous radio frequency signals. Once the signal is lost, blocked, or interfered with, the entire search operation immediately comes to a standstill. Visual search, on the other hand, is highly dependent on the operator's personal experience, and its detection efficiency in the visible light band drops sharply in complex environments such as nighttime, dense fog, rain, snow, and dense forests. Existing methods cannot effectively integrate various heterogeneous sensing information such as visible light images, infrared thermal source characteristics, and radio frequency signal strength, making it difficult to achieve complementary confirmation and dynamic comprehensive assessment of the target. Due to the limited sensing dimensions, the failure of any information source causes the search system to lose its continuous perception capability of the target, revealing significant shortcomings such as low efficiency, poor reliability, and weak environmental adaptability in complex environments.

[0004] Second, the single-machine operation mode suffers from low coverage efficiency and lacks multi-machine collaboration and distributed decision-making capabilities. Whether it's manual ground search or a single drone searching from the air, it's difficult to cover large unknown areas in a short time, and the time from loss of contact to final location is lengthy. Each search unit operates independently, lacking autonomous area division, neighbor information sharing, and task negotiation and dynamic allocation mechanisms, making it impossible to form a unified understanding of the global target situation through clustered distributed computing. Even if a unit captures suspected target features, it's difficult to immediately call other platforms to dynamically converge on the target area for collaborative confirmation. This isolated working mode severely restricts search coverage speed and target acquisition probability, causing the optimal search window to be wasted during point-by-point scanning.

[0005] Third, the search process lacks autonomy and real-time decision-making guidance. Current technology relies entirely on operators making manual judgments and guiding routes based on extremely limited information. The search system itself lacks the ability to dynamically update the probability distribution of target appearance based on real-time sensing data. When a platform observes a weak suspected heat source or a momentary radio frequency signal, the system cannot automatically generate and update a probability heatmap, nor can it autonomously adjust its local search strategy through a dynamic weighted self-searching algorithm and trigger cluster-wide coordinated movement towards high-probability areas. The decision-making process is highly dependent on human experience and the continuous availability of communication links, resulting in a significant lag in response. In changing weather conditions and complex terrain environments, it is extremely easy to miss crucial positioning opportunities.

[0006] Fourth, the end-point interactive analysis and search and rescue execution are disconnected, lacking an integrated closed loop. Traditional handheld direction-finding devices or stand-alone image transmission systems can only achieve simple point marking, lacking multi-dimensional batch data query capabilities such as box selection and circle selection, making it difficult to quickly compare features and analyze debris distribution in suspected target areas. Furthermore, existing systems do not integrate multi-camera perspective collaborative confirmation and rapid mapping and modeling modules, making it impossible to immediately verify the authenticity of a target after its discovery, nor can they simultaneously construct a debris distribution map. In addition, due to the lack of mobile adaptation and augmented reality (AR) assistance, decision-making information generated by the command post cannot be transmitted to on-site search personnel in real time, and on-site personnel also find it difficult to accurately correlate the visualized situation on the screen with the actual environment. This results in search results being unable to directly support subsequent accident analysis and payload recovery work, creating an information gap between command decisions and on-site execution. Summary of the Invention

[0007] This invention provides a method, system, device, and medium for collaborative search of missing drone swarms, which solves the technical problems of insufficient perception fusion, lack of collaborative decision-making, weak autonomous guidance, and weak multi-drone collaborative search and rescue tracing capabilities in existing collaborative search technologies for missing drone swarms, resulting in low search efficiency and low success rate in complex environments.

[0008] The first aspect of this invention provides a collaborative search method for a swarm of missing unmanned aerial vehicles (UAVs), comprising:

[0009] The ground control station divided the initial search area into multiple sub-areas based on the missing drone's identity information and last known location;

[0010] The search cluster autonomously allocates each sub-region to a search drone within the search cluster through a decentralized distributed auction mechanism;

[0011] Each of the aforementioned search drones performs a search within its assigned sub-region, simultaneously collecting multimodal perception data, and performing environmentally adaptive dynamic weighted fusion on the collected multimodal perception data to calculate the comprehensive confidence score in real time;

[0012] When the overall confidence score exceeds a preset threshold, the search drone that has calculated the overall confidence score will be used as the discoverer, and the discoverer will use the perception mode with the highest current weight in the multimodal perception data as the dominant perception mode.

[0013] The detection drone generates an initial probability heatmap based on the dominant sensing mode, and updates the probability heatmap in real time by fusing newly acquired multimodal sensing data during the approach to the target to obtain a target probability heatmap. At the same time, it broadcasts a strong clue alarm to nearby search drones.

[0014] Upon receiving the strong clue alarm, the nearby search drones autonomously decide whether to participate in collaborative confirmation based on their own status. Search drones that participate in the collaboration converge towards the center of the target probability heatmap to confirm the target.

[0015] Once any participating search drone confirms the target, the remaining participating search drones automatically switch to debris search mode to search for scattered debris parts and transmit the collected debris data back to the ground control station. The ground control station then integrates all the transmitted debris data to generate a visual debris distribution map.

[0016] Optionally, the step of the search cluster autonomously allocating each sub-region to each search drone in the search cluster through a decentralized distributed auction mechanism includes:

[0017] After each search drone in the search cluster completes initialization, a temporary auctioneer is generated through a dynamic election mechanism. The temporary auctioneer selects an unassigned sub-region from the sub-region task list issued by the ground control station and broadcasts the auction message to the entire cluster.

[0018] After receiving the auction message, each of the search drones calculates the task cost corresponding to the sub-region to be allocated based on its own status and the matching degree between the sub-region to be allocated, and generates a corresponding bid based on the task cost and sends it to the temporary auctioneer.

[0019] The temporary auctioneer collects all bids from all bidders, selects the search drone with the best bid as the winner, and broadcasts the winning bid message to the entire cluster.

[0020] Once the winning drone confirms receipt of the task, all search drones in the search cluster will uniformly mark the sub-area to be allocated as allocated.

[0021] Repeat the auction process described above until all sub-regions in the sub-region task list have been allocated.

[0022] Optionally, the step of having each of the search drones perform a search within its assigned sub-region, simultaneously collect multimodal perception data, and perform environmentally adaptive dynamic weighted fusion of the collected multimodal perception data to calculate a comprehensive confidence score in real time includes:

[0023] Each of the aforementioned search drones performs a comprehensive search within its assigned sub-region using an extended spiral or a cooperative plowing path.

[0024] During the entire process of performing a comprehensive search, each of the search drones simultaneously collects radio frequency signals, visible light images, and thermal infrared data, and preprocesses the collected raw data to generate radio frequency feature data, visual feature data, and thermal source feature data, respectively.

[0025] Each of the search drones normalizes the radio frequency feature data, visual feature data and thermal source feature data it collects to a unified numerical range to obtain normalized radio frequency features, normalized visual features and normalized thermal source features.

[0026] Based on the current time, lighting conditions, and terrain environment, dynamically allocate radio frequency weights, visual weights, and heat source weights;

[0027] Based on the radio frequency weight, the visual weight, and the heat source weight, the normalized radio frequency features, the normalized visual features, and the normalized heat source features are weighted and summed to calculate the comprehensive confidence score corresponding to each of the search drones.

[0028] Optionally, the step of dynamically allocating radio frequency weights, visual weights, and heat source weights based on the current time, lighting conditions, and terrain environment includes:

[0029] When it is detected that the current environment is a clear and open daytime, the visual weight is set to the highest weight, the radio frequency weight is set to the second highest weight, and the heat source weight is set to the lowest weight.

[0030] When it is detected that the current environment is nighttime or dense fog, the heat source weight is set to the highest weight, the radio frequency weight is set to the second highest weight, and the visual weight is set to the lowest weight.

[0031] When it is detected that the current environment is a jungle or dense vegetation, the radio frequency weight is set to the highest weight, the visual weight is set to the second highest weight, and the heat source weight is set to the lowest weight.

[0032] Optionally, the step of generating an initial probability heatmap based on the dominant sensing mode by the detection drone, updating the probability heatmap in real time by fusing newly acquired multimodal sensing data during the approach to the target to obtain a target probability heatmap, and simultaneously broadcasting a strong clue alert to nearby search drones includes:

[0033] If the dominant sensing mode is radio frequency, the detection machine generates an initial probability heatmap of a preset shape with the direction of arrival of the radio frequency signal as the axis, and uses the initial probability heatmap of the preset shape as an intermediate probability heatmap.

[0034] If the dominant perception mode is visual or thermal infrared, the detection machine calculates the target azimuth angle based on the pixel position of the suspected target in the image and the current pose of the detection machine, generates a Gaussian initial probability heatmap centered on the target's calculated coordinates, and uses the Gaussian initial probability heatmap as an intermediate probability heatmap.

[0035] The detection machine approaches the high-probability region in the intermediate probability heatmap, continuously collecting new multimodal sensing data during the approach process;

[0036] The discovery machine uses a Bayesian update rule to fuse new multimodal sensing data into the intermediate probability heatmap, and introduces an uncertainty diffusion mechanism to simulate the decay of clues over time, thereby updating the target probability heatmap in real time.

[0037] While updating the target probability heatmap in real time, the detection aircraft broadcasts a strong clue alert to all nearby search drones within its communication range.

[0038] Optionally, the step of the nearby search drones that receive the strong clue alarm autonomously deciding whether to participate in collaborative confirmation based on their own status, and the search drones participating in the collaboration converging towards the center of the target probability heatmap to confirm the target, includes:

[0039] The nearby search drone that receives the strong clue alarm extracts the high-precision coordinates of the discoverer, the target probability heat map and the dominant clue type from the strong clue alarm, and obtains its own distance from the center of the target probability heat map, remaining battery power and sensor status parameters.

[0040] The degree of collaborative participation is calculated by the nearby search drone based on the distance, the remaining battery power, and the sensor status parameters.

[0041] If the degree of collaborative participation exceeds a preset participation threshold, the nearby search drone will autonomously decide to participate in the collaborative confirmation.

[0042] The participating search drones autonomously plan conflict-free flight paths and converge towards the center of the target probability heatmap;

[0043] The first search drone to arrive at the pre-defined confirmation range hovers and uses a high-magnification zoom camera to capture and identify the suspected target from multiple angles, thus confirming the target.

[0044] Optionally, the step of automatically switching to wreckage search mode after any participating search drone confirms the target, searching for scattered wreckage parts and transmitting the collected wreckage data back to the ground control station, and the ground control station integrating all transmitted wreckage data to generate a visualized wreckage distribution map, includes:

[0045] Once any participating search drone confirms the target, the confirming search drone immediately broadcasts a target confirmation message containing the target's precise coordinates and image evidence to the entire cluster.

[0046] Upon receiving the target confirmation message, the remaining search drones participating in the collaboration automatically terminate their convergence behavior toward the center of the target probability heatmap and switch to an extended scanning mode centered on the confirmed target point.

[0047] During the extended scan, each search drone collects the precise coordinates of the debris components and takes images to generate debris data, which is then transmitted back to the ground control station in real time.

[0048] The ground control station integrates target confirmation information and debris data transmitted back from multiple aircraft, marks the main debris point and various scattered debris points on the electronic map, and generates a visualized debris distribution map.

[0049] A second aspect of the present invention provides a collaborative search system for a swarm of missing unmanned aerial vehicles (UAVs), comprising:

[0050] The area division module is used by the ground control station to divide the initial search area into multiple sub-regions based on the identity information and last known location of the missing UAV;

[0051] The task allocation module is used by the search cluster to autonomously allocate each of the sub-regions to each search drone in the search cluster through a decentralized distributed auction mechanism;

[0052] The perception fusion module is used by each of the search drones to perform searches in the assigned sub-areas, synchronously collect multimodal perception data, and perform environmental adaptive dynamic weighted fusion on the collected multimodal perception data to calculate the comprehensive confidence score in real time.

[0053] The clue determination module is used to identify the search drone that has calculated the comprehensive confidence score as the discoverer when the comprehensive confidence score exceeds a preset threshold, and the discoverer will take the perception mode with the highest current weight in the multimodal perception data as the dominant perception mode.

[0054] The heat map push module is used to generate an initial probability heat map by the discovery drone based on the dominant perception mode, and to update the probability heat map in real time by integrating newly acquired multimodal perception data during the approach to the target to obtain a target probability heat map, while broadcasting a strong clue alarm to nearby search drones.

[0055] The collaborative verification module is used for nearby search drones that receive the strong clue alarm to autonomously decide whether to participate in collaborative verification based on their own status. Search drones that participate in the collaboration converge towards the center of the target probability heat map to confirm the target.

[0056] The debris mapping module is used so that when any participating search drone confirms the target, the remaining participating search drones automatically switch to debris search mode, search for scattered debris parts, and transmit the collected debris data back to the ground control station. The ground control station then integrates all the transmitted debris data to generate a visual debris distribution map.

[0057] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the above-described method for collaborative search of missing drone swarms.

[0058] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the above-described method for collaborative search of a cluster of missing drones.

[0059] As can be seen from the above technical solutions, the present invention has the following advantages:

[0060] This invention addresses the technical problems in existing collaborative search technologies for missing drone swarms, such as insufficient perception fusion and collaborative decision-making, disconnect between autonomous guidance and closed-loop verification, and weak multi-drone collaborative search and tracing capabilities. It proposes a collaborative search method for missing drone swarms. First, the ground control station divides the initial search area into multiple sub-regions based on the missing drone's identity information and last known location. Then, the search swarm autonomously allocates each sub-region to individual search drones within the swarm through a decentralized distributed auction mechanism. Each search drone performs a search within its assigned sub-region, simultaneously collecting multimodal perception data. The collected multimodal perception data undergoes environmentally adaptive dynamic weighted fusion, and a comprehensive confidence score is calculated in real time. When the comprehensive confidence score exceeds a preset threshold, the search drone with that score is designated as the discoverer, and the discoverer... The drone takes the perception mode with the highest weight in the multimodal perception data as the dominant perception mode. The detection drone generates an initial probability heatmap based on the dominant perception mode and updates the probability heatmap in real time by integrating newly acquired multimodal perception data during the approach to the target to obtain the target probability heatmap. At the same time, it broadcasts a strong clue alarm to the nearby search drones. The nearby search drones that receive the strong clue alarm decide autonomously whether to participate in the collaborative confirmation based on their own status. The search drones participating in the collaboration converge towards the center of the target probability heatmap to confirm the target. When any of the search drones participating in the collaboration confirms the target, the remaining search drones automatically switch to the debris search mode to search for scattered debris parts and transmit the collected debris data back to the ground control station. The ground control station integrates all the transmitted debris data to generate a visualized debris distribution map. This solution processes multimodal perception data through environmentally adaptive dynamic weighted fusion, enabling it to adapt to different search environments and addressing the limitations of single-sensor methods in specific scenarios, thus improving the stability of clue detection. A decentralized distributed auction mechanism enables autonomous task allocation, eliminating the need for centralized scheduling at ground stations and mitigating the impact of single-point failures on the overall search task, while simplifying the control process for cluster collaboration. Autonomous guidance through real-time updates of probability heatmaps during target approach allows for adjustments to the target probability distribution based on continuously acquired new data, improving the accuracy of target approach. Multi-drone collaboration in target confirmation and subsequent debris search and data integration processes provides comprehensive target and debris information, and the generated visualized debris distribution map offers fundamental data support for subsequent drone recovery and accident analysis, enhancing the overall processing capabilities for searching for missing drones. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 A flowchart illustrating the steps of a collaborative search method for a missing unmanned aerial vehicle (UAV) swarm provided in this embodiment of the invention;

[0063] Figure 2 This is a schematic diagram of the structure of a collaborative search device for a missing drone swarm provided in an embodiment of the present invention;

[0064] Figure 3 This is a schematic diagram of the distributed search planning algorithm provided in an embodiment of the present invention;

[0065] Figure 4 A schematic diagram of the process for environmental adaptive dynamic weighted fusion of multimodal sensing data provided in this embodiment of the invention.

[0066] Figure 5 A schematic diagram of the algorithm flow for dynamic weight self-finding provided in an embodiment of the present invention;

[0067] Figure 6 This is a schematic diagram illustrating the specific process of probability heatmap updating provided in an embodiment of the present invention;

[0068] Figure 7 This is a schematic diagram of the collaborative verification and mapping algorithm provided in an embodiment of the present invention;

[0069] Figure 8 This is a structural block diagram of a collaborative search system for a missing drone swarm provided in an embodiment of the present invention;

[0070] Figure 9 This is a structural block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0071] This invention provides a method, system, device, and medium for collaborative search of missing drone swarms, which addresses the technical problems of insufficient perception fusion and collaborative decision-making, disconnect between autonomous guidance and closed-loop verification, and weak multi-drone collaborative search and rescue tracing capabilities in existing collaborative search technologies for missing drone swarms, resulting in low search efficiency and low success rate in complex environments.

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a collaborative search method for a missing drone swarm provided in this embodiment of the invention.

[0074] This method can be applied to, for example... Figure 2 The device structure shown is a simplified core process diagram, only demonstrating key interaction units and data flow. The overall architecture comprises four core components: search personnel, ground control station, search cluster, and search drones. During actual operations, search personnel input information about the missing drone, and the relevant data is synchronized to the global monitoring platform via the task input module. Both the task input module and the global monitoring platform are components of the ground control station. The global monitoring platform, leveraging its control capabilities, delineates the initial search area and sub-regions based on the missing drone's identity information and last known location. The generated sub-region task list is transmitted to the corresponding search drones within the search area via the task distribution channel. The search cluster and search drones work collaboratively, and search results generated during the search process are transmitted back to the global monitoring platform, ultimately forming a complete closed-loop control process through result feedback.

[0075] Furthermore, the search drone is equipped with an onboard processing system, specifically including a multimodal perception module for scanning ground and signal data, a high-precision pose estimation module for providing high-precision positioning information to confirm the precise location of the search drone, a neighbor communication module for exchanging data between search drones based on a preset communication protocol, and a collaborative decision-making control module with built-in integrated algorithms for data processing. These modules provide hardware support for the search drone's operations and are adapted to the collaborative needs of the entire process of searching for missing drones.

[0076] This invention provides a collaborative search method for a swarm of missing drones, comprising:

[0077] Step 101: The ground control station divides the initial search area into multiple sub-areas based on the identity information and last known location of the missing UAV.

[0078] It should be noted that in this scenario, the ground control station refers to the management platform used to coordinate the collaborative search mission of the missing drone cluster, process multi-source data, and issue commands to the search cluster. It is the command center of the search mission and includes a task input module and a global monitoring platform. The identity information of the missing drone refers to the set of attributes that can uniquely identify the missing target, including the model, unique identification code, rated flight parameters, physical feature model, and communication frequency band. The last known location refers to the spatiotemporal coordinate data of the last time the missing drone was recorded before it lost contact. The initial search area refers to the geographical area to be searched, which is centered on the last known location of the missing drone and combined with flight performance, the scenario of loss of contact, and the terrain environment. Sub-regions refer to independent search units adapted to the operational capabilities of a single search drone, obtained by dividing the initial search area according to preset rules. Among them, the preset area division rules refer to the standards for delineating the initial search area based on the model of the missing drone, the duration of loss of contact, and the complexity of the terrain. For example, for drones with strong endurance, the radius of the initial search area can be expanded, while for complex terrains such as mountains and jungles, the area radius can be reduced to focus on high-probability search areas.

[0079] In this embodiment of the invention, search personnel input the identity information and last known location of the missing drone. This data is entered through the task input module of the ground control station and simultaneously transmitted to the global monitoring platform. The global monitoring platform, combining parameters such as the missing drone's rated endurance, duration of loss of contact, and terrain features, determines the initial search area according to preset regional division rules. Subsequently, based on the geographical boundaries and terrain data of the initial search area, as well as the drone operation capabilities of the search cluster, the global monitoring platform uses geographic grid partitioning or terrain-adaptive segmentation to divide the initial search area into multiple sub-regions of roughly equal size and clear boundaries. It then generates a sub-region task list containing the boundary coordinates, terrain data, and search priority identifiers for each sub-region, providing standardized task data for subsequent task allocation by the search cluster. This effectively solves the problems of arbitrary regional division and low task matching in traditional manual searches.

[0080] Step 102: The search cluster autonomously allocates each sub-region to each search drone in the search cluster through a decentralized distributed auction mechanism.

[0081] It should be noted that the decentralized distributed auction mechanism is a complete decision-making and operational logic integrated within each search drone, combining communication interaction protocols and task evaluation functions. Task assignment is completed through autonomous cluster collaboration, without the need for an additional central control node for overall scheduling. The search cluster is a collaborative working group composed of all search drones participating in this search operation. The preset timeout is a parameter introduced for the first time here, referring to the maximum waiting time for the auctioneer to collect all bidding bids. This timeout value can be flexibly set based on cluster communication distance and device response speed. Once this timeout is exceeded, bid collection stops and result determination begins.

[0082] In this embodiment of the invention, after the ground control station completes the search area division, the total number of sub-regions is greater than the total number of search drones deployed on-site. This allows for refined task allocation, ensuring a balanced workload across the entire cluster. All divided sub-regions are then compiled into a standardized sub-region task list and distributed throughout the search cluster. All search drones within the cluster undergo a unified initialization self-check, including basic operations such as onboard processing system function self-checks, high-precision positioning calibration, and cluster network communication connectivity testing. Once all equipment meets the required standards, the auction and allocation process officially begins. Figure 3 As shown, Figure 3 This is a schematic diagram of the decentralized distributed auction task allocation process provided in the embodiments of the present invention. It clearly presents the complete process of task initialization partitioning, auction initiation, cost calculation, bid submission, result determination, task confirmation, and entering parallel search after all allocations are completed. At the same time, it distinguishes between two types of execution branches: the allocation of the best bid and the difficult area without valid bid markings.

[0083] Furthermore, step 102 may include the following sub-steps:

[0084] S11. After each search drone in the search cluster completes its initialization, a temporary auctioneer is generated through a dynamic election mechanism. The temporary auctioneer selects an unassigned sub-region from the sub-region task list issued by the ground control station and broadcasts the auction message to the entire cluster.

[0085] In this embodiment of the invention, the ground control station completes the search area division and task issuance phases. All search drones have a built-in distributed search planning algorithm, which integrates decision logic, communication protocols, and evaluation functions, and can run independently on a single device. The algorithm explicitly sets two types of execution roles: auctioneer and bidder. The temporary auctioneer appearing in this step is the auctioneer set by the algorithm. It can be the drone that first completes initialization preparation in the cluster, or it can be generated through dynamic election within the cluster. The role is not fixed; any device in the cluster can serve as the auctioneer, effectively avoiding the impact of single-point device failure on the overall process.

[0086] exist Figure 3 In the corresponding task initialization phase, methods such as grid segmentation and polygon segmentation are used to divide the entire large search area into N sub-regions of different types that do not overlap (represented as...). These tasks, arranged sequentially to form an ordered queue structure, constitute the sub-region task list mentioned in the process, which is uniformly distributed to the search cluster by the ground control station. After all UAVs within the search cluster complete initialization self-checks, positioning calibration, and network connectivity verification, a temporary auctioneer is selected according to predetermined rules. The temporary auctioneer retrieves the locally stored sub-region task list, filters sub-regions in an unassigned state, and publishes auction-related notifications according to the algorithm's built-in communication rules, completing the network-wide broadcast of the auction message. Figure 3 The process node where the auctioneer selects an unallocated sub-region to initiate an auction.

[0087] S12. After receiving the auction message, each search drone calculates the task cost corresponding to the sub-region to be allocated based on its own status and the matching degree between the sub-region to be allocated, and generates a corresponding bid based on the task cost and sends it to the temporary auctioneer.

[0088] In this embodiment of the invention, the bidders defined by the algorithm cover all drones within the search cluster, and the device currently acting as a temporary auctioneer also has bidding authority. The algorithm is equipped with a dedicated bidding function, which is used to calculate the suitability of a single drone to undertake tasks in the corresponding sub-region. The bid value generated by the device is inversely proportional to the overall task cost generated by the device executing the task.

[0089] After all devices in the cluster receive the auction notification, they enter... Figure 3 The cost calculation process, taking into account the real-time operating status, the performance of onboard sensors, and the terrain and environmental conditions of the sub-region, calculates the task cost according to a unified formula. The overall calculation relationship is as follows:

[0090] ;

[0091] in, This represents the straight-line distance between the current location of the search drone and the geometric center of the sub-region to be assigned. This represents the estimated flight energy consumed to complete the search of the entire area. The value is affected by the distance traveled, the area size, the flight speed, and the remaining battery power of the equipment. The lower the remaining battery power, the higher the energy consumption cost for long-distance operations. The cost of adapting equipment to different terrains is determined based on the drone's detection capabilities and the degree of terrain compatibility, i.e., the cost of using search drones. Sensor capabilities and The degree of matching with the terrain environment, for example, if In dense woodlands, drones equipped with high-frequency detection units are less expensive in this regard; This represents the current workload cost of the equipment, and includes statistics on the number of sub-areas and operating area already undertaken by the equipment, used to balance the overall workload of the cluster, i.e., search drones. The number of sub-regions or the total area currently allocated are used to ensure cluster load balancing and prevent a particular search drone mission from being overloaded.

[0092] After all equipment costs are calculated, a bid is determined using a bidding function and the cost value. The lower the task cost, the higher the bid, indicating that the equipment is more suitable for the corresponding search task. The compiled bidding information is then sent to the temporary auctioneer according to communication transmission standards. Figure 3 The process node where the drone sends out the bidding price after completing the calculation.

[0093] S13. The temporary auctioneer collects all bids from all bidders, selects the search drone with the best bid as the winner, and broadcasts the winning bid message to the entire cluster.

[0094] In this embodiment of the invention, the temporary auctioneer initiates the collection of bidding information and simultaneously starts a timeout statistics process. A preset timeout period, a parameter first proposed in this embodiment, is used to limit the maximum time for collecting all bidding bids. In practical applications, the specific value can be flexibly adjusted based on the cluster communication distance and device response speed. After the waiting period ends, the temporary auctioneer summarizes and organizes all valid bids, completes value comparison and filtering, selects the highest bid as the optimal bid, determines the drone that submitted this bid as the winning device, and then publishes the winning bid announcement to the entire network, i.e., broadcasts the winning bid message to the entire cluster. Figure 3 The process branch for selecting the best bidder to determine the winning equipment.

[0095] If no valid bids are collected after the preset time period ends, it means that there are no matching operating equipment in the current sub-area. The temporary auctioneer will designate this area as a difficult area for temporary storage, pending subsequent unified allocation. Figure 3 Another type of process branch is the one where no valid bids define the difficult area.

[0096] S14. Once the winning drone confirms receipt of the task, all search drones in the search cluster will uniformly mark the sub-areas to be assigned as assigned.

[0097] In this embodiment of the invention, the successfully identified winning drone receives the winning bid notification from the cluster, verifies the task-related information, confirms acceptance of the search task, and simultaneously sends the task acceptance confirmation message outwards. Figure 3 The process nodes for confirming the successful bid for a drone mission are as follows: After receiving the confirmation information, the temporary auctioneer pushes a status update notification to all devices in the cluster, searches all drones in the cluster, and synchronously modifies the locally saved sub-region task list. All sub-regions to be allocated in this round are marked as "allocated," preventing duplicate bidding and allocation of the same region and ensuring a stable and orderly task allocation process.

[0098] S15. Repeat the above auction process until all sub-regions in the sub-region task list have been allocated.

[0099] In this embodiment of the invention, after the auction allocation process for a single sub-region is completed, the temporary auctioneer retrieves the complete sub-region task list again, selects the remaining unallocated sub-regions in the list, and repeats all the above auction operation steps. This loop continues until all sub-regions in the sub-region task list are normally allocated, or areas deemed problematic are retained, at which point the entire decentralized auction allocation process officially ends. All allocation work is completed in sequence. Figure 3 With all process nodes assigned to all areas, each drone in the cluster has an independent and non-overlapping search range. All devices proceed to their designated locations according to their assignments, officially commencing parallel search operations.

[0100] Step 103: Each search drone performs a search within its assigned sub-area, simultaneously collecting multimodal perception data, and performing environmentally adaptive dynamic weighted fusion on the collected multimodal perception data to calculate the comprehensive confidence score in real time.

[0101] Furthermore, step 103 may include the following sub-steps:

[0102] S21. Each search drone performs a comprehensive search within its assigned sub-area using an extended spiral or a cooperative plowing path.

[0103] In this embodiment of the invention, after sub-region allocation, each search drone obtains its own exclusive and non-overlapping sub-region, and then proceeds in parallel to its assigned sub-region to initiate a comprehensive search operation. During the search, two comprehensive search paths are employed: an expanding spiral path and a cooperative plowing path, ensuring no blind spots or duplicate searches within the sub-region and adapting to search needs in different terrain environments. The expanding spiral path starts from the geometric center of the sub-region and gradually expands outwards in a spiral, suitable for relatively flat areas without significant obstructions. The cooperative plowing path advances parallel along the long side of the sub-region, suitable for complex terrain with localized obstructions. The two paths can be automatically adapted and selected based on the sub-region's terrain, ensuring comprehensive and efficient search coverage.

[0104] S22. During the entire process of conducting a comprehensive search, each search drone simultaneously collects radio frequency signals, visible light images, and thermal infrared data, and preprocesses the collected raw data to generate radio frequency feature data, visual feature data, and thermal source feature data, respectively.

[0105] In this embodiment of the invention, when each search drone performs a search operation, it simultaneously collects three types of data through an onboard multimodal perception module. The multimodal perception module specifically includes a radio frequency detection unit, a visual perception unit, and a thermal imaging unit. These three units work together to complete the raw data acquisition and preprocessing. The specific data generation process is as follows:

[0106] Radio frequency (RF) signal acquisition and RF feature data generation: The RF detection unit is woken up by a software-defined radio (SDR) or a specific frequency band receiver. It scans the communication frequency band of the lost UAV (e.g., 2.4 GHz data transmission, 5.8 GHz image transmission) pre-input by the ground control station and measures the received signal strength (RSSI) through the RF front-end circuit. If equipped with a directional antenna or array, the signal source direction (DOA) can also be calculated through phase difference. After noise reduction and filtering preprocessing of the above RSSI and DOA data, RF feature data is generated to locate the signal position of the lost UAV.

[0107] Visible light image acquisition and visual feature data generation: The visual perception unit is an airborne visible light camera that captures ground images below at a predetermined frame rate (a preset parameter appearing for the first time here, referring to the number of frames the camera captures per second, which can be set according to search speed and target recognition accuracy, for example, 10-20 frames / second); the captured image data is sent to a deep learning target detection model running in the airborne processor. This model is specifically optimized for the shape, color, and texture features of the missing drone. After preprocessing the image data by denoising, scaling, and feature extraction, it outputs the bounding boxes and confidence scores of suspected targets in the image, which are used as visual feature data.

[0108] Thermal infrared data acquisition and heat source feature data generation: The thermal imaging unit is an uncooled infrared thermal imager that performs a comprehensive scan of the ground to generate a ground thermal radiation distribution map; the airborne processor performs noise reduction and enhancement preprocessing on the thermal radiation distribution map, identifies abnormal heat source points that have significant differences from the ambient background temperature (the preset parameter that appears for the first time here, with the temperature difference threshold set at 5-10℃), and records the temperature value and pixel coordinates of the abnormal heat source points as heat source feature data.

[0109] Through the collection and preprocessing of the above three types of data, radio frequency feature data, visual feature data, and thermal source feature data are generated respectively.

[0110] S23. Each search drone normalizes the radio frequency feature data, visual feature data and heat source feature data it collects to a unified numerical range to obtain normalized radio frequency features, normalized visual features and normalized heat source features.

[0111] In embodiments of the present invention, such as Figure 4As shown, due to the different dimensions and large differences in numerical ranges of radio frequency feature data (such as RSSI, DOA), visual feature data (such as the confidence level of suspected targets), and heat source feature data (such as abnormal heat source temperature), direct weighted fusion cannot be performed. Therefore, it is necessary to normalize the three types of feature data separately. Among them, the unified numerical range is a preset parameter that appears for the first time here, set to the range [0, 1]. The core purpose of the normalization process is to eliminate the difference in dimensions, so that the three types of feature data are in the same comparable range, ensuring the rationality and accuracy of subsequent weighted fusion.

[0112] The specific normalization process is as follows: For each type of feature data, the maximum and minimum values ​​of that type of data are calculated respectively. A linear normalization method is used to map each original feature data to the interval [0, 1], and finally the normalized radio frequency features, normalized visual features, and normalized thermal source features are obtained respectively. For example, the original value range of radio frequency signal strength (RSSI) is mapped to [0, 1]. The stronger the signal strength, the closer the normalized radio frequency feature value is to 1. The visual recognition confidence is directly retained in the interval [0, 1] (if the original confidence exceeds this range, forced normalization is performed). The original value of the abnormal thermal source temperature and the background temperature difference is mapped to the interval [0, 1]. The larger the temperature difference, the closer the normalized thermal source feature value is to 1.

[0113] S24. Dynamically allocate radio frequency weights, visual weights, and heat source weights based on the current time, lighting conditions, and terrain environment.

[0114] In this embodiment of the invention, radio frequency weight, visual weight, and thermal source weight (the sum of which is 1) are coefficients used to measure the importance of the three types of normalized feature data. These are dynamically assigned by the collaborative decision control module using a self-finding algorithm based on real-time environmental information (time, illumination, terrain), ensuring that the weighted fusion result closely matches the actual search environment and improving the accuracy of the comprehensive confidence score. The collaborative decision control module is the core module of the search drone's onboard processing system, responsible for running distributed search planning algorithms, multimodal information fusion, and target weighted scoring algorithms, adapting to multimodal data fusion and search decision requirements.

[0115] Furthermore, step S24 may include the following sub-steps:

[0116] S241. When it is detected that the current environment is a clear and open daytime, set the visual weight to the highest weight, the radio frequency weight to the second highest weight, and the heat source weight to the lowest weight.

[0117] In this embodiment of the invention, under clear, open daytime conditions with ample sunlight and no significant obstructions, the visible light images captured by the visual perception unit have high clarity, resulting in the highest recognition accuracy of the deep learning target detection model. Therefore, the visual weight is set to the highest weight (e.g., 0.5). Radio frequency signals are less affected by environmental interference and can assist in locating the lost drone; therefore, the radio frequency weight is set to the second highest weight (e.g., 0.3). At this time, the ambient background temperature is uniform, and the thermal imaging unit's advantage in identifying abnormal heat sources is not significant; therefore, the heat source weight is set to the lowest weight (e.g., 0.2), achieving environmentally adaptive weight allocation.

[0118] S242. When it is detected that the current environment is nighttime or dense fog, set the heat source weight to the highest weight, the radio frequency weight to the second highest weight, and the visual weight to the lowest weight.

[0119] In this embodiment of the invention, in nighttime or dense fog environments, insufficient light and low visibility result in poor clarity of the visible light images captured by the visual perception unit, significantly reducing target recognition accuracy. Therefore, the visual weight is set to the lowest weight (e.g., 0.1). The thermal imaging unit is unaffected by lighting conditions and can clearly identify the temperature difference between the lost drone and the environmental background, demonstrating a significant advantage in identifying abnormal heat sources. Therefore, the heat source weight is set to the highest weight (e.g., 0.6). Radio frequency signals are less affected by interference in nighttime and dense fog environments and can stably locate signal sources. Therefore, the radio frequency weight is set to the second highest weight (e.g., 0.3) to ensure accurate fusion of multimodal data even in low visibility environments.

[0120] S243. When it is detected that the current environment is a jungle or dense vegetation, set the radio frequency weight to the highest weight, the visual weight to the second highest weight, and the heat source weight to the lowest weight.

[0121] In this embodiment of the invention, in jungle or densely vegetated environments, the vegetation severely obstructs the line of sight of the visual perception unit, making it difficult to clearly capture the missing drone in visible light images. Therefore, the visual weight is set to the second highest weight (e.g., 0.2). Vegetation strongly blocks heat radiation, affecting the accuracy of abnormal heat source identification by the thermal imaging unit. Therefore, the heat source weight is set to the lowest weight (e.g., 0.1). Radio frequency signals can penetrate vegetation obstruction, transmit stably, and be captured by the radio frequency detection unit, resulting in the highest signal positioning accuracy. Therefore, the radio frequency weight is set to the highest weight (e.g., 0.7) to meet the search requirements in densely vegetated environments.

[0122] S25. Based on radio frequency weight, visual weight, and thermal source weight, the normalized radio frequency features, normalized visual features, and normalized thermal source features are weighted and summed to calculate the comprehensive confidence score corresponding to each search drone.

[0123] In this embodiment of the invention, the collaborative decision control module runs a multimodal information fusion and target weighted scoring algorithm. Based on the radio frequency weight, visual weight, and heat source weight dynamically allocated in step S24, it performs a weighted summation operation on the normalized radio frequency features, normalized visual features, and normalized heat source features obtained in step S23. The specific calculation process is as follows:

[0124] Overall confidence score = Normalized radio frequency features × Radio frequency weights + Normalized visual features × Visual weights + Normalized thermal source features × Thermal source weights

[0125] The overall confidence score ranges from [0, 1], consistent with the normalized feature. A value closer to 1 indicates a higher probability of a missing drone in the area; a value closer to 0 indicates a lower probability. During the search operation, each search drone calculates its overall confidence score in real time and simultaneously transmits the score and corresponding location information back to the ground control station, providing accurate data support for subsequent suspected target confirmation and emergency decision-making.

[0126] Step 104: When the overall confidence score exceeds the preset threshold, the search drone that has calculated the overall confidence score is taken as the discoverer, and the discoverer takes the current perception mode with the highest weight in the multimodal perception data as the dominant perception mode.

[0127] It should be noted that the preset threshold supports two adjustment methods: operators can manually set a fixed value based on the on-site search scenario, or the system can automatically and dynamically adjust it based on historical search data and real-time environmental conditions. In daily actual search operations, 0.6 is generally used as the typical value for this threshold.

[0128] In this embodiment of the invention, each search drone continuously performs a comprehensive search within its assigned sub-area, with the onboard processor constantly monitoring and comparing its real-time calculated comprehensive confidence score. Once the system detects that the current comprehensive confidence score exceeds a preset threshold, it determines that there are valid search clues related to the missing drone in the current area, and simultaneously locks onto the search drone that completed the calculation of that set of comprehensive confidence scores, officially identifying it as the discoverer.

[0129] Once the detection unit is identified, the collaborative decision control module on the detection unit automatically starts the dynamic weight self-finding algorithm. It retrieves the three sets of weight parameters previously allocated based on the real-time environment: radio frequency weight, visual weight, and heat source weight. It then compares the values ​​of the three sets of weights horizontally, selects the weight type with the highest current percentage, and selects the one with the highest current priority from the three types of multimodal sensing data as the dominant sensing mode.

[0130] After selecting the dominant perception mode, the dominant navigation vector is confirmed, using the data feature with the highest weight as the core reference for suspect target identification. If the radio frequency (RF) feature has the highest weight, the signal source direction angle calculated by the RF detection unit is directly used as the dominant navigation vector. If the visual or thermal source features have higher weights, the offset between the coordinates of the suspected feature pixels within the image or thermal imaging and the image center point is first calculated. Then, coordinate transformation is performed based on the real-time pose of the search drone to accurately calculate the azimuth and elevation angles of the target relative to the drone. This is used as the dominant navigation vector to provide accurate directional guidance for subsequent probability distribution model building.

[0131] Step 105: The detection drone generates an initial probability heatmap based on the dominant perception mode, and updates the probability heatmap in real time by integrating newly acquired multimodal perception data during the approach to the target to obtain the target probability heatmap. At the same time, it broadcasts a strong clue alarm to the nearby search drones.

[0132] In an embodiment of the present invention, Figure 5 This is a schematic diagram of the main process of strong clue triggering, dominant navigation vector confirmation and probability heatmap generation provided in the embodiments of the present invention. It fully demonstrates the entire process of comprehensive confidence score threshold determination, initial probability heatmap construction, data fusion and update, time decay diffusion, strong clue alarm broadcasting and navigation point generation. Figure 6 This is a schematic diagram of the probability heatmap iterative update and uncertainty propagation sub-process provided in the embodiments of the present invention, which shows in detail the technical details of new data acquisition, observation model construction, Bayesian update, uncertainty propagation and heatmap output.

[0133] After determining the dominant perception mode, the detection drone first uses its current location as the starting point and constructs a two-dimensional polar coordinate probability distribution model along the confirmed dominant navigation vector. Following the rule that the closer the target is to its own location and the higher the degree of alignment between its travel direction and the dominant navigation vector, the higher the probability of the target's presence, forming a corridor-like target search area with probability decay characteristics. Based on this distribution pattern, and according to the type of dominant perception mode, the detection drone constructs a corresponding initial probability heatmap, also known as an intermediate probability heatmap. Subsequently, it directionally approaches the target along the high-probability areas of the heatmap. During the mobile operation, the drone continuously collects new perception data, integrates the new data through Bayesian update rules, and introduces an uncertainty diffusion mechanism to simulate the decay of clues over time, correcting the probability distribution in real time to obtain the target probability heatmap. Simultaneously, it broadcasts strong clue alerts to neighboring drones through the neighbor communication module, achieving cluster-based collaborative search.

[0134] Furthermore, step 105 may include the following sub-steps:

[0135] S31. If the dominant sensing mode is radio frequency, the detector generates an initial probability heatmap of a preset shape with the direction of arrival of the radio frequency signal as the axis, and uses the initial probability heatmap of the preset shape as an intermediate probability heatmap.

[0136] In this embodiment of the invention, when the dominant sensing mode is radio frequency, the corresponding Figure 5 The radio frequency signal is the dominant clue branch and Figure 6 The initial heatmap construction node for radio frequency (RF) signals uses the confirmed dominant navigation vector type as the basis for building the heatmap model. When the RF signal is the dominant clue, the probability distribution presents two preset shapes: a cone or a fan shape with the signal arrival direction as the axis. The farther away from the detection device, the lower the probability of the target's presence in the area. Simultaneously, the higher the RSSI signal strength value collected by the RF detection unit, the higher the initial probability value is set for the corresponding axis region, intuitively reflecting the directional effect of signal strength on the target's location. The shape radiation angle is a preset parameter appearing for the first time in this step and can be flexibly adjusted according to the degree of environmental obstruction and signal transmission loss, with a typical value of ±30°. The probability distribution follows the rule that the closer to the central axis and the closer to the detection device, the higher the probability of the target's presence, gradually decreasing radially outward along the axis. The generated initial probability heatmap with the preset shape is directly used as the intermediate probability heatmap for subsequent updates and optimizations.

[0137] S32. If the dominant perception mode is visual or thermal infrared, the detection machine calculates the target azimuth angle based on the pixel position of the suspected target in the image and the current pose of the detection machine, generates a Gaussian initial probability heatmap centered on the target's calculated coordinates, and uses the Gaussian initial probability heatmap as an intermediate probability heatmap.

[0138] In this embodiment of the invention, when the dominant sensing mode is visual or thermal infrared, the corresponding Figure 5 Mid-range visual thermal imaging signals are the dominant cue branches and Figure 6 The initial heatmap construction node for the mid-vision thermal infrared class, and the specific data generation and heatmap construction process are as follows:

[0139] 1. Coordinate calculation: The detection unit extracts the pixel position of the suspected target in the image collected by the visual perception unit or thermal imaging unit, and combines it with its current pose, that is, the position and attitude data output by the high-precision pose estimation module, to calculate the ground geographic coordinates corresponding to the suspected target through coordinate transformation, and simultaneously calculates the azimuth angle of the target relative to the detection unit.

[0140] 2. Heatmap Generation: Using the calculated target geographic coordinates as the center, a Gaussian distribution model is constructed to generate an initial Gaussian probability heatmap. The standard deviation of the Gaussian distribution is a preset parameter appearing for the first time in this step. Its value is negatively correlated with the confidence level of visual thermal imaging recognition. The higher the recognition confidence, the smaller the standard deviation, the sharper the Gaussian peak, and the more concentrated the probability distribution range; conversely, the lower the recognition confidence, the larger the standard deviation, the gentler the Gaussian peak, and the wider the probability distribution range.

[0141] 3. Determining the intermediate heatmap: The generated Gaussian initial probability heatmap is directly used as the intermediate probability heatmap for subsequent iterative updates.

[0142] S33. The detector approaches the high-probability area in the intermediate probability heatmap, continuously collecting new multimodal sensing data during the approach process.

[0143] In this embodiment of the invention, corresponding Figure 5 In the process of finding high-probability cells in the heatmap as navigation points to approach the target, and... Figure 6 China Mobile collects new data nodes during the process. The discovery machine identifies the cell with the highest probability in the intermediate probability heat map through an algorithm, uses the geographical coordinates of the cell as the next navigation point, and approaches the target along the direction of the dominant navigation vector to ensure that the direction of travel always points to the area of ​​high probability of the suspected target.

[0144] Throughout the approximation process, the airborne multimodal sensing module remains continuously operational, synchronously acquiring new radio frequency signals, visible light images, and thermal infrared data. It also performs preprocessing such as noise reduction and feature extraction on the acquired raw data to generate new observational evidence, namely new multimodal sensing data, including new DOA measurements, new suspected target pixel locations, and new abnormal heat source data, providing sufficient measured data support for subsequent iterative updates of the probabilistic heatmap.

[0145] S34. The discovery machine uses a Bayesian update rule to fuse new multimodal sensing data into the intermediate probability heatmap, and introduces an uncertainty diffusion mechanism to simulate the decay of clues over time, and updates the target probability heatmap in real time.

[0146] In this embodiment of the invention, the two core logics—the Bayesian update rule and the uncertainty diffusion mechanism—correspond to... Figure 5 Heat map updates, uncertainty diffusion process and Figure 6 The process involves constructing a sensor observation model, updating the probability heatmap based on the sensor observation model, and propagating uncertainty across all nodes based on the motion model. The specific construction process of the target probability heatmap is as follows:

[0147] 1. Construct observation probability models: For newly acquired multimodal data, construct corresponding observation models to quantify the directional nature of the new data to the target location and clarify the actual probability of observing the current evidence when the target is at the corresponding coordinate point; the new radio frequency DOA measurement value corresponds to a normal distribution with the mean of the value, and the new visual and thermal imaging recognition results correspond to a Gaussian distribution centered on the pixel back-projection geographic coordinates.

[0148] 2. Bayesian Update Data Fusion: The observation model is fused into the intermediate probability heatmap using the Bayesian update rule. The calculation formula is as follows:

[0149] ;

[0150] in, This is the posterior probability (updated probability); Proportional to is a simplified expression of Bayes' theorem; For likelihood probability (observation model); This represents the prior probability (historical probability). The target variable is the one being searched for. For the observed evidence variable, i.e., confidence, the new probability of a cell is proportional to the product of the probability of the new observed evidence and the original prior probability. After fusion, the probability values ​​of the entire heatmap are normalized to ensure that the sum of the probabilities of all cells is 1, so that the probability of the new evidence-supported area increases and the probability of the contradictory area is suppressed.

[0151] 3. Uncertainty diffusion mechanism: To simulate the actual search scenario where clues naturally cool down and weaken over time, a Gaussian smooth convolution kernel is introduced as a preset parameter. Typically, a 3×3 convolution kernel is used to complete the probability diffusion processing. When no new observation data is added, the probability peak inside the heat map gradually widens and flattens, completing the clue decay simulation. Finally, iteratively generating a target probability heat map that accurately fits the on-site environment.

[0152] S35. While updating the target probability heatmap in real time, the detection aircraft broadcasts a strong clue alert to all nearby search drones within the communication range.

[0153] In this embodiment of the invention, corresponding Figure 5 In the process of sending strong clue alerts to nearby drones, the discovering drone, while iteratively updating the target probability heatmap, pushes standardized strong clue alerts to all nearby search drones within its communication coverage area, relying on its onboard neighbor communication module. The strong clue alert includes the sending device identifier, the device's real-time high-precision coordinates, the real-time updated target probability heatmap, the data record timestamp, and the currently determined dominant clue type. After receiving the information, nearby devices can quickly adjust their search plans, move closer to high-probability search areas, and achieve cluster collaborative positioning.

[0154] Step 106: Nearby search drones that receive strong clue alerts autonomously decide whether to participate in collaborative confirmation based on their own status. Search drones that participate in the collaboration converge towards the center of the target probability heatmap to confirm the target.

[0155] In the embodiments of the present invention, please refer to Figure 7 , Figure 7 This is a flowchart illustrating the process of a proximity search UAV receiving strong clue alerts, making autonomous decision-making and collaborative participation, and multi-UAV collaboratively approaching and confirming the target, as provided in this embodiment of the invention. It fully presents the entire process of data analysis, state assessment, participation decision-making, path planning, encirclement and approach, hovering identification and confirmation, and false alarm regression. At the same time, the high-precision pose estimation module equipped in this solution provides position support for swarm collaborative flight and accurate positioning throughout the process. When the GPS signal is valid, the pose calculation is completed using a GPS and inertial measurement unit fusion positioning mode. In the scenario where the GPS signal fails, the device can automatically switch to an independent positioning mode based on lidar real-time positioning and map building technology or visual odometry, ensuring accurate and stable UAV flight pose estimation throughout the process and laying a solid positioning foundation for collaborative confirmation operations.

[0156] After all nearby search drones within the communication range successfully receive the strong clue alarm data packet pushed by the detection aircraft, they rely on the onboard collaborative decision control module to retrieve their own operational status data, combine it with the target-related information in the alarm to complete a comprehensive evaluation, and autonomously determine whether to join the collaborative confirmation queue. Search drones that determine to participate in the collaborative confirmation work will uniformly plan a reasonable flight route and converge from multiple directions toward the center of the target probability heat map, ultimately completing the identification and confirmation of the suspected missing drone target on the ground.

[0157] Furthermore, step 106 may include the following sub-steps:

[0158] S41. Upon receiving a strong clue alert, the nearby search drone extracts the high-precision coordinates of the discoverer, the target probability heatmap, and the dominant clue type from the strong clue alert, and obtains its distance from the center of the target probability heatmap, remaining battery power, and sensor status parameters.

[0159] In this embodiment of the invention, after the proximity search UAV completes the reception of the strong clue alarm data packet, it first parses and extracts the information stored in the data packet line by line to accurately obtain three types of core external reference information: the high-precision coordinate data of the discovering UAV recorded in the alarm, the complete target probability heat map distribution data, and the dominant clue type corresponding to the strong clue in this search. At the same time, the UAV simultaneously starts its own status data collection, calculates the actual spatial distance between its current position and the center point of the target probability heat map based on the airborne positioning system, reads the real-time remaining power value of the UAV through the power monitoring module, and then checks the real-time operating status of all airborne sensors such as the radio frequency detection unit, visual perception unit, and thermal imaging unit through the device self-test program. It fully collects the three basic parameters required for autonomous decision-making: distance, remaining power, and sensor status parameters, providing all the raw data for subsequent collaborative participation calculations.

[0160] S42, The cooperative participation degree is calculated by the nearby search drone based on distance, remaining battery power and sensor status parameters.

[0161] In this embodiment of the invention, the collaborative decision control module built into the proximity search UAV retrieves the distance parameters, remaining battery power parameters, and sensor status parameters collected by S41, and calculates the corresponding collaborative participation value of the UAV according to the preset comprehensive evaluation calculation logic. The collaborative participation value is an evaluation value that quantifies the comprehensive ability of the UAV to participate in the collaborative confirmation task. The closer the UAV is to the center of the target probability heat map, the more sufficient the onboard remaining battery power reserve, and the more stable and reliable the operation of various sensing sensors, the higher the calculated collaborative participation value. Conversely, the collaborative participation value is lower, which intuitively reflects the adaptability and operational feasibility of a single UAV participating in the collaborative confirmation work.

[0162] S43. If the degree of collaborative participation exceeds the preset participation threshold, the nearby search drones will autonomously decide to participate in the collaborative confirmation.

[0163] In this embodiment of the invention, the preset participation threshold can be flexibly adjusted according to the urgency of the on-site search task and the overall operational layout of the swarm of drones. In daily on-site search operations, a unified fixed value can be used as a general judgment standard. The drone compares its self-calculated collaborative participation value with the preset participation threshold. Once the collaborative participation value exceeds the preset participation threshold, it means that all operating conditions of the drone meet the requirements for collaborative operation, and the drone will make an autonomous decision to formally join the swarm collaborative confirmation queue to participate in the target verification work. If the collaborative participation value does not reach the preset participation threshold, the drone will abandon participation in this collaborative confirmation task and continue to adhere to its initially assigned search area to perform regular full-coverage search operations, thus fully realizing the intelligent autonomous decision-making of the drone.

[0164] S44. The participating search drones autonomously plan a conflict-free flight path and converge towards the center of the target probability heat map.

[0165] In this embodiment of the invention, all search drones successfully identified for collaborative confirmation independently plan their own dedicated flight trajectories from their current location to the target's center position, based on their real-time high-precision location information and the coordinates of the target's probability heatmap center. Simultaneously, multiple participating drones exchange flight route data in real time via a neighboring drone communication module, negotiating and adjusting their flight headings and altitudes to effectively avoid conflicts such as overlapping or intersecting flight paths. Ultimately, this allows multiple drones to smoothly and orderly approach and converge towards the target's probability heatmap center position from different directions and angles within the search area, successfully completing the swarm encirclement and convergence operation.

[0166] S45. The first search drone to arrive at the preset confirmation range hovers and uses a high-magnification zoom camera to take multi-angle pictures of the suspected target for identification, thus completing the target confirmation.

[0167] In this embodiment of the invention, the preset confirmation range refers to the preset spatial range parameter that appears for the first time in this step. Specifically, it refers to the circular low-altitude operating area centered on the center of the target probability heatmap, which is the limited area for conducting close-range target verification. During the multi-drone collaborative approach process, the search drone that first flies to the center of the preset confirmation range will immediately adjust its flight attitude to maintain a stable low-altitude hovering state. Subsequently, it will mobilize the onboard gimbal equipment in conjunction with the high-magnification zoom camera equipment to conduct multi-angle, multi-magnification magnified real-scene shooting of the area where the suspected target is located on the ground. Combined with the airborne deep learning recognition model, the suspected target will be intelligently identified. Alternatively, it can be combined with remote manual verification by ground personnel to complete the final confirmation of the missing drone target.

[0168] Furthermore, such as Figure 7 As shown, if the drone continues to take pictures and identify within a preset fixed time period but still cannot successfully search for and confirm the missing drone target, the system will automatically determine that this strong clue alarm is a false alarm event. The search drone performing the verification task will end the hovering verification state, return to its initially assigned exclusive search area, and continue to perform regular search operations. Once the missing drone target is successfully and accurately identified and confirmed, the equipment will immediately trigger the subsequent debris search and mapping related operation process.

[0169] Step 107: Once any participating search drone confirms the target, the remaining participating search drones automatically switch to debris search mode, search for scattered debris parts, and transmit the collected debris data back to the ground control station. The ground control station then integrates all the transmitted debris data to generate a visualized debris distribution map.

[0170] In this embodiment of the invention, once any search drone participating in the collaborative confirmation operation successfully verifies and locks onto the target of the missing drone, it will immediately activate the cluster message broadcasting mechanism. All other collaborative drones that are converging towards the target center will simultaneously receive the relevant notification, automatically terminate the encirclement and convergence flight action, and switch the overall operation mode to a dedicated debris search operation mode. The drone will then conduct debris search work in the surrounding area with the confirmed main target location of the missing drone as the core, simultaneously collect debris-related data and transmit it back to the ground control station terminal in real time. Finally, relying on the built-in functions of the ground control station, all debris data will be summarized and organized, and a visual debris distribution map will be automatically generated.

[0171] Furthermore, step 107 may include the following sub-steps:

[0172] S51. Once any participating search drone confirms the target, the search drone that confirmed the target immediately broadcasts a target confirmation message containing the target's precise coordinates and image evidence to the entire cluster.

[0173] In this embodiment of the invention, the search drone that successfully completes the target confirmation work immediately sends a standardized target confirmation message to the entire drone search cluster through the onboard neighbor communication module. The message data packet is strictly edited and composed according to the predetermined format. The data packet is fully labeled with the specific message type, the unique identification number of the target confirmation device, the high-precision coordinates of the missing drone target on the ground, and valid confirmation evidence such as on-site real-scene images and videos of the target. This ensures that all operating drones in the cluster can accurately obtain complete and valid information about the target lock in the first time, providing a unified instruction basis for the subsequent switching of the overall operation mode.

[0174] S52. After receiving the target confirmation message, the other search drones participating in the collaboration automatically terminate their convergence behavior toward the center of the target probability heat map and switch to an extended scanning mode centered on the confirmed target point.

[0175] In this embodiment of the invention, all search drones participating in the collaborative confirmation operation, except for the confirmed target device, immediately cease all flight and convergence actions toward the center of the target probability heatmap after successfully receiving the target confirmation message pushed by the entire cluster. The onboard operation control program automatically switches the working mode, and the entire drone officially enters the wreckage search operation state. All drones uniformly use the confirmed target point of the missing drone that has been verified on-site as the central reference point, automatically retrieve and start multiple preset extended scanning flight paths. Commonly used paths include Archimedes spiral paths, concentric circle scanning paths, and other full-coverage search routes. Relying on the standardized extended scanning mode, a large-scale field area around the target point is thoroughly searched without blind spots, comprehensively searching for various debris components scattered after the missing drone crashed.

[0176] S53. During the extended scanning process, each search drone collects the precise coordinates of the debris components and takes images to generate debris data, which is then transmitted back to the ground control station in real time.

[0177] In this embodiment of the invention, all search drones in extended scanning mode, upon discovering various drone wreckage components scattered on the ground during their full-domain search flights, immediately initiate information collection. First, relying on their onboard high-precision pose estimation module, they accurately locate the precise latitude and longitude coordinates of the wreckage components. Then, they activate their onboard imaging equipment to capture clear, real-world images of the wreckage components, preserving the image data. The precise coordinate information and the real-world image information of the wreckage are integrated and summarized to form complete and standardized wreckage data. After completing the processing of a single set of wreckage data, the drone uploads all the wreckage data back to the ground control station terminal equipment in real time via a long-distance wireless communication link, ensuring that the ground station can grasp the basic information on the distribution of wreckage in the field as soon as possible.

[0178] S54. The ground control station integrates the target confirmation information and debris data transmitted back by multiple aircraft, marks the main debris point and each scattered debris point on the electronic map, and generates a visualized debris distribution map.

[0179] In this embodiment of the invention, the global monitoring and interaction module on the ground control station uniformly receives target confirmation information and various types of field debris data transmitted back by all search drones in the cluster. Staff do not need to manually organize and compile the data; the system can automatically complete the aggregation and integration of multi-source data. The system uses the confirmed crash location of the missing drone as the main debris reference point, marking the point in the system's built-in electronic map interface. Simultaneously, it marks the coordinates of all scattered debris components collected and transmitted by all drones on the map. It can also automatically draw lines connecting the main debris point to each scattered debris point, ultimately automatically integrating and generating an intuitive, clear, and complete visualized debris distribution map. This map presents a clear and complete picture of the overall field distribution of the missing drone's debris, providing complete and intuitive data support for subsequent field debris recovery operations and fault cause analysis.

[0180] This invention completes a fully closed-loop collaborative search process for a missing UAV swarm, encompassing initial region delineation, distributed task allocation, multimodal perception fusion, dynamic self-homing localization, multi-UAV collaborative confirmation, and debris distribution mapping. It achieves the following technical effects:

[0181] 1. Significantly improved operational efficiency: Parallel regional search is achieved through multi-machine collaborative operation, greatly shortening the initial large-scale search time; The built-in weighted scoring mechanism based on multimodal perception fusion and the dynamic weight self-finding algorithm can intelligently reason and judge suspected clues and actively approach them, significantly improving the efficiency and success rate from the discovery of clues to the final confirmation of the target, especially suitable for search tasks in large-area and complex environments.

[0182] 2. Strong adaptability to complex environments: The search drone is equipped with a multimodal perception module that integrates radio frequency, vision and thermal infrared perception modes. The three modes complement each other in different environments: radio frequency signals can penetrate vegetation obstruction, thermal imaging is not affected by light and partial obstruction, and visual perception provides high-precision target recognition. This enables the system to maintain stable and efficient search capabilities even in complex environments where single perception methods fail, such as at night, in dense fog, rain, snow, and dense forests.

[0183] 3. High system reliability: The decentralized collaborative architecture avoids single point of failure. The failure of individual drones or communication interruption will not affect the execution of the overall search cluster task, ensuring the continuity and stability of the search operation.

[0184] 4. High degree of autonomy: From initial sub-region allocation, clue evaluation, decision-making to final collaborative confirmation, the entire operation process is completed autonomously by the drone swarm through built-in algorithms, which greatly reduces the reliance on manual operation and lowers the workload and professional skill requirements of operators.

[0185] 5. High value for subsequent analysis: It is equipped with collaborative confirmation and mapping functions. After the target location is confirmed, it can automatically generate a visual debris distribution map. It not only successfully completed the location and search mission of the missing drone, but also provided complete on-site spatial data for the subsequent investigation of the cause of the accident. It has extremely high engineering application and analysis value.

[0186] Please see Figure 8 , Figure 8 This is a structural block diagram of a collaborative search system for a missing drone swarm provided in an embodiment of the present invention.

[0187] This invention provides a collaborative search system for a swarm of missing drones, comprising:

[0188] The area division module 801 is used by the ground control station to divide the initial search area into multiple sub-regions based on the identity information and last known location of the missing UAV;

[0189] The task allocation module 802 is used by the search cluster to autonomously allocate each sub-region to each search drone in the search cluster through a decentralized distributed auction mechanism.

[0190] The perception fusion module 803 is used by each search drone to perform search in the assigned sub-area, synchronously collect multimodal perception data, and perform environmental adaptive dynamic weighted fusion on the collected multimodal perception data to calculate the comprehensive confidence score in real time.

[0191] The clue determination module 804 is used to identify the search drone that has calculated the comprehensive confidence score as the discoverer when the comprehensive confidence score exceeds a preset threshold, and the discoverer will take the perception mode with the highest current weight in the multimodal perception data as the dominant perception mode.

[0192] The heat map push module 805 is used to generate an initial probability heat map by the detection drone based on the dominant perception mode, and to update the probability heat map in real time by integrating newly collected multimodal perception data during the approach to the target to obtain the target probability heat map, while broadcasting strong clue alarms to nearby search drones.

[0193] The collaborative verification module 806 is used for nearby search drones that receive strong clue alarms to autonomously decide whether to participate in collaborative verification based on their own status. Search drones that participate in the collaboration converge towards the center of the target probability heat map to confirm the target.

[0194] The debris mapping module 807 is used so that when any participating search drone confirms the target, the remaining participating search drones automatically switch to debris search mode, search for scattered debris parts, and transmit the collected debris data back to the ground control station. The ground control station then integrates all the transmitted debris data to generate a visual debris distribution map.

[0195] Since the above is a system corresponding to a method for collaboratively searching for a swarm of missing drones, its implementation principle is the same as that of a method for collaboratively searching for a swarm of missing drones. For the sake of convenience and brevity, those skilled in the art can clearly understand that the specific working process of the system and modules described above can be referred to the corresponding process in the aforementioned method embodiments, and will not be repeated here.

[0196] Please see Figure 9 , Figure 9 This is a structural block diagram of an electronic device provided in an embodiment of the present invention.

[0197] An electronic device according to an embodiment of the present invention includes: a memory 901 and a processor 902. The memory 901 stores a computer program. When the computer program is executed by the processor 902, the processor 902 performs the collaborative search method for missing drone swarms as described in any of the above embodiments.

[0198] Memory 901 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Memory 901 has storage space 903 for program code 913 for performing any of the method steps described above. For example, storage space 903 for program code may include various program codes 913 for implementing the various steps in the methods described above. These program codes may be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When run by a computing processing device, this code causes the computing processing device to perform the various steps in the methods described above. These program codes may be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When this code is run by a computing device, it causes the computing device to perform the various steps in the collaborative search method for the missing drone swarm described above.

[0199] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the collaborative search method for a cluster of missing drones as described in any of the above embodiments.

[0200] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0201] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0202] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0203] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0204] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0205] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A collaborative search method for a swarm of missing unmanned aerial vehicles (UAVs), characterized in that, include: The ground control station divided the initial search area into multiple sub-areas based on the missing drone's identity information and last known location; The search cluster autonomously allocates each sub-region to a search drone within the search cluster through a decentralized distributed auction mechanism; Each of the aforementioned search drones performs a search within its assigned sub-region, simultaneously collecting multimodal perception data, and performing environmentally adaptive dynamic weighted fusion on the collected multimodal perception data to calculate the comprehensive confidence score in real time; When the overall confidence score exceeds a preset threshold, the search drone that has calculated the overall confidence score will be used as the discoverer, and the discoverer will use the perception mode with the highest current weight in the multimodal perception data as the dominant perception mode. The detection drone generates an initial probability heatmap based on the dominant sensing mode, and updates the probability heatmap in real time by fusing newly acquired multimodal sensing data during the approach to the target to obtain a target probability heatmap. At the same time, it broadcasts a strong clue alarm to nearby search drones. Upon receiving the strong clue alarm, the nearby search drones autonomously decide whether to participate in collaborative confirmation based on their own status. Search drones that participate in the collaboration converge towards the center of the target probability heatmap to confirm the target. Once any participating search drone confirms the target, the remaining participating search drones automatically switch to debris search mode to search for scattered debris parts and transmit the collected debris data back to the ground control station. The ground control station then integrates all the transmitted debris data to generate a visual debris distribution map.

2. The collaborative search method for missing drone swarms according to claim 1, characterized in that, The step of the search cluster autonomously allocating each sub-region to each search drone in the search cluster through a decentralized distributed auction mechanism includes: After each search drone in the search cluster completes initialization, a temporary auctioneer is generated through a dynamic election mechanism. The temporary auctioneer selects an unassigned sub-region from the sub-region task list issued by the ground control station and broadcasts the auction message to the entire cluster. After receiving the auction message, each of the search drones calculates the task cost corresponding to the sub-region to be allocated based on its own status and the matching degree between the sub-region to be allocated, and generates a corresponding bid based on the task cost and sends it to the temporary auctioneer. The temporary auctioneer collects all bids from all bidders, selects the search drone with the best bid as the winner, and broadcasts the winning bid message to the entire cluster. Once the winning drone confirms receipt of the task, all search drones in the search cluster will uniformly mark the sub-area to be allocated as allocated. Repeat the auction process described above until all sub-regions in the sub-region task list have been allocated.

3. The collaborative search method for missing drone swarms according to claim 1, characterized in that, The steps of having each of the search drones perform a search within its assigned sub-region, simultaneously collect multimodal perception data, and perform environmentally adaptive dynamic weighted fusion of the collected multimodal perception data to calculate a comprehensive confidence score in real time include: Each of the aforementioned search drones performs a comprehensive search within its assigned sub-region using an extended spiral or a cooperative plowing path. During the entire process of performing a comprehensive search, each of the search drones simultaneously collects radio frequency signals, visible light images, and thermal infrared data, and preprocesses the collected raw data to generate radio frequency feature data, visual feature data, and thermal source feature data, respectively. Each of the search drones normalizes the radio frequency feature data, visual feature data and thermal source feature data it collects to a unified numerical range to obtain normalized radio frequency features, normalized visual features and normalized thermal source features. Based on the current time, lighting conditions, and terrain environment, dynamically allocate radio frequency weights, visual weights, and heat source weights; Based on the radio frequency weight, the visual weight, and the heat source weight, the normalized radio frequency features, the normalized visual features, and the normalized heat source features are weighted and summed to calculate the comprehensive confidence score corresponding to each of the search drones.

4. The collaborative search method for missing drone swarms according to claim 3, characterized in that, The step of dynamically allocating radio frequency weights, visual weights, and heat source weights based on the current time, lighting conditions, and terrain environment includes: When it is detected that the current environment is a clear and open daytime, the visual weight is set to the highest weight, the radio frequency weight is set to the second highest weight, and the heat source weight is set to the lowest weight. When it is detected that the current environment is nighttime or dense fog, the heat source weight is set to the highest weight, the radio frequency weight is set to the second highest weight, and the visual weight is set to the lowest weight. When it is detected that the current environment is a jungle or dense vegetation, the radio frequency weight is set to the highest weight, the visual weight is set to the second highest weight, and the heat source weight is set to the lowest weight.

5. The collaborative search method for missing drone swarms according to claim 1, characterized in that, The steps of generating an initial probability heatmap based on the dominant sensing mode by the detection drone, updating the probability heatmap in real time by fusing newly acquired multimodal sensing data during the approach to the target to obtain a target probability heatmap, and simultaneously broadcasting a strong clue alert to nearby search drones include: If the dominant sensing mode is radio frequency, the detection machine generates an initial probability heatmap of a preset shape with the direction of arrival of the radio frequency signal as the axis, and uses the initial probability heatmap of the preset shape as an intermediate probability heatmap. If the dominant perception mode is visual or thermal infrared, the detection machine calculates the target azimuth angle based on the pixel position of the suspected target in the image and the current pose of the detection machine, generates a Gaussian initial probability heatmap centered on the target's calculated coordinates, and uses the Gaussian initial probability heatmap as an intermediate probability heatmap. The detection machine approaches the high-probability region in the intermediate probability heatmap, continuously collecting new multimodal sensing data during the approach process; The discovery machine uses a Bayesian update rule to fuse new multimodal sensing data into the intermediate probability heatmap, and introduces an uncertainty diffusion mechanism to simulate the decay of clues over time, thereby updating the target probability heatmap in real time. While updating the target probability heatmap in real time, the detection aircraft broadcasts a strong clue alert to all nearby search drones within its communication range.

6. The collaborative search method for missing drone swarms according to claim 1, characterized in that, The step of the nearby search drones that receive the strong clue alarm autonomously deciding whether to participate in collaborative confirmation based on their own status, and the search drones participating in the collaboration converging towards the center of the target probability heatmap to confirm the target, includes: The nearby search drone that receives the strong clue alarm extracts the high-precision coordinates of the discoverer, the target probability heat map and the dominant clue type from the strong clue alarm, and obtains its own distance from the center of the target probability heat map, remaining battery power and sensor status parameters. The degree of collaborative participation is calculated by the nearby search drone based on the distance, the remaining battery power, and the sensor status parameters. If the degree of collaborative participation exceeds a preset participation threshold, the nearby search drone will autonomously decide to participate in the collaborative confirmation. The participating search drones autonomously plan conflict-free flight paths and converge towards the center of the target probability heatmap; The first search drone to arrive at the pre-defined confirmation range hovers and uses a high-magnification zoom camera to capture and identify the suspected target from multiple angles, thus confirming the target.

7. The collaborative search method for missing unmanned aerial vehicles (UAVs) according to any one of claims 1-6, characterized in that, The step of having any participating search drone confirm a target, and the remaining participating search drones automatically switching to debris search mode to search for scattered debris components and transmit the collected debris data back to the ground control station, and having the ground control station integrate all the transmitted debris data to generate a visualized debris distribution map, includes: Once any participating search drone confirms the target, the confirming search drone immediately broadcasts a target confirmation message containing the target's precise coordinates and image evidence to the entire cluster. Upon receiving the target confirmation message, the remaining search drones participating in the collaboration automatically terminate their convergence behavior toward the center of the target probability heatmap and switch to an extended scanning mode centered on the confirmed target point. During the extended scan, each search drone collects the precise coordinates of the debris components and takes images to generate debris data, which is then transmitted back to the ground control station in real time. The ground control station integrates target confirmation information and debris data transmitted back from multiple aircraft, marks the main debris point and various scattered debris points on the electronic map, and generates a visualized debris distribution map.

8. A collaborative search system for a swarm of missing unmanned aerial vehicles (UAVs), characterized in that, include: The area division module is used by the ground control station to divide the initial search area into multiple sub-regions based on the identity information and last known location of the missing UAV; The task allocation module is used by the search cluster to autonomously allocate each of the sub-regions to each search drone in the search cluster through a decentralized distributed auction mechanism; The perception fusion module is used by each of the search drones to perform searches in the assigned sub-areas, synchronously collect multimodal perception data, and perform environmental adaptive dynamic weighted fusion on the collected multimodal perception data to calculate the comprehensive confidence score in real time. The clue determination module is used to identify the search drone that has calculated the comprehensive confidence score as the discoverer when the comprehensive confidence score exceeds a preset threshold, and the discoverer will take the perception mode with the highest current weight in the multimodal perception data as the dominant perception mode. The heat map push module is used to generate an initial probability heat map by the discovery drone based on the dominant perception mode, and to update the probability heat map in real time by integrating newly acquired multimodal perception data during the approach to the target to obtain a target probability heat map, while broadcasting a strong clue alarm to nearby search drones. The collaborative verification module is used for nearby search drones that receive the strong clue alarm to autonomously decide whether to participate in collaborative verification based on their own status. Search drones that participate in the collaboration converge towards the center of the target probability heat map to confirm the target. The debris mapping module is used so that when any participating search drone confirms the target, the remaining participating search drones automatically switch to debris search mode, search for scattered debris parts, and transmit the collected debris data back to the ground control station. The ground control station then integrates all the transmitted debris data to generate a visual debris distribution map.

9. An electronic device, characterized in that, The system includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the collaborative search method for a swarm of missing drones as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the collaborative search method for missing drone swarms as described in any one of claims 1-7.

Images