An unmanned aerial vehicle intelligent inspection system and method based on multi-machine cooperation, and a readable storage medium

Abstract

The application provides a kind of unmanned aerial vehicle intelligent inspection system and method based on multi-machine cooperation and readable storage medium, belongs to unmanned aerial vehicle inspection field.Aiming at the problems of large bandwidth occupation, high delay and blind area of multi-machine change defense boundary in single machine inspection, the application adopts "cloud-edge-end" cooperative architecture: the cloud end divides target area into N main subareas, and generates boundary overlapping subarea along the public edge, and the overlapping area proportion is constrained to not more than the preset area proportion threshold;the unmanned aerial vehicle edge AI unit executes target detection at the end side, and only returns metadata message containing category, coordinate and confidence when detecting target, without returning original video stream;the cloud end executes time window alignment and spatial distance calculation on metadata, judges the same target, and then performs confidence weighted arbitration, and retains the optimal record to generate alarm.The application can reduce bandwidth occupation, eliminate boundary blind area and ensure alarm uniqueness, and is suitable for water area inspection and forest fire prevention scenes.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology for unmanned aerial vehicles (UAVs), and in particular to an intelligent inspection system, method, and readable storage medium for UAVs based on multi-UAV collaboration. Background Technology

[0002] Currently, unmanned aerial vehicle (UAV) intelligent inspection systems are widely used in scenarios such as waterway patrol, forest fire prevention, and environmental monitoring. The closest existing technology to this application typically employs a "single-unit patrol + centralized cloud analysis" working mode: a single UAV, equipped with a visible light or infrared camera, flies along a preset route, transmitting high-definition video streams in real time to the cloud via an image transmission module. A deep learning model on the cloud GPU server then performs frame extraction, decoding, and target detection on the video stream, ultimately displaying the analysis results and issuing warnings at the command and control terminal. In this mode, the UAV only handles image acquisition, with all data processing burden concentrated on the network transmission link and the cloud.

[0003] However, the aforementioned existing technologies have the following technical drawbacks: First, high-definition video streams involve large amounts of data, and in complex electromagnetic environments such as the wild or water, wireless links are easily blocked and interfered with, leading to video stuttering and frame drops, causing gaps or missed detections in cloud analysis; Second, each stage of video encoding, network transmission, cloud decoding, and inference takes time, and there is often a delay of several to tens of seconds from the detection of an anomaly to the generation of an early warning; Third, when a single machine operates continuously along a route, once a fault occurs or the signal is lost, a monitoring vacuum will appear on the entire route, and when multiple machines are switching positions, it is very easy for targets to cross boundaries and be missed at the area cutting points.

[0004] Therefore, it is necessary to study an intelligent inspection system, method, and readable storage medium based on multi-drone collaboration for unmanned aerial vehicles (UAVs) to address the shortcomings of existing technologies and to solve or mitigate one or more of the aforementioned problems. Summary of the Invention

[0005] In view of this, the present invention provides an intelligent inspection system, method, and readable storage medium for unmanned aerial vehicles (UAVs) based on multi-UAV collaboration to solve the following technical problems existing in the prior art: Existing UAV inspection systems adopt a "single-UAV patrol + cloud-based centralized analysis" mode. The high-definition video stream data volume is large, and the wireless link is easily interfered with in complex electromagnetic environments such as the wild or water, resulting in video stuttering, frame loss, and gaps or missed detections in cloud analysis; each stage of video encoding, network transmission, cloud decoding, and inference takes time, and there is often a delay of several seconds to tens of seconds from the detection of an anomaly to the generation of an early warning; if a high-performance computing unit is installed on the UAV side to achieve local analysis, it will significantly increase the load on the aircraft and power consumption, compress the single inspection endurance time, and reduce the recognition accuracy due to heat dissipation limitations; when a single UAV operates continuously along a route, once a fault or signal loss occurs, a monitoring vacuum will appear for the entire route, and the technical problem of target cross-boundary missed identification is very likely to occur at the area cutting point when multiple UAVs are switching positions.

[0006] On the one hand, the present invention provides a drone intelligent inspection system based on multi-drone collaboration. The drone intelligent inspection system includes: a cloud processing center and a drone collaborative cluster. The drone collaborative cluster includes at least two drone nodes, each equipped with an edge AI computing unit. The cloud processing center divides the target area into several sub-regions and generates overlapping sub-regions, constraining the area ratio of the overlapping sub-regions to not exceed a preset area ratio threshold. Each of the aforementioned drone nodes performs target detection on the edge AI computing unit, and only sends back metadata messages containing target category, geographic coordinates and confidence level, without sending back the original video stream; The cloud processing center performs spatiotemporal alignment on multiple metadata messages, and after determining that they are the same target, performs confidence-weighted arbitration to retain the best record.

[0007] As described above and in any possible implementation, a further implementation is provided in which the cloud processing center, combining the nominal capacity of the single-unit battery, the wind speed prediction model, and the peak power consumption of the airborne edge AI computing unit, inversely calculates the effective coverage area threshold S of the single unit while retaining a safe redundancy of power. limit The target area is divided into N main sub-regions based on the number N drones in the drone collaborative swarm, and the area Si of each main sub-region is constrained to not exceed the specified value S. limit , where N is a positive integer; The cloud processing center extends bufferedly to both sides along the common edge of adjacent main sub-regions to generate boundary overlapping sub-regions, with a buffer extension distance D. buffer Satisfy D buffer = k·(σ rtk + σ hover ) + d fov , where σ rtk For RTK positioning accuracy, σ hover For drone hovering accuracy, d fov The horizontal coverage margin is calculated based on the cruise altitude and the camera's field of view, and k is the safety factor.

[0008] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein the cloud processing center is further configured to: calculate the ratio η of the total area of ​​the overlapping sub-regions to the total area of ​​the target region; if η exceeds the preset area ratio threshold, then initiate an iterative optimization loop: first, linearly reduce the buffer extension distance D at a ratio of 20% / η. bufferThe buffer polygon is regenerated; if simply reducing the distance still fails to meet the constraints, the granularity of the principal sub-region division is further adjusted or the curved common edges are straightened to reduce the overlapping area of ​​the buffer zone; in each iteration, it is evaluated whether the new scheme simultaneously satisfies that η does not exceed the preset area ratio threshold and the area Si of all principal sub-regions does not exceed S. limit The iteration continues until the optimal solution that satisfies all constraints is found; the preset area ratio threshold is 10% to 30%.

[0009] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the cloud processing center performs spatiotemporal alignment of multiple metadata messages by: performing time alignment on data transmitted back within the same time window based on a time window threshold, and determining multiple records whose geographical coordinate distance is less than a preset spatial threshold as the same target based on spatial distance calculation; the time window threshold is 1 second to 10 seconds, and the preset spatial threshold is 5 meters to 50 meters.

[0010] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the metadata message includes a keyframe thumbnail corresponding to the target; the geographic coordinates are WGS-84 geographic coordinates obtained based on image coordinate transformation; and each of the drone nodes is further configured to: when the edge AI computing unit does not detect the target, send a heartbeat packet to the cloud processing center at a set time interval, the heartbeat packet including the location information, remaining battery power, and operating status of the drone node.

[0011] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the cloud processing center is configured to: use a Kalman filter algorithm to process the geographic coordinate sequence of multiple records determined to be the same target, and complete the motion trajectory of the target when flying across UAV nodes in the boundary overlapping sub-region.

[0012] In addition to the aspects described above and any possible implementation, a further implementation is provided, wherein the cloud processing center is configured to: receive remaining battery power information and CPU / NPU utilization information sent by each of the drone nodes via heartbeat packets; when the remaining battery power of any drone node is lower than a preset battery power threshold, recalculate the remaining unpatched area of ​​the drone node, dynamically split the tasks of the remaining unpatched area and allocate them to nearby drone nodes with sufficient battery power, and control the drone node to return to its home location; the preset battery power threshold is 10% to 30% of the full battery power of the drone node.

[0013] In addition to the aspects and any possible implementations described above, a further implementation is provided in which each of the drone nodes is configured to: when the local detection confidence of the edge AI computing unit is lower than a preset confidence threshold, send a single-frame image to the cloud processing center and request the cloud processing center to perform secondary target detection confirmation on the single-frame image.

[0014] As described above and in any possible implementation, a multi-drone collaborative intelligent inspection method for unmanned aerial vehicles (UAVs) is further provided. Employing the aforementioned multi-drone collaborative intelligent UAV inspection system, the multi-drone collaborative intelligent UAV inspection method includes the following steps: S1, the cloud processing center combines the nominal capacity of a single battery, wind speed prediction models, and peak power consumption inversion of the airborne edge AI computing unit to calculate the effective coverage area threshold S. limit Based on the number of drones N, the target area is divided into N main sub-regions, and the area of ​​each main sub-region is constrained to not exceed S. limit The boundary overlapping sub-region is generated by adaptively buffering and extending along the common edge of the adjacent main sub-region. The area ratio of the boundary overlapping sub-region is constrained to not exceed the preset area ratio threshold. A bow-shaped or spiral scanning route is generated for each sub-region and route conflict detection is performed. The route tasks of each main sub-region and the boundary overlapping sub-region are sent to the corresponding UAV node. N is a positive integer. S2. All the aforementioned drone nodes take off in coordination and fly to the airspace above the assigned sub-area. The onboard edge AI computing unit performs target detection inference on the edge of the acquired images and only sends back metadata messages containing target category, geographic coordinates and confidence level to the cloud processing center, without sending back the original video stream. S3. The cloud processing center performs spatiotemporal alignment on the multiple metadata messages sent back, and after determining that they are the same target, performs confidence-weighted arbitration, retains the best record, and generates alarm information.

[0015] In accordance with the aspects described above and any possible implementation thereof, a computer-readable storage medium is further provided having a computer program stored thereon that, when executed by a processor, implements the method described above.

[0016] Compared with the prior art, the present invention can achieve the following technical effects: 1. By equipping each drone node with an edge AI computing unit, the edge AI computing unit is directly connected to the airborne camera via MIPI or USB interface and bypasses the image transmission encoding chip. Target detection and inference are completed directly at the source of acquisition. Only lightweight metadata messages containing target category, geographic coordinates, confidence level, and keyframe thumbnails are sent back to the cloud. This completely eliminates the need for raw video streaming at speeds of several megabits per second to tens of megabits per second, compressing communication bandwidth requirements to the level of hundreds of bytes. This eliminates video stuttering, frame dropping, and high latency caused by limited wireless link bandwidth at the physical transmission level, enabling the system to work stably in complex electromagnetic environments in the wild. End-to-end identification and warning latency can be compressed to less than 2 seconds.

[0017] 2. By dividing the target area into N main sub-regions based on the number of drones N, and extending along the common edge buffer of adjacent main sub-regions to generate boundary overlapping sub-regions, and by iteratively optimizing the constraint that the total area ratio η of the boundary overlapping sub-regions does not exceed a preset area ratio threshold, the blind spots at the regional boundaries caused by drone positioning drift and camera field of view deviation are accurately filled with minimal spatial redundancy, achieving full-domain zero-miss coverage. At the same time, since each drone is only responsible for a small sub-region, the proportion of the target to be detected in the field of view is increased and the background interference is reduced, so that the lightweight target detection model running INT8 quantization compression can still maintain a high target recall rate under the condition of limited edge computing power, breaking through the industry bottleneck that lightweight drones cannot perform high-precision real-time analysis.

[0018] 3. The cloud-based fusion analysis server performs deduplication of multi-source returned metadata based on time window alignment and spatial distance calculation, and adopts a confidence-weighted arbitration mechanism to retain the best record, filtering out duplicate alarms caused by multiple cross-shooting in overlapping sub-regions, thus ensuring the uniqueness and accuracy of alarms. Furthermore, the Kalman filter algorithm is used to complete the target's motion trajectory when flying across aircraft, eliminating the "flash" or "disappearance" of the target position caused by the switching of perspectives, and realizing a leap from "single-point perception" to "global cognition".

[0019] 4. By monitoring the remaining power of each drone node in real time through the cloud, a dynamic task redistribution mechanism is triggered when the power of any node falls below the preset power threshold. The tasks of the remaining unpatched areas are split and assigned to nearby nodes with sufficient power. The original nodes return to base in advance, ensuring the continuous and uninterrupted execution of inspection tasks under low power conditions of a single drone. This significantly enhances the overall robustness of the system and avoids monitoring vacuums caused by single-drone failures.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the technical effects described above at the same time. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments 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.

[0022] Figure 1 This is a schematic diagram of the overall architecture of an intelligent inspection system based on multi-machine collaboration provided in one embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the spatial relationship between the target region division and the boundary overlapping sub-regions provided in an embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of a multi-machine collaborative inspection and data processing process provided in one embodiment of the present invention. Detailed Implementation

[0025] To better understand the technical solution of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0026] It should be understood that the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0027] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “described,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. In the following description, references such as “cloud processing center 100,” “wireless communication link 200,” “drone collaborative swarm 300,” “drone node (301, 302…30X),” and “edge AI computing unit (301a, 302a…)” are used interchangeably with… Figures 1-3 The labels in the attached figures correspond one-to-one.

[0028] This invention provides an intelligent inspection system based on multi-machine collaboration, the intelligent inspection system based on multi-machine collaboration comprising: The cloud processing center has a built-in geographic information system module, a fusion analysis server, and an inspection database. The geographic information system module is used to acquire geographic boundary information of the target area, and the fusion analysis server is connected to the inspection database. A wireless communication link is provided, which is connected to the cloud processing center and is used to transmit control commands and identification result data between the cloud processing center and the UAV collaborative cluster. A drone collaborative swarm includes at least two drone nodes, each drone node is equipped with an airborne camera and an edge AI computing unit, the edge AI computing unit is directly connected to the airborne camera, and is used to perform target detection inference on the images captured by the airborne camera to obtain detection results; The cloud processing center is configured as follows: based on the number N of drones in the drone collaborative cluster, the target area is divided into N main sub-regions, and boundary overlapping sub-regions are generated by buffering and extending along the common edge of adjacent main sub-regions to both sides. The ratio η of the total area of ​​the boundary overlapping sub-regions to the total area of ​​the target area is constrained not to exceed a preset area ratio threshold. The coordinate information of each main sub-region and boundary overlapping sub-region is packaged into a flight path task and sent to the corresponding drone node via the wireless communication link, where N is a positive integer. Each drone node is configured as follows: the edge AI computing unit performs target detection inference on the acquired images at the edge. When a target is detected, it transmits a metadata message containing the target category, geographic coordinates, and confidence level back to the cloud processing center via the wireless communication link. The fusion analysis server is configured to perform time window alignment and spatial distance calculation on the multiple transmitted metadata messages. Multiple records within the same time window with geographic coordinate distances less than a preset spatial threshold are identified as the same target. A confidence-weighted arbitration is then performed on these records, and the record with the highest confidence level is stored in the inspection database. Through this "cloud-edge-device" three-level collaborative architecture, each drone node completes inference and transmits lightweight metadata only at the edge, completely eliminating high-bandwidth video stream transmission. Simultaneously, the boundary overlap design with area constraints eliminates blind spots at regional boundaries during multi-drone collaborative coverage. The cloud fusion analysis server deduplicates and aggregates multi-source transmitted data to ensure the uniqueness and accuracy of alarms.

[0029] The cloud processing center is further configured to: calculate the ratio η of the total area of ​​the overlapping sub-regions to the total area of ​​the target region; if η exceeds the preset area ratio threshold, iterative recalculation is performed by reducing the buffer extension distance or adjusting the shape of the main sub-region until η does not exceed the preset area ratio threshold. Through this iterative optimization mechanism, the system can control the redundant area within the minimum effective range while ensuring the integrity of boundary coverage, thus avoiding the ineffective consumption of cluster energy and computing power.

[0030] The preset area ratio threshold is 10% to 30%, preferably 20%; the buffer extension distance is the sum of the positioning error of the UAV node and the field of view coverage margin of the airborne camera. Setting the area ratio threshold to 20% is a constraint boundary determined after comprehensively considering coverage integrity and resource consumption; the buffer extension distance is the sum of the positioning error and the field of view coverage margin, which can accurately fill the potential blind spots caused by the UAV positioning drift and the camera field of view deviation.

[0031] The metadata message includes a keyframe thumbnail corresponding to the target; the geographic coordinates are WGS-84 geographic coordinates obtained based on image coordinate transformation. Attaching the keyframe thumbnail to the metadata message creates a chain of evidence for the target in the cloud, facilitating manual verification. Using the WGS-84 geographic coordinate standard ensures the comparability of coordinates transmitted from different UAV nodes within the same geographic reference system, providing a unified benchmark for spatial distance calculations in the cloud. The cloud processing center performs spatiotemporal alignment on multiple metadata messages by: aligning the data transmitted within the same time window based on a time window threshold, and determining multiple records with geographic coordinate distances less than a preset spatial threshold as the same target based on spatial distance calculations; the time window threshold is 1 to 10 seconds, and the preset spatial threshold is 5 to 50 meters.

[0032] Each of the aforementioned drone nodes is also configured to: when the edge AI computing unit does not detect a target, send a heartbeat packet to the cloud processing center at a set time interval; the heartbeat packet contains the drone node's location information, remaining battery power, and operating status; when the edge AI computing unit detects a target, it immediately triggers the return of the metadata message. This dual-mode communication mechanism of "abnormal triggering + timed heartbeat" occupies only a very low bandwidth to maintain status monitoring when no target is detected, and immediately triggers a return to ensure real-time early warning when a target is detected, achieving an optimal balance between communication bandwidth usage and response speed.

[0033] The fusion analysis server is also configured to use a Kalman filter algorithm to process the geographic coordinate sequences of multiple records identified as the same target, thus completing the target's trajectory when flying across UAV nodes within overlapping boundary sub-regions. The Kalman filter algorithm can optimally estimate the target's motion state based on the returned coordinates at different times, eliminating the "flash" or "disappearance" of the target's position caused by viewpoint switching, and forming a continuous and complete target motion trajectory.

[0034] The cloud processing center is also configured to: receive remaining battery information from each of the drone nodes; when the remaining battery of any drone node is lower than a preset battery threshold, recalculate the remaining unpatched area of ​​that drone node; dynamically split the tasks for the remaining unpatched area and allocate them to nearby drone nodes with sufficient battery power; and control the drone node to return to its home location. This dynamic task reallocation mechanism ensures the continuous execution of inspection tasks even when a single drone has low battery power, enhancing the overall robustness of the system.

[0035] The preset power threshold is 10-30% of the remaining power of the drone node, preferably 20%. Using 20% ​​as the power trigger threshold not only provides sufficient safety margin for the original drone to return, but also ensures timely task handover before the power is exhausted, avoiding monitoring gaps in the inspection area.

[0036] In some embodiments, each drone node is further configured to: when the local detection confidence of the edge AI computing unit is lower than a preset confidence threshold, send a single-frame image to the cloud processing center, requesting the cloud processing center to perform secondary target detection confirmation on the single-frame image. This computing power offloading alternative mechanism is suitable for scenarios where local models cannot reliably judge due to extremely complex backgrounds such as heavy fog. By uploading the single-frame image to the cloud for secondary confirmation, it achieves a flexible inspection combining coarse and fine-grained methods, maintaining the principle of low-bandwidth backhaul while ensuring recognition reliability in extreme scenarios.

[0037] Specifically, the edge AI computing unit is directly connected to the airborne camera via a MIPI or USB interface, allowing the raw image data stream output by the UAV to bypass the image transmission encoding chip in the data path and be directly input to the edge AI computing unit without undergoing video encoding processing by the image transmission encoding chip. The edge AI computing unit runs a target detection model compressed with INT8 quantization. By directly connecting to the airborne camera via MIPI or USB interface and bypassing the image transmission encoding chip, the edge AI computing unit achieves "acquisition as inference" at the physical link level, eliminating the additional latency introduced by the traditional image transmission encoding process. Running the target detection model compressed with INT8 quantization significantly reduces computing power and power consumption requirements, while increasing the proportion of the target in the field of view by reducing the granularity of sub-region tasks, thereby maintaining a high target recall rate under limited computing power conditions.

[0038] The target detection model, compressed using INT8 quantization, is deployed on the edge AI computing unit. Its function is to receive image data from the airborne camera, perform target detection inference on the UAV side, identify targets to be detected in inspection scenarios, and output the target category, image pixel coordinates, and detection confidence score. Based on the image pixel coordinates, it further calculates WGS-84 geographic coordinates, which serve as the core data source for metadata messages. On one hand, the obtained confidence score can be used locally to filter low-confidence detection results; on the other hand, this confidence score is reported to the cloud along with the metadata, allowing the cloud to perform confidence-weighted arbitration to achieve deduplication alerts for the same target in overlapping areas across multiple UAVs. When the local output confidence score is lower than a preset confidence threshold, a flexible verification process of uploading a single-frame image to the cloud for secondary confirmation will be triggered. This model pushes target detection inference down to the UAV side, eliminating the need for the system to transmit the original video stream back, only transmitting lightweight metadata messages, reducing communication bandwidth overhead and warning latency.

[0039] The process of establishing the INT8 quantized compressed target detection model is as follows: First, real-world image samples were collected from water patrol or forest fire prevention patrol scenarios. The samples covered normal scenes as well as complex and harsh conditions such as backlight, heavy fog, and water surface reflection. The target to be detected in the images was labeled with category and pixel box, and the training set, validation set and test dataset were obtained. Secondly, a lightweight single-stage object detection network such as YOLOv5s, YOLOv8n, NanoDet or MobileNet-SSD is selected, and floating-point model training is completed using the training dataset to obtain the original object detection model with FP32 floating-point weights. Next, a portion of real inspection images were selected as the calibration dataset. The floating-point model was compressed using the quantization toolchain provided by the NPU vendor. The model weights and activation values ​​were mapped from 32-bit floating-point numbers to 8-bit integer values, reducing the model size and computational load. After quantization, performance verification is carried out in a simulation environment and on the corresponding embedded AI hardware. The recall rate, false detection rate and inference frame rate of the model before and after quantization are compared. If quantization causes too much loss of accuracy, a hybrid quantization strategy can be used to retain floating-point operations for some network layers. Finally, the quantized model is converted into a hardware inference format that can be loaded by the edge AI computing unit and deployed to the drone's onboard edge AI computing unit.

[0040] In this invention, since the inspection area is divided into multiple main and sub-areas in the cloud, each drone is only responsible for a smaller main and sub-area. The target to be inspected occupies a larger proportion in the camera's field of view and there is less background interference, which can offset the slight accuracy loss caused by INT8 quantization. This allows the lightweight detection model to maintain a high target recall rate even under the limited computing power and power consumption conditions on the edge.

[0041] This invention also provides an intelligent inspection method based on multi-machine collaboration, comprising the following steps: The cloud processing center obtains the geographical boundary information of the target area, divides the target area into N main sub-regions according to the number of drones N in the drone collaborative cluster, and generates boundary overlapping sub-regions by buffering and extending along the common edge of the adjacent main sub-regions to both sides. The ratio η of the total area of ​​the boundary overlapping sub-regions to the total area of ​​the target area is constrained not to exceed a preset area ratio threshold. The coordinate information of each main sub-region and the boundary overlapping sub-regions is packaged into a flight path task and sent to the corresponding drone node. After receiving the flight route task, each of the aforementioned UAV nodes takes off in concert and flies to the airspace of its assigned sub-area to perform image acquisition. The edge AI computing unit of each UAV node performs target detection inference on the acquired images at the edge. When a target is detected, it transmits a metadata message containing the target category, geographic coordinates, and confidence level back to the cloud processing center via a wireless communication link. The purpose of multi-UAV coordinated synchronous take-off is that each UAV node covers its respective sub-area in parallel, compressing the overall inspection coverage time to about 1 / N of that of a single-UAV serial inspection, thereby doubling the inspection efficiency. The fusion analysis server in the cloud processing center performs time window alignment and spatial distance calculation on the multiple metadata messages returned. Multiple records within the same time window and whose geographical coordinate distance is less than a preset spatial threshold are identified as the same target. The server then performs confidence-weighted arbitration on the multiple records identified as the same target and retains the record with the highest confidence to generate alarm information.

[0042] The method to constrain η from exceeding the preset area percentage threshold is as follows: calculate the current η value, and if η exceeds the preset area percentage threshold, reduce the buffer extension distance or adjust the shape of the main sub-region and recalculate, iterating until η does not exceed the preset area percentage threshold.

[0043] When the edge AI computing unit does not detect a target, each drone node sends a heartbeat packet to the cloud processing center at set time intervals. The heartbeat packet contains the drone node's location information, remaining battery power, and operating status. The purpose of this heartbeat packet sending step is to maintain the cloud processing center's continuous awareness of the real-time status of the drone collaborative cluster with extremely low communication bandwidth during flight without any detected events. This provides the remaining battery power data input of each node for the dynamic task reallocation mechanism, ensuring that the cloud can promptly trigger the task takeover process when any node's battery power reaches a preset battery power threshold, avoiding task gaps caused by power monitoring interruptions.

[0044] In one embodiment, the method further includes: processing the geographic coordinate sequences of multiple records identified as the same target using a Kalman filter algorithm to complete the target's motion trajectory when flying across UAV nodes within a boundary overlapping sub-region. The purpose of this trajectory completion step is that when a target switches from the field of view of one UAV node to the field of view of an adjacent UAV node within a boundary overlapping sub-region, there is a time difference and coordinate jump in the reporting of the same target by the two UAV nodes. If trajectory completion is not performed, the target motion record will be broken, manifesting as the target position "flashing" or "disappearing." By using the Kalman filter algorithm to perform optimal estimation interpolation on the discrete coordinate sequence, a continuous and complete target motion trajectory can be generated, providing a complete record of target behavior for subsequent law enforcement evidence collection, supporting the improvement from "single-point perception" to "global cognition."

[0045] In one embodiment, the system further includes: the cloud processing center receiving the remaining power information sent by each of the drone nodes in real time; when the remaining power of any drone node is lower than a preset power threshold, recalculating the remaining unpatched area of ​​the drone node, dynamically splitting the tasks of the remaining unpatched area and allocating them to nearby drone nodes with sufficient power, and controlling the drone node to return to its home location.

[0046] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements all the steps of the above-described intelligent inspection method.

[0047] This invention relates to an intelligent inspection system and method based on multi-machine collaboration, aiming to solve the problems of unstable transmission and high latency caused by long-distance backhaul of high-definition video streams in existing single-machine inspection modes, as well as insufficient recognition accuracy and shortened battery life due to limited computing power on the single-machine end side. The core technical concept of this application is: adopting a three-level collaborative flat architecture of "cloud-edge-end", eliminating blind spots at regional boundaries when multi-machine collaborative coverage is achieved through dynamic boundary redundant grid partitioning with area constraints; after each UAV node completes target detection inference on the end side, it only sends back lightweight metadata messages to the cloud, completely abandoning the transmission of the original video stream; the cloud fusion analysis server then performs spatiotemporal alignment and confidence weighted arbitration on the multi-source backhaul data, thereby realizing a distributed intelligent inspection closed loop with low bandwidth, low latency, and no blind spots in the entire area. The above technical solution is described in detail below through four embodiments.

[0048] Example 1: Intelligent UAV Inspection System Based on Multi-UAV Collaboration with a Three-Tier "Cloud-Edge-Device" Architecture Reference Figure 1 The intelligent inspection system in this embodiment consists of three main layers: cloud processing center 100, wireless communication link 200, and UAV collaborative cluster 300. Together, they form a flat architecture of "cloud-edge-device" collaboration.

[0049] (a) Cloud Processing Center 100 The cloud processing center 100 consists of a fusion analysis server 101 and an inspection database 102, and incorporates a Geographic Information System (GIS) module. Its core responsibility has shifted from cumbersome video decoding and frame-by-frame inference to becoming a task planning engine and a spatiotemporal data fusion hub. Specifically: the GIS module is responsible for parsing the vector map of the target area to be inspected, obtaining the geographic boundary information of the target area, and inputting the boundary data into the area division algorithm; the fusion analysis server 101 is data-connected to the inspection database 102, responsible for receiving metadata messages transmitted from each UAV node, performing multi-source data fusion, deduplication determination, and confidence arbitration, and storing the final alarm records in the inspection database 102. Since the cloud no longer undertakes continuous decoding and inference of the video stream, its continuous consumption of computing power is significantly reduced, and the overall system operating cost is significantly lowered. In a typical deployment implementation, the fusion analysis server (101) can be an x86 server equipped with a GPU accelerator card (e.g., configured with an NVIDIA T4 or an inference accelerator card of equivalent computing power) to meet the real-time fusion processing requirements of multiple concurrent metadata streams. Since the cloud no longer undertakes the continuous decoding and inference of video streams, its continuous occupation of computing power is greatly reduced. Compared with the traditional "single machine + cloud centralized analysis" mode, the cloud GPU computing power requirement can be reduced by more than 90%, and the overall system operating cost is significantly reduced.

[0050] (ii) Wireless communication link 200 The wireless communication link 200 is connected to the cloud processing center 100 for bidirectional data transmission between the cloud processing center 100 and the UAV collaborative cluster 300. Unlike traditional systems, the wireless communication link 200 in this embodiment only retains two types of channels: one is a low-bandwidth control command transmission channel, used to transmit control commands such as flight path tasks and parameter configurations from the cloud processing center 100 to each UAV node; the other is a service metadata feedback channel, used to upload the identification result metadata messages and heartbeat packets of each UAV node to the cloud processing center 100. The wireless communication link 200 no longer occupies the high-bandwidth HDMI / SDI digital image transmission frequency band, greatly reducing the risk of radio frequency interference and enabling the system to work stably in complex electromagnetic environments in the field.

[0051] In one implementation, the wireless communication link 200 can use a 4G / 5G IoT card. The return latency of a single metadata message (in the hundreds of bytes) is typically less than 100 milliseconds, fully meeting the requirement that the end-to-end warning delay does not exceed 2 seconds. In another implementation, a LoRa remote transmission module (with an air speed of 50kbps) can be used to adapt to remote areas without public network signal coverage. In this case, the transmission latency of the metadata message is approximately tens to hundreds of milliseconds, which still meets the real-time requirements of the inspection service. The heartbeat packet and the metadata message share the same return channel. Both are lightweight JSON messages, so there is no channel contention issue.

[0052] (III) 300 Unmanned Aerial Vehicle Collaborative Cluster The drone collaborative swarm 300 includes at least two drone nodes (such as 301, 302, and 30X), and each drone node can be a homogeneous or heterogeneous multi-rotor drone platform. Unlike traditional drones that merely function as "flying cameras," in this embodiment, each drone node embeds a dedicated edge AI computing unit (such as 301a and 302a) in its hardware link. In one typical implementation, the edge AI computing unit can be an embedded AI module based on the RK3588 chip, which has a built-in NPU with 6 TOPS of computing power, supports INT8 quantization inference, and consumes approximately 5W to 8W of power, making it suitable for consumer-grade multi-rotor drone platforms with strict requirements for payload and power consumption. In another implementation, an NVIDIA Jetson Nano module can be used, equipped with 128 CUDA cores, supporting INT8 quantization inference, and consuming approximately 5W to 10W of power. Alternatively, the Horizon Robotics X3 series NPU module can be used, with an INT8 inference computing power of up to 5 TOPS and a power consumption of approximately 2W to 3W, making it suitable for lightweight drone platforms that are extremely sensitive to power consumption. The aforementioned edge AI computing unit is directly connected to the airborne camera via a MIPI CSI-2 interface or a USB 3.0 interface, bypassing the traditional image transmission encoding chip at the hardware link level. This allows image data to directly enter the edge AI computing unit for inference after acquisition, achieving "acquisition as inference" and eliminating the additional latency introduced by the traditional image transmission encoding process. The resolution of the airborne camera can be selected according to the inspection requirements. In a typical implementation, a 1080P or 4K visible light camera can be used. Image data is directly sent to the edge AI computing unit in MIPIRAW or YUV format. The inference input resolution is typically scaled to 640×640 pixels to match the input requirements of the lightweight detection model. The edge AI computing unit runs a lightweight target detection model compressed using INT8 quantization. Taking YOLOv5s as an example, its original FP32 model size is approximately 14MB, which can be reduced to less than 4MB after INT8 quantization compression. The inference frame rate on the RK3588 NPU can reach more than 10 frames per second, far exceeding the acquisition frame rate requirement of 2 frames per second in this embodiment, providing ample computing power margin. The typical target confidence threshold can be set to 0.5, and the target recall rate can reach more than 90% at the sub-region task granularity. Since each UAV node is only responsible for a small sub-region, the target to be detected accounts for a large proportion of the field of view and there is little background interference, enabling the lightweight model to maintain a high target recall rate even under the limited computing power of the edge, effectively reducing the risk of missed detections and overcoming the industry problem that lightweight UAVs cannot perform high-precision real-time AI analysis.

[0053] (iv) Intelligent division of target area and boundary redundancy control Traditional regional meshing often employs mechanical, uniform division, ignoring camera field-of-view (FOV) errors and UAV positioning drift, resulting in blind spots at region boundaries. To address this issue, this invention proposes a dynamic meshing algorithm with dual constraints (geometric and energy constraints).

[0054] The target region is divided into two types of sub-regions. These two types of sub-regions differ fundamentally in their definition, quantity, coverage logic, and constraint relationships. A detailed comparison is as follows: The primary and secondary regions are the initial average division results of the target area. The cloud processing center (100) divides the target area into N primary and secondary regions Zi (i=1,2,...,N) based on the number N drones currently deployed in the cluster and the nominal range of a single drone. Each primary and secondary region is independently covered by a dedicated drone node, and the primary and secondary regions do not overlap during the initial division stage, forming a seamless block of the target area. The shape of the primary and secondary regions can be adaptively adjusted according to the geographical contour of the target area, and is not limited to rectangular grids. The core design goal of the primary and secondary regions is to determine the partitioning granularity based on the single drone's range, so that the workload of each drone node matches its range, and at the same time, to compress the overall inspection time to about 1 / N of the single drone serial inspection time through N drones in parallel coverage, thereby doubling the inspection efficiency. Taking a typical reservoir scenario (approximately 2 square kilometers in area, with 5 drones) as an example, when N=5, the target area is divided into 5 main sub-areas, each with an area of ​​approximately 0.4 square kilometers. Each drone is only responsible for the inspection and coverage of this small area of ​​water.

[0055] Category 2: Overlapping sub-regions (the number depends on the number of adjacent common edges) Boundary overlap sub-regions are redundant coverage bands generated by buffering and extending to both sides along the adjacent common edges of adjacent main sub-regions after the main sub-regions are divided. The number of boundary overlap sub-regions is determined by the number of common edges between adjacent main sub-regions. For example, in the reservoir scenario mentioned above, there are several common edges between the four main sub-regions, corresponding to the generation of several boundary overlap sub-regions (typically two boundary bands). Boundary overlap sub-regions do not belong to any specific drone node, but are achieved by two adjacent drone nodes on either side of the common edge each flying over the region, thus achieving cross-coverage. The buffer extension distance is the sum of the drone node's positioning error and the airborne camera's field of view (FOV) margin, to accurately fill potential blind spots caused by drone positioning drift and camera FOV deviation. Boundary overlap sub-regions are subject to area constraints: the system calculates the total area S of all boundary overlap sub-regions. boundary Total area S of the target region total The ratio η=S boundary / S totalThrough iterative optimization, it is ensured that η does not exceed the preset area ratio threshold (preferably 20%), so as to achieve zero missed detection coverage across the entire area with minimal spatial resource redundancy.

[0056] Comparison of the two types of sub-regions: The primary sub-region and the boundary-overlapping sub-region differ fundamentally in the following four dimensions: First, in terms of coverage logic, the primary sub-region is independently covered by a single dedicated drone node, while the boundary-overlapping sub-region is cross-covered by two adjacent drone nodes; Second, in terms of quantity, the number of primary sub-regions equals the number of drones in the cluster, N, while the number of boundary-overlapping sub-regions is determined by the number of adjacent common edges, usually less than N; Third, in terms of area constraint, the sum of the areas of N primary sub-regions equals the total area S of the target region. total The total area of ​​overlapping sub-regions is subject to a hard constraint that η ≤ a preset area ratio threshold; fourthly, in terms of functional positioning, the main sub-region undertakes the core inspection coverage task, while the overlapping sub-regions are specifically used to eliminate monitoring blind spots at the boundaries of areas during multi-drone rotation, and to provide a data foundation for multi-drone cross-verification of multi-source data fusion in the cloud. It is precisely because the same target may be captured and reported by two drone nodes simultaneously that cloud-based time window alignment, spatial distance calculation, and confidence-weighted arbitration and other fusion disambiguation mechanisms are triggered to ensure the uniqueness and accuracy of alarms.

[0057] Reference Figure 2 The cloud processing center 100 executes a dynamic mesh generation algorithm with dual constraints (geometric and energy consumption constraints) during the mission planning phase. This algorithm not only considers the geometric features of the geographic space but also deeply integrates the physical performance limits of the UAV itself. The specific implementation steps are as follows: Step 1, Initial rasterization and load sensing task pre-assignment: The cloud processing center first calls the GIS module to obtain the target inspection area S. total The vector boundary (usually in WKT or GeoJSON format). The system reads the number N of currently online and fully charged drones, and combines this with the nominal battery capacity of a single drone, the wind speed prediction model, and the peak power consumption of the onboard AI computing unit to inversely calculate the effective coverage area threshold S of a single drone with 20% safety redundancy power. limit .

[0058] The inversion approach is as follows: Given the total available power → deduct all power consumption → how far can the remaining power fly → convert it into the area that can be covered.

[0059] (a) Energy Budget: Total available power per unit (after deducting 20% ​​safety redundancy) is: E available =C bat ×V nom ×0.8, Where C batV represents the battery's nominal capacity (in mAh, e.g., 5000mAh). nom E represents the battery's nominal voltage (in volts, e.g., 22.2V corresponds to a 6S lithium battery pack). available The unit conversion to Wh represents the upper limit of the effective energy available for a single unit for inspection flights. Taking a typical parameter as an example: E available =5000mAh×22.2V×0.8 / 1000≈88.8Wh.

[0060] (b) Components of total power consumption: The total power consumption P of the UAV during inspection flight total It consists of three parts: P total =P hover (v wind )+P fly +P ai , Where P hover (v wind Power consumption during hovering and low-speed flight (affected by wind speed, typical range 200W–400W); P fly Additional power consumption for horizontal propulsion (typical range 50W–150W); P ai This refers to the peak power consumption of airborne edge AI computing units (e.g., RK3588 is approximately 8W, Jetson Nano is approximately 10W, and Sunrise X3 is approximately 3W). ai While its power consumption is relatively small compared to flight power consumption, it cannot be ignored in accurate inversion.

[0061] (c) Wind speed prediction model: P hover (v wind The wind speed v in ) wind For dynamic variables, the system supports the following three models for acquisition: Firstly, the physical model, applicable to low wind speed conditions, is based on the rotor aerodynamic hovering power formula: , where W is the lift requirement (N) generated by the weight of the drone. ; F wind Adding force to wind resistance, air density (kg / m³) 3 (related to temperature and altitude), A is the rotor disk area (m²). 2 A front The area of ​​the drone's body perpendicular to the wind direction is represented by the wind speed data, which is obtained from publicly available meteorological APIs (such as Genting Weather and Rainbow Weather) for the next hour's wind speed forecast or from the drone's own barometric pressure sensor real-time readings. dC is a dimensionless coefficient representing the obstruction effect of airflow on the drone's body. It is related to the drone's shape, the direction of the incoming airflow, and the Reynolds number. In hovering mode, C... d Generally determined through wind tunnel testing or ground-based air blowing calibration; the smoother the shape, the higher the C value. d The smaller; the more edges and corners the body has, C d The larger.

[0062] Secondly, the empirical lookup table model pre-calibrates the actual power consumption under different wind speeds through ground testing, generating a power consumption-wind speed lookup table. Before inspection, P is directly obtained by interpolation from the table. hover (v wind It has the best engineering practicality; Third, a data-driven model is used, which trains a regression model (such as linear regression or XGBoost) using historical flight logs (current × voltage × wind speed). The current wind speed prediction is input, and the expected power consumption is output. This method offers the highest accuracy but requires accumulating a certain amount of historical flight data. In a preferred embodiment, the system combines an empirical lookup table model with meteorological API wind speed forecasts, ensuring both engineering feasibility and prediction accuracy.

[0063] (d) Effective flight time inversion: after obtaining P total Afterwards, the effective flight time for a single aircraft is: T fly =E available / P total Taking typical parameters as an example (E) available =88.8Wh, P hover =340W, P fly =100W, P ai =8W):T fly =88.8Wh / (340+100+8)W≈0.198h≈11.9 minutes.

[0064] (e) Effective coverage area conversion: In the bow-shaped scanning flight path mode, the effective coverage area S of a single aircraft limit Determined by effective flight time and inspection flight speed v cruise Single scan strip width W strip and flight belt utilization η overlap The decision was made jointly, and the calculation formula is as follows: S limit =v cruise ×T fly ×W strip ×η overlap .

[0065] Among them, W strip =2×H×tan( H is the cruising altitude (m). η is the camera's half field of view (i.e., FOV / 2);overlap The utilization rate of the flight strip is usually taken as 0.7 to 0.8 (excluding the overlapping part of adjacent flight strips).

[0066] Taking typical parameters as an example (v) cruise =8m / s, T fly =11.9min=714s, H=50m, FOV=60° =30°, η overlap =0.75): W strip =2×50×tan(30°)≈57.7m; S limit =8m / s×714s×57.7m×0.75≈247,500m 2 ≈0.25km 2 .

[0067] This value represents the effective coverage area threshold for a single machine under current power, wind speed, and computing power conditions. The system uses this threshold to perform subsequent load balancing partitioning, ensuring that the area Si of each master-sub-region does not exceed S. limit .

[0068] Source of each parameter: C bat V nom : Drone hardware specifications, factory-fixed values; P hover (v wind ): Real-time calculation of wind speed prediction model; P fly : Experienced values ​​or actual calibration of the drone model; P ai Edge AI chip specifications, such as the RK3588 with a peak power of 8W; cruise H Task parameters are preset by the operator or the system; η overlap The route planning parameter is typically 0.75.

[0069] The system is based on S limit S total Initial partitioning is performed. Unlike simple geometric partitioning, this step uses a load balancing algorithm: prioritizing the protection of each principal subregion Z. i area S i ≤S limit At the same time, ensure that the shape of each sub-region is as regular as possible (such as rectangle or sector) to reduce the unnecessary time and energy consumption of the UAV during turning. The generated initial mesh set is denoted as {Z1, Z2, ..., Z...} NThe target area is initially divided into N principal sub-regions Zi (i=1,2,...,N). The shape of each principal sub-region can be adaptively adjusted according to the geographical contour of the target area, and is not limited to a rectangular grid. The purpose of this step is to determine the partitioning granularity based on the nominal range coverage area of ​​a single drone, so that the workload of each drone node matches its range capability, avoiding overloading (leading to returning before completing the task) or underloading (causing ineffective waste of cluster computing power and electricity), and at the same time providing an initial partitioning basis for the generation of subsequent overlapping sub-regions.

[0070] Step 2, Boundary extraction and adaptive buffering based on topological relationships: Based on a GIS-based topology analysis engine, the system traverses all master and sub-regions {Z i Extract the set of common adjacency edges E between each pair of adjacent subregions. adj For each common edge e∈E adj The system performs a buffer operation: Error modeling: The system retrieves the RTK positioning accuracy data σ of the current environment. rtk (Typically around 0.1m-1m) and drone hovering accuracy data σ hover At the same time, combined with the field of view (FOV=2) of the airborne camera... Given the cruise altitude H, calculate the horizontal coverage margin d. fov The horizontal coverage margin d fov It can be preferred through formula d fov =2H·tan -2(σ rtk +σ hover The value can be calculated or preset to a fixed value based on actual engineering experience.

[0071] Buffer distance determination: Set the single-sided buffer distance D buffer =k·(σ rtk +σ hover )+d fov Where k is the safety factor (usually taken as 2-3, representing the confidence interval). The system extends D to both sides along the common edge e. buffer Distance generates a buffered polygon that is rectangular or trapezoidal.

[0072] Boundary sub-region generation: Perform geometric union operation (B) on all buffer polygons. union =⋃Buffer(e)), and for items exceeding S total The boundary portion is trimmed to generate a preliminary set of boundary sub-regions B. initialThis step ensures that even if the drone experiences slight positioning drift, its shooting range can still cover the theoretical dividing line. At the same time, the overlapping sub-regions of the boundary cover the potential blind spots caused by the drone's positioning drift and the camera's field of view deviation, ensuring that there is no monitoring vacuum when adjacent drone nodes switch positions.

[0073] Step 3: Area proportion constraints and iterative optimization (core constraint logic) To avoid excessive expansion of the total flight distance and operation time of the drone swarm due to boundary redundancy design (i.e., to avoid "exhausting power for full coverage"), this invention introduces a strict area occupancy threshold control mechanism: Percentage Calculation: The system automatically calculates the total area S of the current boundary sub-region. boundary (i.e. B) initial (area of) and total area S of the target area total The ratio η=S boundary / S total .

[0074] Iterative recalculation: A hard constraint is set: η ≤ 20%. If η > 20%, the system determines that resource consumption has exceeded the limit and immediately starts an iterative optimization loop. Fine-tune the buffer distance: linearly reduce D by 20% / η. buffer Regenerate the buffer polygon; Shape optimization: If simply reducing the distance cannot meet the requirements, the system will adjust the granularity of the main sub-region division. For example, it can merge two narrow adjacent main sub-regions into a larger main sub-region to reduce the number of adjacent edges; or it can straighten the originally curved common edges to reduce the overlapping area of ​​the buffer zone.

[0075] Cost function evaluation: In each iteration, the system evaluates whether the new partitioning scheme satisfies η≤20% and all S i ≤S limit The iteration continues until the optimal solution that satisfies all constraints is found, thus achieving the best balance between "zero missed detections" and "low energy consumption".

[0076] like Figure 2 As shown, this design uses minimal mathematical redundancy to precisely fill visual blind spots in physical space. The preset area ratio threshold comprehensively considers the balance between coverage integrity and the energy consumption and computing power cost of the drone swarm, and can be adjusted within a reasonable range according to the accuracy requirements of the actual inspection scenario.

[0077] Step 4: Route generation and task mapping distribution After optimization and partitioning, the system obtains the final set of primary sub-regions {Zi} and boundary sub-regions {Bj}. For each Zi and Bj, the system automatically generates corresponding UAV waypoints: for primary sub-regions, a "bow" or spiral scanning path is typically used; for boundary sub-regions, a reciprocating scanning path is generated along the centerline of the buffer zone. Each waypoint sequence includes latitude and longitude (WGS-84 coordinates), flight altitude, suggested frame rate, and gimbal pitch angle to ensure that the image acquisition quality at each waypoint meets the target detection requirements.

[0078] Before the mission is launched, the system performs flight path conflict detection to ensure that multiple drones do not intersect or collide in the same airspace. Subsequently, the cloud processing center (100) distributes the corresponding flight path mission package (including sub-area ID, flight path coordinate sequence, triggering conditions, etc.) to the designated drone nodes (301, 302...30X) via wireless communication link (200). After receiving the flight path mission, each drone node loads its local AI model parameters and takes off in sequence or synchronously, flying to its assigned sub-area to perform altitude-holding image acquisition.

[0079] The purpose of this step is to: the cloud processing center (100) uniformly plans and distributes the flight route tasks, which can achieve the optimal task allocation globally and ensure seamless connection of the coverage of each UAV node; compared with the scheme planned autonomously by each UAV node, the unified distribution by the cloud can effectively avoid the flight route conflict, coverage overlap or area omission problems caused by the autonomous planning of multiple UAVs, and at the same time provide a unified task coordinate benchmark for subsequent dynamic task redistribution (such as task takeover when the power is insufficient).

[0080] (V) End-side inference and data transmission workflow for multi-machine collaborative distributed inspection Reference Figure 3 Each UAV node executes the following end-side inference and feedback logic during flight operations, forming a tight parallel control closed loop.

[0081] In the first stage, local perception and edge inference: the airborne camera acquires images at a set frame rate (e.g., 2 frames per second to reduce power consumption). The image data is directly input to the edge AI computing unit (301a, 302a) via MIPI or USB interface, where the edge AI computing unit runs a lightweight target detection model compressed with INT8 quantization to perform inference. Since each UAV node is only responsible for a small sub-region, the target occupies a large proportion of the field of view with little background interference, and the lightweight model can still maintain a high target recall rate under limited computing power. The purpose of using a low frame rate (e.g., 2 frames per second) is to: control the computational load of the edge AI computing unit and the power consumption of the airborne battery within a reasonable range while ensuring the real-time performance of target detection; since the UAV's flight speed is limited and the target stays in the field of view for a long time, low frame rate acquisition will not lead to missed targets, and at the same time, it can significantly extend the single inspection endurance, achieving an optimal balance between detection real-time performance and endurance.

[0082] The second phase involves dual-mode communication of abnormal triggering and regular heartbeat: When the edge AI computing unit does not detect the target, the drone node only sends a "heartbeat packet" to the cloud processing center 100 once at a set time interval (e.g., every 30 seconds). The heartbeat packet contains the drone node's current location information, remaining battery power, and operating status, which is used to maintain the cloud's real-time awareness of the cluster status, while occupying only a very low communication bandwidth.

[0083] In one implementation, the heartbeat packet is a lightweight JSON-formatted message containing the following fields: drone node number (drone id (String type, e.g., "301"), current WGS-84 geographic coordinates (lat, lng, double-precision floating-point type), flight altitude (altitude, in meters), and remaining battery percentage (battery). pct (Integer, e.g., 75), running status flags (status, enumeration, e.g., "normal" / "warning") and timestamps (timestamp, Unix timestamp, integer).

[0084] A complete example of a typical heartbeat message is as follows: {"drone id ":"301","lat":30.2541,"lng":120.1233,"altitude":50.0,"battery pct The message contains approximately 120 bytes of data. With a heartbeat interval of 30 seconds, the average transmission rate of a single drone node is only about 4 bytes / second, resulting in extremely low communication bandwidth usage.

[0085] When the edge AI computing unit identifies a target (such as a person holding a fishing rod on the water surface or a heat source or light spot in the forest area), it immediately triggers the return process of metadata messages without waiting for the next heartbeat cycle, thus ensuring the real-time nature of the warning.

[0086] The third stage involves lightweight metadata packaging and low-bandwidth backhaul: After the edge AI computing unit locks onto the target, it captures a low-resolution keyframe thumbnail of the current frame (typically 160×120 pixels, approximately 3KB to 5KB after JPEG compression), along with the target semantic category label, WGS-84 geographic coordinates obtained based on image coordinate transformation, and model confidence score, and packages them into a lightweight JSON format metadata message.

[0087] The complete field structure of a typical metadata message is as follows: {"drone id ":"301","target class ":"person-fishing","lat":30.2538,"lng":120.1245,"confidence":0.92,"altitude":50.0,"thumbnail base64 ":"...(base64 encoded JPEG thumbnail data)...","timestamp":1700000120}.

[0088] Among them, target class The `tl` and `lng` fields are the target semantic category labels (string type, e.g., "person-fishing" for anglers, "fire-smoke" for forest fireworks); the `lat` and `lng` fields are WGS-84 geographic coordinates obtained based on image coordinate transformation (double-precision floating-point type, accuracy better than 1 meter); the `confidence` field is the model confidence score (floating-point type, range 0-1); `thumbnail` base64The keyframe thumbnail field is a Base64 encoded string (approximately 4KB to 7KB); the timestamp field is a Unix timestamp (integer) of the detection time. The data size of this message is only in the hundreds of bytes; the plain text field without the thumbnail is approximately 200 bytes, and the complete message with the thumbnail is approximately 4KB to 8KB. Compared to the original video stream, which can reach several megabits per second or even tens of megabits per second, the communication bandwidth requirement is reduced by 3 to 4 orders of magnitude, resulting in significant bandwidth compression. The drone node transmits this message instantly to the cloud processing center 100 via the wireless communication link 200 (e.g., a 4G / 5G IoT card or a LoRa remote transmission module). This process completely eliminates the need for original video stream transmission, eradicating frame loss and high latency issues caused by limited wireless link bandwidth at the physical transmission level.

[0089] (vi) Cloud-based multi-source data fusion and disambiguation Due to the overlapping coverage of multiple drones in sub-regions with overlapping boundaries, the same target is highly likely to be reported simultaneously by two adjacent drone nodes. The fusion analysis server 101 of the cloud processing center 100 executes the following four steps to ensure the uniqueness and accuracy of the alarm.

[0090] The first step is time window alignment: The fusion analysis server 101 sets a time window threshold (for example, in a preferred embodiment, it is set to ±3 seconds, meaning multiple messages with a time difference within 6 seconds are grouped into the same batch), and aligns the data transmitted within the same time window into the same batch. The time window threshold can be adjusted according to the target movement speed in the specific inspection scenario. For example, for fishing targets that are stationary or moving slowly on the water surface, the time window can be appropriately widened to ±5 seconds; for fast-moving targets (such as vehicles or speedboats), the time window should be narrowed to ±1 second accordingly.

[0091] The purpose of this step is as follows: Because different drone nodes have different flight positions, processing speeds, and channel delays, when the same target is captured by two drone nodes simultaneously, the arrival time of their corresponding metadata messages in the cloud may have slight differences. If spatial comparison is performed directly without time alignment, two reports of the same target may be misjudged as two independent events at different times due to the time difference, leading to missed or incorrect judgments. Time window alignment, by grouping multiple messages with time differences within a threshold range into the same batch, provides a correct comparison basis for subsequent spatial distance calculations.

[0092] The second step is Spatial Intersection over Union (IOU) calculation: Based on the returned WGS-84 geographic coordinates, the geographic distance between target points reported by different UAV nodes is calculated on the electronic map. If the distance between two points is less than a preset spatial threshold (for example, set to 10 meters in a preferred embodiment, this threshold is determined after comprehensively considering the UAV's GPS positioning accuracy (typical CEP is about 1.5 meters), the target's physical size (person targets are about 0.5 meters), and image coordinate transformation errors), then they are determined to be the same target. The preset spatial threshold can be adjusted according to the UAV's positioning accuracy and the target's physical size. For example, for vehicle targets (physical size about 4 meters), the spatial threshold can be appropriately widened to 20 meters; for RTK UAV platforms with higher positioning accuracy (positioning accuracy better than 0.1 meters), the spatial threshold can be narrowed to within 5 meters to improve the accuracy of deduplication. Therefore, the preset spatial threshold can be adjusted according to the UAV's positioning accuracy and the target's physical size to adapt to the accuracy requirements of different inspection scenarios.

[0093] The purpose of this step is that the same physical target within the overlapping sub-region of the boundary will inevitably be captured and reported by adjacent drone nodes at the same time. If spatial deduplication is not performed, the cloud will generate duplicate alarms for the same target, which will mislead the management personnel. By calculating the geographical distance of the target points reported by different nodes and comparing it with the preset spatial threshold, multiple duplicate records of the same physical target can be accurately identified and merged, which is the core mechanism to ensure the uniqueness of alarms.

[0094] The third step is confidence-weighted arbitration: For multiple records identified as belonging to the same target, the fusion analysis server 101 performs confidence-weighted arbitration, retaining the record with the highest confidence level and storing it in the inspection database 102. This record is then linked with keyframe thumbnails transmitted from multiple machines to form a target evidence chain, facilitating review by law enforcement personnel. Redundant records with lower confidence levels are discarded to avoid duplicate alarms interfering with management personnel.

[0095] The fourth step is Kalman filter trajectory completion: The fusion analysis server 101 uses the Kalman filter algorithm, taking the WGS-84 geographic coordinate sequence returned at different times as the observation input, to make the optimal estimate of the target's motion state and complete the target's motion trajectory when flying across UAV nodes in the boundary overlapping sub-region. In a typical implementation, the state vector of the Kalman filter can be defined as [x,y,vx,vy] (the target's longitude, latitude, and velocity components in the corresponding direction, a total of 4 dimensions), and the observation vector is [x,y] (i.e., the returned WGS-84 longitude and latitude coordinates, a total of 2 dimensions). The process noise covariance matrix Q and the observation noise covariance matrix R can be initialized according to the target's motion characteristics and positioning accuracy. When a target switches from the field of view of one UAV node to the field of view of an adjacent UAV node, there is a time interval between the two observations (usually 1 to 3 heartbeat cycles, i.e., 30 to 90 seconds). The Kalman filter algorithm extrapolates the target position within the observation interval through the prediction step, and then performs an update step to correct the estimated value when a new observation arrives. This eliminates the phenomenon of the target position "flashing" or "disappearing" caused by the switching of the viewpoints of adjacent UAVs, forming a continuous and complete record of the target's motion trajectory, and realizing the improvement from "single-point perception" to "global cognition".

[0096] Example 2: Enhanced Inspection System with Dynamic Task Reassignment Mechanism Based on Embodiment 1, this embodiment further introduces a dynamic task redistribution mechanism based on remaining power to address the task continuity problem under low power conditions of a single machine and enhance the overall robustness of the system.

[0097] The cloud processing center (100) continuously monitors the cluster's battery status by receiving real-time remaining battery information reported by each drone node (301, 302...30X) via heartbeat packets. The heartbeat packet contains the battery information. pct The field (integer, in percentage) is the source of the remaining battery power data. The cloud updates the corresponding node's battery status record every time it receives a heartbeat packet. Additionally, the heartbeat packet may also include a field showing the current CPU / NPU utilization rate (cpu). npu load(Integer, unit is percentage) The cloud comprehensively refers to two indicators, namely the remaining power and the real-time computing load, to evaluate the real-time working condition of each node, so as to more accurately trigger dynamic task redistribution decisions. When the remaining power of any drone node (e.g., 302) is lower than the preset power threshold, or when the power drops suddenly due to increased wind resistance, low temperature, etc., the drone node (302) sends a distress signal to the cloud processing center (100). The preset battery threshold can be set according to the drone model and return distance. For example, in a preferred embodiment, it is set to 20% of the drone node's full battery (for a typical consumer-grade multi-rotor drone, a full battery is about 5000mAh, 20% corresponds to about 1000mAh, and the corresponding battery voltage is about 3.5V / cell, which can support about 5 to 8 minutes of return flight time). This threshold not only provides sufficient safety margin for the original drone to return, but also allows the task handover to be completed in time before the battery is exhausted, avoiding monitoring gaps in the inspection area.

[0098] Upon receiving the distress signal, the cloud processing center 100 immediately executes the following dynamic reallocation process: First, it recalculates the remaining unpatched area of ​​drone node 302 and generates the remaining flight path tasks; second, it selects a neighboring drone node (e.g., 303) with sufficient remaining power in the cluster, dynamically splits the tasks for the remaining unpatched area, and assigns them to that node; finally, it controls the original drone node 302 to return to base early, and the neighboring node 303 seamlessly takes over the remaining sub-area tasks. Because the cloud continuously receives heartbeat packet power data from each node, task reallocation can be triggered immediately when the original node's power reaches a threshold. The entire switching process is imperceptible to the continuity of inspection coverage, ensuring uninterrupted execution of inspection tasks.

[0099] In extremely complex situations (such as heavy fog causing the local model's confidence level to consistently fall below the preset confidence threshold), if the drone node has sufficient capacity and a good communication signal, it can temporarily activate low-frame-rate high-definition image transmission to send single-frame images to the cloud processing center for secondary target detection confirmation (i.e., a computing power offloading backup mechanism). This mechanism maintains the principle of low-bandwidth backhaul while ensuring recognition reliability in extreme scenarios by uploading single-frame images to the cloud for secondary confirmation, achieving a flexible inspection that combines coarse and fine-grained methods.

[0100] Example 3: Alternative Solution for Unmanned Aerial Vehicle (UAV) Mesh Relay This embodiment provides an alternative implementation scheme for the wireless communication link 200, which is suitable for extreme outdoor environments without public network signal coverage (such as deep mountains, islands, remote reservoirs without base station coverage, etc.).

[0101] In this embodiment, the wireless communication link 200 is replaced by an Ad-hoc UAV self-organizing network (Mesh network) channel instead of a 4G / 5G channel that relies on the public network. Adjacent UAV nodes (301, 302...30X) directly exchange their respective identification result metadata through the self-organizing network protocol, without needing to go through the cloud processing center 100. Each UAV node still completes target detection inference and generates lightweight metadata messages on the edge side, and the edge inference logic is completely consistent with that in Embodiment 1. The master control UAV node ("leader") located at the center of the network topology collects the metadata messages of the surrounding nodes and uniformly executes fusion logic such as time window alignment, spatial distance calculation, and confidence weighted arbitration. The fusion algorithm is consistent with the algorithm executed by the cloud fusion analysis server 101 in Embodiment 1. The fusion result is then transmitted back to the cloud processing center 100 or directly pushed to the management terminal by the leader through satellite communication or other available channels.

[0102] The core advantage of this embodiment is that even in the event of a complete public network outage, the drone swarm can still rely on its self-organizing network to complete multi-drone collaborative sensing and data fusion, maintaining the normal execution of inspection tasks. Each drone node still only transmits lightweight metadata (rather than the original video stream), and the bandwidth requirements of the self-organizing network channel are extremely low, further enhancing the system's anti-interference capability in complex electromagnetic environments. In addition, this embodiment can also be used in conjunction with a hybrid fixed-wing and multi-rotor formation scheme: long-endurance fixed-wing drones equipped with powerful computing units are used to handle wide-area main area surveys, while multi-rotor drones are used specifically for close-range verification of complex boundary areas and suspected targets, effectively decoupling the load and computing power, and further expanding the system's applicable scenarios.

[0103] Example 4: Typical Application Scenarios for Inspecting Illegal Fishing in Reservoirs The following section uses typical scenarios of illegal fishing inspections in water environment management to specifically explain the technical solution and beneficial effects of this application.

[0104] The reservoir covers an area of ​​approximately 2 square kilometers, with open terrain and a winding shoreline, containing numerous hidden fishing spots. Traditional single-machine inspection solutions ("single-machine patrol + cloud-based centralized analysis") require approximately 40 minutes of flight time to fully cover the entire water surface. Furthermore, due to limitations in video stream transmission bandwidth, video stuttering and frame drops frequently occur in areas with weak signals at the reservoir's edge. Cloud analysis also suffers from gaps, resulting in a high rate of missed detections.

[0105] The proposed solution is deployed as follows: The reservoir is divided into four main sub-regions (Z1-Z4) and two overlapping boundary zones. These overlapping zones extend bufferably to both sides along the common edge of adjacent main sub-regions. The buffer extension distance is the sum of the UAV positioning error and the camera's field-of-view coverage margin. The system automatically calculates the ratio η between the total area of ​​the overlapping sub-regions and the total area of ​​the reservoir. Through iterative optimization, η is ensured to not exceed a preset area ratio threshold. Once the area constraint is met, the flight path tasks for each sub-region are distributed to the corresponding UAV nodes. Five consumer-grade multi-rotor UAVs equipped with lightweight AI computing units are simultaneously deployed, each responsible for approximately 0.4 square kilometers of water surface area.

[0106] During the inspection, each drone node's onboard camera captured water surface images at a frequency of 2 frames per second. The image data was directly fed into the edge AI computing unit for inference, running a lightweight target detection model compressed using INT8 quantization. When no target was detected, each drone node sent a heartbeat packet to the cloud processing center 100 every 30 seconds, reporting its location, battery level, and status. Approximately 15 minutes into the inspection, drone node 303 detected a fisherman within the boundary overlap zone. The edge AI computing unit immediately locked onto the target, captured a low-resolution keyframe thumbnail, and packaged it, along with the target category label "person-fishing," WGS-84 geographic coordinates, and model confidence, into a lightweight JSON metadata message for immediate transmission. Almost simultaneously, drone node 304, responsible for the adjacent master-sub-region Z3, also captured the same target and similarly transmitted a metadata message.

[0107] After receiving the two messages, the cloud-based fusion analysis server 101 first performs time window alignment, grouping the two messages into the same batch. Then, based on the WGS-84 geographic coordinates in the two messages, it calculates the geographic distance, determining that the distance is less than a preset spatial threshold, thus identifying them as the same target. It then performs confidence-weighted arbitration, retaining the record with the higher confidence level and storing it in the inspection database 102, and linking it with the dual-view thumbnails of drones 3 and 4 to form an evidence chain. Finally, it pushes a unique and accurate alarm to the management personnel's APP: "Suspected illegal fishing personnel found in the boundary area between drones 3 and 4; it is recommended to proceed with the response," along with thumbnails of both drones. The entire process, from edge identification to alarm push, takes no more than 2 seconds. After completing the full-area inspection and returning to base, the drone has more than 60% battery remaining, ready to immediately execute the next mission.

[0108] Compared to traditional standalone solutions, the proposed solution demonstrates the following improvements in the aforementioned scenarios: coverage time is reduced from approximately 40 minutes with a single machine to approximately 15 minutes with multiple machines operating in parallel, representing an efficiency increase of approximately 62%; end-to-end identification and warning latency is reduced from several to tens of seconds to less than 2 seconds; the false negative rate for targets in the boundary area is reduced to 0 through η-constrained boundary redundancy coverage; and the issue of duplicate reporting by multiple machines is accurately addressed through cloud-based spatiotemporal fusion deduplication, generating unique and accurate alarms. These improvements are achieved through a collaborative mechanism that integrates "boundary overlap design with area constraints - edge-side inference and de-streaming video transmission - cloud-based spatiotemporal fusion deduplication," none of which can be omitted.

[0109] Supplementary explanation of specific implementation methods In the above embodiments, the implementation of the computer-readable storage medium is as follows: a computer program is stored on the storage medium. When the computer program is executed by the processor, the processor performs the following steps: the cloud processing center 100 obtains the geographical boundary information of the target area, divides the target area into N main sub-regions according to the number N drones in the drone cooperative cluster 300, extends along the common edge of adjacent main sub-regions to both sides to generate boundary overlapping sub-regions, constrains the ratio η of the total area of ​​the boundary overlapping sub-regions to the total area of ​​the target area to not exceed a preset area ratio threshold, and packages the coordinate information of each main sub-region and the boundary overlapping sub-regions into a flight path task and sends it to the corresponding drone node; each drone node receives... After completing their flight path mission, the drones take off in coordination and fly to their assigned sub-areas to acquire images. The edge AI computing units of each drone node perform target detection inference on the acquired images at the edge. When a target is detected, a metadata message containing the target category, geographic coordinates, and confidence level is transmitted back to the cloud processing center 100 via the wireless communication link 200. The fusion analysis server 101 of the cloud processing center 100 performs time window alignment and spatial distance calculation on the multiple transmitted metadata messages. Multiple records within the same time window with geographic coordinate distances less than a preset spatial threshold are identified as the same target. A confidence-weighted arbitration is then performed on these records, retaining the record with the highest confidence level to generate an alarm message. When the processor executes the above program, additional steps such as iterative optimization of η constraints, heartbeat mechanism, Kalman filter trajectory completion, and dynamic task reallocation can also be driven by corresponding program instructions in the storage medium, thus achieving a complete functional correspondence with the aforementioned system implementation method.

[0110] Variants and Equivalents It is understood that, without changing the aforementioned core technical mechanism, this application also covers the following variant implementations: Variant 1: The edge AI computing unit is not limited to implementations based on RK3588, Jetson Nano, or Horizon Sunrise series NPU modules. It can also be other embedded AI processors with INT8 quantization inference capabilities, as long as they can be directly connected to the airborne camera via MIPI or USB interface and complete real-time target detection inference.

[0111] Variant 2: Lightweight object detection models are not limited to YOLOv5s or MobileNet-SSD. Other quantized and compressed single-stage object detection models, such as NanoDet and YOLOv8n, can also be used. The specific selection can be adjusted according to the computing power and accuracy requirements of the edge device.

[0112] Variant 3: The metadata backhaul channel is not limited to 4G / 5G IoT cards; it can also use LoRa remote transmission modules, NB-IoT modules, or drone self-organizing network channels, as long as they can carry JSON message transmissions of hundreds of bytes. In extreme environments without public network signal coverage, the Ad-hoc self-organizing network relay scheme described in Example 3 can also be used, where the lead unit aggregates and merges the data before transmitting it back via satellite communication.

[0113] Variant 4: The time window threshold, spatial distance threshold, and preset power threshold are all adjustable parameters. They can be adjusted according to the drone positioning accuracy, target movement speed, and drone model in the specific inspection scenario. The above adjustments do not affect the substance of the core technical solution of this application.

[0114] Variant 5: In the area division stage, visual SLAM (simultaneous localization and mapping) technology can be combined to enable drones to accurately determine which subgrid they are in even in areas without GPS signals, ensuring the spatial accuracy of the zoning inspection and further improving the system's applicability in complex terrain.

[0115] Variant 6: The collaborative drone swarm is not limited to homogeneous multi-rotor platforms. It can also adopt a hybrid formation scheme of fixed-wing and multi-rotor drones: long-endurance fixed-wing drones are responsible for the general survey of the main area, while multi-rotor drones are specifically responsible for the close-up verification of complex boundaries and suspected targets, thereby indirectly decoupling the load and computing power and further expanding the applicable scenarios of the system.

[0116] The foregoing has provided a detailed description of a multi-drone collaborative unmanned aerial vehicle (UAV) intelligent inspection system, method, and readable storage medium provided in the embodiments of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and core ideas of this application; furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

[0117] Certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that hardware manufacturers may use different names to refer to the same component. This specification and claims do not distinguish components based on differences in name, but rather on differences in function. The terms "comprising" and "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising / including but not limited to". "Approximately" means that within an acceptable margin of error, those skilled in the art can solve the technical problem and substantially achieve the technical effect within a certain margin of error. The following descriptions in the specification are preferred embodiments for carrying out this application; however, these descriptions are for the purpose of illustrating the general principles of this application and are not intended to limit the scope of this application. The scope of protection of this application shall be determined by the appended claims.

[0118] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a product or system comprising a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a product or system. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the product or system that includes said element.

[0119] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0120] The foregoing description illustrates and describes several preferred embodiments of this application. However, as previously stated, it should be understood that this application is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the application concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of this application should be within the protection scope of the appended claims.

Claims

1. A multi-drone collaborative unmanned aerial vehicle (UAV) intelligent inspection system, characterized in that, The intelligent drone inspection system includes: a cloud processing center and a drone collaborative cluster, wherein the drone collaborative cluster includes at least two drone nodes, each equipped with an edge AI computing unit; The cloud processing center divides the target area into several sub-regions and generates overlapping sub-regions, constraining the area ratio of the overlapping sub-regions to not exceed a preset area ratio threshold. Each of the aforementioned drone nodes performs target detection on the edge AI computing unit, and only sends back metadata messages containing target category, geographic coordinates and confidence level, without sending back the original video stream; The cloud processing center performs spatiotemporal alignment on multiple metadata messages, and after determining that they are the same target, performs confidence-weighted arbitration to retain the best record.

2. The UAV intelligent inspection system based on multi-machine collaboration according to claim 1, characterized in that, The cloud processing center combines the nominal capacity of the single-unit battery, the wind speed prediction model, and the peak power consumption of the airborne edge AI computing unit to calculate the effective coverage area threshold S of a single unit while retaining a safe redundancy of power. limit The target area is divided into N main sub-regions based on the number N drones in the drone collaborative swarm, and the area Si of each main sub-region is constrained to not exceed the specified value S. limit , where N is a positive integer; The cloud processing center extends bufferedly to both sides along the common edge of adjacent main sub-regions to generate boundary overlapping sub-regions, with a buffer extension distance D. buffer Satisfy D buffer = k·(σ rtk + σ hover ) + d fov , where σ rtk For RTK positioning accuracy, σ hover For drone hovering accuracy, d fov The horizontal coverage margin is calculated based on the cruise altitude and the camera's field of view, and k is the safety factor.

3. The UAV intelligent inspection system based on multi-machine collaboration according to claim 2, characterized in that, The cloud processing center is also configured to: calculate the ratio η of the total area of ​​the overlapping sub-regions to the total area of ​​the target region; if η exceeds the preset area ratio threshold, then initiate an iterative optimization loop: first, linearly reduce the buffer extension distance D by a ratio of 20% / η. buffer The buffer polygon is regenerated; if simply reducing the distance still fails to meet the constraints, the granularity of the principal sub-region division is further adjusted or the curved common edges are straightened to reduce the overlapping area of ​​the buffer zone; in each iteration, it is evaluated whether the new scheme simultaneously satisfies that η does not exceed the preset area ratio threshold and the area Si of all principal sub-regions does not exceed S. limit The iteration continues until the optimal solution that satisfies all constraints is found.

4. The UAV intelligent inspection system based on multi-machine collaboration according to claim 1, characterized in that, The cloud processing center performs spatiotemporal alignment of multiple metadata messages by: aligning the data transmitted within the same time window based on a time window threshold, and determining multiple records with geographical coordinate distances less than a preset spatial threshold as the same target based on spatial distance calculation.

5. The UAV intelligent inspection system based on multi-machine collaboration according to claim 1, characterized in that, The metadata message includes a keyframe thumbnail corresponding to the target; each drone node is also configured to send a heartbeat packet to the cloud processing center at a set time interval when the edge AI computing unit does not detect the target. The heartbeat packet includes the location information, remaining battery power, and operating status of the drone node.

6. The UAV intelligent inspection system based on multi-machine collaboration according to claim 1, characterized in that, The cloud processing center is also configured to: use the Kalman filter algorithm to process the geographic coordinate sequence of multiple records identified as the same target, and complete the motion trajectory of the target when flying across UAV nodes in the boundary overlapping sub-region.

7. The UAV intelligent inspection system based on multi-machine collaboration according to claim 5, characterized in that, The cloud processing center is further configured to: receive the remaining battery power information and CPU / NPU utilization information sent by each drone node via heartbeat packets; when the remaining battery power of any drone node is lower than a preset battery power threshold, recalculate the remaining unpatched area of ​​the drone node, dynamically split the tasks of the remaining unpatched area and assign them to nearby drone nodes with sufficient battery power, and control the drone node to return to its home location; the preset battery power threshold is 10% to 30% of the full battery power of the drone node.

8. The UAV intelligent inspection system based on multi-machine collaboration according to claim 1, characterized in that, Each of the aforementioned drone nodes is further configured to: when the local detection confidence of the edge AI computing unit is lower than a preset confidence threshold, send a single-frame image to the cloud processing center and request the cloud processing center to perform secondary target detection confirmation on the single-frame image.

9. A method for intelligent inspection of unmanned aerial vehicles (UAVs) based on multi-machine collaboration, employing the intelligent UAV inspection system based on multi-machine collaboration as described in any one of claims 1 to 8, characterized in that, The intelligent inspection method for unmanned aerial vehicles (UAVs) based on multi-machine collaboration includes the following steps: S1, the cloud processing center combines the nominal capacity of a single battery, wind speed prediction models, and peak power consumption inversion of the airborne edge AI computing unit to calculate the effective coverage area threshold S. limit Based on the number of drones N, the target area is divided into N main sub-regions, and the area of ​​each main sub-region is constrained to not exceed S. limit The boundary overlapping sub-region is generated by adaptively buffering and extending along the common edge of the adjacent main sub-region. The area ratio of the boundary overlapping sub-region is constrained to not exceed the preset area ratio threshold. A bow-shaped or spiral scanning route is generated for each sub-region and route conflict detection is performed. The route tasks of each main sub-region and the boundary overlapping sub-region are sent to the corresponding UAV node. N is a positive integer. S2. All the aforementioned drone nodes take off in coordination and fly to the airspace above the assigned sub-area. The onboard edge AI computing unit performs target detection inference on the edge of the acquired images and only sends back metadata messages containing target category, geographic coordinates and confidence level to the cloud processing center, without sending back the original video stream. S3. The cloud processing center performs spatiotemporal alignment on the multiple metadata messages sent back, and after determining that they are the same target, performs confidence-weighted arbitration, retains the best record, and generates alarm information.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of claim 9.

Images