Park inspection method and system based on unmanned aerial vehicle cooperative scheduling

Abstract

The present application relates to the technical field of unmanned aerial vehicle application, and more particularly to a park inspection method and system based on unmanned aerial vehicle cooperative scheduling, comprising the following steps: acquiring a park basic geographic base map, sequentially performing geographic hierarchical labeling, inspection area classification and problem point label library construction, and outputting an inspection topological map; through inspection point positioning and differentiated path planning, controlling the unmanned aerial vehicle cluster to collect multi-source inspection data and transmit it to the ground control center; through a double deep learning model, extracting and fusing park dynamic and static problem point features, and outputting real problem point feature information; based on the feature information, completing multi-dimensional evaluation and grading of the problem points, and generating an optimal scheduling scheme by fusing maintenance resource information; implementing the optimal scheduling scheme and forming a closed loop through secondary inspection verification by the unmanned aerial vehicle to update the park inspection database. The present application constructs a fine park inspection topological map, combines differentiated unmanned aerial vehicle scheduling and a double deep learning model, and is suitable for various inspection requirements and reduces costs and improves efficiency.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) application technology, and in particular to a park inspection method and system based on UAV collaborative scheduling. Background Technology

[0002] Currently, park inspection work mainly relies on manual inspection, with some parks supplemented by single drones for auxiliary inspection. Meanwhile, drone swarm collaborative scheduling technology has been applied in industrial fields such as power distribution networks and road traffic. Deep learning and computer vision technologies are also gradually being integrated into the feature recognition process of various industrial inspection scenarios, becoming an important means of intelligent upgrading of inspection technology.

[0003] Existing technologies have many obvious shortcomings in park inspection scenarios. Manual inspection is limited by terrain and field of vision, resulting in low inspection efficiency and blind spots. It is difficult to effectively detect problems such as dead branches on high ground and garbage in remote areas of the park, and it cannot quickly complete the inspection of large areas of the park after extreme weather. Single drone inspection has limited coverage and low operating efficiency, making it difficult to complete inspection tasks of multiple areas and types at the same time. Existing drone swarm collaborative scheduling technology is designed for industrial scenarios and cannot be adapted to the characteristics of park operations with dense crowds and complex environments. Its feature extraction model cannot identify fragmented static problems and dynamic human behavior problems in park inspections. At the same time, it lacks path planning and maintenance resource scheduling logic adapted to park scenarios, and cannot meet the actual needs of park inspections. Summary of the Invention

[0004] To address the technical deficiencies in the background technology, this invention proposes a park inspection method and system based on UAV collaborative scheduling, which solves the aforementioned technical problems and meets practical needs. The specific technical solution is as follows: The park inspection method based on drone collaborative scheduling includes the following steps: Obtain the park's basic geographic base map, then perform geographic layer labeling, inspection area classification, and problem point label library construction in sequence, and output a refined digital inspection topology map of the park; Based on the inspection topology map, the system controls the drone swarm to collect multi-source inspection data and transmit it to the ground control center through inspection point positioning and differentiated path planning. Based on multi-source inspection data, after preprocessing, the features of dynamic and static problem points in the park are extracted and fused through a dual deep learning model to output the feature information of real problem points. Based on feature information, a multi-dimensional assessment and classification of problem points is completed, and the optimal scheduling scheme is generated by integrating maintenance resource information. The optimal scheduling scheme is implemented and a closed loop is formed through secondary inspection and verification by drones. The park inspection database is then updated to complete the inspection.

[0005] Furthermore, the specific steps for outputting the refined digital inspection topology map of the park are as follows: Based on the park's basic geographic base map, accurate annotation of three layers of geographic information, including topographic layer, facility layer and control layer, is performed to output a park base map with layered geographic information. Based on the park base map with layered geographic information, the park is divided into fine inspection area, large area inspection area and dynamic monitoring area according to the inspection accuracy requirements, and the park base map with inspection area classification is output. Construct a pre-classification tag library for park problem points that includes static and dynamic problem point types and labels them with urgent, general, and minor handling priorities, and output a standardized park problem point tag library; By integrating park base maps with inspection area classifications and a standardized park problem point label library, a refined digital inspection topology map of the park is generated.

[0006] Furthermore, the specific steps for collecting multi-source inspection data are as follows: Based on the inspection topology map, combined with the daily inspection data of the park and the distribution pattern of historical problem points, the key inspection areas are located and real-time dynamic monitoring points are set up, and the park inspection topology map with inspection point markings is output. Based on the park inspection topology map with inspection point markings, the drone swarm is divided into functional sub-clusters and flight paths matching each inspection area are planned. Obstacle avoidance and human avoidance nodes are embedded in the paths, and differentiated inspection path planning schemes with obstacle avoidance nodes are output. Based on the differentiated inspection path planning scheme, the system controls each functional sub-cluster to perform inspection operations, collects image data and corresponding auxiliary data such as positioning, shooting time, and flight altitude, and outputs multi-source raw inspection data. The raw data from the multi-source inspection is transmitted to the ground control center in real time via wireless communication. After data integrity verification, valid multi-source inspection data is output.

[0007] Furthermore, the specific steps for outputting the feature information of the real problem points are as follows: Based on multi-source inspection data, the image data is preprocessed by image enhancement, deblurring and segmentation along the inspection path in sequence. Blurry and overexposed invalid image data are removed, and the preprocessed valid inspection image data and supporting auxiliary data are output. Based on effective inspection image data, static images are input into an improved convolutional neural network combined with an optimized spatial attention mechanism feature extraction model to extract features of static problem points and generate feature maps, outputting the feature map of static problem points in the park. Based on effective inspection image data, dynamic images are input into a lightweight YOLO model combined with a time-series analysis feature extraction model to extract features of dynamic problem points and generate feature vectors, outputting the feature vectors of dynamic problem points in the park. The feature map of static problem points in the park and the feature vector of dynamic problem points in the park are matched for dimension and fused. Duplicate and invalid feature information is removed and the feature information of real problem points in the park is output.

[0008] Furthermore, the specific steps for generating the optimal scheduling scheme are as follows: Input real problem point feature information, accurately match it with a standardized park problem point label library, complete a comprehensive evaluation from four dimensions, assign feature labels to each problem point and visualize them on the topology map, and output the multi-dimensional evaluation and classification results of park problem points; The system acquires real-time distribution information of park maintenance personnel, tools, and material reserves, as well as accessibility information of walkways and work passages within the park. After data integration, it outputs accessibility information of maintenance resources. The multi-dimensional assessment and grading results of park problem points and the accessibility characteristics of maintenance resources are input into the feature coupling module to complete the deep feature fusion and output the fused feature information of park problem points and maintenance resources. Based on the fusion of feature information, maintenance resources are matched from high to low according to processing priority and the optimal passage path is planned. The output is an optimal scheduling scheme for park maintenance resources that can be dynamically adjusted according to the real-time status of the park.

[0009] Furthermore, the specific steps for updating the park inspection database are as follows: The optimal scheduling plan for park maintenance resources is parsed into standardized scheduling instructions that include problem point information, handling requirements, supporting resources, and access routes, and then sent to the park maintenance management terminal via wireless communication. Maintenance personnel execute maintenance operations according to standardized dispatch instructions, upload work progress and on-site image data to the ground control center in real time, and output maintenance operation execution result data after data verification; Based on the maintenance operation execution results data, control the drone cluster to perform secondary inspection and verification of the processed problem points, collect verification images and extract verification feature information, and output the problem point verification feature information; The verification feature information of the problem point is accurately compared with the original feature information of the corresponding problem point. If it is qualified, the problem point is marked as closed loop. If it is not qualified, the scheduling plan is regenerated. Finally, all the data of this inspection are integrated to update the park inspection database and complete the inspection.

[0010] Furthermore, the construction steps of the improved convolutional neural network combined with the optimized spatial attention mechanism feature extraction model are as follows: An improved convolutional neural network is constructed, consisting of an input layer, a multi-scale convolutional layer, an activation layer, an adaptive average pooling layer, and a feature fusion layer. The multi-scale convolutional layer uses a combination of 3×3, 5×5, and 7×7 convolutional kernels and performs parallel convolution operations. The activation layer uses the ReLU6 activation function. The feature fusion layer integrates multi-scale features through a concatenation method. An optimized spatial attention mechanism module is constructed, which sequentially sets up a channel compression layer, a spatial feature extraction layer, a weight generation layer, and a weight normalization layer. The channel compression layer compresses the feature channel dimension through a 1×1 convolution kernel. The spatial feature extraction layer extracts spatial features by combining pooling and convolution. The weight generation layer generates a spatial attention weight matrix through a fully connected layer. The weight normalization layer normalizes the weight matrix to the 0-1 range through the Sigmoid function. The optimized spatial attention mechanism module is embedded between the adaptive average pooling layer and the feature fusion layer of the improved convolutional neural network base network, realizing feature interaction and linkage between the spatial attention mechanism module and the convolutional neural network base network, and completing the overall construction of the feature extraction model of the improved convolutional neural network combined with the optimized spatial attention mechanism.

[0011] Furthermore, the construction steps of the feature extraction model combining the lightweight YOLO model with time series analysis are as follows: The YOLO base model is lightweighted by replacing the standard convolutions in the backbone network with depthwise separable convolutions, reducing the number of detection heads while retaining the core detection branches. At the same time, anchor box adaptive clustering optimization is performed based on dynamic target sample data in parks to generate a lightweight YOLO target detection sub-model adapted to dynamic target detection in parks. A temporal analysis submodule is constructed, which sequentially sets up an inter-frame feature extraction layer, an optical flow calculation layer, and a temporal feature fusion layer. The inter-frame feature extraction layer extracts target feature points of consecutive frame images. The optical flow calculation layer calculates the inter-frame motion trajectory and displacement data of the feature points. The temporal feature fusion layer integrates single-frame target features and inter-frame temporal features using a feature stitching method. The temporal analysis submodule and the lightweight YOLO object detection submodel are fused on the edge. The output of the lightweight YOLO object detection submodel is matched with the input of the temporal analysis submodule by feature dimension matching, thus completing the overall construction of the feature extraction model that combines the lightweight YOLO model with temporal analysis.

[0012] Furthermore, the operation control of each functional sub-cluster of the drone swarm is performed according to the following steps: Based on the differentiated inspection path planning scheme with obstacle avoidance nodes, standardized operating parameters such as flight altitude, speed and image acquisition interval are set for each functional sub-cluster according to the inspection area type. Each functional sub-cluster performs inspection operations according to standardized operating parameters, and obstacle avoidance sensors detect personnel and obstacles within the flight path range in real time, outputting real-time obstacle avoidance monitoring data; If real-time obstacle avoidance monitoring data triggers a safety warning, the flight control module will dynamically adjust the path, control the drone to fly around the obstacle and update the inspection path, and output the adjusted drone inspection path. Based on the adjusted drone inspection path, the corresponding drone sub-clusters are controlled to complete the image acquisition of uncollected areas. After removing duplicate data, the park inspection multi-source data without omissions is output.

[0013] A park inspection system based on UAV collaborative scheduling includes a distributed multi-UAV cluster data acquisition terminal, a ground control center processing terminal, and a park maintenance and management terminal. The three are connected by a 5G wireless communication network to establish a two-way data interaction connection. The distributed multi-drone cluster data acquisition terminal consists of drone sub-clusters for refined inspection, large-area inspection, and dynamic monitoring. Each drone is equipped with a high-definition visible light camera, GPS positioning module, millimeter-wave obstacle avoidance sensor, flight control module, 5G wireless communication module, and local image storage module, which are used to realize multi-source data acquisition for park inspection, flight obstacle avoidance, and real-time data transmission. The ground control center processing terminal is an industrial-grade server cluster, integrating topology map construction, path planning, data preprocessing, dynamic and static dual feature extraction and analysis, problem assessment and classification, maintenance resource scheduling, closed-loop verification and UAV swarm control function modules. It has built-in improved convolutional neural network combined with spatial attention mechanism and lightweight YOLO combined with time series analysis feature extraction model, which is used to complete the data processing, assessment and scheduling and UAV operation management of the entire inspection process. The park maintenance management terminal includes mobile terminals for maintenance personnel and fixed terminals for the park management office, which are used to receive dispatch instructions, upload maintenance operation data, manage maintenance resources, and visualize inspection data and trace historical data.

[0014] Compared with existing technologies, the park inspection method and system based on UAV collaborative scheduling provided by the present invention has the following beneficial effects: This invention provides a precise geographic information foundation for drone inspections by constructing a refined digital topology map of parks. It employs a differentiated drone swarm scheduling strategy and considers obstacle and pedestrian avoidance requirements in path planning, improving both the efficiency of park inspections and ensuring drone flight safety in densely populated areas. Combined with a dual deep learning feature extraction model, it accurately identifies static and dynamic problem points in parks. Furthermore, through multi-dimensional problem point assessment and grading coupled with the feature-based scheduling of maintenance resources, it achieves optimal allocation of maintenance resources. This, along with a complete inspection loop formed by secondary drone inspections and verification, comprehensively enhances the intelligence and precision of park inspections. It effectively solves the problems of low efficiency, inaccurate identification, and unreasonable resource scheduling in existing inspection methods. It is adaptable to the daily and emergency inspection needs of various parks, significantly reducing the workload and cost of manual inspections. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the park inspection method based on UAV collaborative scheduling in this invention.

[0016] Figure 2 This is a schematic diagram of the park inspection system based on UAV collaborative scheduling in this invention. Detailed Implementation

[0017] In the description of this invention, it should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "middle," and "inner," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, it should be noted that unless otherwise explicitly specified and limited, the terms "installed," "connected," and "joined" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention through specific circumstances.

[0018] The embodiments of the present invention will now be described with reference to the accompanying drawings and related examples. However, the embodiments of the present invention are not limited to the following examples.

[0019] See Figure 1 This invention provides a park inspection method based on UAV collaborative scheduling, comprising the following steps: Step S100: Obtain the basic geographic base map of the park, and then perform geographic layer labeling, inspection area classification and problem point label library construction in sequence to output a refined digital inspection topology map of the park. The park's basic geographic base map is fundamental map data containing the park's original geospatial information. It is typically generated from remote sensing imagery, GIS systems, or surveying data, serving as a data source for constructing higher-level inspection maps and providing geographic benchmarks and terrain contour information. Geographic layer labeling is the process of dividing the park's basic geographic base map into multiple logical layers according to different geographic element categories and semantically tagging it. This enables structured representation of geographic elements, facilitating conditional judgments in subsequent area division and path planning. Inspection area classification is the operation of dividing park space based on regional functional attributes, risk levels, or inspection priorities. This provides a basis for differentiated scheduling strategies and supports on-demand allocation of inspection resources. The problem point label library is a standardized set of labels storing common park problem types. This unifies the problem identification and recording format, ensuring consistency and traceability of problem identification results, and supporting feature extraction model training and evaluation grading logic. The park's refined digital inspection topology map is a multi-dimensional digital map integrating geospatial structure and semantic information. It includes layered geographic labeling, inspection area classification, and problem point label mapping relationships. This provides high-precision spatial reference and semantic context support for UAV inspection tasks, supporting positioning, path planning, and problem matching.

[0020] Step S200: Based on the inspection topology map, through inspection point positioning and differentiated path planning, control the UAV cluster to collect multi-source inspection data and transmit it to the ground control center; Inspection point location determines the specific spatial coordinates of key inspection areas on the inspection topology map, clarifying the target locations for UAV data collection and improving the targeting of problem detection. Differentiated path planning develops differentiated UAV flight path generation strategies based on the characteristics of different areas, balancing inspection coverage and flight safety, and adapting to multi-target constraints in complex park environments. UAV swarms are operational units composed of multiple UAVs with communication and collaboration capabilities, capable of performing distributed inspection tasks, enabling parallel inspections across multiple areas, improving overall operational efficiency and task fault tolerance. Multi-source inspection data is a heterogeneous data set reflecting the park's status, collected by UAVs equipped with various sensors, enriching the information dimensions for problem identification and enhancing the perception of concealed or low-contrast targets. The ground control center is the central control system responsible for receiving, processing, and directing UAV inspection tasks, enabling centralized monitoring and remote intervention of the entire inspection process.

[0021] Step S300: Based on multi-source inspection data, after preprocessing, extract and fuse the features of dynamic and static problem points in the park through a dual deep learning model, and output the feature information of real problem points; The dual deep learning model is a composite feature extraction system consisting of an improved convolutional neural network and a lightweight YOLO model. It aims to balance recognition accuracy and computational efficiency, simultaneously handling static object anomalies and dynamic crowd behavior. Static and dynamic problem point features are quantifiable representations of object anomalies that do not change rapidly over time and instantaneous activity anomalies in the park environment, respectively, providing input for subsequent fusion recognition and assessment grading. Real problem point feature information consists of valid problem point data confirmed after fusion processing by the dual deep learning model, eliminating false detections and noise interference. This data serves as reliable input for problem assessment and resource scheduling, reducing the false alarm rate.

[0022] Step S400: Based on feature information, complete the multi-dimensional assessment and classification of problem points, and integrate maintenance resource information to generate the optimal scheduling scheme; Multi-dimensional problem assessment and grading is a process of evaluating the severity and urgency of identified problems from multiple independent dimensions. This assessment establishes a priority sequence for problem handling and guides resource allocation. Maintenance resource information is a dataset of available maintenance personnel, vehicles, equipment, and their current status. It provides realistic constraints and capacity matching criteria for generating scheduling plans. The optimal scheduling plan is a set of task assignments and action instructions generated by integrating problem assessment results with maintenance resource status. This ensures efficient matching between problem handling tasks and available resources.

[0023] Step S500: Implement the optimal scheduling scheme and verify it through secondary inspection by drone to form a closed loop, update the park inspection database, and complete the inspection.

[0024] Secondary drone inspection and verification is an operation where, after a problem has been addressed, a drone is dispatched again to re-inspect the original problem area to confirm the effectiveness of the rectification. This is used to verify the effectiveness of the handling and prevent omissions or false closures. The park inspection database is a structured data warehouse that stores data from the entire inspection process over a long period of time. It includes problem records, handling logs, and historical images, and is used to support inspection performance evaluation, trend prediction, and knowledge accumulation.

[0025] Taking emergency inspection of urban parks after heavy rain as an example, the application method of this invention is as follows: After heavy rain, the system automatically calls up the park's basic geographic base map, combines historical water accumulation points and drainage facility distribution to complete geographic layer labeling, delineates inspection area categories such as flood-prone areas and high-risk areas for fallen trees, and loads problem labels such as water accumulation, landslides, and broken branches. After generating a refined inspection topology map, dense inspection points are set for high-risk areas around the river, and low-altitude close-fitting path planning is used to avoid obstructions; a swarm of drones takes off according to differentiated paths, carrying visible light and infrared cameras to collect multi-source data, discovering a slight collapse of the slope and water accumulation on the walkway. After the data is transmitted back to the ground control center, dual deep learning models respectively identify structural damage features and water flow retention features, and merge and output the real problem point information. Based on this, the system assesses its impact range and urgency level, and combines the location of nearby standby engineering vehicles and the idle status of emergency repair teams to generate the fastest response dispatch plan. After the repairs are completed, a drone is dispatched to conduct a second inspection of the area to confirm that the collapse has been reinforced and the water has been drained. Once the verification is passed, the database is updated to form a complete closed loop.

[0026] This invention provides a comprehensive park inspection solution that deeply integrates drone swarm collaborative scheduling, deep learning, computer vision, and geospatial information processing technologies into the park inspection scenario. This not only provides precise geographic information and path planning for drone inspections, ensuring flight safety and inspection efficiency in densely populated and complex park environments, but also achieves accurate identification of both static, fragmented, and dynamic behavioral issues through dual deep learning models. Combined with multi-dimensional assessment and feature-coupled maintenance resource scheduling, it achieves optimal allocation and dynamic adjustment of maintenance resources. Finally, a complete inspection loop is formed through secondary drone inspection verification, comprehensively improving the intelligence, precision, and standardization of park inspection work. This effectively solves the pain points of traditional manual inspections, such as low efficiency, numerous blind spots, limited coverage of single drone inspections, inaccurate problem identification, and unreasonable resource scheduling. It is adaptable to the daily inspection needs of various parks and emergency inspections after extreme weather, significantly reducing the workload and management costs of manual inspections. Furthermore, the continuous updating of the inspection database provides data support for the long-term maintenance and management of parks.

[0027] In one embodiment of the present invention, the specific steps for outputting the refined digital inspection topology map of the park are as follows: Step S101: Based on the park's basic geographic base map, accurately label the three layers of geographic information, including the topography layer, facility layer, and control layer, and output the park base map with layered geographic information; Using high-precision electronic geographic base maps of parks retrieved from urban geographic information public service platforms or park management systems as a foundation, vectorized and precise annotations are performed according to three dimensions: topography layer, facility layer, and control layer. Each layer of annotation completes the structured entry and visualization of information, and the information in the three layers is independent of each other but can be retrieved in a linked manner.

[0028] During the topographic layer labeling process, the natural geographical features within the park are accurately labeled, including the direction, width, and material of the trails, the boundaries, area, and depth of water bodies, the slope and direction of the hillsides, and the boundaries, vegetation types, and areas of the green areas, providing a topographic basis for subsequent UAV flight path planning. During the labeling process at the facility level, various man-made facilities within the park are accurately labeled, including leisure facilities such as benches, fitness equipment, and rest pavilions; amusement facilities such as slides and swings; functional facilities such as public toilets, streetlights, and trash cans; and power facilities such as distribution boxes and monitoring poles. The labeling content includes the spatial location, specifications, installation time, and maintenance status of the facilities, thereby achieving digital archiving of the facilities. During the labeling process at the control level, control areas are labeled in accordance with park management regulations and drone flight safety requirements. These include drone no-fly zones, densely populated areas such as children's playgrounds, restricted areas such as green core areas and slope protection areas, and key inspection areas such as concentrated facilities areas and waterfront areas. The labeling content includes the boundaries of the control areas, control requirements, and inspection frequency, providing a basis for subsequent drone operation control. After completing the three-layer geographic information labeling, data verification and integration are performed in GIS tools to remove invalid data with labeling errors and missing information, generating a park base map with layered geographic information. This base map supports single-level viewing and multi-level linked viewing.

[0029] Step S102: Based on the park base map with layered geographic information, divide the park into fine inspection area, large area inspection area and dynamic monitoring area according to the inspection accuracy requirements, and output the park base map with inspection area classification. Based on the park base map with layered geographic information, and combined with the density of people's activities, the distribution of facilities, and the level of inspection needs in each area of ​​the park, the park is divided into three types of inspection areas: fine inspection area, large area inspection area, and dynamic monitoring area. The division results are marked on the base map through vector boundaries to realize the digital definition of the inspection areas. The principle of high inspection accuracy, small coverage area, and high frequency is to delineate refined inspection areas; the principle of moderate inspection accuracy, large coverage area, and regular frequency is to delineate large-area inspection areas; and the principle of frequent dynamic problems, large personnel flow, and the need for real-time monitoring is to delineate dynamic monitoring areas. Children's playgrounds, public toilets, rest pavilions, and areas with concentrated facilities are designated as refined inspection areas; park green areas, suburban areas, artificial water bodies, and green belts surrounding parks are designated as large-area inspection areas; park entrances, main roads, restricted areas, and waterfront walkways are designated as dynamic monitoring areas. In the GIS tool, add exclusive attribute labels and vector boundaries to the three types of inspection areas. After completing the area division, perform data verification to ensure that there is no overlap or omission in each area, and generate a park base map with inspection area classification. This base map can directly provide a basis for the task allocation of subsequent drone sub-swarms.

[0030] Step S103: Construct a pre-classification tag library for park problem points that includes static problem points and dynamic problem point types and labels them with urgent, general and minor processing priorities, and output a standardized park problem point tag library; Based on the actual needs of park inspection, we sorted out various problems that are likely to occur during park operation, constructed a dual classification system of static and dynamic problem points, and assigned three levels of handling priority to all problem points: urgent, general, and minor. We also built a standardized and scalable pre-classification tag library for park problem points through database tools, providing a unified standard for subsequent problem point identification, assessment and classification. First, identify static issues within the park that do not change dynamically, including four main categories: facility damage, green space maintenance, environmental cleanliness, and static safety hazards. Then, identify dynamic issues within the park that arise due to human activities and environmental changes, including three main categories: human violations, temporary illegal facilities, and sudden safety incidents. Each main category is further subdivided into specific subcategories to achieve a refined classification of issues. Establish clear priority criteria: emergency priority is defined based on the possibility of causing personal safety accidents and requiring immediate handling; general priority is defined based on the impact on the normal operation of the park and requiring handling within a specified time; and minor priority is defined based on the minimal impact on the park's operation and can be included in routine maintenance. Match the corresponding priority to each specific problem point. A pre-classification tag library for park problem points was constructed using relational database tools. Each problem point was assigned a unique tag code, problem type, problem description, processing priority, processing specifications, and responsible party, among other structured fields. After the tag library was constructed, data testing and optimization were carried out to ensure that the tag library could be spatially correlated with the subsequent geographic base map. Finally, a standardized tag library for park problem points was output.

[0031] Step S104: Integrate the park base map with inspection area classification and the standardized park problem point label library to generate a refined digital inspection topology map of the park.

[0032] By using geospatial association technology, the park base map with inspection area classification is deeply integrated with the standardized park problem point label library to achieve the integration of geospatial information, inspection management information and problem point label information, and finally generate a refined digital inspection topology map of the park that can support the subsequent full-process inspection. By leveraging the spatial association function of GIS tools, various labels in the standardized problem point label library are associated with corresponding geographical areas and facilities on the park base map, thus achieving a mapping relationship between specific areas / facilities and specific problem point labels. For example, the "facility damage" label is associated with geographical facilities such as amusement facilities and leisure seats, and the "personnel violation" label is associated with geographical areas such as park entrances and restricted area boundaries. After completing the information association, a comprehensive verification of all integrated data is conducted to identify and address issues such as errors and omissions in the association between geographic information and tag information. At the same time, the visualization effect of the map is optimized, and the filtering, retrieval, and visual annotation of problem point tags are supported. The integrated geographic information, inspection area information, and problem point label information are packaged into a single package to generate a refined digital inspection topology map of the park. This map supports online editing, data updates, and visualization, and can be directly used as the core operational basis for subsequent drone inspections.

[0033] This invention, through layered geographic information annotation, inspection area classification, problem point label library construction, and information integration, achieves for the first time the digital, structured, and integrated fusion of park geographic information, inspection management information, and problem point label information. It constructs a refined digital inspection topology map for parks adapted to drone-based collaborative inspections, addressing industry pain points such as the lack of accurate digital geographic information support, the absence of standardized inspection area divisions, and the lack of a unified labeling system for problem point identification in traditional park inspections. The three-layered geographic information annotation enables a refined presentation of park terrain, facilities, and management information, providing geographic assurance for subsequent drone flight safety. Area classification based on inspection accuracy provides a standardized basis for differentiated path planning and task allocation for subsequent drone sub-swarms. The priority-based problem point pre-classification label library lays a unified standard system for rapid problem point identification, multi-dimensional assessment and grading, and maintenance resource scheduling. The integrated topology map becomes the foundational carrier of the entire drone-based collaborative scheduling park inspection method, enabling the linked retrieval and visualization of all basic information, significantly improving the standardization, accuracy, and intelligence of subsequent inspection processes, and providing core support for the efficient implementation of the entire inspection method.

[0034] In one embodiment of the present invention, the specific steps for collecting multi-source inspection data are as follows: Step S201: Based on the inspection topology map, combined with the daily inspection data of the park and the distribution pattern of historical problem points, locate the key inspection areas and set real-time dynamic monitoring points, and output the park inspection topology map with inspection point markings. Retrieve operational data such as daily park inspection records, equipment maintenance ledgers, and historical problem statistics, and spatially correlate them with the park's refined digital inspection topology map to identify core areas such as areas with high incidence of facility damage, areas with concentrated safety hazards, and areas with high frequency of personnel activity within the park; Based on the integrated spatial geographic information and operational data, spatial clustering analysis technology is used to locate fixed key inspection areas in the topology map, mainly including park amusement facilities areas, waterfront walkways, slope protection, and concentrated facility areas; at the same time, in the dynamic monitoring area of ​​the topology map, real-time dynamic monitoring points are set according to the density of personnel flow and the probability of dynamic problems, mainly at park entrances, main road intersections, and restricted area boundaries. In the park's refined digital inspection topology map, exclusive spatial attribute labels are added to fixed key inspection areas and real-time dynamic monitoring points. The labeling content includes the spatial coordinates of the inspection point, inspection frequency, monitoring requirements, and types of core concerns. After the labeling is completed, the map data is verified to ensure that no inspection points are omitted or duplicated. Finally, a park inspection topology map with inspection point labels is output.

[0035] Step S202: Based on the park inspection topology map with inspection point markings, divide the drone cluster into functional sub-clusters and plan flight paths matching each inspection area. Embed obstacle avoidance and human avoidance nodes in the paths and output differentiated inspection path planning schemes with obstacle avoidance nodes. Based on the three-category area division standard of the park, namely the refined inspection area, the large-area inspection area, and the dynamic monitoring area, the distributed drone cluster is divided into refined inspection sub-cluster, large-area inspection sub-cluster, and dynamic monitoring sub-cluster. Each sub-cluster is configured with matching drone hardware parameters, and each sub-cluster independently receives instructions from the ground control center and can work collaboratively. Using the A* algorithm in UAV path planning, combined with a park inspection topology map with inspection point markings, flight paths are planned and matched to the corresponding inspection area operation requirements for each functional sub-cluster. The fine inspection sub-cluster path is a low-altitude, small-area fine coverage path, the large-area inspection sub-cluster path is an equally spaced, gridded full-area coverage path, and the dynamic monitoring sub-cluster path is a cyclical, high-frequency dynamic monitoring path. All paths must cover the fixed key inspection areas and real-time dynamic monitoring points within the corresponding area. By combining the control layer information in the park inspection topology map, obstacle avoidance and human avoidance nodes are embedded in the planned flight path. By setting spatial distance thresholds, fixed obstacles such as trees, buildings, and power facilities in the park, as well as control areas such as densely populated areas and no-fly zones, are set as no-passage zones on the path. The node labels include information such as obstacle avoidance range, detour direction, and safe distance. After completing the sub-cluster division and path planning, the rationality, safety and coverage of the path are simulated and verified, redundant road segments in the path are optimized, and finally a differentiated path planning scheme with obstacle avoidance nodes is generated. This scheme can be directly sent to the UAV cluster control module of the ground control center.

[0036] Step S203: Based on the differentiated inspection path planning scheme, control each functional sub-cluster to perform inspection operations, collect image data and corresponding positioning, shooting time, and flight altitude auxiliary data, and output multi-source inspection raw data. Based on the differentiated path planning scheme, the ground control center configures corresponding flight operation parameters for each functional sub-cluster, including image acquisition resolution, frame rate, flight altitude, speed, etc., to ensure that the acquired data matches the inspection accuracy requirements of each area. The ground control center issues inspection operation instructions, and each functional sub-cluster executes the inspection operation synchronously according to the planned path and operation parameters. The fine inspection sub-cluster collects close-up images of the inspection area, the large-area inspection sub-cluster collects gridded panoramic grid images, and the dynamic monitoring sub-cluster collects continuous real-time continuous frame images. While collecting image data, each drone simultaneously collects supporting auxiliary data through its onboard GPS positioning module and flight status monitoring module, including the GPS positioning coordinates, shooting time, and flight altitude corresponding to each frame of image. The auxiliary data is correlated with the image data to form structured multi-source inspection raw data. Each drone collects raw inspection data from multiple sources and stores it synchronously in the local image storage module to prevent data loss during transmission and achieve local data backup.

[0037] Step S204: Transmit the original multi-source inspection data to the ground control center in real time via wireless communication. After data integrity verification, output valid multi-source inspection data.

[0038] During the inspection operation, each drone sub-swarm transmits the locally stored multi-source inspection raw data to the ground control center in real time and at high speed through the onboard 5G wireless communication module. The transmission process uses data encryption to ensure the security of data transmission. The multi-source data receiving and preprocessing module of the ground control center presets data verification rules to verify the received multi-source inspection raw data from three dimensions: data field integrity, image data validity, and auxiliary data matching. It removes invalid data such as blurred, overexposed, or missing images, auxiliary data that is not related to image data, and data fields that are missing. After data verification, the valid data is structured and stored according to inspection area, sub-cluster type, and collection time. Finally, valid multi-source inspection data is output, which will be used as the basis for subsequent image preprocessing and problem point feature extraction.

[0039] In one embodiment of the present invention, the specific steps for outputting the feature information of the real problem points are as follows: Step S301: Based on multi-source inspection data, perform image enhancement, deblurring and segmentation along the inspection path preprocessing operations on the image data in sequence, remove blurry and overexposed invalid image data, and output the preprocessed valid inspection image data and supporting auxiliary data. Histogram equalization and contrast adaptive enhancement algorithms are used to process image data in multi-source inspection data, improve the brightness, contrast and detail recognition of the image, and solve the problem of blurred image details caused by light and shooting angle during UAV inspection. A non-blind deconvolution algorithm is used to deblur and repair blurred images caused by drone flight jitter and air disturbance, restoring the clear texture and feature details of the image, and ensuring that the image has the basic conditions for feature recognition. Based on the drone's flight inspection path and GPS positioning information, the continuously collected panoramic images and frame images are systematically segmented so that each segmented image corresponds to a fixed geographic location in the park, achieving a precise correlation between image features and geographic coordinates. After completing the above operations, invalid image data that is blurry, overexposed, or lacks texture is removed according to preset standards. At the same time, the corresponding auxiliary data such as GPS positioning, shooting time, and flight altitude of the valid image data are retained to ensure that the valid image data and auxiliary data fields are completely matched. Finally, the preprocessed valid inspection image data and supporting auxiliary data are output.

[0040] Step S302: Based on effective inspection image data, input static images into an improved convolutional neural network combined with an optimized spatial attention mechanism feature extraction model to extract static problem point features and generate feature maps, and output the static problem point feature map of the park. From the preprocessed valid inspection image data, static scene images without dynamic targets such as park facilities, greenery, and environment are selected, and then sorted and organized according to the park's geographical area before being input into a customized feature extraction model. The improved convolutional neural network extracts low-level visual features and high-level semantic features at different scales in static images through parallel operation of multi-scale convolutional kernels. The optimized spatial attention mechanism module assigns high weights to small static problem areas such as damaged facilities, bald patches on lawns, and scattered garbage in the image, thereby enhancing the extraction effect of target features and suppressing the interference of background-irrelevant features. The model integrates the extracted static problem point features in a structured manner to generate a two-dimensional feature map corresponding to the geospatial location of the static image. The feature map contains core information such as the location, type, and feature representation of the problem point. Finally, it outputs a static problem point feature map of the park, and this feature map is associated with the corresponding auxiliary data.

[0041] Step S303: Based on effective inspection image data, input the dynamic image into the lightweight YOLO model and combine it with the feature extraction model of time series analysis to extract the features of dynamic problem points and generate feature vectors, and output the feature vectors of dynamic problem points in the park. From the preprocessed effective inspection image data, continuous frame images with dynamic targets such as park entrance, main road, and restricted area boundary are selected, sorted and organized according to time series, and then input into a customized feature extraction model. The lightweight YOLO model first performs target detection on a single frame of dynamic image, extracting single-frame features of dynamic targets such as people and temporary illegal facilities. The time series analysis submodule then calculates the motion trajectory and displacement data of dynamic targets in continuous frames using optical flow method, extracting the behavioral features of dynamic targets, thus achieving an upgrade from "single-frame target recognition" to "continuous behavioral feature extraction". The model vectorizes the extracted single-frame features and temporal behavioral features of dynamic problem points to generate a fixed-dimensional feature vector for dynamic problem points in the park. The vector contains core information such as the target type, behavioral features, occurrence time, and geographical location of the dynamic problem point. Finally, the model outputs the feature vector for dynamic problem points in the park, and this feature vector is associated with the corresponding auxiliary data.

[0042] Step S304: Perform dimensional matching and data fusion on the feature map of static problem points in the park and the feature vector of dynamic problem points in the park, remove duplicate and invalid feature information, and output the feature information of real problem points in the park.

[0043] Based on the park's geospatial coordinates and supporting auxiliary data, the two-dimensional static problem point feature map and the one-dimensional dynamic problem point feature vector are unified in dimension and spatially matched, so that the dynamic and static problem point features in the same geographical location form a corresponding relationship. An algorithm combining feature splicing and channel fusion is used to deeply fuse the matched static problem point features with the dynamic problem point features to form a comprehensive feature set containing all problem points in the park; By using feature filtering algorithms, duplicate features, background-irrelevant features, and low-discrimination features in the comprehensive feature set are removed, while core features that can truly represent the problem points in the park are retained. The optimized core feature set is then integrated with information such as park geospatial coordinates and problem point types to ultimately output real problem point feature information for the park, including the spatial location, type, feature representation, and occurrence time of the problem point.

[0044] In one embodiment of the present invention, the specific steps for generating the optimal scheduling scheme are as follows: Step S401: Input the feature information of real problem points, accurately match it with the standardized park problem point label library, complete the comprehensive evaluation from four dimensions, assign feature labels to each problem point and visualize them on the topology map, and output the multi-dimensional evaluation and classification results of park problem points. The feature information of real problem points in parks is matched precisely at the field level with a standardized park problem point tag library. Based on the feature representation of the problem points, the problem type and initial processing priority are determined, thereby achieving standardized classification and definition of problem points. Each problem point is quantitatively scored and comprehensively evaluated from four dimensions: problem type, severity, scope of impact, and urgency of handling. Problem type is assigned basic weights according to categories such as facilities, greening, environment, and safety. Severity is classified into levels according to the degree of damage and development trend of the problem. Scope of impact is quantified according to the geographical area involved in the problem and the range of personnel activities. Urgency of handling is defined according to the possible harmful consequences of the problem and the time limit for handling. The comprehensive evaluation score of each problem point is obtained through weighted calculation. Each problem point is assigned a feature label containing a unique identifier, problem type, comprehensive evaluation score, final processing priority, and scope of impact. At the same time, each problem point is visually marked in the park's refined digital inspection topology map, and the marking information is associated with the spatial location of the problem point. The assessment information of all problem points is structured and sorted from high to low according to processing priority, generating a multi-dimensional assessment and classification result of park problem points that includes the spatial location of the problem points, feature labels, and comprehensive assessment results.

[0045] Step S402: Obtain real-time distribution information of park maintenance personnel, tools and material reserves, as well as access status information of walkways and work passages within the park. After data integration, output accessibility characteristic information of maintenance resources. The park maintenance management system can retrieve real-time information on maintenance personnel's location coordinates, work assignments, on-duty status, available personnel, type, quantity, storage location, and availability of tools, as well as the type, specifications, inventory quantity, and allocation route of materials, enabling dynamic archiving of maintenance resources. By using video surveillance, IoT sensing devices, and manually reported information within the park, the access status of walkways and work passages is obtained, including whether the roads are clear, whether there are construction barriers, and whether the efficiency of access is affected by factors such as weather, so as to quantify the accessibility of each area. By spatially linking real-time distribution information of maintenance resources with park access status information, and using the park's refined digital inspection topology map as a base, the spatial distance, access path, and allocation time of different maintenance resources to each problem point are analyzed. Core features including the availability of maintenance resources, path accessibility, and allocation efficiency are extracted, and finally, the accessibility feature information of maintenance resources is output.

[0046] Step S403: Input the multi-dimensional assessment and grading results of park problem points and the accessibility feature information of maintenance resources into the feature coupling module to complete the deep feature fusion and output the fused feature information of park problem points and maintenance resources; The data format is standardized and the dimensions are unified for the multi-dimensional assessment and grading results of park problem points and the accessibility characteristics of maintenance resources. Both types of characteristics are mapped to the spatial coordinate system of the park's refined digital inspection topology map to achieve a one-to-one correspondence of spatial dimensions. The two types of features are input into a preset feature coupling module. This module is based on deep learning algorithms and resource scheduling models. It associates and matches the processing priority, impact range, processing requirements of problem points with the availability, path accessibility, and operational capabilities of maintenance resources. Through feature weighting and correlation analysis, it achieves deep integration, removes irrelevant feature information, and retains the core content that has decision-making value for resource scheduling. The fused feature information is structured and organized to generate a list of available maintenance resources, core parameters for resource allocation, and basic information on work paths for each problem point. Finally, the fused feature information of the park's problem points and maintenance resources is output.

[0047] Step S404: Based on the fused feature information, match maintenance resources from high to low processing priority and plan the optimal passage path, and output the optimal scheduling scheme for park maintenance resources that can be dynamically adjusted according to the real-time status of the park.

[0048] According to the priority of handling the problem points from high to low, the most suitable maintenance resources are matched for each problem point, including assigning corresponding maintenance personnel, work tools and material reserves, following the principle of "nearby allocation, capacity matching and optimal resource utilization" to avoid resource idleness and duplication. Based on the park's refined digital inspection topology map and combined with the park's traffic status information, the optimal passage route from resource storage / personnel location to problem point is planned for each group of allocated maintenance resources. The route planning takes into account traffic efficiency and operational convenience, while reserving redundant space for route adjustment. The scheduling plan is set with dynamic adjustment trigger conditions. When new problem points appear in the park, the status of maintenance resources changes, or the park's access status changes, the plan can be automatically triggered to re-match and re-plan the route, thus achieving dynamic updates. By integrating the results of maintenance resource matching, optimal access routes, operation requirements, completion deadlines, and dynamic adjustment rules, a structured and executable optimal scheduling plan for park maintenance resources is generated. This plan includes specific scheduling information for single problem points and collaborative scheduling information for multiple problem points, and can be directly distributed to the park maintenance management terminal.

[0049] In one embodiment of the present invention, the specific steps for updating the park inspection database are as follows: Step S501: Parse the optimal scheduling plan for park maintenance resources into standardized scheduling instructions that include problem point information, processing requirements, supporting resources and access routes, and send them to the park maintenance management terminal via wireless communication; The maintenance resource scheduling module of the ground control center performs structured analysis on the optimal scheduling plan for park maintenance resources, extracts the core information corresponding to each problem point, such as unique identifier, spatial location, problem type, processing requirements, supporting maintenance resources, optimal access route, and completion time limit, and integrates them into standardized scheduling instructions with unified format and clear content, so as to ensure that maintenance personnel can directly execute operations according to the instructions; Based on the responsible party and the execution end of the maintenance operation, standardized dispatch instructions are simultaneously sent to the mobile terminals of maintenance personnel and the fixed terminals of the park management office through 5G wireless communication technology. The mobile terminals receive specific operation execution instructions, and the fixed terminals receive operation overall supervision instructions, thereby realizing the synchronization of operation execution and supervision. The ground control center stores and records all standardized dispatch instructions in a unified manner, associates them with the characteristic information of the corresponding problem points, and forms an instruction issuance ledger to facilitate subsequent operation traceability and data verification.

[0050] Step S502: Maintenance personnel execute maintenance operations according to standardized dispatch instructions, upload operation progress and on-site image data to the ground control center in real time, and output maintenance operation execution result data after data verification. After receiving standardized dispatch instructions via mobile terminals, maintenance personnel will carry the corresponding work tools and materials to the corresponding problem points according to the instructions, strictly follow the handling specifications to carry out maintenance work, and record key operation nodes as required during the operation. During and after the operation, maintenance personnel use mobile terminals to upload data such as operation progress, on-site operation images, and real-time images after problem handling to the ground control center in real time. The image data must be linked to GPS positioning and shooting time to ensure the authenticity and traceability of the data. The closed-loop verification module of the ground control center performs dual verification on the received maintenance operation data. First, it verifies the completeness of the data, confirming that key data such as operation progress and on-site images are not missing. Second, it verifies the authenticity of the data, eliminating false data through image comparison, positioning information verification and other methods. The verified maintenance operation data is structured and organized, and corresponding problem points and scheduling instructions are linked to generate maintenance operation execution result data that includes the operation subject, execution process, processing results, and on-site scene, providing a basis for subsequent secondary verification of drones.

[0051] Step S503: Based on the maintenance operation execution result data, control the drone cluster to perform secondary inspection and verification of the processed problem points, collect verification images and extract verification feature information, and output the problem point verification feature information; Based on the maintenance operation execution results data, the UAV swarm control module of the ground control center extracts the spatial location information of the problem points that have been dealt with, and plans a targeted secondary inspection and verification path for the UAV swarm. The path accurately covers all the problem points that have been dealt with, ensuring that no verification is missed. The drone swarm is controlled to perform secondary inspections along the planned verification path, and multi-angle, high-definition images are collected for each processed problem point. The collected verification images and the original problem point images maintain consistency in the acquisition dimension and resolution, which facilitates subsequent feature comparison. The ground control center inputs the collected verification images into the corresponding feature extraction model, extracts the visual features of the processed problem points according to the above feature extraction method, and generates problem point verification feature information containing the status of the problem points after processing. This information maintains the same feature dimension as the original problem point feature information.

[0052] Step S504: Accurately compare the verification feature information of the problem point with the original feature information of the corresponding problem point. If they are qualified, mark the problem point as closed loop. If they are not qualified, regenerate the scheduling plan. Finally, integrate all the data of this inspection to update the park inspection database and complete the inspection.

[0053] The verification feature information of the problem point is compared with the original feature information of the corresponding problem point at the pixel level and feature level. Based on the preset processing qualification judgment standard, it is judged whether the problem point has been handled properly, such as whether the facility damage problem has been repaired, whether the environmental cleanliness problem has been cleaned up, and whether the safety hazard problem has been completely eliminated. If the feature information comparison results show that the problem point is handled in accordance with the qualified standards, the problem point is marked as closed, and information such as the closure time, handling effect, and maintenance resource usage is recorded; if the comparison results show that the problem point is not handled in accordance with the standards, the problem point is marked as not closed, and the scheduling plan generation process is triggered. Based on the feature information of the problem point that does not meet the standards, a targeted maintenance resource scheduling plan is regenerated until the problem point is handled in accordance with the standards and closed. After all problem points have been closed-loop identified, the ground control center will systematically integrate the data from the entire inspection process, including topology map data, multi-source inspection data, problem point feature information, assessment and grading results, scheduling plans, maintenance operation execution data, and closed-loop verification results. The integrated inspection data will then be updated to the park inspection database in a unified data format, thus completing the inspection process.

[0054] By parsing the optimal scheduling scheme for maintenance resources into standardized scheduling instructions, the standardized and regulated execution of maintenance operations was achieved, resolving the technical pain points of ambiguous instructions and non-standard execution in traditional inspections. Real-time uploading and ground-based verification of maintenance operation data enabled dynamic control of the maintenance process, ensuring the authenticity and traceability of the data. Secondary inspections and verifications were conducted using drone swarms, extracting standardized verification feature information, achieving objective and accurate acceptance of problem-solving effectiveness, avoiding the subjectivity and bias of traditional manual acceptance. This was achieved through a combination of original feature information and verification features. The precise comparison of information completed the closed-loop judgment, forming a closed-loop management mechanism of "problem discovery - scheduling and handling - effect verification - reprocessing of non-compliance", ensuring that the park's problem points are effectively addressed. Finally, by integrating the data from the entire inspection process to update the park's inspection database, the inspection data was systematically accumulated, providing detailed data support for the park's subsequent inspection planning, maintenance resource allocation, and management of high-incidence areas of problem points. At the same time, the entire park inspection process formed a complete link of "scheduling - execution - verification - closed loop - accumulation", which greatly improved the park's closed-loop management level and refined management capabilities.

[0055] It should be noted that the construction steps of the improved convolutional neural network combined with the optimized spatial attention mechanism feature extraction model are as follows: Step a: Construct an improved convolutional neural network base network, and sequentially set up an input layer, a multi-scale convolutional layer, an activation layer, an adaptive average pooling layer, and a feature fusion layer. The multi-scale convolutional layer adopts a combination design of 3×3, 5×5, and 7×7 convolutional kernels and performs parallel convolution operations. The activation layer adopts the ReLU6 activation function. The feature fusion layer uses a concatenation method to integrate multi-scale features. The improved convolutional neural network is built in the order of input layer → multi-scale convolutional layer → activation layer → adaptive average pooling layer → feature fusion layer. Each layer adopts a data pass-through connection to ensure the effective transfer of image features between layers. The multi-scale convolutional layer adopts a combination design of three different sizes of convolutional kernels: 3×3, 5×5, and 7×7. The three types of convolutional kernels perform parallel convolution operations on the input static inspection image. The 3×3 convolutional kernel extracts fine-grained local features of the image, the 5×5 convolutional kernel extracts medium-scale regional features, and the 7×7 convolutional kernel extracts large-scale global features. After parallel operation, multiple sets of feature maps of different scales are output. The activation layer uses the ReLU6 activation function to perform non-linear mapping on the feature maps output by the multi-scale convolutional layer, suppressing invalid information with excessively large feature values, enhancing the model's ability to learn non-linear features, and avoiding the gradient explosion problem, thereby improving the model's stability. Adaptive average pooling is applied to the multi-scale feature map output by the activation layer. This layer can adaptively adjust the pooling kernel size and stride according to the different sizes of the input feature map. While compressing the feature dimension and reducing the amount of computation, it retains the core features of static problem points to the maximum extent and avoids feature loss. The feature fusion layer is pre-configured using feature stitching and reserves a feature interface for the optimized spatial attention mechanism module. It can stitch together and integrate multi-scale features after attention weighting, and finally output a standardized multi-scale basic feature map.

[0056] Step b: Construct an optimized spatial attention mechanism module, which sequentially sets up a channel compression layer, a spatial feature extraction layer, a weight generation layer, and a weight normalization layer. The channel compression layer compresses the feature channel dimension using a 1×1 convolution kernel. The spatial feature extraction layer extracts spatial features by combining pooling and convolution. The weight generation layer generates a spatial attention weight matrix through a fully connected layer. The weight normalization layer normalizes the weight matrix to the 0-1 range using the Sigmoid function. The spatial attention mechanism module is built in the order of channel compression layer → spatial feature extraction layer → weight generation layer → weight normalization layer. Each layer is connected in sequence. Spatial attention features are extracted and weights are generated for the multi-scale feature map output by the basic network. The channel compression layer performs convolution operations on the input multi-scale basic feature map using a 1×1 convolution kernel. Without changing the spatial dimension of the feature map, it completes the compression and fusion of the feature channel dimension, reduces the amount of subsequent computation, and improves the module's computational efficiency. Spatial features are extracted by combining pooling and convolution. First, the feature map after channel compression is subjected to dual pooling operations of global average pooling and global max pooling to capture the spatial distribution information of the feature map. Then, the pooled features are convolved and fused by 3×3 convolution kernels to extract the spatial attention features of static problem points. The attention features output by the spatial feature extraction layer are input into the fully connected layer. Through linear transformation and operation of the fully connected layer, a spatial attention weight matrix with the same spatial dimension as the input feature map is generated. Each value in the matrix corresponds to the attention weight of a pixel in the feature map. The higher the value, the greater the probability that the pixel is a static problem point. The weight normalization layer uses the Sigmoid function to normalize the values ​​in the spatial attention weight matrix to the 0-1 range, eliminating the influence of excessive numerical differences, and making the weight values ​​more consistent with the probability distribution characteristics, which is convenient for subsequent weighted enhancement of the basic feature map.

[0057] Step c: Embed the optimized spatial attention mechanism module between the adaptive average pooling layer and the feature fusion layer of the improved convolutional neural network base network to realize the feature interaction and linkage between the spatial attention mechanism module and the convolutional neural network base network, and complete the overall construction of the feature extraction model of the improved convolutional neural network combined with the optimized spatial attention mechanism.

[0058] The optimized spatial attention mechanism module is precisely embedded between the adaptive average pooling layer and the feature fusion layer of the improved convolutional neural network base network. This position is a key node after the dimensionality reduction of the base feature map and before the multi-scale feature fusion, which allows the attention module to perform weighted enhancement on the core base features after dimensionality reduction. The multi-scale basic feature map output by the adaptive average pooling layer is input into the optimized spatial attention mechanism module. After the module generates the corresponding spatial attention weight matrix, the weight matrix is ​​multiplied pixel by pixel with the original multi-scale basic feature map to achieve weighted enhancement of features in static problem area and suppression of background area features, thus completing the feature interaction and linkage between the basic network and the attention module. The attention-weighted multi-scale feature maps are connected to the feature fusion layer of the base network. The integration of multi-scale enhanced features is completed by feature splicing. At the same time, the overall network parameters of the model are initialized and configured. The overall construction of the feature extraction model combining the improved convolutional neural network with the optimized spatial attention mechanism is completed. The model can be directly connected to the park inspection data processing system to extract features from the preprocessed static inspection images.

[0059] After the preprocessed static inspection image of the park is input into the model, it first passes through the multi-scale convolutional layers of the improved convolutional neural network to extract basic features at scales of 3×3, 5×5, and 7×7 in parallel. Then, it passes through the ReLU6 activation layer for nonlinear mapping and the adaptive average pooling layer for dimensionality reduction. Subsequently, the dimensionality-reduced multi-scale basic feature map is input into the optimized spatial attention mechanism module. After channel compression, spatial feature extraction, and weight generation, a spatial attention weight matrix is ​​obtained. Then, it is normalized to the 0-1 interval by the Sigmoid function. The weight matrix is ​​then multiplied pixel by pixel with the multi-scale basic feature map to complete the feature weighting enhancement. Finally, the weighted multi-scale enhanced features are input into the feature fusion layer for concatenation and integration. After dimensionality compression, a standardized static problem point feature map of the park is output.

[0060] It should be noted that the construction steps of the feature extraction model combining the lightweight YOLO model with time series analysis are as follows: Step d: Lightweight modification of the YOLO base model, replacing the standard convolution of the model backbone network with depthwise separable convolution, reducing the number of detection heads while retaining the core detection branches, and performing adaptive clustering optimization of anchor boxes based on dynamic target sample data in the park to generate a lightweight YOLO target detection sub-model adapted to dynamic target detection in the park. The standard convolution in the backbone network of the YOLO basic model is replaced by depthwise separable convolution. Depthwise separable convolution separates the channel convolution and spatial convolution of the standard convolution. While significantly reducing the number of model parameters and computation, it retains the model's target feature extraction capability, allowing the model to run efficiently on edge devices with limited computing power. The number of detectors in the YOLO base model is reduced, redundant detection branches are eliminated, and only the core detection branches are retained. At the same time, the feature extraction capabilities of the retained detectors are optimized to ensure that the detection accuracy of small-scale dynamic targets in parks is not reduced while reducing the amount of computation. Collect sample data of various dynamic targets (such as pedestrians, non-motorized vehicles, temporary facilities, etc.) in park inspection scenarios. Based on the sample data, use the K-means clustering algorithm to perform anchor box adaptive clustering, generate anchor box parameters that are adapted to the size and shape of dynamic targets in the park, replace the default anchor boxes of the YOLO basic model, and improve the model's detection matching degree and accuracy of dynamic targets in the park. After completing the above modifications and optimizations, the model was initialized, trained, and its parameters were tuned. The trained model was then encapsulated into a lightweight YOLO object detection sub-model. This model can quickly detect objects in a single frame of a dynamic park inspection image and output the feature data of the dynamic object in that single frame.

[0061] Step e: Construct a temporal analysis submodule, which sequentially sets up an inter-frame feature extraction layer, an optical flow operation layer, and a temporal feature fusion layer. The inter-frame feature extraction layer extracts target feature points from consecutive frame images. The optical flow operation layer calculates the inter-frame motion trajectory and displacement data of the feature points. The temporal feature fusion layer uses a feature stitching method to integrate single-frame target features with inter-frame temporal features. The temporal analysis submodule is built in the order of inter-frame feature extraction layer → optical flow operation layer → temporal feature fusion layer. Each layer is connected serially. The inter-frame temporal features are extracted, calculated and fused for the single-frame target features output by the lightweight YOLO target detection sub-model. For the input continuous frame images of dynamic park inspection, the key feature points of the target in each frame are extracted. At the same time, the position and shape information of the target in each frame are associated to complete the matching and localization of target feature points between frames, which provides a basis for subsequent motion trajectory calculation. Optical flow is used to process the target feature points output by the inter-frame feature extraction layer, calculate the motion trajectory, displacement data, motion speed and direction of the feature points in consecutive frames, capture the motion law of dynamic targets, and transform the static target features of a single frame into continuous dynamic behavior features. The temporal feature fusion layer is pre-configured using feature splicing and reserves a feature interface for docking with the lightweight YOLO object detection sub-model. It can splice and integrate the single-frame object features output by the sub-model with the inter-frame temporal features extracted by this module, and finally output fused data containing static and dynamic behavior features of the target.

[0062] Step f: Perform edge-side fusion of the temporal analysis sub-module and the lightweight YOLO object detection sub-model. After feature dimension matching between the output of the lightweight YOLO object detection sub-model and the input of the temporal analysis sub-module, connect them to complete the overall construction of the feature extraction model combining the lightweight YOLO model and temporal analysis.

[0063] The output feature dimension of the lightweight YOLO object detection sub-model and the input feature dimension of the temporal analysis sub-module are standardized and adjusted so that the single-frame dynamic object feature data output by the sub-model can be directly connected to the inter-frame feature extraction layer of the temporal analysis sub-module, achieving seamless connection of the feature data of the two. The output of the lightweight YOLO object detection sub-model and the input of the temporal analysis sub-module are fused on the edge to build a unified feature data transmission interface, so that the single-frame detection results of the detection sub-model can be transmitted to the temporal analysis sub-module in real time, realizing real-time linkage between dynamic object detection and temporal feature analysis. After completing feature dimension matching and end-side interface integration, the overall computational process of the model is debugged and optimized to ensure the computational synchronization and data consistency between the detection sub-model and the time series analysis sub-module. Finally, the overall construction of the feature extraction model combining the lightweight YOLO model with time series analysis is completed. The model can be directly connected to the park inspection data processing system to extract features from the preprocessed dynamic inspection images.

[0064] After the preprocessed continuous frame images of dynamic park inspection are input into the model, they are first processed by a lightweight YOLO target detection sub-model for fast single-frame target detection. The core features of dynamic targets in each frame are extracted and the single-frame dynamic image target feature data is output. Then, the data is transmitted to the temporal analysis sub-module in real time. The inter-frame feature extraction layer extracts the target feature points of continuous frames and completes the matching. Then, the optical flow method operation layer calculates the temporal behavior features of the feature points, such as inter-frame motion trajectory and displacement data. Finally, the temporal feature fusion layer concatenates and fuses the single-frame dynamic image target feature data with the inter-frame temporal behavior feature data. After vectorization encoding and dimensionality normalization, a standardized feature vector of dynamic problem points in the park is output.

[0065] In one embodiment of the present invention, the operation control of each functional sub-cluster of the UAV swarm is performed according to the following steps: Step S205: Based on the differentiated inspection path planning scheme with obstacle avoidance nodes, set standardized operating parameters for flight altitude, speed and image acquisition interval for each functional sub-cluster according to the inspection area type; Based on the spatial characteristics and data collection needs of each inspection area, the parameter configuration principles are determined. For fine inspection areas, low-height, slow-speed, and high-frequency data collection is required to obtain clear images of small problem points. For large-area inspection areas, moderate-height, medium-speed, and gridded data collection is required to achieve full coverage. For dynamic monitoring areas, low-to-medium height, medium-speed, and ultra-high-frequency data collection is required to capture the continuous behavioral characteristics of dynamic targets. For the precision inspection sub-cluster, set a flight altitude of 3-5 meters, a flight speed of 1-2 meters per second, and an image acquisition interval of 0.5 seconds per image; for the large-area inspection sub-cluster, set a flight altitude of 8-12 meters, a flight speed of 3-5 meters per second, and collect images at equal intervals along a gridded path with an overlap rate of no less than 30% between adjacent images; for the dynamic monitoring sub-cluster, set a flight altitude of 4-6 meters, a flight speed of 2-3 meters per second, and an image acquisition interval of 0.2 seconds per image. The standardized operational parameters are packaged into an operational parameter configuration scheme, which is then distributed to the flight control modules of each functional sub-cluster through the UAV cluster control module of the ground control center. At the same time, the scheme is kept on file at the ground control center to achieve unified management and control of the operational parameters of each UAV.

[0066] Step S206: Each functional sub-cluster performs inspection operations according to standardized operating parameters. The obstacle avoidance sensor detects personnel and obstacles within the flight path range in real time and outputs real-time obstacle avoidance monitoring data. The ground control center issues an inspection operation start command. The flight control modules of each functional sub-cluster receive standardized operation parameters and flight path information, control the UAV to fly along the planned path at the preset altitude and speed, and at the same time activate the high-definition visible light camera at the collection interval to complete image data collection, and synchronously associate auxiliary data such as GPS positioning and shooting time. Each drone is equipped with a millimeter-wave obstacle avoidance sensor that is in operation throughout the entire process. The sensor scans the three-dimensional space within the drone's flight path in real time at a preset detection frequency. The detection targets include fixed obstacles such as trees, buildings, and facilities, as well as moving targets such as pedestrians and non-motorized vehicles. The sensor obtains information such as the spatial position of obstacles / moving targets, their distance from the drone, and their speed in real time. The millimeter-wave obstacle avoidance sensor performs structured processing on the obstacle avoidance-related information detected in real time, generating real-time obstacle avoidance monitoring data that includes the type of detected object, spatial coordinates, distance threshold, and movement status. This data is transmitted to the UAV's local flight control module on one hand, and back to the UAV cluster control module at the ground control center in real time on the other hand, realizing dual monitoring of obstacle avoidance monitoring data both locally and remotely.

[0067] Step S207: If the real-time obstacle avoidance monitoring data triggers a safety warning, the flight control module performs dynamic path adjustment, controls the UAV to fly around the obstacle and updates the inspection path, and outputs the adjusted UAV inspection path. Both the ground control center and the UAV's local flight control module are preset with obstacle avoidance safety warning thresholds. When real-time obstacle avoidance monitoring data shows that the distance between the obstacle / moving target and the UAV is less than the preset safe distance, or when the moving target is approaching the direction of the UAV's flight path, a flight safety warning is immediately triggered, and the image acquisition operation of the UAV is suspended. After receiving a safety warning signal, the flight control module calls the built-in drone path detour algorithm, combines real-time obstacle avoidance monitoring data and surrounding geographical information, and plans a temporary flight path to avoid obstacles / moving targets. The temporary path must follow the principles of "shortest detour distance, no deviation from the original inspection area, and compliance with the park's no-fly requirements", while ensuring that the adjusted path can still cover subsequent inspection points. After the flight control module completes the temporary path planning, it generates the adjusted UAV inspection path. On the one hand, it controls the UAV to fly along the new path, and on the other hand, it transmits the adjusted inspection path back to the UAV cluster control module in the ground control center in real time. The ground control center records and visualizes the adjusted path, and updates the overall inspection path planning scheme of the UAV.

[0068] Step S208: Based on the adjusted drone inspection path, control the corresponding drone sub-cluster to perform image completion collection on the uncollected areas, remove duplicate data, and output complete multi-source data of park inspection.

[0069] The UAV cluster control module at the ground control center compares and analyzes the adjusted UAV inspection path with the original planned path. Combined with the image acquisition records of the UAV, it accurately identifies the areas where image acquisition was not completed during the path adjustment process and generates the spatial coordinates and acquisition requirements of the unacquired areas. The ground control center issues a completion acquisition command to the corresponding UAV, controlling the UAV to complete image acquisition along the planned completion path according to the spatial coordinates of the unacquired area and the acquisition requirements. The operation parameters for completion acquisition still follow the standardized operation parameters of the sub-cluster to which the UAV belongs, ensuring the consistency between the completed data and the original acquired data. The UAV will transmit the supplementary image data and auxiliary data back to the ground control center. The ground control center will screen all the inspection data collected by the UAV, remove invalid data that was repeatedly collected during the path adjustment process, and integrate the valid original data and supplementary data in a structured way, classifying and labeling them according to inspection area, collection time, and spatial coordinates. After data filtering and integration, the inspection data of this drone is summarized with the inspection data of other drones that have not undergone path adjustment, and finally outputs complete multi-source data of park inspection. This data will serve as the core input data for image preprocessing.

[0070] This invention integrates technologies such as geographic information processing, UAV swarm scheduling, and deep learning to construct a comprehensive technical solution for park inspection. By building a refined digital inspection topology map, combined with geographic hierarchical labeling, inspection area classification, and a problem point tag library, it provides accurate geographic information and standard tag support for the entire inspection process. Relying on differentiated path planning by UAV swarms, coupled with standardized operational parameter configuration, real-time obstacle avoidance monitoring and dynamic path adjustment, and image completion acquisition, it achieves secure collection, real-time transmission, and seamless integration of multi-source inspection data, ensuring flight safety and inspection efficiency for UAVs in densely populated and complex park environments. By constructing a dual deep learning feature extraction model, it accurately extracts, matches, and fuses static and dynamic problem point features in the park, effectively improving the accuracy of problem point identification. Simultaneously, through multi-dimensional comprehensive evaluation of problem points and feature coupling with real-time maintenance resource information, it generates a dynamically adjustable optimal scheduling scheme, achieving scientific allocation and efficient scheduling of maintenance resources. Then, through standardized scheduling command issuance and secondary UAV inspection verification, a complete inspection loop is formed and the inspection database is updated, ensuring effective problem point handling. The overall solution effectively solves the problems of low efficiency and numerous blind spots in traditional manual inspections, and inaccurate identification and unreasonable resource allocation in conventional drone inspections. It comprehensively improves the intelligence and precision of park inspections, accumulates structured data for the long-term maintenance and management of parks, adapts to the daily and emergency inspection needs of various parks, and significantly reduces the workload and management costs of manual inspections.

[0071] See Figure 2 The present invention also provides a park inspection system based on UAV collaborative scheduling, including a distributed multi-UAV cluster data acquisition terminal, a ground control center processing terminal and a park maintenance and management terminal, which establish a two-way data interaction connection through a 5G wireless communication network; The distributed multi-drone cluster data acquisition terminal consists of drone sub-clusters for refined inspection, large-area inspection, and dynamic monitoring. Each drone is equipped with a high-definition visible light camera, GPS positioning module, millimeter-wave obstacle avoidance sensor, flight control module, 5G wireless communication module, and local image storage module, which are used to realize multi-source data acquisition for park inspection, flight obstacle avoidance, and real-time data transmission. The distributed multi-drone swarm acquisition terminal is the front-end data acquisition carrier of the entire inspection system and the main body for drone inspection operations. It consists of three types of drone sub-swarms: fine-grained inspection, large-area inspection, and dynamic monitoring. Each sub-swarm cooperates according to the operational requirements of different inspection areas in the park. Each drone is equipped with a high-definition visible light camera, GPS positioning module, millimeter-wave obstacle avoidance sensor, flight control module, 5G wireless communication module, and local image storage module. The functions of each hardware module work together. The high-definition visible light camera and GPS positioning module work together to complete the correlation acquisition of image data and geographic information. The millimeter-wave obstacle avoidance sensor and flight control module realize real-time obstacle avoidance and dynamic path adjustment during flight. The 5G wireless communication module ensures high-speed real-time transmission of multi-source inspection data. The local image storage module realizes local backup of the acquired data, effectively avoiding data loss during transmission. This terminal can complete the acquisition of multi-source inspection data of the entire area without omission according to standardized operating parameters, providing high-quality raw inspection data for the back-end processing terminal.

[0072] The ground control center processing terminal is an industrial-grade server cluster, integrating topology map construction, path planning, data preprocessing, dynamic and static dual feature extraction and analysis, problem assessment and classification, maintenance resource scheduling, closed-loop verification and UAV swarm control function modules. It has built-in improved convolutional neural network combined with spatial attention mechanism and lightweight YOLO combined with time series analysis feature extraction model, which is used to complete the data processing, assessment and scheduling and UAV operation management of the entire inspection process. The ground control center is the core computing, control, and data processing hub of the entire inspection system. Built with an industrial-grade server cluster, it boasts powerful data processing and computing capabilities. It integrates eight functional modules: topology map construction, path planning, data preprocessing, dynamic and static dual feature extraction and analysis, problem assessment and grading, maintenance resource scheduling, closed-loop verification, and UAV swarm control. It also incorporates an improved convolutional neural network combined with spatial attention mechanisms and a lightweight YOLO combined with temporal analysis dual feature extraction model. Each module and the built-in model work collaboratively to cover the technical processing needs of the entire park inspection process. It can process multi-source inspection data transmitted from the acquisition end throughout the entire process, completing problem feature extraction, assessment and grading, and optimal maintenance resource scheduling. It can also control the operation of UAV swarms, issuing various commands such as flight, data collection, and obstacle avoidance. It is a crucial hub connecting the acquisition end and the management end, ensuring the intelligent and automated advancement of the entire inspection process.

[0073] The park maintenance management terminal includes mobile terminals for maintenance personnel and fixed terminals for the park management office, which are used to receive dispatch instructions, upload maintenance operation data, manage maintenance resources, and visualize inspection data and trace historical data.

[0074] The park maintenance management terminal serves as the command interaction, maintenance operation execution, and data management terminal for the entire inspection system. It consists of mobile terminals used by maintenance personnel and fixed terminals used by the park management office. The two terminals achieve real-time data synchronization, forming a collaborative system of front-end operation execution and back-end comprehensive management. The mobile terminals are the direct information interaction carrier for maintenance personnel, mainly realizing functions such as receiving dispatch instructions, uploading maintenance operation progress, and uploading on-site image data, enabling maintenance personnel to accurately receive operation tasks and provide real-time feedback on operation status. The fixed terminals are the back-end management carrier for park inspection and maintenance work, mainly realizing functions such as maintenance resource ledger management, inspection data visualization, and historical inspection data traceability query, providing data support for park management to allocate maintenance resources, review inspection work, and make management decisions. This terminal is a key link in the implementation of the inspection closed loop, ensuring the effective execution of maintenance operations and comprehensive management of inspection data.

[0075] This invention establishes an integrated architecture via a 5G wireless communication network, enabling bidirectional data interaction between a distributed multi-drone swarm data acquisition end, a ground control center processing end, and a park maintenance and management end. Each end has a clearly defined function and works in close coordination. The acquisition end, with multiple types of sub-clusters and specialized hardware modules, achieves secure acquisition, real-time transmission, and local backup of multi-source data from park inspections, ensuring the efficiency and integrity of front-end data acquisition. The ground control center processing end, relying on an industrial-grade server cluster, integrates full-process functional modules and incorporates a dedicated dual deep learning feature extraction model, enabling intelligent processing of inspection data, accurate identification of problem points, and unified operational control of the drone swarm. As the core computing and scheduling hub of the system, the park maintenance management terminal, through the cooperation of mobile and fixed terminals, realizes the rapid issuance of scheduling instructions, real-time uploading of maintenance operation data, and systematic management of maintenance resources and inspection data. This makes the inspection closed loop more efficient. The overall system realizes the intelligentization and standardization of the entire process of park inspection from data collection, analysis and processing to maintenance execution and data management. It is effectively adapted to the inspection scenarios of densely populated and complex environments in parks, greatly improving the efficiency of park inspection operations and the level of precision in maintenance management. At the same time, it realizes the full-link flow and traceability of inspection data, providing stable and reliable system support for the long-term maintenance management of parks.

[0076] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A park inspection method based on UAV collaborative scheduling, characterized in that, Includes the following steps: Obtain the park's basic geographic base map, then perform geographic layer labeling, inspection area classification, and problem point label library construction in sequence, and output a refined digital inspection topology map of the park; Based on the inspection topology map, the system controls the drone swarm to collect multi-source inspection data and transmit it to the ground control center through inspection point positioning and differentiated path planning. Based on multi-source inspection data, after preprocessing, the features of dynamic and static problem points in the park are extracted and fused through a dual deep learning model to output the feature information of real problem points. Based on feature information, a multi-dimensional assessment and classification of problem points is completed, and the optimal scheduling scheme is generated by integrating maintenance resource information. The optimal scheduling scheme is implemented and a closed loop is formed through secondary inspection and verification by drones. The park inspection database is then updated to complete the inspection.

2. The park inspection method based on UAV collaborative scheduling according to claim 1, characterized in that, The specific steps for outputting the refined digital inspection topology map of the park are as follows: Based on the park's basic geographic base map, accurate annotation of three layers of geographic information, including topographic layer, facility layer and control layer, is performed to output a park base map with layered geographic information. Based on the park base map with layered geographic information, the park is divided into fine inspection area, large area inspection area and dynamic monitoring area according to the inspection accuracy requirements, and the park base map with inspection area classification is output. Construct a pre-classification tag library for park problem points that includes static and dynamic problem point types and labels them with urgent, general, and minor handling priorities, and output a standardized park problem point tag library; By integrating park base maps with inspection area classifications and a standardized park problem point label library, a refined digital inspection topology map of the park is generated.

3. The park inspection method based on UAV collaborative scheduling according to claim 1, characterized in that, The specific steps for collecting multi-source inspection data are as follows: Based on the inspection topology map, combined with the daily inspection data of the park and the distribution pattern of historical problem points, the key inspection areas are located and real-time dynamic monitoring points are set up, and the park inspection topology map with inspection point markings is output. Based on the park inspection topology map with inspection point markings, the drone swarm is divided into functional sub-clusters and flight paths matching each inspection area are planned. Obstacle avoidance and human avoidance nodes are embedded in the paths, and differentiated inspection path planning schemes with obstacle avoidance nodes are output. Based on the differentiated inspection path planning scheme, the system controls each functional sub-cluster to perform inspection operations, collects image data and corresponding auxiliary data such as positioning, shooting time, and flight altitude, and outputs multi-source raw inspection data. The raw data from the multi-source inspection is transmitted to the ground control center in real time via wireless communication. After data integrity verification, valid multi-source inspection data is output.

4. The park inspection method based on UAV collaborative scheduling according to claim 1, characterized in that, The specific steps for outputting the feature information of the actual problem points are as follows: Based on multi-source inspection data, the image data is preprocessed by image enhancement, deblurring and segmentation along the inspection path in sequence. Blurry and overexposed invalid image data are removed, and the preprocessed valid inspection image data and supporting auxiliary data are output. Based on effective inspection image data, static images are input into an improved convolutional neural network combined with an optimized spatial attention mechanism feature extraction model to extract features of static problem points and generate feature maps, outputting the feature map of static problem points in the park. Based on effective inspection image data, dynamic images are input into a lightweight YOLO model combined with a time-series analysis feature extraction model to extract features of dynamic problem points and generate feature vectors, outputting the feature vectors of dynamic problem points in the park. The feature map of static problem points in the park and the feature vector of dynamic problem points in the park are matched for dimension and fused. Duplicate and invalid feature information is removed and the feature information of real problem points in the park is output.

5. The park inspection method based on UAV collaborative scheduling according to claim 1, characterized in that, The specific steps for generating the optimal scheduling scheme are as follows: Input real problem point feature information, accurately match it with a standardized park problem point label library, complete a comprehensive evaluation from four dimensions, assign feature labels to each problem point and visualize them on the topology map, and output the multi-dimensional evaluation and classification results of park problem points; The system acquires real-time distribution information of park maintenance personnel, tools, and material reserves, as well as accessibility information of walkways and work passages within the park. After data integration, it outputs accessibility information of maintenance resources. The multi-dimensional assessment and grading results of park problem points and the accessibility characteristics of maintenance resources are input into the feature coupling module to complete the deep feature fusion and output the fused feature information of park problem points and maintenance resources. Based on the fusion of feature information, maintenance resources are matched from high to low according to processing priority and the optimal passage path is planned. The output is an optimal scheduling scheme for park maintenance resources that can be dynamically adjusted according to the real-time status of the park.

6. The park inspection method based on UAV collaborative scheduling according to claim 1, characterized in that, The specific steps for updating the park inspection database are as follows: The optimal scheduling plan for park maintenance resources is parsed into standardized scheduling instructions that include problem point information, handling requirements, supporting resources, and access routes, and then sent to the park maintenance management terminal via wireless communication. Maintenance personnel execute maintenance operations according to standardized dispatch instructions, upload work progress and on-site image data to the ground control center in real time, and output maintenance operation execution result data after data verification; Based on the maintenance operation execution results data, control the drone cluster to perform secondary inspection and verification of the processed problem points, collect verification images and extract verification feature information, and output the problem point verification feature information; The verification feature information of the problem point is accurately compared with the original feature information of the corresponding problem point. If it is qualified, the problem point is marked as closed loop. If it is not qualified, the scheduling plan is regenerated. Finally, all the data of this inspection are integrated to update the park inspection database and complete the inspection.

7. The park inspection method based on UAV collaborative scheduling according to claim 4, characterized in that, The construction steps of the improved convolutional neural network combined with the optimized spatial attention mechanism feature extraction model are as follows: An improved convolutional neural network is constructed, consisting of an input layer, a multi-scale convolutional layer, an activation layer, an adaptive average pooling layer, and a feature fusion layer. The multi-scale convolutional layer uses a combination of 3×3, 5×5, and 7×7 convolutional kernels and performs parallel convolution operations. The activation layer uses the ReLU6 activation function. The feature fusion layer integrates multi-scale features through a concatenation method. An optimized spatial attention mechanism module is constructed, which sequentially sets up a channel compression layer, a spatial feature extraction layer, a weight generation layer, and a weight normalization layer. The channel compression layer compresses the feature channel dimension through a 1×1 convolution kernel. The spatial feature extraction layer extracts spatial features by combining pooling and convolution. The weight generation layer generates a spatial attention weight matrix through a fully connected layer. The weight normalization layer normalizes the weight matrix to the 0-1 range through the Sigmoid function. The optimized spatial attention mechanism module is embedded between the adaptive average pooling layer and the feature fusion layer of the improved convolutional neural network base network, realizing feature interaction and linkage between the spatial attention mechanism module and the convolutional neural network base network, and completing the overall construction of the feature extraction model of the improved convolutional neural network combined with the optimized spatial attention mechanism.

8. The park inspection method based on UAV collaborative scheduling according to claim 4, characterized in that, The construction steps of the feature extraction model combining the lightweight YOLO model with time series analysis are as follows: The YOLO base model is lightweighted by replacing the standard convolutions in the backbone network with depthwise separable convolutions, reducing the number of detection heads while retaining the core detection branches. At the same time, anchor box adaptive clustering optimization is performed based on dynamic target sample data in parks to generate a lightweight YOLO target detection sub-model adapted to dynamic target detection in parks. A temporal analysis submodule is constructed, which sequentially sets up an inter-frame feature extraction layer, an optical flow calculation layer, and a temporal feature fusion layer. The inter-frame feature extraction layer extracts target feature points of consecutive frame images. The optical flow calculation layer calculates the inter-frame motion trajectory and displacement data of the feature points. The temporal feature fusion layer integrates single-frame target features and inter-frame temporal features using a feature stitching method. The temporal analysis submodule and the lightweight YOLO object detection submodel are fused on the edge. The output of the lightweight YOLO object detection submodel is matched with the input of the temporal analysis submodule by feature dimension matching, thus completing the overall construction of the feature extraction model that combines the lightweight YOLO model with temporal analysis.

9. The park inspection method based on UAV collaborative scheduling according to claim 3, characterized in that, The operation control of each functional sub-cluster of the UAV swarm is performed according to the following steps: Based on the differentiated inspection path planning scheme with obstacle avoidance nodes, standardized operating parameters such as flight altitude, speed and image acquisition interval are set for each functional sub-cluster according to the inspection area type. Each functional sub-cluster performs inspection operations according to standardized operating parameters, and obstacle avoidance sensors detect personnel and obstacles within the flight path range in real time, outputting real-time obstacle avoidance monitoring data; If real-time obstacle avoidance monitoring data triggers a safety warning, the flight control module will dynamically adjust the path, control the drone to fly around the obstacle and update the inspection path, and output the adjusted drone inspection path. Based on the adjusted drone inspection path, the corresponding drone sub-clusters are controlled to complete the image acquisition of uncollected areas. After removing duplicate data, the park inspection multi-source data without omissions is output.

10. A park inspection system based on UAV collaborative scheduling, characterized in that, The system includes a distributed multi-drone cluster data acquisition terminal, a ground control center processing terminal, and a park maintenance and management terminal, which are connected to each other via a 5G wireless communication network for bidirectional data interaction. The distributed multi-drone cluster data acquisition terminal consists of drone sub-clusters for refined inspection, large-area inspection, and dynamic monitoring. Each drone is equipped with a high-definition visible light camera, GPS positioning module, millimeter-wave obstacle avoidance sensor, flight control module, 5G wireless communication module, and local image storage module, which are used to realize multi-source data acquisition for park inspection, flight obstacle avoidance, and real-time data transmission. The ground control center processing terminal is an industrial-grade server cluster, integrating topology map construction, path planning, data preprocessing, dynamic and static dual feature extraction and analysis, problem assessment and classification, maintenance resource scheduling, closed-loop verification and UAV swarm control function modules. It has built-in improved convolutional neural network combined with spatial attention mechanism and lightweight YOLO combined with time series analysis feature extraction model, which is used to complete the data processing, assessment and scheduling and UAV operation management of the entire inspection process. The park maintenance management terminal includes mobile terminals for maintenance personnel and fixed terminals for the park management office, which are used to receive dispatch instructions, upload maintenance operation data, manage maintenance resources, and visualize inspection data and trace historical data.

Images